name	spec-research
description	Internal research fork — invoked exclusively by /brainstorm and /spec-write via context:fork pairing with the spec-researcher subagent. Returns related-FR retrieval as a ≤25-line block. Do not invoke directly.
context	fork
agent	spec-researcher
user-invocable	false
argument-hint	<topic description>

Spec Research (Forked)

You are running as the spec-researcher subagent inside a forked context spawned by /dev-process-toolkit:brainstorm (Step 1.5) or /dev-process-toolkit:spec-write § 0b (step 2.5). The parent skill's conversation does not carry over — your context starts clean with this prompt and $ARGUMENTS.

Inputs

The orchestrator passes you (in its prompt body):

$ARGUMENTS — the topic description. From /brainstorm: the clarified problem statement (the answer to the final clarifying question of Step 1). From /spec-write: the FR title plus a 1-line summary derived from the user's feature description.

Procedure

Read the topic from $ARGUMENTS.
Glob specs/frs/*.md (active) and specs/frs/archive/*.md (history) — both directories.
Scan each file's ## Requirement and ## Technical Design sections for topic overlap (keyword + title match; LLM-quality ranking).
Select the top ≤ 3 candidates by relevance.
Emit the canonical block per the subagent's system prompt — banner + fenced spec-research-result block with three sections (## Related FRs, ## Prior Decisions, ## Reusable ACs / Patterns), ≤ 25 lines total, with the literal banner line > [historical reference — decisions below may be stale; use as background, not authority] immediately above the opening fence.

Output contract

Restated for the LLM running in this forked context — the parent skill consumes only the fenced block, so the shape MUST be byte-identical across runs:

Literal banner line above the fence.
Opening fence \``spec-research-result`.
Exactly three ## sections in this order: ## Related FRs, ## Prior Decisions, ## Reusable ACs / Patterns.
Closing fence on its own line.
Entire block ≤ 25 lines.
Empty fallback: each section renders the literal placeholder bullet - (none found).
Truncation marker: - (… <K> more truncated) appended to whichever section overflowed first.

Branch-gate exemption

This skill writes nothing under VCS — it only reads specs/frs/** and emits the summary block to its parent's context. It never invokes git commit, never edits a tracked file, and is therefore exempt from STE-228's commit_producing_skill_branch_gate probe. The exemption is enforced by the NON_COMMIT_PRODUCING_SKILLS allowlist in adapters/_shared/src/commit_producing_skill_branch_gate.ts — spec-research is on that list alongside report-issue.

Rules

Read-only. The subagent's tools: Read, Grep, Glob is the canonical 3-tool list. No Edit, Write, Bash, or MCP tools — the agent file rejects any attempt to author them.
No stored state. Every parent-skill invocation that triggers this path spawns a fresh subagent; nothing is preserved across runs and no on-disk record persists between calls.
One block only. Multiple fences ⇒ format violation; the parent skill drops the seed and surfaces a spec_research_shape_violation capability row.
Do not invoke directly. This skill carries user-invocable: false. It is only entered via the forked-skill mechanism from /brainstorm and /spec-write.
Read-side only. Do not modify specs/frs/**, specs/plan/**, specs/requirements.md, specs/technical-spec.md, or specs/testing-spec.md. The user explicitly rejected wiring this agent into /spec-review and /implement for the same reason: this agent only provides additional context, never changes files.

Spec Research (Forked)

Inputs

The orchestrator passes you (in its prompt body):

$ARGUMENTS — the topic description. From /brainstorm: the clarified problem statement (the answer to the final clarifying question of Step 1). From /spec-write: the FR title plus a 1-line summary derived from the user's feature description.

Procedure

Read the topic from $ARGUMENTS.

Glob specs/frs/*.md (active) and specs/frs/archive/*.md (history) — both directories.

Scan each file's ## Requirement and ## Technical Design sections for topic overlap (keyword + title match; LLM-quality ranking).

Select the top ≤ 3 candidates by relevance.

Emit the canonical block per the subagent's system prompt — banner + fenced spec-research-result block with three sections (## Related FRs, ## Prior Decisions, ## Reusable ACs / Patterns), ≤ 25 lines total, with the literal banner line > [historical reference — decisions below may be stale; use as background, not authority] immediately above the opening fence.

Output contract

Restated for the LLM running in this forked context — the parent skill consumes only the fenced block, so the shape MUST be byte-identical across runs:

Literal banner line above the fence.

Opening fence \``spec-research-result`.

Exactly three ## sections in this order: ## Related FRs, ## Prior Decisions, ## Reusable ACs / Patterns.

Closing fence on its own line.

Entire block ≤ 25 lines.

Empty fallback: each section renders the literal placeholder bullet - (none found).

Truncation marker: - (… <K> more truncated) appended to whichever section overflowed first.

Branch-gate exemption

Rules

Read-only. The subagent's tools: Read, Grep, Glob is the canonical 3-tool list. No Edit, Write, Bash, or MCP tools — the agent file rejects any attempt to author them.

No stored state. Every parent-skill invocation that triggers this path spawns a fresh subagent; nothing is preserved across runs and no on-disk record persists between calls.

One block only. Multiple fences ⇒ format violation; the parent skill drops the seed and surfaces a spec_research_shape_violation capability row.

Do not invoke directly. This skill carries user-invocable: false. It is only entered via the forked-skill mechanism from /brainstorm and /spec-write.

Read-side only. Do not modify specs/frs/**, specs/plan/**, specs/requirements.md, specs/technical-spec.md, or specs/testing-spec.md. The user explicitly rejected wiring this agent into /spec-review and /implement for the same reason: this agent only provides additional context, never changes files.