Runbook Generation Skill

Trigger

• Keywords: runbook, release runbook, deployment handbook, release handbook, operational guide, pre-release checklist, rollback plan

When NOT to Use

Scenario	Alternative
Incident response runbook	v2 (not yet implemented)
Code review	/codex-review-fast
Architecture design	/architecture
Tech spec writing	/tech-spec
Request tracking	/create-request

Usage

/runbook                              # Auto-detect feature, create or update
/runbook <feature-keyword>            # Specify feature
/runbook --update                     # Force update mode
/runbook --check                      # Read-only staleness validation
/runbook --request <path|title>       # Specify target request (multi-request features)

Workflow

sequenceDiagram
    participant U as User
    participant S as /runbook
    participant FR as Feature Resolver
    participant CB as Codebase
    participant RB as runbook-release.md

    U->>S: /runbook [feature] [--update|--check] [--request path]
    S->>FR: node scripts/resolve-feature.js
    FR-->>S: {key, doc_inventory, source sets}
    S->>S: Mode dispatch + Request selection

    alt Create Mode
        S->>CB: Read current_authority + requests/*.md
        S->>CB: Scoped discovery (5-priority cascade)
        S->>RB: Write runbook-release.md from template
    else Update Mode
        S->>RB: Read existing runbook + provenance
        S->>CB: Compare current state vs provenance SHAs
        S->>RB: Edit changed sections only
    else Check Mode
        S->>RB: Read existing runbook + provenance
        S->>CB: Validate per-section SHAs
        S-->>U: Report: Fresh/Stale/Missing/Unknown
    end

Phase 0: Context Resolution

Resolve feature using the 5-level cascade:

The wrapper, not the CLI directly: resolve-feature.js owns the failure payload, so the full shape with scan_error: true arrives however the CLI fails — a nonzero exit, a signal, a partial write, a payload that is not the agreed shape. (It cannot survive node itself being unavailable: nothing running under node can. What it removes is the CLI's failure domain, not the interpreter's.) Calling the CLI with || echo '{}' produces a payload the gate below cannot recognise as a failure.

Decide the branch yourself, then run one command. This skill grants Bash(node:*), which matches a direct node … invocation and nothing else — a shell if/[ … ]/$(…) compound is not a node command and cannot run here. Parse $ARGUMENTS first (Step 1 below), then issue exactly one of:

node scripts/resolve-feature.js --feature <the feature key from $ARGUMENTS>

node scripts/resolve-feature.js

Use the first when $ARGUMENTS carried a positional feature key, the second otherwise. Pass the key as a separate argv token — never interpolate it into a larger shell expression.

Source	Mapping
/runbook auth	Positional key auth → --feature auth (two separate argv tokens)
/runbook (no arg)	No --feature, resolver uses branch/diff/fallback
/runbook --check	No --feature, parse flags only

Step	Action
1	Parse $ARGUMENTS for feature key or --check/--update/--request flags
2	Run feature resolver, get key, doc_inventory, and the four source sets (current_authority, design_records, work_records, history_records)
2b	If scan_error !== false, stop — not === true: a payload missing the field is a failure too. See the gate below
3	Check for runbook-release.md specifically in feature directory (not any runbook-*.md)
4	Determine mode: create (runbook-release.md absent) / update (runbook-release.md exists) / check (--check flag)

scan_error gate. Gate on scan_error !== false, not on scan_error === true. When it is not exactly false the four source sets are unknown, not empty — the corpus could not be enumerated (unreadable directory, broken taxonomy, no repository), or the resolver never ran and a shell fallback supplied a payload with no such field at all. {} is the shape that made the stricter test useless: it has no scan_error, so === true is false and the gate passes a payload that contains nothing. Do not proceed as though the feature has no authority documents — report and take the ⚠️ Need Human exit. A key may still be present, so a non-null key is not evidence the sets are complete.

Note: Mode dispatch keys off the specific file runbook-release.md, not any runbook-typed doc in doc_inventory. A feature may have runbook-deploy.md (a different topic) without triggering update mode for the release runbook.

Request Selection

Condition	Behavior
--request specified	Use specified request
Single active request	Auto-select
Multiple active requests	AskUserQuestion: list requests, let user choose
No active requests	Use most recent request (warn)

Phase 1: Content Discovery (Create/Update modes)

Use scoped discovery cascade — narrow to wide, with confidence degradation:

Priority	Scope	Confidence
1	Request Related Files paths	High
2	current_authority — code, rules/, and the docs that claim to be current	High
3	design_records (tech spec, architecture)	Medium — intent only, mark steps unverified
4	Feature-local paths (docs/features/{feature}/)	Medium
5	Repo-wide grep	Low (tag results)

A P1 path is classified before it is used. Related Files is High confidence because the request author named those paths deliberately — not because a path in that table is exempt from the role split. Resolve each one first: a path landing in design_records (a tech spec, an architecture doc) is treated as P3 — Medium, marked unverified — even though it arrived via P1. Otherwise the row the split removed comes straight back through the front door, since a request's Related Files table routinely names 2-tech-spec.md.

Priorities 2 and 3 used to be one row reading "canonical docs (tech-spec, architecture) — High", which is the confusion this feature exists to remove: a tech spec is a design record, and a runbook built from one describes a procedure that may never have been built.

See references/discovery-heuristics.md for per-section mapping.

Security — Redaction Rules

When mining configs/workflows/logs into committed markdown:

Prohibited	Replacement
API keys, tokens, secrets	${ENV_VAR_NAME} placeholder
Webhook URLs with credentials	<webhook-url> symbolic reference
Internal-only endpoints	<internal-endpoint> placeholder
Database connection strings	${DATABASE_URL} placeholder

Phase 2: Generate / Update

Create Mode

1. Read current_authority first — a runbook describes what operators will actually run, so the sources are code, rules/, and the docs that claim to be current. Fall back to design_records (tech spec, architecture) only for the intent behind a step, and mark any step sourced that way as unverified in the provenance manifest: a design record may describe a procedure that was never built
2. Read active request(s) by enumerating docs/features/{feature}/requests/*.md — not by filtering work_records. That set answers "is this document a work record", and a ticket that resolves to some other role — authority Yes, or a Doc role naming one of the other three — leaves it while staying an open ticket; selecting from the set would drop exactly that ticket's AC, scope and related files
3. Run scoped discovery for each template section
4. Fill template from references/template.md
5. Embed <!-- runbook-provenance --> manifest with source SHAs
6. Write to docs/features/{feature}/runbook-release.md

Update Mode

1. Read existing runbook-release.md and parse <!-- runbook-provenance --> block
2. Compare each sources[].sha against git hash-object <file>
3. Identify stale sections (any source SHA mismatch)
4. Re-run discovery for stale sections only
5. Edit stale sections via Edit tool (preserve fresh sections)
6. Update provenance manifest with new SHAs

Phase 3: Check Mode (--check)

Read-only validation — does not modify the runbook file.

1. Read existing runbook-release.md and parse provenance manifest
2. For each section, compare sources[].sha against current git hash-object
3. Classify: Fresh / Stale / Missing / Unknown (see references/check-output.md)
4. Output report with per-section status and SHA diffs
5. Emit verdict: Ready / Stale / Incomplete

Output

Mode	Output	Location
Create	New runbook	docs/features/{feature}/runbook-release.md
Update	Updated sections	Same file, incremental edit
Check	Console report	stdout only (no file modification)

Verification

• Feature resolved via node scripts/resolve-feature.js, and scan_error was exactly false
• Runbook detected in doc_inventory (ancillary/runbook type)
• Template has all 9 sections (see references/template.md)
• Provenance manifest embedded with multi-source SHA tracking
• Discovery uses scoped cascade (not repo-wide grep as first option)
• Redaction rules applied (no secrets in committed markdown)
• --check mode is read-only (no file writes)

Auto-Loop Integration

This skill produces .md output. Per @rules/auto-loop.md:

Event	Action
Create/Update writes .md	/codex-review-doc auto-triggered
Check mode (no writes)	No review needed

References

File	Purpose
references/template.md	9-section runbook template with provenance block
references/discovery-heuristics.md	Scoped discovery cascade and per-section mapping
references/check-output.md	--check mode output template and verdict logic

Examples

Input: /runbook
Action: Auto-detect feature → create runbook-release.md → /codex-review-doc

Input: /runbook auth --check
Action: Read auth/runbook-release.md → validate provenance SHAs → output report

Input: /runbook --update --request docs/features/auth/requests/2026-04-01-login-fix.md
Action: Read existing runbook → diff stale sections → update → /codex-review-doc