Skip to content

project-status.md

Latest commit

 

History

History
703 lines (616 loc) · 49.1 KB

project-status.md

File metadata and controls

703 lines (616 loc) · 49.1 KB
Raw

CodeMesh Project Status

Document type: current status snapshot

Reviewed: 2026-09-08

Implementation status is reviewed against the current inspected worktree. Current search-scope correction and provider-free retrieval evidence is bound to candidate 4fb9e81. Capped-runner and incomplete campaign evidence remains bound to candidate 0f18071. Earlier configured/onboarded task-adaptive evidence remains bound to fbc1433; other configured agent evidence remains bound separately to one-step one|nine candidate 8844c21, Config.Net candidates b16eee7 and 3804028, and YoutubeDownloader candidate 64d4052. Earlier paid campaigns remain bound to candidates 0bde606 and 0378740.

Retained evidence below remains bound to the exact commits named in each evidence section; it does not inherit this implementation baseline.

This document records current implementation and retained evidence. It does not set execution order; use docs/planning/next-steps.md for that. Passing tests or evaluation harnesses do not by themselves establish generalized agent benefit, release readiness, deployment acceptance, or production fitness.

Implementation Update: 2026-09-10

The owner-activated development-session MCP feedback loop is implemented in the current local source. An immutable, hash-pinned, expiring, separately revocable session authorizes exact client and CodeMesh bindings. The development-feedback profile adds bounded local v2 packet recording with server-derived provenance; the feedback-maintainer profile adds allowlisted list/get intake and deterministic non-authorizing resolution-plan hashes. Installation-plan v2 pins the session path and hash while retaining v1 normal plan loading. The existing normal four-tool and diagnostic surfaces remain separate.

The full 257-test Agent Access suite, including a real-stdio provider-free integration test, passes locally. The integration records a failed client report, permits maintainer intake and resolution planning without a CodeMesh edit, records a passing superseding recheck, derives the original failure as resolved, rejects calls after revocation, and retains both append-only packets. This proves the implemented same-machine workflow and safety contracts only. It does not prove that agents provide useful feedback, that a proposed fix is correct, that a human approved an edit, or that CodeMesh improves agent outcomes.

Review Checkpoint: 2026-09-08

The review began on clean local main at b2c053c, 65 commits ahead of origin/main. Only documentation and retained evidence changed between the packaged source candidate a158bc0 and that review baseline; application source, tests, build files, and dependency locks are unchanged. Historical checks remain bound to their recorded candidates and were not rerun as current-HEAD tests.

Reviewed area Evidence checked in this review Conclusion
Recent implementation Parser design-time output isolation, explicit HTML attribute escaping, exclusive feedback packet creation, and their regression coverage in source/tests Recent corrections are present; no implementation drift requiring a contract or architecture update was identified.
Local distribution and runtime records SHA-256 recheck of all four package artifacts, eight configured-runtime evidence files, and twelve dependency locks against the retained manifests All 24 hashes match. Package/runtime execution remains the retained a158bc0 local evidence, not a new execution or release acceptance.
Product progress Configured campaign records and the review packet's outcome gate One bounded Config.Net improvement coexists with neutral, regressed, and incomplete campaigns; repeatable configured-agent benefit remains unproven.
Public state Read-only GitHub main, workflow, alert, release, and tag queries on 2026-09-08 Remote main remains 8e9c0da; hosted successes cover that older source. There are 39 open CodeQL alerts, zero open Dependabot/secret-scanning alerts, and no releases or tags.
Remaining work Seven-gate source review packet and execution order Local preparation is reviewable. Current-source hosted review, supported environment/resource and upgrade acceptance, repeatable product evidence, and owner release/publication decisions remain open.

This checkpoint refreshes source and evidence inspection only. It does not activate another model campaign, change a pilot binding, start services, or supply new application, dependency-advisory, or environment-acceptance results.

Documentation verification passed Markdown lint, version consistency, publication safety, workflow policy, and whitespace checks. The canonical Markdown-link check failed on two missing docs/experimental.md targets inside ignored downloaded CodeQL C# and Python query packs under .codemesh-evals/; it reported no project-documentation link failures. Those retained packs were left unchanged. Application suites and smoke were skipped for this documentation-only review.

Product Position

CodeMesh is a local-first context and likely change-impact service for coding agents. The initial product wedge is C#/.NET implementation discovery, impact navigation, and validation selection through compact source-backed context.

The current milestone is configured/onboarded evidence that a normal agent host uses the selected CodeMesh v1 repository integration, then performs at least as safely and correctly as normal repository exploration with attributable efficiency or correctness benefit. Spontaneous MCP discovery remains useful usability evidence, but is no longer the v1 acceptance contract.

Capability Status

Area Current status Important boundary
C# parsing Implemented with Roslyn semantic relationships, declarations, test files, and test cases. Strongest parser and current product wedge.
Python parsing Files, classes, functions, async and multiline functions, imports, local/imported calls, test cases, and containment implemented. This is a lightweight structural parser rather than a complete Python semantic model; dynamic imports and type resolution remain limited.
Rust parsing Files, modules, structs, enums, traits, impl blocks, functions, methods, use declarations, test cases, call/type relationships, and PyO3 exports implemented in CodeMesh.Parser.Rust. The parser is heuristic and source-structural rather than rustc-semantic; broader macro and conditional-compilation resolution remains open.
Markdown parsing Documents, sections, anchors, source spans, and containment implemented. Cross-links to code, commands, endpoints, and projects are not modeled.
Deployment parsing Initial Dockerfile and Compose nodes, metadata, containment, and service dependencies implemented. Helm, Kubernetes, CI/CD, overrides, and code-to-runtime links remain planned.
Ingestion Dry-run, incremental writes, staged deterministic snapshots, atomic multi-language parser composition, compare-and-swap slot publication, reference-aware cleanup, refresh, watch, redaction, optional embeddings, and optional summaries implemented. One selected language set publishes one snapshot or fails before slot activation; watch remains polling-based.
Storage Snapshot-namespaced Neo4j graph, MongoDB content/registry/runs/summaries, Qdrant vectors, and in-memory test adapters implemented. Neo4j graph publication uses bounded write batches. Publication is atomic at the registry slot boundary rather than one cross-store transaction; normal service-backed operation requires configured stores.
Agent Access REST, stdio MCP, Python CLI, typed .NET client, repository status, context search/packages, symbols, nodes, neighbors, and deletion implemented. The normal four-tool profile retains exact binding, fail-closed freshness, reviewed installation, provider-free probing, and package guidance. Explicit development-feedback and feedback-maintainer profiles add immutable session authority, server-derived v2 packet recording, allowlisted intake, and non-authorizing resolution hashes. V1 packets and installation plans remain readable. Repository deletion remains outside every MCP profile. Feedback is local diagnostic input, not trusted truth, approval, outcome memory, or product-benefit evidence. Returned relationships and rankings remain navigation evidence, not runtime proof.
Context packaging Agent-format packages, 12,000-character MCP cap, source spans, rationale, relationships, and .NET, Python, and Rust validation recommendations implemented. Source verification remains required for consequential conclusions.
Local UI Read-only health, repository, run, freshness, and search views implemented. It is a local inspection surface, not established product-benefit evidence.
Evaluation Deterministic fixture, live MCP, spontaneous, evaluator-assisted, configured/onboarded paired-agent, and model benchmark harnesses implemented. Agent campaigns require explicit input-plus-output-plus-reasoning reported-token accounting, a runner-declared hard-cap capability, remaining-budget limits on every invocation, boundary stopping, and no automatic retry. Configured mode also verifies the reviewed normal-profile plan, committed guidance, baseline MCP identities, fresh binding, positive runtime probe, and fail-closed rejection probe. Configured outcomes include neutral one|nine, regressed and improved Config.Net, and two duration-regressed YoutubeDownloader campaigns. None establishes repeatable benefit. Default codex capability preflight refuses because it cannot guarantee the hard cap. Explicit capped-codex failed closed in its first authorized campaign when provider credits were unavailable; no pair completed and the campaign is consumed.
Generated summaries Complexity-aware prompts, local/online providers, and a provider-neutral qualification runner/report compiler are implemented. The runner uses the production prompt/parser, requires calibration and held-out qualification partitions, redacts inputs, records immutable suite/profile/source identity, creates typed blinded review packets, and binds isolated live-retrieval plus resource/cost evidence before applying quality, safety, retrieval, and deployment gates. No summary deployment profile is CodeMesh-qualified. Provider-backed execution, a complete reviewed corpus, two blinded reviews, isolated indexes, query-level retrieval assessment, deployment measurements, and model/source/spend authority remain separate requirements.
Security Pre-persistence value redaction, immutable generated/cache exclusions, default known-secret-file exclusion, repository-relative allow/deny globs, fail-fast pattern validation, and path-policy-bound snapshot identity implemented. Explicitly including known secret files requires a separate CLI flag and does not weaken redaction or repository authorization requirements.
Repository identity Stable project/checkout markers, deterministic source snapshots, checkout-current and clean branch-head slots, explicit retained slots, legacy migration, and richer local freshness implemented. Separate clones attach with explicit --project-id; shared-service tenant/authorization/privacy controls remain deferred.

Detailed current interfaces are in Agent Access Contracts, and component boundaries are in Architecture.

Activated one|nine Pilot

The bounded pilot is active only for the clean Primary checkout at fe2f761583bf78601961ff17934185c4b6a632f9. It is bound to project prj_a570a726917d4d0ca596085aaf07af50, checkout chk_70c50c5494de4837ab0627aca021eab2, and source view fea1951d1eeb93ac265112aef7000dcac1a9832b000ea81b360b0851df56e69b. Provider-free ingestion run 98b439c4b0d484297852dae7723be555e75bf8fcf7c565c96a9b0bad43b66062 published snapshot snp_bf026d3f2e2b5a40f3c9fa112ae36b7e026b0f5a0f0483a23796b0a6adc9243e with 45,650 nodes, 95,218 relationships, 12,873 content items, zero embeddings, and zero summaries. It redacted 324 content items before persistence.

The reviewed plan at hash fcadac37ae4df261e8b746e14f9191e22a7b6399b991925d4d1e12a5947daf34 changed only Primary's ignored .codex/config.toml. The normal-profile probe accepted the exact fresh binding and the configured preflight preserved prompt parity. Its installed ignored AGENTS.override.md guidance was retained by source classification and hash, not raw content. The positive gate returned both app/feed_hub/native_replay.py and native/feed_hub_replay/src/lib.rs; the wrong-checkout gate failed closed with the live binding, freshness, and provenance issue codes.

An execution-readiness recheck on 2026-09-01 found the detached CodeMesh candidate still clean at 0bde606 and confirmed that the reviewed plan, installed configuration, installed guidance, and retained preflight still matched their recorded hashes. Primary had advanced to 011f562 on main; its branch tip was preserved while the checkout was temporarily restored to exact fe2f761. The exact configured preflight passed again without repinning or re-indexing, and Primary was returned to clean main at 011f562 after the campaign.

Clean detached CodeMesh candidate 0bde606 passed all eight frozen one|nine provider-free gates, comprising four positive retrieval gates and four rejection gates, and validated all three retained feedback packets. The same candidate passed the pinned six-case, 18-call YoutubeDownloader summary-free live gate with 1.0 tool success and recall, 0.9 MRR, 0.926186 nDCG, and zero secret leaks. Its final local verification comprised 124 Python tests, Ruff check and format, and all five deterministic MCP fixtures. After the missing ASP.NET Core targeting pack and runtime were installed, attended deterministic .NET verification at documentation commit 9c8aca3 also passed: locked restore, build, 64 tests, and format verification completed with zero failures. The build retained one CS8625 nullable warning, and the YoutubeDownloader sample test remained skipped because CODEMESH_SAMPLE_CSHARP_ROOT was not configured. No solution, project, or C# source differed from evaluated candidate 0bde606 at that verification commit.

The attended correction path remains evidence: 4eccd80 missed the Rust expected path, a6ebaa5 expected the wrong live rejection diagnostic, and 854d678 failed the YoutubeDownloader MRR threshold at 0.766667. No model was invoked in those attempts or the final pass. The sanitized exact-candidate record is configured one|nine provider-free evidence.

The status-to-context correction is now frozen in clean detached candidate 0378740. At that exact commit, locked restore, build, 69 .NET tests, format verification, 125 Python tests, Ruff check and format, all five deterministic MCP fixtures, and all repository standards checks passed; one nullable build warning and the unconfigured YoutubeDownloader sample skip remain. Against the same exact Primary source identity, the new normal-profile configured preflight, all eight frozen one|nine gates, all three feedback packets, and the pinned YoutubeDownloader live gate passed with provider mode none. Primary's newer main tip 0297190 was preserved and restored after the gates. The sanitized record is corrected candidate provider-free evidence.

The subsequent one-step entry correction is frozen in clean detached candidate 8844c21. MCP instructions, tool descriptions, and generated onboarding now make the fail-closed context package the first normal-host action while keeping repository status as an explicit diagnostic and retaining its compatibility next-action field. At that exact commit, locked restore, build, 69 .NET tests, format verification, 125 Python tests, Ruff check and format, all five deterministic MCP fixtures, and all repository standards checks passed. The final service-backed .NET rerun passed after an earlier Neo4j cold-start cancellation; the existing nullable warning and unconfigured YoutubeDownloader sample skip remain.

one|nine instance configuration commit f9f199d repinned only Primary to the new candidate and regenerated all managed instance instructions successfully; it did not activate another checkout. Against exact Primary source identity fe2f761, the reviewed configured preflight, all eight frozen one|nine gates, all three feedback packets, and the pinned YoutubeDownloader live gate passed with provider mode none. Primary's newer main tip 162fa515 was preserved and restored after the gates. The sanitized record is one-step candidate provider-free evidence.

The authorized gpt-5.6-sol medium-reasoning configured campaign then completed three paired repetitions. All six executions were correct and safe, with no failed MCP calls or command-policy violations. Two of three treatments called codemesh_get_repository_status, but no treatment called codemesh_get_context_package; the third treatment made no CodeMesh call. The verdict was therefore insufficient, with zero treatment wins and zero CodeMesh-attributable wins. On the two status-adopting pairs, treatment medians were 2.22% slower and used 6.14% more tokens; overall medians were 30.29% slower and used 46.61% more tokens. Those differences are diagnostic because no context retrieval occurred. The sanitized record is configured one|nine agent evidence.

The separately authorized rerun through exact corrected candidate 0378740 also completed all three pairs correctly and safely, with no failed MCP calls, command-policy violations, or treatment regressions. One treatment called codemesh_get_repository_status; no treatment called codemesh_get_context_package. The verdict remained insufficient, with zero treatment wins and zero CodeMesh-attributable wins. Overall treatment medians were 6.07% slower and used 4.17% more tokens. The single status-adopting pair was 44.22% slower and used 21.54% more tokens, but those differences remain diagnostic because task context was never retrieved. The sanitized record is corrected configured campaign evidence.

The separately authorized gpt-5.6-sol medium-reasoning campaign at exact candidate 8844c21 completed all three configured pairs. All six executions were correct and safe, with no failed MCP calls or command-policy violations. Every treatment called codemesh_get_context_package directly, so the one-step entry correction produced full measured adoption. The verdict was neutral: there were no treatment wins, attributable wins, or matched regressions. Median treatment duration was 1.39% lower and median tokens were 4.13% higher, both inside the 10% neutrality band. One treatment used 110 shell commands while the other two used 19; that outlier remains diagnostic. Primary was restored to clean main at 162fa515 after the consumed campaign. The sanitized record is one-step configured campaign evidence.

Together, the provider-free and model-backed records prove the bounded configuration, runtime identity, safety, and adoption observed across the campaigns. They do not establish attributable correctness or efficiency benefit, repeatability, release readiness, deployment acceptance, provider authorization, or production fitness. Secondary, Tertiary, Release, and Hotfix are not activated; Run-only remains excluded.

Retained Product Evidence

Summary-free retrieval

  • The pinned YoutubeDownloader live suite passed all six cases with complete expected target and relationship coverage.
  • Two corrected three-repeat baseline agent campaigns completed all nine pairs with equal 100% task success but zero CodeMesh adoption.
  • A separate complex output-container task reached 100% live recall after a focused ranking correction, while its spontaneous agent campaign still made zero CodeMesh calls.
  • Overall timing or token differences from non-adopted treatments are retained as diagnostics and are not attributed to CodeMesh.

The sanitized live-gate manifest is docs/evaluation/evidence/summary-free-baseline-e84b189.json.

Assisted utility

  • The first valid onboarding-assisted complex campaign achieved full adoption and task success but regressed median elapsed time and tokens.
  • After the MCP agent brief was capped at 12,000 characters and onboarding was focused on default packages plus targeted verification reads, the same three-pair task retained full adoption and correctness. Median elapsed time fell 34.7%, median tokens fell 64.5%, and treatments used 2 to 4 shell commands versus 14 to 18 for controls.
  • The independent Config.Net summary-free live gate passed at clean CodeMesh commit d764270: all 18 calls succeeded, recall and MRR were 1.0, nDCG was 0.93483, and no secret leaks were detected.
  • The Config.Net assisted campaign produced two CodeMesh-attributable correctness wins, full treatment success and adoption, no safety violations, and 11.8% lower median elapsed time. Treatment median tokens were 3.0% higher.

These results replicate utility after explicit onboarding. They do not prove prompt-parity discoverability, spontaneous adoption, or broad generalization. The Config.Net transcription is docs/evaluation/evidence/config-net-assisted-replication-d764270.json.

Prompt-parity discoverability

  • The Config.Net summary-free live gate remained comparable and passed at clean CodeMesh commit 647d83d: all 18 calls succeeded, recall and MRR were 1.0, nDCG was 0.93483, and no secret leaks were detected.
  • The frozen three-pair prompt-parity campaign completed with identical control and treatment task prompts, but no treatment run called or attempted a CodeMesh MCP tool.
  • Control passed two of three runs and treatment passed one of three, with no safety violations. The harness verdict was regressed, but the zero adoption rate means overall timing and token differences are diagnostic and cannot be attributed to CodeMesh.

This is negative normal-host discoverability evidence, not a failure of the assisted CodeMesh workflow or a broad product verdict. The sanitized transcription is docs/evaluation/evidence/config-net-prompt-parity-647d83d.json.

Configured replication readiness

  • Exact clean CodeMesh candidate b16eee7 adds a configured/onboarded suite while preserving the frozen Config.Net prompt, targets, thresholds, and source commit.
  • The reviewed local normal-profile installation and ignored repository guidance bind exact Config.Net commit ae0af7d, project prj_e5d8c8, checkout chk_6d1d16, and source view b4981658.
  • The configured preflight passed with prompt parity, fresh accepted binding, the four-tool normal profile, a positive context package, and fail-closed rejection. The six-case live gate then passed all 18 calls with recall, MRR, and nDCG of 1.0 and zero secret leaks.

This is provider-free configured-integration readiness, not agent adoption or product-benefit evidence. The exact sanitized record is configured Config.Net provider-free evidence.

The separately authorized configured campaign then completed all three pairs with context-package adoption in every treatment and no safety, MCP transport, or command-policy failure. All three controls passed, while two treatments missed required implementation or test citations; the treatment success rate was therefore one of three and the verdict was regressed. On the only pair where both conditions passed, treatment was 21.31% slower and used 117.18% more tokens. This is valid negative configured-product evidence, not replicated benefit. Preserve the consumed result and do not retry it automatically. The sanitized record is configured Config.Net agent evidence.

Independent configured replication

Exact clean candidate 64d4052 adds a configured/onboarded version of the frozen YoutubeDownloader audio-container impact task without changing its prompt, targets, thresholds, source commit, or model settings. Its 125 Python tests, Ruff checks, five deterministic fixtures, reviewed normal-profile plan, configured preflight, positive and wrong-checkout probes, two-case impact live gate, and unchanged six-case live gate all passed. The impact gate covered all six required targets with recall and MRR 1.0, nDCG 0.891323, and zero leaks; the unchanged live gate retained recall 1.0, MRR 0.9, nDCG 0.926186, and zero leaks. Exact identities and hashes are retained in the configured YoutubeDownloader provider-free evidence.

The separately authorized configured campaign completed all three pairs with 100% control and treatment correctness, context-package adoption in every treatment, and no safety, MCP, or command-policy failure. Treatments attempted the context package two, two, and three times. With no correctness win or regression, treatment median duration was 49.79% higher and median tokens were 41.00% higher across all three matched pairs, so the valid verdict is regressed. This replicates configured adoption and correctness on another C# repository, not repeatable product benefit. Raw query arguments were not retained, so focused-query guidance is a diagnosis lead rather than an established cause. The sanitized result is configured YoutubeDownloader campaign evidence.

Task-adaptive query correction

The provider-free YoutubeDownloader diagnosis preserved its six impact targets, source commit, repository identity, and retrieval thresholds. One frozen task-shaped package covered all six targets in one call with primary rank two. Three proactive focused packages also covered all six, but produced 35,991 agent-formatted characters instead of 12,000 and repeated five of 24 ranked item occurrences. The diagnostic comparison is retained separately because its CodeMesh worktree contained the new suite and guidance edits: task-adaptive query diagnosis.

Exact clean detached candidate fbc1433 contains the smallest correction. Agents now start with one concise task-shaped package, stop when it resolves every required facet, and issue a focused follow-up only for a specific unresolved implementation or test facet. The guidance caps each task at three non-overlapping packages. At that commit, 125 Python tests, Ruff checks, five deterministic MCP fixtures, the reviewed normal-profile preflight, positive probe, and wrong-checkout rejection all passed. The adaptive five-case gate passed all 15 calls with recall 1.0, MRR 0.875, nDCG 0.862253, and zero leaks. The unchanged impact gate retained recall and MRR 1.0 and nDCG 0.891323; the unchanged canonical gate retained recall 1.0, MRR 0.9, and nDCG 0.926186. No model provider was invoked. Exact hashes and identities are in the task-adaptive provider-free evidence.

The separately authorized fbc1433 configured campaign then completed all three pairs with 100% control and treatment correctness, one successful context package in every treatment, and no safety, MCP, or command-policy failure. It had no correctness win or regression. Median treatment tokens were 6.63% lower, but median duration was 10.0192% higher, crossing the frozen 10% threshold by 0.0192 percentage points and producing a regressed verdict. The campaign also reported 2,596,056 aggregate tokens, including 2,219,264 cached-input tokens, and therefore exceeded its authorized 1.6M reported-token ceiling. The evaluator at that candidate had no hard aggregate-token stop; this is a control failure, not a reason to discard or retry the consumed result. Raw traces and query arguments were not retained. The sanitized record is task-adaptive configured campaign evidence.

Current Gap

The next search diagnosis now has request-isolated operation timing in the source. It separates source-store fetches, candidate selection, context expansion, declaration and content reads, optional relationships, and final ranking, while retaining the existing package-stage timing and unchanged responses. Injected-delay tests cover attribution, result parity, overlapping requests, and failure cleanup. This instrumentation does not establish a new agent-benefit result.

Clean instrumentation candidate 2879e7f passed the unchanged adaptive live suite in a fresh isolated store with recall 1.0, MRR 0.875, nDCG 0.862253, and zero leaks. The task-shaped package p50 was 451.070 ms. Adding 50,000 synthetic unrelated graph nodes left retrieval unchanged but increased package p50 to 1,504.467 ms, including 1,227.914 ms in lexical fetch. The query plan loaded lexical properties for all 51,685 nodes before repository filtering. An isolated query experiment moved that filter before property projection, loaded properties for only the 1,685 target nodes, and preserved the exact ordered candidate hash. Clean correction candidate 4fb9e81 passed the frozen adaptive, impact, and canonical live suites with unchanged quality and zero leaks. Under the same synthetic store load, task-shaped package p50 fell to 391.808 ms (73.96% lower) and lexical fetch p50 to 124.113 ms (89.89% lower). All adaptive cases retained identical ordered candidates. The correction evidence records exact candidates, checks, hashes, and claim limits. This controlled mechanism does not establish the cause of the historical 16.7-second result or an agent benefit. The search-scope diagnosis retains timings, plans, fixture identity, report hashes, and limitations.

Explicit repository onboarding is implemented for the bounded one|nine pilot: the normal profile exposes only context package, repository status, repository listing, and node lookup; exact binding and freshness are enforced; and installation plus probing are reviewable. Clean one-step candidate 8844c21, exact Primary commit fe2f761, the configured preflight, the eight-gate frozen one|nine pilot, and the YoutubeDownloader live gate all pass provider-free. The configured normal-host campaigns at historical candidate 0bde606 and corrected candidate 0378740 were both valid but insufficient. Across them, three treatments called repository status and none retrieved task context. The subsequent 8844c21 campaign was valid and neutral: all three treatments retrieved a context package, all six executions passed safely, and there were no attributable wins or regressions.

The prompt-parity Config.Net campaign recorded zero MCP calls and zero MCP call attempts. Preserve it as negative spontaneous-discovery evidence; do not change the frozen task prompt, grading, or targets to manufacture adoption. The first two configured one|nine campaigns evaluated the status-first onboarding path and showed that the status-result action did not produce context adoption. The third campaign preserved the same task and comparison conditions and showed that direct context-package entry did produce adoption in every treatment.

The evaluated MCP configuration completed without a harness failure, and the CodeMesh manifest published the intended instructions and tool contracts. The measured transition failure was narrower: a fresh accepted status response did not itself state the next tool action. The current implementation adds an MCP-only codemesh_next_action status field that directs a fresh accepted binding to continue immediately to an agent-format context package and directs every other state to fall back. This correction is committed, covered by deterministic contract tests, and verified through the full provider-free exact-candidate gates at 0378740. Its paid campaign still observed one status-only treatment and no context-package call, so the action field did not produce the required normal-host handoff.

The current implementation removes that two-step dependency. MCP server instructions, context-package and status tool descriptions, and generated repository onboarding now use the fail-closed context package as the single normal-host entry action. Repository status remains available for explicit ingestion, count, and freshness diagnostics, and its compatibility next-action field remains intact. Exact clean candidate 8844c21 passed the full deterministic suite, configured preflight, eight one|nine pilot gates, feedback validation, and pinned YoutubeDownloader live gate with no model provider. The authorized paired campaign then observed context-package adoption in every treatment and a neutral outcome. This proves the measured handoff for the selected task; it does not prove attributable product benefit or cross-repository repeatability.

The configured/onboarded evaluator now preserves prompt parity, uses the exact reviewed checkout-local normal installation, and retains integration mode, guidance, baseline MCP, tool-profile, binding, installation-plan, and runtime- probe identities. It rejects a dirty or stale target, a dirty CodeMesh source checkout, commit or binding mismatch, changed plan files, unexpected tool surface, and a failed rejection probe. Historical spontaneous and assisted modes retain their prior diagnostic behavior.

The provider-free task-shaped context-package coverage diagnosis is now complete at exact baseline candidate b16eee7. The frozen 58-term task query returned eight test paths and only one of seven required targets. Four concise task-derived facets covered all seven targets in a separate five-case/15-call provider-free suite with recall 1.0, MRR 0.660714, nDCG 0.76259, and zero secret leaks. The unchanged canonical live suite also passed all 18 calls with recall, MRR, and nDCG of 1.0. This isolates a bounded guidance and query-decomposition gap without proving which query the paid treatments used. Exact details are retained in the focused-query diagnosis.

Exact clean detached candidate 3804028 contains the smallest correction: the context package remains the single normal-host entry tool, while MCP instructions, tool and status guidance, and generated onboarding direct multi-part work to a small set of focused implementation and test queries instead of one long full-task query. All 125 Python tests, Ruff checks, and five deterministic MCP fixtures pass at that candidate.

The reviewed configuration-only Config.Net plan, separately managed ignored guidance, configured preflight, positive probe, and wrong-checkout rejection all passed with provider mode none. The committed focused suite then passed all five cases and 15 calls with recall 1.0, MRR 0.660714, nDCG 0.76259, and zero secret leaks; LazyVar.cs remains at the accepted rank-eight boundary. The unchanged canonical suite also passed all six cases and 18 calls with recall, MRR, and nDCG of 1.0 and zero secret leaks. This establishes an exact provider-free correction candidate, not normal-agent compliance or product benefit. Exact identities and report hashes are retained in the focused-query provider-free evidence. The separately authorized exact-candidate campaign then completed all three pairs and received an improved verdict. Every treatment used the context package, made two or three package attempts, passed the citation grader, and completed without safety, MCP, or command-policy failure. Controls passed two of three; the second control missed ConfigurationBuilder.cs, producing one CodeMesh-attributable win and no treatment regressions. Across the two jointly successful pairs, treatment median duration was 13.69% lower and median tokens were 3.59% lower. This is positive configured evidence for one task, repository, model, and three-pair campaign; it does not establish that the wording change caused the result or that product benefit is repeatable. The sanitized record is focused-query configured campaign evidence. Preserve all one|nine and Config.Net campaigns in their original evidence classes. This retest is consumed; no further model spend is authorized.

Both configured YoutubeDownloader campaigns are consumed. Candidate 64d4052 preserved full adoption and correctness but regressed duration by 49.79% and tokens by 41.00%. Candidate fbc1433 reduced each treatment to one context package and lowered median tokens by 6.63%, but its 10.0192% duration regression still failed the frozen threshold. It also exceeded its authorized aggregate reported-token ceiling because that historical evaluator did not enforce a hard stop. The current evaluator closes that control gap with explicit reported-token accounting, a mandatory compatible runner capability, remaining-budget limits, boundary stopping, and no automatic retry. The default Codex runner does not claim that capability and is refused before any model call. The explicit capped-codex runner now counts complete request input, reserves worst-case output-plus-reasoning charge, enforces zero provider request/stream retries, retains ambiguous reservations, and reconciles its sanitized ledger against Codex final usage. These records establish configured adoption and correct/safe completion, not repeatable benefit or efficiency causality. Neither consumed campaign may be retried.

Provider-free attribution at exact instrumentation candidate 66033ad measured the task-shaped package at 16,830.931 ms total p50: search accounted for 16,725.370 ms (99.37%) and item hydration for 102.668 ms (0.61%). Hydration did not qualify as dominant, so the proposed concurrency optimization was rejected. Any fresh paid evaluation still requires exact-candidate deterministic, configured, retrieval, safety, and capped-runner gates plus explicit spend authority.

Exact clean candidate 0f18071 passed 149 Python tests, Ruff check and format, all five deterministic MCP fixtures, repository standards, the configured positive and fail-closed rejection preflight, and the frozen adaptive, impact, and canonical provider-free suites. A synthetic provider-free run through the installed Codex executable completed one counted request with zero retries and exact proxy/final-usage agreement. The one separately authorized gpt-5.6-sol medium-reasoning campaign then stopped after its first control execution because the provider reported no available credits. No pair completed. The proxy retained the interrupted call's full 267,528-token reservation, admitted no later request, and recorded zero request or stream retries. The campaign is insufficient, incomplete, and consumed; it is hard-cap failure evidence only, not agent adoption, correctness, efficiency, or spend evidence. The sanitized record is capped YoutubeDownloader campaign evidence.

Expand onboarding to another one|nine checkout only with separate authority, exact binding, fresh ingestion, and a passing checkout-specific probe. The earlier de239b7 stale state and 9f3216c stale-candidate rejection remain useful historical diagnostics; they are not the current candidate state.

Public Repository And Hosted Evidence

At the 2026-09-08 read-only GitHub recheck, the public main branch remains at 8e9c0da, while the local configured-evaluator candidate and its documentation remain unpublished. The latest hosted baseline-verification success applies to remote 8e9c0da; there is no exact-HEAD hosted workflow evidence for the unpublished local commits.

The latest hosted CodeQL workflow completed successfully for remote 8e9c0da, which proves that the scan ran, not that the repository is security-clean. GitHub still reports 39 open CodeQL alerts: 3 critical command-line-injection alerts and 36 high alerts, comprising 35 path-injection findings and one reflected-XSS finding. Dependabot and secret scanning report zero open alerts. These hosted findings apply to the published remote state and do not establish the security state of the unpublished local commits.

The original CodeQL triage is preserved. Local security hardening now validates evaluation suite names and identifiers, contains setup patches and generated artifact paths, and rejects invalid suites before preparation or model execution. Literal argv and all five UI routes have adversarial regression coverage. The full Python suite passes 227 tests with two existing dependency warnings. The local boundary review accounts for all 39 findings, including explicit operator-selected paths and trusted executable inputs. This is local verification, without an updated hosted scan or alert dismissal.

The subsequent UI correction explicitly HTML-escapes dynamic link attributes and the numeric search limit at their render points. All 231 Python tests pass, including hostile attribute values and invalid or out-of-range limit rejection before store access; Ruff lint and formatting pass. Two existing dependency warnings remain. This verification does not dismiss the hosted XSS finding.

The local-input security review then reproduced overwrites of existing packet destinations, including symlinks and hardlinks. Feedback recording now creates packet files exclusively and explicitly rejects linked outboxes before the existing Git ignore check (which already rejected Linux outbox symlinks). The three overwrite regressions fail before correction and pass afterward; an additional test checks explicit outbox rejection. The full Python suite now passes 235 tests with two existing dependency warnings.

The local CodeQL review now binds complete extraction and pinned suites to exact local source. Python b73e19a has zero default-threat findings and 61 local-input findings; unchanged C# source at e0beba9 has zero default findings and 192 local-input findings. All 253 local-input results have retained locations and boundary review. The HTML finding is absent after explicit attribute escaping. Cached first attempts were rejected for local-threat claims and rerun, with original results retained. This is local static-analysis evidence, without hosted dismissal or a blanket security-clean claim.

Local Release Preparation

The source version is now declared in VERSION and consumed by .NET builds. Python package metadata, lockfile, and version export are checked for agreement; REST OpenAPI uses that export. Both CLIs expose a provider-free version command. The retained source value is 0.1.0, preserving the existing Python value and creating no published release. The full local verification passes 70 .NET tests with no skips, 228 Python tests, and five deterministic MCP cases. The .NET build retains its existing nullable warning and Python retains two dependency warnings. The retained 2026-09-05 .NET and locked Python dependency advisory checks report no known vulnerabilities for CodeMesh's own dependency sets; this does not clear the frozen external sample's separate advisory or the hosted CodeQL findings.

Release Preparation defines a reviewable candidate contract, compatibility policy, environment evidence, verification, recovery, and publication gates. Local package verification now binds Python wheel/source artifacts, a .NET publish archive, and their checksums to clean 96b4ebf. Installed-wheel CLI, REST OpenAPI, normal manifest, bundled suite/patch loading, and all five fixture cases pass. Full smoke passes at that candidate. Parser packaging correction 1ef5ca9 adds the required SDK to its runtime and the version files to the Docker allow-list; its project probe loads through MSBuild without filesystem fallback, while retaining a read-only generated-editor-config warning at that historical candidate. The subsequent C# capability 0.2.3 output-isolation correction now passes two concurrent read-only container parses without warnings, preserves generated compiler metadata and cross-project relationships, and cleans temporary output after completion and observed project failures. Its deterministic verification passes 71 .NET tests with no skips and 228 Python tests. See the output-isolation decision. Hosted security review, Windows and other environment acceptance, a supported upgrade/rollback rehearsal, appropriate product evidence, and owner release approval remain separate gates. A disposable Linux recovery rehearsal now passes at f314043: the complete stopped four-volume set restored into new volumes with identical image IDs and matching graph, MongoDB, vector, registry, snapshot, and context-item digests. This closes the local same-version recovery mechanism check; it does not establish a supported upgrade or production recovery.

A development-baseline upgrade and rollback rehearsal now verifies isolated 8e9c0da to ce23f2e reader compatibility, re-ingestion, identity preservation, eight source citations, scoped deletion, and exact whole-set rollback. The initial six-case retrieval gate failed its relationship check without sample dependency assets. Locked sample restore failed on the frozen AngleSharp advisory; the assets it produced enabled a subsequent unchanged six-case/18-call gate to pass, with advisory warnings retained. The local mechanism passes, while sample dependency and supported-environment acceptance remain open. All task containers are down and backups are preserved.

The Linux Python runtime matrix now passes all 235 tests and five deterministic fixture cases on Python 3.12.14, 3.13.15, and 3.14.7 at b73e19a. Each retains two existing dependency warnings. This extends local source verification; Windows and released-artifact acceptance remain separate.

Refreshed local artifacts now bind the completed parser, UI, and feedback corrections to clean a158bc0. The installed wheel passes version/OpenAPI checks, the four-tool normal manifest, all 16 packaged suites and patches, and five fixture cases. The .NET publish contains the design-time targets and reports the exact build SHA. Full smoke and two concurrent read-only parser-container parses pass; the parser emits no warnings and leaves no scratch directories. The retained 2026-09-05 dependency advisory checks report no known vulnerabilities in CodeMesh's 11 .NET projects and 48 Python packages. All task containers are down. These are local artifacts and checks, with the separate sample advisory and release/acceptance gates still open.

The source review packet now supplies draft release notes and a seven-gate evidence ledger. Artifact bytes and all 12 dependency-lock identities were reverified on 2026-09-08. The packet records the 2026-09-05 requirement for a PR and verify for main; no release or remote tag exists at the 2026-09-08 recheck. Owner review, publication authority, supported profiles, and representative resource acceptance remain open.

The fresh configured runtime probe now verifies normal-profile installation and retrieval, rejection of incorrect source hashes, unknown checkouts, and dirty source, and acceptance after source restoration. The installed wheel drove the unchanged source-backed launch; no model campaign was run.

Current Boundaries

  • Generated summaries remain disabled in the core product-proof baseline.
  • No summary model is qualified.
  • Python, Rust, Markdown, and deployment parsers do not match Roslyn's semantic depth.
  • The activated one|nine pilot is pinned to one Primary snapshot and must not be silently expanded or repointed when either checkout advances.
  • Shared multi-user stores, portable snapshot artifacts, outcome memory, Kubernetes/Helm ingestion, and broader SDLC intelligence are not implemented.
  • Current source redaction and path filtering do not replace repository authorization or review of an explicitly broadened ingestion scope.
  • Hosted workflow success for remote 8e9c0da is not evidence for unpublished local commits, and an automated scan completing successfully does not resolve or justify its open findings.

Status Maintenance

Update this page when implemented capability, reviewed evidence, or the immediate product gap changes. Keep detailed procedures in their canonical guides and keep future execution order in docs/planning/next-steps.md.