Workflow Review Changes Loop

[BLOCKING] Execute skill steps in declared order. NEVER skip, reorder, or merge steps without explicit user approval. [BLOCKING] Before each step or sub-skill call, update task tracking: set in_progress when step starts, set completed when step ends. [BLOCKING] Every completed/skipped step MUST include brief evidence or explicit skip reason. [BLOCKING] If Task tools are unavailable, create and maintain an equivalent step-by-step plan tracker with the same status transitions.

Quick Summary

Goal: Converge a review scope to a clean no-op pass by re-running the ENTIRE /workflow-review-changes workflow INLINE, round after round, over a fixed scope combined with the fixes accumulated so far — stopping when a complete round applies zero fixes. From round 2 the severity floor applies: LOW findings are no longer fixed, so a round that surfaces only LOW findings is a zero-fix round and ENDS the loop.

Summary:

• Steps (in order): (0) resolve scope + Goal Contract → (0b) bind the convergence loop (protocol loop primary + optional /goal accelerator) → (1) round loop { run /workflow-review-changes INLINE → detect fixes-applied → log iteration } → (2) converge on a zero-fix round OR escalate on non-progress → (3) recap.
• Convergence: stop ONLY when a whole round applies zero fixes (its fix cycle skipped, working tree unchanged, reviews clean) — not merely one clean review.
• Inline invariant: run /workflow-review-changes via the Skill tool, NEVER the Agent tool — it self-binds its own review-loop obligation (owning the session Stop hook for its /goal gate WHEN available), which a sub-agent cannot own or carry back to this loop.
• Severity floor — from round 2, LOW stops blocking. Round 1 fix every validated severity. From round 2 tell the inner /workflow-review-changes to fix only CRITICAL/HIGH/MEDIUM and to defer LOW findings — a round whose validated findings are ALL LOW therefore applies zero fixes and CONVERGES. Carry every deferred LOW into the recap and Goal Contract; NEVER re-tier a real CRITICAL/HIGH/MEDIUM down to LOW to reach the exit. Severity tiers per SYNC:severity-rubric.
• Bounded: round cap default 3; blocking findings not shrinking across 2 rounds, or cap hit with CRITICAL/HIGH/MEDIUM fixes still landing → STOP & escalate via AskUserQuestion.

Why this skill exists (READ FIRST — it is the whole justification): /workflow-review-changes already converges internally to a clean pass, BUT its inner loop does NOT re-run the 7 specialist reviewers from scratch. Per the Conditional Inline Re-Review Protocol in .claude/skills/workflow-review-changes/SKILL.md, the step-15 inline re-review re-runs only /changes-review's own BE/FE/SCSS/UI dimensions plus scoped re-runs of the specific specialist that raised a finding — /architecture-review, /performance-review, /security-review, /integration-test-review, /production-readiness-review, /domain-entities-review, and /ui-review fire ONCE (steps 4–10). Only a fresh full re-invocation re-runs ALL specialists over the now-fixed code — catching second-order defects the fixes themselves introduced and killing whole-workflow confirmation bias. Without this outer loop those regressions ship unreviewed.

Workflow: resolve scope + Goal Contract → bind the convergence loop (protocol loop + optional /goal accelerator) → round loop { run /workflow-review-changes INLINE → detect fixes-applied → log iteration } → converge when a round applies zero fixes → recap.

Key Rules:

• MUST run INLINE in the main session — NEVER dispatch /workflow-review-changes as a sub-agent. It self-binds its own review-loop obligation (owning the session Stop hook for its /goal gate when available); as a sub-agent that in-session guarantee is silently lost (see the [WORKFLOW-IN-WORKFLOW] execution block in workflow-review-changes/SKILL.md). This loop skill therefore also runs inline.
• Convergence = a whole round applied ZERO fixes (its fix cycle, steps 12–15, was skipped because reviews passed clean). That, not "one clean review", ends the loop.
• Scope base is FIXED across rounds; the working tree grows. Recompute the scope each round as branch-diff base ∪ current uncommitted changes — the diff base never moves, so convergence is measured against a stable target.
• Round cap (default 3) and findings-increasing → STOP & escalate via AskUserQuestion. NEVER loop open-ended. Cap exhaustion escalates only when CRITICAL/HIGH/MEDIUM fixes are still landing — a LOW-only round converges via the severity floor.
• The severity floor bounds ITERATION, never the standard. It ends the loop; it never authorizes shipping a known CRITICAL/HIGH/MEDIUM, and it never applies to a binary gate inside the inner workflow (a failing test is a failure, not a LOW finding).

First Principle — Convergence, Not Motion

A round that changes files is progress only if the next round finds fewer things to fix. The loop exists to reach a fixed point (zero fixes), not to keep churning the diff. If findings stop shrinking, that is a signal to escalate, not to spin another round.

Step 0 — Resolve Scope + Goal Contract (FIRST ACTION)

1. Parse the review scope from the user prompt into a stable, reusable scope string. It has two parts UNIONed:
   • Branch-diff base — a branch-to-branch or PR diff, e.g. fix/timelog-org-business-timezone into develop. Capture it as git diff develop...HEAD (three-dot: changes on the feature branch since it forked from develop) so the base is a fixed merge-base, not a moving target.
   • Current changes — the uncommitted working-tree changes (git status --porcelain, git diff + git diff --staged).
   • Scope string (recompute each round): {branch-diff base} ∪ {current uncommitted changes}. The base commit is fixed for the whole loop; the uncommitted set legitimately grows as fixes land.
   • If the prompt names no branch diff (pure "current changes" review), the scope is just the working-tree changes — the loop still applies.
2. Resolve/create the Goal Contract per SYNC:goal-contract-satisfaction-loop (plans/goals/{YYMMDD-HHmm}-{slug}/goal.md, template .claude/templates/goal-contract-template.md). Its single required Success Criterion:

   A complete /workflow-review-changes run over {scope} applies zero fixes (a clean no-op pass — no /plan-execute file changes; round 1 no validated findings of any severity, round 2+ no validated CRITICAL/HIGH/MEDIUM, with remaining LOW findings deferred rather than fixed). Record the round cap (default 3), the severity floor (LOW non-blocking from round 2), and the scope string in Constraints.

Step 0b — Bind the Convergence Loop (protocol-first; /goal is an optional accelerator)

The convergence loop is bound by TWO layers. The protocol loop (Steps 1–2) is the BINDING mechanism and MUST be self-driven by you, the running agent, on every host — with or without any command or hook. The /goal command is an OPTIONAL accelerator layered on top; it is never the primary mechanism, and its absence NEVER weakens the loop. This mirrors the project rule that hooks/trackers are accelerators only — correctness must not depend on them.

1. Protocol loop — ALWAYS binding (hook/command-independent). You are personally responsible for not stopping until the loop converges or bounded-escalates. This binds Claude, Codex, and Copilot equally, whether or not /goal exists:

Repeatedly run /workflow-review-changes INLINE over {scope} (branch-diff base ∪ current uncommitted changes, recomputed each round). After each round, detect whether it applied any fix. From round 2 on, instruct the inner workflow to fix only CRITICAL/HIGH/MEDIUM validated findings and to defer LOW ones. Do NOT stop while the last round still applied fixes. Converge when a full round applies ZERO fixes (reviews clean, /plan-execute changed no files) — which from round 2 includes a round whose only validated findings were LOW. Cap at {N=3} rounds; if blocking findings do not shrink across 2 consecutive rounds, or the cap is hit with CRITICAL/HIGH/MEDIUM fixes still landing → STOP and escalate via AskUserQuestion. Never loop open-ended.

Treat this as a standing obligation you re-read at every Step 2 checkpoint — NOT a one-time note you can rationalize away after the first round. The Goal Contract's required Success Criterion (Step 0) is its durable, host-independent record.

2. /goal command — invoke as an accelerator WHEN AVAILABLE. If a /goal command exists and you are permitted to run it in this environment, ALSO invoke it (a real tool/command call, NOT a paraphrase, NOT a Goal Contract file substituted for it) with the SAME condition, so a session Stop hook mechanically enforces the loop:

/goal review-changes convergence loop: repeatedly run /workflow-review-changes INLINE over {scope} (branch-diff base ∪ current uncommitted changes, recomputed each round). After each round, detect whether it applied any fix; if fixes>0 → run another round; if a full round applied ZERO fixes (reviews clean, /plan-execute changed no files) → CONVERGED, clear the gate. From round 2 on, the inner workflow fixes only CRITICAL/HIGH/MEDIUM and defers LOW, so a LOW-only round is a zero-fix round and CONVERGES. Do NOT stop while the last round still applied fixes. Cap at {N=3} rounds; if blocking findings do not shrink across 2 consecutive rounds, or the round cap is hit with CRITICAL/HIGH/MEDIUM fixes still landing → STOP and escalate via AskUserQuestion. Never loop open-ended.

The /goal Stop hook blocks stopping until the condition holds and auto-clears when met — do not tell the user to clear it.

If /goal is unavailable, unregistered, or not permitted (e.g. Codex/Copilot, or a Claude run without the command): DO NOT error, DO NOT block, and DO NOT invent a stand-in gate. Record ONE line in the Goal Contract — /goal accelerator unavailable — loop bound by protocol (Steps 1–2) + this Goal Contract — and proceed. The protocol loop above plus the Goal Contract are the same gate, enforced by discipline instead of a hook.

Nested gates (by design, safe): each inner /workflow-review-changes round self-binds its OWN Step 0 review-loop obligation (and installs its own /goal gate WHEN available) that clears when that round reaches its internal clean pass. This OUTER loop persists across rounds and subsumes the inner ones (a converged loop implies every inner round ended clean). All self-clear on satisfaction — no orphaned gate. Do NOT tell the user to clear either.

Step 1 — Round Loop (run → detect → log)

For each round R (starting at 1), do ALL of:

1. Snapshot before: record the working-tree fingerprint — git status --porcelain + git diff --stat (or git rev-parse of git stash create for an exact hash). This is the fixes-applied baseline for the round.
2. Run the workflow INLINE: invoke /workflow-review-changes via the Skill tool (NEVER the Agent tool) with the recomputed {scope} as its prompt. Let it run its full 20-step sequence including its own internal fix→re-review loop.
3. Detect fixes-applied (objective): compare the working tree after the round to the before-snapshot AND read the workflow's own result:
   • Fixes applied (>0) if the working tree changed during the round OR the workflow reported its fix cycle (steps 12–15) ran / /plan-execute modified files.
   • Zero fixes if the working tree is byte-identical to the before-snapshot AND the workflow reported either no validated findings, or (round 2+) validated findings that are ALL LOW with the fix phase explicitly skipped.
   • If a claimed LOW-only exit has a changed fingerprint, re-review; it is not a no-op convergence pass.
4. Append an Iteration Log entry to the Goal Contract: round number, files changed this round (file:line), fixes-applied count, the review verdict, and remaining gaps.

Step 2 — Convergence & Escalation Gate

Evaluate after every round:

Increasing findings = STOP. If round R surfaces MORE findings than round R-1, the fixes are regressing the code — STOP and escalate immediately (mirrors the Issue count increasing rule under Iteration Tracking (Conversation-Scoped) in workflow-review-changes/SKILL.md). Never trade one fix for two new findings across rounds.

Step 3 — Recap

Emit a concise convergence recap: rounds run, total fixes applied per round (the shrinking sequence), the final clean-pass evidence, and the Goal Satisfaction matrix (required criterion PASS). Point to each round's report under plans/reports/ and the Goal Contract Iteration Log. Do NOT commit or push unless the user explicitly asks.

Convergence Detection — Why Two Conditions

A round counts as converged ONLY when both hold: (a) the working tree is unchanged by the round, AND (b) the reviews reported clean at that round's bar — no validated findings in round 1, no validated CRITICAL/HIGH/MEDIUM from round 2 (deferred LOWs listed, not fixed). Both are required because:

• Working-tree-unchanged alone is ambiguous — a round can make no changes because a finding was unfixable/escalated, not because it was clean. That is escalation, not convergence.
• Reviews-clean alone is insufficient — an orchestrator can rationalize a "clean" verdict; the objective git-diff comparison is the backstop that proves no fix actually landed.

When (a) is true but (b) is false → escalate (a real finding the loop cannot close). When (b) is true but (a) is false → the round DID fix things → run another round to re-prove clean.

IMPORTANT MANDATORY sequence: Step 0 (scope + Goal Contract) → Step 0b (bind the convergence loop: protocol loop primary + optional /goal accelerator) → Step 1 (round loop: run /workflow-review-changes INLINE → detect fixes → log) → Step 2 (converge on zero-fix round / escalate on non-progress) → Step 3 (recap).

Executable review policy — one predicate, one durable transition model. Review skills and their tooling MUST use the canonical helper .claude/scripts/lib/review-policy.cjs (policy version 2) for round eligibility. The helper's blockingFindings(round, findings, hardGates) predicate returns every validated finding in round 1, and only CRITICAL/HIGH/MEDIUM findings in rounds 2–3; NOT VERIFIABLE is a separate unresolved-evidence state that remains blocking at every round. Failed binary gates are synthetic blocking findings at every round. evaluateRound retains round-2+ LOWs in deferredLow and never treats a LOW-only round as blocked after the floor applies. Severity is assigned before the predicate and never changed to obtain a PASS.

Round and minimum rules. MAX_ROUNDS is 3 and is a ceiling, never a target. A clean review ends once round >= minRounds; the default minimum is 1. A caller may declare minRounds explicitly (for example, 2) when an independent second pass is a real requirement; the declaration is persisted and cannot be inferred from a round counter. A failing test-green, security-must-fix, required-artifact, or other binary gate is never waived by the severity floor.

Durable run record. A review run MUST identify runId, target fingerprint, policy version, target revision, minimum/maximum rounds, completed rounds, full findings/gate evidence, interruption/resume metadata, and acceptance. Use the atomic, lock-serialized transitions in review-policy.cjs: start, record, accept, interrupt, resume, invalidate, and check. Repeating an identical completed round is idempotent and MUST NOT consume budget twice. A changed target fingerprint invalidates prior evidence and acceptance but MUST preserve the bounded round budget; stale evidence cannot be accepted. A policy-version change (including the round-2 LOW floor) invalidates old records; start a new run rather than interpreting old evidence under new semantics. Interrupted/resumed runs retain completed rounds and findings. The record is bookkeeping, not consent, native permission, or proof that a host actually performed the review.

CLI boundary. The helper CLI accepts JSON on stdin and uses its own real clock; a supplied now is rejected. State directories must be absolute, non-root real directories, records are size-bounded, and malformed/locked state fails closed for the transition. Full reports remain on disk; an inline result envelope is only a transport summary. Any new review policy consumer must add a semantic fixture, boundary counter-cases, a seeded mutant, and a report with the target fingerprint and command exit status.

Goal Contract Satisfaction Loop — Persist the user goal in an external file, execute against it, and loop review/fix until every saved required criterion passes or a blocker escalates. Bounded closed loop — NEVER open-ended autonomous exploration.

1. Resolve the active goal (in order): active plan goal.md → plans/goals/{YYMMDD-HHmm}-{slug}/goal.md → create a new Goal Contract from the current user request (template: .claude/templates/goal-contract-template.md).
2. Required sections: Original Request, Purpose, Success Criteria (checkboxes; mark required vs optional), Constraints, Evidence Required, Iteration Log, Goal Satisfaction matrix.
3. Before work: read the active goal and map planned work to saved success criteria — execution serves the saved criteria, never chat memory alone.
4. After execution/verification: append an Iteration Log entry — result, evidence references (file:line, command output, report path), remaining gaps.
5. Review gate: emit a Goal Satisfaction matrix — | Success Criterion | Evidence | Status | with PASS/FAIL/BLOCKED. Overall PASS requires every required criterion PASS.
6. Loop rule (retry): required criterion FAIL → validate the gap is real → fix → re-review only the affected criteria. Stop cleanly when all required criteria PASS.
7. Escalation rule (stop): two consecutive iterations with no criterion progressing, or a blocker needing user input → mark the criterion BLOCKED with a user-facing reason and escalate. NEVER loop indefinitely.
8. Skip rule: tiny conversational tasks may skip the goal file ONLY with a recorded one-line reason. User-accepted gate skips are recorded in the goal file with reason and scope.
9. Security: NEVER store secrets, tokens, credentials, or private customer data in goal files — store evidence references and redact sensitive values.

Blocked until: active goal resolved (or skip reason recorded) · saved success criteria read before edits · iteration evidence appended after execution · Goal Satisfaction matrix emitted before any PASS verdict.

Critical Thinking Mindset — Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. Anti-hallucination: Never present guess as fact — cite sources for every claim, admit uncertainty freely, self-check output for errors, cross-reference independently, stay skeptical of own confidence — certainty without evidence root of all hallucination.

AI Mistake Prevention — Failure modes to avoid on every task:

Re-read files after context changes. Context compaction, resume, or long-running work can make memory stale; verify current files before acting. Verify generated content against source evidence. AI hallucinates APIs, names, claims, and document facts. Check the relevant source before documenting or referencing. Check downstream references before deleting or renaming. Removing an artifact can stale docs, generated mirrors, configs, and callers; map references first. Trace the full impact chain after edits. Changing a definition can miss derived outputs and consumers. Follow the affected chain before declaring done. Verify ALL affected outputs, not just the first. One green check is not all green checks; validate every output surface the change can affect. Assume existing values are intentional — ask WHY before changing OR flagging one as a defect. Before changing or reporting a constant, limit, flag, cutoff, wording, or pattern, read nearby context and history, the CALLER's ordering, and 2+ sibling call sites of the same convention. A doc stating WHAT without WHY is missing rationale, not proof of a missing guard. Surface ambiguity before acting — don't pick silently. Multiple valid interpretations require an explicit question or stated assumption with risk. Assert the outcome your system owns, not the intermediate state your infrastructure owns. When verifying async work, assert the final business state — never the delivery/retry bookkeeping held in shared infrastructure that any co-running process can write. Such a check passes when run alone and flakes the moment anything else shares that infrastructure. Keep shared guidance role-relevant. Universal guidance must help every receiving skill or agent; code-specific obligations belong only in code-specific protocols.

Trade-Off Interrogation Gate — ALWAYS ask these THREE questions before ANY verdict, score, finding, or recommendation — about the thing under review AND about every recommendation YOU make. — why: naming a benefit without its price is an endorsement, not a review; the costliest trade-offs are the ones nobody wrote down.

1. Is there any trade-off? Name what it SACRIFICES. "None" / "pure win" is an unfinished analysis, NOT an answer — to claim none, state which dimensions you checked and why each is unaffected: future change cost · complexity · performance/latency · memory/cost · coupling · reversibility · migration burden · operational load · blast radius · security posture · testability · team skill/ramp · delivery time · UX.
2. Is it worth it? Weigh gain against sacrifice EXPLICITLY — what is gained (with a metric) · what it costs · WHO pays · WHEN it comes due — then emit WORTH IT / NOT WORTH IT / UNCLEAR. "Better" with no metric and no cost FAILS this question. NOT WORTH IT → withdraw or replace the recommendation, never keep it as-is.
3. Is the trade-off material enough to CONFIRM WITH THE USER? A material trade-off is the user's call, never yours. MATERIAL when ANY holds: irreversible / one-way door (data migration, public contract, storage format, vendor lock-in) · cost shifted onto someone else (another team, ops/on-call, future maintainer, end user) · one quality attribute traded for another (correctness↔speed, security↔convenience, latency↔cost, simplicity↔flexibility) · a boundary crossed (client↔server tier, service contract, event contract, shared library) · a high-consequence path (auth, money, data integrity, breaking change, High/Medium residual risk) · the worth-it verdict is UNCLEAR.

MATERIAL → STOP and confirm via AskUserQuestion BEFORE the verdict stands — state the trade-off, both options, what each sacrifices, and your recommendation. NOT material → record it inline with a one-line justification and proceed.

Non-asking execution contexts — ESCALATE BY HANDOFF, never by silence. AskUserQuestion reaches only the main interactive agent: a sub-agent cannot ask the user, and a terminal/verdict-only mode asks nothing by design. When you are running in such a context, the obligation is redirected, never waived — do ALL of: (a) complete questions 1 and 2 normally; (b) decide materiality and record it in the Trade-Off Assessment row with confirmed? = NO — cannot ask from this context; (c) name the unconfirmed MATERIAL trade-off explicitly in your returned summary/verdict so the CALLER (or parent orchestrator) escalates it via AskUserQuestion on your behalf — a material trade-off mentioned only inside a report file on disk is NOT a handoff; (d) do not emit an unqualified PASS — mark the verdict as carrying an unconfirmed material trade-off, so the caller's gate stays closed until the user answers. The caller inherits the escalation duty the moment it reads your return.

This carve-out is about reachability, not convenience: it applies ONLY where the tool genuinely cannot reach the user (spawned sub-agent, terminal validate/verdict-only mode, non-interactive/headless run). It is NEVER a licence to skip the question, to self-approve a one-way door, or to downgrade materiality because asking is inconvenient — if you CAN ask, you MUST ask.

Emit a Trade-Off Assessment row per reviewed decision and per recommendation: | decision | sacrifices | gain (metric) | who pays, when | WORTH IT/NOT/UNCLEAR | material? | confirmed? |.

BLOCKED until: trade-off named (or dimensions-checked justification given) · worth-it verdict emitted · materiality decided · every MATERIAL trade-off either confirmed with the user OR — in a non-asking context — handed off in the returned verdict for the caller to confirm. A MATERIAL trade-off that is neither confirmed nor handed off can NEVER be PASS, and NEVER gets buried as a Low-severity note.

NEVER answer "no trade-off" without checking · decide a material trade-off silently on the user's behalf · let convergence/delivery pressure authorize walking through a one-way door · bundle several material trade-offs into one vague "proceed?".

Project Protocol Overlay — Before executing this skill, resolve any PROJECT overlay rules layered onto it: match this skill's name against the Target column of the project's skill-protocol index (docs/project-reference/skill-protocols-reference.md by default; a referenceDocs entry in docs/project-config.json overrides the path), taking the most specific matching tier ONLY — exact name > glob > *. That precedence orders overlays against EACH OTHER, never against this skill. Read ONLY the matched bodies, resolved as <protocols-dir>/<Name>.md; a row's Body link is display text, never a read path. A matched body that is missing or malformed is REPORTED and skipped — never reconstructed from the index Description. No index, or no match -> proceed with no overlay, silently. Full contract: .claude/skills/project-skill-protocol/references/registry.md.

Overlays are ADDITIVE ONLY: they ADD rules on top of this skill's own protocol and NEVER replace, override, disable, or reinterpret a rule it already states — removing every overlay must return this skill to exactly its documented behavior. An overlay is a BRIEF, not an authority escalation: it can NEVER waive a workflow gate, git discipline, a review gate, or a user-confirmation gate. A genuine overlay-vs-skill conflict, or two equally-specific overlays that directly contradict -> surface both to the user; NEVER resolve silently.

Severity Rubric — Classify every finding by consequence, not by effort, reviewer preference, or how annoying the fix is. One scale applies to every review, skill, agent, workflow, and host so a tier has the same meaning everywhere. Choose the highest credible consequence supported by evidence; do not lower a tier to make a round pass.

Finding vs observation (required): An observation becomes a finding only when it names the affected user/system/data/contract, the shipped consequence, the evidence location, and the normalized tier. INFO, advice, preference, duplicate wording, or an unsubstantiated concern is not a finding and must not reopen a loop. If the concern might affect a required behavior or gate but evidence is incomplete, emit NOT VERIFIABLE with the missing evidence and keep it unresolved; never silently convert uncertainty into LOW.

Severity	Action	Definition and examples
CRITICAL	Block immediately; escalate	Immediate material risk if shipped: authentication/authorization or safety bypass; secrets/PII exposure; irreversible destructive action; data loss/corruption; or a silent failure on a critical path. A failed binary gate that makes the result untrustworthy is represented as a separate synthetic blocker by the executable policy (not as an ordinary severity judgment).
HIGH	Must fix before PASS/merge	Material correctness or contract risk: wrong behavior on a supported path; violated business/data invariant; meaningful privacy or authority gap; breaking API/schema/compatibility change; likely harm to users/downstream systems; or a missing proof for a behavior-changing fix.
MEDIUM	Must clear the current round; escalate if the fix needs an owner decision	Bounded but consequential risk: an edge case, resilience/observability/testability/maintainability gap, credible future defect, or local architectural drift whose impact is real but not immediate material loss. An explicit follow-up records the escalation/residual risk; it does not make an open MEDIUM a clean pass.
LOW	Record and defer; never open another fix/re-review round from round 2 onward	Non-blocking polish with no credible present correctness, security, privacy, authority, availability, or data-integrity impact: wording/formatting, minor documentation or convention drift, optional defensive cleanup, or a cosmetic/refinement suggestion.

Consequence decision tree (apply in order): (1) Is a binary gate failed? Keep it as a separate hard blocker (the executable helper represents it as synthetic CRITICAL); do not use the ordinary severity label to hide what failed. Otherwise, would shipping permit immediate material security/safety/authority harm, irreversible destruction, data loss/corruption, or a critical-path silent failure? → CRITICAL. (2) Otherwise, does a supported path, invariant, public contract, privacy/authority boundary, compatibility promise, or behavior-changing proof fail with material user/downstream impact? → HIGH. (3) Otherwise, is there a bounded but consequential edge, resilience, observability, testability, maintainability, or architectural gap with a credible impact? → MEDIUM. (4) Otherwise, is the evidence sufficient to show only non-blocking polish with no credible present material impact? → LOW. (5) If the evidence needed to choose between steps 1–4 is missing, → NOT VERIFIABLE, not LOW. When multiple tiers fit, select the highest credible consequence; effort, implementation cost, reviewer discomfort, frequency alone, and proximity to the round cap never decide the tier.

Boundary examples (normalize before applying the round predicate): an auth bypass, exposed secret/PII, destructive command without an authority gate, or failed required test/generation/parity gate is CRITICAL; a wrong supported response, broken invariant/API/schema, meaningful privacy/authority defect, or unproven behavior-changing fix is HIGH; a bounded retry/timeout/alert/testability gap or credible maintainability drift is MEDIUM; a typo, formatting inconsistency, optional cleanup, or cosmetic suggestion proven not to affect present behavior is LOW. A missing fact about any of those boundaries is NOT VERIFIABLE until evidence or an explicitly documented residual-risk decision exists.

Classification procedure (required for every finding): (1) state the affected user, system, data, contract, or gate; (2) assess consequence if the issue ships; (3) assess exposure/likelihood and reversibility/detectability; (4) select the highest tier justified by those facts; (5) cite file:line or equivalent evidence and a confidence percentage. Effort, implementation cost, reviewer discomfort, and proximity to the round cap are never severity inputs. NOT VERIFIABLE is a pending evidence state, not one of the four tiers and never a LOW escape hatch: if the unresolved claim could affect required behavior, security, privacy, authority, availability, data integrity, or a binary gate, it remains an open evidence blocker until resolved or explicitly owner-accepted with documented residual risk. Classify an item LOW only when evidence supports the absence of credible present material impact.

Hard-gate rule: Binary gates (tests, required artifacts, security must-fix checks, generated parity, policy compliance) are not ordinary severity-rated findings. The executable helper records a failed gate as a synthetic CRITICAL blocker solely so one predicate can carry it; the report must still name the gate and failure evidence. A failed gate blocks at every round, including when all ordinary findings are LOW; never disguise a failed gate as LOW.

Score-based skills map their numeric scale onto these tiers — do not invent a parallel vocabulary:

• 0-2 criterion scoring (e.g. production-readiness-review): 0 = CRITICAL/HIGH (criterion unmet, blocks readiness), 1 = MEDIUM (partial, consequential gap), 2 = pass (no finding). If the criterion is only polish, use LOW rather than forcing a 0.
• Two-axis scoring (e.g. performance-review, impact × likelihood): high impact + high exposure → CRITICAL/HIGH; material impact with bounded exposure → HIGH/MEDIUM; low impact and low exposure → LOW. Record the axes and why the selected tier is the highest credible consequence.
• Scorecards / /20 grades (e.g. architecture-scalability-review): the aggregate score and verdict band are separate from finding severity. A sub-80 area is evidence to investigate, not an automatic CRITICAL/HIGH/MEDIUM/LOW label; classify each underlying gap by the consequence decision tree and keep advisory score deductions separate from blocking findings.

Domain-vocabulary normalization (mandatory): Specialized skills may keep a local reporting vocabulary, but it MUST feed this same four-tier round predicate — never a second severity system:

• BLOCKED, HARD FAIL, or FAIL is a blocking local verdict, not an automatic CRITICAL label. Classify the underlying consequence as CRITICAL when it is an immediate material risk or failed binary gate; otherwise classify it as HIGH or MEDIUM with evidence, while preserving the local block until the owning gate is satisfied.
• WARN is not permission to ignore a finding. Map it to MEDIUM when the gap is consequential, to LOW only when evidence supports no credible present material impact, or upward to HIGH/CRITICAL when the consequence warrants it. PASS/compliant is not a finding.
• UI P0/P1/P2/P3/P4 map to CRITICAL/HIGH/MEDIUM/LOW/LOW respectively as a starting point; override upward only when the evidence shows a higher shipped consequence. A P0/P1 accessibility or task-completion floor remains a blocking gate even when a local UI report calls it a priority rather than a severity.
• Numeric SRE/readiness or impact/likelihood scores are evidence inputs, not replacement tiers. Emit the score, the consequence, and the normalized CRITICAL/HIGH/MEDIUM/LOW tier together. INFO/advisory observations are not findings unless the evidence shows a material consequence.

A finding's tier drives the gate: CRITICAL/HIGH/MEDIUM remain actionable and blocking under the round policy; LOW may be tracked as a follow-up and, from round 2, does not by itself justify another fix/re-review. An owner decision may explain or schedule an open MEDIUM but does not turn it into a clean pass; owner acceptance never makes a failed binary gate pass and must record scope, rationale, and residual risk.

• MANDATORY Resolve the active Goal Contract BEFORE work (active plan goal.md → plans/goals/{YYMMDD-HHmm}-{slug}/goal.md → create from current request) and read saved success criteria before editing.
• MANDATORY Append iteration evidence after execution; emit a Goal Satisfaction matrix (PASS/FAIL/BLOCKED) before reporting PASS; loop on validated FAIL; escalate repeated no-progress or blockers. NEVER store secrets in goal files.

MUST ATTENTION apply critical + sequential thinking — every claim needs appropriate traced evidence (file:line for repo/code claims; source URL or artifact section for research, product, content, and docs claims); confidence >80% to act, <60% DO NOT recommend. Anti-hallucination: never present guess as fact, admit uncertainty freely, cross-reference independently, stay skeptical of own confidence.

MUST ATTENTION apply AI mistake prevention — verify generated content against evidence, trace downstream references before deleting or renaming, verify all affected outputs, re-read files after context loss, and surface ambiguity before acting.

• MANDATORY MUST ATTENTION ALWAYS ASK THE 3 TRADE-OFF QUESTIONS — on the thing under review AND on every recommendation you make: (1) is there any trade-off? name what it SACRIFICES (change cost · complexity · perf · coupling · reversibility · migration · ops load · blast radius · security · testability · delivery time · UX) — "none"/"pure win" is an unfinished analysis, so state the dimensions checked; (2) is it worth it? gain (with a metric) vs cost, WHO pays, WHEN → emit WORTH IT / NOT WORTH IT / UNCLEAR; NOT WORTH IT → withdraw or replace it; (3) is it material enough to confirm with the user? irreversible/one-way door · cost shifted onto another team/ops/maintainer/user · one quality attribute traded for another · a tier/service/event/library boundary crossed · auth/money/data-integrity/breaking-change/High-or-Medium-risk path · verdict UNCLEAR → STOP and confirm via AskUserQuestion BEFORE the verdict.
• MANDATORY A MATERIAL trade-off with no user confirmation can NEVER be PASS; NEVER bury one as a Low-severity note, NEVER decide it silently, and NEVER let delivery or convergence pressure authorize a one-way door. — why: an un-walked-back one-way door is the user's call to make, not the reviewer's.
• **M

…(truncated)