subagents

Subagents

Source of truth for agent-to-agent Task spawn — type selection, token-efficient splits, and cost-aware model routing.

Orchestrator — wires coordinator ↔ member agents. Process skills are atoms that compose via context-pack.md; this skill owns how members are spawned and what they receive.

References: context-pack.md · second-opinion-dispatch.md · explore-escalation-dispatch.md · review-council-dispatch.md · subagent-types.md · task-splitting.md · model-routing.md · adversarial.md · task-prompt.md · member-schema.md · output-format.md · agent-discovery.md.

Read references/research-basis.md when calibrating spawn or cost claims. Do not load by habit.

Quick reference

Need	Where
Member context / token rules	references/context-pack.md
Process dispatch recipes	second-opinion · explore-escalation · review-council dispatch refs below
Which subagent_type when	references/subagent-types.md
Token-efficient slice splits	references/task-splitting.md
Cheapest good-enough model	references/model-routing.md
Adversarial / staged debate	references/adversarial.md
Member prompt template	references/task-prompt.md
Per-member output shape	references/member-schema.md
Consolidated report	references/output-format.md
Council agent discovery	references/agent-discovery.md

Non-negotiables

When this skill applies (user attached subagents, an entry skill invokes dispatch, or the plan includes Task members):

1. ≤100k total context — hard ceiling — combined coordinator material + every member prompt + cited excerpts in one dispatch run must stay under 100,000 tokens. If over budget: split into another dispatch run, shrink slices, drop bodies for pointers, or serialise members — never exceed 100k. Budget in the dispatch plan (task-splitting.md).
2. Spawn real members — one host Task per planned member with chosen subagent_type and model per Model assignment. Parallel read_file / grep / other tools are not substitutes for member runs.
3. Synthesis runs after members — merge member outputs before the consolidated report. Writing synthesis without completed Task runs is a violation.
4. Forbidden rationalizations — do not skip spawns because you already read the repo, want lower latency, or want to save tokens when entry skill or plan already committed to dispatch.
5. Valid skips — user declines spawn; When-not-to-spawn passes and no entry-skill carve-out; host cannot run Task; only one member planned and single-pass suffices and no entry-skill carve-out.

Cost default: Cheapest good enough — Auto / omit model under Auto parent; never default to premium or *-fast in parallel.

When-not-to-spawn

Before N ≥ 2, name a single-pass rival: one coordinator with deeper tool use in one context. Spawn only when slices are independent and the rival cannot cover them (task-splitting.md).

Skip subagents when the rival suffices, work is sequential, members would duplicate without added confidence, or the user wants one authoritative pass.

Entry-skill carve-out

When an orchestrator or process skill mandates Task spawn, do not re-litigate “one agent suffices.” Follow that recipe’s member budget and Non-negotiables.

Entry skill	Spawn shape	Recipe lives in
second-opinion	Med or deep (light = no spawn)	second-opinion-dispatch.md
iterate	Single blind member per pass	iterate blind-reviewer-dispatch
handoff	Single compact member (model-invoked)	handoff handoff-subagent-dispatch

Type and model defaults for each → subagent-types.md.

One vs many

Pattern	Members	When
Single subagent	1	Blind review pass, one slice, sequential debate defender, tiebreaker
Parallel coverage	2–6 typical	Independent sources, areas, or topics (task-splitting.md)
Staged debate	2 + 1	Wave attackers → defender (adversarial.md § B)

More than 10 members — split into multiple dispatch runs.

Dispatch modes

• Coverage — split by source, subsystem, or artifact (task-splitting.md).
• Perspective — same material, distinct stance; diversify prompts/stances, not premium models by default.
• Adversarial — kill mandates, context asymmetry → adversarial.md.

Hard rule: never parallel members with identical model and identical prompt. Shared Auto (inherit-auto) is expected — diversify via subagent-types.md and stances.

Workflow

1. Classify

Load the entry skill recipe when one applies; else plan manually:

• Job: research | explore | gather | review | mixed
• Source of truth: web | repo | plan
• Goal: coverage | perspectives | adversarial | adversarial-med | adversarial-deep
• Pick subagent_type per member → subagent-types.md
• Split slices for minimum total tokens → task-splitting.md

2. Plan and spawn

1. Discover — agent-discovery.md when council agents may apply.
2. Dispatch plan — write before spawning (template below). Include single-pass rival unless carve-out applies.
3. Pre-spawn gate — model-routing.md; fail closed on contradictions.
4. Spawn — compose prompts per task-prompt.md. model=inherit-auto → omit tool model.

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

Parent model: [Auto | <named model>]
User model overrides: [none | member=slug, …]
Cheapest path: [inherit-auto | model=auto | explicit slug + why Auto insufficient]

Selected members:

- [subagent_type] · tier=[Fast|Standard|Premium] · model=[inherit-auto | slug] · stance=[id]: [minimal sub-task — see task-splitting.md]

Why these types: [subagent-types.md rationale]
Token budget: [estimated tokens — must sum <100k across all members + coordinator excerpts]
Synthesis plan: [merge / adjudicate]

3. Synthesize

Apply the synthesis gate.

Synthesis gate

Prerequisite: at least one completed Task per planned member (unless valid skip).

1. Merge agreeing findings once.
2. Preserve conflicts — do not flatten.
3. High-stakes contradiction → one sequential tiebreaker (model-routing.md) or ask user.
4. Report per output-format.md; domain shape → entry skill.

Model assignment

Cheapest good enough — not most capable by default. Full procedure → model-routing.md.

Invariant: Parent model = Auto + no override ⇒ every Task call omits model. inherit-auto is a plan sentinel only.

Fallback matrix

Need	Prefer	Fallback
Repo map	explore	generalPurpose
Web docs	docs-researcher	generalPurpose
Plan critique	generalPurpose + stance	council if HOST + plan context

Log fallbacks in availability log.

Output format

Follow output-schema.md. Consolidated runs → output-format.md.

Consumer bindings

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