Imported from AdmiralGallade/harness-repository (harnesses/everything-claude-code/skills/plan-orchestrate/SKILL.md). Install upstream with npx skills add AdmiralGallade/harness-repository --skill plan-orchestrate. Copyright stays with the author.

Plan Orchestrate

Bridge a plan document to /orchestrate custom by emitting one ready-to-paste invocation per step. The skill is generative only — it never executes /orchestrate. The user pastes each line when ready.

When to Activate

• User has a multi-step plan document (PRD, RFC, implementation plan) and wants to drive it through /orchestrate.
• User says "orchestrate this plan", "give me orchestrate prompts for each step", "compose chains for this plan".
• A step-by-step plan exists but the user does not want to manually pick agents per step.

Skip when:

• The work is one ad-hoc step → call /orchestrate custom directly.
• The plan is unreadable or empty. Lack of explicit numbering alone is not a skip condition — see the "No clear steps" edge case below.

Inputs

<plan-doc-path> [--lang=python|typescript|go|rust|cpp|java|kotlin|flutter|auto] [--scope=all|step:<n>|range:<a>-<b>] [--dry-run]

• <plan-doc-path> — required; relative or absolute path (@docs/... accepted).
• --lang — reviewer language variant; defaults to auto (detected from project).
• --scope — limits emitted steps; defaults to all.
• --dry-run — print decomposition + chain rationale only; do not emit final prompts.

Authoritative /orchestrate shape (do not deviate)

{ORCH_CMD} custom "<agent1>,<agent2>,...,<agentN>" "<task description>"

Where {ORCH_CMD} is determined in Phase 0 (see below). The command string in the emitted output always uses one concrete form — never both, never a placeholder.

• custom is a sequential chain; each agent's HANDOFF feeds the next.
• Comma-separated agent list. No spaces preferred; one space tolerated.
• No --mode / --gate / --agents=... flags exist — never invent them.
• Agent names come from the catalogue in this skill. Embedded double quotes in the task description are escaped as \".

ECC install form and namespacing

Two install forms determine the prefix on both the slash command and every agent name. The two MUST stay in sync — one form per output, never mixed:

Let <claude-home> denote the Claude Code home directory: ~/.claude on macOS/Linux, %USERPROFILE%\.claude on Windows. Resolve it the way the host platform resolves the user home directory (do not hardcode ~).

Why this matters: under the plugin install, agents register as everything-claude-code:tdd-guide. Bare names force fuzzy matching, which fails intermittently under parallel calls. Under legacy, the prefixed forms are not registered and fail outright.

Available agent catalogue (must pick from these)

General:

• planner — requirement restatement, risk decomposition, step planning
• architect — architecture, system design, refactor proposals
• tdd-guide — write tests → implement → 80%+ coverage
• code-reviewer — generic code review
• security-reviewer — security audit, OWASP, secret leakage
• refactor-cleaner — dead code, duplicates, knip-class cleanup
• doc-updater — documentation, codemap, README
• docs-lookup — third-party library API lookups (Context7)
• e2e-runner — end-to-end test orchestration
• database-reviewer — PostgreSQL schema, migration, performance
• harness-optimizer — local agent harness configuration
• loop-operator — long-running autonomous loops
• chief-of-staff — multi-channel triage (rarely a fit for plan steps)

Build error resolvers:

• build-error-resolver (generic) / cpp-build-resolver / go-build-resolver / java-build-resolver / kotlin-build-resolver / rust-build-resolver / pytorch-build-resolver

Code reviewers:

• python-reviewer / typescript-reviewer / go-reviewer / rust-reviewer / cpp-reviewer / java-reviewer / kotlin-reviewer / flutter-reviewer

A misspelled agent name fails /orchestrate. Cross-check against this list before emitting.

How It Works

Phase 0 — Detect ECC mode + language

1. Read <plan-doc-path>. If missing or empty, report and stop.

2. Detect ECC install form once and freeze it into ECC_MODE. Algorithm (run in order, stop at the first match):

   1. If <claude-home>/plugins/marketplaces/everything-claude-code/ exists → ECC_MODE=plugin.
   2. Else if <claude-home>/agents/ exists and contains at least one ECC agent file (e.g. tdd-guide.md, code-reviewer.md) → ECC_MODE=legacy.
   3. Else → default to ECC_MODE=legacy and emit a one-line warning at the top of the output: > Warning: could not detect ECC install; defaulting to legacy form. If you use the plugin install, edit the prefixes manually.
   4. If both markers exist (mixed install), plugin wins — the plugin namespace is the only one that resolves agent names without fuzzy matching.

   From this point on, every emitted line uses the matching prefix on both the slash command and every agent name. Never emit both forms in the same output.

3. Resolve --lang. When auto, run a polyglot-aware detection:

   • Probe markers: pyproject.toml / uv.lock / requirements.txt → python; package.json → typescript; go.mod → go; Cargo.toml → rust; CMakeLists.txt or top-level *.cpp → cpp; pom.xml / build.gradle (Java) → java; build.gradle.kts or top-level Kotlin → kotlin; pubspec.yaml → flutter.
   • Polyglot tie-break: if more than one marker matches, pick the language whose source files outnumber the others (count via git ls-files, excluding vendor/, node_modules/, dist/, build/, .venv/, generated files, and obvious test fixtures). On a tie or when no language exceeds 60% of source files, set lang=unknown.
   • No marker matched → set lang=unknown.
   • lang=unknown is a sentinel — it is not an agent name. Phase 2 rules 4 and 5 turn it into code-reviewer / build-error-resolver at chain composition time.

4. Detect a PyTorch sub-profile: when lang=python and any of pyproject.toml / requirements.txt / uv.lock declares a dependency on torch, set pytorch=true. This only affects build chain selection (Phase 2 rule below); the reviewer remains python-reviewer.

5. Normalize any agent names declared in the plan: if the plan text references agents by their plugin-prefixed form (e.g. everything-claude-code:tdd-guide), strip the prefix to get the bare catalogue name before validating or composing chains. Re-prefixing happens only at output time per ECC_MODE (Phase 4). Never let a pre-prefixed name flow into chain composition — it would double-prefix in plugin mode.

Phase 1 — Decompose steps

Identify "step units" in priority order:

1. Explicit numbering: ## Step N / ### Phase N / ## N. ... / top-level ordered list.
2. A "Step" column in a table.
3. ----separated blocks with verb-led headings.
4. Otherwise treat each H2 as one step.

Per step extract id (1-based), title (≤ 80 chars), intent (1–3 sentences), tags.

Phase 2 — Tag and pick chain

Tag by intent (multi-tag allowed; chain built from primary + stacked secondaries):

Trigger words below are matched case-insensitively. Multilingual plans are supported by matching the word stems in any language as long as the meaning aligns with the listed English trigger words.

Chain composition rules:

1. Primary tag selection: when a step matches multiple tags, the first one in table order (top of the table = highest priority) is the primary; the rest are secondaries. Composition rules 2 and 3 below handle specific multi-tag combinations explicitly; otherwise, append secondary chains in tag table order.
2. impl + security → tdd-guide,<lang>-reviewer,security-reviewer.
3. impl + db → tdd-guide,database-reviewer,<lang>-reviewer.
4. Deduplicate the resulting chain (preserve first occurrence). E.g. review + lang=unknown would yield code-reviewer,code-reviewer after rule 5; deduplication collapses it to code-reviewer.
5. <lang>-reviewer resolves to code-reviewer when lang=unknown.
6. <lang>-build-resolver resolves to build-error-resolver when lang=unknown. Special case: if Phase 0 set pytorch=true, use pytorch-build-resolver for build chains regardless of <lang>. There is no python-build-resolver; --lang=python without pytorch=true resolves to build-error-resolver.
7. Zero-tag steps: if no trigger word matches, set chain to code-reviewer and write no tag matched; default review-only chain under "Chain rationale".
8. Chain length ≤ 4 after deduplication. If exceeded, drop weakest tag (lookup and docs first).
9. Do not pair planner and architect in an impl chain (token waste). Pair them only on design steps.
10. Steps tagged impl, refactor, or migration end with a reviewer-class agent — any of <lang>-reviewer, code-reviewer, security-reviewer, or database-reviewer. The most domain-specific reviewer wins the tail position (e.g. rule 2's impl+security ends with security-reviewer; rule 3's impl+db ends with <lang>-reviewer because database-reviewer already gates the migration earlier in the chain). test and build steps are gated by their own validators (e2e-runner and the build resolver respectively) and do not require an additional reviewer.

Phase 3 — Compress task description

Each emitted <task description> must:

• Be self-contained (the first agent does not need the plan document open).
• Start with [Plan: <path>#step-<id>].
• Include 1–3 verifiable Acceptance criteria.
• Include a Scope guard (Out of scope: ...) only if the plan declares one for this step. Inherit verbatim. If the plan has no out-of-scope statement, omit the clause entirely — do not invent one.
• Be 200–600 characters; one line; embedded " escaped as \"; no literal newlines.

Phase 4 — Output

Emit Markdown using the form determined by ECC_MODE. The output uses one form throughout — every {ORCH_CMD} and every agent name is rendered with the matching prefix from Phase 0. Do not emit both forms; do not include "this is plugin form" / "strip the prefix" instructions in the rendered output.

Concrete rendering rules:

• {ORCH_CMD} = /everything-claude-code:orchestrate under plugin, /orchestrate under legacy.
• {AGENT(name)} = everything-claude-code:<name> under plugin, <name> under legacy.
• The overview-table "Chain" column uses the same {AGENT(name)} rendering.
• Per-step bash blocks contain only the runnable command. No # plugin form or # legacy form comments — the form is implicit and uniform across the whole output.

Output structure:

# Plan-Orchestrate Result

**Plan**: `<path>`
**Lang**: `<detected-or-given>`
**ECC mode**: `<plugin | legacy>`
**Steps**: <N>
**Scope**: <all | step:n | range:a-b>

## Steps overview

| # | Title | Tags | Chain |
|---|---|---|---|
| 1 | ... | impl, db | `{AGENT(tdd-guide)},{AGENT(database-reviewer)},{AGENT(python-reviewer)}` |
| ... | | | |

---

## Step 1 — <title>

**Intent**: <1–3 sentences>
**Tags**: <a, b>
**Chain rationale**: <why this chain; which agent closes the loop>

```bash
{ORCH_CMD} custom "{AGENT(tdd-guide)},{AGENT(database-reviewer)},{AGENT(python-reviewer)}" "[Plan: docs/foo.md#step-1] <compressed task description>; Acceptance: <1–3 items>; Out of scope: <…>"
```

The {ORCH_CMD} and {AGENT(...)} notation above describes the substitution this skill performs at runtime. The actual emitted Markdown contains the resolved strings, never the placeholders.

Append a final "Batch execution" block aggregating every step's command in order so the user can paste them all at once. Skip the Batch block in overview-only mode (see "Large plan" edge case): when only the overview table is being emitted, there are no per-step commands to aggregate.

Phase 5 — Self-check (run before emitting)

• Every agent in every chain comes from the catalogue (after stripping any everything-claude-code: prefix that appeared in the plan; see Phase 0 step 5).
• Resolved {ORCH_CMD} and every resolved {AGENT(...)} use the same form (plugin or legacy) — never mixed in one output.
• No # plugin form / # legacy form annotations and no "strip the prefix" instructions remain in the rendered output.
• No invented --mode / --gate / --agents=... fields.
• Each task description is single-line, double-quoted, with embedded " escaped.
• Each task description begins with [Plan: <path>#step-<id>] and includes Acceptance (1–3 items). The Out of scope: clause is present only when inherited from the plan.
• No duplicate agent in any chain after Phase 2 dedup.
• Chain length ≤ 4.
• Steps tagged impl/refactor/migration end with a reviewer-class agent (<lang>-reviewer, code-reviewer, security-reviewer, or database-reviewer). test and build are exempt — see Phase 2 rule 10.
• Zero-tag steps emit code-reviewer with the rationale no tag matched; default review-only chain.
• Overview table lists every step in the plan, regardless of --scope.
• Per-step detail block count matches the resolved --scope (full plan when --scope=all; one block for step:n; range size for range:a-b). In overview-only mode, no per-step blocks and no Batch block are emitted.

Edge cases

• No clear steps: prefer H2/H3 splitting; if still ambiguous, report "no structured steps detected" with the document outline and ask the user to confirm running by outline.
• Large plan (>1500 lines): enter overview-only mode — emit only the overview table and ask the user to narrow with --scope before re-running for details. In this mode, skip per-step detail blocks and skip the Batch execution block.
• Step too broad (e.g. "complete all backend work"): do not force a single chain. Suggest splitting into N.a and N.b and propose a split.
• Plan declares agents (rare): first strip any everything-claude-code: prefix to get the bare catalogue name (Phase 0 step 5), then validate against the catalogue. Replace invalid agents and explain under "Chain rationale". The bare name is re-prefixed at output time per ECC_MODE.
• Polyglot project where --lang=auto cannot pick a winner: set lang=unknown; reviewer resolves to code-reviewer and build resolver to build-error-resolver. Mention the fallback under "Chain rationale".

Examples

Example 1 — Plugin mode, Python plan

Input:

plan-orchestrate @docs/plan/example-feature.md --lang=python

Excerpt of expected output:

## Step 2 — Encrypt sensitive UserProfile fields

**Intent**: Introduce an `EncryptedString` SQLAlchemy type and AES-GCM encrypt `birth_datetime` / `location` before persistence; load the key from an environment variable.
**Tags**: impl, security, db
**Chain rationale**: Security-sensitive write path, so `security-reviewer` closes the chain; `database-reviewer` validates the alembic migration; `python-reviewer` covers typing and PEP 8.

```bash
/everything-claude-code:orchestrate custom "everything-claude-code:tdd-guide,everything-claude-code:database-reviewer,everything-claude-code:python-reviewer,everything-claude-code:security-reviewer" "[Plan: docs/plan/example-feature.md#step-2] Implement EncryptedString SQLAlchemy type and migrate UserProfile.birth_datetime/location columns; key from ENV APP_DB_KEY; Acceptance: encrypt/decrypt roundtrip tests pass; alembic upgrade/downgrade clean on empty DB; no plaintext in DB after migrate; Out of scope: cross-tenant profile sharing logic"
```

Example 2 — Legacy mode, same step

If ECC_MODE=legacy were detected, the same step would be emitted as a single uniform command (no plugin-prefixed forms anywhere in the output):

/orchestrate custom "tdd-guide,database-reviewer,python-reviewer,security-reviewer" "[Plan: docs/plan/example-feature.md#step-2] ..."

The two examples above illustrate the two possible outputs for two different environments. A single skill invocation produces only one of them, end to end.

Notes

• Generative only. Never invoke /orchestrate from inside this skill.
• Match the language of the plan document for task descriptions (agent names always remain English).
• Do not insert "Co-Authored-By" lines or emoji in the output unless the user explicitly asks.