v0.41.50: module evidence for qualified references, one argument contract, one extraction census #574

Merged

buildagent merged 137 commits from integration-1111 into master 2026-11-12 11:04:34 +01:00

Conversation 0 Commits 137 Files changed 81 +6890 -998

buildagent commented 2026-11-12 05:46:24 +01:00

Member

Copy link

Release train for v0.41.50. Seven lanes, each reviewed from its diff and re-run by the lead, integrated on integration-1111.

What changes for a user

• Resolution (#566). A qualified reference is not bound across files on name evidence when the file's imports or module declarations contradict the candidate; where module evidence names one candidate, it is bound. Against the last release on the nine pinned repositories: 19 binds removed (all wrong before), 228 added (every candidate addition audited; a blind second audit of 65 rows agrees), 1 retargeted, 133 rule-retagged.
• MCP arguments (#570). One policy per argument drives the published schema, the runtime check and the error. invalid_argument_value replaces ten older codes; the four malformed-argument codes share one field set. Every other error body is unchanged and nineteen are compared with bodies recorded from the last release. Reference: docs/mcp-arguments.md, checked against the schema both ways.
• Incomplete extraction (#545, host side). One census behind project_overview, resolution_gaps, index_coverage, list_files and evidence_gaps; host refusals are counted; non-UTF-8 source files are recorded as content_refused.
• Small fixes. read_code with rev for a path whose directory is gone (#568); application/json on every JSON resource (#569); code-index query on an invalid configuration prints the refusal and exits 1; search_text on a permanently excluded literal path says so (#521); installer check reports a code-index elsewhere on PATH, and the release notes' build line names the run's tag (#572).

Behaviour changes to call out

• Scripts matching invalid_max_depth, invalid_token_budget, invalid_direction, invalid_role_filter, invalid_ref_kind, invalid_resolution, invalid_response_format, empty_seeds, too_many_seeds or invalid_arguments need invalid_argument_value.
• code-index query with an invalid .code-index.toml: exit 1 with a JSON refusal (was exit 2, no output).

Evidence

• Content, cost and stage baselines are blessed in one commit; each reason carries the bind account or the attribution. Content movement (rust-ripgrep +18, ts-zod +13, js-express +1, php-guzzle −2) equals the audited account.
• Cost against the last release: VM steps +0.7 % to +1.7 % on seven repositories, −14.0 % on ts-zod, −2.5 % on rust-analyzer; full scans +0.7 % to +2.0 %. Two costs were attributed per statement and removed before the bless (five module-evidence statements; one capability check that flipped a plan). The nine bind accounts are byte-identical across every reduction.
• Every package-supplied input of the new rules passes a capability grant; the pool, ref-kind and visibility registries are updated and a test shows an import without reachability_anchor cannot disambiguate a pool.
• tools/list 59,167 bytes (ceiling 59,311), skill resource 3,949 tokens (ceiling 4,000); no ceiling was raised.

Qualification

One full gate run on 05ea92f9 (empty package store), 43 steps: 40 green at once, three red, each attributed, fixed in its own commit and its step re-run green.

Step	Cause	Fix
agent_bench	A change in this train (713b1423) emptied change_impact's test_role_files at limit: 0; the benchmark's which-tests questions use that call. Recall 1.000 → 0.974 (rust-ripgrep), 0.970 (python-flask).	3339b685 restores the release behaviour and keeps witness_paths_total. Mutation run red, restored green. Recall 38/38 and 33/33.
tier3	Record not re-recorded with the other baselines. Drift equals the audited account: rust-analyzer +179 resolved (194 added, 15 removed), 120 retagged; py-django 6 retagged.	7db6ebbe bless with the account as reason; green without the bless flag.
pkg_javascript_full	Declared unbound-by-scope 42, measured 41 on js-express. One site left the set: test/app.request.js:12:19, a write the builtin bound to an unrelated local; the host now refuses it, the package never bound it. Other three declared counts unchanged.	f4d07dbb declares 41. The file is a test declaration, not in the package or its digest. Full JavaScript leg green.

After the fixes: code-index-mcp suite 1,082 passed on each leg (snapshot and daemon), fast gates 519 passed, fmt, clippy and docs exit 0. The whole workspace was not re-run after the three commits; they touch one MCP source file, one MCP test, one test declaration and one record.

CI round 1 (f4d07dbb): Windows 3 failed of 5,404, Linux 1 failed of 5,632

Test	Cause	Fix
mcp_smoke::revision_read_survives_removed_parent (Windows)	The new test used /outside-every-root/lib.rs, which is not absolute on Windows; the product's path_outside_root there is correct.	b4f7b0b4: the path comes from a second temp directory.
store_repair_e2e::a_running_daemon_stops_serving_a_publisher_whose_key_was_withdrawn (Windows)	Established defect, also in the last release: ProjectExtractors::claims re-acquired the slot read lock it held; with an activation writer queued in between, path_claims never returns. That it caused this run's failure is a hypothesis (128 s fits the 120 s RPC timeout).	0a658f8d: claims answers from one snapshot; Linux test fails without it; the e2e now fails on the first RPC error and prints the daemon's exit record and log tail.
argument_reference_e2e::frozen_legacy_acceptance_matches_the_installed_baseline (Windows)	Round 2 named the call: the 400,000-character whole_word literal, server alive and silent for 30 s. Stack samples on Linux with longer literals put the time in SQLite's fts5ParseTokenize, which reallocates the phrase every eight tokens; the Windows allocator copies the block each time. Pre-existing in the product (#575). My round-1 attribution to the regex compile was wrong; that change is reverted (f1cfa9db).	b28850b0: the case uses the shortest literal the regex engine refuses (40,000 x ι, limit 31,208), same refusal body as the installed release, a tenth of the FTS tokens. 4ef10253: the replay names the case and prints server stderr on a stall.
query_cli::two_one_shot_queries_at_once_start_one_daemon (Linux)	Pre-existing race (reproduced on the last release by forced interleaving; 0 of 60 natural local runs): a one-shot reader beside a live daemon reported its read-only snapshot as ready and answered symbol_not_found over a reconciling index.	ff266d0a: that snapshot stays reconciling, so the existing handoff qualifies the answer.

After these: code-index-cli, code-index-daemon and code-index-mcp suites green on both legs locally, fmt, clippy (host and x86_64-pc-windows-gnu) exit 0. CI round 2 (ff266d0a): Windows 5,406 of 5,407 (only the replay above); the store-repair and path tests pass. CI round 3 (f1cfa9db, the final head): all four workflows green; Linux 5,634 passed, Windows 5,406 passed. Windows behaviour is verified by CI only.

Not in this train

• #545 guest side (string clamp, depth diagnostic, member-chain re-walk, lossy decode) and the requested bytes of a memory-ceiling refusal.
• #570: output filters on eleven tool cells, aliases, limit on the diff tools, the registry for configuration keys, environment variables and CLI flags.
• 33 known wrong binds this change does not touch (#566).
• #571 (agent adoption) waits for operator decisions.
• #575: the cost of a very long search_text literal (found in CI round 2, pre-existing).

🤖 Generated with Claude Code

https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

Release train for v0.41.50. Seven lanes, each reviewed from its diff and re-run by the lead, integrated on `integration-1111`. ## What changes for a user - **Resolution (#566).** A qualified reference is not bound across files on name evidence when the file's imports or module declarations contradict the candidate; where module evidence names one candidate, it is bound. Against the last release on the nine pinned repositories: 19 binds removed (all wrong before), 228 added (every candidate addition audited; a blind second audit of 65 rows agrees), 1 retargeted, 133 rule-retagged. - **MCP arguments (#570).** One policy per argument drives the published schema, the runtime check and the error. `invalid_argument_value` replaces ten older codes; the four malformed-argument codes share one field set. Every other error body is unchanged and nineteen are compared with bodies recorded from the last release. Reference: `docs/mcp-arguments.md`, checked against the schema both ways. - **Incomplete extraction (#545, host side).** One census behind `project_overview`, `resolution_gaps`, `index_coverage`, `list_files` and `evidence_gaps`; host refusals are counted; non-UTF-8 source files are recorded as `content_refused`. - **Small fixes.** `read_code` with `rev` for a path whose directory is gone (#568); `application/json` on every JSON resource (#569); `code-index query` on an invalid configuration prints the refusal and exits 1; `search_text` on a permanently excluded literal path says so (#521); installer check reports a `code-index` elsewhere on `PATH`, and the release notes' build line names the run's tag (#572). ## Behaviour changes to call out - Scripts matching `invalid_max_depth`, `invalid_token_budget`, `invalid_direction`, `invalid_role_filter`, `invalid_ref_kind`, `invalid_resolution`, `invalid_response_format`, `empty_seeds`, `too_many_seeds` or `invalid_arguments` need `invalid_argument_value`. - `code-index query` with an invalid `.code-index.toml`: exit 1 with a JSON refusal (was exit 2, no output). ## Evidence - Content, cost and stage baselines are blessed in one commit; each reason carries the bind account or the attribution. Content movement (rust-ripgrep +18, ts-zod +13, js-express +1, php-guzzle −2) equals the audited account. - Cost against the last release: VM steps +0.7 % to +1.7 % on seven repositories, −14.0 % on ts-zod, −2.5 % on rust-analyzer; full scans +0.7 % to +2.0 %. Two costs were attributed per statement and removed before the bless (five module-evidence statements; one capability check that flipped a plan). The nine bind accounts are byte-identical across every reduction. - Every package-supplied input of the new rules passes a capability grant; the pool, ref-kind and visibility registries are updated and a test shows an import without `reachability_anchor` cannot disambiguate a pool. - `tools/list` 59,167 bytes (ceiling 59,311), skill resource 3,949 tokens (ceiling 4,000); no ceiling was raised. ## Qualification One full gate run on `05ea92f9` (empty package store), 43 steps: 40 green at once, three red, each attributed, fixed in its own commit and its step re-run green. | Step | Cause | Fix | |---|---|---| | `agent_bench` | A change in this train (713b1423) emptied `change_impact`'s `test_role_files` at `limit: 0`; the benchmark's which-tests questions use that call. Recall 1.000 → 0.974 (rust-ripgrep), 0.970 (python-flask). | `3339b685` restores the release behaviour and keeps `witness_paths_total`. Mutation run red, restored green. Recall 38/38 and 33/33. | | `tier3` | Record not re-recorded with the other baselines. Drift equals the audited account: rust-analyzer +179 resolved (194 added, 15 removed), 120 retagged; py-django 6 retagged. | `7db6ebbe` bless with the account as reason; green without the bless flag. | | `pkg_javascript_full` | Declared `unbound-by-scope` 42, measured 41 on js-express. One site left the set: `test/app.request.js:12:19`, a write the builtin bound to an unrelated local; the host now refuses it, the package never bound it. Other three declared counts unchanged. | `f4d07dbb` declares 41. The file is a test declaration, not in the package or its digest. Full JavaScript leg green. | After the fixes: `code-index-mcp` suite 1,082 passed on each leg (snapshot and daemon), fast gates 519 passed, fmt, clippy and docs exit 0. The whole workspace was not re-run after the three commits; they touch one MCP source file, one MCP test, one test declaration and one record. ### CI round 1 (`f4d07dbb`): Windows 3 failed of 5,404, Linux 1 failed of 5,632 | Test | Cause | Fix | |---|---|---| | `mcp_smoke::revision_read_survives_removed_parent` (Windows) | The new test used `/outside-every-root/lib.rs`, which is not absolute on Windows; the product's `path_outside_root` there is correct. | `b4f7b0b4`: the path comes from a second temp directory. | | `store_repair_e2e::a_running_daemon_stops_serving_a_publisher_whose_key_was_withdrawn` (Windows) | Established defect, also in the last release: `ProjectExtractors::claims` re-acquired the slot read lock it held; with an activation writer queued in between, `path_claims` never returns. That it caused this run's failure is a hypothesis (128 s fits the 120 s RPC timeout). | `0a658f8d`: `claims` answers from one snapshot; Linux test fails without it; the e2e now fails on the first RPC error and prints the daemon's exit record and log tail. | | `argument_reference_e2e::frozen_legacy_acceptance_matches_the_installed_baseline` (Windows) | Round 2 named the call: the 400,000-character `whole_word` literal, server alive and silent for 30 s. Stack samples on Linux with longer literals put the time in SQLite's `fts5ParseTokenize`, which reallocates the phrase every eight tokens; the Windows allocator copies the block each time. Pre-existing in the product (#575). My round-1 attribution to the regex compile was wrong; that change is reverted (`f1cfa9db`). | `b28850b0`: the case uses the shortest literal the regex engine refuses (40,000 x `ι`, limit 31,208), same refusal body as the installed release, a tenth of the FTS tokens. `4ef10253`: the replay names the case and prints server stderr on a stall. | | `query_cli::two_one_shot_queries_at_once_start_one_daemon` (Linux) | Pre-existing race (reproduced on the last release by forced interleaving; 0 of 60 natural local runs): a one-shot reader beside a live daemon reported its read-only snapshot as `ready` and answered `symbol_not_found` over a reconciling index. | `ff266d0a`: that snapshot stays `reconciling`, so the existing handoff qualifies the answer. | After these: `code-index-cli`, `code-index-daemon` and `code-index-mcp` suites green on both legs locally, fmt, clippy (host and `x86_64-pc-windows-gnu`) exit 0. CI round 2 (`ff266d0a`): Windows 5,406 of 5,407 (only the replay above); the store-repair and path tests pass. CI round 3 (`f1cfa9db`, the final head): all four workflows green; Linux 5,634 passed, Windows 5,406 passed. Windows behaviour is verified by CI only. ## Not in this train - #545 guest side (string clamp, depth diagnostic, member-chain re-walk, lossy decode) and the requested bytes of a memory-ceiling refusal. - #570: output filters on eleven tool cells, aliases, `limit` on the diff tools, the registry for configuration keys, environment variables and CLI flags. - 33 known wrong binds this change does not touch (#566). - #571 (agent adoption) waits for operator decisions. - #575: the cost of a very long `search_text` literal (found in CI round 2, pre-existing). 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

buildagent added 129 commits 2026-11-12 05:46:24 +01:00

fix: require visible targets and module evidence for cross-file qualifiers ... d58559cfcb

Checkpoint recovered #566 work. Workspace all-target check, indexer check and all 13 file_isolation_566 tests pass. Corpus accounting and remaining verification pending.

perf: share script routing snapshots across distinct importer targets ... c7ae37ac22

Keep unresolved disclosure target-specific and deduplicate aliases to the same unresolved import. Revision and epoch changes invalidate the routing snapshot. Indexer check and 16 script-import correctness tests pass.

fix: retain exact script import paths as module evidence ... 389e03bc55

Preserve named imports whose raw path resolves to the candidate module. Add an exact-target control with a same-stem decoy. Update the reverse-dependency fixture schema for the script-language existence probe without changing its assertions. Indexer check, new fixture and both reverse-dependency tests pass.

fix: follow explicit public extern-crate module aliases ... e12ef77a4c

Include Rust library hubs when assembling module proofs and gate their explicit extern-crate aliases as reachability anchors. This retains real cross-crate re-exports without file-stem evidence. Exact-target extern-crate fixture, indexer check and all 17 pool registry tests pass.

fix: distinguish Rust type paths from value receiver shadowing ... ce8aeab932

A same-spelled local value does not shadow a Rust type path, including constructor type heads. Mixed value/type groups stay conservative. Strengthen the positive control to require three exact remote type targets; its shadowing mutation goes red. Indexer check and the strengthened fixture pass.

fix: retain qualifier binds unless indexed facts contradict them 8da9c70d38

fix: veto lexical module shadows and defining assignments per site 008cef46e4

test: restore original module fixtures for permissive qualifier fallback 8a986a9836

fix: preserve instance members and release candidate visibility pools aa45260d27

fix: keep foreign C# members when receiver ancestry is unproven d13bfbc792

fix: retain structural module evidence before permissive fallback d09f27a70d

fix: reject PHP imports contradicted by declared namespaces 9db81e1b6a

test: distinguish module proofs from permissive receiver fallback 8393338f92

test: restore population record with the reverted Rust fixture line 663dd690bf

refactor: centralize adopted module and script dialect policy 82a45f5af0

refactor: read module policy through the core profile descriptor 99c7d60199

perf: restrict module proofs to ambiguous pools and applicable refusal sites f106ad8af4

fix: include module import reads in the generation audit 64774bbafa

Freeze script routing fixtures and repair resolver integration oracles 37273bd69c

Preserve unknown visibility and grade only positive scope exclusions 7ec441ab5e

Merge released resolver and fixture changes from a7ce9603 481ddbe9aa

refactor(mcp): distinguish rejecting and applied argument ranges ... 8757cf2fca

Stage 1. Targeted strict_and_clamped_ranges_have_different_contracts passed. Ran clamp-to-reject mutation: same test failed at the applied-range assertion (exit 101); restored snapshot and verified md5.

refactor(mcp): centralize strict numeric argument policies ... 0b4e238abf

Stage 2 strict ranges. Published-router parity test passed. Ran schema-only depth maximum mutation (20 to 21); that test failed (exit 101), with runtime constants unchanged. Restored snapshot and verified md5.

fix: read historical paths after their parent directory is removed (#568) 732b1cc82c

fix: derive release build metadata from the selected core tag (#572) debfff5ef1

fix: report external PATH installs during installer checks (#572) 1c2a01aabe

fix: share resource MIME types between listings and content (#569) 395b880828

fix: emit shared invalid-configuration refusal for CLI queries (#572) babca49d4d

fix: disclose coverage for literal excluded text-search paths (#521) 43a3bfb0f6

fix: omit impact path lists from count-only replies (#557) 713b142399

refactor(mcp): centralize batch cardinality policies ... 53527b9f71

Stage 2 batch caps. Published-router cardinality test passed. Ran schema-only maxItems +1 mutation; same test failed (exit 101), runtime caps unchanged. Restored snapshot and verified md5.

feat(mcp): publish applied ranges from shared argument policy ... f14f1e25dd

Stage 2 overlay/clamps. Published-router clamp test passed. Ran mutation adding rejecting maxima to clamps; same test failed at repo_map.token_budget (exit 101). Restored snapshot and verified md5. No runtime domain changed.

docs: measure hostile host extraction disclosures before unifying them 41b6b6c626

feat(mcp): share format vocabularies between schema and handlers ... 23c5b17ec1

Stage 5 formats. Legacy spelling test passed for null omission, empty/Unicode whitespace, mixed case, and distinct read formats. Ran trim-removal mutation: same test failed (exit 101); restored snapshot and verified md5. Publish canonical enum plus null and static effective defaults.

feat(mcp): share canonical envelope vocabulary without static default ... b0d68877de

Stage 5 envelope. Legacy test passed for existing aliases, whitespace and case-sensitive rejection. Ran trim-removal mutation: same test failed (exit 101); restored snapshot and verified md5. Environment-dependent defaults remain unpublished.

feat(mcp): share exact direction and role vocabularies ... b4928932f5

Stage 5 exact vocabularies. Legacy/schema test passed. Ran mutation replacing out with outgoing; same test failed (exit 101). Restored snapshot and verified md5. Existing exact-case and whitespace rejection preserved; nullable role stays nullable.

feat(mcp): publish shared nullable filter vocabularies ... 523eeda7b9

Stage 5 optional vocabularies. Resolution/category/ref-kind schema and legacy resolution spelling test passed. Ran trim-removal mutation with verified snapshot: same test failed (exit 101); restored snapshot and verified md5. Category runtime and schema use one daemon constant.

fix: share the incomplete-source census across host disclosure readers 148d3cc884

docs(mcp): replace redundant vocabulary prose with schema metadata ... d8aaea5f08

Stage 5 description budget. Policy unit tests and clamp transport test passed. Ran schema maximum mutation: strict parity test failed; ran applied-range removal mutation: updated transport test failed (exit 101). Restored verified snapshots. Retain concise projection behavior in descriptions.

fix: route absolute historical reads before lexical validation (#568) 992d40c01f

fix: return refusal exit status for invalid query configuration (#572) 575e405192

fix: share configuration diagnostics without deserializing ToolError (#572) 69508528f2

docs: name excluded literal paths in refused zero vocabulary (#521) 53621b676d

docs: confirm internal impact consumers bypass count-only projection (#557) 043aa0bc8f

Preserve dependency identity across shared Rust source ... 3a77f1e026

Retain the manifest-only closure and compiling roots alongside permissive layout reachability. Refuse imports, qualified type paths and their members when a complete dependency library contradicts the foreign target; preserve declared and shared-root access. The alias regression fails with the refusal disabled. Manifest layering, Cargo scope, file isolation, scoped equivalence, pool registry, workspace check and warning-denied documentation pass.

test: grade hostile extraction reasons across every disclosure surface 4adc55acb9

test(mcp): verify page caps through applied-range metadata ... 31b3c0b6bf

Stage 5 follow-up. Full graph target exposed the old prose-only assertion (Option unwrap at line 501) on both legs. Updated assertion test passed. Ran applied-range removal mutation: this test failed at its own metadata assertion (exit 101). Restored snapshot and verified md5.

chore: remove worker notes from the repository ... 690dc81fcc

Notes belong in the lane's scratch directory; the entries were appended there.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

Record the nine-repository account against the release ... ac4bd51fb2

Frozen release-flags candidate: 19 removals, 229 additions, one retarget and 133 rule retags; net 210. File bytes, imports, producer identities and grants match for all nine pinned heads. Named-import rule 141 has zero corpus facts and targets on both sides. Detailed positional/rule tables and immutable-binary provenance are in /tmp/claude-1000/codex-r20/bind-account.md and account/.

refactor(mcp): consolidate argument policy lookup ... 28a941bd3c

One Policy lookup and schema_fragment replace seven semantic lookup helpers. Preserve the published tools/list byte-for-byte (59,243 bytes before and after; cmp exit 0).

Validation: arguments::tests 5 passed. RUN mutation changed direction default both to out: exact_vocabularies_preserve_direction_and_role_acceptance failed with the expected default mismatch (exit 101); restored from md5-verified cp snapshot. No behavior change.

Merge codex-r19: historical reads under removed directories, JSON resource MIME types, structured CLI refusal for invalid configuration, release build line, installer check reports PATH installs, excluded literal path disclosure, count-only impact without path lists ... cf160f38c1

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

test: prepare hostile package fixtures before MCP warm-up 116d6c9a3a

feat(mcp): complete flat argument errors at the transport boundary ... 79ed705ad9

D3 substep: add tool, received, accepted and static default to name/presence/type refusals. Wrap rmcp-marked residual decode tool errors in the same flat shape, with no query guess and the decoder message as hint. No redundant typed-decode pass: rmcp already identifies these with its fixed deserialize prefix.

Targeted unknown_args_e2e flat/residual test passed. RUN mutation bypassed boundary completion: this test failed on missing tool (exit 101). Restored from md5-verified cp snapshot. Value-code migration and policy validation follow separately.

Record source audits of release bind movement ... 4b5f4c1b35

193 distinct audited rows: 143 correct additions; 19 wrong-before removals; one wrong-before/correct-after retarget; 28 correct and two wrong unchanged controls. All required integer-stride selections and all small cells are covered. All 31 inherited wrong targets remain bound. Four historical LineIndex mistakes remain refused generically on both sides. Evidence: /tmp/claude-1000/codex-r20/audit-rows.tsv and companion JSON files.

Record the completed delta correction audit ... 5d0a4309cf

The release-flags rerun has no audited wrong addition or lost correct bind, so no further rule change is justified by these delta audits. The earlier Ruby-only population mismatch was discarded and all nine repositories remeasured. Known unchanged wrong targets remain a release blocker; nothing is blessed or pushed.

feat(mcp): unify malformed value refusal codes on the wire ... fd0b62bb17

D3 substep: translate the 22 malformed-value codes at the common transport boundary, name the actual argument, and retain domain refusals. Give plugin selector refusals explicit argument names. Update wire assertions; mixed read-entry refusals remain for stage 8.

Targeted malformed_values_use_one_code_and_name_the_argument passes (18 cases plus a preserved domain error). RUN mutation removed invalid_resolution translation: this test failed with invalid_resolution instead of invalid_argument_value (exit 101). Restored from md5-verified cp snapshot. Full release inventory: /tmp/claude-1000/codex-r17/error-codes.md.

fix: cache incomplete-source census by persistent read snapshot ff5c7e7c8c

fix: reuse census rows for the overview error sample f2f1339759

feat(mcp): validate integer representations and semantic bounds ... d2d9f7178d

D3 substep: use the common policy for rejecting ranges and cardinalities, while preserving clamped acceptance. Check integer representation from schemars formats before decoding, preserving name/presence/type ordering. Update the human error contract and canonical envelope documentation.

unknown_args_e2e: 16 passed, including clamped 0/65/u32::MAX. RUN mutation bypassed the uint representation guard: integer_representation_and_policy_bounds_name_the_field failed because negative limit reached the residual decoder and lost its argument name (exit 101); restored by md5-verified cp snapshot. Preserve plugin confirmation isolation while allowing D3 received in its refusal body.

Record released-to-candidate SQLite cost attribution ... e1ac3be7e0

All nine repositories exceed one percent VM growth against a7ce9603. Exact signed deltas for 1,562 normalized statements reconcile every measured dimension; full SQL, source locations and cheaper statement shapes are recorded in /tmp/claude-1000/codex-r20/cost-comparison.md and cost-statement-deltas.tsv. The cost gate remains red for ripgrep scans and zod work; the structural ratchet remains red for the audited movement. No baseline is blessed. Diagnostic printer changes were restored with cp and matching MD5 receipts.

feat(mcp): validate vocabulary values with the existing parsers ... f571b18bb2

Complete stage 6: each vocabulary Policy carries its legacy acceptance predicate. Schema values stay canonical while runtime validation preserves the existing aliases, whitespace and case rules. Bad category values now receive the same flat value error instead of a raw backend error.

Targeted vocabulary transport test passed, including normalized formats/resolution, blank filters, null and envelope aliases. RUN mutation replaced row-format parsing with exact canonical matching: this test rejected the formerly accepted CONCISE spelling and failed (exit 101); restored by md5-verified cp snapshot.

feat(mcp): complete argument errors within partial read batches ... a656ccae00

Stage 8: complete each malformed read entry from the published item schema, preserve domain refusals and successful siblings, and leave text content items untouched.

Mixed read-batch transport test passed for json and text. RUN mutation eagerly rejected boolean entries at the batch gate: this test failed because valid siblings were refused too (exit 101). Restored by md5-verified cp snapshot.

feat(mcp): disclose applied snippet widths from argument policy ... 594023bba0

Stage 9 adapter: report arguments_applied.context_tokens only when the accepted request is clamped. Refusals carry no applied-value metadata, and existing limit/line/budget disclosures keep their shapes.

Targeted 0/1/64/65/u32::MAX snippet-width transport test passed. RUN mutation omitted context_tokens completion: this test failed on its missing requested/applied object (exit 101); restored from md5-verified cp snapshot.

fix: classify file refusals through the shared coverage vocabulary 7fe85d586a

docs: pin the measured parsed-code scope of UTF-8 refusal rows ... 7e76f8a25d

Review point 4: the required release corpus_ratchet run passed all 9 tests without any baseline drift; no per-repository changed-file list exists. No baseline was blessed. The run_task branch receives extractor-claimed ParseTasks only; search_text-only metadata follows run_text_only_task and still skips undecodable bytes. Runtime artifacts in /tmp/claude-1000/codex-r18/populations-r2.json and populations-search-r2.json show 14 files, 8 parse_errors, 9 incomplete files; utf8.rs is inventoried with content_refused and per-path evidence, while undecodable sql/csv/txt have no rows. po is unclaimed without extra_text_extensions. No behavioral decision or worker protocol change.

fix: preserve measured symbol floors for partial extraction 438cf24dd4

test(mcp): cross-check concurrent applied-value disclosures ... 221f7de23b

Complete stage 9: snippet completion must preserve limit_clamped, lines_widening and the repo_map token_budget field. Classify snippet bounds by their runtime applied-value disclosure.

Targeted combined-disclosure transport test passed. RUN mutation wrote snippet metadata into limit_clamped: this test failed on the changed limit_clamped shape (exit 101). Restored from md5-verified cp snapshot.

docs: state the shared extraction contract in operations ... 26d51aba0a

Move historical measurements to /tmp/claude-1000/codex-r18/host-extraction-census-before-after.md for the issue comment. Name the clean-leg assertion requiring verdict normalization, and exercise a real subsequent UTF-8 refusal without clearing it. Its classifier mutation fails at the intended diagnostic assertion. Observed memory bytes remain open; the worker/supervisor protocol is unchanged.

fix(mcp): publish effective static defaults and omit routed defaults ... 8cba9cb3d2

D7/reference substep: publish depth 20, map budget 1500 and the daemon match-line default; omit project defaults whose effective routing depends on the call. Preserve explicit null and omission behavior. Share the map default with its handler.

Targeted defaults transport test passed, including runtime omission controls and refusal default detail. RUN mutation retained the inherited project default: this test failed on the routing/session default assertion (exit 101). Restored from md5-verified cp snapshot.

docs(mcp): check argument reference against the published schema ... 478e7c53fd

R substep: hand-maintained defaults, bounds and canonical vocabularies for the first 15 tools (98 arguments), checked against tools/list in both directions. No pass-through registry rows.

Targeted argument_reference_e2e passed. RUN mutation changed only documented max_depth default 20 to 99: this test failed on that contract cell (exit 101); restored from md5-verified cp snapshot. Remaining tools and family/exemption tables follow in the next small commit.

Skip Cargo route preparation without crate-dialect files 2b76e07605

docs(mcp): complete the argument and family reference ... a0da9fe0c6

Complete reduced stages 3+4: all 142 published arguments, exact schema/reference names and contract cells in both directions, plus 23 tool families and explicit exemptions. Record exactly eleven deferred output-filter cells as planned, #570. Pending response_format/limit cells are removed as their implementation lands.

Full argument_reference_e2e passed. RUN mutation added a defaulted GetSymbolArgs field without a reference row: this test failed at the reference-name equality assertion (exit 101). Restored from md5-verified cp snapshot.

Skip extern-route scan without usable Rust roots 895a9ffeb6

feat(mcp): add a substantive concise file inventory format ... 1dd27a184a

Stage 10/list_files: keep indexed-file identity, live size/count fields and paging; omit the potentially wide unindexed-file sample, retaining its measured count and scan scope. Default detailed behavior is unchanged. Keep the reference and exemption table synchronized.

Formats and reference transport tests passed. RUN mutation ignored concise projection: list_files_concise_omits_the_wide_sample_and_preserves_inventory failed on the retained sample (exit 101). Restored from md5-verified cp snapshot.

perf: collect sparse census candidates before grouping files ... 33d6942ed6

Keep the scoped path seek and revision cache. Read diagnostic, refusal and unclaimed legacy-error candidates once, then aggregate only their file IDs.

Compare totals and per-path codes to the grouped oracle on mixed generations and the hostile registry. Dropping the refusal arm fails the equivalence test (5 versus 6). Release cold 20-run medians: 25.21, 27.02, 28.88 ms for 200,000 contributions.

feat(mcp): add concise dependency rows without dropping edge identity ... 2f869091db

Stage 10/get_dependencies: concise omits optional aliases while retaining file/module/line identity and both side counts, cursors and uncertainty disclosures. Default detailed behavior remains. Update the checked reference and exemption table.

Dependency-format and reference tests passed. RUN mutation ignored response_format: dependencies_concise_omits_wide_aliases_and_preserves_edges failed on the retained alias (exit 101). Restored from md5-verified cp snapshot.

Merge codex-r18: one incomplete-extraction census behind every disclosure surface, refusals counted beside diagnostics, partial files keep a floor count ... ea3ba8283d

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

test: account for three new suites in the settle barrier population ... a2237f0bd5

resource_mime_e2e, literal_path_coverage_e2e and hostile_extraction_e2e use
the shared initialize on both legs.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

Bound module ambiguity probes to used groups and index route pairs 98175871d7

feat(mcp): add concise resolution diagnostics with complete aggregates ... 53c6e87a4b

Complete stage 10: omit per-name example paths/lines in concise resolution_gaps, keeping identity, reasons, counts, scope and degradation. All three new formats preserve their existing detailed defaults and update the checked reference/exemptions.

All three format tests and the full reference test passed. RUN mutation ignored response_format: gaps_concise_omits_example_locations_and_preserves_diagnostics failed on retained example locations (exit 101). Restored from md5-verified cp snapshot.

Restrict contradiction sites by module owner, kind and role e377d84d0f

Cache namespace-route membership by its complete proof key d7f2800de4

test: bound accepted hostile controls and minimize fuel fixture a21ec19fe8

test: drop accepted controls from hostile refusal registry 8f08f0bd28

Cache Rust type-namespace proofs per file, name and written qualifier 15b906bd73

fix(mcp): explain diff page bounds when limit is refused ... 5be137dcad

D4 permitted fallback: keep limit exempt. DiffCursor stores a whole-file offset and both collector paths roll files and tallies back atomically; a strict changed-symbol row maximum needs intra-file continuation, cursor compatibility and collector/rendering changes beyond the 120-line allowance. A smaller symbol_budget alone can still exceed the requested maximum on the first file.

Only unknown limit on changed_symbols/review_diff gains a hint naming fixed response token bounds, whole-file paging and cursor/next_cursor. Update both exemptions with this reason. Targeted hint and reference tests passed. RUN mutation suppressed the special hint: the hint test failed (exit 101); restored from md5-verified cp snapshot.

test(mcp): gate every family capability and explicit exemption ... 31d7b07653

Stage 12/families: compare every family capability with the published schema, requiring the exact missing-cell exemption set. Pin the eleven planned #570 filter cells and reject pending decisions.

Reference and family gates passed. RUN mutation removed the list_files/lang exemption: family_capabilities_have_exact_named_exemptions failed on the missing-cell set (exit 101). Restored from md5-verified cp snapshot.

test: minimize refusal registry without losing reasons or languages 8fca883533

Reject unsupported contradiction languages before site joins 13ff49e065

test: pin the first refused PHP signature size c220f269ce

Cache local qualifier shadows and skip inapplicable module dialects 99c8f45c6c

test(mcp): freeze legacy acceptance independently of argument policy ... 54826d3414

Stage 12/legacy: 38 cases observed on the installed binary, whose SHA-256 is recorded in the fixture. Pin normalized spellings, aliases, nullable defaults, integer boundaries, applied caps, batching and metadata bypass; preserve partial read acceptance. The working-tree gate does not derive expectations from Policy.

Installed baseline probe: 38 cases, zero mismatches. Working-tree legacy test passed. RUN mutation turned applied caps into rejecting maxima: the legacy test failed on the formerly accepted usize maximum (exit 101). Restored from md5-verified cp snapshot.

Restrict anchor candidates to names used by each qualifier ed96b27d21

test(mcp): enforce the release tools-list byte ceiling ... d7fa51281e

Complete stage 12: tools/list of the built server must fit 59,311 compact UTF-8 bytes. Reclaim only description prose now represented by schema vocabulary, defaults and ranges, keeping behavioral guidance.

Native size gate passed at 59,021 bytes. RUN mutation padded an argument description: this same gate failed above the ceiling (exit 101). Restored from md5-verified cp snapshot. Final cross-leg and workspace verification follows.

Restrict package-tail candidate inserts to used qualifier names 504cf29df9

Skip dotted layout identities without live dotted contributions d87b0165e2

test: serialize hostile registry against host deadlines 55aa25cd0b

Skip native dotted initializer reads without dotted contributions 3e22812ad1

Merge codex-r21: hostile extraction registry reduced to six fixtures and run serially against the host deadline ... dd7a08e10d

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

fix(mcp): preserve actionable value errors across query batches ... 1117a5ee5a

Verification fixes: derive range/vocabulary/cardinality hints from the same looked-up policy, complete malformed query-batch entries without changing successful siblings, and use received for the original whole-word query. Register the shared map default and keep skill guidance inside its unchanged resource budget.

Targeted contract tests, format/range hints, the skill budget and all seven resource-grading tests passed. New query-batch transport test passed. RUN mutation bypassed query-batch completion: that test failed on the old whole_word_expression_unsupported code (exit 101). Restored from md5-verified cp snapshot. Initial verification reds were introduced by this train and corrected here.

fix(mcp): preserve semantic argument refusal codes 9924a80c8a

refactor(mcp): construct value refusals from argument policy 6ff857ae2e

fix(mcp): render argument hints as policy sentences 94b4f50253

fix(mcp): omit empty acceptance and null defaults 35d36d201e

docs(mcp): render human argument reference from shared schema cells b47408661f

Merge codex-r17: one argument policy behind schema, validation, errors and reference; structured value errors; applied-value echo; concise formats on three tools; size, family and legacy-acceptance gates ... d25d2684a9

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

Refuse foreign nominal type anchors in native inline Rust modules 18dfcf0cc4

Clean qualifier name caches between resolver passes 995c268135

Merge codex-r20: file-isolation rules on positive contradictions (#566) ... fdf3e4b94f

Against v0.41.49 on the nine pinned repositories: 19 binds removed, all
audited wrong before; 228 added, all audited correct; 1 retargeted from a
wrong to the correct target; 133 change only their rule tag. A blind second
audit of 65 rows (all removals and the retarget, 45 sampled additions)
agrees on every row. SQLite work per cold index stays within +1.74 % of the
release on every repository (rust-analyzer -2.1 %, ts-zod -13.5 %).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

perf(indexer): probe module groups by file identity e8df7e00f0

perf(indexer): reuse loaded module files for imports 419c1c88e9

perf(indexer): reuse loaded module files for externs b4962dfc38

perf(indexer): key module file population by file and qualifier a8e4b89bb7

fix(mcp): restrict argument completion to malformed arguments 40893733b5

docs(mcp): trim merged skill and tools-list prose budgets 3ccd55092c

fix(mcp): treat current content refusal rows as terminal in diffs c043c62f89

test: align refusal census and schema prose fixtures with decided behavior 95d7f210ee

test(mcp): pin retained semantic refusals to release bodies 7468763ccc

perf(indexer): key qualifier-name projection by file 579c3f99a7

Merge codex-r23: module-evidence statements probe by file key (#566 cost) ... aae42441f2

The isolation rules raised full-scan steps past the cost gate's ceiling on
two repositories (rust-ripgrep +7.7 %, ts-zod +9.6 %; rust-analyzer +10 %).
Per-statement attribution named five statements in the module-evidence
build; they now drive from the file key or reuse the loaded file set
instead of rescanning the group and anchor tables. Full scans end between
+0.7 % and +2.0 % of v0.41.49 on all nine pinned repositories; VM steps,
sorts and automatic indexes are no higher than before the change. The
bind account is byte-identical on all nine repositories after every commit.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

perf(indexer): inline the loaded module file set in id order ... 366322b0dc

The statement text followed the map's iteration order; sorted, it is
the same text for the same index.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

fix(indexer): admit module evidence before routes and refusals ... 5f5e910d74

Gate producer identities before aggregation, lexical and namespace inputs, and Cargo import propagation. Register the derived relations and ref readers, remove the obsolete visibility exception, and prove the import admission with an ungranted/granted capability fixture.

fix(mcp): preserve kept release refusal bodies ... b7c850bc53

Remove argument-name queries from rename, malformed read-batch entries, and cross-project graph refusals at their construction sites. Capture eleven reachable kept codes from the installed release, including a bounded-matcher refusal and linked-project mismatch, and document why the guarded syntax mapping has no ordinary call trigger.

Merge codex-r22: merge repair, release-identical error bodies, admitted module evidence ... e986cc28ae

Repairs the 18 failures of the interim run on the integrated tree and
the seven registry and contract tests behind them:

- The flat argument fields apply to the four malformed-argument codes
  only. Every other error body is the release's: the fixture now holds a
  release-recorded body for eleven kept codes and eight domain errors.
- changed_symbols and review_diff report a current content_refused row
  as terminal not_indexed.
- Skill resource 3,949 tokens, tools/list 59,206 bytes; no ceiling moved.
- Every producer-supplied input of the module-evidence rules passes a
  capability admission before it can route or refuse a bind; a test shows
  an import without reachability_anchor cannot disambiguate a pool.
  All nine corpus bind accounts are byte-identical with the admissions.

Merge-time change: the contradiction-site admission is a trailing probe
(probe_join). Joined freely it drove the statement from the grant and
cost 26x the VM steps on rust-ripgrep and 90x on cs-dapper.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

test(corpus): bless content, cost and stage baselines for the isolation rules (#566) ... 68dc0d3fcf

Content: rust-ripgrep +18, ts-zod +13, js-express +1, php-guzzle -2
resolved, equal to the audited bind account against v0.41.49 (19 removed,
228 added, 1 retargeted, 133 rule-retagged on the nine pinned
repositories).
Cost: ts-zod -14.0 % VM steps; the other tier-1 repositories +0.7 % to
+1.7 % VM steps and +0.8 % to +2.0 % full-scan steps, after per-statement
attribution and two reductions.
Stage: tier1q_pass1, tier1q_tiebreak and tier3_import_boost move by the
account's per-rule numbers.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

chore: v0.41.50 ... 05ea92f9e1

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

fix(mcp): a count-only change_impact reply keeps test_role_files ... 3339b685c8

`limit: 0` cleared `test_role_files` since 713b1423 (this train). The
list is the file-level run-set, and `limit: 0` with `response_format:
"concise"` is the cheapest call that names the tests to run: the agent
task benchmark asks its which-tests questions that way and lost one
hand-verified answer on rust-ripgrep and one on python-flask (recall
1.000 -> 0.974 / 0.970). The last release kept the list; so does this.

`witness_paths_total` stays (the count of the list `limit: 0` clears);
`test_role_files_total` is gone with the suppression it described.

Mutation run: clear the list in `apply_count_only_impact` ->
`count_only_keeps_the_run_set_and_counts_the_witness_chains` red
(left 0, right 7); restored by snapshot and md5, green.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

corpus: re-record the tier-3 baseline after the #566 bind account ... 7db6ebbece

The drift equals the audited account against the previous release:
rust-analyzer 15 binds removed and 194 added (count.resolved +179),
120 retagged tier1q_tiebreak -> tier1q_pass1; py-django 6 retagged, no
bind added or removed. No parse., stage. or activation key moved.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

test(javascript): the declared unbound-by-scope count on js-express is 41 ... f4d07dbba4

One site left the set: test/app.request.js:12:19, `querystring` in
`app.request.querystring = function () { … }`. The builtin bound that
write to an unrelated local in lib/request.js (tier1q_pass1); the host
now refuses a foreign qualified bind at a script write (#566), so both
sides leave it unbound and the scope verdict has nothing to refuse.
The package side is unchanged (its account row was unbound before and
after). scope-verdict 3425, declared-local 1392 and rebound-by-local 70
are measured unchanged. The file is a test declaration and is not part
of the packed package or its digest.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

buildagent added 6 commits 2026-11-12 08:13:38 +01:00

test: use a portable absolute path outside revision roots b4f7b0b4af

test: identify legacy replay stalls and capture server stderr 4ef10253bd

fix: release extractor slot before reporting path claims 0a658f8d8f

fix(mcp): refuse an over-long whole_word literal without compiling it ... 3d7463f8a8

A `whole_word` literal past the regex engine's compiled-size limit is
refused `whole_word_needle_unsupported`. The engine found that out by
building the pattern first: 0.24 s for 400,000 characters in a release
build, 1.1 s unoptimised, growing with the request. On the Windows CI
runner (about 18x slower on the rest of the same test) that one reply
missed the harness's 30 s deadline in
`frozen_legacy_acceptance_matches_the_installed_baseline`.

`WordMatcher::NEEDLE_MAX_BYTES` (384 KiB) refuses up front. Measured:
the longest needle the engine compiles is 327,671 bytes (characters
without case variants; `a` stops at 131,068), so no needle that worked
is refused, and the refusal body is unchanged. The replayed case drops
from 1.46 s to 0.46 s locally.

Mutation run: bound = 64 KiB ->
`the_needle_bound_refuses_nothing_the_engine_compiles` red (the engine
compiles a needle of that length); restored by snapshot and md5, green.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

test(daemon): state what the slot-snapshot test protects and its mutation ... 5409aa2fd2

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

fix(mcp): a one-shot reader beside a live daemon keeps startup pending ... ff266d0a66

`code-index query` decides between the daemon and a local `--no-daemon`
server before that server starts. When a daemon appears in between
(two concurrent first queries), the local server opens the index
read-only beside the live owner. That snapshot was built with
`LocalIndex::new`, which reports lifecycle `ready`, so a negative
lookup was answered as a measured absence over an index the owner was
still reconciling: `symbol_not_found` with 201 of 202 files lagging,
where `index_incomplete` was due.

A one-shot reader beside a live owner now keeps the snapshot in the
`reconciling` state, so the existing handoff names the owner and
qualifies negative answers. A long-lived session is unchanged. With a
settled daemon the one-shot call answers through the handoff as before
(probed: positive and negative lookups, with and without --no-daemon).

Pre-existing: a forced interleaving reproduces the stale answer on the
last release. It surfaced once in CI on this branch in
`query_cli::two_one_shot_queries_at_once_start_one_daemon` (0 of 60
local runs).

Mutation run: build the owned one-shot snapshot with `LocalIndex::new`
-> `a_one_shot_owned_snapshot_keeps_startup_pending` red (lifecycle
`ready`); restored by snapshot and md5, green.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

buildagent referenced this pull request 2026-11-12 09:20:15 +01:00

search_text: cost of one query grows quadratically with the literal's length (Windows: a 400,000-character literal blocks the server for more than 30 s) #575

buildagent added 2 commits 2026-11-12 09:44:18 +01:00

test(mcp): replay the needle refusal with the shortest literal the engine refuses ... b28850b0d7

CI round 2 on Windows named the stalled call: legacy case 55, the
400,000-character `whole_word` literal, server alive and silent for
30 s while every other case answered in about 1 ms.

Cause, from stack samples of the server on Linux with longer literals
(6.4 M characters: 23 s): SQLite's `fts5ParseTokenize` reallocates the
phrase every eight tokens. glibc remaps the block in place; the Windows
allocator copies it, 50,000 times up to 16 MB for this literal. The
regex compile was not the cause.

The case now uses 40,000 x `ι`: the engine's size limit is reached at
31,208 of them (four case variants each), so the refusal is the same
and the FTS phrase is ten times shorter. The installed release answers
this literal with the same code, hint and reply size.

The unbounded cost of a long `search_text` query is in every release
and needs a length bound in the argument policy; filed separately.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

Revert "fix(mcp): refuse an over-long whole_word literal without compiling it" ... f1cfa9db17

This reverts commit 3d7463f8. Its premise was wrong: the Windows stall
it was meant to remove is in SQLite's FTS5 phrase construction, not in
the regex compile (see the previous commit). The bound was correct but
is not needed for this release, and a query-length limit in the
argument policy will make it unreachable.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014XLNPtTbPwnHBYfKMb3f1W

buildagent merged commit 3d0a584c15 into master 2026-11-12 11:04:34 +01:00

buildagent referenced this pull request from a commit 2026-11-12 11:04:35 +01:00

Merge PR #574: v0.41.50: module evidence for qualified references, one argument contract, one extraction census

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels   

code-review

  

correctness

  

dos

  

performance

  

security

  

severity/high

  

severity/low

  

severity/medium

  

tech-debt

  

Kind/Breaking

Breaking change that won't be backward compatible

  

Kind/Bug

Something is not working

  

Kind/Documentation

Documentation changes

  

Kind/Enhancement

Improve existing functionality

  

Kind/Feature

New functionality

  

Kind/Security

This is security issue

  

Kind/Testing

Issue or pull request related to testing

  

Priority

Critical

The priority is critical

  

Priority

High

The priority is high

  

Priority

Low

The priority is low

  

Priority

Medium

The priority is medium

  

Reviewed

Confirmed

Issue has been confirmed

  

Reviewed

Duplicate

This issue or pull request already exists

  

Reviewed

Invalid

Invalid issue

  

Reviewed

Won't Fix

This issue won't be fixed

  

Status

Abandoned

Somebody has started to work on this but abandoned work

  

Status

Blocked

Something is blocking this issue or pull request

  

Status

Need More Info

Feedback is required to reproduce issue or to continue work

No labels

code-review

correctness

dos

performance

security

severity/high

severity/low

severity/medium

tech-debt

Kind/Breaking

Kind/Bug

Kind/Documentation

Kind/Enhancement

Kind/Feature

Kind/Security

Kind/Testing

Priority

Critical

Priority

High

Priority

Low

Priority

Medium

Reviewed

Confirmed

Reviewed

Duplicate

Reviewed

Invalid

Reviewed

Won't Fix

Status

Abandoned

Status

Blocked

Status

Need More Info

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

h-dv/code-index!574

No description provided.