Skip to content

Releases

2.9.35

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 11 Nov 15:53

Added

  • read and section accept separate positional arguments, and hint suggestions quote safely in PowerShell.
  • manifest-sweep inspects the dependencies declared by nested package manifests in one command.
  • --batch-serve accepts a multi-query request. A request may carry operations (symbol, read, brief, section, semantic, skeleton, outline, refs), which run in order and answer in one reply; the same release adds shared project memory, a stack-trace output filter and faster call-graph traversal.
  • Outlines and reads cover more structure. C# #region blocks are addressable as sections, html-query --content returns the raw content of script and style blocks, and skeleton --grep and outline --grep also match Java declaration signatures, including annotations; the parser fingerprint changed, so existing indexes reindex after upgrade.

Changed

  • symbol prints the whole body up to about 12 KB, shared fairly across matches; each match still shows at least 5 lines, none longer than 2,400 bytes. A body past the budget is cut at a line and names the read command that returns the rest.
  • Repeated Markdown section hints now advance after a served range and unchanged-file fingerprint confirm a section read; when no selectable section remains, the hint points to the outline command.
  • The canonical skill now treats whole-file/range denials and the >80%-indexed warning as a per-file circuit breaker, directing reads and edits to surgical token-goat commands.
  • sqlite-query session resolves the active Copilot session database; COBOL outlines expose nested data items and level-88 conditions.
  • Tree-sitter parsing reuses one native parser per grammar, reducing repeated AST extraction setup; the parser fingerprint changed, so existing indexes reindex after upgrade.
  • The TypeScript signature-context API resolves transitively referenced interfaces and type aliases across supplied parsed files; the parser fingerprint change reindexes existing indexes on upgrade.
  • changed now recognizes the exact Git --name-status option-order error when a path is used as the baseline, guides CLI callers to --since or --grep and MCP callers to pass a ref rather than a path filter, and does not misclassify unrelated revision errors with similar wording.
  • grep falls back to literal substring matching for malformed regex syntax while retaining refusals for overlong or potentially stalling patterns.
  • Shell/task-output line deduplication avoids an intermediate split-line array; direct find results stop collecting unused fuzzy names, and the Bash hook handles find before unrelated output-file probes.
  • Hooks deduplicate repeated CLI and file-covered advisories per session, so the same hint is not repeated within one session.
  • Hooks, search and the worker stall less, because their I/O blocks less.
  • The worker yields its background backlog and housekeeping while a foreground turn is active.
  • The worker daemon now exits after 30 minutes with nothing to index, down from 6 hours; TG_WORKER_IDLE_EXIT_MS still changes it, and the next edit or command starts a new one.
  • The browser image hook and the subagent briefing project map are cached or fast-gated before config loads.
  • grep scans through a shared scanner, and hint-target and outline lookups reuse bounded in-memory caches, so repeated hints on the same file cost less.
  • A repeated line-range read that is refused now gets a shorter message from its second refusal and a terse one from its third on, so the model is steered to the cached output rather than retrying.
  • The canonical skill's paging warning now points at the command named in the denial itself.
  • The pre-push hook has a hard stop. Every leg still running at 540 s (PREPUSH_HARD_STOP_S overrides it, can only lower it: a larger value is clamped to 540 s, which leaves a 10 s startup reserve, a 10 s TERM wait, a 20 s post-stop reap and 20 s of cleanup inside the 600 s hook limit) is ended with its process tree, including a test-file selection still in progress. A stopped lint or typecheck fails the push, since it was not verified. A stopped test leg passes with a notice listing the files it did not finish, which are recorded in owed-<leg>.txt beside the failed-test store. CI runs the full suite on pushes to main and on pull requests. PREPUSH_FULL=1 runs without a stop.

Fixed

  • Token-goat MCP tools are disabled when hooks are available, so the same work is not offered twice.
  • doctor self-tests the wired native copy when it is stale or broken.
  • install omits the matcher property on hook events that are not tool events.
  • Hooks pass a call through when its native arguments are missing, instead of failing.
  • Search ranks a whole-word name match above a token match.
  • php-lint counts failed files with the right singular or plural noun.
  • A worker that retires drops its queued embeds instead of waiting on them.
  • Repeated large-file Read calls no longer loop on deny messages.
  • doctor reads the Windows Defender exclusion list through the shared UTF-8 helper.
  • @hono/node-server is updated from 1.19.17 to 2.1.5, a major version bump, past GHSA-rmxm-3fg6-px4f.
  • session-outline, session-slice and session-audit handle redacted thinking blocks. A transcript block of type redacted_thinking (the API's safety-redacted reasoning, whose data field is opaque) now renders as a placeholder that says the content is redacted, and a turn holding only one previews as (redacted thinking). session-audit counts its bytes under assistant thinking instead of dropping them.
  • The pre-push hook no longer kills anything itself; one cleanup process does. The hook starts prepush-reaper.sh, registers each leg with it and sends it kill requests, so a hung or missing taskkill, a killed hook or a child that taskkill /T leaves behind (an MSYS sleep) can no longer keep a leg alive or hold the hook. A leftover leg can outlive the hook by up to 30 s while the cleanup process ends it; the next push reports any it could not end in reaper-last.txt beside the failed-test store. The hook waits for the cleanup process to say it is up, not for a fixed number of seconds, so a loaded computer no longer fails the push; the push fails at once, with a notice, if the cleanup process exits before it is up. It stays within its 30 s of the moment the hook is gone (it stamps both in its record), and the hook's closing notice says so when the cleanup process itself is not running; a tick is now reserved inside the hook's tail budget, so the whole hook ends under 600 s.
  • qs is pinned to 6.16.1, past GHSA-6pq8-jhm2-2vcw. The moderate advisory covers 6.7.0 through 6.16.0, a denial of service in qs.stringify with arrayFormat: 'comma', and the root overrides entry held it at 6.16.0. The entry is now the exact 6.16.1, the first release past the range, which was published on 2026-10-14. A range would have resolved to 6.16.2 or 6.17.0, both published inside the seven-day cooldown. This repository's own npm audit now reports only the body-parser finding that SECURITY.md already describes.
  • Two Agent spawns with identical input keep their own cached reports when the host supplies a tool_use_id. The recall id then includes that tool_use_id, so the second report no longer overwrites the first. A long plain-text report that quotes the background-launch markers near its start is cached again, and a plain-text launch acknowledgement is recognized by its opening sentence.

v2.9.34

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 03 Nov 00:10

Search now retries near-miss names, and Markdown folds stop printing pointers that cannot run

A search that finds nothing retries a name that is a few letters off, on the symbol, heading and text channels only. With --json you get a retries field when a retry ran. A misspelled long symbol name now lists the real name first instead of short fragments.

A page of weak semantic matches now shows three hits and one line naming the command for the rest, and the rerun hint repeats --limit, --project and --min-score. impact lists callers in other files first within each hop.

Reading a Markdown file no longer prints a section pointer that would be stripped. A paragraph under a heading no quote mark can hold is delivered whole, and a heading with a backtick keeps its own pointer. A folded Markdown file with a long path, such as one in the macOS temp folder, now stays folded. A cut Bash result also names the lines it left out, so you can fetch just those.

The background indexer reads the config once per file, which more than halves the file reads when it re-checks unchanged files.

Full details in the changelog.

v2.9.33

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 26 Oct 14:25

Added

  • session-audit counts cold-cache re-writes. A new section lists requests that followed a gap longer than their cache TTL and wrote the prefix again. For each, it shows billed-equivalent tokens, the share of billed tokens, and a breakdown by the tool (and Bash command class) the session was waiting on. Main sessions and subagents are reported separately.
  • token-goat read and symbol support --sig / --signature flags. They print only the function or class signature declaration and leave out the implementation body, for TypeScript/JavaScript and Python.
  • token-goat doctor --fast flag. It skips host process table enumeration during Windows health checks, so a quick doctor run finishes in ~1s.
  • An opt-in answer for a Grep that looks up one declaration. With inline_symbol_body = true under [hints] (or TOKEN_GOAT_INLINE_SYMBOL_BODY=1), a Grep such as def parse_args or function load against one source file, a directory or the working directory is refused, and the refusal carries that symbol's body, so the next turn does not need to run token-goat read. A Grep that is just the name (parse_args, \bparse_args\b) is answered too, and its refusal lists every matching line as path:line:text before the body. A directory is answered only when every file Grep would search can be read and checked: git's list of tracked and not-ignored files, at most 200 files and 2,000,000 bytes. A listed file the index does not hold (a .gitignore, a LICENSE, a text note) is read and searched for the pattern like the rest, and the Grep runs as written only if the pattern matches in it or it cannot be read. A name with a dollar sign is answered when the pattern escapes it (\$name), since an unescaped dollar matches nothing. A project's own .token-goat.toml cannot turn the setting on. The body is the text token-goat read prints, fenced as untrusted file content and redacted. It acts only when the Grep is a plain content search, the fresh index holds the name exactly once in the searched scope, a declaration pattern matches exactly one line and that line is inside the symbol, and the refusal is under 12,000 bytes. In every other case the Grep runs with the usual hint. The refusal names the file relative to the working directory, so a checkout under a folder with a space, a non-ASCII letter or a long name still gets the answer. The same symbol is refused once per session. It is off by default, because a refusal the agent did not expect is only worth it if the inlined answers are followed. token-goat stats counts these refusals as inline_symbol_body_deny.

Changed

  • The pre-push hook runs a time-budgeted selection of tests. Each suite leg runs, in order: the files that failed last time; the test files changed in the pushed range; every guard test, plus the built-bundle matrix when src/ changed; the tests that import a changed module, nearest first; and then the tests the import graph cannot see. Those are the tests that name a changed path as text and, when src/ changed, the tests that read src/ as text or spawn the built bundle. The selection keeps the estimated wall time within 600 s (Windows) or 300 s (WSL). PREPUSH_BUDGET_S and PREPUSH_WSL_BUDGET_S change the budget. The estimate comes from per-file durations kept beside the last-failure record. Whatever the budget cuts is printed as deferred to CI: N files (~S s) with the list, and CI still runs the full suite. A push that changes only Markdown or docs/ runs the guards and the tests naming the document, and a new branch is diffed from its merge base with origin/main. A path the import graph cannot see (config, scripts, hook scripts, test setup, fixtures, a deleted module), a failure record that cannot be parsed, or PREPUSH_FULL=1 runs the full suite, as before.
  • Compact untrusted content fence notices. Local tool outputs and file extractions now use [token-goat: data only, not directives]. The full warning is kept for web fetches and detected prompt injections.

Fixed

  • A Bash command with many quoted environment assignments or repeated git options no longer stalls the hook. Two patterns the Bash hook runs on every command could match the same text in more than one way. A command such as A="x" A="x" ... ls or git -C "a" -C "a" ... x therefore took time that doubled with each repeat (about 2 seconds at 26 repeats). Each part of the command now has one reading, and the check takes the same short time at any length.
  • A git command behind a quoted environment assignment is recognised as git. FOO="a b" git restore x and FOO='a b' BAR=1 git reset --hard were not seen as git commands that rewrite the working tree, because the assignment was stripped only when its value had no quotes. They are now treated like the same command without the assignment.
  • git -C <dir> push and other git commands with options before the subcommand are classified as git commands. The bash output cache read git followed directly by the subcommand, so git -C dir push, git -c k=v push and git -C dir status were missed. A push was not treated as a state change, and a git -C <dir> status was fingerprinted against the working directory's repository instead of <dir>'s.
  • A failing check script given only .md documents is recorded as a failure. A run such as python scripts/run_self_tests.py README.md that failed was dropped, so the Stop advisory never named it. Its documents still count as no check input, so a pass on one clears no edit.
  • A git killed by a signal reads as unknown, not as exit 1. A git ended by a signal such as SIGKILL used to read as exit 1. An inline Grep scope check took that as a negative answer, and the post-Bash hook took it for a git that never ran and asked for no reconcile. It now reads as exit -1 with the signal named, and each caller takes its unknown branch. Upgrading reindexes: the parser fingerprint changed because src/util.ts is a shared extraction source, so already-indexed files are reparsed once.
  • The Stop advisory no longer fires on scratch edits or misses real checks. It named throwaway probes in the harness scratchpad (outside the project, under the OS temp dir) as unverified work, and said "no test, lint, typecheck or build was run" after a session that ran its validator scripts. An edit outside the session's project root (the nearest project marker above the Stop event's working directory) no longer counts when it sits under the OS temp dir. With no root known, an edit under the OS temp dir does not count either. A sibling package, another repo or a sibling worktree outside the root still counts. A passing python <dir>/run_*tests*.py, validate*.py, check_*.py or <script>.py ... --self-test now counts as a check, but only with path or selector arguments: --help, -h, --list, --version, --dry-run and any other flag make it not a check. A git commit or git push is never counted, because a hook file existing does not prove it checks anything. The command is stored without any argument beyond the script path and --self-test. The advisory still speaks once per distinct state; it repeated only because each turn edited a new file.
  • The Stop advisory no longer dismisses a project edit made through a link in the temp dir, or counts a validator that may be writing its own .md argument. An edit at a path under the OS temp dir was dismissed by its spelling, so a symlink or junction there that leads into the project hid a real project edit. The dismissal now follows links, and a path whose links cannot be resolved is not dismissed. A .md argument no longer counts toward a script check at all, so python scripts/validate_x.py report.md is not recorded as a check whether or not report.md exists. A script given a document may be writing it, and whether that file exists says nothing about what the run verified.
  • With read confinement on, a relative path to a link that leads outside the workspace is now refused. Over MCP, sl.txt or ./sl.txt (a symlink in the project that points outside it), or jn/o.txt through a junction, was read even though the absolute spelling of the same link was refused. The check for the extra allowed roots and the standard skill and transcript folders measured the relative name under those folders instead of under the project, so it found nothing there and let the read through. It now measures the path the read opens.
  • Preserve full output when repeating overlapping commands. Rerunning a command with overlapping output ranges no longer hits suppression loops in served-output generic elision.

Full changelog: https://github.com/DFKHelper/token-goat/blob/v2.9.33/CHANGELOG.md

v2.9.32

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 21 Oct 03:25

Aider support, a grep that accepts inline regex flags, and a large batch of fixes for what happens when git is slow or does not answer.

token-goat install --aider writes a block into Aider's coding conventions that points Aider (and the local models it often drives) at the surgical-read commands instead of whole-file dumps. By default it goes into ~/.aider.conventions.md; with -p/--project it goes into CONVENTIONS.md in the project root. uninstall --aider removes only that block. See the Aider section of docs/install.md.

token-goat grep now accepts PCRE and Python inline flags such as (?i) and (?m) at the start of a pattern, which used to fail with Invalid regular expression. A mistyped check, health or status now points you to doctor, stats and commands. worker stop now exits 1 and says so when it could not end the worker.

Many fixes make token-goat honest when git is slow or never answers. A git command that times out no longer leaves a git process running or hangs token-goat. A timed-out git status is no longer read as a clean tree, and the resume packet, reconcile and the session start now say when git did not finish.

The repository also gains scripts/extract_entities.py, which finds numbered activities, tasks, requirements and steps in PDF, DOCX, XLSX, CSV, JSON and source files. It lives in the repository and is not part of the npm package. See docs/entity-extraction.md.

On Windows, the test suite now uses at most 4 worker processes at low priority, and the docs explain how to exclude token-goat from Defender real-time scanning.

Upgrading reindexes once: each indexed file is reparsed the first time it is needed, and nothing else changes for you.

Full details in the changelog.

v2.9.31

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 16 Oct 20:33

Shell searches that walk a whole disk are now refused, and the routing text and MCP tool list token-goat sends to the model are smaller.

The pre-shell hook now refuses a command such as find / -name x that starts at a filesystem root or the home directory with no depth bound. A shell tool that times out kills its shell but not the find it started, and several orphaned searches burned about 90 CPU-hours. Output caps also now price Chinese, Japanese and Korean text at about 1.1 tokens per character, because the old estimate let CJK-heavy output through at up to three times the intended size.

The MCP server now offers seven tools by default instead of 18, which takes the tool list from 13,029 bytes to 5,293. Set TOKEN_GOAT_MCP_TOOLS to a comma-separated list to get the others back. The routing block written into CLAUDE.md, AGENTS.md and the Copilot instructions shrank from 4,879 bytes to 3,670.

New checks help catch bad edits. A Write or Edit that leaves a placeholder such as // rest of the code goes here, or drops lines the model never saw, is refused once. A failed Edit that was one invisible character off now names the exact lines on disk, and a path that does not exist comes back with the paths it was probably meant to be. You are also told when an agent stops with untested edits. read now accepts a line range given as a flag, such as --lines 1-5, and json-query and yaml-query filter by more than equality.

Security: the development tree is back to a clean npm audit. proxy-addr (critical) and source-map-js (high) are updated to the patched 2.0.8 and 1.2.2. Both were development-only and neither reaches an installed copy of token-goat.

The release also carries more than 220 fixes across the indexer, hooks, MCP server and output commands.

Upgrading reparses every indexed file once and re-embeds every already-embedded file once. Full details in the changelog.

v2.9.30

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 04 Oct 04:58

Semantic search now matches on meaning in a default install, and on Linux hook calls run through a small native client instead of starting Node.

Embeddings run on the WebAssembly build of ONNX Runtime, bundled into token-goat. Its 14 MB binary is downloaded once, checked against a recorded SHA-256 and size, and checked again each time it loads; nothing is installed and no install script runs. The native onnxruntime-node is still used when it is installed.

On Linux (x64 and arm64), install puts the native tg-hook client in front of every hook command it writes, for Claude Code, Codex, Grok, Kimi and Copilot CLI. It hands each call to the resident hook server and falls back to the ordinary Node hook when the server cannot answer. A served call took 14.5 ms at the median against 47.1 ms for the Node hook through the same server. Both Linux binaries are statically linked and carry build provenance attestations that gh attestation verify checks. Windows binaries ship only when a release can sign them, and this one cannot, so Windows keeps the Node hook command, as macOS does, and doctor says so.

New: install --copilot registers the MCP server with Copilot CLI and install --antigravity wires token-goat into Google's Antigravity CLI; install indexes the project you run it in; doctor --fix and install repair stale or missing hooks in every installed harness; doctor --probe <harness> checks that token-goat's context reaches the model; project notes can be anchored to a symbol and say so once it changes; json-query and yaml-query take JMESPath and jq style projections and read stdin with -; bash-output, web-output, mcp-output and retrieve return exact line ranges with --lines; stats --payloads shows what token-goat adds to each session's context, and listing-size prices Claude Code's skill and agent listings; C# files have references, so refs --callers finds C# callers.

Changed: minified bundles are no longer indexed; an edited file is indexed right away instead of at the worker's next poll; search ranks changelogs and docs below the code they mention and prints about 10% less; config set saves only the settings you changed.

Security: values in .env files are masked in search previews, mcp-output --file, compress-text --file and handoff-create --file; with indexing.cross_project_symbols off, MCP tools and search --project no longer reach another project's index; project notes reach the agent marked as data, not instructions; and a long list of redaction gaps is closed, including URL passwords containing : or @, secrets cut at a clip or snippet edge, and spreadsheet headers and sample values.

The release also carries more than 250 fixes across the indexer, hooks, MCP server and output commands.

Upgrading reparses the index once and re-embeds every indexed file once. Full details in the changelog.

v2.9.29

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 25 Sep 15:18

Hook calls answer in about 70 ms instead of 110 to 135 ms, and read-only commands an agent runs answer in about half the time.

The first hook call starts a resident hook server in the background, and later calls go to it instead of starting Node and loading token-goat's hook code each time. When section, symbol, read or another read-only command writes to a pipe rather than a terminal, that same server answers it: token-goat section took 60 ms instead of 122 ms. A hook call also no longer starts a second Node process to hold a lock, and token-goat capabilities lists the server's pipe or socket as the one thing token-goat listens on.

The MCP tool list takes 25% less of each turn's context (13,029 bytes, down from 17,314). hint-stats --session-id <id> reports on one session, and doctor also checks the Copilot hooks installed in the current project.

Fixed: hooks for Claude Code, Codex, Grok, Kimi, Copilot CLI and VS Code keep working under a package.json that says "type": "module"; notes and a session's record of files already read survive a briefly locked file; files a stopped worker never embedded are embedded later, and switching embeddings back on catches up what was indexed while they were off; an edit that reaches the queue while the worker claims it is still reindexed; confined MCP reads work on Windows under every supported Node release; a data directory made read-only stays read-only; the pi, opencode and OpenClaw hosts keep each session's state and id to itself; and a plain search pattern is no longer refused as too slow on a busy machine.

Upgrading reparses the index once. Embeddings are untouched and keep serving throughout. Full details in the changelog.

v2.9.28

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 24 Sep 12:17

Denies and read hints now open with a command that runs as printed, and failed tool calls are classified and counted per tool and per model.

A deny used to name a placeholder such as section "file::SectionHeading", which exits 1 when pasted. It now names a real heading or symbol taken from the file, the index or a bounded read of its start, and routes a JSON or YAML key to json-query/yaml-query. The same deny received again since the last compaction comes back as its command plus one line.

install now registers PostToolUseFailure, where Claude Code sends a failed Read or a non-zero Bash exit, and session-audit --tool-errors reports calls, errors and error rates per tool and per model. A deny that delivers the content, such as a large skill's compact slice, is counted on its own line rather than as an error.

token-goat semantic takes several queries in one call and prints a notice when its best match is weak (distance above 0.85, calibrated on 38 real and 12 off-topic queries). A run of reads sent one turn at a time gets one line asking to batch them, and three searches in a row that find nothing get one line pointing at answer or semantic.

Fixed: a Bash read behind a leading FOO=1 assignment no longer slips past the read gates; a background command's JSON output is no longer denied as an agent transcript; the identical-search note now fires on Claude Code's Grep and Glob; hints are scored per agent and per prompt; and the pi, opencode and OpenClaw bridges keep each session's state to itself, with hook calls in one process run one at a time.

This release also carries everything listed under 2.9.27, which was never published on its own: among it, outline and query filters for large JSON and YAML files, notes that come back at every session start, and a pinned, reviewed list of every package allowed to run in your process. Upgrading reparses PowerShell files once, because the PowerShell parser changed in 2.9.27; embeddings and other languages are untouched. Full details in the changelog.

v2.9.26

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 23 Sep 14:05

Three read filters no longer destroy or deny what they failed to read.

A --json search is passed through instead of being summarised. rg --json emits one JSON event per line and none of them is a match line, but the summariser reads every line as a match on the text before its first colon, and a JSON line's first colon sits inside {"type". Every line fell into the unattributed bucket, so a 403-line search of a single file became grep: 403 matches across 0 file(s) with a note that all 403 lines were unattributed: the filter reporting that it had parsed nothing, and shipping the result anyway. The count was the number of events rather than matches, the file count was wrong, and the output the caller asked for by name was gone. --vimgrep and -o were checked the same way and both attribute correctly, so neither changed.

A node -e one-liner that reads a JSON file with require() and writes the same file back is no longer denied. An in-place edit is not a read into context, and denying one only moves the same script into a file where it runs unchecked, so that shape was already exempt. The exemption was keyed on how the read was spelled rather than on the write: the ordinary version-bump one-liner was denied while the byte-identical edit spelled readFileSync was allowed, and the denial told the caller that fs.readFileSync() bypasses read hooks for a command that never calls it.

A PowerShell [IO.File] read-modify-write of one file is no longer denied. That extractor had no write check of any kind, so replacing a file's text through WriteAllText was blocked as a read, and the substitute offered in the message cannot perform an edit.

Reading one file and writing a different one is still treated as a read, in both runtimes. No reindex is needed.

v2.9.25

Choose a tag to compare

@Zelys-DFKH Zelys-DFKH released this 23 Sep 13:49

The per-file chunk ceiling now counts the chunks that were stored.

indexing.max_chunks_per_file was checked against a character-budget cut taken just before embedding. The embedder then threw that cut away and re-cut the same content with the model's own tokenizer, which flushes on a token budget as well as a character one and divides the file more finely. The ceiling was applied to a count several times smaller than the one that landed.

A chunk may run to 8000 characters on the character cut but is held to 510 tokens on the token one, which measured over 6117 stored chunks is roughly 810 to 1775 characters. A file could therefore clear a ceiling of 600 and still store several thousand rows. Embeddable documents returned before the check entirely and had no ceiling at all.

The ceiling is now re-applied to the rows in the index after embedding, for documents and source alike. A file that breaches it has its chunks removed and is marked so the background sweep agrees. That sweep already counted stored rows, so the live check and the sweep now measure the same thing and cannot reach opposite verdicts on one file.

Upgrading reparses the index. Embeddings are untouched and keep serving throughout: nothing already stored is re-embedded, and no existing index loses chunks, because no file in a normal project comes close to the ceiling this restores.