todo2code

TODO

Active tickets

• ticket-054 — restore skills-agent discovery, prove a todo2code → Repair PR → independent Validator hand-off, then add three bounded todo2code-grounded skills. Current state: IN_PROGRESS / PUBLICATION; executable changes remain in target repositories.

Backlog tickets

• ticket-058 — synchronize release, runtime, SDK and generated-provenance version identity and add a deterministic drift gate. Plan and owner-ticket creation are approved; current state: BLOCKED / WAIT_FOR_DEPENDENCIES while ticket-054 reserves integration.
• ticket-051 — wire scripts/github-event-log.mjs into CI with explicit flags only. Current state: PLAN / WAIT_FOR_APPROVAL; ticket-048 is now merged.
• ticket-052 — promote the Validator autonomy operator checklist into agent-facing governance (AGENTS.md). Current state: PLAN / WAIT_FOR_APPROVAL.

Completed tickets

• ticket-076 — publishes independently callable code2dsl, docs2dsl and config2dsl facades that emit strictly validated, source-limited t2c.intent/v1 records. Full host/Docker checks, Koru, Validator App and protected plan-first governance passed on exact head 944288f; PR #92 merged as main@24ca3a1 and its branch was deleted. Current state: DONE / DONE.
• ticket-074 — routes Subactor-owned semantic calls through central SubLLM with direct Z.AI GLM 5.2, OpenRouter fallback ordering and provider-visible identity. Koru, independent GLM-5.2 review and protected checks passed; PR #88 merged as main@ceae696. Current state: DONE / DONE.
• ticket-073 — defines canonical t2c.analysis-policy/v1 selection, exact-evidence caching and nested request/token/time ceilings while requiring LLM for selected semantic stages. Koru, independent GLM-5.2 review and protected checks passed; PR #86 merged as main@f0df52c. Current state: DONE / DONE.
• ticket-072 — redacts credentials, contextual credential identifiers and provider management URLs once at the OpenRouter boundary while preserving actionable diagnostics and fail-closed require-llm. Koru, independent GLM-5.2 review and protected checks passed; PR #84 merged as main@790b867. Current state: DONE / DONE.
• ticket-062 — adopted immutable new-project v0.14.0 at exact SHA a22eb47, preserved all eight target workstreams, assigned test/python-runtime.test.ts to sdk, aligned the reusable workflow caller and passed protected PR #77. Current state: DONE / DONE; merge main@a762580.
• ticket-055 — identifies OpenRouter usage with explicit OPENROUTER_APP_NAME, falling back to the analysed project’s folder name. Koru, Validator App and protected checks approved exact head 7aae643; PR #74 merged as main@d0659ca. Current state: DONE.
• ticket-056 — restored both Docker E2E profiles by supplying make and respecting the unlocked Rust library contract. Koru, Validator App and protected checks approved exact head 1752f3d; PR #72 merged as main@5f8e831. Current state: DONE.
• ticket-050 — adopted immutable new-project v0.13.2, assigned CHANGELOG.md and .env.example to the governance workstream, and made governance checks recoverable while retaining fail-closed exact-head approval. Koru, Validator and protected checks passed; PR #70 merged as main@f60d3cc. Current state: DONE.
• ticket-049 — Validator autonomy audit, operator guide and refactor plan accepted by the user on 2026-08-08. Sibling tickets 050–052 retain independent approval and ownership gates. Current state: DONE. See AUTONOMY_AND_REFACTOR_PLAN.md.
• ticket-053 — documents the existing CI-form governance check as a blocking pre-push gate. Koru and Validator approved exact head 079de29; protected PR #67 merged as main@100a7d2 and its implementation branch was deleted. Current state: DONE.
• ticket-048 — republished ticket-047’s GitHub acquisition adapter with explicit-only inputs and no ambient process.env fallback. Validator approval and all protected checks passed on exact head 92f99dc; PR #66 merged as main@f1b3d5d. Current state: DONE for publication purposes; the ticket-local historical status remains part of the merged audit record.
• ticket-047 — built the first GitHub acquisition adapter for the t2c.event-log/v1 codec, mapping one bounded push, pull_request, pull_request_review or completed workflow_run payload onto the closed event vocabulary. Host, governance, Docker and focused checks passed locally, but the work never reached protected main: it carries no Koru or Validator approval, and its squashed commits are rejected by CI governance. Republication is tracked as ticket-048. Current state: DONE.
• ticket-046 — generates a canonical, atomic logs.dsl.txt beside every succeeded, degraded and failed pipeline manifest. Koru and Validator approved exact head 1180e45 with openrouter/z-ai/glm-5.2; protected PR #62 merged as main@c1decdb and its implementation branch was deleted. Current state: DONE.
• ticket-045 — defined the canonical, tamper-evident t2c.event-log/v1 contract and validated its 17-event logs.dsl.txt fixture. Koru and Validator approved exact head 46210e1 with openrouter/z-ai/glm-5.2; protected PR #60 merged as main@a66eb40 and its implementation branch was deleted. Current state: DONE.
• ticket-044 — adopted immutable new-project 0.11.0 and its canonical work-classification package through local Goal. Koru and Validator approved exact head 80f860c with openrouter/z-ai/glm-5.2; protected PR #58 merged as main@aae9ec5 and its implementation branch was deleted. Current state: DONE.
• ticket-043 — exposed ticket-040’s deterministic read-only observer through make preflight, with canonical stdout, stable exit codes and complete Git-state non-mutation tests. Koru and Validator approved exact head 87ce55c with openrouter/z-ai/glm-5.2; protected PR #56 merged as main@4a7445a. Current state: DONE.
• ticket-042 — bounded the aggregate semantic deadline of compare-workspace across both pipelines and all document chunks with 2x scaling and a 40-minute ceiling. Koru and Validator approved exact head 5b51880 with openrouter/z-ai/glm-5.2; protected PR #52 merged as main@2c16449. Current state: DONE.
• ticket-040 — bounded, read-only workspace preflight binding exact local Git state, managed governance and the active ticket before governed edits. Ten focused tests plus host, Docker, governance and complexity gates passed; Koru and Validator approved exact head 567424b with openrouter/z-ai/glm-5.2; protected PR #49 merged as main@008bee5. Current state: DONE.
• ticket-041 — strict immutable Git materialization validation and exact-tree graph/truth-map evidence assembly into the existing t2c.branch/v1 projector. Focused, host, Docker, governance and complexity gates passed; Koru and Validator approved exact head 3779613 with openrouter/z-ai/glm-5.2; protected PR #48 merged as main@f188025. Current state: DONE.
• ticket-021 — assigned root-level generated analysis formats to the integration workstream without changing ticket-directory ownership or the managed wrapper. Koru and Validator approved exact head 9e2feff with openrouter/z-ai/glm-5.2; PR #33 merged as main@8ebeb66. Current state: DONE.
• ticket-039 — bounded, read-only local Git materializer for exact branch/tree/merge-base/ahead-behind, stable patch identity and textual collision evidence. Focused, full host/Docker, governance, complexity and live branch checks passed; Koru and Validator approved exact head 29df450 with openrouter/z-ai/glm-5.2; PR #44 merged as main@2948f4a. Current state: DONE.
• ticket-037 — deterministic, fail-closed t2c.branch/v1 projection over immutable branch evidence. Fourteen focused tests, full host/Docker verification, Koru and Validator approved exact head 50d6dba with openrouter/z-ai/glm-5.2; PR #42 merged as main@b5d2417. Current state: DONE.
• ticket-038 — reconciled ticket-036’s stale pending-review narrative with protected completion evidence. Koru and Validator approved exact head 670894d using openrouter/z-ai/glm-5.2; PR #40 merged as main@ed35d3f. Current state: DONE.
• ticket-036 — deterministic, provenance-preserving t2c.truth-map/v1 core projection. Koru and Validator approved exact head 65b4bc1 with openrouter/z-ai/glm-5.2; PR #35 merged as main@15d2b26. Current state: DONE.
• ticket-034 — scales each OpenRouter request timeout from bounded input, output and structural-complexity pressure, retaining the configured default at baseline and increasing by 2x steps up to 8x and 600 seconds. Exact-head Validator approval used openrouter/z-ai/glm-5.2; implementation merged in PR #31 as main@6116961. Current state: DONE.
• ticket-019 — published the dependency-free Python SDK from one root pyproject.toml, passed package, application, SDK, Docker and protected exact-head validation, and merged PR #28 as main@e333ace. Current state: DONE.
• ticket-035 — declared the five atomic Python publication paths as integration-owned shared contracts, preserved the immutable standard provenance and passed exact-head Validator review plus all protected checks. Implementation merged as main@a2441f8; current state: DONE.
• ticket-020 — role-bound trusted intake with persistent manager/user/dev assignments, CQRS/event sourcing, strict schemas, dependency-free Protobuf codecs and Python/TypeScript CLI, MCP and A2A parity. Implementation commit 06a2faa is contained in protected main; current focused validation passes 9/9, full verification has 0 failures and policy 0.10.0 governance passes.
• ticket-018 — adopted hardened wellmanifest/new-project 0.10.0 at immutable merge 9706e63, moved trusted App authority outside PR control, switched Koru to z-ai/glm-5.2, passed exact-head App review and protected checks, and merged as main@6ad85bd.
• ticket-017 — repaired mutating command help, Polish prohibition polarity and repository-bound path resolution; independently audited the concurrent path/action-planning baseline and added isolated Docker E2E core and full-toolchain gates. Current state: DONE.
• ticket-009 — canonical structured response contracts. All seven production OpenRouter boundaries now generate their provider schema and runtime parser from one typed source, retain rejected-response metadata and fail closed without semantic coercion. Published as d0fc143; current state: DONE.
• ticket-008 — upstream governance hardening. wellmanifest/new-project 0.6.0 now uses role-typed participant templates, explicit unresolved ownership, active-ticket reuse and a separate project/TICKETS.md index that cannot overwrite generated analysis. Current state: DONE.
• ticket-007 — explicit unresolved response routing. Every communication issue now names a known participant or the role sentinel unresolved:human / unresolved:agent; no participant is guessed or created. Current state: DONE.
• ticket-006 — canonical structured-output conformance. Retained exact fail-closed diagnostics and a schema drift gate; rejected both tested Qwen routes before graph mutation. Current state: DONE.
• ticket-005 — audited communication and cross-language reranking. Retained section-aware user-*/ai-* Intent DSL plus explicit response ownership; rejected the live semantic candidate after three fail-closed provider contract violations and zero demonstrated coverage improvement. Current state: DONE.
• ticket-004 — language-independent topic matching benchmark. Rejected unsafe raw embeddings and added a separately reported 6-positive/6-negative cross-language gold cohort. Current state: DONE.
• ticket-003 — deterministic audit of residual changelog diagnostics and one evidence-gated correction. Removed 547 false review findings with stable graphs. Current state: DONE.
• ticket-002 — cross-repository semantic benchmark and iterative todo2code quality improvements. Current state: DONE.

P1 — improve semantic quality and signal

• Build a versioned gold dataset for NL/docs/Markdown → DSL, linking and DSL2TODO. Report precision, recall, citation completeness, deduplication rate and stability between repeated runs.
• Add audited structured LLM enrichment for communication records and a grounded per-participant synthesis. Preserve human/agent identity, ticket, source lines and epistemic class as runtime-owned fields.
• Convert governance user-* and ai-* sections to typed Intent DSL without ingesting ticket evidence/logs, reject unstructured migration silently losing content, and attach the required role plus participant IDs to every communication divergence.
• Route a required response when the required role has no participant record. Ticket-006 correctly returns responseRequiredRole=human but an empty responseRequiredFrom because the agent refused to fabricate a human-owned user-* file. Ticket-007 now emits the explicit role sentinel unresolved:human or unresolved:agent, never a guessed identity.
• Add a participant identity registry mapping stable IDs to Git authors, A2A agent IDs and optional human aliases without guessing identity from display names.
• Aggregate low-level AST calls and symbols into module/capability topics before coverage metrics and team-summary prioritization, while preserving links to the original evidence records.
• Link NL, TODO and documentation declarations to module aggregates through normalized capability topics with a precision-oriented three-topic floor. On the repository audit this raised AST↔NL from 0 to 10 and AST↔TODO from 1 to 71 without increasing the 617 AST↔AST relations.
• Distinguish unmeasured documentation coverage from a measured 0% in Intent-vs-Reality and workspace trends when no document records exist.
• Restrict extracted file paths to recognized extensions so dotted Intent DSL fields such as statement.object remain symbols rather than false paths.
• Require runtime-owned provenance on every Intent DSL record: converter name/version and todo2code version for deterministic generation, plus provider/model/response ID for LLM generation and explicit fallback state.
• Add a deterministic documentation → Intent DSL baseline for headings, code blocks and explicit path/symbol references, keeping LLM extraction as optional audited enrichment rather than the only source of document records.
• Extend the gold linking dataset with prose-to-module positive and hard negative cases, then publish precision/recall separately for exact-target and capability-topic relations.
• Read prohibitions and periphrastic obligation as deontic modality. Before this, nie wolno scored unknown, nie może scored optional because the permissive rule matched the może inside the ban, and \bmuszą\b could never match at all: JavaScript’s \b treats ą as a non-word character.
• Fold regular English plurals in topicKeywords, guarded against ss/us/ is/as/os endings and folding -ies to -y. On its own this added relations without moving coverage, because aligned never read the graph; the topic-anchor change below is what converted relations into the metric.
• Split large TODO/CHANGELOG LLM enrichment into bounded 32-record batches with a shared audit, per-record response provenance and deterministic ordering.
• Measure the complete live pipeline against explicit structured-output models. Qwen failed documentation/communication after correction, GPT-5.4 Mini failed NL after correction, and Gemini 3.6 Flash passed all six stages in 125.486 s for $0.412363. The redacted comparison is in ticket-012 and the opt-in live history; offline CI remains provider-independent.
• Make summary generation structurally validated through t2c.conclusion/v1 before rendering Markdown, rather than accepting free-form narrative as the primary result.
• Add an explicit CLI summary mode deterministic|prefer-llm|require-llm, consistent with NL and Markdown extraction modes.
• Improve workspace trend metrics so AST line movement and source-identity churn do not dominate business-topic coverage changes.
• Add incremental AST and documentation-chunk caches keyed by content hash for very large repositories. TypeScript is cached per file, external adapters per language manifest and Markdown per chunking input; corrupt cache fails open and provider responses are deliberately never cached.
• Resolve unambiguous bare source filenames such as markdown.ts against the repository tree before linking, while preserving the current rejection of prose fragments and refusing ambiguous basename matches. The linker indexes basenames owned by exactly one full path; validation.ts, types.ts and git.ts name several files here and keep requiring a directory.
• Treat documentation as a plan only when it is prescriptive. Descriptive prose used to raise PLANNED_NOT_IMPLEMENTED: once the deterministic converter emitted document records the count went to 579, of which 574 came from documentation and 555 had modality: unknown. Scoping the rule to required/recommended brought it back to 22.
• Generate TypeScript runtime validators and OpenRouter response schemas from one canonical schema source to prevent manual contract drift.
• Add deterministic converters for repository configuration and infrastructure inputs (JSON/YAML/TOML, Docker and CI workflows) so their declared behavior is represented in DSL instead of only appearing as AST or documentation references. The shipped ignore policy explicitly re-includes .github/workflows/, so the advertised CI conversion is exercised in real repositories, not only in an isolated fixture.

P2 — modularity and operations

• Split the task-synthesis god module into provider orchestration, structured-response contract, grounded materialization and bounded payload selection without changing its public API or fail-closed behavior.
• Make project.sh correct and validate generated README metadata against package.json; external code2docs fallback values (0.1.0, MIT and a TypeScript runtime badge) can no longer be published silently.
• Isolate project.sh generators in a detached tracked-only snapshot, reject references to untracked inputs, temporary paths and unavailable parser downloads, and make source-changing prefact -a explicitly opt-in.
• Formalize the backward-compatible project/ namespace: root-level code2llm/redup output is analysis, while communication requires a recognised ticket directory, participant registry or explicit front matter. This keeps project/analysis.toon.yaml stable without creating agent_log noise.
• Run clean-room deterministic pipelines against three external mixed repositories (code2llm, domd, pactfix) and fix Python ignore-scope overflow plus false-positive project/ communication discovery.
• Re-run clean-room offline pipelines on three further semcod repos (code2logic, code2docs, redup): all succeeded with deterministic code-change planning; fix configuration ??/|| SyntaxError that blocked the batch.
• Split language adapters from src/extractors/ast.ts into independently testable TypeScript/JavaScript, Python, Go, Java and Rust modules behind the existing common adapter envelope.
• Add a first-class dependency-free PHP syntax adapter. It uses token_get_all(..., TOKEN_PARSE), preserves the shared fact envelope and fails open when PHP is absent; A/B on redsl converted 40 tracked files into 2,127 unique records and removed 18 warning diagnostics.
• Add first-class syntax adapters for other languages present in analyzed repositories.
• Report discovered unsupported PHP/Ruby/C#/Kotlin/C/C++ and other source files explicitly instead of implying complete code-to-DSL coverage.
• Add a CI job with JDK 17+ so the Java fixture cannot be skipped in the required validation matrix.
• Add scheduled opt-in live OpenRouter contract checks with redacted audit, latency/cost thresholds and no dependency of offline CI on provider uptime.
• Extend the scheduled live check from two stages to all six, with per-stage and per-run budgets and a recorded result history. It now measures the manifest of a require-llm pipeline run instead of calling selected stages itself, which is how it drifted to two: a bespoke caller cannot go out of date against a pipeline it does not run. History is reported and never gates — t2c.live-contract-check/v2, 50 runs, restored from CI cache and published as an artifact.
• Collect enough scheduled live runs for the recorded median latency and cost per stage to describe anything. The contract and storage exist; the trend does not yet.
• Make make demollm execute and verify all six semantic LLM stages end-to-end, with no deterministic fallback and a manifest-based PASS gate.
• Give task synthesis one corrective retry when the runtime rejects a fabricated citation. Measured before the change: 1 of 6 make demollm runs completed, with LLM_RESPONSE_INVALID sinking three of them on well-formed but non-existent record IDs. The retry quotes the validation error back and keeps both attempts in the audit; a second fabrication still fails the run.
• Extend the corrective retry to the summary stage, which rejected fabricated citations the same way but had no second attempt. Both attempts stay in the audit and a second fabrication still fails the run.
• Remove the remaining make demollm flakiness without accepting invented evidence: derive conclusion recordIds only from the cited diagnostics, continue to reject unknown diagnosticIds. The historical runtime assigned blank response-local proposal keys; ticket-009 now rejects them and relies on the corrective retry because inventing a key changes provider intent. Live contract and full six-stage run 20260730T185205Z-312a0535 passed with generator version 2 before that hardening.
• Bound the task-synthesis payload consistently. compactSynthesisPayload collected record IDs from diagnostics and then truncated records to 500, so a large repository could ship a diagnostic citing a record the model never saw. Diagnostics whose evidence did not survive the record budget are now dropped from the payload instead of inviting a fabricated citation.
• Run a clean-room deterministic pipeline against the external subactor/platform repository (144 Markdown files, 43 configs, .mjs sources) and fix what it exposed: an absolute host path aborted code-change planning, HTTP routes and hostnames were extracted as repository paths and code symbols, and configuration declarations formed a quadratic subgraph.
• Give configuration declarations a file-level aggregate, mirroring module_fact for AST. configuration_file_fact carries a bounded inventory of declared keys, but links as file evidence only through an explicit path. A capability-topic prototype created 288 cross-source links from five files on code2llm and was rejected; the hardened run keeps exactly 2 explicit path links and 0 system~system relations.
• Count configuration as observed reality in Intent-vs-Reality and in implementation diagnostics. A system record is a fact with lifecycle implemented, and for infrastructure repositories the implementation can be the configuration. Semantic impact must be measured again after the capability-topic hardening rather than inferred from the rejected noisy run.
• Reconsider the system lane in the Intent-vs-Reality SVG. The lane was added to LANE_ORDER, but the eight-lane table is now wide enough that the topic column truncates earlier; check whether a combined “evidence” column or a wider viewBox reads better.
• Measure whether counting configuration as evidence hides genuine gaps, and grade alignment by evidence kind. Measured on tracked HEAD: 16 of 46 aligned topics on subactor/platform rest on configuration alone against 4 of 89 here, so an undifferentiated headline reads the same for a repository whose evidence is a third weaker. RealityRow.evidence and totals.alignedByEvidence now report the split; neither changes what counts as aligned, because a behaviour whose implementation is configuration is implemented.
• Scope detectPolarity so “without / bez + dopełnienie” does not negate the governing intent (for example “Document X without inventing files”); covered by the offline gold set and a focused unit test.
• Add A2A streaming/push notifications and a shared transactional task-store backend beyond the current atomic filesystem snapshot.
• Validate public extraction options before they reach path.resolve. Missing root, sourcePath or text now produces a named option error at the public deterministic and audited NL boundaries.
• Guard .github/workflows/ci.yml against duplicate top-level keys. Two concurrent edits produced a second schedule: block, which YAML silently resolves to the last one — verify:workflows now rejects that state before a workflow can silently lose a trigger.

P3 — NL → DSL → sourcecode (plan wdrożenia)

Cel: domknąć pętlę od intencji w języku naturalnym, przez Intent Evidence DSL, do planowanej zmiany kodu i weryfikacji po re-analizie. System pozostaje guardem intencji: LLM może proponować, runtime waliduje, człowiek zatwierdza.

Faza 1 — wiarygodna analiza (fundament)

• Linker: bare basename pliku tylko gdy unikalny w repo (markdown.ts); niejednoznaczne nazwy (validation.ts, types.ts) wymagają katalogu.
• Diagnostyka: dokumentacja jest planem tylko przy modality required/recommended; opisowe prose nie podnosi PLANNED_NOT_IMPLEMENTED.
• Gold v2 (część): capability hard-negative (2 tematy), partial symbol still evidenced; bare basename, multi-clause NL, route/host non-paths, ticket exact-target.
• Gold v2 (reszta): evaluation/gold/v2/dataset.json z kanałem documentation-deterministic (prescriptive vs descriptive w ekstrakcji), zakresem diagnostics (false DONE bez dowodu vs DONE z dowodem, partial implementation jednego ticketu), ośmioma pozytywami capability-topic, czternastoma parami zabronionymi oraz osobnym kohortem cross-language (6 oczekiwanych, 6 zabronionych). npm run evaluate:gold liczy teraz v2; v1 pozostaje pod evaluate:gold:v1.
• Powiększyć próbę capability-topic na tyle, by zmiana progu trzech tematów dawała mierzalny spadek w obie strony. Siedem pozytywów wykrywa regresję progu, ale nie pozwala go stroić.
• Dopasowanie ponad barierą językową, krok pierwszy: słownik dziedzinowy PL→EN wraz z polskimi końcówkami na angielskich zapożyczeniach (ticketu, foundera), odfiltrowanie polskich słów funkcyjnych (nie 175, jest 110 i jako 54 były jednymi z najczęstszych „tematów” korpusu platformy) oraz kotwiczenie deklaracji bez własnego celu w powiązanym module. A/B na tej samej treści: aligned 25 → 43 i coverage 5,9% → 10,0% na subactor/platform, 74 → 82 i 22,0% → 24,4% tutaj.
• Dopasowanie niezależne od ręcznego słownika. Gold v2 mierzy obecnie 0/6 oczekiwanych relacji PL/DE/ES/FR i 0/6 naruszeń par zabronionych. Ticket-004 odrzucił surowy próg embeddingów: E5 uszeregował 6/6 par syntetycznych, ale zaproponował dwa błędne nowe linki na subactor/platform; wzajemny top-1 usunął błędy kosztem zerowego wzrostu pokrycia. Ticket-005 dodał osobno mierzone, ugruntowane kontrakty rerankera i osiągnął 6/6 na przejrzanych fixture’ach, lecz trzy rzeczywiste odpowiedzi Qwen/OpenRouter złamały wymagany kształt lub typ danych. Żadna relacja nie powstała, a eksperyment nie został wyeksportowany produkcyjnie. Następny kandydat musi najpierw dowieść stabilności provider/schema na śledzonym repozytorium, a dopiero potem wzrostu pokrycia.
• Zbadać 69 deklaracji subactor/platform, które dotykają kilku modułów i dlatego celowo nie dostają kotwicy. Rozstrzygnięcie ich wymaga dowodu mocniejszego niż wspólne tematy — bez niego wybór modułu byłby zgadywaniem.
• Zmierzone i odrzucone: wpisy changelogu nie powinny trafić do isModuleTopicSource. Na subactor/platform ze 111 diagnostyk CHANGELOG_WITHOUT_IMPLEMENTATION dopasowanie tematyczne sięgnęłoby dokładnie jednego modułu w 8 przypadkach (7%) przy 15 niejednoznacznych; w tym repo ze 121 diagnostyk — 7 (6%) przy 105 niejednoznacznych (87%). Wpis wydania opisuje zwykle zmianę w kilku modułach naraz, więc dopuszczenie go do dopasowania tematycznego kupiłoby kilka procent mniej diagnostyk za setkę arbitralnych relacji.
• Rozważyć węższy wariant tego samego: pozwolić kotwicy z Intent-vs-Reality (indexModuleAnchors) objąć wpis changelogu, gdy trafia w dokładnie jeden moduł. To 8 i 7 przypadków wyżej, bez wpuszczania niejednoznacznych do grafu — ale zmienia znaczenie CHANGELOG_WITHOUT_IMPLEMENTATION, więc wymaga decyzji, nie samego pomiaru.
• Wzmocnić NL→target: odrzucanie absolutnych/HTTP/traversal path i hostnames; ticket binding w gold exact-target.
• Dalsze NL→target: resolution symboli względem AST bez zgadywania między modułami; missingFields / AMBIGUOUS_REQUIREMENT wskazują konkretne pole i kandydujące ścieżki (ticket-011).
• Jeden source schematów → TS validators + OpenRouter response schemas (patrz też P1).
• Live latency/cost batch TODO/CHANGELOG: ticket-013 porównał trzy modele, wybrał Codestral 2508 i dodał współbieżność 3; weekly 218,741→53,362 ms, a nlp2uri przeszedł 619 rekordów / 20 żądań w 194,750 ms.
• Rozstrzygnąć jakość, nie tylko koszt, przy wyborze modelu dla batcha TODO/CHANGELOG. npm run live:models na 267 rekordach tego repo pokazał, że mistralai/codestral-2508 i google/gemini-3-flash-preview zgadzają się co do action/modality/polarity/lifecycle tylko w 169 z 267 rekordów (63,3%) — tańszy model nie odpowiada tak samo, a porównanie nie mówi, który ma rację. Potrzebne są przypadki gold nad polami wzbogacenia, nie kolejny przebieg live.
• ticket-014 — nie uznawać nowej możliwości za zaimplementowaną wyłącznie dlatego, że wskazany plik istnieje; wymagany dodatkowy dowód symbolu/tematu albo jawna abstencja. Gold ma parę negatyw/pozytyw, a Koru PLF-003 przeszedł wykrycie, patch, pytest i re-analizę.
• ticket-015 — zachować złożoną intencję implement ... and verify ... w tytule code-change; klasyfikacja czasownika pomocniczego nie może tworzyć Implement Implement ... and it ....

Faza 2 — kompletne reality

• Incremental AST + documentation-chunk cache po content hash (patrz P1).
• First-class PHP syntax adapter (patrz P2 i ticket-016).
• First-class adapters dla pozostałych języków w analizowanych repo.
• Opcjonalnie: testy/CI facts jako evidence implemented obok AST+Git.

Faza 3 — DSL → plan zmiany kodu (wdrożenie bieżące)

• Kontrakt t2c.code-change-plan/v1: paths/symbols, evidence (recordIds/diagnosticIds/conclusionIds/proposalIds), acceptance criteria, risk, rollback, status proposed, content-bound ID/hash i runtime provenance.
• Deterministyczne budowanie planów z PLANNED_NOT_IMPLEMENTED / CHANGELOG_WITHOUT_IMPLEMENTATION i powiązanych propozycji TODO (bez auto-apply, bez udawania codegen).
• Filtrowanie celów planu do konkretnych, użytecznych plików: bez katalogów, globów, ścieżek hosta/URI, vendoringu, cache/build, binariów oraz dumpów project/ i artefaktów runu todo2code; SVG, lockfile i zwykła dokumentacja pozostają dozwolone. Audyt własnego repo zmniejszył szum z 9 do 3 planów.
• Kontrakt t2c.code-change-acceptance/v1: re-diagnose po zmianie, cleared vs remaining diagnostics, fail gdy pojawią się nowe blocking.
• CLI/MCP/A2A: propose_code_change / evaluate_code_change na grafie before/after; SDK TypeScript convenience methods.
• Pipeline zapisuje code-change-plans.json w każdym udanym runie (etap deterministyczny codeChangePlanning w manifeście).
• Deterministyczny review brief CODE_CHANGE.review.md + audit t2c.code-change-review/v1 (hash-bound, bez apply źródeł); CLI render-code-change, MCP/A2A/SDK i zapis w pipeline.
• NL/target hardening: odrzucanie HTTP routes, host paths, .. oraz hostnames jako fałszywych symboli (gold + unit).
• Structured source patch t2c.code-change-source-patch/v1: instrukcje per path, opcjonalny unified diff z walidacją path/sekretów, content-bound hash; CLI propose-source-patch, MCP/A2A/SDK i zapis w pipeline; propozycja sama nigdy nie wykonuje apply.
• Apply source patch po hash approval (apply-source-patch / MCP): tylko gdy każdy edit ma unifiedDiff; instruction-only jest odrzucany; receipt idempotentny, ma runtime provenance i weryfikuje stan plików przy powtórzeniu; pełny preflight, ochrona symlink/root i rollback chronią przed częściowym zapisem. LLM wypełniający diff — nadal opcjonalna przyszła gałąź.
• detectPolarity nie traktuje „without / bez + dopełnienie” jako negacji całego zdania (gold + unit).
• Orkiestracja acceptance: t2c close-code-change (plan lub plan-set + before/after graph) → t2c.code-change-close-result/v1 bez auto-DONE.
• detectModality nie traktuje nagłówków „(recommended)” / gołych przymiotników jako deontycznych (mniej fałszywych PLANNED_NOT_IMPLEMENTED na code2logic).

Faza 4 — ops / multi-agent

• A2A streaming/push + transakcyjny task-store (patrz P2).
• Branch protection / work permits jako warstwa integracyjna repo/CI (poza core runtime).

Definition of done (ścieżka end-to-end)

1. NL/ticket → Intent DSL z provenance.
2. Multi-source graph + diagnostyka Intent vs Reality.
3. Code-change plan ugruntowany w diagnostyce (nie free-form).
4. Po zmianie kodu re-analiza zamyka targeted diagnostics bez nowych blocking.
5. Brak auto-DONE: człowiek lub uprawniony klient zatwierdza.

This site is open source. Improve this page.