multi

Multi

Source of truth for parallel subagent orchestration.

Parallel independent subagents via the host Task tool (Cursor: Subagent). Orchestration kernel only — entry skills own job recipes and domain-specific synthesis.

Quick reference

Need	Where
Must-spawn invariants	Non-negotiables
Model routing (cost + fit)	references/model-routing.md
Adversarial / staged debate	references/adversarial.md
Generic task prompt	references/task-prompt.md
Per-member output shape	references/member-schema.md
Generic consolidated report	references/output-format.md
Agent discovery (mechanical)	references/agent-discovery.md
Research basis (calibration)	references/research-basis.md

Read research basis when calibrating a move or making a research claim. Do not load by habit.

Non-negotiables

When this skill applies (user attached multi, an entry skill invokes parallel dispatch, or the plan chose N ≥ 2 members):

1. Spawn real members — Use the host Task tool once per planned member with chosen subagent_type and model per Model assignment. Parallel read_file / grep / other tools are not substitutes for member runs.
2. Synthesis runs after members — The synthesis gate merges member outputs. Writing a consolidated report without running those Task calls first is a violation, not an optimization.
3. Forbidden rationalizations — Do not skip spawns because you already read the repo, expect overlapping findings, want lower latency, want to save tokens, or the diff is “docs-only,” “skills/agent-infra,” or “single theme.”
4. Valid skips — Omit parallel spawns only when: the user declines or asks for a single pass; the job matches Fit check and no entry skill already invoked parallel dispatch; the host cannot run Task; or only one member was planned.

Model routing: Apply Model assignment / model-routing.md per member. On usage-limit / credit exhaustion → Usage-limit retry.

Fit check

Before planning N ≥ 2, name a single-pass rival: the same job done by one agent with a deeper primary pass and more tool use in one contiguous context. Spawn parallel members only when slices are independent and the rival cannot cover that independence (different sources of truth, kill-mandate asymmetry, or true breadth).

Use multi when members are independent and parallel work improves coverage, speed, or confidence — e.g. multi-source gathering, broad exploration, orthogonal research topics, or mixed gather passes.

Skip multi when the single-pass rival suffices, work is sequential, members would duplicate without adding confidence, or the user wants a single authoritative pass.

Entry-skill carve-out: When code-review (or another entry skill) escalated to parallel dispatch / council, or when second-opinion mandates staged debate, Fit check does not apply. Do not re-litigate “one agent suffices.” Follow that skill’s member budget and Non-negotiables.

Dispatch modes

• Coverage — Split by source, subsystem, domain, or artifact. Example: one member maps data flow, another maps call sites.
• Perspective — Same material, distinct stance. Model diversity alone is not enough if prompts are identical.
• Adversarial — Specialization of Perspective: kill mandates, context asymmetry, convergent/divergent synthesis. Full recipe → adversarial.md.
  • Parallel — one wave; independent members (code-review when escalated; investigate when contested).
  • Staged debate (Goal: adversarial-staged) — wave 1 attackers → wave 2 defender who may receive coordinator-composed wave-1 briefs. Allowed sequential exception to same-wave isolation; not live inter-member chat.

Hard rule: Never run parallel members with identical model plus identical prompt. When all members inherit Auto, diversify via distinct prompts and/or stances — shared Auto is expected. Adversarial cross-model diversity → adversarial.md § Model routing overlay (named parent or user request only; never escalate tier just to diversify).

Workflow

1. Classify

Load the entry skill's recipe when one applies; otherwise plan manually:

• Job type: research | explore | gather | mixed
• Source of truth: web | repo | plan
• Goal: coverage | perspectives | adversarial | adversarial-staged | both
• Independence: if members in the same wave need each other's output, that is sequential — do not parallel-spawn them. Exception: staged debate (adversarial.md § B) runs wave 2 after wave 1 with coordinator-composed briefs.

2. Plan and spawn

1. Discover — agent-discovery.md when council agents may apply; otherwise pick host built-in subagent_type values directly.
2. Dispatch plan — write before spawning. Include Single-pass rival (why one agent + deeper pass is insufficient) unless the entry-skill carve-out applies.

Task: [What the user asked]
Classification: [research / explore / gather / mixed]
Source of truth: [web / repo / plan]
Goal: [coverage / perspectives / adversarial / adversarial-staged / both]
Single-pass rival: [one agent + deeper pass — why insufficient, or "entry-skill carve-out"]

Parent model: [Auto | <named model>]
User model overrides: [none | member=slug, …]
Auto reachable: [inherit-auto | model=auto | no]
Host supports: [Task model enum values checked]
Billing pool: [first-party | API | mixed]
Explicit model slugs used: [none | slug + slice-fit reason + cost note]
Fast variants used: [none | slug + explicit latency reason]

Selected members:

- [subagent_type] · tier=[Fast|Standard|Premium] · model=[inherit-auto | slug] · stance=[id or n/a]: [sub-task and expected output]
- [subagent_type] · tier=[Fast|Standard|Premium] · model=[inherit-auto | slug] · stance=[id or n/a]: [sub-task and expected output]

Why these members: [brief justification]
Synthesis plan: [how outputs will be merged or adjudicated]

inherit-auto is a dispatch-plan sentinel only — it means omit the Task/Subagent model argument. It is not a model slug. Never translate inherit-auto into an arbitrary explicit slug. Pass model=auto only when the host Task enum explicitly includes auto.

1. Pre-spawn gate — run Pre-spawn model-routing gate and fail closed on contradictions. Do not issue Task/Subagent calls until every member has a resolved model action that agrees with the plan.
2. Spawn — one Task per member in parallel. Compose prompts per task-prompt.md. Plan vs tool syntax: plan model=inherit-auto → omit model on the tool call; plan model=<slug> → pass model="<slug>" only when that slug is in the host enum. On usage/credit failures → Usage-limit retry.

Member planning defaults:

• Research → one member per independent topic
• Exploration → one member per area, subsystem, or artifact type
• Gathering → one member per source of truth or coverage lens
• Mixed → combine roles only when slices stay independent

3. Synthesize

Apply the synthesis gate.

Synthesis gate

Prerequisite: At least one completed Task per planned member (unless valid skips above apply). If no members ran, do not fabricate a multi report.

After members return:

1. Merge findings that agree; state once with the highest shared confidence.
2. Preserve conflicts — do not flatten disagreements.
3. High-stakes contradiction → single sequential tiebreaker per model-routing.md (prefer Auto, else most appropriate stronger model) or escalate to the user.
4. Write one consolidated report per output-format.md.

Domain-specific synthesis (review filing, investigate verdicts, second-opinion sections) → entry skill recipe.

Fallback matrix

When council agents are unavailable or the job uses host built-ins only:

Need	Prefer	Fallback
Repo map	explore	generalPurpose
Web research	docs-researcher	generalPurpose
Plan structure critique	generalPurpose + stance	council agent if in HOST and dispatch.contexts includes plan

Log skipped council agents and chosen fallbacks in the availability log.

Model assignment

Use the Task tool's allowed model enum from the current host. Never invent slugs. Optimize for cheapest good enough — not most capable by default.

Full procedure → model-routing.md: Routing precedence (canonical order), Parent-aware routing, Plan vs tool syntax, Pre-spawn model-routing gate, Fail closed (do not spawn), Explicit routing (named parent only), Anti-fast (parallel), Reach Auto, Usage-limit retry.

Invariant: Parent model = Auto + no user model override ⇒ every Task/Subagent call omits the model property. inherit-auto is a dispatch-plan sentinel only — never translate it into an arbitrary explicit slug. Tier labels never select a slug under an Auto parent.

Agent count

Scope	Count	When
Small	2–3	Focused feature, explore + verify
Medium	4–6	Multiple related areas or mixed gather
Large	7–10	Broad exploration or comprehensive research

More than 10 members — split into multiple multi runs.

Limitations

• 10 member maximum
• Same-wave members work independently; no inter-member communication (staged debate wave 2 may receive coordinator-composed wave-1 briefs — adversarial.md)
• Best for parallel independent work, not sequential workflows (except adversarial-staged)
• Startup overhead makes it wasteful for tiny tasks

Output format

Follow output-schema.md. Consolidated runs use references/output-format.md.

Consumer bindings

Project recipe index and council agent paths arrive as project-specific injected context on skill read. Do not edit synced copies in place.