Runtime compatibility (read first)

These instructions apply to Claude Code, Codex and DeepSeek Harness (dsh). Resolve CYCLER_ROOT from the environment, CLAUDE_PLUGIN_ROOT, or this skill's directory (../..). CLAUDE_PLUGIN_ROOT in the legacy instructions below is a compatibility alias for that root; CLAUDE_CODE_SESSION_ID may be replaced with CYCLER_SESSION_ID or the current runtime session ID. Use the tools available in your runtime; ToolSearch is not a prerequisite for research or Linear.

When the native Workflow tool is unavailable (Codex, dsh, or a Claude install without it), run the SAME lifecycle through the portable runner and report its result:

node "$CYCLER_ROOT/workflows/run.mjs" --runtime '<claude|codex|dsh>' --cwd "$PWD" --workflow workflow-feature --task '<complete request or issue key>'

Pass --contract <path> to reuse a contract; --no-commit leaves changes for inspection. Do not continue into the native Workflow(...) instructions after this runner completes. The runner owns independent stages, gate enforcement, GitHub comments and implementer replies.

/workflow-feature — full task lifecycle in a single workflow

Inputs

• The user's request (the current message) — or an existing contract path.
• The request may be a Linear issue URL or bare key instead of a description (see below).
• READ ${CLAUDE_PLUGIN_ROOT}/harness/HARNESS.md first.

Actions (in order)

1. Run git status --short (respect in-flight changes — do not touch them).
2. If the argument is a Linear reference — a https://linear.app/<workspace>/issue/APL-1/... URL or a bare key like APL-12 — resolve it before invoking the workflow:
   • Read it with "${CLAUDE_PLUGIN_ROOT}/poller/lin" issue view <KEY> — one Bash call, no tool schema to load. Prefer this over the Linear MCP: ToolSearch + get_issue costs a schema load and a tool round-trip, and the MCP server is not always connected (in dispatched sessions the Linear MCP never finished connecting, so the follow-up-filing step there was unsatisfiable by construction).
   • lin is linear-cli reading the OAuth token the Linear poller keeps refreshed. Do not use a pinned LINEAR_API_KEY or .linear.toml: Linear access tokens expire every 24h, so a copied one works for a day and then fails with a 401 that reads like a network fault.
   • Pass the issue's title + full description as task (e.g. task: 'APL-1 — <title>\n\n<description>'). The description is the spec: the issues in this workspace are written with Goal / Evidence / Acceptance sections, which map almost directly onto the contract template. Do not paraphrase or summarise the acceptance criteria — pass them through.
   • Pass the key as issueId. It drives the branch (claude/APL-1), the contract filename (apl-1-<slug>.md), the commit subject, and the PR body's Closes APL-1.
   • If lin fails (no token, revoked app), pass the reference through as-is: the workflow's CONTRACT AUTHOR resolves a reference-only request itself as a fallback. Say which path you took.
   • A non-Linear argument is unaffected — pass the request verbatim exactly as before.
3. Post a start comment before invoking the workflow (Linear reference only):

"${CLAUDE_PLUGIN_ROOT}/poller/lin" issue comment list <KEY> | grep -q 'workflow run started' \
  || "${CLAUDE_PLUGIN_ROOT}/poller/lin" issue comment add <KEY> --body "'feature' workflow run started in session \`${CLAUDE_CODE_SESSION_ID:0:8}\`."

Every other Linear write in this skill happens in step 5, after the workflow returns — which is 20-40 minutes later. Without this one the board shows nothing at all for the whole run, and a delegated issue sitting silent is indistinguishable from a dispatch that never fired. That ambiguity is what the poller's own failure comment exists to remove; the run itself needs the same. The grep makes it idempotent, and a failure here is never a reason not to start the workflow.

4. Read the cycler config first, so the run uses this repo's settings rather than defaults:

node "${CLAUDE_PLUGIN_ROOT}/harness/read-config.mjs" repo.base main
node "${CLAUDE_PLUGIN_ROOT}/harness/read-config.mjs" repo.branch_prefix claude/
node "${CLAUDE_PLUGIN_ROOT}/harness/read-config.mjs" repo.path ""
node "${CLAUDE_PLUGIN_ROOT}/harness/read-config.mjs" --json
echo "${CLAUDE_PLUGIN_ROOT}"

Pass those as prBase, branchPrefix, cwd, config and pluginRoot below. A repo with no config gets the defaults, which are the values shown in those commands — nothing breaks, nothing is silently shaped like somebody else's repo.

cwd and pluginRoot are not optional. The Workflow runtime exposes args, agent, parallel, pipeline, log, phase and budget — there is no process, so the workflow has nothing to fall back on and throws immediately if either is missing. It used to fall back to process.cwd(), which is a ReferenceError in that runtime: the script died on its first line and every unattended dispatch left no branch and no PR.

The workflow's audit and review stages are subagents, and you must let them run. A repo may carry a rule against spawning subagents to work an issue — that rule is about not spawning a local agent instead of delegating a separate issue, and it is never about your own run's auditor, review lenses or refuters. Those are the only readers in the run that did not write the diff.

If any of them genuinely could not run, the run is not complete: say so in the PR body and in the Linear comment, in those words — "review self-performed, not independent" — so nobody reads an unreviewed diff as a reviewed one. Do not quietly substitute your own judgement for the lenses and report it as a clean review.

5. Call the single workflow:

Workflow({ scriptPath: '.claude/workflows/task-orchestration.js', args: {
  githubReview: true, // shared standalone harness: publishes GitHub review threads and implementer replies
  task: '<the user request verbatim — or, for a Linear reference, the resolved issue title + description>',
  cwd: '<repo.path from the cycler config — REQUIRED, the workflow throws without it>',
  pluginRoot: '<the ${CLAUDE_PLUGIN_ROOT} you echoed above — REQUIRED, the workflow throws without it>',
  config: <the parsed --json output above; omit only when there is no config file>,
  executorModel: '<lower Claude model id like "sonnet"/"haiku", or omit to inherit>',
  noCommit: <true if the user wants to review before commit; omit for full automation>,
  stopAtContract: <true when plan mode is active; omit otherwise>,
  prBase: '<repo.base from the cycler config, default "main"; pass the previous task\'s branch when stacking tasks>',
  branchPrefix: '<repo.branch_prefix from the cycler config, default "claude/">',
  sessionId: '<output of: echo $CLAUDE_CODE_SESSION_ID — names this session in the board comments>',
  issueId: '<the tracker issue key when the task is driven from one, e.g. "APL-10" — sets the branch name>',
  branch: '<explicit branch override; omit unless you need to force a specific name>',
  models: <per-stage overrides, e.g. { verify: 'haiku' }; omit to use the defaults below>,
  worktree: <true to run in a dedicated git worktree — REQUIRED when launching runs in parallel>
}})

• If a contract already exists and the user wants to use it, pass contractPath instead of task.
• When the user invokes /workflow-feature <ISSUE-KEY> (e.g. /workflow-feature APL-10), pass that key as issueId. The workflow also parses an uppercase ABC-123 key out of the task text, so this is belt-and-braces — but pass it explicitly whenever you know it. It determines the branch name (<branchPrefix><ISSUE_ID>).

6. Perform the Linear writes the workflow planned. The result carries linearWrites[] — the workflow no longer performs these itself (it used to spend 3-4 subagents per run on writes that failed against an unauthenticated MCP; see HARNESS.md -> Linear round-trip). You hold a working Linear connector, so you do them:
   • Walk the array in order. Each entry has kind, issue, stateType, statePreference, assignSelf, links[], body.
   • Use ${CLAUDE_PLUGIN_ROOT}/poller/lin, not the Linear MCP (see step 2 for why): lin issue comment list <KEY> to check, lin issue comment add <KEY> --body '...' to write, lin issue update <KEY> --state '<name-or-type>' to move state.
   • Idempotency: list the comments first and skip a write whose body is already present as a comment with identical text. Never add HTML comments or hidden markers to a body. A resumed or re-run workflow must not spam.
   • States by TYPE: lin issue update takes a state by name or type, so pass statePreference when set and fall back to stateType (started/unstarted/completed). Never hardcode a name that only exists in one team.
   • Never move the issue to Done — the harness does not merge, so only the human who merges can honestly close it. Never reassign an issue that already has an assignee.
   • Never fail the task over Linear. If a write errors, report it and move on; the task's own outcome is unaffected. Entries with skipped: true need no action — the note says why.
   • Tell the user which writes you performed, and which failed.
   • Post nothing else on the issue. No answers to questions, no recaps, no next steps: the human answers in this session, and the board only carries the workflow's own one-line comments.
7. Read the returned structured result: { status, stage, branch, branchCreated, worktree, lockHeld, cleanupNote, contractPath, changedFiles[], summary, commandJournal[], audit, report, confirmed[], notes, fixLog[], commit, commits[], pr, linearWrites[] }. worktree is null once removed, or a path still on disk. lockHeld: true means the shared-tree lock could NOT be released — remove it by hand (path in cleanupNote) or the next shared-tree run refuses to start. Note: the workflow self-corrects — audit/verify/review failures loop back to the implementer with the exact issues, up to args.fixMax (default 2). A blocked result means the loop gave up at the failing stage (fixLog + lastFailure), or the implementer/contract stage failed. On done, pr.prUrl is the opened pull request (never merged by the harness).

Plan mode: contract-as-plan

Detect plan mode from your own context: a "Plan mode is active" reminder is present and the ExitPlanMode tool is available. (No env var exists — detection is reading your context; a deterministic alternative via a hook's permission_mode is documented in HARNESS.md.) When active, pass stopAtContract: true.

The workflow then returns { status: 'plan', stage: 'contract', contractPath, openQuestions } instead of implementing. Present the contract as the plan:

1. Read the contractPath file; append any openQuestions.
2. Write it to the plan file path from your plan-mode reminder — that file is the ONLY file you may write in plan mode.
3. Call ExitPlanMode — the user reviews the contract as the plan.
4. Approved → re-invoke the SAME workflow with contractPath: <path> and WITHOUT stopAtContract; the implement → audit → verify → review → commit stages run normally.
5. Rejected → do not re-invoke. The user edits the contract file (or re-runs /workflow-feature); they may then pass contractPath to continue.

If you set stopAtContract but ExitPlanMode is unavailable (not actually in plan mode), present the contract in a normal message and ask for approval before re-invoking.

Output

• Relay to the user: stage reached, status (done / blocked), the gate report, the review verdict (when a review ran), any fix rounds (fixLog), the commit hash(es), and the opened PR url if one was created.

Stop conditions

• status === 'plan' → present the contract as a plan and re-invoke with contractPath on approval (see Plan mode section). Do NOT start implementing without that approval.
• status === 'blocked' → relay the blockers verbatim (with the stage: contract / implement / audit / verify / review). Do NOT auto-retry in a loop. The user fixes or re-runs.
• status === 'done' → show the summary + commit hash + PR url (if opened). Then save memory: only decision / constraint / failure records — never chat history.

Notes

• Models are routed per stage, not one model for the whole run. contract runs on opus (one agent, but every later stage reads it, and its Non-goals are what let the auditor catch scope creep); the mechanical stages branch/commit/pr run on haiku (fixed command sequences whose success the workflow re-checks itself); review:test-gaps and followups run on haiku. Everything else inherits executorModel. verify, audit and refute are deliberately NOT downgraded — verify plans the gate before running it, audit is what catches leaked secrets, and a refuter defaults to isReal:false when unsure, so a weak one silently drops real findings. Override any stage with args.models ({ verify: 'haiku' }), or pass null for a stage to force it back to executorModel.
• The contract author answers its own questions (APL-47). openQuestions blocks the run and waits for a human, so it is reserved for genuine PREFERENCE decisions — a product stance, a trade-off only the user can price. Anything OBSERVABLE (how the code is structured, what copy the feature already uses, what the current screen looks like) the author resolves itself: by reading, or for macOS UI questions by building and running the app and taking a screenshot. Resolved-by-looking decisions are recorded in the contract's Risks & assumptions so a reviewer can challenge the observation. This is what keeps stopOnOpenQuestions: true (the default) affordable — before it, a well-specified UI issue still cost two round-trips.
• Parallel runs need worktree: true (APL-45). Without it a run works in the shared checkout and takes an exclusive lock; a second concurrent run is refused outright, naming the branch that holds the lock — loud, not interleaved. With it, the run gets its own worktree under .claude/worktrees/ and every stage is pointed there, so several /workflow-feature runs on different branches are safe at once. Two runs on the SAME branch are still impossible: git refuses to check one branch out in two worktrees, and the harness will not --force past that. A clean run removes its worktree; a blocked run keeps it and reports the path in result.worktree — inspect it, then git worktree remove --force <path>. Two things worktrees do NOT fix: the token budget is shared across concurrent workflows (pass an explicit budget when batching, or each run's guard sees a pool the others are also spending), and each run still costs up to ~70 agents.
• A Branch stage runs before the implementer: it creates or reuses claude/<ISSUE_ID> off an up-to-date prBase, and refuses to run if that branch would equal prBase. The harness therefore never commits onto the base branch. See ${CLAUDE_PLUGIN_ROOT}/harness/PIPELINE.md for the naming convention (and why it differs from Linear's suggested feature/apl-N).
• A blocked result carrying fatal: true means a gate stage did not run (an agent returned no result — usually a terminal API error), not that the diff failed. Never re-read it as a pass; re-run.
• The workflow commits the gated diff, pushes the branch, and opens a PR to prBase (default main). It NEVER merges — merging is the user's call. Review-fix rounds add commits to the same PR.
• The workflow cannot take user input mid-run by design (single-shot). To keep sign-off control, either pass noCommit: true (review the diff, commit/push/PR later manually), or author a contract first and pass contractPath.
• Full automation is the default: it commits, opens the PR, and pushes review fixes to it.