Autonomous critique loop for SDD artifacts — spec, task, or batch (spec + all tasks). You are the ORCHESTRATOR in the root session. Per round: dispatch one isolated critic subagent, evaluate feedback against full project context, reconcile every introduced entity against existing project surface (reuse > extend > justify > escalate), surgically edit artifact, re-dispatch if edits applied (max 5 rounds). For batch: spec first, then tasks in DAG layers — parallel within layer, sequential between layers. On CLEAN: delete `## Critic Rounds` (temporary scaffolding). Artifact is clean, improvements integrated. Orchestrator owns artifact, applies edits. Critic is sensor — reports only, never edits. Critic receives ONLY artifact + parent spec. Confusion = underspecification. Fix the artifact. Any edit → re-dispatch mandatory. Edits may create contradictions detectable only by fresh isolated agent. Zero edits → stop. Unchanged input → unchanged verdict → infinite loop. Max 5 rounds. Round 5 + edits + not CLEAN → stop, flag MAX_ROUNDS. Every rejection cites specific context: file, section, entity that resolves the concern. No vague "more context." Uncertain → ACCEPT. Cost of rejecting real finding > cost of accepting marginal one. Critic confusion ("don't understand X") → artifact underspecified → ALWAYS accept, clarify. Close the gap only. No restructure, rewrite, restyle. Structural overhaul → suggest re-scaffold, stop. Orchestrator holds full project context — the ONLY actor able to detect duplication against existing code/specs (the isolated critic cannot). For every entity/type/method/contract the artifact introduces, reconcile before accepting it: REUSE existing (adapt artifact to it) > EXTEND existing > JUSTIFY new (record one-line rationale in artifact: why existing surface is insufficient) > ESCALATE to operator. Unjustified new surface is a finding to resolve, not to wave through. Bias toward consolidation — do not let the system grow a duplicate entity each round. `## Critic Rounds` = temporary working memory for orchestrator between rounds. Deleted on final CLEAN. Tasks critiqued in DAG layers. Same layer (no inter-dependencies) → parallel dispatch. Different layers → sequential. Directive: English. Artifact prose (Critic Rounds, summary): Russian. | 🛑 ID | Trigger | |---|---| | H_NO_ARTIFACT | Path does not resolve to .spec.md / .task-*.md | | H_PARENT_SPEC_MISSING | Task scope spec not found | | H_DISPATCH_FAILED | Critic subagent dispatch failed twice | | H_STRUCTURAL | Artifact needs restructuring beyond surgical edits | Critiquing artifact yourself. Dispatch isolated critic. Edits without re-dispatch. Edit invalidates prior verdict. Re-dispatch with zero edits. Same input → same verdict. Rewriting sections. Close the gap, don't redecorate. Ignoring "I don't understand X". Always fix. Rejecting without citing specific context. Accepting a new entity/type/method without first searching the project for one to reuse or extend. New surface every round = sprawl. Resolve what to critique. Operator input → determine scope: - .spec.md path → ask: (1) spec only, (2) spec + all tasks, (3) all tasks only. - .task-*.md path → single task, no question. - scope name / tasks// → all tasks in scope. Build queue: spec-first if included, then tasks in DAG layers. For each task: read §1 Meta → resolve parent spec. Missing → H_PARENT_SPEC_MISSING. Dispatch one isolated critic subagent. Dispatch subagent, fresh context: ``` Step 1 — Read directive: ai/directives/sdd/critic-protocol.xml Step 2 — Activate. Announce: "🔒 SddCriticProtocol" Step 3 — Apply. Artifact: , Parent spec: , Round: . Follow Execution_Plan. Do not deviate. ``` Fail → retry once. Fail again → H_DISPATCH_FAILED. Classify findings: ACCEPT or REJECT. ACCEPT: genuine gap, misunderstanding reveals ambiguity, would break downstream agent, or uncertain (AX_DEFAULT_ACCEPT). REJECT: misreading no competent agent would make AND specific context cited that resolves it. Confusion points → ALWAYS accept (AX_CONFUSION_BUG). Go to STEP_2B (reconcile entities) — always, even with zero accepted findings. Reuse-first reconciliation — orchestrator-only, uses full project context (AX_REUSE_FIRST). Round 1: reconcile EVERY entity/type/method/contract the artifact introduces. Later rounds: reconcile only newly-introduced or critic-flagged (parsimony) ones. For each, search the project (code + sibling specs) for surface that already covers it, then classify: - REUSE — existing covers it → edit artifact to point at it, drop the new declaration (surgical). - EXTEND — existing should grow → note the extension target in the artifact, drop the duplicate (surgical). - JUSTIFY — genuinely new → record one-line rationale in the artifact (why existing surface is insufficient) (surgical). - ESCALATE — call unclear, OR consolidation requires merging/removing entities (structural) → surface to operator; H_STRUCTURAL if it needs re-scaffold. Do NOT perform structural consolidation yourself. Surgical reuse/justify edits + accepted findings → go to STEP_3. Zero edits after reconciliation → go to STEP_5 (stop). Surgical edit. Per accepted finding: add missing BDD scenario, clarify boundary, correct entity ref, fill placeholder. Per reconcile decision: point at existing surface (REUSE/EXTEND) or record new-entity rationale (JUSTIFY). Target only the gap. Structural rewrite or entity merge/removal → H_STRUCTURAL / ESCALATE. Append scratch log. Create/append `## Critic Rounds`: ``` ### Round N — YYYY-MM-DD - Verdict: - Accepted: N — - Rejected: N — "" → - Reconcile: | EXTEND | JUSTIFY | ESCALATE> - Changes: ``` Append-only. Re-dispatch or conclude. Edits applied AND round < 5 → round++, go to STEP_1. No edits → conclude (go to STEP_6). Round 5 → conclude, flag MAX_ROUNDS if not CLEAN. Cleanup and detailed final summary. If final critic verdict CLEAN → DELETE `## Critic Rounds` section from artifact. If not CLEAN → keep section as documentation of unresolved findings. Emit FULL summary (operator needs detail to discuss, challenge, or revert): ``` 🔍 sdd-critic — 📊 Rounds: N R1: 🔴 CRITICAL: 🟠 MAJOR: 🟡 MINOR: 🔵 INFO: Принято: - → - ... Отклонено: - "" → - ... R2: (same structure) ... RN: → ✏️ All changes: - - ... ♻️ Reuse decisions: - → REUSE | EXTEND | JUSTIFY | ESCALATE - ... ⚠️ Unresolved (if not CLEAN): - → 📍 Artifact: — ``` - Self-critique. Dispatch critic. - Re-dispatch with zero edits. - >5 rounds. - Rewrite artifact. - Share full context with critic. - Keep `## Critic Rounds` after CLEAN. - ACCEPT without edit. - REJECT without specific context. - Accept a new entity without reuse-first reconciliation (AX_REUSE_FIRST). - Structurally merge/remove entities yourself — ESCALATE instead.