Cohorte

Appearance

Shipped scripts ​

Installed to <core>/pipeline/scripts/ (~/.claude global, .claude bundled). Two families: shipped executables used as-is by the commands, and templates rendered per project by /cohorte-init-pipeline. Every call site chains them with || true, so a missing script is a silent no-op — /cohorte-doctor check 1 is what catches a half-copied core.

preflight.sh — the deterministic phase gate ​

sh
preflight.sh <report-file> "<cmd>" ["<cmd>"...]

Runs each command through sh -c, all output appended to <report-file> (the bulk never enters any agent's context). First failure ⇒ prints the last 40 lines of the report raw to stderr, exit 1 — the caller (/cohorte-review §0, the workflow preflight phases) must stop there and spawn no agents. All green ⇒ writes <state>/preflight.ok (<epoch> <HEAD sha> <tree digest>), the stamp gate.py checks before letting review dispatches through. It is a local, gitignored file — never commit it (see gate.md).

kanban-move.sh — board updates outside agent context ​

sh
kanban-move.sh auto <feature_id> <stage> [--pr <num>] [--title <title>]
kanban-move.sh <board.md> <feature_id> <column> [--pr <num>] [--title <title>]
kanban-move.sh --check

Moves (or creates) the #<feature_id> card under the target ## <column> heading — find, dedupe (keeps the first), sub-notes carried along, the %% kanban:settings %% block preserved, — PR #<num> appended with --pr. Exists so no agent ever reads a whole board (which grows with every feature ever tracked) into context.

auto resolves the board itself — name from PIPELINE.md, then kanban.enabled / obsidian.vault_path / boards[name] from ~/.claude/cohorte.config.yaml (override with COHORTE_CONFIG; skip the profile with --project <name>) — and maps a stage key (ideas, brainstorm, spec, ready, building, review, fix, ship, shipped) to that board's heading via boards[name].columns → kanban.columns → the built-in default. An explicit board path and a literal heading still work.

When nothing resolves it prints kanban: <reason> — naming the missing link — and exits 0. A board that is configured but can't be moved is loud instead: exit 2 usage, exit 3 missing board file or unknown column. Callers report which line they got; none of them may infer "no board" on their own. --check does the resolution and nothing else, which is what /cohorte-doctor and /cohorte-update-pipeline use to tell "not configured" from "configured and broken".

new-feature.sh / remove-feature.sh — worktree isolation (templates) ​

Shipped as .template files; /cohorte-init-pipeline renders them to scripts/ in your repo, substituting the __TOKENS__ (project slug, DB pattern, port bases, compose file, branch prefix, install/dev/migrate commands, per-surface env stanzas).

  • new-feature.sh <id> — creates the sibling worktree ../<slug>-<id> on branch <prefix><id>, assigns the next free slot in .worktrees/slots.tsv, creates the per-feature DB (when db_per_worktree), writes the env with port_base + slot, runs install + migrate.
  • remove-feature.sh <id> [--drop-db] — removes the worktree, deletes the merged branch, frees the slot; keeps the DB unless --drop-db.

Repo-side helpers (not installed) ​

  • scripts/validate-core.mjs — CI's structural validator for the core itself: command/agent frontmatter + model pins, template placeholders, cross-references, the no-telemetry ratchet, installer coverage of every shipped script, packaging negations, and workflow-script invariants (parses each core/workflows/*.js exactly as the runtime does — async body wrap — plus meta literal, profile-reader phase 0, no Date.now()).
  • scripts/test-workflows.mjs — behavioural tests for core/workflows/*.js: runs each script with stub agents and asserts the returned verdict object. Structural checks cannot see verdict logic, and one failure mode needs exactly this — agent() returns null when a subagent dies, so a crashed reviewer yields zero findings, indistinguishable from a clean surface. review.js scored that as SHIP until this test existed; the same suite now pins the loop workflow's reducer (loop.js) and the refactor retry round.
  • scripts/test-gate.mjs — behavioural tests for hooks/gate.py, driving its real stdin→stdout contract with PreToolUse payloads: the deny/ask tiers, chained-command splitting (&&, ;, |, ||, newlines), branch-conditional gating resolved at the payload's cwd, the bypassPermissions ask⇒deny escalation, config robustness (missing/unparseable ⇒ silent), the preflight phase gate (missing/stale/HEAD-moved/unreadable stamp), and worktree-aware HEAD matching. The gate is the only component that can block a user's command; py_compile proves only that it parses.
  • scripts/test-adapter.mjs — tests for the runtime adapter (core/adapter/render.js + core/runtimes/*.json), which turns one set of source prompts into five runtimes' worth of commands and agents. Its failures are silent by construction — a dropped conditional or a leaked marker still installs cleanly and only misleads the model at run time — so this installs for every runtime into a scratch home and asserts what lands on disk: no unresolved markers, the dispatch doctrine matching that runtime's capabilities, the gate described as blocking or advisory according to whether it really is, only supported frontmatter keys surviving, the reviewer's read-only restriction re-emitted in that runtime's spelling, and Claude Code's output unchanged.
  • scripts/test-lib.mjs — tests for lib/*.js, the shipped runtime code behind cohorte doctor and cohorte specs: the block-YAML parser (against the real PIPELINE.template.md), the JS port of /cohorte-doctor (each check green, then each one made to fail), and the runtime-layout resolver — including a non-Claude layout and a registry whose absolute paths point at someone else's checkout, both of which used to report a healthy install as several false reds.
  • scripts/assert-gate-hook.mjs — post-install assertion on a settings.json: exactly one gate.py registration (callers install twice first), with a matcher covering both Bash and Task. Both invariants are shipped regressions.
  • .github/workflows/ci.yml — runs the validator, syntax-checks CLI/lib/hook/installers, and dry-runs both installers into scratch homes, asserting the exact postconditions (agents present, scripts executable, workflows shipped, templates-not-rendered).
  • .github/workflows/publish.yml — publish-on-main: version-diff check against the registry, a required CHANGELOG entry, npm publish --provenance (token-less OIDC trusted publishing), tag + GitHub release.