stories

name	stories
description	Use when the user wants to ship work across a corgi-compose workspace — EITHER a batch of tracker issues (Linear/Jira links/keys like ABC-123; "do these stories", "implement these tickets", "ship X and Y") OR a free-text feature ("build a feature that …", "add X across the services") OR "find/pick what to work on" with nothing named ("pick the stories", "what should I work on", "grab some agent tickets"). NOT for authoring/running corgi-compose itself (use the corgi skill) or a trivial one-line edit you'd just make directly.
metadata	{"author":"Andriiklymiuk"}

Corgi stories

Work items — tracker issues (Linear/Jira) or a free-text feature — → spec each → isolated branch(es) → tested + reviewed → draft PR/MR per repo → grouped report. Services, dirs, dependency order: all from corgi-compose.yml. Never hard-code.

Speed model

Lighter than a full multi-agent pipeline. One blocking gate on the adjustment/bug fast path; complex stories add superpowers checkpoints.

• Gate (blocking): spec sign-off (Phase 2). Confirm intent before any branch. Cheap, guards the whole batch. Never skip — a clear up-front directive collapses it to inline diagnosis (Phase 2 fast-path): still confirms intent, no separate pause.
• No final gate. Draft PR/MR: push, open draft, scoped review, report diff + link; human flips to ready. Draft = not ready-for-review, no merge; CI still runs if the repo runs it on drafts.
• Pre-authorized autonomous run (opt-in). Blanket approval up front — "I approve all changes", "just do it and open the PRs/MRs", "ship it and watch CI" — is standing sign-off for the whole batch: collapse the gate regardless of tier/span (Phase 2), open the draft PR/MR, and watch CI to green (Phase 5). Don't re-ask at each step. Guardrails still hold — draft-only, never merge, never force-push, blocked stories still surface as questions.

Story tiers — set per story (Phase 1), drives rigor

Tier	What	Extra rigor
Adjustment	clear-spec UI/copy/flag/config; unambiguous	just a test for the new behaviour
Bug	broken / regressed	regression test FAILS on base branch before fix, passes after
Feature	new behaviour, real design, or new/changed cross-service contract	hand to superpowers if installed, else equivalent inline (below)

Tier ≠ span. Complexity axis vs single/multi-service (Phase 4). Multi-service adjustment is still an adjustment. Most stories = adjustments → fastest path.

Express lane — small-surface adjustment. Reuses the Phase 2 fast-path's tier + span bar — adjustment/bug, single service, no cross-service contract — but NOT its "one item / one sentence / no open question" half: that gates the approval pause; this gates machinery. They're independent — a small change with an open design question (which variant? what scope?) still takes the express lane: skip the subagent, just don't skip the gate. Fires whenever the surface is locatable up front — one component or a few files you can name (asset/copy/flag/style swap, small wiring). Lighter steps, same guardrails:

• No Explore subagent — grep + read the 2–3 files inline; a subagent returns a full report to map one component, the orchestrator's own search is cheaper. Reserve Explore for unclear surface or real fan-out.
• No local docs/stories/*.md — the ## Spec tracker comment IS the spec, don't write both (multi-service/feature still gets the file).
• Right-size proof — gate once (typecheck/lint/test) + the single proof the change needs (visual swap → one screenshot/serve check). Don't stack every proof.

Still mandatory: gate sign-off, the spec comment (QA section folded in), branch, per-story review, draft PR, report. Unsure on scope → a quick inline grep settles it; don't default to the heavy path.

Bug sub-type — logic vs visual. "FAILS on base" assumes a unit test can see it. Visual/layout bug (z-index/stacking, overflow, position, breakpoint, CSS specificity/cascade) → jsdom can't catch it; a green unit test proves nothing. Don't fake one.

• Repo has a visual/e2e harness (Playwright, Cypress, Storybook visual diff; Maestro for Expo/RN — references/expo-verification.md) → red check there, same FAILS-on-base.
• None → manual-only is legit, not a skip. Spec + PR body carry repro steps
  • before/after screenshot; report says manually verified, no auto guard. Still include the QA "what to test" section in the spec comment — a human must re-check visual.
  • Expo/RN service on a macOS host → not manual-only: drive the simulator with Maestro + screenshots even without a committed harness (references/expo-verification.md).

Complex story → superpowers

Bigger than adjustment (real design, unclear approach, large surface, new contract) → don't force one-shot. superpowers installed, via Skill:

• superpowers:brainstorming — settle intent + approach before code.
• superpowers:writing-plans — becomes the spec doc.
• superpowers:test-driven-development + superpowers:executing-plans — build, tests first.
• superpowers:verification-before-completion — prove before draft PR.

Not installed → equivalent inline: settle approach with user, plan into spec, tests first, verify before push. Either way flows back here — same spec doc, one gate, per-story review, draft PR, grouped report.

Guardrails (non-negotiable)

• Never touch manualRun services/db_services. Reference-only — corgi doesn't start them, this flow doesn't change them. Fix lands there → STOP, flag out-of-band.
• Draft PRs/MRs only. Never non-draft, never merge, never force-push.
• One blocking gate — spec sign-off (Phase 2); the sign-off is the branch authorization.
• No destructive git without explicit OK — checkout off a dirty tree, branch deletes, force-push, pushing shared branches.
• Don't push the workspace meta repo unless asked — only service branches.
• No AI attribution anywhere. Ticket comments, PR/MR bodies, commit messages all read as human-written. Never add "generated by AI / Claude / corgi", a co-authored-by / AI trailer, a bot signature, or a robot-emoji badge.
• Minimal code comments. Don't narrate the change in code comments — it reads from the diff and the PR/commit. Add one only for a genuinely non-obvious invariant that would otherwise be lost, and only where the file already comments. No "// added X" notes.
• Don't spam the ticket. One comment per story (spec, with its QA section folded in), updated in place on re-runs — never a pile of new comments.

Optional tooling (degrade gracefully)

Stands alone. Used if present, never required:

• superpowers:* (separate plugin) — complex-story engine + nicest review. Missing → inline. Don't block on a missing plugin.
• expo:* (separate plugin) — when a service is an Expo/RN app, its skills (expo:expo-dev-client, expo:building-native-ui, …) deepen the simulator verification in references/expo-verification.md. Missing → the reference alone suffices.
• A code-review command (e.g. /code-review) — Phase 3.5 if present; else a review subagent works everywhere.

Always available, all the flow needs: git, gh/glab, corgi, Explore/Task agents, tracker MCP.

---

Phase 0 — Read workspace from corgi-compose.yml

1. Locate (ls corgi-compose.yml *.corgi-compose.yml). None → /corgi-new first, or ask which repos; don't guess a layout.
2. Read the yaml, extract only needed keys — services.<name>.{path,cloneFrom,manualRun}, depends_on_services, exports (schema: skills/corgi/references/yml-schema.md). Don't render the whole project: /corgi-describe is too many tokens; corgi --describe dumps JSON then still runs the command. Build:
   • Service → dir map. path: (local) or cloneFrom: (clone target) = the repo you branch in. cloneFrom not on disk → corgi init clones first.
   • Dependency/order graph from depends_on_services + exports/ ${producer.VAR}. Depended-on service (schema/contract owner) first; consumers follow. Cycles → flag.
   • manualRun set → exclude (Guardrails).
3. Per repo: forge, base, commands.
   • Forge: git -C <dir> remote get-url origin → *github.com* = gh; *gitlab* = glab. A batch may span both.
   • Base: git -C <dir> symbolic-ref --short refs/remotes/origin/HEAD (or gh repo view --json defaultBranchRef -q .defaultBranchRef.name / glab repo view). <base> for branch, red test, PR target.
   • Test/typecheck/lint/build: discover from package.json scripts, Makefile, pyproject/go.mod, service start/beforeStart/scripts. Don't assume a runner. Also find the CI gate the PR will actually face — a coverage-threshold script (test:cov:check, a check-coverage step, a --coverage floor) or the CI workflow (.github/workflows, .gitlab-ci.yml). Note it; Phase 3 runs that same gate before the PR, not just a scoped test.
4. Detect tracker. linear.app URL → Linear (). /Jira → Jira (; for sites). Bare key + both connected → ask.

Phase 1 — Investigate (once), then spec

Route the intake first — what am I building?

• Explicit tickets (keys/links) or a free-text feature → continue below.
• "Find/pick what to work on", nothing named → don't guess tickets. Resolve via the tracker skill's pickup — the agent queue (label agent, not In Progress/Done), drift-skipped — confirm picks, then build here. Same selection as /corgi-queue: selection in tracker, building here.

Tracker issue: fetch, view screenshots, read real code paths.

• Fetch with relations — Linear get_issue includeRelations: true; Jira issue links. duplicateOf / relatedTo / parent epic often already carries or shipped the work — see Already shipped? below.
• Screenshots — visual/QA bug from text alone = guess. Get the image.
  • Linear = curl the uploads.linear.app URL (signed, expires ~5 min — re-fetch issue for a fresh URL), read.
  • Jira = getJiraIssue gives attachment metadata only (filename + content URL /rest/api/3/attachment/content/<id>), no bytes. mcp__atlassian__fetch won't get bytes — ARIs only; an attachment id mis-resolves to the wrong issue. No MCP tool returns bytes. Jira's URL needs an auth header you don't hold. So: 1. Have creds ($JIRA_EMAIL/$JIRA_API_TOKEN)? curl -u "$JIRA_EMAIL:$JIRA_API_TOKEN" -L -o /tmp/<name> "<content-url>", read. Usually not in env. 2. No creds / 401 / 403 → stop guessing. Ask the user to download + share the path (~/Downloads), ls -t ~/Downloads | head, read.
• Local / user-pointed design assets. User names a folder/file (design-screenshots/, Figma export on disk, "see mockup") → read direct, outranks prose. Build as drawn — exact icon/emoji, colour, copy, spacing, placement. No near-equivalent; a human compares the result to the design.

Reuse an existing spec — then re-verify it (may be stale)

Before speccing from scratch, check three places a prior spec may live — a human's, a past stories run's, a decompose ticket's acceptance criteria:

1. Ticket description — the body often carries the intended approach.
2. Ticket comments — list them (Linear list_comments; Jira via getJiraIssue / dev panel). A comment whose first heading is ## Spec is a prior run's spec.
3. Local docs/ — docs/stories/<issue-key>-*.md (this skill's output) + any hand-written design doc whose name/heading matches.

Found one → starting hypothesis, not ground truth. Re-resolve every file:line against the current tree, re-confirm the contract — code moves, specs rot. Say what drifted, rewrite stale parts, keep what holds. Nothing found → spec from scratch.

Bug tier — check for a dead prior fix first. git blame / git log -S"<symptom>" the suspect lines. Prior fix present but bug persists → inert (overridden, lost specificity, runtime style wins, wrong selector, behind a flag). Explain why it's dead, revive/correct it — don't stack a second half-fix. Note the why in root-cause. Re-opened ticket whose prior fix already merged? Read the prior PR/MR's state (gh pr view <n> --json state / glab mr view): MERGED + closed → its branch is dead (often deleted) and the fix already sits on <base> — branch fresh off <base>, never reopen/continue the merged branch; its old ## Spec comment is the insufficient fix, so update that comment, don't follow it.

Already shipped? — verify before spec or branch

Reuse-spec check (above) finds a prior spec; this finds the deliverable already built on <base> — by a sibling / related / duplicate ticket, or an earlier PR that bundled it. Re-build wastes the batch + opens a dup PR. Before speccing an actionable story, prove it is NOT already done:

• Read relations (fetched above). Open any duplicateOf / relatedTo / parent epic — its description or merged PR often already covers this ask.
• Grep <base> for what the story adds — component, copy string, route, flag: git -C <dir> log --oneline -S"<symbol>" -- <area>, git grep "<copy>" origin/<base>. Present + wired → shipped.
• Open the real file to confirm — not an Explore summary. Explore hands you code that already satisfies the story; "found it" ≠ "needs building."

Shipped → do NOT branch. Mark Status: ALREADY DONE, comment the lineage (commit/PR that delivered it) + recommend close as duplicate, report no-op. Comment ≠ close — ask before changing ticket state. Phase 1 finding → surfaces at the gate, not after a wasted branch.

Free-text feature (no ticket) — locate work first

Description, not links → no fetch, nothing says where code goes. (First check docs/ for an existing design doc — Reuse an existing spec.) Find target service(s) before speccing:

1. Map intent → service(s) from corgi-compose.yml (names, paths, depends_on_services) + the README next to the compose + per-service READMEs. Don't guess.
2. Confirm with Explore scoped to candidate service(s) — find real files.
3. Genuinely ambiguous → spec-gate question (ask, or superpowers:brainstorming); don't guess.

Described feature = usually Feature tier: superpowers:brainstorming (or inline Q&A) to settle scope → superpowers:writing-plans for the spec. After sign-off (Phase 2), offer to create a tracker issue (Linear mcp__linear-server__save_issue with title+team and no id / Jira mcp__atlassian__createJiraIssue) for a key + auto-link; declined → spec stays local + on PR, branch drops the key segment (Phase 3). A caller (e.g. suggest) that already created the issue + hands you key + spec → use that key, don't re-create.

Investigate once — don't re-research

Batched stories overlap. Re-exploring per story doubles tokens. So:

1. Cluster by service + area before dispatching.
2. One Explore sweep per area, not per story — all that area's questions in one agent. Never per-story over the same files. (Micro adjustment → no subagent at all; grep inline — Express lane.)
3. Orchestrator = the cache. Subagents can't share context mid-flight: scope sweeps to not overlap, collect each into one investigation note (scratch — memory or a gitignored file), specs reference it.
4. Reuse ledger — shared components/contracts recorded once; stories cite, don't re-derive.
5. Need runtime/deployed data (staging/prod error, request trace, logs you can't get locally)? Invoke the debug skill (Step 4 — provider data), fold findings into the note; don't hand off the whole flow.
6. Workspace memory (read) — .corgi/memory/ exists → read first (corgi memory list --json / index.md, then matching facts; see memory skill): honor a decision constraint, reuse an incident fix for a regression, ground a free-text feature in domain facts. Absent → skip.

Write the spec — every story

docs/stories/<issue-key>-<slug>.md, actionable or not (micro adjustment → skip the file; the ## Spec comment is the spec — Express lane):

• Problem (quote issue) + which services (drives branch/PR count).
• Tier — adjustment/bug/feature.
• Root cause / current behaviour, file:line refs.
• Change plan (snippets) grouped by service, tests, manual verification, risks. Multi-service: ## Contract + cross-service order.

Triage: actionable vs blocked — controls POSTING, not writing

• Actionable → ONE comment on the issue (human-readable, not a .md attachment). Open with a ## Spec heading; when the change is user-testable, end it with a short ## What to test section (plain QA: clicks + outcome, no code/file refs, end Expected:) — folded into the same comment, not a second one. No HTML comment marker (trackers render <!-- … --> as visible text), no footer badge. Linear mcp__linear-server__save_comment({ issueId, body }) (omit id to create, pass id to update); Jira mcp__atlassian__addCommentToJiraIssue. Literal newlines/markdown. The ## Spec heading is how a later run (new session, no comment id) finds it: list comments, match the one whose first heading is ## Spec, and update it (save_comment({ id, body })) instead of duplicating. Skip the QA section for non-testable stories.
• Blocked → do NOT post. Spec local only; mark Status: BLOCKED + Decision needed; surface the choice. Hold it; rest of batch proceeds.

superpowers:brainstorming / superpowers:systematic-debugging (if installed) to resolve ambiguity before blocking.

Phase 2 — Gate: spec sign-off (the one blocking gate)

Present all actionable specs in one round; sign-off before any branch — batch-level, not per-branch. Re-present only changed specs. Blocked held out. Superpowers-escalated stories pass here too: their writing-plans output is the spec.

Fast-path — collapse the approval round to inline diagnosis ONLY when all hold. Gate just confirms intent before branching. Collapse it (state diagnosis + the exact change inline, then go — that is the gate, they can still stop you) only when every one is true:

• exactly one named item (one ticket, or one specific fix the user named),
• tier adjustment or bug (never feature),
• single service, no cross-service contract,
• you can state the exact change in one sentence, no open question.

Any one false → full gate: feature tier · >1 story · multi-service · ambiguous target · you had to ask yourself a clarifying question · can't name the change in one sentence. Unsure → full gate. "User said 'just fix it'" is not clearance if you can't say exactly what "it" is — a vague directive = open intent → gate. The fast-path skips the pause, never the thinking.

Fast-path drops the approval pause, NOT the tracker artifacts. Even on a one-liner, post the spec comment (with its QA "what to test" section) (Phase 1 triage) — durable record, and a visual/QA defect is what a human must re-verify. Don't skip.

Blanket pre-authorization collapses the gate for ANY tier/span. The single-item fast-path above is for unambiguous changes. Separately, when the user has explicitly pre-approved the changes sight-unseen and told you to proceed to PR/MR ("I approve all changes", "do them all and open the MRs") — that is the sign-off, even for a feature / multi-service / multi-story batch. State each spec inline (diagnosis + change) and go; the tracker still gets the spec comment (QA section included). NOT the same as a vague "just fix it" with no scope (that stays open intent → full gate). Signal = explicit approval + a directive to proceed, not mere impatience.

Phase 3 — Branch + implement + verify per story

Branch: feature/<issue-key>/<kebab-slug>, same name in every affected repo.

Get <issue-key> from the tracker, don't invent it (auto-link token):

• Linear — get_issue → identifier (ABC-123) + suggested gitBranchName. Use identifier; Linear links any branch containing it (case-insensitive). (Or gitBranchName verbatim.)
• Jira — getJiraIssue → key (PROJ-123). Dev panel / Smart Commits link by that token.
• No ticket — feature/<kebab-slug>, no key segment (or the key of an issue you created in Phase 1).

Key also in commit + PR/MR title (Phases 4–5). Same branch name across repos so multi-repo PRs group.

Move the ticket to in-progress + assign when work starts. As each actionable, ticketed story's branch is created (post sign-off):

• Transition to the team's started state — resolve, don't hardcode "In Progress": Linear save_issue({ id, state }) to the started-type state (state takes a type/name/id; resolve via list_issue_statuses); Jira transitionJiraIssue to the transition whose target is In-Progress (mcp__atlassian__getTransitionsForJiraIssue).
• Assign to the mover — current tracker user: Linear save_issue({ id, assignee: "me" }) (assignee accepts a user id/name/email/"me"); Jira editJiraIssue assignee = current user (mcp__atlassian__atlassianUserInfo). Don't steal — assigned to someone else → leave it, note who; unassigned → take it.

Idempotent — skip if already set; skip no-ticket + blocked. The in-progress move also stops a looping /corgi-queue re-grabbing a story in flight (auto-pick takes only not-In-Progress). The review transition fires later, at draft PR — Phase 5.

Pick branch vs worktree per repo — check the working tree first: git -C <dir> status --porcelain --untracked-files=no — empty = clean, any output = dirty. (Ignore stray untracked; checkout -b doesn't disturb them.)

Repo state	Stories touching this repo	Mode
clean	one	branch in place
dirty	one	worktree — don't disturb the user's uncommitted work, and skip the destructive base checkout on a dirty tree
any	several	worktree per story (parallel isolation)

Count "stories touching this repo" across the whole batch up front — two single-repo stories both hitting one repo = "several" → both worktree.

Dirty + overlap guard. A worktree branches from clean origin/<base>, so it excludes the user's uncommitted edits. Before routing a dirty repo to a worktree, check whether those edits touch the story's files. Overlap → work silently diverges → STOP, ask the user to commit/stash or confirm. No overlap → worktree safe.

Must-run producer in a worktree → run it with --service-dir. A producer that must be running for a consumer to verify (Phase 4) can live in a worktree: corgi run --service-dir <producer>=/tmp/corgi-wt/<wt-id>-<service> (below). Only if corgi lacks the flag (corgi run --help | grep service-dir) must it go in place — dirty → ask the user to stash/commit first.

• Branch in place (clean tree; or a must-run producer when --service-dir is unavailable, after stash/commit). Branch straight off the fetched remote base — no checkout <base>/pull dance, no local-divergence trap: git -C <dir> fetch origin && git -C <dir> checkout -b <branch> origin/<base>.
• Worktree (dirty tree, or several stories in one repo). Path /tmp/corgi-wt/<wt-id>-<service> — <wt-id> = <issue-key> (or <kebab-slug> for no-ticket), <service> = service name, so a multi-repo story's repos never collide. Branch off origin/<base> — never touches <dir>'s tree:

  git -C <dir> fetch origin
  git -C <dir> worktree prune                                # drop stale entries
  rm -rf /tmp/corgi-wt/<wt-id>-<service>                     # clear a leftover dir (re-run/crash)
  git -C <dir> worktree add -b <branch> /tmp/corgi-wt/<wt-id>-<service> origin/<base>
  # deps dir (node_modules / vendor / target / .venv) gitignored → symlink main
  # checkout's for SEQUENTIAL runs; real install for CONCURRENT runs.
  ln -s "$PWD/<dir>/node_modules" /tmp/corgi-wt/<wt-id>-<service>/node_modules
  

  The worktree dir is now this repo's working dir — implement, gate, review, commit, push, open the PR/MR from it (Phases 3.5–5).
  • Run a worktree'd service with --service-dir (only services in corgi-compose.yml). corgi resolves a service from its path: (main <dir>); to run the worktree's code — e.g. a producer a consumer verifies against (Phase 4) — pass --service-dir <svc>=/tmp/corgi-wt/<wt-id>-<svc>; corgi runs that service's env, beforeStart/afterStart, process from the worktree, main checkout untouched. Per-service, repeatable — mix worktree + compose path: services:

    corgi run --detach \
      --service-dir api=/tmp/corgi-wt/ABC-200-api \
      --service-dir web=/tmp/corgi-wt/ABC-200-web
    # services not named (admin, worker, db_services) run from their compose path:
    

    --service-dir runs the exact worktree code. (corgi also has --service-branch <svc>=<branch> — its own reused worktree off a branch — and --service-checkout <svc>=<branch> for an in-place checkout; handy for ad-hoc "run this branch", but for stories point at your impl worktree.) Needs the flag (); without it, run such a producer in place. A branched repo that a corgi service → no ; run its runner in the worktree dir.

Implement to spec; reuse before building. Minimum diff — no opportunistic refactor, no over-engineering, no code comments unless the file already comments heavily. Run the per-service gate (tests + typecheck + lint) BEFORE commit. Tests for every change, matching existing patterns.

• Run the gate through corgi when the service is in corgi-compose.yml — gives the worktree full resolved env, deps, cwd, so you don't guess the runner or hand-build env:
  • test script → corgi test --service <svc> --service-dir <svc>=<worktree-dir> (worktree'd) or corgi test --service <svc> (in place).
  • Other command (typecheck/lint/migrate/one-off) → corgi exec <svc> --service-dir <svc>=<worktree-dir> --ensure-deps -- <cmd>.
  • Not in compose, no test script, or no compose → run the discovered runner (Phase 0) in the worktree dir. Same --service-dir <svc>=/tmp/corgi-wt/<wt-id>-<svc> mapping as corgi run (Phase 3); drop for in-place. Needs the flag.
• Bug tier: red test first — write it, confirm FAILS on base, then make it pass. Adjustments skip.
• Pre-existing red baseline. Repo typecheck/lint may already fail on <base>, unrelated. Don't chase a whole-repo green that never existed; don't let baseline noise hide your breakage. Gate on no NEW errors — filter the run to changed files, or diff the base error set. Touched files clean; baseline left as-is, not "fixed" (scope creep).
• Scoped test run can false-green. Path/pattern selector can match nothing yet exit 0 — jest reads app/(app)/… parens + [id] brackets as regex, so the target suite silently never runs while another file prints PASS. Confirm the intended suite ran (assert test count > 0); match by filename substring or escape the path. Green exit ≠ tests ran.
• Scoped run misses downstream importers — prefer the repo's FULL suite. Running only the suites for the files you touched can pass while a barrel/index/sibling suite that imports your changed module fails to even load — a new import you added pulls an unmocked native/heavy dep into that suite's graph. corgi gives you the resolved env, so the full suite is cheap: run the repo's whole test script via corgi test --service <svc> (what CI runs) instead of cherry-picking files. If you must scope, also run any suite that imports the module you changed. A green cherry-picked run that skips the importer is how a red CI slips through.
• Run the repo's CI gate, not just tests — a coverage threshold counts. Many repos fail CI on a per-changed-file coverage floor ( / ), computed over the diff vs and often — one 0%-covered file sinks the average. A green test run (scoped or full) says nothing about coverage; run that exact gate before pushing (Phase 0 found it). Key trap: subject to the floor — a pure refactor/extraction into (or an edit of) an uncovered file fails the gate though behaviour is unchanged. Editing an uncovered file = add a test for it, or don't touch it.

Phase 3.5 — Per-story review (scoped, right after gate is green)

Review each story as it finishes, scoped to only its diff — incremental, bounded context, NOT one giant end-of-batch review (re-reads everything, burns tokens).

• Review git -C <branch-dir> diff <base>...HEAD — a review subagent passed only that diff + the spec (works everywhere), or /code-review / superpowers:requesting-code-review if present.
• Fix blocking findings (correctness, missing test, scope creep), re-run gate. Cap ~1 extra round; still blocked → Stop rule. Non-blocking → PR body.

Phase 4 — Commit, then multi-repo ordering

Commit: match the repo's git log style (Conventional prefix only if the repo does). Concise subject + issue key; body only if truly needed, never a wall. No Co-authored-by / AI trailer. Let pre-commit hooks format; re-stage if rewritten.

Record a fact when it'll matter later (confirm first). A non-obvious bug root cause that could recur, or a cross-service contract decision, earns a memory fact — draft, show, write on OK via corgi memory add … then corgi memory index (see memory skill). After a fix fact, run the recurrence check (corgi memory list --type fix --json; a pattern ≥ 3× → write a proposal, stop — never auto-install). No .corgi/memory/ → offer to create; declined → skip. Never put a secret in a fact.

Multi-service (one issue → N repos → N PRs):

• Same branch name in every repo.
• Contract owner first. Consumer regenerates types/clients from a producer (GraphQL codegen, OpenAPI, protobuf, shared schema)? Producer must be implemented, committed, AND running before the consumer verifies:

  corgi run --services <producer> --with-deps --detach   # or: corgi run --detach
  corgi status --ready --service <producer>               # block until healthy
  

  Until up, consumer's generated types are stale → won't typecheck. Producer from the depends_on_services/exports graph (Phase 0). corgi stop when done. Producer in a worktree? corgi run serves its path: (main checkout) by default — add --service-dir <producer>=/tmp/corgi-wt/<wt-id>-<service> to run the worktree's code (Phase 3). Without the flag, implement the producer in place.
• Consumers regenerate, commit generated output, finish their slice.
• Merge order: producer PR first, consumers after. State in spec + every PR body.

Phase 5 — Push + draft PR/MR per repo

Per repo, forge from Phase 0. <dir> = the repo's working dir — the worktree dir (/tmp/corgi-wt/<wt-id>-<service>) if worktree'd in Phase 3, else the checkout. Run gh/glab from inside that dir (they read the repo from cwd; git -C only sets git's dir, and the spec --body-file/cat path is relative to cwd).

GitHub (gh):

git -C <dir> push -u origin <branch>
gh pr create --draft --base <base> --head <branch> \
  --title "<subject> [<issue-key>]" --body "<what / how / tests / issue link>"
gh pr comment <n> --body-file docs/stories/<issue-key>-<slug>.md   # spec on PR
# flip ready: gh pr ready <n>   (re-draft: gh pr ready <n> --undo)

GitLab (glab):

git -C <dir> push -u origin <branch>
glab mr create --draft --source-branch <branch> --target-branch <base> \
  --title "<subject> [<issue-key>]" --description "<what / how / tests / issue link>" --yes
glab mr note create <iid> -m "$(cat docs/stories/<issue-key>-<slug>.md)"   # spec on MR
# flip ready: glab mr update <iid> --ready

• Draft only. Report each PR/MR's diff summary + link; human flips to ready.
• Watch CI to green (when asked, or in a pre-authorized autonomous run). After the draft is up, poll checks to conclusion and report pass/fail per PR/MR — don't stop at "opened".
  • GitHub: gh pr checks <n> --watch --fail-fast (or poll gh pr checks <n> --json name,bucket until none are pending).
  • GitLab: glab ci status --branch <branch> / glab ci view (or poll glab mr view <iid>) until the pipeline leaves running/pending.
  • Batch spanning both forges → watch each PR/MR on its own forge. On red, surface the failing job (name + log tail), fix on the branch, push, re-watch (Phase 3 Stop rule); never flip to ready to dodge a red check. Still draft-only — watching CI is not merging.
• Move the ticket to the review state once its draft PR/MR is up — resolve, don't hardcode: Linear a Code Review/In Review state (later started-type or custom, from list_issue_statuses); Jira the transition whose target is named In Review/Code Review (getTransitionsForJiraIssue). No such state → leave In Progress. Idempotent; skip no-ticket/blocked. Multi-repo → move once all PRs are open, not per-repo. Best-effort: a tracker↔forge automation may treat a draft PR as in progress and revert this move — the review state only sticks once the PR is marked ready. Set it once; don't fight a revert.
• Cross-link siblings + merge order in each multi-repo PR/MR body.
• Run-locally line in the body — the same one-paste corgi run --service-branch <svc>=<branch> … --with-deps (Grouped report) so a reviewer spins the branch up without hunting.
• Canonical spec already on the tracker (Phase 1); PR/MR comment is a convenience copy.

Grouped report (final output)

<subject> = the PR/MR title without its trailing [<issue-key>] (Phase 5 puts the key in the title — don't print it twice). No (draft) suffix. One blank line between stories.

• Single-repo → one line [<issue-key>] <Service>: <subject>, link directly below.
• Multi-repo → a header [<issue-key>] <story description>, then each repo on its own <Service>: <subject> line with the link below — no key repeated, no blank line between repos.
• No-ticket → swap [<issue-key>] for a short [<slug>] tag so the lines still group.
• Run line → after the link(s), one copy-paste corgi run spinning up every impacted service on its branch via --service-branch <svc>=<branch> (corgi builds the worktree from the pushed branch — reviewer needs nothing else). Same <branch> across repos. --with-deps so deps/dbs come up. One --service-branch per service. Skip blocked/failed. Needs the flag (corgi run --help | grep service-branch); else git checkout <branch> && corgi run --services <svc>. Same line for you + reviewer — the branch is committed now, no --service-dir variant needed. (--service-dir at the live impl worktree belongs to the Phase 3 gate, code still uncommitted.)
• Review hint → after the link(s) per actionable (non-blocked) story, one line per PR/MR: ↳ review it: /corgi-review <pr-or-mr-link> — hands the reviewer to the review skill (checks the diff against repo standards + the ticket, posts inline suggestions). Skip blocked/failed.
• CI line (when watched) → one line per PR/MR after the link: ✓ CI green or ✗ CI red — <failing job>. Omit if CI wasn't watched.
• Blocked / failed → no link, one line: [<key>] <Service>: BLOCKED — <decision needed> (or needs attention — <reason>, + the worktree /tmp path if partial work is parked there).
• Review-channel blurb (only when asked) — user asks for a message for the team's review channel → exactly two lines: <Service>: <short title> then the bare PR/MR link. No pitch, no root-cause paragraph, no emoji, no "please review" — the link unfurls; the channel convention is terse.

[ABC-123] web: Remove address step from mobile signup
https://github.com/<org>/<repo>/pull/<n>
↳ review it: /corgi-review https://github.com/<org>/<repo>/pull/<n>
▶ corgi run --service-branch web=feature/ABC-123/remove-address-step --with-deps

[ABC-200] Add phone field to user
api: Add phone field to user
https://github.com/<org>/api/pull/<n>
web: Add phone field to user
https://github.com/<org>/web/pull/<n>
↳ review it: /corgi-review https://github.com/<org>/api/pull/<n> https://github.com/<org>/web/pull/<n>
▶ corgi run --with-deps --service-branch api=feature/ABC-200/user-phone --service-branch web=feature/ABC-200/user-phone

[ABC-125] api: BLOCKED — which auth scope gates the endpoint?

---

Scenarios & scaling

• Big batch → bound context. >~4–5 stories → dispatch per-branch implementation to subagents (superpowers:subagent-driven-development / dispatching-parallel-agents), one per branch, scoped to its spec + the shared note. Orchestrator stays gate-keeper, collects reports + reviews. Chunk a huge batch.
• Concurrent same-repo test runs need a real install per worktree, not the symlink (build caches race).
• One blocked story never blocks the batch — held aside; actionable ships; blocked surface as questions.
• Mixed forges/trackers in one batch fine — resolved per repo / per issue (Phase 0).

---

Source: Andriiklymiuk/corgi — distributed by TomeVault.