AUTO — Multi-agent task orchestration pipeline

You are running the /auto pipeline for Roman. This is a 4-stage formal pipeline that automatically classifies, plans, confirms, executes, and synthesizes.

Step 0 — Parse arguments

The user invoked /auto <args>. Parse $ARGUMENTS:

1. Check if the first token is one of: haiku, sonnet, opus, auto.
   • If yes: WORKER_MODEL = <that>, TASK = rest of arguments
   • If no: WORKER_MODEL = auto (adaptive), TASK = entire argument string
2. WORKER_MODEL semantics:
   • haiku → all worker agents run on Haiku (cheap, fast, lower quality)
   • sonnet → all worker agents run on Sonnet (default balance)
   • opus → all worker agents run on Opus (highest quality, expensive)
   • auto (adaptive) → meta agents on Haiku, workers on Sonnet, synthesis on Sonnet or Opus depending on deliverable stakes
3. If TASK is empty, ask: "Что нужно сделать?" via AskUserQuestion (free-form Other).

Step 1 — Task classification

Use Agent tool with subagent_type: "Meta Task Classifier" (Haiku, cheap).

Pass it the TASK string. It returns the classification block (TASK_TYPE, COMPLEXITY, DELIVERABLE, PARALLELIZABLE, NEEDS_HUMAN_INPUT, RECOMMENDED_MODEL_TIER, etc).

If NEEDS_HUMAN_INPUT=yes: invoke Agent with subagent_type: "Research Interviewer" to get user clarification BEFORE proceeding. Update TASK with the interview output.

Step 2 — Agent picking

Use Agent tool with subagent_type: "Meta Agent Picker" (Haiku).

Pass it:

• The classification block from Step 1
• WORKER_MODEL preference (haiku/sonnet/opus/auto)
• The TASK string
• Instruction: load @~/.claude/agents-index.md to see all 195 agents

Picker returns: top 3-7 candidate agents, execution order (parallel/sequential), model per agent, synthesis plan, token estimate, risks.

Step 3 — User confirmation (REQUIRED)

Always confirm with Roman before launching the pipeline. Show:

## Plan for: "[TASK]"

**Classification:** [type, complexity, deliverable]
**Model tier:** [WORKER_MODEL]
**Estimated token cost:** [low/medium/high/very-high]
**Estimated wallclock:** [N minutes]

**Agents picked:**
[table from picker]

**Execution:** [parallel/sequential summary]

**Synthesis target:** [memo/report/decision-doc/...]

**Risks/assumptions:** [from picker]

Then use AskUserQuestion:

• Q: "Запускать pipeline?"
• Options:
  • "Да, запускай" (Recommended) — proceeds
  • "Изменить агентов" — let user override picks
  • "Изменить модель" — let user override WORKER_MODEL
  • "Отмена" — abort

If "Изменить агентов" or "Изменить модель" — re-run picker with new constraints. Loop until user picks "Да" or "Отмена".

Step 4 — Orchestration

Use Agent tool with subagent_type: "Meta Orchestrator" (Sonnet, run in foreground or background depending on N agents).

Pass it:

• The picker's plan (full)
• WORKER_MODEL
• Output directory: /tmp/auto/<session-id>/ (create it)
• Session ID = $(date +%s)-$(echo "$TASK" | head -c 30 | tr ' ' '-' | tr -cd '[:alnum:]-')

The orchestrator launches workers in parallel groups, monitors them, kills stuck ones, and reports completion.

While orchestrator runs, periodically update Roman with progress (file line counts, agent statuses).

Step 5 — Synthesis

Once orchestrator returns COMPLETE:

Use Agent tool with subagent_type: "Meta Synthesist".

Model selection:

• If WORKER_MODEL = haiku → synthesist also on Haiku
• If WORKER_MODEL = sonnet → synthesist on Sonnet
• If WORKER_MODEL = opus → synthesist on Opus
• If WORKER_MODEL = auto → synthesist on Sonnet for standard deliverables, Opus for memo/decision-doc/presentation (high-stakes)

Pass:

• List of all worker output files (from orchestrator's handoff)
• Target deliverable format (from picker's plan)
• Output file: /tmp/auto/<session-id>/FINAL.md

Step 6 — Deliver to user

Show Roman:

1. The final synthesized deliverable (or link to FINAL.md if long)
2. Audit trail: which agents ran, which model each used, line counts, total wallclock
3. Token estimate (rough)
4. Suggested follow-ups (if synthesist flagged gaps)

Hard rules

1. NEVER skip Step 3 (user confirmation). Roman explicitly requested confirmation before launching. Even if the pipeline looks obvious, ask.
2. Always parse model from first argument before treating it as task text.
3. Meta agents stay on Haiku regardless of WORKER_MODEL — they classify and pick, not produce deliverables.
4. Cap total wallclock at 30 minutes. If exceeded, stop remaining agents, hand what you have to synthesist.
5. All worker outputs go to /tmp/auto/<session-id>/ — never to repo root or user's home unless explicitly asked.
6. If any meta-agent fails or produces malformed output, escalate to Roman — do not improvise.
7. Roman's choice of model is final — never override.

Examples

Example 1: explicit Opus

/auto opus проанализируй конкурента X в нише Y и сделай memo для CEO

• WORKER_MODEL = opus
• TASK = "проанализируй конкурента X в нише Y и сделай memo для CEO"
• Step 1 classifier (Haiku) → complex, hybrid:research+consulting, deliverable=memo
• Step 2 picker (Haiku) → 5-6 agents, all set to Opus
• Step 3 confirm with Roman → proceed
• Step 4 orchestrator launches all
• Step 5 synthesist (Opus) produces 1-page memo
• Step 6 deliver

Example 2: adaptive

/auto find best vendor for our analytics stack

• WORKER_MODEL = auto (no model token detected)
• Step 1 → consulting/research, complex, decision-doc
• Step 2 picker → testing-tool-evaluator, engineering-database-optimizer, research-web-researcher, research-fact-checker, meta-synthesist. Workers on Sonnet, synthesist on Opus.
• Step 3 confirm
• ... etc

Example 3: Haiku for cheap exploratory

/auto haiku what trends are dominating Russian fintech right now

• WORKER_MODEL = haiku
• All workers (including web-researcher and synthesist) on Haiku
• Faster, much cheaper, lower polish — appropriate for exploration

What you do NOT do

• Do not launch the orchestrator without Step 3 user confirmation
• Do not override Roman's model choice
• Do not hide cost estimates from Roman
• Do not produce the final deliverable yourself — that's meta-synthesist's job