Skip to content

Changelog

Every ComplyEaze Bridge release, in the order it shipped, plus what is already in source and not yet published.

What changed

In source, not yet published #

Published builds are MCPB packages that are not yet code-signed (tags mcp-preview-* and, from 0.4.0, mcp-v*): so far mcp-preview-0.2.0, mcp-preview-0.3.0, mcp-v0.4.0, mcp-v0.4.1, mcp-v0.4.2, mcp-v0.5.0, mcp-v0.5.1, mcp-v0.6.0, mcp-v0.6.1 and mcp-v0.7.0. The number of the next build is chosen when it is released. The version boundary between the published MIT-licensed v0.1.0 release and Apache-2.0 builds from current source stays unambiguous.

0.7.0, #

In plain words: ComplyEaze Bridge 0.7.0, since 0.6.1 (27 Oct 2026)

These changes are in ComplyEaze Bridge 0.7.0. Most lines name the pull requests or issues they come from. Like every build so far, it is not yet code-signed. The desktop app has no published installer, so lines about it apply only to a build from source. Sales invoice posting is not available in 0.7.0, whatever the posting setting: a Sales invoice is refused before anything is sent until the checks its lab rehearsal left owed are done (#2162), so the lines about invoices apply only once it is offered.

Should I upgrade?

  • If you read stock, statements or a ledger's movement: yes. stock_summary now reads as of the last day of any month, not only 31 March, and gives the quantity of an item inside a stock group (#1233). ledger_movement reads only the named ledger's vouchers, so a large book no longer fails as too large for a ledger with a small share of it (#485). A Trial Balance row that does not add up is marked, and profit_and_loss and balance_sheet name its ledger and gap when they refuse (#2068, #692).
  • If you use posting: what Bridge writes into Tally is still only what you approve in the approval window, and the kinds of voucher it posts are the same as in 0.6.1 (Journal, Payment, Receipt and Contra). On Windows the approval boxes for a post and for a ledger create now stay above Tally when you click into Tally while one is open; this was seen with the server started by the lab's own test client, not yet with Claude Desktop as its parent, and the review, release and settle boxes share the change but were not shown on a screen, and every box shows dates as Tally does, for example 14-Sep-2026 (#1780, #2101). A post's answer that reported STATUS 0 no longer counts toward the voucher mark it proved (#2035), and a date that names no calendar day is refused in four checks (#2009).
  • If you want Bridge to create a ledger: it can now create one ledger at a time, through its own approval box, only when BRIDGE_AGENT_ENABLE_MASTER_WRITES and posting are both on. The desktop extension has no setting for the first, so a default install does not offer it (#690).
  • Before you upgrade: once 0.7.0 has recorded a person's check of a post matched by content only (acknowledge_post_review with doubt content_only), or a person's settlement of a post Tally answered as creating none of its vouchers (doubt not_created), do not go back to 0.6.1: an older build refuses the whole import journal once either record is written (#1039, #1709).
  • Everyone else: nothing else needs doing. Install steps are the same as for 0.6.1: installation guide; after installing, check the extension shows 0.7.0, the Terms setting is ticked and the posting setting is as you want it; quit Claude Desktop completely and reopen it.

What you can do now

  • prepare_ledger_creates and post_ledger_creates create ONE new ledger in a company, the first write a published build makes to Tally's masters (part of #690). They are listed only when BRIDGE_AGENT_ENABLE_MASTER_WRITES and posting (BRIDGE_AGENT_ENABLE_WRITES) are both set to true; the desktop extension has no setting for the first, so a default install does not offer them. The ledger is a party under Sundry Debtors or Sundry Creditors, a Suspense ledger, or a GST duty ledger whose head has been measured, from typed fields; ComplyEaze Bridge builds the request itself, accepts no XML, and can set no opening balance or parent. prepare_ledger_creates reads the company's ledger and group names, refuses a name that already exists as a ledger, a group or an alias or that differs from one only in case, spaces or symbols (Tally can take such a Create for the existing ledger and change it), writes nothing to Tally, and returns a batch_id and the preview the person will be shown (its ledger name masked under mask_parties). post_ledger_creates asks the local user in a native dialog and sends the create once; a batch already sent is never sent again. After the send the book is read back and one state is returned, posted_verified only when Tally's counters show one create and no alteration, the company's own master mark moved by exactly one (on a later call, the ledger is the first master change after the last read before the send), exactly one new ledger of the name exists with the parent, tax fields and GSTIN as sent, no ledger or group that existed before changed, and no other loaded company's master mark moved (bill-wise is not read back). Measured: one ledger created and read back as posted_verified through the tools on a synthetic company (TallyPrime 7.1 Silver, 14 Oct 2026), and the approval box shown whole, declined with nothing sent, accepted and then answered from the journal with no second box (TallyPrime 7.2, a continuous-integration package of an earlier head of the change, 239896f8, 28 Oct 2026). Not run: a create whose answer is lost, and other TallyPrime releases and editions; the build of this release ran it once (see the record of live runs below). A name Tally resolves to an existing ledger is reported loudly and is not retried.
  • A ledger create with no answer from Tally on file (it could not be recorded, or Tally never answered) can now settle. A later post_ledger_creates call for the batch answers posted_unattributed when it finds the ledger as sent and the ledger's ALTERID shows it as the first master change in the company after ComplyEaze Bridge's last read of it before sending the create, and not changed since. That releases the ledger name the batch held. It is never posted_verified: with no answer from Tally on file, nothing shows the change was ComplyEaze Bridge's own create rather than the same ledger made by someone else, so the answer asks the person to check the ledger in Tally. A ledger changed since stays open with the reason ledger_altered_since_create (#2003).
  • acknowledge_post_review with doubt set to content_only records, after the native review dialog, that a person checked in Tally a native post matched by content only: its binding to its own vouchers was refused, or no answer from Tally was recorded. The record changes nothing in Tally and no status or verdict, binds no voucher, lifts no invoice stop and releases no row; verify_import reports it as content_review, and local_data_report counts the batch as not_settled_reviewed_not_attributed. A later posted_verified, a later answer from Tally or a rollback of the book voids it. It is a new kind of import journal record: a build older than this one refuses the whole journal once one is written, so do not downgrade after recording one (part of #1039).
  • acknowledge_post_review with doubt set to not_created lets a person settle a native post that Tally answered as creating none of its vouchers, in its own native dialog, once they say each voucher is in the book or will not be posted. ComplyEaze Bridge never sends those vouchers again. The record changes nothing in Tally and no status or verdict; the rows stay held, so they are never built or posted again, and a Sales invoice keeps its stop until the separate release. verify_import reports it as person_settlement, and local_data_report counts the batch as settled_by_person, outside not_settled. A later posted_verified, a later answer from Tally, a binding or a rollback of the book voids it. It is a new kind of import journal record: a build older than this one refuses the whole journal once one is written, so do not downgrade after recording one (part of #1709). On Windows its box opens system-modal like the other approval boxes. Before a settlement, its dialog counts only the matching vouchers that are in the accounts; after one, a voucher's next step no longer asks for the settlement, and says not to import the voucher's file through Tally's Import menu.
  • gst_by_counter_ledger totals, for a period, the GST the books record on ledgers under Duties & Taxes per counter ledger (the one Purchase Accounts, Direct Expenses, Indirect Expenses or Fixed Assets ledger on a purchase, debit note, journal or payment with a supplier on it, named with its GUID), debit and credit side apart, per head, and sets every other voucher with GST aside under a named reason with its own totals (a voucher with two such ledgers, a reverse-charge or set-off journal, a sale or credit note, a tax payment with no supplier, a duty ledger not typed GST, TDS or TCS, a ledger under another group). Books only: not a return figure, and no eligibility or availment decision. Measured on fourteen vouchers typed on TallyPrime 7.2's screen in a synthetic book, item invoices among them. The same synthetic book, after its company and one supplier were registered and December's vouchers keyed, was read with an earlier build of this tool that did not yet admit Direct Expenses or Fixed Assets ledgers or name the ledger GUID: none of December's six vouchers was set aside, and a supplier invoice entered twice (the same reference, date and bill) was totalled twice with nothing flagged, so the answer does not detect a duplicate invoice. Gold, other releases and a registered book on the build of this release are not measured (part of #1866).
  • vouchers takes bank_allocations: true to add, on each ledger entry, the bank date, instrument number and date, transaction type and status of its bank allocations, each field saying whether Tally sent it absent, empty or with a value; a transfer reference typed in a Contra's bank allocation is its instrument number. status is Tally's own STATUS element, not a reconciliation state, and nothing returned says whether an entry is reconciled. On two entries reconciled on Tally's screen with a bank date of 6 April 2026, Tally's raw answer filled the bank date and set BANKMANUALSTATUS and BANKRECONSTATUS to Reconciled, while STATUS stayed No; this read returns the bank date but neither of those two elements. The built tool has not been run against a live Tally on a reconciled entry; its parser, run on the saved raw answer, read the bank date (#2006). Without the argument, answers are as before. Measured on one synthetic book of five hand-keyed Contras (TallyPrime 7.2, keyed in double entry), and, for the reconciled entries, by a raw read of five Contras of a second synthetic book (TallyPrime 7.2); no request changed (#1231, #1392).
  • vouchers takes goods_lines: true to add, on each ledger entry of an invoice voucher, its goods lines: the stock item, actual and billed quantity, rate and amount, and each godown and batch row, each field saying whether Tally sent it absent, empty or with a value. An entry with a line whose batch amounts do not add up to its amount shows no lines, and says why. A Stock Journal showed no lines in the one read of it (TallyPrime 7.2, not committed); a Delivery Note or Receipt Note has not been captured. Without the argument, answers are as before. Measured on Sales and Purchase invoices of two synthetic books (TallyPrime 7.1, one also on 7.2); no request changed (#2015).
  • vouchers takes summarise_by: cost_centre to sum a window's cost-centre allocations into one bucket per cost centre under each cost category, with each category's own totals (category_totals: a category allocates the whole entry, so categories overlap) and the entries that carry no allocation (entries_unallocated). An entry whose allocations cannot be read, or whose allocations under one category do not add up to the entry, refuses the summary (summary_cost_centre_unreadable, summary_cost_centre_allocation_mismatch). The summary does not read the company's Cost Centres setting: allocations were read on a book with it off, so an empty summary means nothing was allocated in the window. Measured on the saved raw answers of one synthetic book with Cost Centres on (TallyPrime 7.2): a Payment allocated to two centres of one category and a Journal entry allocated in full under two categories; the summary was run through the tool once, on the build of this release (see the record of live runs below); no request changed (#1231).
  • A party's mailing address, PIN code, state and country can now be read from Tally's dated mailing details (LEDMAILINGDETAILS.LIST) (#1490). On the measured book (one synthetic company, TallyPrime 7.1 Silver, 12 Oct 2026) the flat address, PIN code and state fields came back empty or absent for a party that held an address, and the dated list held it: two address lines, the PIN code, the state, the country, the date it applies from and the mailing name. ledger_masters fields=compliance takes include_mailing_details (default false): without it no request names the list and no row has the key; with it each row carries the dated entries as compliance.mailing_details, oldest first. It is refused together with include_gst_details, as the request naming both has not been captured. Under mask_parties the mailing name, every address line and the flat address field are shortened like a party's name, to their first two and last two characters; the PIN code, state and country are returned under every redaction mode. A second mailing entry with a later date was read on TallyPrime 7.2 (one party, oldest first); three or more entries, an entry with no date and other releases are not measured. The registers do not read the mailing details and keep their request and their estimate.
  • The party/ledger master workbook, a local file written by the desktop app, gains five columns at its end: the mailing address, PIN code and state in force on the Balances-as-of date, the date they apply from, and a status that says why they are blank (no mailing details, none in force on that date, not observed, or unreadable). The workbook asks Tally for the mailing details only for a book whose master-alteration mark (an upper bound on its ledgers) is at most 3,200; for a larger book it sends the same request as before, every ledger's status reads not observed, and a note in the workbook's header says so, so a book that exported before is never refused for the larger per-ledger estimate. For a book past that mark the five columns are still there but empty, and every status reads not observed: the choice is made on the mark, which counts every master alteration, so a book with few ledgers and many alterations is read without the mailing details too (#1490).
  • stock_summary now reads stock as of the last day of any month, not only 31 March. The period runs from 1 April of that financial year, or the book's start if later; any other date is refused as stock_summary_as_of_not_measured, as before. On two lab books on TallyPrime 7.2, month ends inside the year matched Tally's own Stock Summary, items inside stock groups included; a voucher dated on the end date was counted on a 15 October read, and a month end with a voucher on it was not measured. A company whose financial year does not start on 1 April is still read from 1 April (part of #1233).
  • stock_summary now returns the closing quantity of a stock item inside a stock group. Tally's plain Stock Summary has no line for such an item, so the quantity is the item's own in the stock item list and is not compared: the item says inside_stock_group, with item_list_amount and item_list_unit, and is never agreed. On one lab book with groups three deep, each such quantity equalled Tally's Stock Summary screen with every level expanded. An item whose parent reads Primary without Tally's root marker now says parent_not_resolved and carries no quantity, and checks.closing_quantity is checked_where_agreed (part of #1233).
  • ledger_movement with a ledger now reads only that ledger's vouchers for the period, instead of every voucher in the book, so a large book no longer fails as too large for a ledger that holds a small share of it. The window is planned from a count of the ledger's vouchers, and the ledger is taken from the book's own spelling of its name. The result says how it was read (window_scope: ledger_scoped, with the vouchers read, or whole_book with the reason). When the ledger's name has a character the scoped read does not send or holds a word Tally's formulas read as an operator (such as "and" or "not"), when a day of its vouchers is too heavy for one read, or when the scoped read finds nothing, the whole book's vouchers for the period are read as before and the result says so. A returned voucher with no line on the ledger refuses the read. The lab has compared the scoped read with the Trial Balance for one ledger and one month only, where its debit matched; ledger_movement as a whole has not yet been read on a Tally with this change (part of #485).
  • The outstandings party detail can now answer for a party on a book whose window is too large to read. Where the read of the whole company's vouchers is refused as needing too many requests, a party matched to one ledger alone, whose vouchers carry the ledger's name as stored and whose name holds only ASCII letters, digits and single spaces and no TDL operator word, is counted by its own ledger and its vouchers are read by that ledger in planned parts, each checked against the count voucher by voucher. Such a detail says so in window_scope (basis ledger_scoped, with a note on what that list was compared with) and counts party_vouchers_read; every other detail is read as before and keeps company_vouchers_read. Where the ledger is not read, or its read finds nothing or would need a day divided by AlterID, the original refusal stands and its window_scope gives the reason; a refusal of the ledger read itself (a part not admitted, a row with no line on the ledger) keeps its own code (part of #1356).
  • egress_log no longer refuses a read whose receipts are too large for one response. When the receipts asked for (they can be about 11 KiB each) would not fit the response size cap, it returns the newest whole receipts that do fit, measured on the response as it is sent, marks the answer truncated, and counts the receipts left out in result.omitted_for_size. A single receipt that alone is over the cap is still refused (agent_response_too_large). The result now says order is newest_first. Older receipts are still not reachable by paging (part of #2125).
  • The desktop Trial Balance and its Excel workbook now mark a ledger whose closing is not its opening plus its debit and credit totals, as the trial_balance answer does: the row is shown as Tally sent it with the difference beside its closing (a movements_differ_by column in the workbook), and the same explanation the answer gives appears above the table and in the workbook. Such a row was seen on a Closing Stock ledger whose closing balance was keyed on Tally's screen; the screen and the workbook were tested on its captured answers and not opened in the running app. A read where every ledger adds up looks as before (#2132, part of #2068).
  • parse_bank_statement_csv reads a local CSV bank statement in the one export layout measured, requires controls_source (prior_statement, bank_ledger_in_tally or this_file) and refuses every file with a typed code: one of fifteen shape categories, or statement_csv_account_line_unmeasured for a file that reads, because the export's account line is not yet established; it proposes nothing yet (part of #1320).

Safer or fixed

  • On Windows, the boxes that ask a person to approve a post, to record a review, to release an invoice stop or to create a ledger are now opened system-modal, to keep each above Tally when the person clicks into Tally while it is open. When a box opened a few seconds after the person's last input to Tally, Windows had not kept it on top, and one click into a maximised Tally covered it until it was brought back from the taskbar; in one lab run on Windows 10, the post and ledger-create boxes, opened system-modal, stayed above Tally after that click, and Tally still took a click and a key while a box was open; the other two boxes were not shown on a screen. In that run the taskbar showed no task buttons while a box was open; the cause was not tested (#2161). That run and the later lab runs started the server from the lab's own test client; Windows 11, and a box with Claude Desktop as the server's parent, were not run (part of #1780).
  • On Windows, the box that asks a person to create a ledger is now shown as the post's box is: brought to the front once when it opens, with its taskbar button flashing if Windows refuses, so that the Y and N its first line names can reach it. In one lab run it opened some seconds after the person's last input to Tally and never had the keyboard, so a Y pressed as its first line says would have gone to Tally. It is also measured against the screen: a box that does not fit is refused as master_review_too_large, with nothing sent. Like the other approval boxes, it now needs Windows 10 version 1703 or later; on an older Windows the answer is master_approval_unavailable (part of #1780).
  • The approval and review boxes now show dates in Tally's day-month-year form with the year in full, for example 14-Sep-2026 and Books from: 1-Apr-2026, where they showed Tally's request form 20260914. This applies to the one-voucher, invoice and batch approval texts, the one-voucher and batch review texts, and the box that asks to create a ledger, on Windows and macOS. Only the text a person reads changes: what is sent to Tally stays YYYYMMDD. Each date is up to three characters longer, so a text that sat exactly at a box's size limit before may now be shown as a batch's totals, or refused, as any text over the limit is (fixes #2101).
  • A native post whose recorded answer reported STATUS 0 no longer counts that answer's CREATED toward the voucher mark it proved, so whichever of the company's posts is checked next, a mark read unchanged does not record a rollback for it (#2035). After the update, a rollback record that read a mark at or above such a post's pre-POST mark no longer covers it: the post no longer reads book_rolled_back because of that record; its batch no longer carries restored where the transfers section or a withdraw_skips refusal names it, and a held row may name another batch that holds it; and the post stays unverified, as an answer reporting STATUS 0 always leaves it. The readback in that post's own post_import call still reads book_rolled_back when the mark falls between the one read before the post and the one read just after it; that post's own floor then records no rollback, though another native post of the company at that Tally address can still record one in the same call. The answers the tests use are written by hand: no import answer with STATUS 0 and a CREATED count has been captured from a live Tally, and the native posts measured answer with a RESPONSE root that reports no status, so this has not been seen on a live post.
  • Four checks that accepted any eight digits as a Tally date now refuse a value that names no calendar day (for example 20251340 or 20250229): a mirror batch's requested date range, which HSN item is in force on an invoice date, whether a sales ledger's GST rate is in force on an invoice date, and a company context's from and to dates. Before, such a value passed and, in the HSN and rate checks, could be chosen as the item or rate in force. A batch, an invoice build or a parsed export that carried an impossible date is now refused where it was accepted. No captured answer holds one, and whether Tally ever sends one in these fields is not measured; tested on invented facts (fixes #2009).
  • When post_import or verify_import cannot confirm a post because the company's voucher mark moved by another amount than the post created, or the post's span later held another number of vouchers, the post's summary now also says what was observed and what can cause it (for example a GST return exported from Tally, an edit of one of the vouchers, or another user's entry), and that ComplyEaze Bridge sent the post once and Tally reported creating each voucher once. That sentence is shown only for those two refusals, never for an answer that reported STATUS 0 or whose counters were not clean. Measured on TallyPrime 7.2, on one synthetic book each: two GSTR-1 JSON exports each raised the voucher mark by 3 without changing a voucher, while one local GSTR-3B JSON export of one month left it unchanged (17 before and after) and moved no voucher's alter id; a voucher saved by hand, once, inside a post's span made the mark move by 51 for 50 created, and the span binding was refused. Nothing about how a post is proven changes (part of #1633).
  • The outstandings party detail (#1960) now reads the company's two change marks before the statement of open bills and again after the voucher window, and refuses (voucher_window_changed_during_read, the same refusal and next step as every read that brackets the book) when either moved; the voucher window no longer reads the marks a third time. Before, the statement, the ledger catalogue and the vouchers were tied by the party's spelling with nothing checking that the book was the same between them, so a ledger renamed in Tally between two of those reads could answer as a finding (for example no bill for the party). It adds one marks read net (two, and the window reads none of its own), and any voucher or master edit in Tally during the call refuses it; run the same request again once. A change that moves neither mark is not seen. Measured: a rename moves the master mark and a voucher edit moves the voucher mark (TallyPrime 7.1 Silver, 21 Sep 2026), but one rename that changed only a ledger's letter case moved neither mark (one book), so that kind of rename is still not caught; this read has not been run against Tally.
  • Under mask_parties, ledger_masters with fields=compliance now masks the flat address of a ledger (and, from the same change, every mailing address line) like a party name: before, the flat address was returned as Tally sent it under every redaction mode. An address Tally returned empty stays empty. The flat PIN code and state, which every compliance read returns, and the mailing entry's PIN code, state and country, which are returned only when the assistant asks for the mailing details, are still not masked under any mode (#1490).
  • The vouchers snapshot shows voucher_alter_id, and the audit read's manifest alter_voucher_id, as null, not 0, when Tally's answer carries no voucher mark, which it omits for a company that has never held a voucher; ledger_masters already showed null for the same book. A mark Tally sends, 0 included, is shown as sent. Every comparison of the marks still reads an absent voucher mark as 0, so no read or post decides differently (fixes #1984).
  • A trial_balance row whose closing is not its opening plus its debit and credit totals (an empty amount counted as zero) now carries movements_differ_by (the exact difference), the first line counts such ledgers and a limitation explains, and the row is returned as Tally sent it. Seen on a Closing Stock ledger whose closing balance was keyed on Tally's screen: Tally kept the opening as the closing and showed the change only in the debit total, so the closing is not necessarily the figure its statements use. Every row of the committed captures adds up, so no other ledger is flagged; profit_and_loss and balance_sheet already refuse a read where a row does not add up, by the same check (part of #2068, #692).
  • When profit_and_loss or balance_sheet refuses because a Trial Balance row does not add up, the refusal now names that row's ledger and its gap, says what to tell the user, and carries the evidence of the Trial Balance and statement reads already made, not only the company read (#692).
  • balance_sheet and profit_and_loss now say when Tally's own Balance Sheet does not balance: on the one book measured (a stock item's opening value edited, TallyPrime 7.2) the export left out the Difference in opening balances line Tally's screen showed, so when its lines do not sum to zero the result carries tally_balance_sheet_sides (the credit and debit totals), the first line gives both and nothing is established; the new reason tally_balance_sheet_does_not_balance is given when no other reason applies (#2063).
  • A refused trial_balance read now carries a cause when the Trial Balance answer could not be used (the rule it failed, such as trial_balance_source_identity_missing or trial_balance_amount_invalid), where it gave the code trial_balance_read_failed alone. profit_and_loss, balance_sheet and cash_flow show the same cause, and no other reason or step changes. Every trial_balance_read_failed refusal that has no step of its own now carries one (a timeout, an HTTP error, an oversized answer or an unusable one): it tells the assistant to report what the cause says, or that the read failed when none is named, and not to repeat the same read in a loop. The refusals that have their own codes (before the books began, a period not honoured, Education mode, the base-currency codes) keep their own text and get no step from this. On the one book with foreign-currency vouchers measured (TallyPrime 7.2), the Trial Balance was refused with no cause before this change; a build with it, read in two sittings on 29 Oct 2026, refused with the cause trial_balance_amount_invalid and its next step. The later change that sets TallyPrime 7.2's composite layout aside clears that cause on the book's captured answer; it has not been read on that book through the tool. A failure in the totals after the parse still names no cause (part of #2071, part of #683).
  • When a read of the company's ledgers (the compliance read of ledger_masters, the sales and purchase registers, or the invoice import's admission) is refused because Tally wrote a ledger's foreign-currency balance in a layout ComplyEaze Bridge does not recognise yet (party_ledger_master_read_failed, cause native_foreign_currency_ledger_balance), the refusal now says what to do: that the read is refused rather than guessed at (#683), that a retry refuses again, and to tell the user that this read cannot give the company's ledgers while Tally writes a balance that way. The refusal was seen in the lab on one TallyPrime 7.2 book, on a build without this next step; that book's layout (no space after the foreign currency symbol) is now recognised and set aside by name like the 7.1 layout, from its captured answers, not from a read through the tool. What still refuses is a layout no capture shows, such as a rate with decimals; the next step has not been seen on a live refusal (part of #2071). No request changed.
  • A ledger balance that TallyPrime 7.2 writes as a foreign-currency composite with no space after the foreign symbol is now set aside by name, as the 7.1 layout is, instead of refusing the whole read. This applies to the compliance read of ledger_masters, the sales and purchase registers and the invoice import's admission (which read the same ledger source), to trial_balance on a book with several Currency masters, where such a ledger was named as left out only for the 7.1 layout, and to outstandings, where on a book with several Currency masters a party whose balance is in that layout leaves every figure and is named in mixed_currency_ledgers_excluded, so the result is partial, as it is for the 7.1 layout (on a book with one Currency master the read refuses, as before). A voucher entry in that layout is still refused, as no 7.2 voucher answer has been captured. On the lab book the trial_balance parse that refused with trial_balance_amount_invalid now returns the plain rupee ledgers and names the four it sets aside; none of trial_balance, outstandings and ledger_masters has been read on that book through the tool with this change (part of #683). No request changed.
  • When a request of Tally's Bills Receivable or Payable report takes longer than the 20-second limit, outstandings is now refused with its own cause, native_bills_report_deadline_exceeded, in place of request_deadline_exceeded. Its next step now also says that Tally can go on building the report after the limit, so the next read of that company should wait until tally_status answers. Another request of the same read cut at the limit (the groups, the ledgers, the company checks) keeps request_deadline_exceeded and no longer gets the Bills report's next step. A call the host withdraws while the read runs now finishes the request in flight and sends nothing more to Tally. The cause was tested with a simulated Tally, and a withdrawn call was not run against a live TallyPrime (part of #2119).
  • When the outstandings read of Tally's Bills Receivable or Payable report is refused because one of its requests took longer than the 20-second limit, the refusal now names a next step from what was measured on one synthetic book of about 30,000 vouchers: about 15,000 open bills answered and about 22,500 and 29,900 did not, and an earlier as_of with fewer bills open answered there, which on another book depends on the bills open then (#2119). A party detail refused by a step after its party was read by its ledger (the closing marks read, a book that moved, or the detail's checks) now says so in window_scope, and the next step for party_ledger_row_without_line says that only a bill trail can be narrowed by a reference (part of #1356).
  • When ledger_movement reads a ledger by its own vouchers and Tally returns a voucher with no line on that ledger, the refusal now names a next step: the figure was not released, the call is not repeated in a loop, and where Tally is not in Education mode trial_balance gives that ledger's opening, debit, credit and closing for the dates. When the read falls back to the whole book after a ledger-scoped attempt was sent and given up, a result that carries a read_cost says in window_scope that the block leaves that attempt out. No request changed (part of #485).
  • When ledger_movement reads the whole book's vouchers for a ledger instead of the ledger's own (a name the scoped read does not send, a name holding a word Tally's formulas read as an operator, a day too heavy for one scoped read, or a scoped read that found nothing) and that read is then refused, the refusal now carries window_scope with the reason and its own sentence (no figure was taken). Only a refusal of that read is marked; an argument or catalogue refusal is not. No request changed (part of #485).
  • A parent that is Tally's root is now  Primary, the text Tally sends, in the masters rows (cost centres, godowns, stock groups, groups), the stock_summary items, the ledger_masters, ledger_movement, gst_by_ledger and trial_balance rows (the Profit & Loss A/c ledger). It was a replacement character followed by #4; Primary, which read as a corrupted string. A group a user named Primary is still written plain, so the two stay apart, and gst_by_ledger no longer masks the root under mask_parties. The working form that the readers compare is untouched (fixes #2070).
  • A folded twin's parent in an import report's live_ledgers is now  Primary when it is Tally's root, the text Tally sends, where it was a replacement character followed by #4; Primary; a group a user named Primary is still written plain. The vouchers summary's Primary bucket label is unchanged: it is a label for the root's bucket, kept apart from a user group by the bucket's own key (part of #2070).

What works differently

  • A skipped bank-statement row that no statement carries is now tally_unchecked, not not_found, in the transfers section of a build and of a verification. not_found said the other account's Contra does not exist, but ComplyEaze Bridge reads only the statements it built and verified: a Contra entered in Tally another way is not in its records. not_found is now said only after a read of every voucher of the row's date in Tally finds no Contra between the two accounts, and no such read exists yet, so every row that read not_found reads tally_unchecked. Its next step tells the assistant to read that date's vouchers with vouchers, to look for a voucher of the Contra voucher class, or of any other type such as a Journal or Payment, that carries the row's amount on the bank ledger the other account's statement was built for, to tell the person what it found and to keep the skip; a voucher dated another day or booked against another ledger is not looked for, a transfer may print on different days, and a skip is withdrawn (withdraw_skips), on one side only, solely when the person confirms that no entry in Tally carries the transfer. Reading a day sends its vouchers to the AI provider: mask_parties shortens party and ledger names, drop_narration drops narrations, and neither hides amounts. A row the other account's statements do not cover stays not_established. Tested on invented facts. No tool makes the whole-day read yet, and it has not run against a live TallyPrime. The requests it would send were run by hand on TallyPrime 7.2 against one synthetic book (four days, 3 to 6 April 2026): each day's two censuses were byte-identical and agreed with the vouchers answer, which held the book's five Contras and none on 6 April (part of #1392).
  • The post_import tool text no longer tells the assistant that a row of a batch Tally rejected, or a genuinely different transaction that shares a voucher's fingerprint, is entered in Tally by hand. Like the refusals' own next steps, it now says that ComplyEaze Bridge cannot post that row yet and to tell the user so (#1709).
  • The desktop Journal review no longer answers a Journal too large for it by pointing to the manual workflow. It now says that ComplyEaze Bridge cannot post that Journal from the review and names the next step. When the review itself is too large, that is to ask the assistant to build it again with a shorter narration, or to decide whether to split it into smaller Journals, each built as its own batch; when the Journal's file is larger than the review reads, a shorter narration cannot help, so it says only to split it (#1709).
  • The tax-audit financial statements limitation about closing stock on books not integrated with inventory no longer calls the Stock-in-Hand ledger value "manually-entered"; it describes the balance the client recorded in the books, and its meaning is unchanged (#1709).
  • The voucher-type masters read's note on how sub_method was measured now says a voucher was entered "on Tally's voucher screen" where it said "by hand"; the note only describes how the reading was taken, and its meaning is unchanged (#1709).
  • The held_elsewhere step of the transfer section now names both outcomes of sending the batch that holds a skipped row: once the skip is live, post_import refuses that send as skip_row_paired_elsewhere, and a skip that was only built still lets it through, which leaves the row posted_against_skip once found posted (#1988).
  • The vouchers tool description and the agent guide say that a row's party is Tally's own PARTYLEDGERNAME, which on a Receipt, Payment or Contra can name the bank or cash ledger rather than the counterparty, and that the counterparty of those vouchers is read from amounts. The parse is unchanged (fixes #1986).
  • The text about cost-centre allocations in voucher reads that 0.6.1 shipped is corrected in three places. The vouchers tool no longer says that each category restates the entry's full amount and that summing rows across categories double-counts: that was not measured (only one category had been captured then), so it now says the rows are not checked against the entry and not to add rows across categories, each category's rows having added up to the whole entry on the one book since measured with two categories on one entry. The privacy page no longer says that an allocation in a shape the read does not recognise is not sent and not flagged: it is flagged (cost_centre_allocation_shape_unrecognised), and the tool text, the guide and the protocol reference now list all eleven reasons an entry's list can be unreadable. On an item invoice the sales or purchase ledger's entry carries the goods line: an allocation list beside the goods line is read like any other and one inside it is flagged; no captured item invoice carries allocations. No request and no answer changed (part of #1231).
  • The validate_masters description and the voucher schema's ledger field now say what the alias work already returns: a missing or near_miss name that is exactly one ledger's alias carries alias_of and is still not bound, and the twin check counts aliases, naming through_alias (#1688).
  • negative_stock now says which sign a value carries when it is present: the export's, as in the Stock Summary and the trial balance, where a debit is negative and a credit positive (the screen shows a credit as (-)). On the one book measured, the item at -3 Nos and a rate of 120.00 was exported with the value 360.00 on TallyPrime 7.1 and on 7.2, and the 7.2 screen showed (-)360.00; the other two short items captured had no value. The sentence is in the tool text and is the second of the result's limitations; no value changes (fixes #2051, part of #1358).
  • Reading the mailing details makes the answer larger, and the read is sized for it (#1490). In the one capture (12 Oct 2026, TallyPrime 7.1 Silver, eight ledgers) the answer grew from 19,094 to 20,896 bytes: 948 bytes for the one party that held a mailing entry and 122 bytes for each of the seven rows that carried an empty placeholder. With include_mailing_details the read is sized at 5,000 bytes a ledger instead of 3,750, above the largest row measured (4,902 bytes: that party with a second dated entry, TallyPrime 7.2); it is marked unverified in the source and not measured on a book of many populated parties, so a large book may be read in more parts, or refused, sooner, and a party with three entries or long address lines can still be larger than the figure. A dense book reads about 950 bytes more for each party that holds an address; how that compares with the 20-second read deadline is not measured.
  • The README now says what egress_log returns when its answer is trimmed for size: it names its order (result.order), counts the receipts cut for size in result.omitted_for_size and says they cannot be reached by this tool, and its evidence response digest hashes every receipt line the call read, which on a trimmed answer is more than the lines returned. A test pins the digest on a trimmed answer. No request or answer changed (part of #2125).

What has been run against a live TallyPrime, and what has not

  • Ledger create (#690): created and read back as posted_verified once through the tools on a synthetic company (TallyPrime 7.1 Silver, 14 Oct 2026); the approval box shown whole, declined, accepted and answered from the journal once on a CI package of the change (TallyPrime 7.2, 28 Oct 2026). On the build of this release, on a new synthetic company (TallyPrime 7.2 Silver, 30 Oct 2026): a declined create sent nothing (master_approval_declined); an approved create read back as posted_verified, the company's master mark up by one; and the same call made again answered from the journal with no box. The box was whole on the screen, above Tally, and had the foreground when it opened (#2171). A create whose answer is lost (#2003) is not run.
  • The approval boxes opened system-modal (#1780): one lab run on Windows 10 on a CI build of the change; the post and ledger-create boxes stayed above Tally after a click into Tally. Windows 11 and the review, release and settle boxes on a screen were not run.
  • Not run: the approval boxes with Claude Desktop as the server's parent, the way most people run ComplyEaze Bridge. The lab runs that opened a box, the live pass of writes and the masters row among them, started the server from the lab's own test client. A fallback for a box whose topmost flag is not kept (#2193) is held for 0.8.0.
  • stock_summary at month ends (#1233): two synthetic books on TallyPrime 7.2, one sitting each; items inside stock groups on one book with groups three deep.
  • A Trial Balance row that does not add up (#2068, #692): seen on a master build's Windows package on two synthetic books (TallyPrime 7.2, 29 Oct 2026): only the Closing Stock ledger was flagged, and the statements refused naming it and its gap. One sitting per book.
  • Refusal causes on a book with foreign-currency vouchers (#2071): the Trial Balance cause and next step were seen in two sittings on TallyPrime 7.2; the 7.2 composite layout set aside by name (#683) is read from the captured answers only, not through the tools.
  • The Bills report refusal past the 20-second limit (#2119): measured on one synthetic book of about 30,000 vouchers; the new cause was tested with a simulated Tally.
  • vouchers with bank_allocations, goods_lines and summarise_by: cost_centre (#1231, #2015): measured on saved raw answers of synthetic books; the cost-centre summary through the tool ran once on the build of this release, on one synthetic book (TallyPrime 7.2 Silver, 30 Oct 2026): the four cost-centre buckets in order, the category totals, the totals and the unallocated entries were as expected; with the bank ledger selected there were no buckets; the marks did not move. Masking was not run.
  • A ledger asked for by its alias (#1688): run once on the build of this release, on one synthetic book (TallyPrime 7.2 Silver, 30 Oct 2026). validate_masters gave the alias as near_miss with alias_of naming its ledger, and the ledger itself as exact. An absent name that shares a word with the book's other ledgers was also listed as near_miss, with candidates and no alias_of; a near miss never binds a ledger. Not run: the twin check through an alias, a build naming the alias, aliases of groups, Gold.
  • Settling a native post Tally answered as creating none of its vouchers (#1709): not run on a whole refusal. The release-candidate row stopped before its post, because locking a period in TallyPrime needs user security (a user and a password), which the lab's book does not have; no post was sent and nothing was saved. The settle dialog is covered by tests only.
  • Not run on a live TallyPrime for this build, with only captures or tests behind them: a post answer reporting STATUS 0 (#2035), an impossible calendar date (#2009), the outstandings party detail of a book that moved during the read (#1960), the desktop Trial Balance screen and workbook (#2132), a withdrawn outstandings call (#2119), and a ledger create whose company's master mark stepped by more than one (#2165), which goes to reconciliation and is tested only, because another master change inside the seconds of a create cannot be timed in the lab.

Live pass of writes: run on 30 October 2026 on the CI package built from master commit 026ba163, which holds every code change of this release; the version change merges after it, so that package reports 0.6.1 as its own version. Tally was TallyPrime 7.2 Silver, on the synthetic company BRIDGE TIE OUT LAB 3. The post_import text (#1967) and the held_elsewhere next step (#1995) read as written. A batch that would post a row a statement of this company skipped as carried by the other account's Contra was refused before any box opened (skip_row_paired_elsewhere), with no attempt recorded and the voucher mark, the batch journal and the skip records unchanged. An approved one-voucher post, answered by the person in the approval box, posted and verified (posted_verified, the voucher mark up by one step, the post bound to its voucher); the box opened above Tally, held the foreground, and showed its dates as Tally does (Date: 30-Sep-2026, #2105). verify_import then said only that each voucher was matched to the Tally voucher this post created (#1896). The company was then restored from its backup taken before the post: verify_import reported the post as rolled back (book_rolled_back) and recorded the rollback, and asked again it read the same state without recording it twice. The book was left as it was before the post. The binding line and the two rollback checks would pass on 0.6.1 too, so they show no regression rather than proving a change. The file that ran is the CI package; the published file is built again by the release workflow and is checked after publication against the release's checksum, with no write.

Known limits

  • Not run on a live TallyPrime for this build (unless the lines above say otherwise): anything that writes to Tally other than the rows of the live pass of writes and the masters row, Gold, macOS, other TallyPrime releases than 7.1 and 7.2 Silver, a book of about 30,000 vouchers for posting (#1029), and the Microsoft Store install of Claude Desktop (#1584, #1459).
  • Sales invoice posting is not available in 0.7.0, whatever the posting setting: a Sales invoice is refused before anything is sent (#2162).
  • Ledger create makes one ledger at a time and alters none (#690). A create that Tally runs after Bridge stopped waiting is possible: in the lab Tally ran every abandoned create it was sent (#1713), so a create with no answer is never sent again.
  • A ledger create whose answer was lost can later be closed as confirmed_absent, which releases its name, when a read sent after the create finds the ledger absent and the company unchanged. That rests on Tally running a create it was sent before a read sent after it. The lab measured this on TallyPrime 7.2 Silver, for ledger creates sent while a long read kept Tally busy, on new and on reused connections (whether ComplyEaze Bridge itself ever reuses a connection this way is not established): one observation per pattern, repeated once, on one synthetic book. Not measured: a create sent while another connection's long import holds Tally's gateway, Gold, other releases and other kinds of master (#2176, #1713).
  • Foreign-currency books: no read reports a foreign-currency amount; each sets such a ledger aside by name or refuses (#683).
  • `stock_summary` still refuses a book whose master mark is above 874, however few stock items it holds (#1233), and dates other than a month's last day.
  • `outstandings` at a year end on a dense book: the Bills report can take longer than the 20-second limit; the read is refused with its own cause and next step, and no figure is released (#2119). A longer limit for the Bills Receivable and Bills Payable report requests (#2158) is not in this release.
  • Settling a post Tally created none of (#1709): the settle dialog has not been shown after a whole refusal from a live TallyPrime; it is covered by tests only.
  • The approval boxes under Claude Desktop: on Windows the boxes were seen above Tally when the server was started by the lab's own test client, but not yet with Claude Desktop as its parent (#1780; the fallback #2193 is held for 0.8.0).
  • On Windows, while an approval box is open, the taskbar may show no task buttons (#2161). The macOS alert is unchanged (#1780).
  • The whole-day read for a skipped transfer row is not called by any tool yet, so no row reads not_found (#1392).
  • The file tested is not the file that will be tagged. Each live row names the build it ran on; the published file is built again by the release workflow and checked after publication against the release's checksum, with no write.
  • What has been run against a real TallyPrime is listed in the README.

0.6.1, #

In plain words: ComplyEaze Bridge 0.6.1, since 0.6.0 (16 Oct 2026)

These changes are in ComplyEaze Bridge 0.6.1. Most lines name the pull requests or issues they come from. Like every build so far, it is not yet code-signed. The desktop app has no published installer, so lines about it apply only to a build from source. Sales invoice posting stays switched off in 0.6.1; the lines about invoices apply only once it is offered.

Should I upgrade?

  • If you read ledger GST details (`ledger_masters` with `include_gst_details`): yes. 0.6.0 named Tally's input tax credit flag itc_eligible and read Yes as "eligible". A ledger whose "Eligible for Input Tax Credit" setting was keyed No came back as itc_eligible: true. The key is now ineligible_itc, an ineligibility flag (measured on one ledger under Purchase Accounts; read a Yes on any other ledger as unconfirmed). Keys named itc_eligible in anything you built on 0.6.0 must change to ineligible_itc, and their meaning is the opposite of what the old name suggested (#1741).
  • If you use posting: what Bridge writes into Tally is still only what you approve in the approval window, and the kinds of voucher it posts are the same as in 0.6.0 (Journal, Payment, Receipt and Contra; Sales invoices and ledger creation are not offered). What a post is checked against did change, and this build refuses more than 0.6.0 did. A post is refused before anything is sent when its last read before the send does not show the company's voucher mark (#1945). A verification whose voucher mark Tally leaves out twice, after its own window read saw one, is refused, not read as a rollback (#1819). A bank-statement batch is checked against the company's skipped rows and Contras at the moment it is sent, not only when it was built (#1714): if that check refuses, nothing is sent, any approval of the batch is withdrawn, and the batch has to be built again (after transfer_view_not_established the batch need not be built again, but it stays refused until the record is repaired; report it). The approval window and the review window now open with a line saying what each answer does (#1832). On Windows a window too tall for the screen it would open on is cancelled and the post is refused as too large, with nothing sent (#1879), and a batch listed line by line that does not fit is shown as totals when asked again (#1961). That fit check was measured on one Windows screen only (1920x1080 at 125 percent). On a Mac the alert's first line is no longer wrapped twice (#1854), but a Mac batch window can still be taller than the screen. While the approval window is open on Windows, do not type in Tally: a typed Y or N answers the window. Click the window with the mouse (#1784). The live pass of writes below ran one approved post and one statement batch that the send-time check refused, on a live book; the other new mark refusals (#1945, #1819) have run in tests only.
  • Before you upgrade: once 0.6.1 has seen a restored book, do not go back to 0.6.0 (see the restore paragraph below). Saved batches that name a ledger whose name equals another ledger's alias, and some voucher_presence calls that declare a type manual, are refused now where 0.6.0 answered them.
  • Everyone else: nothing else needs doing. Install steps are the same as for 0.6.0: installation guide; after installing, check the extension shows 0.6.1, the Terms setting is ticked and the posting setting is as you want it; quit Claude Desktop completely and reopen it.

What you can do now

  • audit_test_run, off by default behind the extension's new Allow tax-audit test runs (audit_test_run) setting (BRIDGE_AGENT_ENABLE_TAX_AUDIT; off, it is not listed and a call is refused as audit_tool_disabled), runs one tax-audit test on a company's own books for a date range, today only cash_44ab: figure and finding ids with their values, confidence and clauses, and the count of any failed book check, with no ledger, party or voucher name; you name the cash and bank groups. Books only, not a tax opinion; refused while a redaction setting is on (audit_redaction_unsupported), and for a period outside 1 April 2025 to 31 March 2026, the year its rules are for (audit_period_outside_ay_2026_27). The engine's test passed a real-book parity check on 21 Oct 2026, on one book with no cash payments (that side was checked at zero), at reference pin 6168aeaa (crate tree 53938eb8); the tool is replayed from the engine's synthetic book and not yet run against a live Tally (part of #697).
  • gst_uncertain_vouchers lists the vouchers of one voucher type (a Sales type is what was measured) in a window that Tally marks Uncertain in its GSTR-1, and the vouchers a person accepted as they stand (Accept As Is), each with its party ledger, its invoice total (without its sign, with its side) and Tally's identifiers, and the line a CA reads before filing. It counts the window's vouchers with the status read first and refuses a window above 1,000 vouchers (provisional) before it reads amounts, and refuses an amounts read that returns fewer vouchers than the count found; the line names the voucher type it covers, and is withheld when no voucher was read or none is uncertain, and, with the result partial, when a voucher's flags or a listed voucher's party or amount were not read. Its answer carries party names and amounts to the assistant; mask_parties masks the names and not the amounts. The reason for uncertainty is not read, and bills missing an HSN/SAC are not split out. Measured on one synthetic book of TallyPrime 7.1 Silver, before and after one Accept As Is; item invoices, a registered buyer, credit notes, foreign currency and several accepts at once are not measured. It reads at most 93 days and refuses a book whose voucher mark is above 25,000 (part of #1342, #1692).
  • profit_and_loss and balance_sheet can now be established on a book with stock items (part of #692), using Tally's own Less: Closing Stock figure from its Profit and Loss, shown as closing_stock (applied, with its value and source, or declined, with the reason) and not derived or taken from stock_summary. It is used only when that one figure closes every gap at once (Current Assets, the Balance Sheet's Profit & Loss A/c line and the trading result, and the Cost of Sales : heading), the company's own flags show inventory on and integrated, the Opening Stock line is shown empty and no Stock-in-Hand ledger carries an amount; otherwise the read refuses as it did. Both tools now read Tally's own Profit and Loss (one more pair of requests for balance_sheet, which stays the sheet it was if that read fails), and a book that offers a closing stock also has its inventory flags read (a failed read declines the figure as inventory_flags_not_read). The derived lines must also sum to zero. Measured on one synthetic book; a book with opening stock is declined (shown on captured answers), and a costing method, a book that is not integrated and a window spanning two years are not measured.
  • masters can list a company's stock items (kind stock_items): each row carries the item's own dated HSN entries (hsn) and GST rate entries (gst) as Tally sent them, nothing choosing the code or the rate for a date (part of #1233). An empty or missing list is not "no HSN or GST": after a code was keyed on a stock group, the group's four other items still carried empty lists (one book); a GST rate or an HSN code on the company, and whether an item that defers to a coded group inherits its code, were not measured; the group's and the company's setting is not read, and every row says so (inherited_from_group_or_company: not_read). A defect in one item's list marks that list unreadable with its defect and the read stands; a Tally ERROR, markup inside a value or a cut-off list refuses the read. Item names and parents are masked under mask_parties, and so are HSN descriptions (a person can edit them); HSN codes and rates are returned as sent. Tally's GSTINELIGIBLEITC is returned as sent under ineligible_itc: it was Yes where the screen read Eligible for Input Tax Credit: No and absent where it read Yes (one item, two rows, PARTIAL), which fits an ineligibility flag; null is neither eligible nor ineligible, and what an absent element means for the tax is not known. On an entry whose source is As per Company/Stock Group (the entry Tally adds itself on a first save; a person can also choose that source) a flag Tally sent is withheld: ineligible_itc is null and ineligible_itc_withheld is tally_default_entry, because the Yes seen on such an entry was never compared with the screen; a flag Tally did not send stays null with no marker, and an entry with any other source is sent as it is. ledger_masters names the same element the same way (#1741). A book is read only up to a master mark of 489 (a limit of 489 refuses, by arithmetic, both client books measured; not tried on them). Measured on one synthetic book of TallyPrime 7.1 Silver (three captures, one item with two dated entries in each list); a state row other than Any, slab rates, another release, a larger book and a zero-row stock-item answer are not measured. stock_summary is unchanged.
  • negative_stock (new tool; part of #1358): Tally's own Negative Stock list for a window, as item, quantity, rate and value per listed stock item, returned as Tally prints them and compared with no other report (checks says so). Being listed is not a finding that the quantity is below zero, and an empty answer is not_established, never "no negative stock": Tally's empty answer cannot be told from a report that was not rendered. Tally's report was captured from synthetic books: a short item listed for a whole financial year, the same item at another quantity for a window ending mid-year (with its rate and value empty), and a book with none short answered an empty envelope. Any other window is named as unmeasured, a book whose master mark is over 874 is refused before the report is asked for (a stated guess, as for stock_summary), and the cost on a large book is not known. The tool itself has not yet been run against Tally.
  • negative_ledgers (new tool; part of #1358): Tally's own Negative Ledgers list for a window, as ledger, side and amount per listed ledger, returned as Tally lists it and compared with no other report (checks says so). An empty list is not_established, never "no negative ledger": Tally's empty answer cannot be told from a report that was not rendered. Tally's report was captured from three synthetic books, each for one whole financial year (on the first, a creditor with a debit balance and a debtor with a credit balance were listed and a Suspense ledger with a credit balance was not; on the other two, the one ledger listed matched the trial balance); any other window is named as unmeasured, and the cost on a large book is not known. The tool itself has not yet been run against Tally.
  • Voucher reads now carry cost-centre allocations: an entry that has them gets cost_centre_allocations (category, cost_centre, amount; names trimmed, the amount as sent; both names are shortened under mask_parties, in the vouchers read and the two registers). A list that cannot be read whole (a centre with no name or no amount, an amount that is not a decimal, a centre or a category named twice, an empty allocation container, an allocation list in a shape not recognised) does not refuse the read: the entry carries cost_centre_allocations_unreadable with a reason and no partial list. No request changed; the allocations were already in the answers and were dropped (part of #1231).

Safer or fixed

  • A batch built from a bank statement is checked against the company's skipped rows and Contras when post_import sends it, as it was when it was built, with the same refusals (withdraw_skip_paired, skip_row_paired_elsewhere, skip_row_already_posted, withdraw_skip_invalid, transfer_double_booked, transfer_view_not_established): before the post (before the dialog, or in the call that redeems an approval) and again under the admission lock. The refusals at send speak of the saved batch: nothing was sent, any approval of it is withdrawn and, except for transfer_view_not_established, the batch must be built again. A row the batch withdrew is judged on the withdrawal it recorded (withdrawn_skips), whether it was built from a statement or from vouchers given directly, and it counts only if the batch was built after the newest record of that skip: a skip recorded since refuses the batch (part of #1703).
  • When a batch that names skipped rows in withdraw_skips is refused at send as withdraw_skip_paired, its next step now ends by saying which build to make: this statement again, with the rows in paired_rows still skipped and not named in withdraw_skips, instead of a bare "build it again" after steps that may speak of the other account's statement (#1714).
  • A company restored from a backup after a native post is remembered once it is seen (part of #1050). When verify_import reads the company's voucher mark below the floor a natively posted batch had proved (a mark Tally omits twice counts as a company holding no voucher), the import journal records it, a second restore too, and from then on no natively posted batch of that company at that Tally address that it covers reads posted_verified, even after vouchers keyed again by hand raise the mark past it: a voucher it finds reads as matching content, one it does not as book_rolled_back, and the answer says to check them in Tally. The answer that records it names the company's other native batches it covers, to verify again. Before, a voucher keyed again with the same content could read verified once the mark had risen. Limits: a restore that is keyed or edited past the mark before any check is still not seen; a wrong backup restored and then the right one leaves the batches it covered reading rolled back (as would a repair or rewrite that lowered the mark, not measured); and a covered Sales invoice batch stops that company's invoices until released. Do not go back to an earlier release after this one has seen a restore: an earlier release refuses an import journal holding the new record, and stops every build, post and verification until ComplyEaze Bridge is upgraded again.
  • The check of a bank-statement batch against skipped rows and Contras now treats a batch that a recorded restore covers as neither found posted nor verified, whatever its last status (part of #1050). Before, a skipped row such a batch held could be retired, or listed as posted_against_skip, for good, and its Contra could pair the other account's skip, although the restore may have removed the voucher. Now the row stays open (held_elsewhere, naming the covered batch), its Contra is an unverified candidate, and its days give no coverage. Every row that names such a batch is marked "restored": true, and the next step says that no step inside ComplyEaze Bridge settles it: do not withdraw the row or build or amend its rows, check in Tally and tell the person, and do not enter the transfer by hand or import that batch's file. Nothing else in the journal changes and no new record is written. Limits: the note is advice, which the code backs with refusals. A withdrawal of the row is refused (withdraw_skip_paired, import_txn_already_posted or withdraw_skip_invalid), so is an amendment of the covered batch (import_amend_natively_posted), and so is a new build that posts a row the covered batch holds (import_txn_already_posted). The covered statement's days stay blocked for the other account's coverage, so skips there stay not_established or matched_unconfirmed. A voucher that comes back in Tally after the restore is no longer flagged as posted_against_skip. Not run against a live TallyPrime.
  • voucher_presence now checks a voucher type declared manual against the book's own numbering before it reads the window, because a voucher number is an identity only where the book's series for the type says Manual and prevents duplicate numbers. A call that declares a type manual now makes 20 more requests (34 become 54 on a small book: the company's master mark, then the voucher-type numbering) and is refused as a whole, with presence_numbering_method_differs_from_book and the types in candidates, when the book does not number a declared-manual type Manual with exactly one series that prevents duplicate numbers (the check now also reads the duplicate setting, and shows it in book) or holds no type of that exact name; a numbering that cannot be read refuses with the read's own code (masters_education_unqualified in Education mode, for example) or presence_numbering_unreadable. Calls that were answered before are refused in those cases. A sub-method printed on a Manual series is ignored (a sub-method printed twice still refuses the read as field_invalid). Names match exactly. Every type declared manual is checked, whether or not a proposal names it. The check reads the book's numbering as it is now and holds for the moment of the read. A call that declares no type manual reads no numbering and its result says numbering_read not_read; a call that does read it returns what the book said for each declared type, or only their counts when the rows would not fit the response budget. Request counts are from test doubles; not yet run against a live Tally.
  • ledger_masters with include_gst_details now names Tally's GSTINELIGIBLEITC as ineligible_itc (true for Yes, false for No, null when the element is absent), where 0.6.0 named it itc_eligible and so told an assistant that an entry with a keyed "Eligible for Input Tax Credit: No" was eligible. Measured on one ledger created under Purchase Accounts in a synthetic book (TallyPrime 7.1): the screen's Yes read as the element absent and its No as Yes, so the element is an ineligibility flag (PARTIAL: one book, one release, one ledger, one dated entry; part of #1741). A Yes on any other ledger says nothing either way: the Sales ledgers of one captured book carry a Yes on their keyed entries, and Tally's own default entry (source As per Company/Group) carries one with no keyed No; neither was read against the screen. The value mapping is unchanged: do not read true as eligible.
  • The tool list is answered whole under its own fixed ceiling of 1,000,000 bytes, apart from the cap on what a tool returns from the books (#1798). Before, a BRIDGE_AGENT_MAX_BYTES set below about 197 KB made the list itself too large (agent_response_too_large) and left the assistant with no tools. Every tool answer still keeps the configured cap, and a test holds the tool list under 225,000 bytes.
  • A ledger name that is an alias of another ledger is now named as one (part of #1688). validate_masters and the refusal of a build (masters_not_exact) add alias_of to a requested name that is, character for character, one alias of exactly one live ledger, naming that ledger as Tally lists it, and say that an operator should confirm it is the ledger meant before the name is replaced; ComplyEaze Bridge still never sends an alias. The check for two ledgers whose names read alike now counts a ledger's aliases too (through_alias names the alias). Before, a name that was one ledger's own name and folded equal to another ledger's alias matched the first exactly and could be built and posted; it is now refused as a twin, and so is a batch built or approved before this change that names it, at post or in the queue (ledger_has_folded_twin), until the alias or the name is changed in Tally and the batch is built again. A name that is only an alias used to read as missing or near. On Tally's own screen an alias equal to another ledger's name was refused in the cases tried, so such a twin is probably rare (PARTIAL: one company; lowercase not tried). Not yet run against Tally through the built tool.
  • Checking a native post with verify_import no longer records a rollback when Tally leaves the company's voucher mark out of its answer twice in a row while the check's own window read did see a mark. That case is now refused as voucher_mark_contradiction, with nothing recorded, and the refusal says to verify again once. A window read that saw no mark, or a mark of 0, records mark 0 as before. No omission of this kind has been seen from a live Tally (part of #1050, fixes #1819, #1853).
  • audit_test_run now refuses a company whose books begin after the audit period ends (C5-books-from) before it reads the books (it first asks Tally which companies are loaded); before, it read the whole book first and refused afterwards. The refusal's code is unchanged and it still may not be retried. Books that begin inside the period (a first year) still read (fixes #1934).
  • When a row is refused as already posted and verify_import reports the blocking batch book_rolled_back (the company's voucher mark was seen below that batch's), the refusal's next step now says that ComplyEaze Bridge cannot write the row again and cannot tell whether its voucher is gone, and to tell the user (#1038).
  • On Windows, the approval box for a post, a review and a stop release is now measured when it opens. A box that is not wholly inside the screen it opened on is cancelled and never answered, whatever key was pressed, and the call is refused as import_review_too_large (the post) or ack_review_too_large (the review and the release), with nothing sent to Tally. Before, Windows left a box taller than the screen with its buttons below the edge, where a person could press Y on lines they could not see. A Windows that cannot be made per-monitor DPI aware shows no box at all, and the call is refused as import_approval_unavailable (ack_review_unavailable for the review and the release). A long batch or a small or highly scaled screen can now be refused at the box where it was shown before; a larger or less scaled screen, or a shorter narration, helps, and the post_import and acknowledge_post_review texts say so. A saved batch or Journal whose window is too large to show whole is refused with a message that says so: nothing was sent, a larger or less magnified screen may help, a batch whose window listed its vouchers one line each is shown as totals only when asked again soon (see the totals paragraph below), and a batch can be split into parts (a single voucher this long cannot be posted from here). The refusal message was run in tests only. The box is also brought to the front when it opens, and its taskbar button flashes if Windows refuses the foreground to a background process (the flash was not observed). Measured on one Windows PC with a work area 1,030 pixels high (1920x1080 at 125 percent, one monitor), in a CI build of that change: boxes of 40 and 44 lines (935 and 1,011 pixels high) were shown and answered normally, and boxes of 48, 60, 80 and 120 lines (1,087 to 2,455 pixels) were cancelled in about a quarter of a second, not seen on that PC, and refused. That screen holds about 44 lines. Scales of 100, 150 and 175 percent, 1366x768, a second monitor and the Windows text-size setting are not measured. The macOS alert is not covered by this change (part of #1780, #1832, #1879, #1889).
  • An amendment now refuses a reference that differs from any build of its batch, whether added, changed or removed (import_amend_voucher_reference_changed, #1941). Before, an amendment that changed only a Journal's reference could read posted_verified once anything else altered the voucher, without its new reference ever being read back, and one that removed a reference could not clear it. Change a reference in Tally after the amendment is imported and verified.
  • A native post refuses before it sends anything (post_mark_unrecorded) when its last read before the send does not show the company's voucher mark, instead of reading the missing mark as zero and binding its vouchers from it; the voucher was never misattributed, but it could be posted and then left for reconciliation. A read after the post that omits the mark is likewise no mark, so the post is bound from its own answer instead of being refused as a mark that went back, and its post_location reads after_snapshot_unavailable with marks_readback_voucher_mark_omitted instead of a step to zero. A company that has never held a voucher is still refused when the batch is built (fixes #1945).
  • On Windows, a batch whose approval box with one line per voucher does not fit the screen is refused as before (import_review_too_large, nothing posted), and asked again it shows the totals only: the per-ledger totals, the company and the cautions, with the sentence that the voucher lines are not shown. Without this the same batch could only be split into parts. The approval is for that text, and a later call posts exactly it. It applies to a batch of two or more vouchers once its box was seen not to fit, for as long as an approval would be kept; one voucher and the review window have no shorter text and are refused as before. If the totals box does not fit either it is refused again, and a further ask shows the totals again for as long as an approval would be kept, so the batch has to be split. Run on a real screen in the lab for the check that finds the box too tall (part of #1780); the second ask and the posting after it have been run in tests only.
  • Dependencies were updated to clear the advisories RUSTSEC-2026-0382 and RUSTSEC-2026-0383 (#1793) and two more (#1849), and zip moved to 9.0.5 (#1782).

What works differently

  • The three cash-line answers ComplyEaze Bridge cannot build yet (cash paid straight to someone, a deposit from the cash box, unbooked cash sales) no longer say to record the line by hand or to record the sales in Tally first; each keeps its reason and offers dont_know, which builds the line to the suspense ledger, marked for the CA to move (part of #1709).
  • On macOS the first line of the approval alert (what each button does) is no longer broken by ComplyEaze Bridge at 100 characters: the alert wraps its own text, so a break of ours was wrapped a second time and left a ragged second row. The Windows box keeps its 100-character lines. That change does not alter the box's size caps (part of #1780, #1854).
  • The next steps for five situations no longer send a person to enter a voucher in Tally themselves; each now says what ComplyEaze Bridge does and that it cannot put the voucher in the book yet: a voucher Tally reported as not created (and one later not found after that report), a row a rejected batch carried, a statement day that is partly posted, the extra row of same-day twins, and the first voucher of a company that has never held one (part of #1709).
  • A saved batch built while voucher posting is off now says whether turning on Allow voucher posting would give it a route: the build checks the batch as post_import would at the batch limit of 50, and the assistant is told to say that, with the setting on (it means accepting the limits in Terms of Use section 9.2), ComplyEaze Bridge would check the batch again and ask for approval, and may still refuse it; the file import stays the route until then. Every text that offers both routes says to use one only: once the file is imported, do not post the batch, and once the batch is posted, do not import the file (an import after a post is not checked against it). With posting on, a batch post_import can take no longer suggests importing its file as well; a batch over 50 is told to be built again in batches of at most 50, and never to import the original file once one of them is posted or imported. The desktop review names the assistant's batch posting for a saved batch it cannot post, says a saved invoice is posted only that way, and no longer suggests choosing a Journal without its number. Wording only otherwise; the file routes for a voucher that carries its own number and for text the approval window cannot show are unchanged (part of #1709).
  • The approval window, and the review window, now open with a line that says what each answer does, before any detail (part of #1780). On Windows, whose box takes the keyboard, it reads "Press Y (or click Yes) to post this voucher into Tally. Press N (or click No) to cancel. Nothing is posted until you answer." (a batch says "these N vouchers"; the review window says "record that you reviewed" and "recorded"); on a Mac it names the buttons to click. The invoice-stop release, which shares the review window, says that Yes releases the stop, under its own statement. The text is wrapped under 100 characters a line, so the window's caps grew by its size (3 lines and 150 characters at most), and every text that fitted the caps before still fits them, though a box 3 lines taller can now be too tall for a screen it fitted before. The post_import tool text carries the same sentence, with a note to say it before calling. The keys are unchanged: Y and N still answer the Windows window. The new first line has not been seen above Tally, and on a non-English Windows the box's own button letters may differ, which was not measured. Boxes of the new size were shown on one Windows screen (see the Windows box paragraph below).
  • A read refused because the company it names is not loaded in Tally, was renamed since list_companies, or is listed more than once now says so and says to open the company in Tally and ask again (#1699). Before, an assistant saw only an operation code such as ledger_movement_read_failed, and the all-companies outstandings sweep and the All clients screen reported the same cases as a failed currency probe. The refusal carries company_identity_absent_or_ambiguous, company_identity_presentation_collision or company_identity_not_matched as its cause. Nothing about what is refused changes; only the reason and the next step are named. Not yet run by us against a renamed company in a live Tally. A presentation collision (company_identity_presentation_collision) is named as the cause but carries no next step.
  • A read that stops because the company's books changed in Tally while it was being read (voucher_window_changed_during_read and the *_extent_changed codes) now says that nothing is wrong with the request and to run the same request again once the saving or exporting in Tally has finished; a later page of a held listing that cannot be served (listing_snapshot_changed) says to start again from the first page without snapshot_id (part of #1633). One GST return export in TallyPrime 7.1 Silver raised the voucher mark by 3 with no voucher change, which is why a routine export can trip these. Not changed: a correct post that reads as unverified after such an export is still open on #1633.
  • A snapshot request that reaches its 20-second deadline is recorded as request_deadline_exceeded instead of tally_unreachable (#1671), so a person is no longer told to check that Tally is running when it was only slow. The deadline and the rule that such a request is not sent again are unchanged. Covered by tests only.
  • The Proof of Sync screen no longer shows a rejected-row count (#1634): it could only ever read zero, because only accepted observations enter the local mirror. The summary line drops the "rejected rows" clause and the table column reads "Accepted returned rows"; a run that rejected source records is still shown by its gap. The count stays in the proof's data and in the export.
  • The summary of a post whose binding was refused now says that ComplyEaze Bridge cannot close the batch or amend it (part of #1039). Before, it said only that the batch "stays open". The notice for a multi-voucher batch whose voucher mark step was not confirmed no longer suggests that recording a review with acknowledge_post_review clears it: the record is kept on this computer and the notice stays. No behaviour changes. A way to close a batch with a refused binding from ComplyEaze Bridge is not built: a person's confirmation cannot stand for a voucher's identity before it is measured whether Tally names the post's own vouchers when another user is writing.
  • When a withdrawn skipped transfer row's other statement may sit in a batch that is not verified (other_side_unverified), the next step now offers to build that statement again only if every batch named for the row was never sent to Tally, verify_import listed every one of their vouchers as not_found, and the person confirms none of their files was ever imported; if any voucher was found or a file was imported, it says to check in Tally which of their vouchers are there (#1820). Wording only.
  • The desktop app now shows a window asking the person to accept the Terms of Use, before it shows anything else and before it reads from or sends anything to Tally. Until the box for the current Terms version is ticked and accepted, every read and every send from the app is refused. The acceptance is a line in a small file in the app's own data folder (the Terms version, the time and that it came from the desktop app); if that file cannot be read or written the window stays closed and names the file. A new Terms version asks again. Reads of data already kept on the machine and saved drafts are not gated. The extension and the MCP server keep their own acceptance, unchanged. Tested with unit, component and browser tests; the packaged app and the acceptance window were not run (part of #1106).
  • When the local mirror refuses to commit a snapshot run because of something in the run's own saved batch, the Proof of Sync screen now says so and offers no resume. Before, it showed the same "mirror failed" title as a locked or full database and offered a resume, which could never succeed, because the same commit is refused every time. The run stays listed and inspectable: its status line reads "Commit refused: a resume cannot succeed. Start a new read.", the recent-runs table reads "Commit refused", and nothing is lost from the last verified baseline. Other mirror failures keep their title and their resume. After the app is restarted, a refused run shows again as one that can be resumed until the resume is tried, and is then refused again with the new text. Only unit and component tests ran; no run was refused this way by a real mirror, since no release writes such a batch (fixes #1677, #1900).
  • A call to audit_test_run that names a test the tool does not admit is refused as argument_invalid:test, before any request to Tally; the guide now names that code (it had named audit_test_not_admitted, which a call could not reach), and the list of admitted tests in the tool's schema is taken from the tool's own list instead of a second copy (fixes #1915).
  • The receipt of a failed call now keeps the error code of a tax-audit refusal (a rule code such as BIND-GROUP-UNKNOWN) and of a refused argument (argument_invalid:test). Before, only lower-case tool codes were kept, so those two kinds of refusal left a receipt with no error. A code is kept only if it is one of three kinds (a lower-case tool code, or one of two closed lists) and is at most 64 bytes; anything else is still dropped. The rule codes are the 61 tax-audit codes, and an argument name is kept only when one of the registered tools publishes it (fixes #1916, #1923).
  • The entry for a change that is not yet in a published build is now one file under docs/changelog/unreleased/ (a folder each for added, changed and removed), not lines added to CHANGELOG.md, so two changes no longer edit the same lines. The install site's "What changed" page and the release notes list these files under the unreleased changes, and cutting a release moves them into its section (#1929).
  • The window that asks a person to release an invoice stop is titled for that, "ComplyEaze Bridge — release the invoice stop" (on Windows, "… release the invoice stop for this company?"), and on a Mac its button reads "Release the stop". It was titled as recording a review of one voucher, with the review's "I reviewed it" button. It runs in its own dialog mode and answers with its own token, so a review window cannot answer it (#1828).
  • When vouchers, voucher_presence, ledger_movement or a register comes back empty for a window on a large book, the second look at the window widened by a day each side now counts it instead of reading its vouchers, so it usually sends fewer requests (#1240; the same number where the count takes several requests and the widened window is small). The labels do not change: an empty window with vouchers on a neighbouring day is still partial (empty_window_wider_read_nonempty), and a voucher inside the window still refuses the read. What differs: a count that took several requests is refused (voucher_window_changed_during_read) when the company's marks differ from the ones the first read opened on, as a divided read is; a count that took one request is not bracketed, where a widened window that used to be read in several parts was; and because no vouchers are read, a refusal that came from reading them (a part that failed or was not admitted, a part Education cannot serve, a read plan that could not be made or went over its allowance, a disagreement between the count and the read) no longer happens. A count of several requests can still be refused when it passes its own allowance. A book that fits one request is read as before. One live run on TallyPrime 7.2 gave the same answers and the request counts the test doubles predict, except one window of a book counted in several requests whose neighbouring day held vouchers, which sent fewer requests than predicted (#1999).
  • A read for a company that is not among those open in Tally, refused as company_identity_not_found, now carries a next step: call list_companies, use an ID it returns, and ask the user to open the company in Tally if it is not listed. The refusal itself is unchanged. Checked on a test double, not a live Tally (#1903).
  • The Download page, the install guide and the top of each release now help a person check a download. The Download page gives the two commands that print a file's SHA-256 code (certutil -hashfile "<file>" SHA256 on Windows, shasum -a 256 "<file>" on a Mac) with the file name in double quotes, says to run them in the Downloads folder, shows beside each download the code as GitHub reports it, when GitHub returns one, and says where a newer version appears. A GitHub release now opens with a short header (which file to take, the install guide, the checksum step) before the change list. What an older release needs moved to the last section of the Download page, below the steps for a new install, and the README gained an "At a glance" block with what can be asked and what cannot yet. This release opens with that header, and the Download page changes when its deploy is next run (part of #1148, #1046, #1785, #1786, #1788, #1790, #1837, #1850, #1882).
  • The install guide and the Download page now say that a Windows computer needs the Microsoft Visual C++ Redistributable (x64) to start ComplyEaze Bridge, and link Microsoft's own installer for it. The package does not carry it: the Windows program of 0.6.0 uses it, and a computer without it should not start the program (no clean Windows computer was tried). Only the texts changed, not the build (fixes #1860, #1861).
  • The NOTICE file in every package, and the README's License section, now say that Tally and TallyPrime are trade marks of Tally Solutions Pvt. Ltd., that ComplyEaze Bridge uses the names only to describe compatibility, and that it is not affiliated with or endorsed by Tally Solutions (part of #1865, #1868). NOTICE also names the copyright holders of the code in the AWS-LC cryptography library that the packages link, as that library's own licence file states them (BoringSSL, OpenSSL-derived code, fiat-crypto, s2n-bignum and the Jitter Entropy library), for version 0.44.0 of the crate; a newer version should be read against it (fixes #1858, #1859).
  • The amendment warning now says that saving a voucher unchanged in Tally, or cancelling it, moves its ALTERID (one run each), and so did changing a bill allocation's reference, adding a cost-centre allocation and setting a bank-reconciliation date (one run each), where it said an edit made on Tally's own screens had not been shown to move it; the amendment check itself is unchanged (part of #685, #1927).
  • The party and ledger master export now reads the CA's stored Schedule III grouping decisions from the encrypted local mirror, so the first export after the app starts can raise the macOS keychain prompt where it did not before. When the decisions cannot be read, the workbook says "Could not be read" and is NOT FINAL instead of saying none were applied; and when the mirror holds decisions for this company GUID under another company number or books-from date (for example books restored into another data folder), the export reports that count, is NOT FINAL, and applies none of them (part of #737, #1878).
  • parse_bank_statement's tool text now tells the assistant it may ask the person once about all the open cash lines, listing each line's date and amount, then answer each line on its own and ask again about any line the person did not confirm; one answer is never applied to another line for them. What the tool returns is unchanged: an open cash line still carries only its id, date, amount and party name. Not run against a live TallyPrime (part of #1761).
  • Checking a native post with verify_import no longer binds the post to the vouchers in its range of Tally AlterIDs when the post's recorded answer reported STATUS 0, whatever the answer's counters say. The check now reports post_span_binding as refused with the code span_response_status_failure; the post itself already treated that answer as not clean (#2008).

What has been run against a live TallyPrime, and what has not

  • voucher_read_cost (#1465): run once on the published 0.6.0 build (the code is unchanged since) on one synthetic company with a voucher mark of 68. The preview said one census read for vouchers and none for ledger_movement and verify_import, with a floor of 0 seconds; its own requests were identity, marks and status reads only. One vouchers call over a one-day window (11 vouchers) came back complete with one census request, equal to the preview's. The floor formula was not exercised (one census read gives a floor of 0). The refusal above a voucher mark of 2,097,152 was not measured (no book that size).
  • Approval window size (#1780, #1832, #1879): run on one Windows screen only (1920x1080 at 125 percent, one monitor); see "What works differently" for the box sizes shown and refused. Not run: Windows at 1366x768 or at 100 or 150 percent, a second monitor, a moved box. Not run on a Mac against this build. A capture of the Mac alert on 23 October, before #1854 (a 1512x982-point screen) showed the one-voucher worst-case box fitting and the 43-line batch box not fitting and not scrolling; this build was not measured on a Mac.
  • The first line of the approval window (#1832): seen above Tally once, on the approval window of the approved post in the live pass of writes; not looked at on any other window.
  • Empty-window count (#1240, #1999): one live run on TallyPrime 7.2 (on a CI build of #1897) gave the same answers and the request counts the test doubles predict, except one window of a book counted in several requests whose neighbouring day held vouchers, which sent fewer requests than predicted. The live pass of writes and the restore row (#1814) also ran on 7.2; no other read did.
  • Reads on TallyPrime 7.2 (ineligible_itc #1741, voucher numbering check #1704, the not-loaded cause #1787): not run. The 0.6.1 reads were measured on TallyPrime 7.1, apart from the empty-window count above; the live pass of writes, the restore row and that count ran on 7.2.
  • audit_test_run (#1733): not run against a live TallyPrime in this build, replayed from the engine's synthetic book only.
  • statement_tie_out (#1211): not run.
  • Voucher-type numbering (#724): run on one approved post and one refusal of a Receipt under Manual numbering (see the live pass of writes below). Not covered: Gold, Contra under Manual, a type that already held numbered vouchers.
  • A restore remembered after a native post (#1814): run on 27 October 2026 on master build fd6f7c0d (of 26 October), which holds #1814 (not the cut build), on TallyPrime 7.2 Silver, on one synthetic company (BRIDGE TRANSFER PAIR LAB), ten steps. After the book was restored from a backup, the verification of the posted batch named the restore (book_rolled_back), and kept naming it after a Bridge restart and after a voucher with the same number was keyed by hand; its records were unchanged across the restart, and the book ended at its starting mark.
  • Guidance after a post Tally refused (#1108): the wording is corrected on every path that can be reached (one voucher, a whole batch, the summary line, a later check, a post whose answer was not clean). It was not run live: no way is known to make Tally refuse a post after Bridge's own checks pass, so the issue stays open for 0.7.0. A ledger rename and back is not run. Labelling a batch Tally created only part of is not built.
  • Port locks on a Microsoft Store install of Claude Desktop (#1584) and where the import file really is on a Store install (#1459): not run.
  • Voucher ranges and direct posts on a book of about 30,000 vouchers (#1029) and a batch post that completes in Tally but whose reply is not kept (#852): not run.
  • Not run on a live TallyPrime for this build, with only captures or tests behind them: gst_uncertain_vouchers, negative_stock, negative_ledgers, masters stock_items, the closing stock used by profit_and_loss and balance_sheet (#692), cost-centre allocations in voucher reads (#1231), ledger aliases (#1688), the check of a statement batch when it is sent (#1714), except the one refusal named in the live pass of writes below, the refusal of a post whose last read shows no voucher mark (#1945), the amendment reference check (#1941) and the invoice-stop window title (#1828).

Live pass of writes: run on 27 October 2026 on the CI package built from commit 67f647e1, the commit this release is cut from; that package reports 0.6.0 as its own version; Tally was TallyPrime 7.2 Silver. Two rows ran, each on its own synthetic company. (1) On BRIDGE NUMBERADM LAB: one approved post under automatic voucher-type numbering, and the preparing of a post refused under manual numbering (#724): both passed. For the approved post, two Receipts were keyed first (numbers 1 and 2). The approved batch post, answered by the person in the approval window, posted and verified (posted_verified, the voucher mark up by one step, the post bound to its voucher, no rollback and no mark contradiction), took number 3 and left the two keyed vouchers unchanged; the approval window's first line (#1832) was seen above Tally. For the refusal, with the Receipt type set to Manual numbering, building the post was refused before anything was sent (import_voucher_numbering_not_automatic, refused type Receipt), the voucher mark and the vouchers were unchanged, and the company was restored from its backup afterwards. (2) On BRIDGE TIE OUT LAB 3: the send-time check of a statement batch that would post a row which a statement of this company skipped as carried by the other account's Contra (#1714): The batch was refused at send, in 0.17 seconds, as skip_row_paired_elsewhere. Its next step says nothing was sent to Tally, any approval of the batch is withdrawn and the batch must be built again. No attempt was recorded; the voucher mark, the batch journal and the skip records were unchanged, and no voucher appeared on that day. A control post of a fresh one-voucher batch was not completed: its approval window timed out unanswered and nothing was posted. The approved post in (1) shows a post going through on the same build. The check against Contras was not run live. The file that ran is the CI package; the published file is built again by the release workflow and is checked after publication against the release's checksum, with no write.

Known limits

  • Not run on a live TallyPrime for this build (unless the lines above say otherwise): anything that writes to Tally other than the approved post and the refused statement batch of the live pass of writes above, including a batch of Journals, a batch near 50, a post Tally refused or that landed partly, amendments, the review acknowledgement window (no voucher raised a doubt), a post that needs an On Account approval (a bill-wise party; the build refuses one as designed) and an import file taken to Tally's Import menu; Sales invoice posting and ledger creation (not offered); the desktop app and its Terms of Use window (unit, component and browser tests only; the packaged app was not run); Gold; macOS; any Windows or Tally release other than Windows 10 with TallyPrime 7.1 Silver for the reads and TallyPrime 7.2 Silver for the three rows that ran on it (the live pass of writes, the restore row and the empty-window count; the other reads listed above have not run on 7.2); a second monitor. The earlier live passes and the 0.6.0 package check are recorded in the 0.6.0 section below, and say nothing about the file in this release.
  • The file tested is not the file that will be tagged. Each live row names the build it ran on: voucher_read_cost on the published 0.6.0 build, the window fit on a CI build of #1879, the empty-window count (#1240) on a CI build of #1897, the restore row (#1814) on master build fd6f7c0d; the write pass ran on the CI package of the commit the release is cut from, which reports 0.6.0 as its own version. The release workflow builds bridge_mcp again on its own, with a plain release build of that binary, where the CI package takes it from the Tauri build; the two files are the same source but not byte-identical, and no build is claimed to be reproducible. The published file is checked after publication against the release's checksum, with no write.
  • Approval window fit. Measured on one Windows screen (1920x1080 at 125 percent). On other screens a window can be refused as too large, or fit, where it was not measured; on a Mac the alert's size is not checked and a batch window can be taller than the screen (#1780).
  • The headline of a batch post has not been seen on a live Tally post (#1090, #1532).
  • What has been run against a real TallyPrime is listed in the README.

0.6.0, #

In plain words: ComplyEaze Bridge 0.6.0, since 0.5.1 (16 Oct 2026)

These changes are in ComplyEaze Bridge 0.6.0. Most lines name the pull requests or issues they come from. Like every build so far, it is not yet code-signed.

Should I upgrade?

  • If you use posting, or want to post batches: yes, read this first. One setting, Allow voucher posting, one voucher or a batch of up to 50, now covers a single voucher and a saved batch of 2 to 50. It is a new setting and is off by default: if you turned posting on in 0.5.1 or earlier, turn it on again after installing 0.6.0. Every post still waits for your approval in a separate window, and for a batch you approve its summary (#1090, #1630).
  • If you use the bank-statement tool: a skipped row must now name the other account's ledger, and a transfer between two of your own accounts is checked so that it is not booked twice (#1585, #1625). Several refusals are new; see "What works differently" below.
  • If you only read your books: the new reads below are the reason to update, if any. Nothing in 0.6.0 writes to Tally unless posting is on.
  • How: ComplyEaze Bridge does not update itself. Follow the installation guide: (1) close any other program that runs ComplyEaze Bridge; (2) in Claude Desktop, Settings, Extensions, install the new file; (3) keep ComplyEaze Bridge's data folder; (4) check that the extension shows 0.6.0, that the Terms setting is ticked, and that Allow voucher posting, one voucher or a batch of up to 50 is as you want it; (5) quit Claude Desktop completely and reopen it.

What you can do now

  • month_register is a new read-only tool: Tally's own month-wise Sales Register or Purchase Register for whole months, one row per month with Tally's debit, credit and closing columns exactly as printed. A window that is not whole months, or that starts before the books, is refused. The rows are not tied to the vouchers, to the trial balance or to sales_register and purchase_register, and not compared with Tally's register screen. Its request renders as the lab sent it, and its parser reads the answers BRIDGE PILOT LAB gave on 13 October 2026 (one synthetic book, TallyPrime 7.1 Silver). The Receipt and Payment Registers are not read (part of #1232, #1759).
  • A transfer between two of a firm's own bank accounts, printed on both statements and booked once, is now followed from one statement to the other. A skip row must name the other account's ledger (a skip with no ledger, or one naming the statement's own bank ledger or the suspense ledger, is refused at its row), and the parse records each row's network transfer reference and printed instrument, for the transfer references both banks print. A build records each Contra and each skipped row, and a successful build and the first page of a verification carry a transfers section: each skipped row is paired with the other account's Contra by reference, then by date, or listed in one of its open states (not found, ambiguous, other side unverified or not established, among others), with counts only, no amount, reference, party or ledger. A build that would book a transfer a second time is refused (part of #1392, #1585, #1592, #1625). Run once, on 15 October 2026, on a Windows lab PC (TallyPrime 7.1 Silver, one synthetic company) with the Contras imported by hand. Not measured: which banks print the same reference on both statements of a transfer, how many days apart two banks print one transfer, and posting a batch that holds a transfer with post_import.
  • masters with kind voucher_types now returns, on each row, series_numbering: the numbering method (automatic, manual, automatic_manual_override, or the raw text under unrecognised, never read as manual) and whether duplicate numbers are prevented, read at the number-series level from a second paired read of the voucher-type export and joined to the row by GUID with the name as a second equality; a type with no series or several returns series_count and nothing is read from it, and a series missing or repeating a field returns field_invalid. Under automatic, and under no other method, a row also carries sub_method, the series' own setting for the numbers of vouchers already entered: auto_retain ("Retain Original Voucher No."), auto_renumber ("Renumber Vouchers"), absent where the series printed none, or the raw text under unrecognised. A type found in one read only (masters_numbering_names_differ), or a GUID and name that disagree between the two reads (masters_numbering_identity_differs), refuses the whole read; so does the numbering read changing between its two reads (masters_numbering_collection_changed) and a numbering answer whose row count or size breaks the mark's premise (masters_bound_premise_violated), and one that does not parse (a missing or repeated GUID among the causes) refuses at once; a moved book is reported as masters_extent_changed in place of a join or premise refusal. The type-level numbering_method stays as it was, with its limitation. Measured on one synthetic book (TallyPrime 7.1 Silver, Windows): 24 voucher types, each with one series named Default, of which three (Journal Automatic, Payment Manual with duplicates prevented, Receipt Manual with duplicates not prevented) were checked against Tally's screen. Automatic (Manual Override) through this tool (it is mapped from the series text measured on a synthetic book on 13 October 2026), more than one series, Multi-user Auto, a single series not named Default, a type created new and other releases are not measured. The two requests were run live as one tool call on 14 October 2026 on two synthetic books (TallyPrime 7.1 Silver, Windows; 24 voucher types each, exactly one series per type, its name not read): the read was not refused, and on the six types whose numbering was set or checked on Tally's screen it returned that numbering (Manual with duplicates prevented, Manual with duplicates not prevented, Automatic); the other types, left at a new company's defaults and not checked on screen, read Automatic (part of #724). sub_method was measured on a second synthetic book of the same release and edition, on one voucher type (Receipt, set to "Renumber Vouchers" and back); any other wording, a series that prints none, Gold and other releases are not measured.
  • ledger_masters with fields=compliance takes an opt-in include_gst_details (a boolean, default false; with any other fields it is refused as ledger_masters_gst_details_requires_compliance): each ledger's row then carries compliance.gst_details, Tally's dated ledger-level GST rate details as sent, per entry and state row, with a rate that is absent, empty, zero or a value (the text as sent and the decimal), never merged (part of #1342). HSN/SAC and type of supply are not returned because no capture shows them, the place of supply stays in the registration history, and slab rates are only flagged. Without the flag the requests, the sizing and the result are unchanged. With it the read is admitted at 8,000 bytes a ledger instead of 3,750, a margin over the one capture (a ledger with an empty list 2,138 to 2,704 bytes; one dated entry 5,248 to 5,394; UTF-16 bytes of the ledger's own element), not a measured ratio; a large book may be read in more parts or refused as too large. Measured on one 14-ledger lab book of TallyPrime 7.1 Silver and not yet run through this tool on a live Tally.
  • stock_summary now returns an item's closing quantity, closing.quantity with its amount and unit, where Tally's own plain Stock Summary has exactly one line with the item's name and shows the same quantity, unit and amount; every other item says why none is returned (none_sent, unread, inside_stock_group, report_has_no_line, report_name_not_unique, report_differs) (part of #1233). An item inside a stock group returns none: in the captures, the report listed only what sits directly under the root (one capture holds only groups, the other only items under the root; none holds both). Checked against two real captures of one synthetic book taken after a sale of an item that had no stock: Tally kept it as -5 Nos with no value, in the item rows and in the report alike, and the parser and the tool's tests read it. Then run through the tool on a live Tally by an outside contributor, on a new synthetic book built to the same list (TallyPrime 7.1 Silver): the answer equalled Tally's own Stock Summary for 31 March 2026 on every point, the closing quantity and unit of both items, the empty value of the negative one, and the book's stock-item count (2 of 2). Not measured: a book whose report lists a group and an item under the root side by side, an item inside a stock group, a compound unit, other dates, other releases, macOS, and a group that carries an item's name.
  • statement_tie_out is a new read-only tool: after parse_bank_statement, it says whether the bank ledger in Tally stands where the statement says it stood at the start and at the end of the statement's dates, as three gaps (book less statement): the opening gap, the closing gap, and the change in between, before an import or after one made by hand. It covers a whole statement, or whole calendar months of it, names no cause and writes nothing. It returns the two dates and the gaps, not the balances, though a gap plus the statement's balance at that date (for a whole statement, the balance you supplied) gives the book's. The parse's proposals file now also records the statement's first and last dates (part of #1211). It was run once in the lab on 14 October 2026 (TallyPrime 7.1 Silver, one synthetic company): whole months before and after their post, a whole statement whose last month was not posted, and reads after the ledger's last voucher all gave the gaps expected; a read at a date in the middle of a month is not measured (see the pull request).
  • statement_tie_out can check whole calendar months of a longer statement: a file parsed with from and to (or one of them) from the 1st of a month to the last day of a month, within the statement's rows, is compared with the book at those dates, against the statement's balance there as worked out from its own rows. Any other narrowed window is not_established (window_not_whole_calendar_months, window_beyond_statement_rows). With the book's figures, the gaps then give the statement's balance at those month ends and its net over those months: never a part of a month except the part-months at the statement's own start and end (whose net a month-end balance gives together with the opening and closing supplied to the parse), though a month in which every other row's amount is already known (one row alone, or one row beside open cash lines, whose amounts the parse returns) gives that row's amount (part of #1211). No month window has been run against a live book.
  • gst_status reads Tally's own GSTR-1 status of the vouchers of one voucher type in a window. For each voucher it says whether Tally lists it as included, uncertain or excluded, or does not list it in the return (measured for Receipts only), or is cancelled, optional or post-dated (decided first from Tally's own flags, never counted as included, with what its status flags read returned beside it; #1655, #1664), and whether a person accepted it as it stands (Accept As Is). The status is Tally's own flag and not a judgement of ComplyEaze Bridge; the reason for a status is on Tally's screen only, so none is returned. A voucher whose flags are absent or odd, or whose deleted flag reads Yes (never seen), is returned as unread, never as included. The voucher type is named exactly (a type of the same class under another name is not read), a window is at most 93 days, and a book whose voucher mark is above 25,000 is refused. Measured on one synthetic book of TallyPrime 7.1 Silver, for the unregistered buyer of its Sales invoices; a registered buyer, other releases and Gold are not measured, and an excluded voucher has not been observed.
  • company_features reads three settings that a company's record in Tally holds today (cost centres, GST and batch-wise stock) and its currency symbol (part of #1231). Each setting is yes, no or not reported, and says what it rests on: each of the three was compared with Tally's own F11 screen on three synthetic books (nine values, all equal; not this company). A setting Tally did not send is never read as no, and an empty or odd one refuses the read. The settings are the ones Tally holds today, not what the books contain or what they were during a year. The currency is a symbol, not a currency code. Measured on three synthetic books of one TallyPrime 7.1 Silver; a setting changed inside a sitting, Education mode and other releases are not measured.
  • voucher_read_cost says, before a call, the most census reads one call of vouchers, ledger_movement or verify_import will send on a company (a read can be refused before its census ends), and the least time a call that is not refused earlier spends on the gaps between them (part of #1465, after #1239). It reads the company's identity and then its marks (the opening every windowed call makes) and sends no census and no voucher request. The census follows the book's voucher mark, not the window, so the figure is for one call: a narrower window does not lower it, and N calls pay it N times. The count and the floor are derived from the mark and the read's own rules, not measured on the book, and they leave out the data parts and the wider re-read an empty window gets in vouchers and ledger_movement; no estimate is made. Beside them are the whole calls the protocol reference records, with their census reads and data parts (one book, about 250,000 vouchers, 31 census reads), as data marked as not a bound. Nothing is refused on it. Not measured: Windows hosts, and any book other than that one.
  • parse_bank_statement reads Bank of Baroda statements in one measured layout (bank: "bob"; Part of #1457). The layout was measured on two real statements, held privately, and the profile follows what was seen there: the statements print neither an opening nor a closing balance nor totals, so both balances come from you, and every page must print Page N of M and every page after the first must start with a row. They print the account number masked (first and last three digits clear), so account_label must carry the whole number; only those six digits are compared, the masked middle cannot be. With no printed totals those balances and the footers are the only checks on which rows were read: missing rows whose amounts net to zero would not be seen. They print the newest row first; the rows are read oldest first and numbered that way. A statement is read as Bank of Baroda only if page 1 prints an IFSC Code with the bank's prefix on its IFSC Code: line; otherwise it is refused as statement_bank_not_recognised. The measured statements cut a narration at about 50 characters, so a name can end mid-word. A party is named only where the shape settles it: a UPI row names its VPA (when the VPA is whole), NEFT and RTGS rows name what follows the UTR (not when a line was broken at the cell edge inside the narration), and bank charges and a cash deposit are recognised by their wording; a loan recovery is named as printed, loan number included, so two loans are two parties. Two payers whose names agree up to the cut read as one party. IMPS, mobile-banking and internet-banking rows, interest and every other wording go to suspense. A statement that opens without a password needs password_file to be an empty file (on a Mac it must be owned by you with no group or other permission bits; on Windows it must be owned by you or Administrators, and no one but you, SYSTEM and Administrators may read or change it). Not measured: other account types and branches, statements printed oldest first, other wordings, and a statement downloaded another way.
  • vouchers can keep only the vouchers of a ledger whose narration ends in a suspense tag ComplyEaze Bridge writes (suspense_tagged: true with ledger, #810 slice 2): "purpose not confirmed" for a cash line answered "don't know", or "UNIDENTIFIED - reallocate from" the ledger for a party no mapping names; each item says which in matched.suspense_tag. It reads the tag from the narration, so a voucher whose narration was edited after the tag is not listed and the same text typed by hand matches: read the ledger without the filter too. Not measured: a tagged narration read back from a real Tally. The same tag test now also counts, in a build's suspense_lines, a voucher whose tag names its ledger in another case, spacing or hyphenation.
  • post_import (after a send was read back) and verify_import now begin with a headline for a batch ComplyEaze Bridge sent: whether the batch is confirmed in Tally (a batch of several vouchers also says how many of its vouchers read back and how many are not confirmed), with the counts, and a count for each other status, in batch, counted from the result's own rows. A batch whose vouchers all read back but whose readback as a whole is not confirmed is not called confirmed, and neither is a batch of which only some vouchers read back: its lead says that they read back and that the batch is not confirmed. A batch not yet posted, a file imported by hand and a refusal before anything was sent have none. The result is unchanged (part of #1090).

What works differently

  • One setting now covers all posting: with posting on, ComplyEaze Bridge posts one voucher per approval, or a saved batch of 2 to 50 Journals, or of 2 to 50 Payments, Receipts and Contras, after one approval of the batch's summary (#1090). In the extension it is a new setting, Allow voucher posting, one voucher or a batch of up to 50, off by default: it is a new setting, so if you turned posting on in 0.5.1 or earlier, turn it on again under the new text. On the command line, BRIDGE_AGENT_ENABLE_WRITES now admits batches too, and the separate BRIDGE_AGENT_ENABLE_BATCH_POST is retired: while it is set and posting is on, post_import refuses as setting_retired until you remove it. A batch saved earlier that was refused as more than one voucher now posts after its own approval.
  • post_import and build_import_xml now refuse a voucher sent without a voucher number into a voucher type whose voucher numbering does not read Automatic, as import_voucher_numbering_not_automatic, and into one that reads Automatic without the sub-method Auto Retain (the screen's "Retain Original Voucher No."), as import_voucher_numbering_not_retain. Either names each such voucher type and what was read in refused_types, and a batch that holds both is refused as import_voucher_numbering_not_automatic (part of #724). Under Manual voucher numbering, a second voucher sent without a number was seen to take a number already used, or to be refused whole. With a voucher type set to "Renumber Vouchers" (the series reads Auto Renumber), Tally was seen to change the numbers of vouchers already in the book: at the switch, when a voucher with an earlier date was entered by hand, and when a voucher with an earlier date was sent by import without a number, without moving any voucher's AlterID and with the import's answer reading CREATED 1, ALTERED 0; a voucher dated after every other changed nothing (one synthetic book, one voucher type, TallyPrime 7.1 Silver). A series that prints no sub-method, or one ComplyEaze Bridge does not recognise, is refused the same way. An amendment's vouchers are checked the same way. It is refused before its own check of the book, so its next step says that whether the earlier batch's vouchers are in the book was not checked, and has the assistant call verify_import for that batch first: a voucher found in the book is never entered again, one that is not found is not built as a new voucher either, and the vouchers of the refused voucher types are left out until ComplyEaze Bridge numbers vouchers itself; no refusal sends anyone to enter a voucher in Tally. A Journal that carries its own voucher number (one built for Tally's Import menu) is checked by build_import_xml too: it is written into a voucher type that reads Manual with duplicate numbers prevented, where Tally keeps the number and refuses one a Journal already holds, or Automatic with Auto Retain, where Tally discards the number and assigns the next, and it is refused otherwise under the same two codes, Manual with duplicates allowed included, where a Payment carrying a held number was created as a second voucher. It is refused on a voucher type set to "Renumber Vouchers": there Tally was seen to discard the number a Receipt carried and give it the number of its place in date order, and the later-dated Receipts' numbers went up by one. So a build that holds only Journals carrying their own numbers, which made no numbering read before, now makes that read: it can fail with that read's own codes, or as pre_import_mark_unobserved when the company's master mark cannot be read, and it is refused on a book in Education mode, as a build of vouchers without numbers is. On the desktop's Journal screen the refusal names the voucher type and what was read of it. The next step, and a post's message naming the voucher types, are sent only where BRIDGE_AGENT_MAX_BYTES is at least 8,192; from 4,096 up to that a refused build still returns its reason and refused_types, and a refused post its code and refused_types. On a type set to "Renumber Vouchers", a file of more than two vouchers, a mixed file, a Journal that carries a number or is dated after every other, Payment and Contra are not measured; nor are Gold, other releases, a series not named Default and the same through ComplyEaze Bridge's own post: on a release, an edition or a book whose series prints no sub-method, or words it differently, every Automatic voucher type is refused.
  • Before approval a post now reads the company's master mark and then the voucher numbering, and compares the mark again just before the post is sent: any master saved in that span, while the person decides included (a ledger created, a voucher type changed), refuses the post as post_masters_moved and withdraws the approval; post again. Each of thirteen saves of a voucher type's screen in Tally moved the mark by 2 (nine of a type holding no vouchers, and two in each of two runs on a type holding vouchers, all on one synthetic book, TallyPrime 7.1 Silver). A voucher type saved during the read itself refuses as masters_bound_premise_violated (cause masters_alter_id_above_master_mark). A book in Education mode is now refused, as masters_education_unqualified, for a post and for a build. Each pass of a post sends 20 more requests to Tally before the approval (12 of them POSTs), so a held approval redeemed in a later call sends them again; a build sends 14 more (8 POSTs), under its admission lock.
  • `verify_import` now compares a voucher's narration. A posted voucher whose narration in Tally differs from the one approved reads posted_divergent, with narration in its diffs; before, every other compared field matched and it read posted_verified, and a divergent voucher never listed a narration difference. A narration edited in Tally after the post therefore leaves the batch unsettled until the narration in Tally matches the approved text again. The read-back, the post's span binding and the amendment check now share one comparison: the text as written (trimmed, with its [BRIDGE:…] tag for a file imported by hand) against the text read, byte for byte; a narration Tally holds none of equals only an empty one written. Whether Tally changes trailing spaces or line breaks on import is unmeasured; the first live run will show it (#1608).
  • On Windows, parse_bank_statement now refuses a password file owned by neither you nor Administrators, or that anyone but you, SYSTEM and Administrators can read or change (statement_password_file_permissions), as a Mac already refuses one that other users can read. The refusal says how to make the file private: make a new copy in a folder of your own and delete the original, because moving a file keeps who can read it. A password file on a drive that keeps no access list, such as a FAT or exFAT stick, is expected to be refused as well; that is not yet measured.
  • The hand-import script scripts/bank_statement_import.py (not part of the downloaded package, #1430) now writes the same | Statement party: <name> segment under the same rules as the app, and refuses a run whose narration holds a control character or the reserved marker, or runs past 2,000 characters, as narration_not_admissible, as the app does (party_not_admissible when the party's own segment is what breaks the rule). Importing its file again for a statement already imported is expected to rewrite those vouchers' narrations in the book: their keys do not depend on the narration the script writes. A same-key re-import has been measured over the gateway; through Tally's own Import menu, which is how this script's file is imported, only an unchanged re-import of a file the app built has been seen, on one Windows computer (#1413).
  • A character that draws nothing (Unicode Default_Ignorable_Code_Point or Format, except ZWJ, ZWNJ and the prepended concatenation marks) is not read from a bank-statement PDF, and a soft hyphen is read as the hyphen it prints. A narration holding such a character is refused as voucher_text_invalid when a batch is built, and before a native post of a batch saved earlier; the reference is not checked. A statement that held one gets new bridge_txn_id labels, so a batch built from it by an earlier build does not line up with a new run's amends_batch_id.
  • verify_import no longer reads its window a second time when the first read was divided, every part was checked against a count of the window that names each voucher, and the company's marks were unchanged when read again after the last part (#1241). That is a window of more than about 170 vouchers in a book whose voucher mark is above about 170; it holds every voucher of the book in those dates, not only the batch's. The proof's voucher_read_corroboration then reads {"state": "not_sent", "reason": "counted_and_bracketed_read"} instead of a second read. A window read in one request, or divided without a count, is read twice as before. ledger_movement still reads its window twice. The request counts come from test doubles, not from a live Tally.
  • Verifying a batch no longer replaces its saved proof, so the host need not ask before each read-back. Each verify_import, and each of the two verifications acknowledge_post_review runs, now saves its proof as a new pair of files, <batch>.proof.<time saved>.<SHA-256>.json and .md, beside every earlier pair, and the status it adds to the import journal names which pair is current. The finished masters check and the first verified ALTERIDs go to files of their own, written once. Nothing these two tools find in the folder is replaced or deleted, so they are now marked as writes that are not destructive. The folder therefore grows with every verification; nothing deletes old proofs. A batch last verified by an earlier build keeps its single <batch>.proof.json, which no longer changes. After the first verification with this build, a connector older than it refuses the import journal (a downgrade was already refused after the first native post), so keep this version for recovery (#911).
  • On Windows, the approval window now opens above a full-screen Tally (seen once). It could open behind TallyPrime when Tally filled the screen, so the person saw nothing and the approval lapsed unanswered. Both windows, the post approval and the review acknowledgement, are now drawn above every window that is not itself kept on top. Nothing else changes: the text, the buttons, No as the default and the two-minute limit. Seen once, on one screen, in front of a full-screen Tally and in front of Claude Desktop; the same run found that the window takes keyboard focus (see "Known limits") (#1589).
  • A one-voucher approval now shows a long narration in parts, where it refused it before. A bank-statement voucher's narration is often too long for one line of the dialog, so posting it alone was refused and the file had to be imported by hand. Such a narration is now shown under the heading in numbered parts, each quoted on its own line and split only at a space, so that the parts joined in order are the text that is posted. A line above the heading counts the parts, and the date line says the voucher's own text is below the dashes. It is still refused whole, never cut, when no space splits it into parts that fit, when it holds a run of dashes or a number in brackets, or when the whole text passes the dialog's caps. A reference is never split: one too long for its line is refused as before. Each part past the first, and the counting line, leave one line fewer for entries. The review dialog of acknowledge_post_review now shows the narration last, under the same heading, on one line or in parts, which leaves it room for nine entries instead of ten. An approval text that fits as before is unchanged. Not yet measured: how the Windows and macOS dialogs draw the parts (#1535).
  • A `post_import` refusal made before anything was sent now says what to do next. Its error.next_step carries the same step every other tool gives for that refusal: its code's, or else its cause's. Before, a code raised inside the post, such as a date range too large to check in one request, came back with no step. Only when the journal shows nothing was sent: a refusal after a send, or when a send could not be ruled out, still carries no such step, and is reconciled with verify_import. The steps the post already gave (a busy port, a row another batch posted) are unchanged (#1573).
  • A batch approval now shows its totals first. The window for a batch shows, in this order: the count and company, the company number, the count by type, the dates, the total debit and credit and the money lines; then the company GUID and the Tally address, the voucher lines or the sentence in their place, the per-ledger lines, the batch id and the cautions. Every line keeps its wording. The window for one voucher is unchanged. This order has not yet been seen in a running window on Windows or macOS (#746).
  • A batch approval's Journal lines now name a ledger on each side, and the window no longer says narrations are shown exactly. In a batch of up to ten, each Journal's line named only its first debit ledger, so the person could not see what each Journal credited; it now shows Dr "<first debit>" +N more Cr "<first credit>" +M more, with up to 20 characters of its narration (other types keep 40), cut shorter, never under 10, when that makes the line fit. A Journal line that fits at no such cut leaves the totals and the reason line, as any line that does not fit does. The line above the voucher lines said each narration was quoted "exactly as it will be posted", above narrations cut at 40 characters; it now says a cut narration is marked. Payment, Receipt and Contra lines still name one ledger. On every line, the further entries on a side now read "+N more", never a bare "+N" that could be read as an amount.
  • The approval dialog of post_import now takes a voucher on a ledger whose name in Tally ends in one line break; until now such a voucher could be built and was refused for posting. The approval dialog, and the review dialog if a doubt is recorded after the post, show such a name quoted with the line break written out and the words "Line break" before the quote, on every line that shows the name (Line break "Name\r\n"), and add one line saying what the mark means and that no other name in that text has one. A line break anywhere else in a name is refused as before, and no dialog shows a line break in any other value. The desktop Journal screen takes no saved Journal that names such a ledger, to review or post, or to reconcile once it was sent: it says so and says to use the assistant for it. An approval text that carries both this line and the On Account sentence holds one entry fewer. The mark takes 12 of a line's 100 characters and the written-out break 4: a text with a line over the cap is refused, never cut, except that a batch's per-voucher lines give way to a reason line. Measured once on a licensed TallyPrime 7.1 Silver on Windows: one Journal on such a ledger posted through the dialog, Tally read the name back with the break as &#13;&#10;, and the post ended verified; the read-back is now the test's fixture. Not measured: how macOS draws and wraps the mark and the quoted name, and the review dialog against a live Tally (#626).
  • The review dialog of one voucher (acknowledge_post_review) now shows the voucher's type quoted (Record that you reviewed ONE "Journal" in "Books"), as it shows every other value read from Tally, so every Tally value in that text is escaped (#626).
  • An empty window whose neighbouring days hold vouchers now says so in its reason. When the read of a window returns nothing and the read widened by a day on each side finds rows, none inside the window, vouchers, voucher_presence, purchase_register, sales_register and ledger_movement now give empty_window_wider_read_nonempty where they gave nonempty_uncorroborated, which read as if the window itself held rows. Only the name changes: the state stays partial (a refusal, for ledger_movement), and an empty window next to a busy day still cannot reach complete. The other reasons (empty_uncorroborated, company_has_no_vouchers, window_contradicted) are unchanged, and the vouchers and voucher_presence tool descriptions and the agent guide now list all three outcomes of an empty window.
  • The install steps are rewritten for a first-time accountant. The installation guide gives one short action per step, the exact words on each button, what the screen shows next and what to do when yours differs; the download page and the home page give the same steps in short form, and the download page points to the guide. They name the three things a new user meets in Claude Desktop (the notice that the extension is not verified by Anthropic, the confirm box, and the permission prompt before a tool runs), give TallyPrime's path (F1, Settings, Connectivity, Client/Server configuration), and say to leave posting off. The guide now lists what the data folder holds, says the list is not complete, and includes the last four digits of a statement's account number kept with its proposals, the ledger names chosen for a statement, the sign-in name a recorded review can include, and the record of accepting the Terms. TallyPrime's Step 1 screens, the browser check and the Alt+K and Alt+Y menus were checked by looking on Windows (TallyPrime 7.1, 10 October 2026); the guide says which steps were not tried (#1017).
  • The texts that say what Response redaction leaves unhidden now also say that the last four digits of a bank statement's account number (account_last4 in the answer of parse_bank_statement) are always sent: the extension's description and its redaction setting, the two tool descriptions that describe redaction, the README, both guides, the security page, the FAQ and one article. Only the wording changed; what is sent is the same (#1616).
  • parse_bank_statement, build_import_xml and the agent guide now tell the assistant to import a bank statement in whole-day batches: cut a batch only between days, at most 50 vouchers. post_import refuses a whole batch with import_preexisting_identity when a voucher has a twin with the same date, type, ledgers and amounts that an earlier batch posted, while twins inside one batch pass together; a statement cut across two batches on one day hit that refusal. A day with more than 50 rows does not fit one batch, and the text says to ask you. Batch posting is off by default, so same-day twins post together only with it on or through a hand import of the built file. Wording only: no new refusal and no change to the check (#1614).
  • verify_import's description now says in plain words that after a restore from a backup, a native voucher keyed again with the same content before any check can read verified, that no identifier a read returns tells the two apart, and that after restoring a backup over a company you should check its posted vouchers in Tally. Tally gave the vouchers keyed again the post's MasterIDs, GUIDs and voucher numbers (measured once, on one synthetic company on TallyPrime 7.1 Silver; the run that compared every identifier had Tally's security off); the description had called that reuse unmeasured. Behaviour is unchanged (#1612, #1636).
  • The installation guide now says that a batch posted from the command line is approved on a summary of per-ledger totals and, when there are ten or fewer vouchers and the lines fit, a line for each voucher; it said the summary showed no voucher's date or narration. The window has listed each voucher of such a small batch since 0.5.0 (#1632).
  • A native post whose binding was refused now says what the book holds. When verify_import cannot bind a batch to the vouchers its own post created, post_span_binding.summary says the batch stays open (except when every voucher reads tally_reported_not_created, where the summary differs) and, when every voucher reads matching_content_observed against a voucher that is neither cancelled nor optional (or only some do), that the book holds, for each voucher sent (or for some), one with the same date, voucher type and ledger entries, and that ComplyEaze Bridge cannot tell whether that voucher is this post's; it never says the post's vouchers are in the book. local_data_report counts such batches as not_settled_binding_refused, and not_settled_not_verified no longer includes them. The post_import description now says no review can be recorded for such a batch: acknowledge_post_review refuses it, because its vouchers are matched by content only, so the person checks them in Tally and the batch is not rebuilt (#1469).
  • A refusal because Tally did not answer now carries a next step. endpoint_unreachable tells the assistant to ask you to confirm that Tally is running with the XML server enabled on the port named in the error, and not to repeat the request until you have; endpoint_circuit_cooldown (a request held back after repeated failed requests to Tally, nothing sent) says it can be repeated after about ten seconds. A code's own step still comes first, and the other unanswered reasons still name none. The parse_bank_statement description and the security and privacy notes in the repository also now say that | Statement party: <name> sits at the end of the narration text, and that a file imported by hand adds the [BRIDGE:...] tag after it (#1483).
  • The Terms of Use setting is now told as "tick" wherever a person is asked to set it: the refusal a tool gives before the Terms are accepted, the installation guide, the connection guide, the download and home pages and the FAQ said "turn on" or "switch on". "Turn on" stays for the HTTP gateway and the posting setting (#1484).
  • The masters tool no longer calls a voucher type's numbering value its numbering method. Its description and its result's limitation now say the value is the one Tally reports at the type level, that it read None for types set to Automatic and to Manual on Tally's screen (one synthetic book, Windows), that it can differ from the number series that decides what a supplied number does, and that it is not to be passed to voucher_presence's numbering. That input now says the method is the one you confirm for the series used. What is read is unchanged (#1487).
  • The Outstandings screen's ledger-master export now ends its balances at the screen's "As of" date, so a workbook can be exported as of 31 March; it ignored the field and used the date of the computer. With no date chosen nothing changes, and the assistant's ledger_masters still uses the computer's date. A date before the company's books begin is refused as as_of_precedes_books_from before any master, balance or group is requested. Only the 1st, 2nd and 31st of a month are accepted as a balance-period end, so a typed 30 June ends on 1 July and that day's vouchers are inside the balances; the notice says so (#1531).
  • The security and privacy notes in the repository now say the 0.5.1 Terms refusal (terms_not_accepted, naming the setting) was the answer the tools the assistant called gave on one Mac with a test package and on one Windows computer with the build our checks made for 0.5.1; they had said it was not yet checked on Windows (#1543).
  • The README's table of tools now has a row for each tool offered with posting off, adding company_features, gst_status and voucher_read_cost (#1672).
  • The FAQ answer and the security page about a second computer reaching Tally's gateway now state the rule: do not assume the gateway is closed to your office network; the firewall of the computer that runs Tally is what keeps other computers out, so confirm it blocks incoming connections to Tally's port. They keep the scope of the one test on TallyPrime 7.1 (a second computer got the status answer; nothing else was checked) and no longer say it was "not recorded in the repository" (#1685).
  • The README and two blog pages no longer say that the record of a test run is our dated notes, "kept privately"; what was and was not run is stated as before (#1528).
  • The core for posting one GST Sales invoice is in this build and switched off. Sales is not in the qualified list, so every Sales build is refused before anything is sent to Tally, no file is written, the tools' schema does not offer the type, and a saved Sales batch is refused at post. A Sales invoice that was sent and is not verified posted would stop every further Sales invoice of its company until it verifies or a person releases it; that cannot happen while Sales is off. For Journal, Payment, Receipt and Contra one change is accepted: "invoice": null is read as absent. A first rehearsal on 7 October 2026 posted two invoices and found what must change before Sales can be switched on. In 0.6.0 Sales stays switched off, and a Sales invoice would be built only when voucher posting is on, so there would be no hand import of one (part of #691, #1342, #1463, #1488, #1491).

Safer or fixed

  • On Windows, every request to one Tally port is now also serialised through a lock file under your user profile (.complyeaze-bridge\coordination), in addition to the one under AppData. Windows writes a Store-installed Claude Desktop extension's AppData files into the package's private folder, so the extension and an unpackaged ComplyEaze Bridge process could each hold "the" lock in a different folder and a post could overlap a read on the same Tally. A process that cannot create that folder refuses every send as import_admission_lock_unavailable. Not measured: the cost of the extra lock per send (#1584, #1641).
  • build_import_xml shows where the import file really is when a Microsoft Store install of Claude Desktop keeps Bridge's files in the package's private folder, so File Explorer and Tally can open the path it prints; a new path_state field says whether the path was resolved. Where files are stored and read is unchanged. Measured once, partly, on one Windows PC with Claude Desktop from the Microsoft Store, on 15 October 2026 (#1459, #1569).
  • When the local Tally mirror's key is missing from the computer's credential store, or is stored but is not a valid key, the app now says so with its own codes (tally_mirror_key_missing, tally_mirror_key_invalid) and tells you that a retry will not help, instead of asking you to approve a credential prompt and retry. Restoring the credential (the macOS login keychain, or that user's Windows Credential Manager) or opening the app as the same computer user is what recovers the mirror. The app deletes nothing, and no reset exists yet (#1563). A key is also no longer created when the mirror file's existence cannot be checked. Not measured on a real credential store: what the macOS keychain or the Windows Credential Manager returns after a keychain reset or a profile move.
  • A later check (verify_import, a repeated post_import, or the check before a review is recorded) of a post that Tally answered as creating none of its vouchers no longer suggests the voucher was edited in Tally. Its status is unchanged (sent_not_attributed, the batch still reconciliation_required), but the voucher's next step now says that Tally reported it as not created when it was sent, that it is not found for these dates now, and to look for it in Tally before entering it by hand, since it may have been entered in another form meanwhile. This reads what the post left on record: its saved answer (every counter reported, none created, one exception for each voucher sent, nothing else counted) and, for a batch, its own verdict that the voucher step matched. A partly created batch, and a batch whose voucher mark moved across the post, keep the earlier line. Measured live once, on a single voucher refused whole; a batch is not measured live (#1108).
  • A voucher of an unbound native post that ComplyEaze Bridge cannot find in the book no longer has its next step suggest that it was edited in Tally when Tally's own answer to that post was not clean (for example a batch Tally created only partly, or an answer that did not parse): the line now says only that it cannot be matched, and to check in Tally before posting it again (#1108).
  • build_import_xml refuses a voucher whose own voucher_number begins or ends with white space, or is only white space, as voucher_number_invalid, before any file is written: on a raw import Tally read a number sent with a space at its end back without it, so such a voucher could never verify as posted (#1730). A space inside a number is still admitted.
  • parse_bank_statement no longer reads a statement PDF's text in render mode 3 (invisible), which is how a scanner app lays a hidden OCR layer under a page image, so that text is not parsed as if the bank had printed it. A scan whose only text is a hidden layer, or whose drawn text yields no statement row while most of the document's characters were hidden, is refused as statement_no_readable_text (before: its text was parsed, or the refusal was statement_no_account_number_line or statement_page_sequence_unproven). A character PDFium cannot classify refuses the file as statement_unreadable_pdf. No OCR is attempted. Neither PDFium failure was seen on the real files checked (#1637).
  • A Tally read that timed out, or whose response body failed after Tally had begun answering, is now sent once instead of up to three times: a client timeout does not stop the work Tally is doing, so a further send could queue behind the first. This covers the company-list reads, the ledger listing's extent read and the desktop snapshot's small reads. A read whose connection or send failed, or that Tally answered with a server error or a rate limit, is still retried (#1538, #1556).
  • The desktop core snapshot sends each of its group, ledger, voucher-type and voucher-window exports once: when the connection closes after the request was sent, it returns the error instead of sending the export up to three times. Its small reads (the pair of book-extent reads around each export, and the company list) keep the three-attempt retry, so one lost extent read does not make you start the run again (#1674).
  • The desktop core snapshot no longer sends the ledger period-balance report it used after each window to corroborate balances: one TallyPrime release could not evaluate it (others were not tried), and it was never seen returning a usable answer. A core run therefore cannot reach Verified; every window it reads records report_tie_out_unavailable, as one already did when the report was refused (Education mode). A resumed run no longer reopens a completed window, though a run resumed from an earlier build's saved state keeps a tie-out already stored. No published package sent this read, only the unreleased desktop app's Proof of Sync (#1643).
  • A desktop core snapshot that stops because the book changed while its groups, ledgers or vouchers were read is now recorded as source_changed_during_run ("Run again during a controlled quiet period"); one that stops because the two reads of a book extent differ, or because Tally gave no master-change marker, is recorded as company_book_extent_drifted or company_altmstid_missing, each with Gap Map guidance, where all showed "Response parse failed". A well-formed extent that names no row for the selected company is now company_identity_mismatch where it showed "Application response rejected". The two native-export drift conditions still use the catch-all (#1554).
  • The import journal's first record is now made durable in its folder: when the journal is new or empty, the data folder is synced after the write, and if that sync fails the record is taken back and the append refuses with import_ledger_unavailable (import_ledger_rollback_failed if the take-back also fails). A failed folder sync on the acknowledgement, the proofs, the verified-baseline additions and the doubt record is now reported instead of ignored: the acknowledgement refuses with import_file_write_failed, and publishing proofs refuses before the status record is appended. The masters-check verdict still ignores a failed folder sync, and the built import file is not folder-synced, as before. Before, a power loss soon after the first record could leave no journal and so no local record of a post. On Windows the folder sync does nothing, as before (#1548).
  • Desktop setup now refuses a selected company GUID that is not in the strict form of a Tally GUID, as company_selection_invalid (retry not recommended), and the snapshot identity check refuses a stored one as company_identity_invalid; before, the setup accepted any short trimmed ASCII text. Every company GUID in the committed captures (some of them synthetic) is in the strict form, so none is expected to be refused; a TallyPrime release that writes another spelling would be refused until it is observed (#1627).
  • The changed_since tool, which no build offered, and the code only it reached are deleted (#1769).
  • Locked dependencies moved, with no source change: rustls 0.23.45 to 0.23.46, which clears four advisories published on 14 October 2026 (RUSTSEC-2026-0346 to -0349: secrets in Debug output of client session values, a Brotli certificate decompression that can allocate up to 1 GiB before authentication, and a ServerHello downgrade of later connections to classical key exchange), and tokio, uuid, zeroize, libc and objc2 by a patch or minor release (tokio 1.53.2 fixes a Windows process-handle wait and a blocking-pool shutdown bug, per its release notes). The Rust licence inventory is regenerated (#1718, #1580).

Known limits

  • Live pass of what 0.6.0 writes into Tally (16 October 2026). The release candidate was installed as the extension in Claude Desktop (Microsoft Store build 2.37337.1.0) on Windows 10 Home 22H2 (build 19045), and run against TallyPrime 7.1 Silver on port 9001, on one new synthetic company, BRIDGE RC GATE LAB. Build: commit a3b7316947cdf3c5ce46615c8b89e45d538345f3 (pull request #1775, CI run 38877254488), bridge-tally.mcpb SHA-256 6b83ca74377d7ed86a2b524f4d53bd2e9753b3318bfa6634078d5f5645184022, containing bridge_mcp.exe SHA-256 e9780806594e60afabdb83d35bb751f0aeefba708b2b63c33b11452bbf0e5a84; the installed copy was checked against that hash before the run, and the server answered every call (Bridge's own log). With the one posting setting on, each of the four voucher types was posted once with no voucher number after its approval window, and one saved batch of three (a Payment, a Receipt and a Contra) was posted after one approval. In every post Tally's answer reported every counter, created 1 (3 for the batch) and nothing else counted; each voucher then read posted_verified in the post's own readback and again in verify_import, with no duplicates, and the company's voucher mark moved by exactly the number created. Tally gave the numbers itself: Journal 1, Receipt 2 (the book's seed receipt held 1), Contra 1, Payment 1, then in the batch Payment 2, Receipt 3, Contra 2. On Windows the approval window showed in front of a full-screen Tally and in front of Claude Desktop, with No as the default (one screen, one run). In that run a key meant for Tally (the Down arrow) moved the selection inside the approval window (No, Cancel, Yes) and Tally did not react: the window takes keyboard focus. No is still the default, so Enter alone refuses; arrow keys followed by Enter can reach Yes. A typed letter was not tried in that run (later that day, on the published file, a typed n declined the window and a typed y approved it: #1780) (#1148, #1589, #1630).
  • Not run in that pass: a batch of Journals; a batch larger than 3 or near 50; a post Tally refused or that landed partly; amendments; the review acknowledgement window (no voucher raised a doubt); a post that needs an On Account approval (a bill-wise party; the build refused one as designed); the retired batch variable (setting_retired); an import file taken to Tally's Import menu; Sales invoice posting (not offered in 0.6.0); the desktop app; Gold; macOS; any other Windows or Tally release. Also not run: the voucher-numbering rows (#724, #1668: a back-dated numberless Journal, a second numberless Payment, the refusal of a numberless voucher into a Manual-numbered type); voucher_read_cost against a real vouchers call (#1465); the Store-install proof of the lock (#1584, #1459: lock contention, local_data_report from the chat, which folder the desktop app uses); a second monitor. The candidate was installed while 0.5.1 was present, and its settings were not compared with 0.5.1's; what Claude Desktop passes for a new setting on an upgrade is not measured (#1090, #1630).
  • The file tested above is not the file that will be tagged. The pass ran on the CI package of the release pull request. The release workflow builds bridge_mcp again on its own, with a plain release build of that binary, where the CI package takes it from the Tauri build; the two files are the same source but not byte-identical, and no build is claimed to be reproducible. Between that commit and the commit the release is built from, no file that goes into the package or the desktop bundles changed (the check is recorded on the pull request that adds this paragraph).
  • Release package check (16 October 2026, 19:13 local time). A package built by the release workflow's own package steps (rehearsal run 38885882661, pull request #1777, source commit d6561887130d4fd4265121d6ef40db650b538280, which is a3b73169 plus one workflow file) was installed over the release candidate in Claude Desktop (Microsoft Store 2.37337.1.0) on the same Windows 10 computer, TallyPrime 7.1 Silver and company. bridge-tally-mcp-v0.6.0-windows-x64.mcpb SHA-256 3e76e8f6edf4d00fa8a1420b696661be9ddbe0b6ab3c84465f9cff58cc3fda05 (13,821,469 bytes), holding bridge_mcp.exe SHA-256 10e4614f74eab975897834ec1c9750edbe9fc7ad91584b802e5cde57a8180899, which is not the release candidate's file; its manifest is byte-identical to the candidate's. The settings were kept by that replace (port 9001, posting on); one run, same version, and it says nothing about an upgrade from 0.5.1. One Journal (12 August 2026, debit BRIDGE RC Expense, credit BRIDGE RC Suspense 5.00, no voucher number) was posted after its approval window: Tally answered created 1, every other counter 0 (538 bytes, SHA-256 42e4de026f72a1e7d882ad6f19067a8f9a6cf549dc1c40f90c752aefdfebfec5), the company's voucher mark moved from 8 to 9, the post read posted_verified in its own readback and again in verify_import, and Tally gave it number 2. The four voucher types and the batch ran on the release candidate, not on this file. The published file is built by the release workflow's own run, after these notes are written, so its bridge_mcp.exe is not compared with the one above here.
  • The headline of a batch post has not been seen on a live Tally post (#1090, #1532).
  • Nothing was run on a Mac against a live TallyPrime for this release; the Mac package is built and started by our checks without Tally.
  • What has been run against a real TallyPrime is listed in the README.

0.5.1, #

In plain words: ComplyEaze Bridge 0.5.1, since 0.5.0 (9 Oct 2026)

These changes are in ComplyEaze Bridge 0.5.1. Each line names the pull requests it comes from, except where it names an issue.

Should I upgrade?

  • If you updated from 0.4.2 to 0.5.0 and Claude Desktop did not start ComplyEaze Bridge: yes. 0.5.0 marked the Terms of Use setting as required. On one Mac, after an update from 0.4.2, 0.5.0 was not started and Claude Desktop showed "Unable to connect to extension server". 0.5.1 no longer marks the setting required. After installing 0.5.1, turn on the Terms setting in the extension's settings, then quit and reopen Claude Desktop. Until the setting is on, every tool answers terms_not_accepted and names the setting; nothing is read from Tally. What we saw: a package built from the Terms-setting change (#1436) on one Mac (only its version set to 0.5.1), installed over 0.4.2 and over 0.5.0, was started by Claude Desktop (in one run only after Claude Desktop was quit and reopened), and with the setting off the tools the assistant called gave that answer. On one Windows 10 computer with Claude Desktop 2.26454.2 from the Microsoft Store, the published 0.5.0 file over 0.4.2 was not started (no ComplyEaze Bridge tool in the chat and no error), and the build our checks made for this release (commit d744d25a, not the published file) was started after a fresh install, after an update from 0.4.2 and after an update from 0.5.0 that was switched on (as the update leaves it, and with the Terms ticked); with the Terms off the tools the assistant called gave that answer, and with them on the Tally check ran. Not tried: the published 0.5.1 file, and a release older than 0.4.2 (#1413).
  • If 0.5.0 already starts for you: the other changes below are the reason to update, if any.
  • If you use the bank-statement tool and build vouchers for mapped ledgers: such a voucher now carries the statement's party at the end of its narration text (see "What you can do now" below). Vouchers already imported keep their narration, but amending a batch built before this change rewrites the narrations of that batch's vouchers in the book (#1430).
  • After updating, if the assistant does not find ComplyEaze Bridge, check that it is switched on in Settings, Extensions. On one Windows computer, an update kept the extension's switch as it was: an extension switched off before the update was still off after it, and the assistant showed no ComplyEaze Bridge tools and no error. Switching it on brought the tools back there, and they then answered terms_not_accepted until the Terms setting was on (#1413).
  • If you use ComplyEaze Bridge on Windows: the build of this release made by our checks was tried on one Windows 10 computer, with TallyPrime 7.1 (Silver) and a company made for the test; see "Known limits" below for what was and was not tried.
  • How: ComplyEaze Bridge does not update itself. Follow the installation guide: (1) close any other program that runs ComplyEaze Bridge; (2) in Claude Desktop, Settings, Extensions, install the new file (on one Mac a new install opened no settings window: open Settings, Extensions, ComplyEaze Bridge yourself); (3) keep ComplyEaze Bridge's data folder; (4) check that the extension shows 0.5.1, turn on the Terms setting if it is not on, and check that "Allow voucher posting" is as you want it; (5) quit Claude Desktop completely and reopen it.

What you can do now

  • No new tool.
  • A voucher built from a bank statement for a mapped ledger now carries | Statement party: <name>, the counterparty the statement printed, at the end of its narration text (#1430). A file you import by hand adds ComplyEaze Bridge's [BRIDGE:...] tag after it. A line sent to suspense and a cash line are written as before. A party that contains | or a control character, that carries the reserved marker, or that makes the narration longer than 2,000 characters now refuses the statement run (party_not_admissible); skip that party or map it to the suspense ledger. A label the parser gives a line by its wording (bank charges, a card fee, an instalment) is not written as a party. Vouchers already imported keep their narration, but amending a batch built before this change rewrites the narrations of that batch's vouchers in the book. The name is part of the narration, so mask_parties does not shorten it and drop_narration removes it.

Safer or fixed

  • An update now starts ComplyEaze Bridge before the Terms are turned on, and its tools say why they refuse. The Terms of Use setting is no longer marked required and still starts off. Claude Desktop now starts ComplyEaze Bridge, and until you turn the setting on every tool answers terms_not_accepted, naming the setting. That the app starts and the tools the assistant called answered was seen on one Mac with a package built from the Terms-setting change (#1436), and on one Windows computer with the build of this release (#1413). The setting no longer shows "(required)" beside its title, and its description now begins "Tick to use ComplyEaze Bridge: it refuses every tool call until you accept the Terms of Use." A test checks that wording, and it was seen in Claude Desktop on that Windows computer.
  • The schema of every paged read no longer names a default limit of 500. When limit is left out the server uses its configured row limit (500 unless it is configured otherwise), as the tools' own text already said; the schema said 500 even when the row limit was lower or higher (#1148).
  • A State Bank of India statement line whose UPI, NEFT, RTGS or transfer text prints an empty name field now names its party as UNNAMED, as the IMPS reader already did. Before, the party was empty and the suspense line's narration named no one. UNNAMED cannot be mapped to a ledger, so the line still goes to the suspense ledger (#1429).
  • The import journal (agent-import-ledger.jsonl) is now read with each of its dates as a checked date. A record holding a date that is not YYYYMMDD refuses the whole journal (import_ledger_invalid): building, posting, verifying and reconciling every batch, and the desktop journal screen, stop until the record is put right, and the local data report shows the journal as not read. The current code writes only that form, so only a record edited by hand is expected to hold such a date, and the stored bytes do not change (#1307).
  • The PDF library is loaded from the folder that holds the program (#1422).

Known limits

  • Windows was tried on one computer (Windows 10, Claude Desktop 2.26454.2 from the Microsoft Store, TallyPrime 7.1 Silver, a company made for the test), with the build our checks made for this release at commit d744d25a. The published 0.5.1 file was not tried, and nothing was tried on other Windows versions or other Tally editions (#1413).
  • On that computer, a bank-statement import file built for six vouchers was imported by hand into Tally and read back: the verification found all six, and a voucher posted to a mapped ledger showed its narration with | Statement party: <name> before the [BRIDGE:...] tag. An unchanged re-import of the same file was matched to the six vouchers already there and added none. A changed re-import was not tried (#1430).
  • On one Windows computer where Claude Desktop was installed from the Microsoft Store, the folder shown for a built import file did not exist: the file was under the user's AppData\Local\Packages\Claude_<id>\LocalCache\Local\Bridge\agent\imports folder. If Tally's Import screen does not find the file at the path shown, look there. Seen with 0.5.0 as well (#1459).
  • The library that reads Tally's replies was updated (#1198). Its output for the saved replies in our test record is unchanged. The Windows runs above used it for a company list, the read-back of six vouchers and an August trial balance on a small company; it was not run against a large book.

0.5.0, #

In plain words: ComplyEaze Bridge 0.5.0, since 0.4.2 (3 Oct 2026)

These changes are in ComplyEaze Bridge 0.5.0. Each line names the pull requests it comes from, except where it names an issue. A "mark" below is a counter Tally keeps that moves when vouchers or ledgers change.

Should I upgrade?

  • After updating, accept the Terms again, and check the posting setting. 0.5.0 names the setting that accepts the Terms for version 2026-10.1, a different setting from the one you accepted for 0.4.2. After updating, tick the new Terms setting in the extension's settings, check that "Allow voucher posting" is as you want it, then quit and reopen Claude Desktop. What we saw: we installed the package our checks built for this release (not the published file) over 0.4.2 on one Mac. It replaced 0.4.2 and kept the Tally port, Response redaction and Allow voucher posting as they were (a posting setting that was on stayed on); the old Terms setting was gone and the new one had no value. Before the tick, and before any quit, Claude Desktop did not start ComplyEaze Bridge, and a request to one of its tools failed with a 400 error that did not mention the Terms. After the tick was saved Claude Desktop started it at once and a new chat got an answer from it, but the assistant first called a leftover second entry for the same extension, which failed with the same error. After Claude Desktop was quit and reopened there was one connector and no error. Not tried: quitting and reopening without the tick, installing over a release older than 0.4.2 (remove such a release first), and Windows. If Settings, Extensions lists two ComplyEaze Bridge entries (as happened when 0.4.2 was installed beside an older release), remove the older entry before you use either. Then enter your settings in the new one, tick the Terms setting, and quit and reopen Claude Desktop (#1010).
  • If you turn posting on: do not post a batch built by 0.5.0 with an older version. Versions 0.3.0 to 0.4.2 refuse to post a batch built with 0.5.0, and nothing is posted. Their message says to build the batch again; do not do that there: reinstall 0.5.0 (or a newer version) and post the batch from it. An older version makes none of 0.5.0's checks on bill-wise and cash-in-hand ledgers for what it builds and posts itself. As before, from the first post attempted with 0.4.2 or later, versions before 0.4.2 stop reading the import history. 0.2.0 and earlier have no such refusal: do not run them over a data folder 0.5.0 has built in (#1234, #1343).
  • A batch built by an earlier version must be built again once before it can be posted (the bill-wise and cash-in-hand records); check first whether its file was already imported by hand (#815, #1234).
  • If you only read your books: nothing here is urgent. The reasons to update are the search and summaries in vouchers, the small-company count, the new cash_flow tool and the safer refusals below.
  • If you use ComplyEaze Bridge on Windows, treat it as untried: no one on our side has installed the Windows package of this build in Claude Desktop.
  • How: ComplyEaze Bridge does not update itself. Follow the installation guide: (1) close any other program that runs ComplyEaze Bridge; (2) in Claude Desktop, Settings, Extensions, install the new file (remove a release older than 0.4.2 first, and see the first bullet above if two entries are listed); (3) keep ComplyEaze Bridge's data folder; (4) tick the new Terms setting, check that the extension shows 0.5.0 and that "Allow voucher posting" is as you want it, and enter your other settings again only if they were cleared; (5) quit Claude Desktop completely and reopen it.

What you can do now

  • masters can list a company's cost centres (cost_centres, with each centre's category) and cost categories (cost_categories, with their three allocation flags). They are returned whether or not the company's Cost Centres setting is on: a book whose setting read No still returned its two centres, so a No setting is not "no centres", and an empty list does not say whether the feature is off or none is defined. A centre without a category, and an answer whose collection does not carry the type of the kind asked for, are refused rather than read as "none". An empty cost-category list is refused (masters_cost_categories_empty) on the expectation that the predefined Primary Cost Category always exists, which was seen in both books whose categories were captured and was not captured for a book with no centre defined. ComplyEaze Bridge does not return how a voucher was allocated to a cost centre. Measured on three synthetic books for cost centres and two for cost categories (the setting reading No on two, Yes on one: that book, captured in a separate sitting, returned its three centres, one under another, and the default category, in the same shape); a larger book is refused when its master mark is too large for the response budget (1,037 for cost centres, 1,264 for cost categories). Not measured: centres nested more than one level deep, an alias beyond a centre's own name, a larger book (#1398).
  • vouchers and the sales and purchase registers now return reference_date, a voucher's REFERENCEDATE (YYYYMMDD), where the voucher has one (#1257). Checked once against a live TallyPrime 7.1 on one synthetic book: the element came back on every voucher, empty on all but one and filled on that one (a Purchase) and on a voucher keyed for the check; it was not checked on a sale or on another voucher type, and a value that is not a date refuses the read.
  • vouchers and the outstandings party detail now reach a ledger by either spelling of its name when Tally's own name for it differs in case or symbols from the spelling its vouchers carry (26 of 4,017 ledgers in a separate census of 13 books, not reproducible from this repository), and ledger_match shows Tally's own name, with ledger_row_spelling when the vouchers spell it differently. A spelling that is two ledgers' is refused as ledger_ambiguous. The voucher filter and the trail still use the spelling the vouchers carry; ledger_movement and imports are unchanged. Not measured: that Tally accepts the voucher-row spelling when it differs from the ledger's own name, what a voucher row carries for such a ledger on a book other than the one measured, and what the outstandings report that the trail reads carries for it, so a party detail that finds no bill or no unallocated row for such a ledger says report_spelling: not_established; that the first name list is the primary language (#1085).
  • ledger_movement now says what a slow read cost, as vouchers does: when its two reads of the voucher window together took 20 seconds or more, or the census was 16 reads or more, the result carries read_cost with both reads' observed seconds, a floor derived from the gaps between census reads (derived, not measured), and whether those two voucher reads together stayed inside the one measured host limit, which leaves out the two ledger catalogue reads, the company check and the wider read an empty window gets (#1239). It makes no estimate for another window, adds no request and refuses nothing. Checked once live on a synthetic book whose voucher mark is 29,900: a one-day window (82 vouchers) took 24.6 seconds and carried no block; August (2,542 vouchers) took 80.9 seconds and carried a block of 65 seconds over the two voucher reads, 4 census reads and a floor of 1 second, its voucher and census counts equal to the relay's own log. Not measured: the sixteen-census-read case (a much larger book) and a call that is slow for reasons the block leaves out.
  • cash_flow reads Tally's own Cash Flow for whole months (the month-wise movement of the cash and bank ledgers, not a cash flow statement under AS 3). The months are returned only after their net total has been compared with the trial balance's cash and bank ledgers and found equal; the split into months is Tally's own and is not checked, and Tally's debit and credit columns are not returned. A ledger under Bank OD A/c or Bank OCC A/c with movement refuses the result: Tally was seen counting one such ledger, which the check does not yet count, and the refusal is covered by tests and has not been seen against Tally. The tool was run against Tally on two synthetic books: the net total tied on five windows (three with only debits, two with a credit amount, one of them with a positive closing as well) and a quiet window was answered as nothing to compare. Tally's own answers were captured on four synthetic books in all, one of them a year of empty months on a large book. A Bank OD credit or a Bank OCC ledger, a contra, a month with entries that cancel, a window ending in February, a window crossing a financial year, optional or post-dated vouchers, a several-currency book and a large book with cash activity are not measured (#1232).
  • vouchers can now find a voucher by its number, reference, a phrase of its narration or an amount (voucher_number, reference, narration_contains, amount), and can add a window up by ledger, month, voucher type, group or primary group (summarise_by), with debit, credit and voucher counts per bucket and the cancelled, optional and entry-less vouchers counted apart. Search and the ledger, month and voucher-type summaries run on the rows the window read already holds and send no new request to Tally; the group and primary-group summaries also read the ledger list and the group list, before and after the window. They refuse when any ledger's group chain, or the set of ledgers, changed meanwhile, and when a ledger the window touches cannot be placed, so no entry is left out of a group total, naming the ledger it could not place. A book whose master-alteration mark is above 11,983 (a provisional limit, computed and not measured, about half of what the transport admits for the same list) is refused before the ledger list is read. A group summary also gives each group's whole figure, descendants included, as subtree_totals, and every row says what it covers. A zero from a counted window is a checked zero. A narration phrase is refused where narrations are withheld from the assistant. A summary sums post-dated vouchers (counted, with the vouchers Tally sent no flag for counted apart) and any non-posting voucher type a book uses, and says so in the result. Checked once against a live TallyPrime 7.1 on a synthetic book of 67 vouchers: every bucket equalled the sums over the listed vouchers, the ledger buckets equalled trial_balance for the same year, and each search returned what the same criterion selects from the listing; it did not cover a large book, a memorandum, a reversing journal or a voucher withheld for a foreign-currency amount. That check covered search and the ledger, month and voucher-type summaries. A second check on 7 October 2026 on the same book read the whole year once as group and once as primary_group: 12 and 8 buckets over the 64 vouchers that count (a cancelled, an optional and an entry-less voucher left out); every group bucket equalled the ledger buckets added up by the trial balance's own parent column, the 18 subtree_totals equalled the trial balance rolled up the group tree, and each call took 64 requests (the window's 34 and the group modes' 30, as counted from the code) and about 11 seconds. It did not cover a large book, a group renamed or moved while a window is read, or a held later page (#1230).

Safer or fixed

  • purchase_register and sales_register now take their state from the rule vouchers uses. A non-empty window is complete only when every voucher read was checked against a separate count of the window. Before, the registers called such a window complete on the company marks and the ledger masters alone, so the same window had two answers. A row's status is unchanged. Not measured live: a window admitted against a census through the registers on a large book (#1031).
  • vouchers, voucher_presence, purchase_register and sales_register now count a small book's window too (a company with a few dozen vouchers or fewer), where they used to read it without a count. A nonempty window of such a book can now be complete, and voucher_presence can say absent for it, which it could not before. The cost is one more read of the book's vouchers: six more requests to Tally on every call that counts (a vouchers call on a small book goes from 16 to 22 requests, voucher_presence from 28 to 34); the added time was measured on windows of one voucher, on debug builds, at 0.5 to 2 seconds a call (one run of the day and week calls, three of the presence call), and not on a larger book. absent still needs a window proven complete, nothing resembling the proposal, the voucher number of a manual type, and a party that was supplied and compared completely: a proposal with no party (a Journal), no manual number, or a party whose comparison was cut short reads possibly_present with its reason, however complete the window. A voucher created, altered or deleted between the count and the read now refuses a small book as voucher_window_part_not_admitted (cause part_census_mismatch), as a large one always did. The Ledger Entries screen reads through vouchers' window read, so it counts too and meets the same refusal. An empty window of a small book is answered as before, by the check of an empty window: the count adds its requests there and changes nothing else. ledger_movement, verify_import, outstandings and that check are unchanged. Not measured against a live book: what Tally does when it cannot serve the count, which refuses the call as voucher_window_volume_unestimated where the small book used to answer partial (it served the count every time in the run above), a larger book, and the registers and the Ledger Entries screen, which use the same count but were not called live (#1029).
  • profit_and_loss now compares Tally's Cost of Sales : heading with the derived cost of sales (Purchase Accounts plus Direct Expenses) even when the heading reads zero or empty. Over a non-zero cost of sales such a heading refuses gross and net as tally_profit_and_loss_differs, naming the heading; over a zero one it ties. Before, a zero or empty heading passed whatever the cost of sales was. Both committed Profit and Loss captures still tie; whether a real Tally ever prints the heading zero or empty over a non-zero cost of sales is not measured (#1070).
  • post_import and verify_import proofs now measure alter_id_delta from the company's voucher mark read just before the POST, for a native post, instead of from the mark recorded when the batch was built. Anything posted between the build and the POST no longer counts as this post's, so the delta agrees with what Tally created. alter_id_delta gains from (pre_post_mark, or build_mark for a file imported by hand, which has no POST of its own); pre_import_mark still reports the build-time mark (#1087).
  • build_import_xml now refuses a batch that names a ledger which keeps bills in Tally, as bill_wise_party_unapproved, until the person approves each such party (the build is repeated with on_account_approvals): an entry on it carries no bill allocation, so Tally lands it On Account and the person must match it to a bill by hand. Before, the build only printed a warning. post_import and the queue admission read the flag again, from the ledger list they already read, and refuse import_bill_wise_changed for a ledger that became bill-wise since the build. A batch built before this change is refused for posting as import_batch_predates_bill_wise_record and is rebuilt (check first whether its file was already imported by hand). The approval is the assistant's word, not proof that a person said yes, and a hand import of the file is not checked. The approval dialog of a native post shows the person which ledgers these are: "On Account" stands before the name of each approved ledger, on each of its entries for one voucher and on its totals line for a batch, with one sentence saying the ledger was bill-wise when the batch was built and its entries carry no bill allocation; the desktop Journal review shows the mark and the sentence too. A batch's per-voucher lines carry no mark, and their heading then says it is marked below; in such a batch a narration that holds the sentence's opening is withheld on its per-voucher line. A dialog that cannot show the mark refuses. The mark adds no line, only characters, so a batch already at the dialog's 100-character line limit or its character limit can now be refused as too large (import_review_too_large). The dialog for one voucher ends with the voucher's own text: its reference and narration are the last two lines, quoted as they will be posted, under a line saying that no line below it is an entry, a total or an instruction; every line the dialog writes itself stands above that line, and the date line says where the two are. That line and the note on the date line count toward the dialog's 100-character line limit and its 1,600-character limit. Looked at once on each system, with a one-voucher text at all three limits (24 lines, 1,600 characters, a 100-character narration line): on one Mac and on one Windows PC the whole text, its last two lines included, and the buttons were visible without scrolling. Not measured: a smaller screen or a larger display scaling, a batch dialog on either system, a large book, and the bills of a ledger whose flag reads No (#1234).
  • Versions 0.3.0 to 0.4.2 refuse to post a batch this version builds. They do not read the cash-in-hand and bill-wise records a saved batch now carries, so after a rollback they could have posted it with neither check; the batch's ledger binding is now saved under a name they do not read, and they refuse a batch with no binding before any approval window or request to Tally (import_batch_predates_ledger_binding, nothing posted). Their message for that refusal says "Build the batch again"; for a batch this version built, reinstall this version (or a newer one) and post the batch from it instead, because a batch built again on an older version has neither check. A batch an earlier version saved is read as before. Which versions can read the import history is unchanged: from the first post attempted with 0.4.2 or later, only 0.4.2 and later read it. 0.2.0 and earlier have no such refusal: do not run them over a data folder this version has built in (#1234).
  • The posting setting is shorter and names its four known limits in plainer words, and Terms of Use section 9.2 (version 2026-10.1, effective 7 October 2026) now lists the same limits and more: a company or ledger renamed or opened at the moment of a post, another Tally connector that can change entries without the approval window, and that dealing with a voucher that reaches the wrong company remains yours. The setting that accepts the Terms is now named for version 2026-10.1, so the new setting has to be accepted after an update (see "Should I upgrade?"); version 2026-10 stays published for the builds that asked for it (#1010).
  • When a ledger name you gave is not in the book, the candidates now list ahead of the others the ledgers that hold every word you typed (rule: shared_every_distinctive_token), then those that share only some; a ledger matched by a stronger rule still comes first. The same ledgers are found and counted as before and none is chosen for you; when more than 25 are found, the ones left off the list are the ledgers holding only some of the words before those holding all of them (#1076).
  • Five lists of a batch's ledgers in refusals now come in the batch's own order, not in name order. masters of masters_not_exact, ledger_twins and ledgers_changed list a ledger where the batch first names it; refused_ledgers of cash_bank_ledger_not_established lists a ledger where one of its entries is first refused, and refused_ledgers of the cash-answer refusals (in build_import_xml and post_import) where an answer first names it. The order is the same whether or not party names are masked. The same ledgers are found and refused as before. Where a list is cut (ledgers_changed names eight; a small response cap leaves rows of refused_ledgers out), which ledgers are named, and how many refused_ledgers_omitted counts, can differ from before. The ledgers a post_import or verify_import answer names in masters_after_post.ledgers come in the same order, where the batch first names each, in the answer and in the proof saved from it (the desktop screen names the first eight of them); the recorded verdict, and the review dialog that reads it, keep the order they were recorded in. Other lists are unchanged (#1234).
  • The approved-party list in the build result is new in this release and follows the masking setting, as the refusal's list does. Under mask_parties the party lists build_import_xml returns come in the order of each party's digest, not of the names, so the order does not follow the alphabet of the real names; the order can change from one call to the next (#1234, #1338, #1345).
  • What redaction does and does not hide is now stated in the extension. Its description, its Response redaction setting and the descriptions of the two log tools say that redaction only shortens party and ledger names and bank account numbers (or drops narrations), and does not hide amounts, company names, PAN, GSTIN, IFSC, MSME or Udyam numbers, contact details or references, nor a name written inside a narration or reference. This replaces the earlier wording "can only mask party names or drop narration". Not measured: the shortening of bank account numbers rests on code, not on a live capture (#1046, #1364).
  • post_import refuses a business-cash entry whose cash ledger was moved out of Cash-in-Hand after the batch was built, before any approval and again when the post is queued (cash_ledger_not_cash_in_hand); a batch built before this release is built again once (import_batch_predates_cash_ledger_record). Tested on captured data only; not measured: a live regroup between build and post (#815, #1228, #1229).
  • build_import_xml refuses a bank statement built onto a Cash-in-Hand ledger (statement_bank_ledger_not_a_bank), and warns, without refusing, about a suspense ledger that is outside Suspense A/c, whose group is not established, or that is not in the book. No live run; the read-only check of the bank ledger in Tally against the statement is not in this build (#1211, #1214).
  • An HDFC statement that 0.4.2 refused with unparseable_date at the last row of a page is read with the statement-period line skipped, so its dates are no longer mixed into the previous row. No real statement has been run through a package with this fix; one whose period line does not match still ends in the old refusal (#1310, #1325).
  • When Tally rejects a post entirely, the post's summary now says so and no longer suggests posting again. Not measured: a live post that Tally rejects whole (#1108, #1213).
  • A busy Tally port gets four answers. Busy before post_import sends anything: the refusal tells the assistant to call post_import again with the same batch after 5 seconds, that you will be asked to approve again, and not to rebuild it. Busy after the send: it tells the assistant to use verify_import with the original batch and never to post again. Whether the post was sent could not be observed: the refusal says to use verify_import with the original batch. The second pass of an approved post has no time left: it refuses at once as tally_endpoint_busy with nothing sent (#869, #893, #1219, #1244).
  • The approval dialog for a batch of up to ten vouchers lists each voucher on one line (type, date, amount, ledger, narration cut at 40 characters); a larger batch, one with an unsafe narration or reference, or one whose lines do not fit the dialog shows the totals with a line saying why. A narration that reads like a line of the dialog itself is not shown. A single-voucher dialog shows the narration without its outer spaces, as posted. The per-voucher lines were seen on macOS only (#1055, #1063, #1223, #1226, #1278, #1283).
  • ledger_movement no longer refuses a read because of a zero entry written as -0.00 and flagged with Tally's debit-side sign flag. We have not seen Tally send one (#1251, #1265).
  • ledger_masters: a refused native ledger export now names a fixed cause (native_collection_not_success, native_collection_status_absent, native_collection_malformed_response, native_collection_identity_mismatch, native_collection_bounds_exceeded, native_collection_row_unusable); a company ledger-count answer past the response cap refuses as ledger_count_company_response_too_large; the catalogue read is sent once and a timeout refuses as ledger_export_invalid with request_deadline_exceeded (#718, #937, #1033, #1144, #1167, #1184, #1216).
  • A refused date carries one next step: give each date as YYYYMMDD or YYYY-MM-DD, a real day, with from no later than to (#1268, #1273).
  • verify_import lists, on a voucher cancelled in Tally, any effective unmarked copy entered again by hand (effective_copies_observed), as a report only that does not attribute it. Proven on Journal vouchers only (#806, #1221).
  • The tool descriptions the assistant reads are clearer: every input is described, six easily confused pairs end with "use X instead", and the two log tools say when to use each and that reading changes nothing but its own receipt line. The ledger_masters description is about 4 percent shorter than in 0.4.2 (6,714 to 6,448 characters in the source) without dropping a limit or safety phrase, and the vouchers description says a bill allocation carries its bill date and credit period when Tally sends them. No claim is made that assistants choose better (#945, #1155, #1157, #1161, #1164, #1166, #1176, #1180, #1191, #1260).
  • The extension's listing text (a long description, keywords and links) is added to the manifest; it says the extension is untested outside Claude Desktop and that ComplyEaze Bridge is not affiliated with Tally Solutions (#1375).
  • ComplyEaze Bridge is named in full in the text of the desktop app's screens and messages that these changes reached. The app's product name and window title in its configuration still read "Bridge", and some error strings in its source still begin with it (#962, #1163, #1165, #1168, #1170, #1179, #1190).

Known limits

  • Search and summaries were checked on one synthetic book of 67 vouchers; large books, masking and foreign-currency vouchers are not checked, and a group summary is refused above a provisional limit that was not measured (#1230).
  • On a large book a voucher read is practical a day at a time: one day took about 170 seconds on the largest book measured, and a month is read day by day. What a read cost is stated after it, not before (#485, #595, #703, #1239).
  • The either-spelling ledger lookup was checked live on one book and one ledger; a second check on a book where a ledger's spelling differs from its own name is pending (#1085).
  • Bill-wise approval is the assistant's word, not proof that a person said yes; batch posting is not offered (#1090, #1234).
  • A bill's trail and an unallocated amount's detail read the whole date window; a party-scoped request is not built (#945, #1356). The slice L captures of #1257 (a duty-head change or delete moving the master change counter; an item purchase on the class-entry shape) have not been run.
  • Whether a company has cost centres or GST switched on, how a voucher is allocated to a cost centre, and bank-reconciliation fields are not read (trial_balance and cash_flow do return the company's currency, and stock_summary the inventory settings) (#1231). Tally's Funds Flow, Ratio Analysis, Negative Stock and Negative Ledgers reports are not read (#1358).
  • GST Sales and Purchase posting are not in this build (#628).
  • No Windows install of this build has been tried in Claude Desktop; the per-voucher approval lines were seen on macOS only.
  • Posting limits: Terms of Use section 9.2 lists the four known limits.
  • What Claude's AI provider receives is stated in the README and the security page; the list there is open, and redaction shortens names rather than hiding them (#1350, #1354, #1361, #1368).
  • A bank statement is not checked against the bank ledger in Tally (#1211); the HDFC fix has not been run on a real statement.

0.4.2, #

In plain words: ComplyEaze Bridge 0.4.2, since 0.4.1 (2 Oct 2026)

These changes are in ComplyEaze Bridge 0.4.2. Each line names the pull requests it comes from, except where it names an issue. A "mark" below is a counter Tally keeps that moves when vouchers or ledgers change.

Should I upgrade?

  • Remove the old extension first. 0.4.2 installs beside 0.4.1 or earlier instead of replacing it, because its author line changed from "Bridge contributors" to "ComplyEaze contributors" and Claude Desktop builds an extension's identity partly from its author name (seen with 0.4.1 on a Mac; earlier releases carry the same author line, so the same is expected; we did not try Windows). Before installing 0.4.2, in Claude Desktop open Settings, Extensions and remove the older ComplyEaze Bridge (release 0.2.0 shows as "Bridge Tally"). Keep ComplyEaze Bridge's data folder: both versions use it, so do not delete it. After installing, enter your settings again: the Tally port, the posting setting (posting starts off), the Terms setting, which every tool needs, and Response redaction, which starts at none (set it again if you had shortened or masked names). If you already have two, remove the older one: once 0.4.2 has tried a post, the older one can no longer prepare, post or check vouchers (see the posting points below).
  • The number is a patch number on a larger change. 0.4.2 adds a tool, asks you about ledger names it used to read, and once you post with it you cannot go back to 0.4.1.
  • If you use 0.3.0 or 0.4.0 on Windows: yes. 0.4.2 contains everything in 0.4.1, including the fix for the network-path forms of the bank-statement path issue published as advisory GHSA-vm5g-r3p7-wxx7. Some paths still pass; the advisory lists which. If you cannot upgrade yet, follow the advisory's steps.
  • On a Mac with 0.3.0 or 0.4.0: recommended, though not urgent. The advisory rates the issue low there; some paths still pass, and the advisory lists which.
  • If you use 0.4.1: nothing below is urgent. The reasons to upgrade are that a ledger name that differs by a symbol, an accent or run-together words now makes the assistant ask you instead of being read, later pages of a long voucher list can be served faster, and a post that Tally refuses is now named plainly.
  • If you used 0.4.1 or earlier: check answers for a ledger named loosely. If you asked about a ledger by a name that was not its exact spelling, the figures may be for a different ledger with the same letters and digits, and an earlier vouchers answer did not show which ledger was matched. Ask again with 0.4.2, which asks you instead (#1092).
  • If you turn posting on, the way a posted voucher is matched has changed. A post no longer adds a ComplyEaze Bridge tag to the narration, and each voucher is matched by its place in the run of changes Tally records for that post (#1054). From the first post 0.4.2 sends to Tally (its record is written just before sending, so this includes a post Tally refuses or never answers), do not run 0.4.1 or earlier, of the extension or of the desktop app, on this computer: the record of what was sent (the import journal) then holds fields they refuse to read, and they stop preparing, posting and checking vouchers. Before that first post, going back is harmless. The one-voucher post that the extension makes was not run live with this matching (see Known limits).
  • If you turn posting on: `voucher_presence` cannot recognise these vouchers. It cannot identify a voucher this version posted (it matches a ComplyEaze Bridge voucher by the tag only for files you imported by hand), so a voucher posted with 0.4.2 that someone then edited or re-dated in Tally can read absent there. Check what ComplyEaze Bridge posted with verify_import, never with voucher_presence, before entering any of it again by hand. In Tally's own screens these vouchers can no longer be told from hand-entered ones by their narration. Posting is off in a new install, and 0.4.2 installs as one: check the setting after you install it.
  • How: ComplyEaze Bridge does not update itself. Follow the installation guide: (1) close any other program that runs ComplyEaze Bridge; (2) in Claude Desktop, remove the older ComplyEaze Bridge in Settings, Extensions, then install the new file from the same screen; (3) keep ComplyEaze Bridge's data folder, which holds its record of what it has sent to Tally (from 0.4.2 it is the only record of which vouchers it posted); (4) enter your settings again, including Response redaction (it starts at none), and check that the extension shows 0.4.2 and that "Allow voucher posting" is as you want it; (5) quit Claude Desktop completely and reopen it.
  • What was tried is under "Known limits" below. No one on our side installed the Windows package of this build in Claude Desktop on a Windows PC.

What you can do now

  • A new tool, `sales_register`. It reads a register of the tax in the books for sales: the Sales and Credit Note vouchers of a date window that touch a Duties & Taxes ledger, with each entry's tax taken only from the GST duty head on that ledger's master, never from a name or an amount. It reads as purchase_register does, and it is a register of the books, not a GST return. A Credit Note keeps Tally's signs (nothing is netted or flipped, so add signed amounts), and may be a sales return or a credit note to a supplier: each row says the party's group and the tool does not choose. A Debit Note is listed apart, without an amount. The state-side tax head has two recognised forms, state_tax and sgst_utgst, for the same side. A row is marked not_measured_live only for kinds the row itself shows (for example inter-state, cancelled, optional or post-dated); an unmarked row is not thereby measured, because a sale typed on Tally's screen, a tax Tally computed, a duty head such as cess, an invoice with several goods lines and a currency other than the book's base cannot be marked. It needs a book whose base currency is INR and refuses a book too large to list. It reads the whole date window once and the ledger masters twice (before and after), and all of it again for every page. One call sent 96 requests to Tally on one small book and 118 on another. Use a narrow date range: its cost on a large book was not measured. It was run against a live Tally, once each, on synthetic companies: one taxed Sales item invoice, one untaxed one (read once by an earlier build; its masters and the tool's answer were not kept) and one Credit Note. What was not shown is under "Known limits" (#1009).
  • Later pages of a long voucher list can be served from the first page's read. vouchers no longer reads the whole window again for every page when the window was read complete, which means counted first. On one synthetic book, one month of 2,542 vouchers read in a release build, the first page took about 66 to 68 seconds and a later page about 1 second (one run each). A later page that names the first page's snapshot_id is refused if the company's marks moved or the held read is gone, instead of continuing from a different read. A read is held for ten minutes and is also dropped after a write through ComplyEaze Bridge, when the size cap that all held windows share (64 MiB) evicts it, or when a newer read of the same question replaces it. A small or new company is not counted, so its window reads partial and every page reads it again. A later page that does not name the snapshot_id reads afresh when the books moved, and says so (earlier_snapshot: its offsets do not continue the earlier pages, so start again from the first page); when the hold ran out it reads afresh without that flag, and only a page whose snapshot.reused is true continues the earlier ones. A date is the same question as 2026-08-01 or 20260801, while the ledger argument must be repeated exactly as typed (#1053, #1118).
  • A refused ledger name now tells the assistant to ask you which ledger you mean. When ledger_movement, vouchers (with a ledger) or the party detail of outstandings cannot find the name you gave, the refusal can now list the ledgers it may mean (not when party names are masked), and the assistant is told to ask you, even when there is one candidate, and never to choose. A name is now read without asking only when it matches a ledger's spelling exactly, or differs only in upper or lower case English letters or in spaces. Before, a name that differed by a dropped symbol, an accent or run-together words was read if one ledger had the same letters and digits; in a lab list of synthetic ledger names that read the truncated name Input Cess ( as a different ledger, Input Cess. Each answer for a named ledger now says how the name matched. A name containing a no-break, figure, narrow no-break or ideographic space asks, as a test now pins (#1057, #1092, #1127; issue #1095).
  • The assistant is told how to pick a company, and to say which one it used. list_companies now begins "Start here" and states the rule (use the open company only if it is the only one open and you named no client, or if exactly one open company matches the client you named; otherwise ask, and never guess), and Claude Desktop is sent the same rule at start-up, with the instruction to name the company, dates and ledger in the first line of an answer, and what to do after a refusal. The start-up text follows your posting setting: with posting off or import only, it says this connection cannot post to Tally. It also tells the assistant to ask before reading vouchers over more than a month unless you gave the period, and before an outstandings party detail (#1058, #1113).
  • A plain opening line on three reads. trial_balance, profit_and_loss and balance_sheet now begin with a headline when they answer: the company, the exact period and, in words, whether the figures were read for every ledger, partial, or established or not. A refusal has no headline yet. The codes and figures stay in result unchanged (#1062, #1066).
  • `outstandings` says what it counted and which date it used. It returns open_bills_total (every open bill in that direction, counted before paging) and open_bills_shown, so a cut page no longer reads like the whole list (on a partial read these cover base-currency ledgers only). In every answer that reads (a refusal carries none) it returns the as_of date it used, which is today's date on this computer when you gave none; tally_status returns that date as today (#1049, #1079).
  • Pages that are not in the package: a Questions page on the install site with a table of what it can and cannot do, a before-you-start list and plain answers, including how far our liability goes and how to reach us (#1061, #1084); a "What it costs" section in the README and a line on the Download page: ComplyEaze has set no price and sells no licence today (#1059); the Privacy Policy (version 2026-10.1, effective 3 October 2026) and every footer now name the publisher with its registration number and registered office, the policy adds a purpose for publishing written quotes you consent to, and says a grievance is answered within one month (it said 30 days) (#1104). The site also gained canonical addresses, a sitemap and an llms.txt (#1047).

Safer or fixed

  • A withdrawn approval is named. If an approval is withdrawn after the post has taken it but before anything is recorded or sent, post_import now answers import_approval_revoked; before, it answered import_dispatch_outcome_unknown although nothing had been recorded or sent. The message ("No posting attempt was recorded. Review the error before requesting approval again.") and the attempt_recorded: false field are unchanged, and no next step is attached. It was tested on a scripted Tally, not a live one (#1131; issue #791).
  • A post that Tally refuses is named plainly. If Tally's answer to a post counted nothing created (CREATED 0, an exception for each voucher sent and every other counter reported as 0) and the voucher is not found, post_import now says Tally reported it as not created, to check it is not in Tally, and to enter that one voucher in Tally's voucher entry screen, never through Tally's Import menu; in the answer to that post itself the voucher's status reads tally_reported_not_created. This covers a single voucher, which is what the extension posts, and a saved batch of two or more vouchers rejected whole when the company's mark was read before and after and did not move (batch posting is an environment setting, BRIDGE_AGENT_ENABLE_BATCH_POST, that the extension's settings do not offer). Tally's answer was captured live only for missing ledgers, on TallyPrime 7.1 Silver (one Education answer to a bad date was also used, for its counter shape only), and the reading of it was tested on those saved answers, not in a live post; the code applies it to any answer of that shape, including on Gold, where it was not measured. After you enter the voucher by hand as told, the saved batch stays "reconciliation required" and ComplyEaze Bridge cannot close it (#1039 is open; #1116, #1126).
  • Where a posted voucher cannot be matched, the status changed. In 0.4.1 an unfound voucher of a post read not_found. In 0.4.2 a voucher of a post made with 0.4.2 that cannot be matched reads sent_not_attributed ("check in Tally before posting it again", for example after an edit), never not_found; posts made with 0.4.1 or earlier are read as before. That is also the reading for a batch that landed in part, and for a later verify_import, including of a voucher Tally refused. Other new readings name why a voucher is missing: bound_not_in_window (it may have been deleted, re-dated or restored from a backup; whether an edit does this was not tried: check in Tally) and book_rolled_back (the books look older than the post). A post is refused as post_mark_unrecorded, with nothing sent and the approval withdrawn, when the company's voucher mark cannot be read before it (#1054; #1116, #1126 and #1107 for the refused and part-landed cases).
  • Clearer next steps on refused batches. When no attempt was made, a saved batch refused because a narration, reference or voucher number cannot be posted now says that nothing was sent, and to correct the text, build the batch again and post the new one (before, it said to review the error before asking for approval again, which fails the same way each time) (#1105). A batch that mixes a Journal with Payment, Receipt or Contra vouchers is refused as before, and now says to split it into one batch of Journals and one of the others (#1094). A date or company ID refused before anything is read now says what it should have been: a date as text in the form YYYYMMDD or YYYY-MM-DD, and the 36-character company ID from list_companies, not its name (#1115). The refusal for a changed ledger count names an altered ledger as a possible cause, marked as reasoned and not reproduced (#1032).
  • The posting path says what is sent. Narrations are sent as given, and the approval dialogs say "narrations sent as prepared, nothing added" instead of saying a batch reference is added. A narration that contains the character U+FFFD (a diamond with a question mark) followed by #, digits and ; is now refused when the batch is built, because it is compared byte for byte. The file you import by hand keeps its tag. A bound voucher whose amount was changed in Tally reads posted_divergent, and acknowledge_post_review finds a bound voucher by its Tally ID (#1054).
  • `purchase_register` explains a cancelled Purchase. A cancelled Purchase or Debit Note sits in the list of vouchers without a tax entry, with cancelled: true, because the cancelled vouchers measured came back from Tally with no ledger entries, so being listed there does not mean the voucher was untaxed. A cancelled voucher that keeps its entries has not been measured (issue #1013; #1027).
  • A statement line that reads 0 and 0 is compared, not skipped. profit_and_loss and balance_sheet now compare a Tally statement line that has 0.00 present in both columns with the line derived from the trial balance; a derived amount over such a line is reported as a difference and the statement is marked not established, with its derived lines withheld, instead of being called established. On four synthetic lab books every zero line came back empty in both columns, so this has not been seen in a real answer (issue #1067; #1071).
  • `outstandings` now reads one far-future due date. A due date that Tally prints with a four-digit year of 2100 or later (such as 1-Dec-2108) is read as written, and gives no overdue days in the ageing; before, that one bill failed the whole read with no reason. A bill row whose dates cannot be read still refuses, and the refusal now carries a named cause, the report, the row number when a row is what failed (also for an error found while the report is read, such as an unreadable amount or a repeated field, but not for a value that appears before any bill row), and a next step, never the bill's party, reference or date. A whole call on the synthetic book that printed such a date has since completed once (see Known limits) (#1098, #1128, #1136; issues #1091, #1096).
  • The two log tools now say what they do not show. read_evidence and egress_log say what each holds and that neither shows what your AI provider received, and tell the assistant never to say that no data has left the computer. What the assistant reads, amounts included, goes to your AI provider (#1045, #1113).
  • Source only, not in the extension: the desktop app's "Show in folder" opens only a file it exported since it started. ComplyEaze Bridge publishes no desktop installer, and the extension does not contain the desktop app. Any other path is refused before the file manager opens. A file replaced by a link after export would still open, the record keeps the 1,000 newest exports, and the Windows and Linux paths and the button itself were not tried by hand (#1051; issue #915).
  • Wording. The extension's settings are shorter overall (the redaction setting now says more) and the three longest read descriptions lead with the outcome; assistant texts say "ComplyEaze Bridge" in full, and so do the desktop app's messages (source only); the extension's author line reads "ComplyEaze contributors", which is why 0.4.2 installs beside an older version (#1042); acknowledge_post_review names the refusal for a review recorded for a different doubt (#1026, #1028, #1124, #1129, #1130; issues #962, #1010, #830).

Known limits

  • `sales_register` was not shown on an invoice-view Credit Note, an inter-state (IGST) line, a cancelled or optional sale, an unrecognised or missing duty head on a sale, more than one voucher in a window, paging, a company with a registration, a tax Tally computes itself, a duty head no sales run showed (such as cess), an invoice of another shape than the one run (for example several goods lines), a sale typed on Tally's screen, accounting-invoice mode, a post-dated sale, a REFERENCE or a filled PARTYGSTIN on a sale, a REFERENCEDATE on a sale, or a currency other than the book's base. The invoice it was replayed on was imported by this project. Its complete state rests on the company marks and the ledger masters reading the same before and after, not on a separate voucher count (#1009).
  • Long voucher lists: a change that moves neither company mark is not seen between pages, a window larger than that cap is not held and is read again for each page, and what happens if someone changes the book at Tally's screen between two live pages, or a later page of a window with a voucher type or ledger filter, was not tried against a live Tally. Not established: whether a save by another Tally Gold user, a restored copy of the company, or a company setting change moves the marks at once (#1053, #1118).
  • Ledger names: a spelling that contains the mask mark (… or ...) and is not typed exactly as the ledger is spelled is refused as ledger_name_masked, whatever the setting. When mask_parties is on, a spelling that shares a key with another ledger's masked form is refused the same way, which can also refuse an ordinary name typed in a different case until it is typed exactly, and the refusal lists no candidates. The candidates were tested on saved Tally answers, not a live book, and the ledger-name changes on a lab list of 144 synthetic names; a real book was not tried (#1057, #1092).
  • Posting, what was tried: two live posts through ComplyEaze Bridge's own path, of 3 vouchers each (3 Journals; a Payment, a Receipt and a Contra), made with the environment batch setting, and raw runs against Tally of 10 untagged Payment, Receipt and Contra vouchers and of 5 Journals, all on a synthetic company in TallyPrime 7.1 Silver. The extension posts one voucher per approval; that one-voucher path was not run live with 0.4.2's untagged matching (#1054).
  • Posting, what is not seen or not closed: if a post's vouchers cannot be matched (for example one was edited in Tally before the matching was made, or Tally stored a narration differently; only one Hindi-and-rupee-sign narration was checked), that is final: the batch stays "reconciliation required", acknowledge_post_review does not apply, and you check its vouchers in Tally yourself (#1039 is open). Another Tally Gold client writing during a post is not seen. After a backup is restored, if the books are keyed past the post's mark before the next check, the restore is not detected as one (the voucher reads bound_not_in_window); if Tally reuses the lost vouchers' IDs (not measured), a different voucher that took a lost voucher's ID and has the same content can read posted_verified (#1050). Such vouchers cannot be posted again through ComplyEaze Bridge and must be entered by hand (#1038). Not tried: whether a voucher keeps its Tally ID when someone edits it in Tally's screens (if it does not, verify_import reads that voucher as bound_not_in_window, which does not mean it was deleted), whether deleting a voucher can lower the company's mark, a Windows approval dialog, Education mode, and more than 3 vouchers through ComplyEaze Bridge's own path (#1054).
  • Statement gate (#1071): whether a real Tally ever prints 0 in both columns is not shown; the Cost of Sales : heading is not compared when it reads zero or empty (#1070).
  • `outstandings` on a book with such a date: one whole call on the synthetic book that printed 1-Dec-2108 completed in a lab run (state complete, 52 requests, 4.0 seconds, a page of 500 of 1,498 open bills). The bill with that date was not on the page returned, later pages were not read, and other books were not tried. A bill dated after as_of still refuses, and a two-digit due year can be read in the wrong century, as before. A bill cannot be left out and the rest returned as partial, because that would misstate the party's balance (#1136, #1098; #1091 stays open).
  • Assistant text: whether Claude Desktop passes the start-up instructions to the model at all, and whether any of the new wording changes what an assistant does, was not measured; the studies were plan-only, without tool calls or real users (#1058, #1113, #1115). The new opening lines were tested on saved Tally answers, and no live Tally was read for them, for the ledger candidates, for the date echo or for the open-bill counts (#1049, #1062, #1066, #1079).
  • Tried against TallyPrime (7.1 Silver in a lab, synthetic companies, one run each unless a line above says otherwise): the sales_register reads, the paging run, the untagged-post runs, and the rejected-voucher answers. The other changes are covered by automated tests, with a saved real answer behind the due-date item; the statement fix is tested on hand-built lines, because no saved answer has a 0 / 0 line. The release check starts each package, lists its tools and reads a sample bank statement, and does not run against TallyPrime. No one on our side installed the Windows package of this build in Claude Desktop on a Windows PC.
  • Pages: no lawyer or CA has read the Questions page answers on liability and claims or the changed Privacy Policy text; the registration number and address were checked against the certificate of incorporation, not the current record (#1084, #1104).

0.4.1, #

In plain words: ComplyEaze Bridge 0.4.1, since 0.4.0 (2 Oct 2026)

These changes are in ComplyEaze Bridge 0.4.1. Each line names the pull requests it comes from, except where it names an issue.

Should I upgrade?

  • If you use 0.3.0 or 0.4.0 on Windows: yes. 0.4.1 closes the network-path forms (a share, its WebDAV form, and the long-path prefixes) of a medium-severity security issue in how the bank-statement tool opens file paths, published as advisory GHSA-vm5g-r3p7-wxx7. Some paths still pass; the advisory lists which. If you cannot upgrade yet, follow the advisory's steps.
  • On a Mac: recommended, though not urgent. The advisory rates the issue low there, and the other changes below apply there too.
  • How: ComplyEaze Bridge does not update itself. Follow the installation guide: (1) close any other program that runs ComplyEaze Bridge; (2) in Claude Desktop, install the newer file from Settings, Extensions; (3) keep ComplyEaze Bridge's data folder, which holds its record of what it has sent to Tally; (4) check that the extension shows 0.4.1 and that "Allow voucher posting" is as you want it, since an earlier default may still be saved as on; (5) quit Claude Desktop completely and reopen it.
  • What was tried is under "Known limits" below. No one on our side installed the Windows package of this build in Claude Desktop on a Windows PC.

What you can do now

  • No new tool. On a company large enough to be counted first, voucher_presence can now answer absent for a voucher it finds nowhere in a date range that holds vouchers and was checked against that count. An empty range is called absent only on a company that has never held a voucher, as before (#985, #1020).

Safer or fixed

  • parse_bank_statement accepts a statement file or password file path only when its text names a local disk. A path that starts with two separators (two slashes, two backslashes or one of each), such as the Windows long-path form (starting with \\?\) or a network share (\\computer\folder), is now refused: copy the file to this computer and give its path there (#1024).
  • vouchers and voucher_presence now call a date range with vouchers in it complete by one rule: only when every voucher read was checked against a separate count of that range. Before, vouchers called any such range complete (unless it withheld a voucher), and voucher_presence called none complete. On a small or new company, which ComplyEaze Bridge does not count first, both now say partial (#985, #1020).
  • A voucher changed in Tally between the count and the read now refuses the read (voucher_window_part_not_admitted, cause part_census_mismatch), also when the range is read in one request, where it used to return without that check. Call again once while the book is quiet; if it repeats, read the range in Tally. ledger_movement, purchase_register, build_import_xml, verifying an import and the party detail of outstandings can give the same refusal. After a post was sent, this refusal on the read-back means the voucher may already be in Tally: use verify_import on the same batch and never post it again. In some cases, verifying an import for one day that Tally cannot serve in one request is now read in parts instead of refused, which sends more requests to Tally (#985, #1020).
  • When post_import refuses a batch before it is sent because some of its rows may already be in the book (import_preexisting_identity), the answer now names those rows by their transaction ids and says what to check in Tally next, instead of returning a bare code, unless the response size limit leaves no room for the list. The same batches are refused as before (#901, #908).
  • vouchers returns a voucher's master_id (Tally's internal voucher id) as the plain number ("1"), not as Tally sends it with a leading space (" 1"), so it matches the same id returned by verify_import (#989, #1021).
  • If voucher posting is turned on, you approve a batch in its approval dialog, and the post is then refused before it is sent (for example because the import journal is busy, the batch is not found, or the call names a different company from the batch's), your approval is withdrawn, and the next post that gets past those checks asks you again. Before, in these cases the approval stayed held for up to 15 minutes, and posting any other batch was refused until it was used or lapsed (#857, #904).

Known limits

  • A date range is called complete when the vouchers read match a separate count of the range. If Tally's own date selection leaves a voucher out, the count and the read both miss it, and the range is still called complete. A small or new company is not counted, so it cannot get absent; adding a count for it is tracked in #1029, with the multi-day and after-post cases (#985).
  • parse_bank_statement checks the text of the path. A mapped drive letter still passes, and so, on a Mac, does a path under a mounted network volume (#1024).
  • A batch refused just before it is sent can still return no rows and no next step. The rows named before that can be more than the vouchers the book holds, when several rows look the same; count the vouchers in Tally before leaving any row out. A row is still refused when a voucher that an earlier batch or a hand entry put in the book has the same date, type, ledgers, amounts and sides (#865, #901).
  • Tried against TallyPrime 7.1 Silver in a lab on 2 October 2026, before #1020 merged: one-day ranges on two synthetic books, and the read-back of one posted Journal. Not tried: a range of several days read in one request, the new refusal itself (seen only in tests), how often it appears on a busy book with several users, and other Tally editions. The other changes have not been run against TallyPrime; each is covered by automated tests. No one on our side installed the Windows package of this build in Claude Desktop on a Windows PC. The release check starts each package, lists its tools and reads a sample bank statement, and does not run against TallyPrime.

0.4.0, #

In plain words: ComplyEaze Bridge 0.4.0, since mcp-preview-0.3.0 (26 Sep 2026)

These changes are in ComplyEaze Bridge 0.4.0. Each line names the pull requests it comes from, except where it names an issue.

What you can do now

  • Read a company's masters as a list: voucher types (with their numbering method), godowns, units, stock groups, or ledger groups, in pages. The book is checked before and after the read. For godowns, units and stock groups, a book with very many masters is refused, with the size named, rather than answered in part (#952).
  • See what a party's unallocated amount is made of, from data the tool already reads, and what that data cannot tell apart. Each bill now carries its own date and credit period, and payable and receivable follow the sign of the bill's balance (#945, #946, #957, #959, #961).
  • See what ComplyEaze Bridge keeps on this computer. local_data_report (and bridge_mcp --local-data-report on the command line) counts files, bytes and the age of the oldest file by kind, and gives the state of the import journal and how many saved batches are not settled. It changes no book and deletes nothing; like every call, it records its own receipt line in the local log. It names no file path unless you ask on the command line (#925).
  • Read the purchase register of tax in the books: the Purchase and Debit Note vouchers of a date window that touch a ledger under Duties & Taxes, with the tax amount each entry records under the ledger's GST duty head. Other voucher types that touch those ledgers (Sales, Journal, Payment) are listed apart with exact counts (at most 100 are listed), and an entry whose ledger has no recognised GST head is listed and never given one. Nothing is posted and nothing is inferred: whether an entry belongs in a return is for you to decide (#971).
  • Ask outstandings for one party, with detail, to see why a bill is open and what an unallocated amount holds, party by party. A bill trail lists the allocations of each of the party's bills, oldest first, leaving out those in cancelled and optional vouchers, and says whether they add up to Tally's own balance for the bill. The unadjusted view lists the party's on-account, advance and pending credit or debit note allocations and compares the on-account total with its unallocated amount; a note is recognised only when its voucher type is named exactly Credit Note or Debit Note. A bill that does not tie, or whose identity is ambiguous, is shown as such and nothing is merged. The detail reads the company's vouchers from the start of the books, so a large book can be refused, with a next step (#981).
  • Read Profit and Loss and Balance Sheet. A figure is shown only when it ties line for line to Tally's own statement. Otherwise ComplyEaze Bridge's own statement is withheld; Tally's own amount for each line is still returned, with the derived amount on each line that differs, and the lines that differ are named. A book with stock items is expected to be refused, because no such book has been measured, and the tie-out itself has been measured on two synthetic books only (#774).
  • Read closing stock values per item. An item is returned only when the values add up to Tally's own Stock Summary and ComplyEaze Bridge read exactly as many items as Tally's own count; otherwise no item is returned, with the reason and the next step. A company with no stock items is told so. A value keeps Tally's sign, as in the Trial Balance: stock held is a negative number. Quantities are not returned, because nothing checks them, and names, parents and units are returned but not checked. Only a 31 March and small books are read (#980, #979, #1001).
  • Read books that define more than one currency. Outstandings set aside foreign-currency ledgers, and a rupee ledger with a foreign-currency balance, and name them. Compliance ledgers and the Trial Balance are read through the book's base currency. The vouchers tool withholds a voucher whose amount Tally stored in a foreign currency instead of refusing the whole date window. The desktop app follows the same read (#642, #647, #649, #715, #781, #824, #825).
  • Read a book too large for one read, in parts, including ledgers under parents that cannot be named, and have the ledger count cross-checked against the company's own count. A movement read names an oversized ledger catalogue and refuses an unknown ledger before it reads any voucher (#679, #885, #891, #936, #938, #939, #960).
  • Recognise SGST/UTGST as a GST duty head (#968).
  • The app and the extension carry the ComplyEaze Bridge tick logo as their icon (#950, #978).
  • A ledger read on a several-currency book, or one whose base currency is not INR, is refused before any request to Tally, instead of returning bare numbers (#751).
  • A recognised bank-statement cash line is now asked its purpose. An unanswered line blocks the import file; "don't know" posts to a suspense ledger, tagged and listed. Other unmapped cash lines still go to suspense, tagged and counted (#817).

Safer or fixed

  • profit_and_loss and balance_sheet no longer open with "state": "observed" when nothing was established. The top-level state is observed only when the tool's result is established; otherwise it is not_established, with the same reason as the nested result. This changes the tool's output; no figure, check or withheld line changes (#984).
  • Every tool is refused until the Terms of Use are accepted, and the extension carries a privacy policy (#943).
  • The party-master balance snapshot ends at today's date, not at a stray voucher date (#896).
  • The Markdown proof shows the verification status and any duplicate vouchers in the batch (#811).
  • ComplyEaze Bridge sends one request to Tally at a time, across all its running processes (#883).
  • Each send to Tally is recorded in the local log with its place, kind, size, outcome and time, and no book content (#918, #941).
  • ComplyEaze Bridge refuses to build or post a row that another batch already sent to Tally (#876, #898).
  • A bank statement's result returns no amount except the figures you supplied and an open cash line's amount, and names the saved proposals file by its id alone (#848, #850).
  • The configured redaction also applies to ledger names in posting and verification results (#851).
  • The native posting windows say ComplyEaze Bridge (#970).
  • Each package's build is attested, and the attestation is checked before it is published (#923).
  • If you are slow at the approval window, the agent's call no longer waits on it. The agent is told the approval is pending and asks again. A click made while no call is waiting is posted by the next call, not one call later (#792, #854).
  • The approval window for a batch names how many vouchers it covers, in its title and its button. The review window shows a debit the way the post window does (#757, #762).
  • A voucher you cancel in Tally after ComplyEaze Bridge posted it reads back as posted but not effective, not as changed. Cancelled vouchers are no longer reported as duplicates of each other (#771, #789).
  • A recorded review can no longer answer for a different doubt than the one it covers (#755, #769, #809, #813, #831).
  • Tally's own error text on a rejected line is read safely, including text with an & in it (#763).
  • Each tool now says whether it changes anything. Reads are marked read-only and say they only record local receipt lines for the call, never book content. Some assistants may now run the read tools without asking each time (#909, #921).
  • The tools that save a file on this computer are marked as writes: preparing an import file and reading a bank statement add new files, while verification and acknowledgement replace the batch's saved proof and are marked destructive. Posting stays marked destructive, and no tool is marked as reaching outside this computer. parse_bank_statement had been marked read-only by mistake (#909, #921).

Known limits

  • Stock: the closing-value total is checked against Tally's own Stock Summary on one synthetic book, and the item count against the rows on one synthetic company, and the sign of a value was measured on one synthetic company. A book with many stock items is refused: the read is for small books, and typical stock-heavy books refuse today. A book in which no item carries a value returns nothing. A book with stock items is expected to be refused by Profit and Loss and Balance Sheet, because no such book has been measured (#774, #980, #1001).
  • A book with very many masters is refused by masters for godowns, units and stock groups, and how common that is across real books has not been measured (#952).
  • local_data_report has not been run on a Windows host, and how it reports a Windows junction is unverified. It does not cover the desktop app's other settings, its database or its logs (#925).
  • The purchase register has been read from a disposable synthetic book only. A purchase typed on screen, item invoices whose purchase ledger sits in an inventory allocation, and books with several currencies are not covered, and it does not decide whether a Debit Note is a purchase return or a debit note to a customer (#971).
  • The party detail of outstandings has been measured on one synthetic book. Its cost on a large book is not measured, and a window that needs too many requests is refused (#981).

Also in source

  • The tax-audit engine gained its stock and party-monthly tests. The engine is still not in the extension, and its accuracy is not yet proven publicly (#738, #744, #788).
  • The desktop app moved to a Tauri release that fixes GHSA-w28w-mhc8-qvjv (#805).

Removed

  • The unfinished document-sync feature is gone from the source, and so from both the desktop app and the binary the Claude Desktop extension runs: the AXAL sign-in, the code that scanned a folder and uploaded the files you chose to ComplyEaze cloud storage, and the two hidden screens for them. No ComplyEaze Bridge tool could reach them, and no navigation led to them. After this change the only network client in ComplyEaze Bridge's own code connects to Tally on your own computer. The CI egress gate fails if a first-party crate other than the Tally transport depends on reqwest in its shipped dependencies (the app crate keeps it only as a dev-dependency) or if any first-party crate depends on hyper. It also counts the calls that send, connect or start a process in library and binary code, and refuses any outside the reviewed Tally files; test code is not counted. The unused delivery types in the portable core crate (no network code) went with it. The bulk party-statement export no longer records a hash for each file it writes and no longer refuses to write a file it cannot record; a record file left by an earlier development build is never read. The feature can be rebuilt from the git history if it is needed again (#914).

0.3.0, #

In plain words: mcp-preview-0.3.0 (26 Sep 2026), since mcp-preview-0.2.0 (16 Sep 2026)

These changes are in the published mcp-preview-0.3.0 build. Each line names the pull requests it comes from, except where it names an issue.

What you can do now

  • Post a Payment, Receipt or Contra voucher, as well as a Journal. Each one still waits for your approval in a separate ComplyEaze Bridge window (#585, #600).
  • Read password-protected bank statement PDFs into voucher proposals, including Union Bank of India statements (#444, #481).
  • See a party's GSTIN as it stood on a date you choose, from its dated registration history (#657, #700).
  • See each ledger's group ancestry, and the date its opening balance is as of (#498, #499, #569).
  • See the post-dated flag, invoice and reference details, and the party GSTIN on voucher reads (#498).
  • Pick vouchers by voucher class the way Tally classifies them (#663).
  • Work with books of more than 1,000 ledgers in the ledger catalogue (#643).
  • Get a whole verify_import verdict at once, with the verified vouchers in pages (#673).
  • Source builds only: with BRIDGE_AGENT_ENABLE_BATCH_POST on as well as posting, post 2 to 50 vouchers of one saved batch after one approval, which shows a summary by ledger. It is off by default and not in the extension until a live batch post is proven. A batch counts as clean only when Tally's counts match exactly; otherwise you review it (#712, #721).

Safer posting

  • Voucher posting is off by default in the Claude Desktop extension. Turn it on in the extension settings (#577).
  • An approval counts only when the approval window returns a fresh one-time token. How the window closes no longer decides it (#665, #704).
  • ComplyEaze Bridge confirms the company as its last step before posting, and reports where the voucher landed. A post can still land in a company that was renamed after that check (#607, #574).
  • It refuses to post if a ledger changed since the voucher was prepared, if the company's masters changed between its final checks and the post, or if the saved file no longer matches its record (#578, #615, #616).
  • If the company's masters change while a post is landing, the post is reported as not verified and needing reconciliation, so you check it in Tally (#623).
  • It records each voucher's import identity before sending it, and never sends one twice: it refuses an import identity its own journal already records (#582, #678).
  • It refuses to post into a book with more than one currency defined (#613).
  • When Tally rejects a line, you see Tally's own error text, kept short and safe to display (#695).
  • When you correct a voucher it built, ComplyEaze Bridge checks the correction against the book as it prepares the file. It does not post corrections itself; you import the file in Tally (#439, #620, #639, #660).

Clearer answers when it cannot answer

  • A refusal now names its cause instead of a generic failure. Causes include:
    • a read that ran out of time (#493);
    • an empty book (#489);
    • a missing company identity (#514);
    • no reply from Tally (#659);
    • a foreign-currency opening balance, or a book with more than one currency (#606, #706, #720);
    • a voucher type that doesn't exist (#677);
    • a list Tally would not serve (#719);
    • two ledgers whose names differ only by a hidden line break (#708).
  • Education-mode Tally: ComplyEaze Bridge refuses a read that Education would return empty without saying so (#588, #598).

Big books

  • ComplyEaze Bridge splits a long voucher window when Tally cannot serve it in one request, and sizes each request by expected response size, not date span (#494, #506).
  • The vouchers tool reports how long each request to Tally took (#608).

Tax-audit engine (in development, not in the extension)

  • A Rust tax-audit engine, bridge-tax-audit, now covers:
    • cash payments and receipts (s.40A(3), s.269ST, s.269SS/T);
    • depreciation;
    • s.44AB applicability;
    • financial statements;
    • the trial balance;
    • stale balances;
    • ledger scrutiny;
    • the cash book;
    • TDS payees;
    • s.43B and s.43B(h);
    • book-keeping quality;
    • loans and interest;
    • partners (s.40(b), s.194T);
    • bank reconciliation;
    • a high-value register. (#501, #504, #508, #560, #561, #571, #592, #593, #618, #636, #710, #713)
  • Its tests are mutation-tested: a check limited to the paths a change touches runs in the pull request, and the recorded list is re-run nightly (#646, #682). Coverage of every module is not claimed.
  • Its accuracy is not yet proven publicly. Issue #738 proposes how to prove it.

Security and upkeep

  • The tools workspace moved off a yanked chacha20 and a vulnerable rustls, and CI audits its lockfile (#455, #458).
  • CI now fails if an HTTP call appears outside the files allowed to make one. This checks where network code may live; it does not by itself prove that no data leaves the machine (#701).
  • We deleted legacy code that nothing reached (#473, #495).

Known limits of 0.3.0

  • The package is unsigned. It is for evaluation, not production.
  • Native Windows validation with Tally and Claude Desktop is outstanding (the Tauri desktop app's own Windows catalog flow is #293). The posting approval's decision also has no behavioural test on Windows (#702). Intel Mac is not qualified.
  • How long Claude Desktop waits on one tool call is not measured (#703).
  • There is no tool to delete or undo a posted voucher, and a ledger that was replaced is not always noticed (#623).

0.2.0, #

These are the detailed entries that were in this file when mcp-preview-0.2.0 was published. They were written before the plain-words summaries began, and there is no summary for that build.

Removed

  • DSC (digital-signature certificate) hardware-token detection, certificate extraction, and their AXAL sync path have been withdrawn from Bridge's scope, along with the pkcs11 and cryptoki dependencies that reached the PKCS#11 driver. pkcs11 0.5.0 was unsound (RUSTSEC-2022-0034) and unmaintained; the capability may be rebuilt properly later if needed. AXAL's Tally and Documents integrations are unaffected.

Changed

  • build_import_xml now reports live_evidence as an array of {observation, report, voucher_types} records rather than a single string, and no longer emits live_evidence_report. The previous shape could name only one source for a whole batch, so a Payment build cited a report that records Payment being refused. A client branching on the old string value needs updating; the accompanying voucher types make the provenance readable without one.
  • Relicensed future Bridge distributions from the MIT License to the Apache License, Version 2.0. The previously published v0.1.0 release remains available under the MIT License that accompanied that release.
  • Core snapshot canaries now authorize attempts through stable observed sealed-profile execution evidence without claiming field support from an incidental first-day dataset; snapshot rows are always re-fetched after the durable run starts.
  • A probe that no longer returns the selected company now clears and invalidates every company-scoped evidence, proof, mirror, diagnostic, and snapshot view before installing the replacement probe, so its fresh review remains usable without displaying stale company data.
  • Snapshot lifecycle probes no longer replace interactive setup-review state; restart admission uses the exact sealed Core receipt, and ambiguous duplicate live company identities fail before any snapshot read with their concrete terminal proof reason preserved.
  • Snapshot recovery now durably replays backward-clock abandonment evidence, enforces a 100,000-record aggregate hydration ceiling, and recovers an exact already-committed receipt from compact hash-bound proof authority without rehydrating canonical membership.

Added

  • Local import files may now carry Payment, Receipt and Contra vouchers as well as Journals, so a bank statement can be expressed in the voucher types Tally files it under. Each of the three is admitted only as two entries over two distinct ledgers carrying neither a voucher number nor a reference. The side that must hold money is refused unless that ledger's live group ancestry reaches a reserved Bank Accounts or Cash-in-Hand identity — the two where a captured ledger is observed sitting under a captured group. The counterparty side must be established as holding no money: any money group there means the voucher is really a Contra, and a ledger whose group ancestry cannot be resolved is refused as well, because neither leg is admitted on an absence of evidence. A build that names a counterparty warns that its amount lands On Account. Native posting is unchanged and still accepts only one unnumbered Journal.
  • A local-first Tally Truth Layer with capability passports, explicit truth states, encrypted mirror evidence, resumable/adaptive snapshots, Proof of Sync and Gap Map output, and a safer operator console. The migrations are additive; rollback requires restoring the prior application and retaining the encrypted database for forward recovery rather than deleting evidence.
  • Portable, bounded Tally protocol, canonicalization, transport, runtime, compatibility, incremental-policy, qualification, observability, and write-safety crates backed by a synthetic loopback protocol simulator.
  • Reviewed single-use setup authority, exact selected-read qualification, and fail-closed compatibility manifests/runbooks. Live Education behavior and every write capability remain unknown or disabled until exact reviewed evidence exists.
  • Native Windows and macOS CI coverage for formatting, tests, builds, and Clippy.
  • Repository-local Windows and macOS application icons.
  • Open-source contribution, security, review, and rectification guidance.
  • Reproducible Node and Rust toolchain baselines, installer smoke builds, and complete lockfile-to-license-inventory checks.
  • Automated legal-resource inspection for Windows MSI/NSIS installers and the staged and DMG-packaged macOS app bundles.

Security

  • SQLCipher/keyring-backed local Tally state, immutable proof/checkpoint receipts, loopback-only HTTP without a proxy, bounded incremental decoding, cancellation and lease enforcement, idempotent crash replay, and sealed no-write qualification boundaries.
  • SQLCipher pool replacement connections now receive raw key bytes from zeroizing storage without retaining a key-derived pragma string. Proof contract v3 binds detailed record counts, and historical crash recovery no longer depends on current checkpoint ownership.
  • File-backed snapshot ownership now uses per-run kernel advisory locks, so a crash can be reclaimed after wall-clock rollback without allowing a live owner to be stolen. Persisted/live company profiles correlate through an opaque endpoint-scoped identity key, and macOS qualification reports ru_maxrss in its native byte units.
  • Losing checkpoint compare-and-swap decisions terminalize as durable failed proofs and close staging attempts instead of remaining falsely resumable; unrelated checkpoint advances do not rewrite Failed or Cancelled outcomes.
  • Compatibility claims now require verified synthetic-fixture identity before an explicit parsed Tally application rejection can establish Unsupported; fixture, context, sentinel, parser, malformed-response, and transport failures remain fail-closed observations rather than incompatibility claims.
  • Updated the XML parsing graph and removed unused Linux-only dialog dependencies from the supported Windows and macOS build graph.
  • Updated the Tauri runtime to 2.11.5 and tauri-runtime-wry to 2.11.4.
  • HTTPS-only AXAL endpoints with redirect blocking, bounded responses, and credential validation.
  • Safer DSC PIN transport and PKCS#11 library discovery without exposing arbitrary native-library loading to the webview.
  • Bounded Tally and document responses with endpoint and upload validation.

0.1.0, #

Added

  • Initial open-source Bridge application with React, Rust, and Tauri support for Tally, GST, DSC, document, sync, and local database workflows.