Implementation status

This repository implements the local deterministic part of the Cohorte V3 specification. It is a pre-release foundation, not a V2-parity or provider-support claim.

Implemented and locally testable

• Python 3.12 package and cohorte CLI.
• Strict Pydantic contracts for profiles, specs, task plans, runs and protocol envelopes.
• Pure workflow reducer with evidence-gated advancement, pause, resume, cancellation, fix and uncertain-effect states.
• SQLite WAL store with foreign keys, optimistic run updates, immutable hashed artifacts, durable events, requests, decisions and idempotent operations.
• Read-only project discovery for basic Python, JavaScript and workspace layouts.
• DAG and coverage validation, bounded argv-only check execution, filtered environment and output.
• JSON-RPC stdio handshake, bounded frames, structured errors, deduplication and replay.
• Passive Codex auth inspection and an explicit bounded live probe using the pinned SDK/runtime.
• Live Codex evidence for subscription routing, read-only enforcement, interruption and resume on Darwin arm64 with SDK/runtime 0.155.1.
• A loop vertical with a frozen spec, deterministic task plan, isolated Git worktree, path ownership enforcement, bounded checks, independent read-only review, bounded fix cycles and candidate-bound ship evidence.
• Live G1 vertical evidence on a disposable repository: workspace-write implementation, two passing standard-library tests, complete surface review and a blocker-free ready_to_ship verdict. The final run is re-readable from SQLite as SHIP / waiting_user, with a candidate-bound ship request.
• Durable phase checkpoints and resume, pause and cancel controls. A live controller was interrupted during review, then resumed from the same worktree and candidate hash. The recovered journal contained one build, two safe check executions, one completed review and one ship request.
• G2 delivery core: candidate-bound approve/deny, freshness and remote-base preflight, marked Git commit, non-force push, GitHub PR and GitLab MR adapters, CI status refresh, and durable effect reconciliation. Local bare-remote tests prove commit/push identity and PR-created-before-ack recovery without a duplicate.
• Live GitHub G2 evidence: private disposable repository, exact approved candidate committed and pushed, PR confirmed open with a clean merge state, and all three delivery effects reconciled in SQLite. A second run based on a real GitHub Actions workflow reached ci_passed with the unit check confirmed successful.
• G3 multi-surface execution: criteria are mapped to a validated task DAG, contract dependencies run before consumers, independent tasks use isolated worktrees and bounded parallel Codex sessions, and Cohorte creates and serially integrates marked task commits before global checks.
• Live Codex G3 evidence on Darwin arm64: one contract task followed by concurrent backend/client tasks (max_parallelism=2), five integrated files, four passing unit tests, per-surface plus integration review, and exact-hash durable evidence ending at SHIP / waiting_user.
• Durable mid-wave recovery: SQLite stores task attempts, generation-fenced leases, worktree and branch identities, produced commits and integration commits. A live controller-crash scenario recovered the still-running client task from its existing commit, completed with exactly three attempts and zero remaining leases, then passed checks and the bounded review/fix/re-review cycle.
• Provider auth expiry and exhausted quota suspend runs in waiting_auth or waiting_quota while preserving their stage and candidate. Transient overload retry is limited to two attempts.
• Active task leases cannot be replaced before expiry. A POSIX SIGKILL qualification leaves a worker alive after controller death, rejects concurrent recovery, then resumes only after worker termination and lease expiry.
• A live Codex interrupt probe observes a marked descendant before interruption and proves it is gone afterward. A killed reviewer yields TransportClosedError and no accepted review.
• G3 Fleet: cross-feature write-set overlap matrix, derived and explicit dependencies, bounded parallel feature waves, isolated feature worktrees, serialized integration, per-feature revalidation after every base change, global checks and cross-feature review.
• Live Fleet evidence: API and Web ran in parallel, the overlapping API follow-up ran in the next wave from the integrated API+Web head, all three revalidation checkpoints passed, six global unit tests passed and the final review covered both surfaces with a ready verdict.
• Intake and Patch workflows: text, file and URL sources retain their locator, timestamp and content hash; deterministic triage routes explicit reproduced bugs to Patch and ambiguous reports to questions. Ticket content remains untrusted data. Patch freezes a minimal, source-linked scope, narrows write ownership, requires an existing automatic regression to fail, then reuses the isolated build/check/review/ship pipeline.
• Live Patch evidence: an immutable intake source became a frozen patch; a committed regression failed with 6 != 5, Codex changed only calc.py, the same regression passed, review returned ready, and SQLite ended at SHIP / waiting_user on the exact candidate hash.
• Audit/refactor/retro deterministic core: bounded Audit rejects any source-tree mutation and emits a prioritized backlog; Refactor requires an exact-hash persisted approval, a green behavioral baseline, narrowed paths and the common build/check/review/ship loop; Retro extracts repeated finding fingerprints and updates the profile only after a proposal-bound approval decision.
• Live maintenance evidence: two read-only audits retained the exact source-tree digest and emitted the same deterministic duplicate-branch fingerprint; the exact-hash approved refactor kept its regression green before and after changing only calc.py, passed independent review and reached SHIP / waiting_user. Retro then required its own proposal-bound decision before producing profile revision 2 with the ratified convention.
• Design/retrieval ports and local adapters: disabled design is explicitly skipped; file design sources produce immutable versioned snapshots; inaccessible configured design is blocked. Retrieval searches bounded file roots when disabled or configured for files, exposes configured provider failures, and labels an allowed file fallback with its reason.
• Align-ds local workflow: compare the captured source with the committed JSON snapshot, persist exact add/update/remove deltas, bind approval to the plan hash, narrow write ownership, then reuse the durable build/check/review/ship loop.
• Live align-ds evidence: four exact deltas from a versioned local design source changed only design-snapshot.json; the declared design check passed, independent review returned ready, and the run reached SHIP / waiting_user on the candidate hash.
• Obsidian Kanban projection: one configured board path, exact-hash concurrency guard, backup before atomic replacement, stable feature markers and durable projection versions. A retry is a no-op, stale state is rejected and no vault scan is performed.
• Metrics report: run outcomes, elapsed duration, human interventions and fix cycles are computed from SQLite, with project, run, phase and provider groups. Provider tokens/cache and estimated cost carry explicit availability; missing usage remains null rather than becoming zero.
• Opt-in V2 migration: an inspectable hash-bound plan lists exact files, mappings, exclusions, losses, warnings and ambiguities. Credentials are excluded, YAML/frontmatter remains inert data, symlinks are rejected, imported metadata is historical/non-certified, old active runs are never resumed, and apply rechecks every source hash before one transactional import. The CLI creates and verifies a database backup and supports an exercised rollback.
• Monotonic SQLite schema upgrades: schema 1 upgrades to schema 2 under an exclusive transaction after a verified backup, existing data/configuration remains intact, and a future schema is refused without a silent reset. An installed 0.1.0a1 wheel is upgraded in place to 0.1.0a2 in hosted Ubuntu, macOS and Windows CI on Python 3.12/3.13; external configuration, project data and artifact bytes are compared across the upgrade.
• Design/retrieval availability: absent external connections return explicit CLI errors; runtime port failures block design capture or trigger only an explicitly configured retrieval file fallback. Provider error text is redacted before it enters captures or CLI responses. No external design/retrieval provider is claimed as live-qualified.
• François-facing stdio contract: capability-negotiated handshake, project/feature/request read models, grouped metrics, run controls, request responses and durable event replay/unsubscribe. Passive account status exposes no credentials. This is the Cohorte protocol surface; the external François UI remains unvalidated.
• Unified agent events: Claude and Codex workflow turns persist agent.turn.started, agent.turn.finished, agent.tool and agent.usage with the same provider/phase/access fields; SDK permission-denial counts use agent.permission.denials. The SQLite journal and events.subscribe expose the same content-free data. Tool decisions are observed at Claude's pre-tool hook and from Codex's terminal item list; an unknown decision stays unknown. Token counts follow each SDK's accounting, cache tokens mean cache reads, and Claude's estimated_cost is not a billing fact. Live single-surface workflows on both providers persisted build/review usage and tool events on Darwin arm64. Brainstorm, Audit and Fleet use project-scoped events when no durable run ID exists; see docs/evidence/g3-unified-agent-events-darwin-arm64.json.
• Persistent POSIX service host: the CLI starts, probes and stops a background Unix-socket host; filesystem permissions and kernel peer credentials restrict it to the local user. Connections reuse the same RPC implementation, enforce bounded frames and writer drain timeouts, and a client disconnect leaves durable workflow state unchanged. Durable replay continues into live event notifications without a query gap. This path is validated on Darwin arm64.
• Windows named-pipe backend: user-SID-only DACL, remote-client rejection, single-service mutex, bounded frames and the same start/status/stop RPC lifecycle. The Windows-only integration test is green in the Python 3.12/3.13 GitHub-hosted Windows CI matrix. This validates the runtime lifecycle, while release support still requires real-host and load-level qualification.
• Claude Agent SDK adapter for brainstorm, build, review and fix, selected by the project profile (or brainstorm --provider claude). Native CLI auth status is inspected passively; ambiguous API, gateway and token overrides fail closed under subscription_only. SDK tool lists and pre-tool path guards restrict reads and edits to the worktree. The adapter now observes durable run pause/cancel state during active turns and calls native interrupt(). Live CLI probes stopped an active run as cancelled and paused then resumed another run to waiting_user. A live Claude no-tool smoke, structured read, workspace edit, outside-write denial, guarded read-only write denial and single-surface build/check/review workflow passed on Darwin arm64 after the user connected a new account. The prior account was refused by its organization policy, and Cohorte did not switch to API billing. See docs/evidence/g0-claude-sdk-darwin-arm64.json, docs/evidence/g1-claude-vertical-darwin-arm64.json and the earlier blocked-account evidence.
• Live brainstorm preparation: product, architecture and QA run in distinct ephemeral read-only Codex sessions over one factual bundle, followed by a separately identified synthesis session. The brief retains contribution references, divergences, strong objections and user answers in a content-addressed SQLite artifact; only user answers become decisions. Invalid contribution references receive at most two bounded correction attempts.
• Complete spec freeze workflow: draft schema and semantic completeness, surface/check/reference resolution and task-plan coverage are verified before an approval request is created. The user decision targets the canonical hash of the exact final frozen document and current profile; any later draft, profile, reference or plan change refuses the freeze. The CLI writes the approved document atomically and advances the durable feature status to frozen.
• Machine-checked AC01–AC30 qualification ledger with explicit passed, partial, blocked, deferred and not-started states. The first tranche adds a clean wheel-install verifier, multi-stack discovery coverage and an out-of-scope ownership rejection regression.

Deliberately unverified or incomplete

• AC14 passed on a disposable GitLab repository with a bounded SHIP fixture, configured release notes in the MR description, and a successful CI pipeline; see docs/evidence/g5-ac14-gitlab-darwin-arm64.json.
• AC19 remains partial: the merged François client still targets the older TypeScript CLI and needs adaptation and live validation against the Python cohorte/1 service.
• Native mid-turn session reuse and exactly-once effects after a provider crash are not established. AC30's read-only retry proof is bounded to the Darwin reviewer-style probe; other hosts and the full review workflow have not received the same live negative-tool test.
• AC23 passed on bounded local evidence: Serena and Graphify-Labs MCP searches, a user-run Figma CLI snapshot, and an agent-run Figma invalid-token rejection. Larger projects, other hosts, connector process outages, and a least-privilege Figma token remain hardening work.
• Windows real-host and load-level slow-client behavior remain outside the hosted CI proof.