Orchestrations

When to Use

Multi-phase work that chains several foundational skills against one or more MCP server projects. Typical triggers:

• "Build N new servers" / "scaffold and ship X, Y, Z" → workflows/greenfield-build.md
• "Update and release these servers" / "run maintenance and ship" → workflows/maintenance-release.md
• "QA / field-test / find-and-fix bugs in these servers" → workflows/field-test-fix.md
• "Fix these issues and ship" / handoff document with findings to act on → workflows/fix-wrapup-release.md

Single-skill work — running just maintenance, just git-wrapup, just release-and-publish — invokes the foundational skill directly. Use this orchestrations skill when at least two phases need to chain.

Mental Model — Three Tiers

Workflows in Tier 2 sequence Tier 1 skills with gates and verification. They never duplicate Tier 1 content — they direct to it. A workflow file says "Phase N: agent reads and runs skills/git-wrapup/SKILL.md," not "here's how to wrap up a release."

The orchestrator is the agent driving the workflow — the one reading this SKILL.md. Sub-agents the orchestrator spawns receive prompts pointing at Tier 1 skills directly; they do not receive this skill or the workflow file. That boundary prevents recursive sub-agent spawning.

Pick a Workflow

Identify the workflow from user intent first, then sanity-check against project state if intent is ambiguous.

If intent is ambiguous (no clear signal), surface the candidate workflows to the user and confirm. Don't pick silently.

A workflow file is the orchestrator's playbook for one run. Read it end-to-end before kicking off the first phase.

Universal Rules

These apply to every workflow. Workflow files don't restate them; the orchestrator carries them forward and restates them in sub-agent prompts where applicable.

1. No commits, pushes, tags, branch creation, or destructive ops without explicit user authorization. Work phases leave the working tree dirty for orchestrator review. Wrap-up and release phases run only after the user authorizes — though once authorized, the authorization is durable through the workflow's end (no re-asking at each phase boundary).
2. No git stash, no git reset --hard, no git restore ., no git clean -f, no git checkout -- .. These bypass safety and risk silent data loss. Read-only git (status, diff, log, show, blame) is always safe.
3. No --no-verify, no --no-gpg-sign, no bypassing commit hooks. If a hook fails, investigate the underlying issue.
4. bun run devcheck is the handoff gate between phases. Work phases must hand back a green devcheck. If a phase can't reach green, halt and report the failing step verbatim rather than carrying broken state forward.
5. No marketing adjectives in commits, tags, READMEs, or changelog entries — no "comprehensive", "robust", "enhanced", "seamless", "improved". State the change, not its quality.
6. One workflow per orchestration run. Don't interleave two workflows in the same session. If a target needs both (e.g., maintenance surfaces a bug fix that needs field-testing first), sequence them as two workflow runs with a clean handoff in between.
7. gh release create --notes-from-tag is incompatible with --repo. Always cd into the target repo directory for gh release commands.
8. Annotated tags only (git tag -a), never lightweight, created with --cleanup=whitespace — the default (strip) deletes #-leading lines as comments; --cleanup=verbatim glues the SSH signature into the message (unparseable-as-signed tag, signature block leaks into the release body). Tag annotation subject omits the version number — GitHub prepends v<VERSION>: to release titles when using --notes-from-tag, so including the version in the subject creates stutter. The tag body's final line is a Markdown link to this version's changelog file — [CHANGELOG v<VERSION>](https://github.com/<OWNER>/<REPO>/blob/main/changelog/<major.minor>.x/<VERSION>.md), separated from the gates line by a blank line — giving the release a one-click jump to the full entry.
9. Conventional Commits subjects (feat|fix|refactor|chore|docs|test|build(scope): message). One logical concern per commit. The release commit (version bump + changelog + regenerated artifacts) lands on top of a stack of feature/fix commits, never collapsed alongside them.
10. Email on any artifact is the user's domain email, never a personal address that might appear in git config.

Orchestrator Posture

The orchestrator owns the goals. Workflow phases are not "run skill X" — they are "achieve goal Y, using skill X as the path." Sub-agents (when used) are instruments for hitting the goals, not the work itself. The same posture applies in linear mode — the orchestrator runs the phase directly, but the goal is still the contract.

Before running a phase (or spawning a sub-agent for it), write down four things:

1. Goal — the verifiable end state this phase must produce. Concrete and testable: "v0.5.2 tag exists at HEAD with structured-markdown annotation; bun run devcheck green; npm view <pkg>@0.5.2 resolves." Not fuzzy: "ran the release-and-publish skill."
2. Primary sources — the specific files, GH issues, and reference docs the sub-agent must read directly. Inlining content into the prompt is a paraphrase that loses nuance; agents grounded in the source catch details the orchestrator's summary missed. For GH issues, instruct both gh issue view N --comments (the comment thread) and the timeline cross-reference query in the Orient block (what references the issue, including cross-repo) — the body alone misses both. The orchestrator reads these sources too (to construct the prompt), but that's prompt construction, not a substitute for the sub-agent reading them.
3. Path — the Tier 1 skill(s) and steps that get to the goal. This is what gets handed to the sub-agent.
4. Verification — the read-only checks that confirm the goal was hit. Defined upfront, not as an afterthought.

Why the framing matters:

• Verification follows from goal definition. If the goal is concrete, the verification is obvious — check that exact state. If the goal is fuzzy, verification degrades to "did the sub-agent say it worked?"
• Sub-agent self-reports describe intent, not always reality. A goal you wrote down beforehand is the falsification target — the sub-agent's report is a hypothesis to verify against it.
• Replanning is local. When verification fails, the goal is unchanged; the orchestrator picks a different path (re-spawn with the failure context, re-slice the work, intervene directly). Phase rework doesn't cascade.

Inform without inlining. An enhanced sub-agent prompt names the specific primary sources and the goal — it does NOT paraphrase them. "Review GH issue #123 (read it via gh issue view 123 --comments); the goal is X; verify with Y" is the right shape. Pasting the issue body into the prompt forces the sub-agent to work from a paraphrase. Let the sub-agent read the source and explore for additional context as needed.

Sub-Agent Strategy (if available)

Sub-agents are optional. Match the mechanism to your platform's capability — three tiers, in increasing order:

1. No fanout — run phases linearly. The phase structure is the value; parallelism is the optimization.
2. Parallel sub-agents — compose N prompts, launch concurrently, collect, verify (the pattern below). Single-target workflows usually run linearly anyway; multi-target workflows get one sub-agent per target.
3. Programmatic orchestration — if your platform offers deterministic multi-agent control flow, use its primitives: schema-validated sub-agent returns, automatic concurrency management, resumable/journaled runs, and barrier-free pipelining across phases.

Phases, gates, goals, and constraints are identical across all three tiers — only the fanout mechanism changes. Use the most capable tier available, and don't hand-roll what the platform does natively (e.g., rolling concurrency). Choose by scope and capability, not by default.

Model tier ≠ orchestration tier. A higher orchestration tier is not automatically the right choice. On some platforms, programmatically-orchestrated or nested sub-agents (an agent spawning agents) silently run on a cheaper/downgraded model, while the strongest model is reachable only by sub-agents the main loop spawns directly (tier 2). When a phase needs the top model (heavy generation, design, framework adoption), prefer direct main-loop fanout even when a more "capable" orchestration primitive exists — the primitive can cost you the model. Verify the model your platform actually assigned via its UI/telemetry; never infer it from a sub-agent's transcript, which interleaves auxiliary calls (titles, summaries) on cheaper models and will mislead you.

The decision tree below is orthogonal to tier — it governs whether a given phase fans out, by target count and conflict risk:

Orient block

Every sub-agent prompt opens with this block. Sub-agents do not inherit the orchestrator's CLAUDE.md/AGENTS.md chain or skill registry — both must be reconstructed in the prompt. Substitute the bracketed values per target.

You are working on `[project name]` at `[project absolute path]`.

Orient first. These steps are required before any task work — do them in
order. If any file does not exist, note it and continue.

1. Read the global agent protocol at `~/.claude/CLAUDE.md` (or your agent's equivalent — `~/.codex/AGENTS.md`, etc.).
2. Read the workspace-level protocol if one exists at `[workspace agent protocol path]`
   — skip this step if no workspace-tier protocol applies.
3. Read the project protocol at `[project absolute path]/CLAUDE.md` (or `AGENTS.md`, whichever the project keeps).
4. Run `cd [project absolute path] && bun run list-skills` to see the project's
   available skills with descriptions and locations.
5. Read the skill file(s) for this task: `[Tier 1 skill paths]`.
6. Read the primary sources for this task directly — design docs (`docs/design.md`),
   GH issues, handoff documents, reference/gold-standard files. For a GH issue, read
   both the comment thread and its cross-references — the body alone misses both:
     - `gh issue view <N> --comments` — description + comment thread
     - `gh api 'repos/{owner}/{repo}/issues/<N>/timeline' --paginate --jq '.[] | select(.event=="cross-referenced") | .source.issue | "\(.repository.full_name)#\(.number) — \(.title)"'` — issues/PRs that reference this one, including from other repos
   List each source explicitly: `[primary source paths and gh commands]`. Skip this
   step only if no primary source applies (rare).

Only after that, begin the task below.

**Goal:** [the verifiable end state this phase must produce — concrete, testable]
**Path:** [Tier 1 skill(s) and steps the sub-agent should follow]
**Constraints:** [no-go list — restate git/commit rules and other invariants verbatim]
**Expected outputs:** [report shape you want back — e.g., "Step 8 numbered summary", "list of files touched with one-line rationale per fix"]

The sub-agent reads the primary sources directly during orient (step 6) — do not paste their contents into the prompt. The orchestrator names them; the sub-agent reads them.

Isolation rules

1. Bash git in each sub-agent's own CWD. Parallel sub-agents run git through the shell against their own working directory — independent, no shared state across agents.

2. Sub-agents do not receive this orchestrations skill or workflow files. Their prompts include Tier 1 skill paths only. This prevents recursive sub-agent spawning — if a sub-agent decides it needs to fan out work, that's a signal the orchestrator sliced the work too wide. Re-slice; don't let the sub-agent recurse.

3. Sub-agent prompts must restate the no-git-write and no-stash rules verbatim. The orchestrator's CLAUDE.md/AGENTS.md rules aren't visible to sub-agents at prompt time.

4. Narrow scope per fanout. A sub-agent doing "implement everything, write tests, run devcheck, polish, commit, tag" will exhaust its context window before finishing — the work lands on disk but the agent can't continue. Split phases so each sub-agent finishes well under the context limit. Plan a follow-up "finish" phase as a normal backstop, not a fallback for failure.

Parallel fanout pattern

For N targets in a phase:

1. Compose N sub-agent prompts (one per target) with the orient block + task body + workflow's phase-specific constraints
2. Launch them as parallel sub-agents in a single orchestrator action
3. Collect their reports
4. Verify with a read-only orchestrator check before advancing to the next phase

Barriers only where gates sit. Step 4's "advance to the next phase" implies a barrier — collect every target's phase-N result before any target starts phase N+1. That barrier is only required when a gate sits between the phases: a human decision (authorization, version-bump intent) or cross-target synthesis (the roll-up). Where no gate intervenes, a target may flow through consecutive phases independently — tier-3 platforms pipeline this for wall-clock, and even hand-spawned runs can let one sub-agent carry a target across adjacent gate-free phases. Keep the barrier at gate boundaries; drop it elsewhere. Each workflow's phases table encodes this directly: the Gate after column marks every boundary as barrier (with a terse reason) or gate-free so the spawn/round structure is derivable without re-derivation.

Editor / wrap-up separation

Editing phases and wrap-up phases never go in the same sub-agent. Editing sub-agents make file changes and run devcheck — they do not commit, tag, or push. Wrap-up sub-agents read the working tree, commit, tag, and (when releasing) push and publish — they do not edit source. This separation lets the orchestrator review diffs before they become permanent and keeps the commit graph clean.

Normalization

Independent sub-agents diverge on incidental choices — scoped vs. unscoped package names, script invocation form, README hero structure, badge ordering. When choices should be uniform across targets, plan an explicit normalization step after the fanout — don't expect alignment for free.

For small N or small diffs, the orchestrator normalizes directly. For large N or non-trivial fixes, spawn a narrow-scope fanout with an explicit rule list.

Rolling concurrency

If your platform manages sub-agent concurrency automatically (tier 3), rely on it rather than hand-rolling the below. Otherwise: rate limits on parallel sub-agent spawning are intermittent — sometimes 15 concurrent agents work fine, sometimes 3 get throttled. Don't hard-cap; use rolling concurrency. Launch an initial batch, then as each agent completes, kick off the next in line. If a wave gets rate-limited, shrink the window for the next wave.

Cross-project naming hygiene

When N targets share a phase, never name other targets in a sub-agent's prompt — even as examples. Sub-agents pattern-match on everything in their prompt, and cross-project names leak into commits, messages, and variable names. Each sub-agent's prompt references its own target only.

Verification (orchestrator)

Verification runs against the goal you defined for the phase — not against the sub-agent's self-report. A sub-agent that reports "done" without producing the goal state is not done. The artifact checks below are the means of confirming the goal; pick the ones that exercise your specific goal definition.

Sub-agent self-reports describe intent, not always reality. After every phase that touched the filesystem or remote services, run a read-only check against the goal:

• Files — ls, git status, git diff --stat
• Commits — git log --oneline -5
• Tags — git tag --points-at HEAD, git ls-remote --tags origin
• GitHub — gh repo view --json visibility, gh release view v<VERSION>, gh issue list, gh issue view <N> --comments to confirm the fix comment landed
• npm / registries — npm view <pkg>@<version>, registry-specific checks
• Build state — re-run bun run devcheck if the previous phase was supposed to land green
• Quality — tag annotation is a headline digest covering every change (flat bullets — notable ones named, minor ones in one grouped bullet; no changelog section headers, deps ≤1 line, no gates line), subject omits the version number, no marketing adjectives, issue backlinks where applicable, changelog link as final line

If verification disagrees with the sub-agent's report, that's the signal to re-spawn with the actual state and the unmet goal in the prompt — not to trust the report. The goal hasn't changed; only the path needs to.

Authorization Flow

Pipeline authorization is durable through to completion. Once the user authorizes a workflow run, don't re-ask at each phase boundary — proceed automatically through gates that pass. Conditions that always require a fresh check-in: destructive ops on shared resources, external actions without sign-off, errors that need human judgment.

Workflow File Discipline

Workflow files are thin by design. Each phase row in a workflow's phases table maps to a Tier 1 skill or a thin orchestration step. Phase notes are for orchestration overrides only — sequencing rules, fanout-specific constraints, non-obvious instructions, decisions the foundational skill leaves to the caller. Never paraphrase what a foundational skill already documents. A phase that runs a Tier 1 skill end-to-end with no orchestration override needs no phase note — just the row in the table.

The same discipline applies to gotchas: workflow-specific gotchas are about the orchestration pattern itself (e.g., parallel sub-agent context exhaustion, normalization gaps). Gotchas about a Tier 1 skill's internals belong in that skill, not the workflow.

When the Workflow List Doesn't Fit

For scenarios that don't map cleanly to one of the four workflow files — security audits across N servers, framework-wide migrations, design-only extensions, ad-hoc multi-step work — the universal rules and sub-agent strategy above still apply. Author a new workflow file at workflows/<scenario>.md when the pattern is repeatable enough to codify. Follow the shape of the existing workflow files. Open with the back-pointer every workflow carries — a "Read ../SKILL.md first" tail on the frontmatter description, plus a "Use after reading ../SKILL.md." line under the H1 — so an orchestrator that opens the file directly (not routed through this skill) still picks up the universal rules and sub-agent strategy. Then, when applicable: Tier 1 skills referenced, pre-flight, phases table, phase notes, workflow-specific gotchas, checklist. Apply the workflow file discipline above.

Pre-flight Checklist (every workflow)

Verify before kicking off the first phase. Workflow files add their own pre-flight items on top of these.

• Target list captured with absolute paths
• Intent and state signals point to a single workflow (or confirmed with user if ambiguous)
• Selected workflow file read end-to-end
• Phase objectives understood (the Objective column of the phases table is the goal contract — verification runs against these)
• Plan surfaced to user: workflow, targets, phase objectives, applicable universal rules
• User authorization captured for the workflow's commit/push/publish phases (if any apply)
• Sub-agent capability confirmed (or fallback to linear execution noted)