Cohorte V3

Install the Python CLI from PyPI (not the legacy npm package):

uv tool install 'cohorte-engine==1.0.0a3'
cohorte --version

User documentation: VitePress site (sources). Only to preview the documentation site locally, run npm ci --prefix docs && npm run dev --prefix docs. npm does not install the Python CLI.

For Python dev releases, see the release runbook.

Cohorte is a local, evidence-driven workflow engine for official coding-agent clients. This repository is a Python 3.12 rewrite built at the repository root. It does not depend on the former TypeScript/Pi implementations.

The current pre-release provides the deterministic core, bounded feature and Patch workflows, and overlap-aware Fleet execution: strict contracts, a pure workflow reducer, SQLite persistence, immutable artifacts, project discovery, DAG validation, isolated Git worktrees, controlled checks, independent read-only review, a review/fix loop, a JSON-RPC stdio bridge, and a CLI. Codex authentication and live execution are capability-gated. A Claude Agent SDK adapter is available through the optional claude dependency and agent_defaults.provider: claude; a bounded single-surface Claude workflow has passed live build, checks and review on Darwin arm64. Passive auth status claude checks the native CLI without exposing credentials. An explicit auth verify claude --live probe is available after confirming the account to use. The connected native account passed the Claude SDK smoke, structured read, workspace edit, outside-write denial, a guarded read-only write denial, and CLI cancellation/pause/resume probes. G0 remains partial: the full login interaction and safe native session recovery inside a workflow are not yet qualified. Both adapters now persist common turn, tool and usage events for workflow runs. The event payloads contain provider, phase, access and bounded metadata, without prompts, responses, commands or file paths. SDK token accounting can differ, and Claude's reported cost is an estimate rather than a provider billing statement.

uv sync --all-extras
uv run cohorte doctor
uv run cohorte --json init /path/to/project
uv run cohorte --json status
uv run pytest

From a project directory, the guided CLI needs no project ID or JSON flags:

cohorte init .
cohorte profile show
cohorte profile edit
cohorte status
cohorte intake
cohorte brainstorm
cohorte brief show FEATURE_ID
cohorte spec
cohorte start

init inventories pnpm/npm workspaces, including nested package patterns, internal dependencies, root checks and shared files. It returns the stored profile on repeat runs. Review its questions and edit the stored JSON profile before relying on it for a workflow; cohorte init . --refresh explicitly replaces edits with a fresh discovery. cohorte profile apply reviewed.json applies a validated profile from a file, with a revision check. The guided brainstorm asks for the idea, audience, observed problem, desired outcome and constraints, then prints a short synthesis. Its full brief remains stored in Cohorte and can be read with cohorte brief show FEATURE_ID without rerunning the panel. For scripts, use cohorte --json brief show FEATURE_ID to read the complete stored document; new brainstorm runs still require explicit flags and --live. The generated commands are candidates: confirm service setup and migrations before running project-wide checks in a monorepo.

status shows the current project's features, runs and pending run decisions. intake can capture a short text, file or URL through terminal prompts and shows its triage result. spec uses a stored brainstorm brief to collect a single-surface draft, keeps unanswered questions open, then shows the exact candidate hash before an explicit freeze decision. start verifies the frozen spec and the approved profile snapshot before asking to launch a live worktree run. These two commands require an interactive terminal; the explicit file-based commands remain available for automation and multi-surface specs. See the CLI experience audit for the remaining UX gaps. The guided commands were introduced in 1.0.0a3. Verify the installed version with cohorte --version before using them in another project.

For external context, set integrations.retrieval.provider to serena or graphify in the project profile. Serena needs the installed serena MCP executable. Graphify-Labs needs the graphify optional extra and a prebuilt <repo>/graphify-out/graph.json; for a code-only graph, run graphify extract <repo> --code-only --no-cluster --out <repo> explicitly before retrieval. Both providers fail visibly when unavailable; file fallback requires integrations.retrieval.fallback_to_files: true. Figma design snapshots use a file or node URL in integrations.design.source and a locally supplied FIGMA_ACCESS_TOKEN with file_content:read scope. The token is never part of the profile. To verify a real Figma snapshot without printing its contents, run python tests/live/verify_figma_snapshot.py --source <file-or-node-url> in a shell where the token is already set, or enter it at the hidden prompt.

Prepare a feature with independent product, architecture and QA sessions, then approve the exact completed spec before freezing it:

cohorte --json --data-dir /path/to/data brainstorm PROJECT_ID \
  --feature-id safe-export --idea "Add a safe run export" \
  --answer "Keep data local and require atomic output" \
  --repo /path/to/project --output brief.json --live
cohorte --json --data-dir /path/to/data spec-freeze-request draft.json \
  --profile project.json --repo /path/to/project
cohorte --json --data-dir /path/to/data approve REQUEST_ID
cohorte --json --data-dir /path/to/data spec-freeze draft.json \
  --profile project.json --repo /path/to/project \
  --decision-id DECISION_ID --output frozen.json

The brainstorm keeps each native session reference, contribution and disagreement. Freeze rejects incomplete drafts and approvals for a different spec hash, profile, reference set or generated plan.

Run the disposable Codex vertical with the included frozen example:

cohorte --json loop examples/g1/spec.json \
  --profile examples/g1/profile.json \
  --repo /path/to/disposable/repository \
  --worktrees /tmp/cohorte-worktrees \
  --run-id demo-1 \
  --live

Run several frozen feature specs as an overlap-aware fleet:

cohorte --json fleet specs/feature-a.json specs/feature-b.json \
  --profile project.json \
  --repo /path/to/disposable/repository \
  --worktrees /tmp/cohorte-fleet-worktrees \
  --fleet-id release-train-1 \
  --live

Fleet computes the cross-feature dependency graph, parallelizes disjoint features, serializes overlapping write sets, integrates each candidate in order and reruns its declared checks after the base changes.

Capture a ticket, freeze a bounded patch, and run it with a pre-existing regression:

cohorte --json --data-dir /path/to/data intake PROJECT_ID --file ticket.txt
cohorte --json --data-dir /path/to/data patch-spec \
  --source-artifact-id SOURCE_ID --source-revision 1 \
  --profile project.json --patch-id fix-total --title "Fix total" \
  --reproduction "Run the regression test" --observed "It returns 6" \
  --expected "It returns 5" --surface python --write-path calc.py \
  --check regression --in-scope "Correct total" --rollback "Revert calc.py" \
  --output patch.json
cohorte --json --data-dir /path/to/data patch patch.json \
  --profile project.json --repo /path/to/repository \
  --worktrees /tmp/cohorte-patch-worktrees --run-id patch-1 --live

Patch refuses to build unless every declared automatic regression is red on the original commit. The agent can edit only the frozen patch paths; the same checks and an independent review then gate the candidate-bound ship request.

The command creates a branch and isolated worktree, but does not commit, push, open a pull request, or modify the source checkout.

Runs checkpoint every completed phase in SQLite. After an interrupted controller process, resume the same worktree with:

cohorte --json --data-dir /path/to/data resume RUN_ID --live

pause RUN_ID and cancel RUN_ID are cooperative: an active provider turn finishes, then Cohorte records the completed phase boundary and stops before starting another phase.

Delivery remains separately authorized:

cohorte --json --data-dir /path/to/data approve REQUEST_ID
cohorte --json --data-dir /path/to/data ship RUN_ID --live
cohorte --json --data-dir /path/to/data delivery-status RUN_ID --live --watch

ship rechecks the candidate hash and remote base, creates a commit with a run marker, pushes without force, then confirms the GitHub PR or GitLab MR through the provider CLI. Every effect is journaled before execution and reconciled on retry. It never merges or deploys.

When integrations.release_notes.enabled is true in the project profile, ship appends a release-notes section to the PR/MR description. Its optional heading defaults to Release notes; its optional template defaults to {title}\n\n{problem}. Templates may use only {title}, {problem}, and {acceptance} (a bulleted list). Notes are omitted when the integration is disabled, and do not modify the reviewed candidate.

Inspect and apply a bounded V2 metadata migration with an explicit rollback point:

cohorte --json --data-dir /path/to/data migrate \
  --from-v2 /path/to/v2-copy --plan migration-plan.json
cohorte --json --data-dir /path/to/data migrate --apply migration-plan.json
cohorte --json --data-dir /path/to/data migrate --rollback /path/from/apply/backup.bak

The plan contains exact hashes, mappings, exclusions, losses and ambiguities. Apply fails if a source changed, imports no credentials or active run, and stores specs as historical artifacts that require V3 validation before build.

Run the persistent local service used by protocol clients:

cohorte --json --data-dir /path/to/data service start
cohorte --json --data-dir /path/to/data service status
cohorte --json --data-dir /path/to/data service stop

On POSIX this uses a private Unix socket and verifies the connecting process belongs to the same user. On Windows it uses a local named pipe with a DACL restricted to the current user SID. No network port is opened. The lifecycle passes on GitHub-hosted Windows 3.12 and 3.13; release support still requires real-host and slow-client qualification.

By default, Cohorte stores configuration and state outside target repositories using platform standard directories. Pass --config-dir and --data-dir for isolated automation. Cohorte never stores provider tokens.

See the implementation status, qualification matrix and protocol reference.