aidd-loop

Follow feature-dev-aidd:aidd-core for output contract and DocOps.

Loop discipline

1. Read order: loop pack -> review pack (if any) -> fix_plan.json when verdict=REVISE -> rolling context pack.
2. Excerpt-first: do not read full PRD/plan/tasklist/spec when the loop pack excerpt covers Goal/DoD/Boundaries/Expected paths/Size budget/Tests/Acceptance.
3. No questions in loop-mode; record blockers/handoff instead.
4. REVISE: reuse the same scope, follow fix_plan.json, and do not widen boundaries.
5. Scope guard: out-of-scope -> WARN + handoff; forbidden -> BLOCKED. Never expand boundaries yourself.
6. Tests: follow loop pack or reviewer policy; do not invent new test requirements.
7. No large logs/diffs in chat; link to aidd/reports/**.
8. Block policy: strict fail-fast by default; strict_recoverable_reason_codes may explicitly enable bounded retries for selected reasons (for example no_tests_hard).
9. ralph uses policy matrix v2 (hard_block|recoverable_retry|warn_continue) with bounded retries for recoverable reasons.
10. For recoverable blocked retries, emit observability fields: recoverable_blocked, recovery_path, retry_attempt, plus ralph_reason_class|strict_recoverable_reason_class.
11. Scope mismatch fallback is telemetry-only (scope_mismatch_non_authoritative=1): emit expected_scope_key/selected_scope_key; do not treat mismatched fallback as authoritative transition input.

Preload matrix v2

• Roles: implementer, reviewer, qa.
• This skill must not be preloaded for other subagents.

Canonical shared Python entrypoints

• python3 ${CLAUDE_PLUGIN_ROOT}/skills/aidd-loop/runtime/loop_pack.py
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/aidd-loop/runtime/loop_step.py
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/aidd-loop/runtime/loop_run.py
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/aidd-loop/runtime/preflight_result_validate.py
• python3 ${CLAUDE_PLUGIN_ROOT}/skills/aidd-loop/runtime/output_contract.py
• Preflight preparation is stage-chain-internal and not an operator command.

Command contracts

python3 ${CLAUDE_PLUGIN_ROOT}/skills/aidd-loop/runtime/loop_pack.py

• When to run: before loop stage execution to assemble bounded evidence input.
• Inputs: ticket/scope context and available stage artifacts (readmap, review pack, context pack).
• Outputs: deterministic loop pack summary for implement/review/qa orchestration.
• Failure mode: non-zero exit when required upstream artifacts are missing or malformed.
• Next action: repair missing prerequisites and rerun pack generation.

python3 ${CLAUDE_PLUGIN_ROOT}/skills/aidd-loop/runtime/loop_run.py

• When to run: stage-chain-orchestrated loop execution for bounded stage retries.
• Inputs: loop mode/profile, ticket/scope context, and stage-specific payloads.
• Outputs: loop execution status with deterministic retry/blocked metadata.
• Failure mode: non-zero exit on invalid loop payload, policy violation, or unrecoverable blocked condition.
• Next action: apply fix plan or handoff policy action, then rerun with unchanged scope.

Additional resources

• Loop stage-chain reference: reference.md (when: stage-chain path/fallback behavior needs clarification; why: confirm canonical chain and bounded recovery policy).
• Loop pack template source: templates/loop-pack.template.md (when: loop pack format is unclear; why: confirm canonical sections seeded by init/runtime).