Quality gates and delivery tooling that extend Spec Kit.
Spec Kit Extras is a collection of AI agent commands that add delivery-quality disciplines on top of Spec Kit's core specify → plan → tasks → implement workflow. These are engagement-portable — drop them into any GitHub-hosted project that uses Spec Kit.
| Command | What it does | Works with |
|---|---|---|
/speckit.selfreview |
Rigorous self-review of a PR before requesting human review | Any GitHub repo |
/speckit.pr |
Create a PR with closing-claim accuracy gate | Any repo with a PR template |
/speckit.dora |
Generate weekly DORA metrics from GitHub data | Any GH repo with PRs + Actions |
/speckit.decompose |
Decompose design docs into right-sized GitHub issues | Any repo with design docs |
A bolt is a faster sprint — a tightly time-boxed delivery cycle (default 1 week) that groups features, tracks decisions, and archives outcomes. Think of it as a sprint with less ceremony and more momentum.
| Command | What it does | Works with |
|---|---|---|
/speckit.bolt |
Create and manage bolt cycles — start, add features, track status, complete | Any Spec Kit project |
/speckit.archive |
Archive completed bolt with summary, decisions, and retrospective prep | After bolt completes |
/speckit.retrospective |
Clarify past decisions, identify improvements, document lessons | Any archived bolt |
/speckit.roadmap |
Project-level roadmap across bolt cycles | Any project with bolts |
| Kit | What it adds | Link |
|---|---|---|
| Spec Kit | Core specify → plan → tasks → implement workflow | Required foundation |
| Intent Kit | Intent capture and feature decomposition upstream of Spec Kit | Optional — feeds /speckit.decompose |
| ADM Kit | Architecture Decision Modelling — lineage, invariants, thesis | Optional — complements design docs |
Spec Kit's core workflow ensures features are well-specified and planned. But shipping quality software also requires:
- Knowing your PR is ready before someone else reviews it —
selfreviewcatches the patterns that waste reviewer time (stubs, contract drift, closing-claim mismatches) - PR descriptions that actually explain what's happening —
prgenerates descriptions from spec docs + git history and validates issue-closing claims - Visibility into delivery health —
doracomputes the four DORA metrics so teams can see if their process is healthy - Getting from design to backlog without losing traceability —
decomposeturns architecture docs, ADRs, and design documents into well-formed issues with acceptance criteria
If your project uses Spec Kit, install as an extension:
# From speckit-extras repo
./scripts/install.sh --extension
# Or specify target
./scripts/install.sh --extension --target .specify/extensions/extrasThis installs extension.yml + commands into .specify/extensions/extras/ and enables hooks (auto self-review after /speckit.implement).
# Install bolt cycles extension
./scripts/install.sh --extension --pack boltInstall commands directly into your agent's commands directory:
# Claude Code (default)
./scripts/install.sh --standalone
# Any of 15 supported agents
./scripts/install.sh --standalone --agent gemini
./scripts/install.sh --standalone --agent q
./scripts/install.sh --standalone --agent copilotSupports: claude, gemini, copilot, cursor-agent, qwen, opencode, codex, windsurf, kilocode, auggie, roo, codebuddy, q, amp, shai
TOML-format agents (Gemini, Qwen) get automatic format conversion.
Performs a structured self-review of the current branch's PR diff before requesting human review. Calibrated to catch the recurring patterns that waste review cycles.
What it checks:
- Contract adherence (does implementation match design doc?)
- Issue closing accuracy (does PR fully satisfy every AC of linked issues?)
- Stubs and no-ops (is any function claiming to do something it doesn't?)
- State management (phase transitions atomic? prior state leaking?)
- Error handling (failures visible or silently swallowed?)
- Convention alignment (lint, imports, test patterns)
- Downstream risks (what assumptions do consumers inherit?)
Severity classification:
- C (Critical) — functionality doesn't do what it claims, breaks something
- H (High) — material bug, unsafe pattern, doc/code mismatch
- M (Medium) — design concern, missed edge case, convention violation
- L (Low) — style, naming, minor clarity
- I (Info) — observation, no action needed
Output: Markdown report with findings table, contract adherence summary, issue closing assessment, verdict (APPROVE / REQUEST CHANGES / COMMENT), and recommended fixes.
Action: Fixes all C and H findings immediately. Fixes M unless documented reason to defer. Notes L for follow-up.
/speckit.selfreview
/speckit.selfreview #42 # specific PR numberCreates a pull request with a generated description and a closing-claim accuracy gate.
What it does:
- Generates PR title from commit history
- Generates PR body from spec docs + git log + PR template
- Validates closing claims — if
Closes #Nis claimed but not every AC of issue N has evidence in the diff, downgrades toRefs #Nand lists unmet AC - Runs
/speckit.selfreviewand posts findings as a PR comment - Refuses to open if C-severity findings exist (would block merge anyway)
The closing-claim accuracy gate is the load-bearing discipline: PRs that claim to close an issue when acceptance criteria remain unmet get downgraded automatically. This prevents the "Closes #N" lie that leads to issues being closed without full delivery.
/speckit.pr # PR against main
/speckit.pr develop # PR against develop branchGenerates a weekly DORA metrics snapshot from GitHub API data.
What it computes:
- Deployment Frequency — successful deployments per day (from workflow runs)
- Lead Time for Changes — median time from first commit to merge
- Change Failure Rate — fraction of deploys followed by revert within 24h
- Mean Time to Restore — median time from failure to green build
Output: A Markdown report at doc/reports/YYYY-MM-DD-weekly-dora.md with metrics, DORA band classification (Elite/High/Medium/Low), and prior-window comparison.
/speckit.dora # last 7 days
/speckit.dora 2026-05-01..2026-05-14 # explicit window
/speckit.dora --dry-run # stdout only, no file writeDecomposes design documents, ADRs, or architecture docs into reviewed GitHub issues per a structured playbook.
Three modes:
- Create (default) — propose-then-confirm. Drafts issues, presents for review, creates on confirmation via
gh issue create. - Validate — read-only audit of an existing issue against the decomposition playbook.
- Link-back — append
## Related Worksection to source doc with permanent links to produced issues.
What it enforces:
- Every issue has acceptance criteria (rejected without)
- Every issue references the source document it was decomposed from
- Issues are right-sized (XL items must decompose further)
- Multi-axis labels applied (type, priority, size, epic)
- Epic-shaped issues caught and split
- Dependency relationships made explicit
/speckit.decompose doc/design/my-architecture.md # Create mode
/speckit.decompose validate #42 # Validate mode
/speckit.decompose link-back doc/design/arch.md #42 #43 #44 # Link-backWhen installed as an extension, hooks fire automatically:
/speckit.specify → /speckit.plan → /speckit.tasks → /speckit.implement
│
[after_implement hook]
│
▼
/speckit.selfreview
│
▼
/speckit.pr
And for project-level planning:
Design docs / ADRs → /speckit.decompose → Issues → /speckit.specify (per feature)
Weekly health check:
/speckit.dora → doc/reports/ → team visibility
If you use Intent Kit upstream:
/intent.decompose produces feature list → /speckit.decompose creates issues → /speckit.specify per issue
- Git
- GitHub CLI (
gh) — required by all four commands
- Fork the repo
- Create a feature branch
- Open a PR — CI must pass (all 18 checks)