| name | do |
| description | Meta-orchestrator — takes a natural-language task, classifies it (research / docs / feature / review / bug / refactor / status), builds a pipeline from other skills (research → write-doc → wave-sprint → audit), shows the plan, and executes with approval checkpoints. Use when the user wants "do X" without naming a specific skill. **MCP-first by delegation** (per PRD-022 Tier B): does not call forgeplan directly — delegates to sub-skills (research, refine, sprint, audit, briefing) that are themselves MCP-first with CLI fallback; the orchestration layer is forgeplan-agnostic. Triggers (EN/RU) — "do X", "implement and document", "research and write RFC", "сделай", "разберись и реализуй", "проведи ревью ветки", "/do". |
Task Orchestrator
A "meta-command": takes any task phrasing and assembles a pipeline from other
skills. Doesn't do the work itself — delegates. Goal: remove the mental overhead
of "which skill do I call?"; the user says what, the orchestrator decides how.
Project context (read first)
/do is an orchestrator. It needs every available config source:
@docs/agents/paths.md — where RFCs, TODOs, sources live
@docs/agents/build-config.md — build/test/lint commands
@docs/agents/issue-tracker.md — which tracker
@CONTEXT.md — domain glossary
Check with test -d docs/agents. If present — sub-skills (research, sprint,
audit, briefing, ...) pick up the specifics automatically through their own
@imports; you just pass them the task. If docs/agents/ is missing — sub-skills
auto-detect through project files (glob, package.json, etc.).
When to use
- User described a task in one sentence and didn't name a specific skill.
- The task obviously needs multiple steps (research + RFC, research + sprint + audit, etc.).
- User said: "сделай", "разберись", "проведи ревью", "реализуй и задокументируй".
When NOT to use
- A specific skill clearly fits and the user named it — call it directly.
- Trivial task (read file, fix typo) — don't orchestrate, just do it.
- Single-shot research / single-shot audit — call the skill directly.
Workflow
Phase 1: UNDERSTAND — classification
Parse $ARGUMENTS into one or more categories:
| Category | Signals | Pipeline |
|---|
| research | "разберись", "изучи", "что есть", "какой статус", "сравни" | research → report |
| documentation | "напиши RFC", "сделай guide", "доку", "report", "ADR" | research → write-doc |
| feature | "добавь", "реализуй", "implement", "создай фичу" | research → plan → wave-sprint → tests → docs |
| review | "ревью", "review", "проверь код", "аудит" | audit (review squad) → report |
| bug | "баг", "не работает", "сломалось", "investigate" | research → bug-hunt team → fix |
| refactor | "рефакторинг", "refactor", "переделай", "migrate" | research → plan → wave-sprint (refactor) → tests |
| analysis | "анализ", "analyze", "gap analysis", "state of" | research → write-doc (report) |
| status | "статус", "что сделано", "прогресс" | lightweight research (3 explore) → report |
Multiple categories are fine:
- "design webhooks v2 and write an RFC" = research + documentation
- "add SCIM and document it" = feature + documentation
Phase 2: RECALL
If memory is available — quick check:
memory_recall("$TOPIC")
memory_recall("$TOPIC architecture decisions")
Often research was already done — can be skipped.
Phase 3: PLAN — build the pipeline
Pick a template (see below) based on classification. Show the pipeline to the user
before execution:
## Task: $ARGUMENTS
Classified as: [{categories}]
### Proposed Pipeline:
Step 1: RESEARCH — multi-agent research (5 agents in parallel: code, docs, status, reference, knowledge)
→ Output: Research Report
Step 2: WRITE-DOC (RFC) — RFC based on research findings
→ Output: RFC-XXX-WEBHOOKS-V2.md
Step 3: SAVE — File + Memory + RFC-INDEX update
Estimated agents: 7 (5 research + 2 doc)
Approval checkpoints: after research, after draft
Proceed? (yes / adjust pipeline / skip steps)
When the pipeline includes an implementation step (Template C), fold the
project's hard constraints (red lines from CLAUDE.md/AGENTS.md) into
the plan shown at this checkpoint — see sprint
"Bake project red lines into the plan" (marketplace#169). Catching a
spec-vs-constraint conflict at the plan-approval checkpoint is cheaper than
catching it after Step 4 spawns teammates.
Wait for the answer. The user can:
- "yes" / "давай" / "1" → execute all
- "skip research" → jump to step 2
- "only research" → stop after step 1
- Adjust any step
Phase 4: EXECUTE — step by step
Run steps sequentially. Output of one step → input of the next (pass as a brief
summary, not raw data — otherwise context overflows).
Pipeline Templates
Template A: RESEARCH → REPORT (pure research)
For: "разберись", "изучи", "what's the status", "сравни"
Step 1: RESEARCH
→ [`research`](../research/SKILL.md), 5 agents
→ Output: 5 reports
Step 2: SYNTHESIZE
→ Cross-reference, identify gaps, conflicts
Step 3: DELIVER
→ Present report
→ memory_retain (if memory exists)
Step 4: CLEANUP
Template B: RESEARCH → WRITE-DOC (research + docs)
For: "write RFC", "make a guide", "report"
Step 1: RESEARCH (like Template A) → research report
Step 2: WRITE-DOC
→ Determine doc type (RFC / guide / report / ADR)
→ Determine path and next number (see [`rfc`](../rfc/SKILL.md))
→ Write the doc, using research as input
→ Follow the format from [`rfc`](../rfc/SKILL.md)
Step 3: APPROVAL CHECKPOINT
→ Show draft to user
→ Wait: approve / edit / reject
Step 4: SAVE (only after approval)
→ Write file
→ memory_retain summary
→ Update RFC-INDEX (if RFC)
→ Update TODO (if applicable)
Step 5: CLEANUP
Template C: RESEARCH → PLAN → WAVE-SPRINT (feature implementation)
For: "add", "implement", "реализуй"
Step 1: RESEARCH (like Template A, focused on implementation context)
→ What exists, what's needed, reference patterns
Step 2: PLAN
→ Plan from research → via [`sprint`](../sprint/SKILL.md) Step 2
→ File ownership map
→ Task breakdown (5-6 per teammate)
→ Dependencies
Step 3: APPROVAL CHECKPOINT
→ Show plan
→ Wait approval
Step 4: WAVE-SPRINT execute
→ [`sprint`](../sprint/SKILL.md) Step 4 — wave-by-wave
→ Backend, frontend, tests teammates with research context
Step 5: VERIFY
→ typecheck, build, tests (project-specific commands)
Step 6: DOC (optional, if the task included docs)
→ Update TODO with [x]
→ memory_retain decisions
→ Update RFC (Implementation Log + Insights)
→ If a red-line conflict forced omitting part of the mandate, reconcile
the RFC/PRD (As-Built Reconciliation note) before doc step closes —
see [`sprint`](../sprint/SKILL.md) Step 6 (marketplace#169)
Step 7: CLEANUP
Template D: AUDIT (code review)
For: "ревью", "review", "аудит"
Step 1: SCOPE
→ Determine review target (branch diff, files, PR)
→ git diff main...HEAD for branch
Step 2: AUDIT
→ [`audit`](../audit/SKILL.md), 4-6 agents
Step 3: SYNTHESIZE
→ Cross-validate, prioritize Critical > High > Medium > Low
Step 4: DELIVER
→ Present audit report
→ memory_retain key issues
Step 5: CLEANUP
Template E: RESEARCH → BUG HUNT (bug investigation)
For: "bug", "doesn't work", "investigate"
Step 1: RESEARCH (focused on the bug area)
→ memory_recall known issues
→ Read KNOWN-ISSUES.md
→ Understand architecture around the bug
Step 2: BUG HUNT TEAM
→ Spawn 3-5 teammates, each with its own hypothesis:
· hypothesis-input: validation, parsing
· hypothesis-state: race condition, mutation
· hypothesis-config: env mismatch
· hypothesis-timing: async, ordering
→ See [`team`](../team/SKILL.md), Recipe 3.
Step 3: SYNTHESIZE
→ Which hypothesis is confirmed? Root cause?
Step 4: FIX (if simple)
→ Apply fix, add test, verify
Step 5: DOCUMENT
→ Update KNOWN-ISSUES.md
→ memory_retain root cause + fix
Step 6: CLEANUP
Template F: LIGHTWEIGHT STATUS CHECK
For: "status", "what's done", "progress"
No team — parallel sub-agents:
Step 1: PARALLEL SEARCH (3 sub-agents without TeamCreate):
→ Task(Explore): TODO files for topic
→ Task(Explore): codebase for topic
→ Task(Explore): memory_recall for topic (if memory exists)
Step 2: SYNTHESIZE → status report
Step 3: DELIVER
Decision Logic (template selection)
1. Parse $ARGUMENTS for signals (Phase 1).
2. Multiple categories?
- research signals only → A
- research + doc → B
- feature → C
- review → D
- bug → E
- status → F
- feature + doc → C, then add doc step from B
- research + feature → A, then C with research as input
3. If unclear — ask:
"I see you want {X}. Should I:
a) only research and report?
b) research, then write a doc?
c) research, plan, and implement?
d) something else?"
Approval Checkpoints
The orchestrator runs autonomously, but must pause at:
| Checkpoint | When | What we show |
|---|
| Pipeline approval | After Phase 3 | Proposed pipeline + estimated agents |
| Research review | After research | Summary findings, proceed? |
| Draft review | After doc/RFC | Full document, approve/edit/reject? |
| Implementation plan | Before spawning dev team | File ownership, task breakdown |
| Final review | After implementation | Diff summary, tests passing? |
At each one the user can:
- "давай" / "yes" → continue
- "стоп" → abort
- "пропусти" → skip step
- "измени X" → adjust + retry
Context passing between steps
Output of one step → input of the next. Rules:
- Pass summary, not raw data — otherwise context overflows.
- Each step compresses its output for the next.
- For long pipelines — save intermediate artifacts to files (
research/reports/X.md)
and reference them instead of inlining into the prompt.
RESEARCH output (5 reports → summary)
↓
WRITE-DOC (uses summary for content)
↓
WAVE-SPRINT (teammates get research summary + doc as requirements)
↓
MEMORY (retain synthesis)
Error handling
| Situation | Action |
|---|
| Team agent fails | Retry once. If it fails again — self-verify (re-check the affected work independently) + sweep any claim it held (marketplace#169), then partial report, ask user |
| Research empty | "Nothing found", ask to widen keywords |
| Doc draft rejected | Get specific feedback, rewrite |
| Implementation fails tests | Report failures, attempt fix; if blocked — ask user |
| Pipeline drags on | Report progress between steps; user can abort |
| Memory save failed | Report error, continue pipeline |
Integration with related skills
The orchestrator doesn't implement logic itself — it delegates:
| Step | Delegates to | How |
|---|
| RESEARCH | research | 5-agent team, same recipe |
| WRITE-DOC | rfc | RFC format, index, log |
| PLAN+EXECUTE | sprint | Wave-by-wave |
| BUILD from existing plan | build | IMPLEMENTATION-PLAN-driven |
| REVIEW / AUDIT | audit | 4-6 reviewers |
| TEAM ops (foundation) | team | Mode A/B, cleanup |
| BRIEFING / RECALL | briefing / restore | Before/during pipeline |
Examples
"design webhooks v2 and write an RFC"
Categories: research + documentation
Template: B (research → write-doc)
Pipeline:
1. research → webhook patterns, our code, references (n8n, trigger.dev)
2. rfc → RFC-XXX-WEBHOOKS-V2.md
3. SAVE → file + memory + RFC-INDEX
Checkpoints: after research, after draft
"what's the status on SSO SAML?"
Categories: status
Template: F (lightweight)
Pipeline:
1. 3 parallel Explore sub-agents → TODO + code + memory
2. SYNTHESIZE → status report
No team. Fast.
"add SCIM provisioning — full cycle"
Categories: feature + documentation
Template: C + B hybrid
Pipeline:
1. research → SCIM standards, our IAM, references
2. sprint Step 2 → plan
3. sprint Step 4 → backend-dev + test-writer
4. rfc → SCIM-INTEGRATION guide
5. SAVE → files + memory + TODO
Checkpoints: research, plan, implementation, doc
"security review of the current branch"
Categories: review
Template: D
Pipeline:
1. SCOPE: git diff main...HEAD
2. audit → security + perf + tests focus
3. SYNTHESIZE → prioritized findings
4. SAVE → memory + KNOWN-ISSUES if bugs found
Tips
- Be specific in the task — "design webhooks for real-time event delivery with retry and DLQ" beats "do webhooks".
- Say what you want at the end — "and write an RFC" tells the orchestrator to add a doc step.
- You can interrupt — at any checkpoint adjust the pipeline.
- Pipeline is always visible — user sees the plan before execution.
- Memory accumulates — every next call on a related topic is faster.
- Lightweight by default — status checks don't spawn full teams.
Related skills
research — research step.
rfc — write-doc step.
sprint — implementation step.
build — when research is done and an IMPLEMENTATION-PLAN exists.
audit — review step.
team — foundation for all team ops.
restore — recall before a complex pipeline.
briefing — task-tracker briefing outside the code.
Anti-patterns
- Don't run a pipeline without approval — the user must see the plan.
- Don't pass raw output between steps — compress to summary.
- Don't orchestrate trivial tasks — if one skill is enough, call it directly.
- Don't skip RECALL — research is often already done.
- Don't forget cleanup — every step must close cleanly (TeamDelete, memory_retain).
Forgeplan integration
If the forgeplan CLI is on $PATH, this skill is forgeplan-aware — interactive variant of /autorun. Like /autorun, it probes for forgeplan-workflow and delegates if found.
Probe at start
command -v forgeplan
ls ~/.claude/plugins/cache/marketplaces/ForgePlan-marketplace/plugins/forgeplan-workflow 2>/dev/null
| Detected | Behaviour |
|---|
forgeplan-workflow plugin installed | Delegate to /forge-cycle "<task>" for full route → shape → build → evidence → activate. Keeps interactive checkpoints. |
forgeplan CLI but no forgeplan-workflow | Run the standard pipeline (research → sprint → audit) but insert manual forgeplan calls between phases — forgeplan route before research, forgeplan new prd before sprint, forgeplan new evidence + activate after audit. Pause for user confirmation at each. |
| Neither | Standard pipeline without artifact lifecycle. Tell the user this once at the start. |
Why /do differs from /autorun
/do pauses at every phase boundary for user approval.
/autorun runs end-to-end without checkpoints.
- Both are forgeplan-aware in the same way (probe → delegate → fallback).
Use /do when you want to review the artifact lifecycle steps as they happen (and intervene if needed). Use /autorun when you trust the methodology to run unattended.
Want this orchestrated for you?
If forgeplan-workflow is installed, /do already delegates to /forge-cycle. If not, install it:
/plugin install forgeplan-workflow@ForgePlan-marketplace
/reload-plugins