| name | shamu-dev |
| description | Use at the start of any shamu development session, when continuing work on the multi-agent orchestrator, when advancing through phases, spawning sub-agent swarms, or deciding what to do next. Cements the development pipeline (HANDOFF → PLAN → swarm → verify → commit) and the invariants that survive across sessions. |
| license | MIT |
Shamu — Development Pipeline
Shamu is a multi-agent coding-orchestrator being built by a swarm of coding agents. That recursion is the product; it's also the process. This skill is the contract for how that process runs across sessions.
First 3 minutes of any session
Do these in order, no agent-spawning until they're done:
- Read
HANDOFF.md at the repo root — current state, what's next, what's in flight, user-owed answers.
git log --oneline | head -25 — shape of what's landed; every phase commit has a rich body.- Read
PLAN.md §§ "Core architecture" + "Phased delivery → Phase N" (where N = current phase from HANDOFF).
- Scan the followups list in HANDOFF for the current phase — those MUST be absorbed.
Then (and only then) begin planning the current phase.
The phase pipeline
Each phase in PLAN.md has Tracks A, B, C… labeled Parallel or Serial. One cycle per phase:
Intake
- Confirm prerequisite phases are ✅ (PLAN checkboxes + HANDOFF agree).
- Identify Parallel vs Serial tracks for this phase.
- Review carry-over followups from prior phases.
- Announce to the user: here's what's landing, here's what I'm spawning.
Plan the swarm
- For each unblocked Parallel track, draft a sub-agent prompt (template below).
- For Serial tracks, plan sequential execution with explicit dependencies.
- Use
TaskCreate for in-session tracking (session-scoped; dies at session end — don't rely on it for persistence).
Spawn
- Fire agents with
run_in_background: true in a single message containing parallel tool calls — this is what makes the swarm a swarm.
- Each agent is scoped to non-overlapping paths. Sibling tracks do not share files.
- Agents never commit. Agents never git-push. Agents never run destructive git (
git reset, git checkout -- ., git clean -f, git stash drop, git branch -D, force-push). Parallel agents share the working tree — an agent that "reverts unexpected changes to keep its hand-off clean" will destroy the sibling's uncommitted work. Every parallel-agent prompt MUST explicitly forbid destructive git and MUST explicitly tell the agent to leave files it didn't create alone.
- Before landing parallel work: stage per-branch files selectively. If siblings touched
bun.lock, separate via stash / branch-switch rather than accepting both in one branch.
Review
- When an agent returns, read its writeup/deliverable.
- Re-run the gate locally:
bun run lint && bun run typecheck && bun run test && bun run agent-ci — all must exit 0.
- If the agent flagged a deviation from PLAN.md: either accept + fold into PLAN.md, or reject + re-spawn.
- Spot-check key files. An agent's summary describes what it intended to do, not necessarily what it did.
Commit + land (PR-based since Phase 5.C's branch protection)
main is protected: direct pushes are rejected. Every change lands via a feature branch + PR.- Branch naming:
shamu/<kind>/<slug> — <kind> ∈ {phase,chore,fix,docs}. Examples: shamu/phase/6a-linear-auth-client, shamu/chore/pr-workflow-adoption.
- Targeted
git add <specific files> on the branch, never -A while sibling tracks are in flight.
- One track per branch per PR where possible. Sibling tracks become sibling branches + sibling PRs.
- Detailed commit body: what landed, why, notable decisions, followups for later phases. PR body summarizes the same, structured as
## Summary + ## Test plan.
- Update PLAN.md checkboxes in the same PR the track is delivered in (tick as you go).
- Merge method: squash. Classic protection enforces linear history; rebase-merges can't be auto-signed by GitHub and get rejected by
required_signatures. Squash lets GitHub sign the resulting commit via web-flow.
- Self-merge is intended. Ruleset has
required_approving_review_count: 0 by design (solo author + solo reviewer; see scripts/setup-branch-protection.sh rationale comment). The ubuntu-latest status check is the load-bearing quality bar.
Close the phase
- After all track PRs merge, open a close-out PR:
- Any PLAN.md section edits made in response to this phase's findings
- HANDOFF.md refresh (status table, "next" section, followups)
- Exit-criterion check against PLAN.md; if not met, the phase isn't done.
Sub-agent prompt template
Every prompt includes:
- Identity. "You are executing Track X.Y of Shamu…" + project root path.
- Context. Which PLAN.md sections to read, which
docs/phase-N/ writeups are relevant.
- Mission. One paragraph: what's being built, why, what already exists.
- Deliverables. Concrete file list + per-file expectations + acceptance criteria.
- Constraints (always include):
- Work only in
<assigned paths>. Don't touch sibling packages, PLAN.md, or docs/phase-*/.
- Don't commit. Don't push. Don't run destructive git. Leave work unstaged for parent review.
- NEVER run destructive git operations. No
git checkout -- ., no git checkout <file> against another agent's work, no git reset --hard, no git clean -f, no git stash drop, no git branch -D, no force-push. If the tree contains unstaged changes you didn't make (a parallel sibling track's in-flight work), LEAVE THEM ALONE — do not "clean up" the working tree. If you need to switch branches and your current branch has uncommitted changes, stop and flag it instead of forcing a checkout. A 2026-04-18 incident destroyed a parallel sibling agent's ~1000 lines of uncommitted work when one agent did a defensive git reset; the other agent had to be re-run from scratch. Assume any file you didn't create belongs to someone else.
- Don't run
bun install unless a deliverable genuinely requires a new dep. A sibling track's concurrent bun install creates a lockfile race that another agent may then "revert" on branch switch — the resulting coordination bug masks real work loss.
- Don't
git checkout <branch> mid-task. Stay on your assigned branch the entire time. If the branch isn't where you expect, stop and flag.
- Before handoff: run
git status and explicitly verify every expected file is present. If something's missing (especially brand-new files), stop and report — don't assume it's a display quirk.
- Run
bun run lint && bun run typecheck && bun run test && bun run agent-ci before declaring done; all must exit 0.
- Pin dependencies exactly (no
^/~). Install with bun install --frozen-lockfile --ignore-scripts.
verbatimModuleSyntax is on — import type for type-only imports.- No
any (Biome rejects).
- Hand-off format. ≤ 400 words: what landed, gate outputs, coverage, design decisions, followups.
Verification gates — non-negotiable
Before any commit:
bun run lint # Biome, must be clean
bun run typecheck # TypeScript strict across all packages
bun run test # full workspace (Vitest + bun test under turbo)
bun run agent-ci # @redwoodjs/agent-ci local CI run
If any gate is red, do not commit. Fix or re-spawn.
Commit discipline
- Body-heavy commits. Subject ≤ 72 chars; body explains what, why, decisions, followups.
git log is a secondary decision log — keep it useful.
- Never
--no-verify, never --no-gpg-sign. The gate is a wall, not a suggestion.
- Targeted
git add. git add -A is dangerous while parallel tracks are unfinished — you'll snapshot half-state.
- Never amend published commits. Always a new commit. Amending pre-push commits to fix a missed signature is allowed (and in fact necessary when the signing config was absent at first commit time).
- No destructive git (force-push,
reset --hard, branch -D, clean -f) without explicit user approval.
- Signed commits. Local setup:
gpg.format=ssh + user.signingkey=<path to public key> + commit.gpgsign=true. The signing SSH key must be registered on GitHub as --type signing (separate from the auth SSH key). Verify on push with git log -1 --show-signature; %G?=N is a local-verification display quirk when gpg.ssh.allowedSignersFile is unset and does not mean the commit is unsigned — check the raw commit with git cat-file commit HEAD | grep ^gpgsig instead.
Known recurring constraints
From Phase 0 spikes + Phase 1 experience. Internalize these:
- Node-style
drain backpressure on vendor subprocess stdin. Every vendor CLI is Node-based; Bun fire-and-forget writes break Claude/Codex under load.
- git 2.50 rejects
-q on git revert and git worktree prune. Redirect output instead.
bun:sqlite cannot load under Vitest workers. @shamu/persistence uses bun test. Wired through turbo run test. Don't port back.- Capabilities are immutable. Loaded from a frozen manifest; adapters cannot upgrade themselves at runtime (T8/G8).
runId is orchestrator-owned from Phase 2 onward. Supervised adapters accept it via SpawnOpts; never mint their own.from_agent on mailbox writes is orchestrator-assigned from authenticated run context (G6). Never accept from payload.- Path-scope enforced at tool-dispatch time in the adapter permission handler (G4). Pre-commit is defense in depth only.
- Shell AST gate via
shell-quote. Default reject-list: $(), backticks, eval, pipes-to-shell, process substitution.
GITHUB_REPO env must be set from the worktree's origin remote before spawning agent-ci.agent-ci run status is derived from workflow + job statuses, NOT from top-level run-state.status (fire-and-forget save).- Docker containers need reaping (
agent-ci-<n>) on SIGTERM to agent-ci.
- Egress broker lands by Phase 7, not Phase 8. Full-autonomy-as-goal means G2 is earlier than originally planned.
When to pause for user input
Stop and ask before proceeding when:
- A new architectural decision isn't already answered in PLAN.md or HANDOFF.md's "already-answered decisions."
- Secrets, API keys, or credentials are needed and can't be reached via keychain.
- A destructive action seems necessary (force-push, branch delete,
--no-verify).
- An adversarial-review finding contradicts the plan and requires re-scoping.
- A vendor SDK rejects the documented auth path.
- A manual-testing step is needed (the user has a standing offer to manually test).
Never silently work around a blocker. Stop, write up, hand back.
Checkbox + HANDOFF discipline
- Tick PLAN.md boxes in the track's own commit, not in a batch at phase end. Drift accumulates otherwise.
- Update HANDOFF.md at every phase boundary — status table, "next" section, followups list, open questions. Commit separately from track work.
- Fold small decisions into PLAN.md where they're contextual (adapter contract, mailbox section, etc.). Don't create a separate
DECISIONS.md — those files rot.
- Session-mid handoff: if the session ends mid-phase with work in flight, note the in-flight agents + expected outputs in HANDOFF under "What's in flight."
Meta-rules
- Trust but verify sub-agent output. Their summaries say what they intended. Run the gate yourself; spot-check files.
- Propose PLAN.md changes openly. If a phase reveals the plan is wrong, edit it, justify in the commit body, proceed. Don't silently drift.
- Prefer editing existing files over adding new ones. Most tracks should touch existing structures.
- No emoji in code or logs. Liability in CI + screen readers. Plain text.
Cross-session continuity
Things that persist (rely on these):
HANDOFF.md — fresh-session entry point.PLAN.md — architecture + phased delivery + decisions + open questions.docs/phase-N/*.md — spike writeups, evidence logs.- Commit bodies — secondary searchable decision log.
AGENTS.md/CLAUDE.md — GitNexus-generated code context.
Things that DON'T persist (don't rely on them):
TaskCreate/TaskList state — session-scoped, dies with the session.- In-session agent history — gone on session end.
- Ephemeral scratch files not committed to git.
HANDOFF.md replaces the task list for anything that must survive the session.
