autospec-sweep

autospec-sweep (harness-neutral)

Create and maintain the tracked project-level autospec configuration at .autospec/autospec.yml, then use it to run recurring sweeps that keep specs, documentation, tests, and code aligned with reality.

<!-- autospec-block:startup-self-update SKILL_NAME=autospec-sweep -->

Self-update mode

If the feature-request argument is exactly update after trimming whitespace and lowercasing, enter self-update mode and do not run a sweep:

1. Detect whether autospec-sweep is installed in any harness.
2. Run the canonical suite installer once with --update: curl -fsSL https://raw.githubusercontent.com/berlinguyinca/autospec/main/bootstrap.sh | bash -s -- --skill all --harness all --update
3. Show a short before/after summary and stop.

Invocation

/autospec-sweep init
/autospec-sweep configure
/autospec-sweep run

• init creates .autospec/autospec.yml on first use and verifies that every supported harness (Claude Code, Codex CLI, OpenCode) is wired up consistently — shared persistent memory at docs/memory/, a Memory inventory section in AGENTS.md, a CLAUDE.md pointer if missing, and the ~/.claude/projects/<slug>/memory symlink resolving to docs/memory/.
• configure reads current repo findings and edits the tracked config.
• run executes the configured sweep and feeds resulting fixes back through autospec-review, specs, issues, and autospec-run.

Harness detection

Resolve model tiers once at skill start:

1. Claude Code: TIER_A = Opus + maximum thinking; TIER_B = Sonnet + medium.
2. OpenCode: TIER_A = top task model + high reasoning; TIER_B = smaller task model + medium.
3. Codex CLI: TIER_A = current top GPT + high reasoning; TIER_B = current spark/cost-optimized Codex + medium.

If TIER_B is unavailable, silently retry the same delegated work with TIER_A while preserving the task handoff.

Required capabilities & harness adapter

Capability	Claude Code	OpenCode	Codex CLI	Fallback if missing
Repo inspection	Agent Explore	read-only task	rg, git, gh	Inspect locally with bounded reads
Foreground delegation	Agent	task	native subagent/CLI	Do the synthesis in-thread
Ask questions	AskUserQuestion	inline prompt	inline prompt	Ask one concise question
Subagent model tier	Tier A/B per harness detection	Tier A/B per harness detection	Tier A/B per harness detection	Fall back UP, never down
<!-- autospec-block:harness-adapter-core -->

First-run config

Before any sweep work, check for .autospec/autospec.yml.

If it is missing:

1. Run: bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/autospec-sweep-wizard.sh" init
2. Ask only the wizard's small starter set unless repo findings require a project-specific question:
   • sweep profile, default full
   • environment safety, default strict_isolation
   • team personality, default auto
   • competitor research, default false
3. Commit .autospec/autospec.yml with the next autospec spec/config commit. It is project source, not local runtime state.

If the config exists, read it first and treat it as the source of truth. Existing .autospec/test.yml or .autospec/clone.yml files are compatibility inputs only; do not make them the new source of truth.

Cross-harness setup verification (init mode)

init always runs the harness-setup doctor after the wizard, regardless of whether .autospec/autospec.yml already existed. The doctor is idempotent — re-running it on a fully-configured repo is a fast no-op:

bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/verify-harness-setup.sh"

It ensures all three supported harnesses see the same project guidance and the same persistent memory:

Harness	Entry file	Memory source
Claude Code	CLAUDE.md	~/.claude/projects/<slug>/memory → symlink → docs/memory/
Codex CLI	AGENTS.md	docs/memory/ (linked from AGENTS.md ## Memory inventory)
OpenCode	AGENTS.md	docs/memory/ (linked from AGENTS.md ## Memory inventory)

The doctor delegates the memory migration + AGENTS.md inventory block to auto-init-memory.sh, then verifies five invariants:

1. docs/memory/ exists and contains a MEMORY.md index.
2. ~/.claude/projects/<slug>/memory is a symlink resolving to docs/memory/.
3. AGENTS.md exists at the repo root.
4. AGENTS.md contains a ## Memory inventory section.
5. CLAUDE.md exists at the repo root — created as a one-line pointer to AGENTS.md if missing, so Claude Code converges on the same source of truth.

Run verify-harness-setup.sh --check (read-only) as a doctor at any time; run without flags to apply fixes. Exit code is non-zero only when an unfixable invariant fails.

Configure mode

Use this mode when the user wants to modify sweep behavior or when the repo has changed enough that the config is stale.

1. Inspect tracked specs, AGENTS.md, package manifests, test commands, E2E config, deployment docs, and existing .autospec/* compatibility files.
2. Compare findings with .autospec/autospec.yml.
3. Ask the minimum project-specific questions needed to resolve missing values.
4. Edit only the relevant config sections.
5. If a config change contradicts a tracked spec, update the spec in the same autospec design/spec PR before any implementation issue is filed.

Parallel area dispatch

Sweep runs fan out into 4 parallel area subagents per the AGENTS.md subagent dispatch decision matrix. Areas are defined under skills/autospec-sweep/areas/ and dispatched via sweep-area-dispatch.sh:

1. spec-vs-code-drift — reuses explore-research/spec-vs-code.sh.
2. docs-drift — reuses dogfood-adapter-doc-drift.sh, which invokes /autospec-doc --full (full regen + repo-wide completeness audit via doc-orchestrator.mjs) and surfaces missing parts as findings.
3. code-health — reuses explore-research/codebase-signals.sh.
4. dependency-health — new explore-research/dependency-health.sh researcher (also available to autospec-explore — extends the 6-researcher set to 7).

Each area subagent emits research-cycle JSON. The dispatcher aggregates into .autospec/sweep/area-findings.json (schema: autospec-sweep.area-findings.v1) which feeds the existing sweep report and autospec-review.

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

Sweep execution

Model tier: TIER_A for sweep synthesis and gap classification. Any fix implementation created by downstream autospec-run uses that skill's Phase 4 Tier B/Tier A routing.

Start with the executable runner:

bash "${AUTOSPEC_SCRIPTS_DIR:-$HOME/.autospec/scripts}/autospec-sweep-run.sh" run

The runner validates .autospec/autospec.yml, writes .autospec/sweep/latest.json, accepts --gaps <json> for already-emitted gap contracts, and executes the bundled autospec-sweep-review.sh by default. Set AUTOSPEC_SWEEP_REVIEW_CMD="<command>" to replace the deterministic wrapper with a harness-specific or richer review command.

Run the sweep from current reality back to source artifacts:

1. Read .autospec/autospec.yml.
2. Run the enabled review surfaces: spec coverage, docs drift, test gaps, code health, security footguns, clone/test readiness, and implementation status.
3. Emit concrete gaps with file paths, commands, and expected behavior.
4. Update or create specs before creating code-fix issues when reality and specs disagree.
5. File bounded auto-implement issues for accepted gaps and hand them to autospec-run.

Continuous improvement

The default sweep is not a one-time bug hunt. It continuously improves:

• Documentation: keep user docs, API docs, runbooks, and generated docs in sync with implemented behavior.
• Documentation targets: treat documentation.audiences[] and documentation.scopes[] in .autospec/autospec.yml as the source of truth for deep docs. Emit separate gaps for missing target files or missing required autospec-doc-scope markers so user, developer, operator, security, runbook, API, and troubleshooting docs can be built as bounded reviewable work.
• Tests: add missing regression, unit, integration, E2E, invariant, and negative path coverage where the sweep finds weak confidence.
• Code: reduce duplication, dead code, complexity, brittle boundaries, and security-sensitive shortcuts in small reviewable issues.

Every sweep run must execute the configured full test commands. Do not treat review-only sweeps as sufficient. If E2E or integration tests require the software to be deployed or started first, run project.findings.commands.deploy before those tests. A failing deploy, unit, integration, or E2E command fails the sweep and must be fixed before filing success claims.

Prefer small, reversible changes. File specs and issues for improvement work; do not silently refactor broad surfaces inside the sweep itself.

Spec sync

Specs must stay synchronized with reality:

• If code behavior is correct and the spec is stale, update the spec first.
• If the spec is correct and code behavior is stale, file implementation issues.
• If both are ambiguous, ask one project-specific question and record the answer in the config or spec before proceeding.

Safety

Never clone raw secrets or API keys. The config may store secret references, environment variable names, and allowlisted settings, but not secret values. Production-like sweeps default to strict isolation. Scoped production access requires explicit config, backup/restore, and the existing autospec-test safety rails.

Verification

After changing config or sweep behavior, verify:

autospec validate
bats tests/unit/test_autospec_sweep_config.bats
bats tests/unit/test_autospec_sweep_run.bats
ajv compile -s schemas/autospec-config.schema.json --spec=draft2020

Report changed config sections, generated questions, and any remaining unknowns.