kirei-chain

You have been invoked via /kirei-chain. Follow this workflow precisely.

You orchestrate multiple research agents in parallel on the same target, then merge their findings into one combined doc and one combined handoff.

You do not spawn execute agents. The user (or a follow-up /kirei) decides how to act on the merged findings.

---

1. PARSE THE INVOCATION

Extract from the user's prompt:

• Task description — the actual thing to investigate (everything that isn't a flag).
• --types <comma-list> — optional explicit lens set. Valid values: security, ui, refactor, perf, arch, test, data, observability, bundle, error, general.
• --research-only — accepted and implied (this skill is research-only by definition).

If --types is not provided, auto-detect by scanning the task for trigger keywords from the table below. Pick all that match (this skill is for multi-lens questions). If only one lens matches, tell the user that /kirei is the better tool and ask whether to continue anyway.

Lens	Trigger keywords
security	auth, vulnerability, injection, XSS, CSRF, secret, permission, CVE, exploit, token, session
ui	component, layout, design, styling, UX, accessibility, visual, responsive, a11y, animation
refactor	cleanup, refactor, technical debt, extract, dead code, duplication, smell, restructure
perf	slow, performance, bundle, optimize, memory, N+1, latency, cache, render speed
arch	architecture, structure, dependencies, coupling, module boundaries, system design
test	tests, coverage, untested, flaky, edge cases, regression
data	schema, migration, query, index, ORM, FK, table, integrity
observability	logging, metrics, tracing, telemetry, PII in logs, log levels, correlation id
bundle	bundle size, code splitting, lazy load, tree-shake, vendor chunk, gzipped size
error	error handling, swallowed catch, generic Error, error contract, no timeouts, unhandled rejection
general	(fallback for one of the slots if the question is broad)

Not lenses — migrate, debug, review, deps, license, and eval are single-purpose research flows, not multi-angle audits. If the user's prompt is really about "upgrade pkg X" or "find why test Y fails" or "review PR #123" or "audit our dependencies" or "license compatibility" or "audit our eval suite", recommend /kirei (or /kirei-deps) instead — chain mode adds noise without adding insight.

Cap at 4 lenses maximum. When more than 4 match, drop in this priority order (drop earliest first):

1. general — almost always redundant when 2+ specific lenses match
2. ui — usually orthogonal to security/perf/data; survives only if explicitly raised
3. refactor — overlaps with arch for structural concerns
4. test — coverage is rarely the most urgent angle in a multi-lens audit

security, data, perf, and arch survive ties by default. Tell the user which 4 you picked and why; offer to swap.

---

2. ANNOUNCE PLAN

One line:

"Running kirei-{a} + kirei-{b} + kirei-{c} in parallel against [target]. Will merge into a single combined findings doc."

---

3. SPAWN THE RESEARCH AGENTS — IN PARALLEL

This is the whole point of the skill. Send all Agent calls in a single message so they execute concurrently. Do not spawn them sequentially.

For each lens, pick the agent:

Lens	Agent
security	kirei-security
ui	kirei-ui
refactor	kirei-refactor
perf	kirei-perf
arch	kirei-arch
test	kirei-test
data	kirei-data
observability	kirei-observability
bundle	kirei-bundle
error	kirei-error
general	kirei

Each agent writes its findings to its own category folder (e.g. docs/security/, docs/perf/, docs/arch/ — see the per-agent prompts).

Prompt structure for each spawned agent (each agent has no shared context — include everything):

Task: [exact task description from the user]

Working directory: [current working directory]

Context:
[Any relevant context from the conversation — file paths, symptoms, recent changes, constraints]

You are part of a parallel kirei-chain investigation alongside: [other agents in this run].
Stay in YOUR lens — do not duplicate the others' work. Note in your handoff if you see something obviously in another lens.

Deliver: structured KIREI HANDOFF block + write findings to your category folder
(e.g. docs/security/ for kirei-security, docs/perf/ for kirei-perf — see your agent prompt for the exact path).

Run them in the foreground. You need every result before the merge step.

If kirei-arch is one of the lenses, note that it's advisory only — that's fine for chain mode (we're not auto-executing anyway).

---

4. WAIT FOR ALL → REVIEW EACH HANDOFF

When all parallel agents complete:

• For each one, read its KIREI HANDOFF block.
• Verify each agent wrote a findings file. Glob the lens-specific folder filtered by today's date (e.g. docs/security/YYYY-*.md where YYYY is the current year). If any agent printed FINDINGS FILE NOT WRITTEN, write that one yourself from its handoff content using the Write tool to the correct category folder.
• Spot-check 1-2 file paths each agent referenced — make sure they exist.

If any agent failed entirely (errored out, no handoff), report that to the user before merging. Do not silently drop a lens.

---

5. WRITE COMBINED FINDINGS DOC

Write a new combined doc that links to and synthesizes the per-lens docs. Don't duplicate the entire per-lens content — link to the originals and pull out the cross-cutting takeaways.

python "${CLAUDE_PLUGIN_ROOT}/scripts/write-findings.py" "<short-topic-slug>" --category chain << 'FINDINGS'
# Combined Findings: <Topic>

**Date:** YYYY-MM-DD
**Skill:** /kirei-chain
**Lenses:** [list — e.g., security, perf, arch]
**Scope:** [what was investigated]

## Per-Lens Reports
- **Security:** docs/security/YYYY-MM-DD-<slug>.md
- **Performance:** docs/perf/YYYY-MM-DD-<slug>.md
- **Architecture:** docs/arch/YYYY-MM-DD-<slug>.md

## Cross-Cutting Themes
[Patterns that appeared in 2+ lenses — these are the highest-leverage findings.
Example: "Both security and perf flag the unbounded admin query at orders.ts:54 —
security as data-exposure risk, perf as DoS risk. Single fix addresses both."]

## Conflicts Between Lenses
[Where lens recommendations disagree.
Example: "kirei-refactor wants to inline the cache wrapper for clarity;
kirei-perf wants to keep the wrapper for memoization. Recommend keeping the wrapper."]

## Unified Priority Order
Rank fixes considering ALL lenses, not just severity within one lens:
1. [Cross-cutting blocker — addresses multiple lens findings] — owner lens(es)
2. [Critical security finding] — security
3. [Critical perf finding] — perf
4. ...

## Recommended Execution Strategy
[One PR per lens? Single bundled PR? Stagger over a sprint?
Justify based on the conflicts and cross-cuts above.]

## Out of Scope (Surfaced but Not Investigated)
[Anything the agents flagged outside their own lens that wasn't picked up by another lens —
recommend a follow-up /kirei or /kirei-chain run.]
FINDINGS

If CLAUDE_PLUGIN_ROOT is not set: mkdir -p docs/chain via Bash, then use Write to write docs/chain/YYYY-MM-DD-<slug>.md.

---

6. OUTPUT COMBINED HANDOFF

---
## KIREI-CHAIN HANDOFF

**Combined report:** docs/chain/YYYY-MM-DD-<slug>.md
**Lenses:** [list]
**Per-lens reports:** linked in the combined doc (one file per lens, in its own category folder)

**Cross-cutting blockers:**
- [Item that appeared in 2+ lenses] — `file:line`

**Unified priority order (top 5):**
1. ...
2. ...

**Suggested next step:**
- Single broad implementation: `/kirei [task] --findings docs/chain/YYYY-MM-DD-<slug>.md`
- Per-lens implementation: spawn `/kirei` separately for each lens's top item

**This skill does NOT auto-execute — user decides how to act on the merged findings.**
---

---

7. REPORT TO USER

One short paragraph:

• Lenses run, in parallel
• Top 1-2 cross-cutting findings (one sentence each)
• Where the combined report lives
• Suggested next move (single /kirei follow-up vs. multiple)

---

RULES

1. Always parallel. All Agent calls go in a single message. Sequential defeats the point.
2. Cap at 4 lenses. More agents in parallel → noisier merge, longer wait, higher cost.
3. Never auto-execute. This is a research-only skill. The merged findings hand off to a user-invoked /kirei or /kirei [type].
4. Surface conflicts, don't silence them. If two lenses disagree, that's a real signal — name it in the merged doc.
5. One lens = wrong tool. If only one lens matches, recommend the user use /kirei instead.