autospec-release

name	autospec-release
description	Use when a repo needs an end-to-end release readiness sweep across specs, docs, implementation, tests, QA proof artifacts, legacy cleanup, and merge readiness using existing autospec skills.

autospec-release workflow

Run the repo-level release readiness loop. This is the "I want to ship the current repo" wrapper over the existing autospec surfaces: sweep, review, run, test, QA, proof validation, documentation drift checks, and legacy-code cleanup.

Use this when the operator wants one explicit release gate instead of invoking each lower-level skill by hand.

Self-update mode

If the feature-request argument matches update after trimming and lowercasing, re-install the full autospec suite from main, show the before/after diff if the harness exposes it, then stop. Do not run the release gate.

Required capabilities & harness adapter

Capability	Claude Code	OpenCode	Codex CLI	Fallback if missing
Subagent model tier	Tier A: opus + ultrathink	Tier A: top-tier task + max reasoning	Tier A: current top GPT + reasoning_effort=high	Run inline, but keep the same PASS/PARTIAL/FAIL contract

| Existing autospec skills | Slash skills | agent prompts | skills/prompts | Run the matching workflow instructions inline | | GitHub operations | gh through shell | shell | shell | Mark merge/issue steps BLOCKED with exact command | | Test and QA execution | Browser/E2E or shell | browser/task or shell | shell/browser when available | Mark app-only checks NOT TESTED with blocker |

Model tier: TIER_A for release synthesis and final readiness verdicts because a false PASS can ship broken behavior.

Harness detection

Detect the harness once at skill start:

1. Claude Code: Agent with subagent_type is available.
   • TIER_A = opus + ultrathink.
   • TIER_B = sonnet.
2. OpenCode: task tool is available.
   • TIER_A = top-tier task model + high reasoning.
   • TIER_B = smaller-tier task model + medium reasoning.
3. Codex CLI: apply_patch is the primary edit tool.
   • TIER_A = current top GPT + reasoning_effort=high.
   • TIER_B = current cost-optimized Codex model + reasoning_effort=medium.

Prefer a Tier A reviewer/verifier subagent for the final release verdict when the harness supports it. If TIER_A is unavailable, silently fall back to the next available top-tier model. If delegation is unavailable, run the verdict inline.

Area dispatch (parallel subagents)

/autospec-release parallelizes Stages 3-8 across 6 area subagents instead of walking them sequentially in the orchestrator (issue #731). The canonical orchestrator script is release-area-dispatch.sh, which:

1. Reads area definitions from skills/autospec-release/areas/<name>.md.
2. Dispatches all 6 area subagents in parallel via the harness-aware dispatcher (lib/autospec-harness-detect.sh, PR #725).
3. Aggregates per-area findings into .autospec/release-verdict.json honoring the schema consumed by compute-release-verdict.sh (PR #636).
4. Optionally re-probes stale findings via the qa-finding-filter (qa-finding-filter.sh, PR #650) when AUTOSPEC_RELEASE_VERIFY_FIRST=1.

The 6 areas:

Area	Stage	Description
spec-completeness	3	Specs match live code; no stale/superseded blockers.
docs-freshness	3	README/AGENTS/USER_MANUAL describe current behavior.
implementation-completeness	5	Every acceptance criterion has live code.
test-coverage	6	Release-gate suites green; mutation + density floors met.
qa-artifact-integrity	7	proof-matrix + control-ledger + mutation-proof + canary-results fresh + schema-valid.
legacy-cleanup	8	No dead routes/caches/configs masquerading as current.

Each area subagent receives a tight prompt bounded to 50-120k context and 25 tool calls. Findings carry verified_at: <head_sha> since each subagent is its own verifier; the orchestrator merges them and surfaces live_app_proof from qa-artifact-integrity so compute-release-verdict.sh continues to consume the merged verdict unchanged.

Model tier: Tier A (spec work) for spec-completeness, implementation-completeness, and legacy-cleanup; Tier B (implementation work) for docs-freshness, test-coverage, and qa-artifact-integrity.

Invoke the orchestrator directly:

bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/release-area-dispatch.sh"

Or, when iterating on a single area:

bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/release-area-dispatch.sh" --area spec-completeness

When to use

• The repo has existing specs, issues, code, and docs, and the operator asks "can we release this?"
• A project needs QA, review, testing, documentation drift checks, and legacy cleanup in one coordinated pass.
• Lower-level autospec skills have been run piecemeal and the operator wants a single release-readiness answer.
• The team wants a repeatable gate before tagging, deployment, or handoff.

When not to use

• Do not use this as the first step for a brand-new feature idea. Use /autospec or /autospec-define.
• Do not use it as a replacement for /autospec-stop while a monitor is active.
• Do not mark release PASS with mocked tests only when a no-mock smoke path is required by the repo contract.
• Do not paper over legacy code by leaving stale spec/docs references behind.

Release pipeline

Run these stages in order. If a lower-level skill is installed and invocable, call that skill. If not, follow its documented workflow inline and report that the explicit skill invocation was unavailable.

1. Preflight

   • Record branch, dirty files, remotes, GitHub repo, open PRs, open auto-implement and needs-classify issues, and installed autospec skill versions if visible.
   • Do not revert user work. If dirty files exist, treat them as current repo reality and avoid overwriting them.

2. Sweep

   • Run /autospec-sweep or the equivalent sweep workflow.
   • Ensure .autospec/autospec.yml and the latest sweep report exist when the repo is configured for sweeps.
   • Turn recurring drift into concrete spec, issue, test, or docs work.

3. Spec, docs, and story sync

   • Compare docs/specs/**, README docs, user-facing docs, config docs, and implementation reality.
   • Remove stale claims for legacy code, legacy storage, deprecated APIs, removed flags, dead routes, and old workflows.
   • If behavior still exists only for compatibility, document that explicitly as compatibility, not as the preferred path.

4. Review

   • Run /autospec-review.
   • File or surface high-priority regression issues for spec-vs-code gaps.
   • Do not proceed to PASS while untriaged release-blocking review findings remain.

5. Implementation queue

   • Run /autospec-classify if needs-classify issues exist.
   • Run /autospec-run until there are no ready release-blocking auto-implement issues, or until a blocker is documented with the exact issue number and command to resume.

6. Test gate

   • Run /autospec-test where available.
   • Require lint, typecheck, unit, integration, smoke, E2E, and static checks that the target repo defines as release gates.
   • Failed tests are work, not report decoration: fix them or create a release-blocking issue with exact reproduction.

7. QA proof (MANDATORY — required for PASS)

   • Run /autospec-qa against a live booted app. Either: (a) boot the app locally (dev or prod build) and exercise it, or (b) point QA at an already-deployed environment whose commit SHA matches HEAD (or is explicitly recorded as the SHA under test).
   • Mocked unit/integration tests do NOT satisfy this gate. Stage 7 must produce browser- or app-level interaction artifacts (screenshots, DOM snapshots, network traces, or recorded sessions) covering the user- facing surfaces in the release range.
   • Require control-level proof for text boxes, selects, dropdowns, buttons, validation, API effects, failure banners, fallback behavior, accessibility, and no-mock smoke paths where the repo contract requires them.
   • Validate QA artifacts with skills/autospec-shared/scripts/validate-qa-artifacts.sh when available.
   • Do not skip silently. If the app cannot be booted or reached, record the reason, mark Stage 7 as NOT EXECUTED, and apply the Stage 9 verdict cap.
   • After /autospec-qa completes, confirm .autospec/qa-verdict.json exists with head_sha matching current HEAD and live_app_proof: true. If absent, stale, or live_app_proof: false, the QA gate has not been satisfied for this commit — treat as NOT EXECUTED for Stage 9.

8. Legacy cleanup gate

   • Search code, specs, docs, tests, fixtures, config, infra, and examples for removed or deprecated behavior.
   • Delete dead code when tests prove it is unreachable.
   • Remove stale spec/docs references in the same change that removes the code.
   • If a legacy path cannot be removed yet, create a deprecation issue that names the owner, replacement path, and removal trigger.

8a. Repo quality audit

• Run the shared read-only audit before computing the release verdict:

  bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/repo-quality-audit.sh" \
    --repo . \
    --json ".autospec/repo-quality-audit.json" \
    --markdown ".autospec/repo-quality-audit.md" \
    ${AUTOSPEC_QUALITY_AUDIT_FILE_ISSUES:+--file-issues}
  

• Include audit status, filed issue links, suppressed findings, and unfiled residual risks in the release report.
• Release PASS is blocked by unresolved current-branch-regression security or verification findings unless they are explicitly classified as inherited accepted debt in .autospec/quality-audit-accepted.json.

9. Verdict computation and release iteration loop

   The deterministic computation is implemented by compute-release-verdict.sh and is the canonical source of truth. Invoke it with the current verdict path; stdout is one of PASS / PARTIAL / FAIL and the exit code mirrors that (0/1/2; 3 for infrastructure errors). The prose rules below describe what that script does — if they diverge from the script's actual behavior, fix the script and the bats tests in tests/compute-release-verdict.bats, not the prose. Never re-derive the verdict by hand when the script is available.

   Compute the verdict deterministically from .autospec/qa-verdict.json:

   • If the file is missing → verdict is FAIL with a synthetic finding "qa-verdict.json not produced by Stage 7".
   • If head_sha does not match git rev-parse HEAD → treat as missing. Synthesize a FAIL finding "QA verdict stale relative to HEAD" and require Stage 7 to re-run in the same iteration before recomputing the verdict.
   • If live_app_proof: false → cap at PARTIAL regardless of findings.
   • For each finding with release_blocking: true:
     • status: FAIL → verdict is at least FAIL for that finding.
     • status: PARTIAL or NOT_TESTED → cap at PARTIAL.
   • Otherwise → verdict is PASS.

   Hard caps that override the file contents:

   • "No merged PR in this range requires it" is NOT a valid reason to skip live-app QA or any release-blocking finding. This gate evaluates the cumulative shippable state of the repo, not the diff of the most recent PRs.

   Iteration loop:

   • If verdict is PASS, emit the release report and stop.
   • Otherwise, do NOT emit a final report yet. Pick the smallest remediation that converts the highest-priority release_blocking finding into evidence:
     • benchmark_overfit / mutation_proof_missing / automated_test_gap → re-enter /autospec-qa's regeneration loop to rewrite the offending tests.
     • no_mock_smoke / live_backend_blocker → boot the app or unblock the live backend, then re-run Stage 7.
     • spec_contradiction / legacy_residue → fix docs/specs/code in one commit, then re-run Stages 3-4.
     • outsourced_implementation → replace the third-party API/SDK call with an in-repo implementation of the spec's described algorithm; the external service MUST be either removed or gated behind an off-by-default flag scoped to evaluation-only use, with cached payloads carrying explicit redistribution permission. Then re-run Stages 4-7. If the same offending client/URL recurs in the next iteration's findings, escalate to human review rather than re-iterating on the same remediation.
     • Untriaged auto-implement issue or queued blocker → run /autospec-run on it, then re-run Stages 4-7.
     • Any other release_blocking finding (e.g. accessibility, cross_browser, duplicate_code, new_code_intent, presentational_misclassified, other) → file a tracked GitHub issue with reproduction steps, link it from the release report, then re-run Stage 7. After the remediation lands and is committed, re-enter at Stage 2 (sweep) and recompute. Each iteration MUST re-run Stage 7 before recomputing the verdict — a verdict file whose head_sha does not match the iteration's current HEAD is treated as missing.

   Termination conditions (any one ends the loop):

   • Verdict reached PASS.
   • Iteration cap hit. Default: 5 iterations. Override: AUTOSPEC_RELEASE_MAX_ITERATIONS=<n>. On cap, emit PARTIAL with the unresolved findings and the exact next remediation command.
   • ~/.autospec/stop.flag is present (operator halt). Emit PARTIAL with the current state.
   • No-progress detector: two consecutive iterations produced an identical set of release_blocking findings (by category + summary). Emit FAIL with the stuck findings and an escalation note.

   Verdict values:

   • PASS: all gates green; verdict file confirms live-app proof; no blocking findings remain.
   • PARTIAL: useful progress landed but the iteration cap or operator halt ended the loop with at least one non-FAIL blocking finding still open.
   • FAIL: at least one release_blocking finding has status: FAIL after the loop terminated, or the no-progress detector tripped, or the loop aborted on infrastructure failure.

Spec supersession (recency)

When docs/specs/ contains two specs that overlap on a behavior, the spec whose last-modifying commit on the current branch is most recent wins (issue #635: implicit-by-recency supersession; untracked specs fall back to filesystem mtime). Operators do NOT write Supersedes: frontmatter — recency alone decides.

Release verdict computation MUST consult the resolver before treating any finding as release_blocking. A finding whose source_spec is older than another spec covering the same behavior is downgraded from release_blocking to informational — the newer spec has already replaced the requirement.

Use the shared helper for every finding under review:

bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/resolve-spec-supersession.sh" "<behavior-key>"

If the resolver returns a spec path different from the finding's source_spec, the finding is referencing a superseded behavior. Downgrade the severity and annotate the finding with the winning spec path so the operator can confirm the supersession intent. Do NOT block the release on a superseded finding.

This rule applies to release report rows, QA proof-matrix carryover, and legacy-cleanup blockers alike.

Legacy cleanup prompt

Use this prompt whenever code, config, docs, tests, specs, infra, or examples mention a deprecated path:

Before accepting this release:

1. Is this path still reachable in the product, CLI, API, deployment, tests, or
   docs?
2. Is it the preferred path, a compatibility shim, or dead code?
3. Does the spec still describe it as current behavior?
4. Does any test fixture, README, runbook, env var, script, or infra module keep
   it alive accidentally?
5. What is the replacement path and where is that replacement proven?
6. Can I delete it now? If not, what exact issue tracks removal and what event
   will make removal safe?

Required outcome:
- Preferred path: keep it and prove it.
- Compatibility path: label and document it as compatibility only.
- Dead path: delete it and remove matching spec/docs/tests references.
- Unknown path: mark release PARTIAL and create the smallest investigation or
  cleanup issue needed to settle it.

Worktree guard (commit gate)

Before committing any remediation fix (stage 9 iteration) or any other change produced by this skill, run worktree-guard.sh assert and require exit 0. A non-zero exit (in_primary_checkout / dirty / stale_base) is NEVER worked around — stop the release flow and surface the emitted code_health:* identifier to the operator.

# MANDATORY assert gate: MUST exit 0 before any commit in the release flow.
if ! bash "${AUTOSPEC_SCRIPTS_DIR:-scripts}/worktree-guard.sh" assert; then
  echo "worktree-guard assert failed (see code_health identifier above); aborting release commit" >&2
  exit 1
fi

Cleanup after a release-flow commit lands: git worktree remove the linked worktree and git worktree prune. Never work in the primary checkout.

Release report contract

Return a concise report with:

• Verdict: PASS, PARTIAL, or FAIL.
• Branch and commit range reviewed.
• Skills/stages run, with command evidence.
• Changed files and simplifications made, if edits were performed.
• Test, QA, docs, review, and legacy-cleanup evidence.
• Remaining blockers with exact issue numbers or next commands.
• Known NOT TESTED areas, separated from tested evidence.

Never say the repo is release-ready unless the evidence above is complete.