Intelligence breadth and depth workstream

Accepted scope: 36 language/format adapters, 36 pinned source packs, 144 unscoped retrieval questions, six workflow families, and deeper cross-file navigation in three successive groups. This is offline product evidence, not an agent study.

Version 0.1.30: Java call evidence preserves literal dollar signs. Four unchanged Gson files at the existing source pin add explicit production and test import bindings, eight navigation checks and a same-package unresolved-call control. Java signature metadata retains varargs names/types/rest flags and method return types, including trailing array dimensions; constructors have no return type. Named C parameters retain source-written pointer/array/callback types and order through native parsing, chunk metadata and cached restart. Four signatures in one unchanged Redis file provide a real source witness. Go grouped inputs, variadic component types/rest flags and written results reach function, receiver method and interface method metadata; blank and unnamed parameters add no binding names. Three headers in one unchanged CLI file provide real Go signature witnesses and cached-restart coverage. A separate Go variadic case preserves rest metadata and named result syntax through restart without parsing. The acquired-grammar gate requires 92 cases, with standalone Java, C and Go grammar file provenance recorded around the native signature checks. These supplemental source witnesses add no retrieval questions or historical compatibility claims.

Go module bindings retain all declared var/const names while excluding local bindings. Grouped named types and aliases keep individual ranges/documentation, and generic or parenthesized receiver methods bind to same-file named types. Imports inside a declared Go standard module resolve through the nearest owning root; nested module configurations retain their target boundaries. Direct and readiness-first startup check parser/indexer identity before persisted symbol reuse. Real source fixtures cover these paths and full cache restarts.

AST signature diffs compare detailed parameter metadata as well as legacy name lists. Refresh also checks the decoded source hash, so a body-only edit updates cached source and caller edges even when size, mtime and signatures stay unchanged. These edits do not create structural signature-change events. An absent source digest does not trigger a content-only refresh; unchanged source keeps its cached AST. Same-source restart coverage and live refresh checks are separate evidence.

C parameter metadata follows declared names in the native syntax tree. When the grammar represents parenthesized syntax such as int (name) as an abstract function type, the parameter stays unnamed. Resolving that ambiguity needs C type-name context; inner type names and callback arguments are not treated as outer parameter names. Declaration attributes and storage classes are excluded from written type annotations. Existing C function-name and return-type discovery limits remain outside this parameter-metadata change.

Current release: attocode-code-intel 0.1.30. See the 0.1.30 release notes for changes, validation and limits.

Version 0.1.29 recognizes Perl .t scripts, which share the canonical and generic Perl parser identity. Three unchanged pinned Carp sources exercise native definition/caller/import witnesses, test candidates, scoped inclusion through refresh/restart and eight distribution-root navigation checks. Its release verified 83 acquired-grammar cases. Generated-directory exclusions and conservative import ownership remain unchanged; syntax and heuristic evidence retain their existing limits. See the 0.1.29 release notes. Version 0.1.28 summarized duplicate compact-search metadata diagnostics while preserving full data.ranking.index, results and excerpts. Version 0.1.27 fixed namespace-cache iteration during concurrent AST hydration. Version 0.1.26 added the shared five-second Rust precision deadline and bounded cancelled startup cleanup; other languages retain two seconds. Version 0.1.25 made OpenAI embeddings explicit with ATTOCODE_EMBEDDING_MODEL=openai; 0.1.24 added the openrouter preset for the opt-in hosted rerank.

Development optional-language depth, 2026-10-29: causal native controls restore omitted OCaml/Perl call candidates, OCaml native open_module extraction and conservative workspace import candidates for OCaml/Perl/Crystal. All 406 controls pass, including three authored workspaces with all eight workflow checks. Perl's existing scoped inclusion and Crystal's configured definitions were already working; the initial incorrect Crystal hypothesis/control remains recorded as a failed first attempt. Explicit source-built Crystal acquisition and all22 depth cases are maintained in optional CI; default384/required397 inventories remain mandatory. The corrections are included in the 0.1.20 target; dependencies and published 0.1.19 assets remain unchanged until the separate release gates complete. Syntax candidates and heuristic tests make no compiler, historical nativecause, relevance or performance claim.

Routine compatibility with native capture, 2026-10-28: one unchanged-engine run completes same-run 373 Linux controls and all 36 packs/288 checks/144 existing operational anchors. OCaml capture retains 1047 native calls, 0 failures and a complete 1000162-byte trace with no loss. The original payload and preflight ZIPs, executed source and independent raw/source/ Git/origin/lifecycle/resource/cleanup admissions are archived. Historical causes remain unproven; consumed questions provide no fresh relevance evidence. Runtime, harness, tests, dependencies and version019 remain unchanged; no release follows.

Routine OCaml native evidence, 2026-10-25: bounded capture and its original proof retain native outcomes before fallback/reset in opt-in fixture workers, including generic parsers and sequential thread handoffs. All 373 controls pass; keep 42 new regressions and 386 required regressions. One separately frozen pinned OCaml worker completes eight existing checks (four passed/four limited), zero retrieval questions, and 1,028 native parses with no capture loss. All native results are trees: the historical failure causes remain unproven. Raw/source/grammar/origin/resource/cleanup proof is archived and temporary wiring retired. Product 0.1.19, dependencies, native policy and Rust tracing remain unchanged; no package is published by this change.

Maintained dispatcher qualification, 2026-10-25: causal online controls replace inferred operator filenames with explicit source-bound handlers. Preflight validates every registration without running trial bodies; invalid synchronous signatures, async/generator entrypoints, missing modules and shadowed origins are rejected. Keep 26 regressions in 331-case maintained preflight and 344 required regressions. All 349 product runtime files/version/dependencies remain published 0.1.19 exact; this harness correction publishes no package or new experiment result. Earlier release counts below remain historical.

Compact search snippet budgets, 2026-10-25: the online correction restores source text only after the complete page is fitted. Four causal boundary cases pass and all 380 online controls preserve delivered hits/text/coordinates, continuations, actual budgets and bounded probe work. Keep 19 new regressions in maintained preflight (305 cases). Temporary control wiring is retired. The change is included in the 0.1.19 target; no ranking, relevance or speed claim follows.

Optional local file context, 2026-10-25: the completed development comparison rejects the candidate after 1,080 pages/12 fresh workers: one distinct visible-file gain per round in one repository misses the two-file/two-repository requirement; six excerpt-loss observations and three ordinary-control snippet changes also fail guards. Latency passes. Original zero-query import abort and explicitly approved corrected launch remain separate. All 24 validation questions are unissued. Candidate/tests/temporary trial wiring are removed; product and version remain published 0.1.18 exact. No publication or agent-quality/speed claim follows.

The following checkpoints preserve earlier release and experiment history.

Recovery readiness, 2026-10-24: the correction republishes ready when a failed snapshot restore is followed by a completed full fallback rebuild. The phase previously stayed skeleton despite complete parsed and reference counts, causing stale readiness and unnecessary hydration scheduling. Nine regressions and 313 online controls preserve reconstruction/checkpoints, partial recovery, minimal/empty initialization and interrupted-write behavior. Temporary causal-control wiring is retired; maintained preflight now has 286 cases. Parsing, native limits, retry and source/cache validation policy remain unchanged. The fix is unreleased on development; it makes no performance or relevance claim.

Single repair index, 2026-10-24: the matched comparison rejects the candidate: median restart 28.622→26.718s, ratio 0.933462; every matched restart faster true, cold ratio 1.011936. All fourteen fresh phases preserve complete navigation and valid routes at full 4,980-file readiness; 351 online controls pass. The rejected candidate bypassed old-index construction only during persisted repair; all parsing, validation, LSP and staged-publication guards remain. The product candidate is removed. Original attempts/source are retained and temporary trial wiring retired.

Unprofiled stages, 2026-10-24: the three-process unchanged-engine diagnostic measures 14.268s restart initialization task: persisted index load 3.750s followed by repair 6.480s, which includes 1.074s encoding. Raw clocks distinguish nested work, thread CPU, hydration and selected background calls; whole lexical CPU remains unmeasured. The source suggests reducing redundant index/object construction as an independently testable next candidate. No speed benefit is proven or adopted. All 298 controls, source/raw/canonical/route/readiness/resource/cleanup checks pass; temporary instrumentation is retired, with all 349 engine files and version unchanged.

Bounded JSON, 2026-10-24: the new unprofiled comparison rejects the serializer candidate: median restart 30.332→30.034s, ratio 0.990178, short of the required ≤0.90; one candidate restart round is slower. Cold ratio 1.001210 passes. All 14 fresh phases have equal navigation data, full 4,980-file readiness and validated physical routes. The distinct prospective 10% / all-round / cold 5% rule is unchanged; the older fingerprint rejection remains. Cache bounds and validation/repair policy stay unchanged. The active study wiring is retired, original evidence retained, and no new release is published.

Startup attribution, 2026-10-24: the single online diagnostic uses the unchanged 0.1.17 engine and admits equal data for all ten cold/restart navigation operations. Snapshot repair and bounded summary re-encoding dominate the initialization profile. Source-bound cache observations record identical before/after hashes despite a repair/save pass. Main, executor, process and parent-admission clocks are separate; raw profiles and all failure controls are retained. The next target is bounded snapshot JSON encoding, followed by a new unprofiled comparison. No product optimization or release is adopted here.

Ordered follow-up, 2026-10-22: the published 0.1.15 wheel and source distribution passed all 18 clean-install/previous-version-upgrade phases on a GitHub runner; all 18 release jobs passed. No global tool or host was restarted. The subsequent unreleased navigation fix makes equal-scoring symbol pages deterministic before their result limit, including after persisted reopening. Match scores and LSP priority remain unchanged. See the restart investigation and ordering correction. Archived evaluation results stay in the tagged repository and are excluded from future source distributions after the new bundles exceeded the unchanged 32 MiB expanded-size verification limit. The online build and install checks verify the source distribution and the wheel built from it before integration.

The contiguous preview trial closed incomplete on a validator decoding error before any candidate page. The matched Cockroach restart trial stopped on differing baseline cold/restart symbol pages before a candidate timing. Both preserve their unsuccessful attempts and exact source evidence. Neither establishes a quality or speed gain; the fingerprint optimization and experiment wiring are omitted. First-page relevance, restart cost and historical Rust/OCaml causes remain open. All runtime validation in this follow-up ran online; no scored trial was replaced or rerun.

Included in 0.1.15: cached native parsers are now owned by each thread. A real-parser regression reproduces a separate timeout-continuation leak that returned a wrong function name and source span for another worker's document. The investigation distinguishes that demonstrated defect from the historical OCaml and Rust causes, which remain unproven. The 0.1.14 artifact does not contain this fix.

The subsequent OCaml/Rust follow-up fixes competing stderr readers in the language-server client. Event-controlled regressions demonstrate the conflict in both reader orders. The existing bounded OCaml recovery also restores the exact historical source after a deliberately forced native interruption. Neither result establishes the original intermittent cause. The stderr correction is included in 0.1.15; production deadlines and recovery policy remain unchanged.

Latest compatibility evidence, 2026-10-22: all 36 pinned packs completed, with 157 passed, 124 limited and seven unsupported checks, and no failed or incomplete checks. Limited and unsupported outcomes are not verified support. The intermittent OCaml cold-parse trigger and Cockroach restart cost remain open; see the final reliability report. The six-language native-source checks pass 12 positive cases and four negative controls, with two dynamic/macro controls limited. The unmeasured restart shortcut is omitted; the single callback-span candidate is rejected on consumed development.

The local lean profile is shipped and opt-in; daily remains the default. The 2026-10-22 investigation pilot stopped incomplete after its first scored pair: malformed structured-output/tool calls failed the frozen protocol. Fourteen scored requests were never launched; there were no retries or replacements. Quality and cost/time comparisons remain unavailable. Beta positioning is unchanged; no agent benefit is established. R7 demonstrated no completion benefit and ended with a stop decision. These product checks do not establish better agent outcomes or authorize further study runs.

Earlier batches and studies

The dated entries below preserve their original checkpoints. Older release, pending-validation and proposal statements describe those checkpoints.

Earlier search batch, 2026-10-20: test cases before recognized support paths. The fixed candidate moved helpers off two first pages but recovered no judged test on 13 consumed Express/Mojo questions: files at five stayed 13, at 24 stayed 17, with no individual losses. The development stop rule applies. No fresh validation was sampled or run, and no product ranking change is retained.

Earlier experiments, 2026-10-19: function-unit BM25 failed its frozen quality rule. CPU function embeddings failed indexing time and memory limits; int8 also failed vector fidelity. Neither result supports a new local search mode.

Earlier search batch, 2026-10-18: source hit after the first five tests. The candidate gained two judged files at 24 on 20 new questions with no loss, but failed the frozen p95 limit in one run. A second trial with a three-round latency rule passed latency, but lost one judged test file and gained none on 20 more new questions. The search code does not change.

Earlier search batch, 2026-10-18: paired source for "test" questions. The candidate gained no judged file on 16 new questions, so the search code does not change.

Previous search batch, 2026-10-17: test intent for "tested" questions. Short "tested" and "testing" searches keep the source order and show the best test hit second. On 16 new frozen questions, judged files at five went from 14 to 16 with no loss. "test" and "tests" searches still miss every implementation label at five.

Previous search batch, 2026-10-17: file-level coverage and "tested" test intent. Both candidates failed their predeclared rules on 32 new questions, so the search code does not change. File-level concept coverage had no net first-five gain on validation. "tested" intent found more test files but pushed out implementation files.

Previous search batch, 2026-10-17: dogfood fixes and one search candidate. Three defects found by using Attocode on this repository are fixed. On 12 fresh validation questions, the fixes keep every file and quote judgment unchanged. The one retrieval candidate failed two predeclared rules and does not ship.

Previous search batch, 2026-10-16: explicit test search and transient precision recovery. One frozen test-intent partition improved labeled hits at five from 3/8 to 6/8, with four unchanged production controls. Transient precision replies retry within the existing deadline, and partial caches expire. Six fresh real Rust fixtures passed after an actual intermittent null-response failure was captured. The bounded gate passed 829 tests (nine skips), 120 repeated navigation checkpoints, seven source packs and frozen ranking in 171.081 seconds; scratch cleanup passed. PostgreSQL and full upstream CI gate merge. No new cloud rental or release.

Previous follow-up, 2026-10-15: document lifecycle and body-file trial. Shared LSP document state prevents duplicate opens and edit versions across callers; strict protocol and six fresh Rust fixtures pass. The intermittent CI failure's cause is not proven. Body-file fusion improved consumed development 33→40/48 at five, then fresh validation 18→20/24, but lost one existing top-24 hit and failed the full gate. Ranking remains unchanged. Two local Cockroach component prototypes were not adopted; no new cloud rental or release occurred.

Previous follow-up, 2026-10-15: Cockroach and fresh retrieval cycle. The merged fixes passed Cockroach at 167.970/180 seconds, still a narrow margin. All 48 new source-backed retrieval questions are complete: baseline 33/48 at five and 44/48 at 24. Neither fixed candidate improved development Hit@5; no ranking policy was added. All 22 workers are deleted, cumulative estimated compute is $0.9264 before tax. The unused conditional retry is not launchable after the passing target. No additional release or model study was performed.

Release status, 2026-10-15: 0.1.6 is released and verified. The full gate passed all 36 pinned packs, 144 retrieval observations, ranking, 591 contract tests (22 skips), lint, scale and lifecycle. The controller finished in 21.368 minutes under its unchanged 90-minute limit. All 11 release jobs and 11 actual-PyPI clean-install/upgrade checks passed; the wheel's 268 Python files match the validated source commit c3624b2.

Cockroach keyword/body indexing took 28.037/68.810 seconds, with a ready index and 19.691-second restart. Measured checks and queries totaled 177.579 seconds; the margin under the unchanged 180-second pack deadline is narrow. Discovery now receives validated cached ASTs before building imports. The earlier full-path no-reparse claim was corrected because its counter omitted discovery parsing.

The full gate distinguishes 127 passed assertions, 124 label limitations and 37 unsupported checks. Retrieval yielded 99/144 labeled hits at five, 128/144 at 24; labels are non-exhaustive development targets, not agent-benefit evidence. No ranking or model default changed.

At the release checkpoint, all 21 workers were deleted; baseline ledger rows and reservations are retained. The three-worker cycle added $0.0898 estimated compute, cumulative $0.9119 before tax (not an invoice). All three approved reservations are used under the $21 cap; the separately approved follow-up is reported above. Only 1,601,899 report/log bytes returned to the Mac. Protected studies/caches, global tools and the running host MCP are unchanged.

Local follow-up: retrieval audit and file-scoped navigation repair. All 144 saved pages are classified; bounded FastAPI/Laravel tracing shows a candidate present below the page cutoff and a relevant file omitted by the original labels. Six fresh questions yielded 3/6 at five, 5/6 at 24 and 6/6 at 96. File symbol lookups now exclude same-named definitions in other files and avoid scanning their global name lists. This repair merged in PR #210 and is not part of published 0.1.6. Search ranking and the narrow Cockroach timing margin remain open.

Second local follow-up: expansion comparison and importance publication. Disabling expansion improved development hits at five from 9/14 to 11/14, but both arms reached 12/12 on fresh validation. The predeclared improvement rule failed, so no ranking policy changed. Profiling exposed a separate publication race: queries could rediscover and reparse the workspace while keyword cache persistence was still running. Publishing importance first removes that extra work; a controlled 128-file overlap retained identical scores and order while reducing query time from 49–55 ms to about 7 ms. This is component evidence, not a new Cockroach timing result. These follow-up repairs merged in PR #210 and do not change the published 0.1.6 artifact.

Completed: intelligence-only repository (PR #222)

Status: Implemented on 2026-10-16 in PR #222. The agent is preserved at the agent-archive-2026-10-16 tag. The text below is the original proposal, kept as history.

The requested next direction is to retire the agent after this validation batch. This is a transition proposal, not authorization to delete code or publish a new package. The standalone attocode-code-intel distribution already owns the engine; the root attocode distribution depends on it. Its clean-wheel check rejects imports of the agent, swarm, TUI and model-client packages. That gives the transition a practical starting point, but does not prove every repository tool is independent. In particular, quality/ranking tooling still uses root workspace configuration and preserved historical harness inputs.

1. Preserve and inventory. Create a final documented agent tag/archive before removal. Map imports, entrypoints, tests, CI, Docker, documentation and release automation. Preserve frozen studies and their grading worktree. Existing intelligence experiments do not establish that the agent itself is worthless.
2. Make intelligence the primary product. Keep the engine, MCP, CLI, API, storage, useful dashboard/deployment surfaces, source adapters and quality harness. Move root documentation, development commands and CI around those supported scenarios. Keep the existing package/import/CLI identity initially to avoid bundling a rename with the migration.
3. Remove agent maintenance deliberately. Archive agent/swarm orchestration, TUI and agent-only examples, model-provider integrations and jobs after their useful intelligence tests have moved. Document migration for existing agent users; do not silently repurpose the attocode package.
4. Verify the smaller repository. Require clean standalone installation, MCP/CLI/API workflows, PostgreSQL, platform/language-server CI, incremental recovery and bounded pinned-corpus checks. Record dependency, checkout, install-size and CI-time deltas before making size/speed claims. Keep unrelated shared caches and historical evaluations untouched.

The intended result is one clearly supported code-intelligence product, with measurable retrieval and navigation quality and a smaller maintenance surface. There is no deletion, rebrand or agent retirement in the current batch.

The entries below are historical evidence.

Historical milestones (0.1.6)

Milestone	Status	Required evidence
Coverage foundation	Validated locally and in CI	Frozen source-backed manifest; honest outcomes; bounded selection, shards and resume; complete baseline
Python, JS/TS, Go, Rust	Implemented; bounded gate passed	Module/alias/re-export and negative controls; real language-server checks
C/C++, Java, C#	Implemented; bounded gate passed	Include/package/namespace resolution, ambiguity and selected-definition identity
Ruby, PHP, Kotlin, Swift, Scala, Elixir, Lua, Bash, Zig	Implemented; bounded gate passed	Explicit module bindings, qualified references and dynamic-case limitations
Protocol and operation coverage	Implemented; bounded gate passed	Catalog-to-test inventory, real MCP/HTTP, profiles, isolation and failure cases
Release and final matrix	0.1.6 released and verified	Coherent full 36-pack gate, required release CI, actual PyPI clean install/upgrade, verified cleanup

The existing six source packs and frozen ranking baseline remain evidence inputs. New labels must be frozen before evaluating product changes. Missing labels, unsupported capabilities and incomplete execution are not successful assertions. All mutations of source packs happen in disposable copies, never cached inputs.

Historical execution bounds

These allowance changes record earlier stages, not current launch authority. All 22 workers were deleted at the latest checkpoint; the last passing target closed its conditional retry. No new rental is authorized by these old caps.

Quick/local remain capped at 120/600 seconds. Managed disk is capped at 5 GiB, process-tree RSS at 4 GiB, with a 2 GiB free-space floor. Broad worker tests have 60 minutes and the controller has 90 minutes. One worker at a time; $1 reservation per batch. The separately approved repair cycle raised the monthly reservation cap from $10 to $14 and allowed four additional batches; all four are now used. The approved structural-write follow-up allows at most two additional batches, raising the monthly reservation cap to $16, with one worker at a time and early stop on success. Reservations are admission limits, not actual charges. Preserve the existing 14 ledger rows; no further increase is authorized by this follow-up. Temporary tools and optional grammars belong on disposable workers or CI.

Preserve shared caches, evaluation artifacts and frozen studies. No model calls, model downloads or Claude usage-page checks; R7's STOP remains in force.

Historical progress evidence

Implementation starts from main ecc2c3e19d37ed6e802e39eece30fdd5692faaa6. The existing symlink fix is merged but not in released 0.1.5. The first bounded local milestone passed; broader evidence remains explicitly partial. Record each completed milestone's report location, candidate hash, corrected examples, limitations, runtime/storage and cleanup here.

Implemented foundation and language work

• Frozen manifest SHA-256: e254cb987f782483d241d93b929d24d2582d74c51627a223638a489d4ff4bc9a. 36 pinned repositories, 144 unscoped questions, 141 target files. Labels were frozen from native source before running the expanded candidate.
• Pack/shard selection, coherent resume/aggregation, five evidence outcomes, per-pack subprocess deadlines/checkpoints/cleanup, JUnit and runtime operation inventories are implemented. Cloud selection and bounded resume-report upload are supported; resume checks source hashes before rental and full environment identity on the worker.
• Module ownership, static aliases and selected references are implemented across all three language groups with negative controls. These remain partial static capabilities. See the quality guide for unsupported dynamic/build cases.
• Parser work covers 36 languages/formats, no implicit grammar acquisition, bounded slow grammars, C++/Objective-C headers and F# recursive type declarations. Bulk persistence avoids scanning all references once per file.
• Real TypeScript cold and installed Rust warmed identity/edit/shutdown checks passed. Python/Go worker checks, explicit acquisition of 11 available optional grammars, and Linux/macOS 3.12/3.13 upgrade checks have passed CI. Crystal is unavailable in the pinned grammar release and remains explicit unsupported evidence.
• Full standalone diagnostic: 525 passed, four skipped, 52.8 seconds, 746.6 MiB sampled RSS, 95.5 MiB managed disk, cleanup verified. Seven earlier test-location failures remain recorded; study fixtures require external scratch paths.
• Immutable stage A: all 36 attempted, 32 completed without failed assertions, two failed and two timed out. 136 retrieval observations: 95 labeled hits at 5, 124 at 24. Limited/unsupported checks are not positive assertions. The ranking gate passed; raw reports remain under .attocode/quality/runs/.
• Immutable stage B local: 20261014T203352-572ed8e3, 469 passed/four skipped, scale and lifecycle passed in about 40 seconds. Broader stage B results are being collected. Later fixes require a fresh final candidate.
• The first isolated 0.1.5-to-0.1.6 wheel upgrade check passed all seven phases; all three project-local client configurations and persisted index/edit freshness were verified. Its temporary environment/cache were removed. The final source wheel also passed all seven phases in release-0.1.6-final; this is not publication.

Historical remaining delivery gates (subsequently completed)

At this checkpoint, implementation was merged and remote results/cleanup were recorded, but the full cold matrix still failed on CockroachDB indexing. The six-batch allowance was exhausted. A separately approved later cycle passed the full gate and published 0.1.6, as recorded at the top of this page. Keep failures, unsupported cases, missing relationship labels and untested advanced operations visible. No model benefit, ranking improvement or universal reference resolution is claimed.

CI corrections and final verification

• Real Python, TypeScript, Go and Rust servers passed selected-definition, duplicate identity, edit and shutdown checks. Pyright exposed missing workspaceFolders during initialization; the client and protocol regression now provide/check it.
• Root tests exposed a relative Go import compatibility regression, now fixed with unique-target handling; registered Bash/optional-Crystal expectations were updated.
• Full standalone run 20261014-full-final: 684 passed/four skipped, 62.7 seconds, 730.6 MiB sampled RSS, 98.5 MiB managed disk, scratch cleanup verified. Additional CI correction/export regressions were checked separately.
• Stage C observed all 144 questions (99 known-target hits at 5, 126 at 24). Two qualified-name harness failures are retained. Stage D corrects exact-name matching without accepting fuzzy symbols; its full matrix and unchanged ranking gate follow.
• Runtime inventory: 30 operations with successful response observations, 109 with rejection-only evidence, zero unmapped. This does not establish semantic coverage of all 139 operations; subprocess MCP calls are outside this inventory.
• Cloud reports compact JSON losslessly before the unchanged 2 MiB/file, 20 MiB total export limits. No corpora or indexes are downloaded. PR no-model-review prevents automatic model review for this explicitly model-free workstream.

Final stage D source matrix: all 36 packs and 144 questions completed across three coherent shards. Outcomes: 157 passed assertions, 124 limited, seven unsupported, zero failed/incomplete. Known-target hits: 99/144 at five, 126/144 at 24; these are non-exhaustive development labels. Local contracts: 493 passed/four skipped. Final ranking and cloud/release evidence are tracked separately. Stage D predates the report-export-only harness correction; product source is unchanged by it.

Prior-push CI root suite on Python 3.12: 8,477 passed, five skipped, one expected failure; Python 3.13 also passed. Final commit gates must still run before merge.

The local checks plus source matrix took 504.3 supervised seconds, peaking at 1,249.2 MiB managed storage and 830.4 MiB sampled process-tree RSS. Cleanup was verified. The unchanged ranking gate is an additional separately recorded check.

Stage D ranking run 20261014T210555-39a71a6c passed in 55.5 seconds: the frozen 68-query/five-method gate remains unchanged. Total supervised final local checks, source matrix and ranking time: 559.8 seconds; all cleanup verified.

Clean Linux cold-start failures and follow-up

PR #181 merged as fbd5211f7f8b9c47d7bc37fc18471792a8e6fe41 after its required checks passed on the final head. The first clean cloud batch, 293a9ef68f1547fd, tested that published source. Lint, 480 contracts (18 explicit skips), scale and lifecycle passed. Of 36 source packs, 34 completed; CockroachDB exceeded the 180-second pack limit and Pandas exceeded the 90-second cold search warmup wait. Eight questions are missing, not failed retrieval observations. The runner correctly stopped before ranking, so this is not a passing broad run.

Retained check outcomes: 123 passed, 116 limited, 37 unsupported; no failed assertions were observed before the two pack interruptions. Optional grammars were absent on this clean worker. The report's check-level incomplete count does not count interrupted packs; their explicit statuses and missing questions govern completeness. See the remote record.

The cloud worker and its scratch were deleted and an independent API query found zero active quality workers. Estimated compute cost: $0.0733; one of this workstream's six allowed batches used, $4 cumulative monthly reservations. Reports exported 1.30 MiB, with no corpora or indexes downloaded.

The follow-up batches symbol/reference persistence in atomic groups of at most 100 files and reuses a bounded set of content-hash-checked symbol spans between keyword and body indexing. It retains no ASTs or source text. Regression controls check rollback, selected-definition keys, identical source chunks, bounded-cache fallback and same-size/same-mtime edits. No ranking weights, labels or timeout limits changed. Final performance and remote revalidation remain required.

2026-10-16: fresh timing check and intelligence-first transition

The fresh CockroachDB cloud run on 9518a3d passed its unchanged 180-second pack check in 159.621 seconds. The independently verified worker was deleted; all 23 reserved batches have verified cleanup, with no active quality workers. Estimated cumulative compute is $0.9405, not an invoice. The $23 reservation cap is fully used; another rental would need new authorization.

The repository transition is implemented: the complete agent is preserved at agent-archive-2026-10-16, useful tests now import the standalone engine, and the root workspace, CI and documentation target intelligence. See the archive and migration guide. Historical experiment artifacts, private client settings and shared caches remain separate from this migration. Final validation and merge evidence are recorded in the transition results.