ica-seam-swarm

name	ica-seam-swarm
description	Run ICA seam swarm, seam assessment, candidate factory, adversary review, prompt pack, or resume/compress.
role	tool-workflow

ICA Seam Swarm

Run read-only ICA-style architecture assessments over folders treated as seams. This skill owns routing, sharding, prompt packaging, synthesis, and confidence filtering. $improve-codebase-architecture owns the ICA vocabulary and deepening judgment.

Load Order

1. Read ROUTING.md first when the request is terse, typo-heavy, scope-ambiguous, prompt-pack ambiguous, or dispatch authorization is unclear.
2. Read RECIPES.md after choosing the mode.
3. Read ORCHESTRATION.md before file listing, preflight, sharding, dispatch, or compression work.
4. Read PROMPTS.md only when building worker or synthesis prompts.
5. Read CANDIDATES.md only for Candidate Factory lanes.
6. Read PERSONAS.md and PATTERNS.md only when assigning shard personas or prompt mechanics.

Routing Rules

1. Prompt Pack Only: explicit request for dispatch prompts, a prompt pack, manual dispatch material, a swarm plan to run elsewhere, or a request not to use agents. Produce compact shared context, shard deltas, and prompts. Do not spawn.
2. Standard Seam Swarm: tight-seam assessment. Use a solo read-only pass unless agent dispatch is authorized; when authorized, run core shards and synthesize architecture tightness.
3. ICA Candidate Factory: discover cross-package opportunities, entropy lanes, DRY candidates, vocabulary candidates, ownership ambiguity, or shared Interface candidates. Produce candidates, not plans.
4. Adversary Review: pressure-test an existing seam report, candidate list, plan, or ADR. Produce weak/dropped/kept claims, falsification evidence, a confidence filter, and one next review prompt.
5. Resume / Compress Swarm: resume after a long or interrupted swarm. Build a compression handoff first, then continue from the smallest useful shard.

When a request matches multiple modes, prefer the narrowest output that answers the current uncertainty. If prompt-pack and candidate language appear together, use ROUTING.md to disambiguate intent before gathering files.

Terse Invocation Router

For shorthand such as ba browse plugin, stop before file listing. Present a numbered router with a recommendation. First run cheap target resolution: verify an exact user-supplied path exists, or verify a shorthand / alias maps to one unambiguous path or concept. If target confidence is ambiguous, list the plausible target alternatives before asking for a number.

Do not print context packets, file inventories, worker prompts, synthesis prompts, or preflight summaries in a router turn.

Agent Dispatch Authorization

Only dispatch subagents when one of these is true:

• The user explicitly asks to run a multi-agent swarm, spawn agents, dispatch agents, use architecture agents, or run review agents.
• The user chooses a router option that explicitly says it will dispatch read-only subagents.
• The user resumes an already-dispatched swarm and asks to continue it.

Bare mentions of this skill, seam swarm, or swarm without an action verb are routing ambiguity, not dispatch authorization.

If the user merely asks whether a folder is a tight seam, use a solo read-only pass by default, or present the router when budget or scope is a real choice.

Execution Flow

1. Identify the seam and selected mode.
2. Confirm dispatch authorization or choose solo / prompt-pack behavior.
3. Gather the complete file list with explicit generated/vendor exclusions.
4. Read nearby AGENTS.md, CONTEXT.md, package maps, and relevant ADRs or runbooks.
5. Build the XML-shaped context packet from PROMPTS.md.
6. Run the orchestrator preflight from ORCHESTRATION.md.
7. Split into independent shards, or run the solo pass.
8. Synthesize with the mode-specific output contract in PROMPTS.md.
9. Apply the adversary / weak-finding filter before finalizing.
10. Stop at evidence and follow-up prompts. Do not modify code or docs unless the user explicitly asks for follow-up implementation.

Bridge Rules

Use $decision-mode only for real choices: mode, budget, seam scope, shard shape, dispatch-vs-prompt-pack, custom lanes, or follow-up owner.

Use $grill-with-docs before dispatch only when scope or language is blocking prompt quality. Use it after synthesis only when a surviving candidate may become durable vocabulary, an ADR, package CONTEXT.md, package AGENTS.md, package-map, or runbook work.

Do not run grill-with-docs inside worker agents.