Make mach6 review assessments enforce issue and plan scope

[m-aebrer]

Closes #395

Make mach6 review assessment scope-aware, and clarify that the parent model owns implementation reasoning while feature-dev delegation is an optional context-preservation tool requiring detailed task instructions.

Implementation plan posted as a comment below.

[m-aebrer]

Implementation Plan

Problem analysis

mach6 currently treats factual validity as sufficient for a genuine finding. Because every genuine finding enters the fix action plan, technically valid but irrelevant observations can silently become new PR requirements across repeated automated review cycles.

The corrected contract has two independent gates:

1. Factual gate: Does the finding accurately describe a real problem in the code?
2. Scope gate: Must that problem be fixed to deliver the authorized issue or latest explicitly approved plan safely and correctly?

A factually valid finding can pass the first gate and fail the second; it should then normally be deferred, not marked genuine. Authoritative scope comes from the original linked issue and acceptance criteria, the latest explicit mach6-plan, and subsequent human-approved scope updates. Review findings and prior automated assessments are evidence, not scope authority, and cannot expand scope by repetition. Regressions or correctness, security, safety, or integrity failures introduced by the PR remain eligible for genuine classification because safe delivery is part of the authorized work.

The human-approved scope has also been expanded for mach6-implement: the parent model owns design and implementation reasoning. Delegating to feature-dev is optional and is primarily useful for high-volume, repetitive, mechanically scoped grunt work—for example, applying a content-dependent transformation across dozens of files after the parent has defined the rules. The distinction is not that delegated work reads or writes a lot—nearly all implementation does—but that the reasoning is settled and repetitive execution would consume parent context. When delegating, the parent must first turn the approved plan or finding into clear, detailed, specific instructions and decision rules; vague handoffs are not acceptable. Direct implementation by the parent is generally valid in both implement and fix modes.

Formal mach6 review is also an accountability checkpoint, not an agent-owned continuation step. Implementation must be committed, pushed, and recorded before formal review so work cannot be lost or repeatedly rewritten by unsupervised review/fix cycles. The user explicitly triggers each mach6-review; agents stop at the boundary and may use suggest_next to offer it. This does not prohibit a parent from launching a focused one-off reviewer or checker for a narrow correctness question or second opinion during implementation. The prohibited behavior is autonomously starting the full mach6 multi-agent review/assessment cycle, or repeatedly reviewing and rewriting unsaved work without user supervision.

Prior art supports these boundaries:

• Google Engineering Practices recommends self-contained changes with related tests, while its review guidance separates blocking correctness concerns from personal style nits and warns against speculative over-engineering.
• PR scope-creep guidance recommends moving unrelated review requests into separate work rather than diluting the current change.
• Spec-first AI review guidance treats acceptance criteria as the review rubric and rejects unauthorized additions.

Deliverables

1. Make independent assessment explicitly scope-aware

   • Require the assessor to establish authoritative scope before classification.
   • Apply factual and scope gates separately to every finding.
   • Define factually valid but out-of-scope observations as normally deferred, while retaining nitpick and false-positive distinctions.
   • Explicitly exclude optional hardening, speculative edge cases, unrelated pre-existing defects, architecture preferences, and broader cleanup from genuine classification unless authorized.
   • Preserve genuine eligibility for regressions and correctness, security, safety, or integrity problems introduced by the PR.
   • Require every genuine justification to explain both factual evidence and scope relevance.
   • Keep the classification table comprehensive, but include only genuine findings in the action plan.

2. Pass authoritative scope through mach6-review

   • Gather and provide the linked issue, acceptance criteria, latest explicit plan, and subsequent human-approved scope updates to the independent assessor.
   • Repeat the two-gate classification rule in the orchestration instructions.
   • State that review findings and automated assessments cannot become scope updates by novelty or repetition.
   • Require two-part reasoning for genuine classifications and constrain the action plan to findings necessary for the scoped PR to merge.

3. Clarify parent ownership in mach6-implement

   • Apply the same contract in both implement and fix modes.
   • State that the parent model owns implementation reasoning, design choices, decomposition, and detailed execution planning.
   • Present direct parent implementation as generally acceptable, not as an exception limited to trivial changes.
   • Present feature-dev delegation as an optional context-management tool for high-volume, repetitive, mechanically scoped execution after the parent has settled the design and decision rules.
   • Require delegated tasks to include clear, detailed, specific direction: intended changes, relevant files and existing patterns, constraints, tests, and expected verification. Delegation must not substitute for parent reasoning.
   • Preserve existing parallelism, mandatory test coverage, and validation requirements.

4. Enforce a user-controlled formal-review checkpoint

   • Make mach6-review explicitly user-invocable but unavailable for model invocation through the skill tool, so agents offer it with suggest_next rather than starting it themselves.
   • Before launching formal review agents or changing PR review state, verify that the working tree is clean and the reviewed commit is pushed as the PR head. If work is uncommitted or unpushed, stop and suggest mach6-push.
   • Make mach6-implement self-contained about the checkpoint because an implementing agent may never load the push or review skills: explain that committing, pushing, and posting durable progress protects accountability and recovery from lost or repeatedly rewritten work; require it to stop after implementation and direct verification; tell it to suggest mach6-push; state that only the user starts formal mach6-review; and prohibit autonomous review-fix-review cycles. Include the focused one-off-checker exception in this skill itself.
   • Make mach6-push stop after committing, pushing, and posting durable progress, then suggest the explicit mach6-review command; it must not invoke review itself.
   • Explicitly preserve focused, one-off reviewer/checker subagents for narrow correctness checks or second opinions. Distinguish those checks from a formal mach6 review/assessment or an autonomous review-fix-review loop.
   • Treat the pushed commit and GitHub progress comment as the accountability and recovery boundary before review.

5. Update public workflow documentation

   • Document the scope-aware meaning of genuine findings and the authoritative scope sources.
   • Document the parent-versus-feature-dev implementation contract.
   • Leave the root README, package README, generic skills page, feature-dev definition, and cross-skill command references unchanged unless implementation reveals a concrete inaccuracy; their current wording is generic or already consistent.

6. Add prompt-contract regression coverage

   • Validate independent-assessor as a shipped built-in agent and protect its two-gate, scope-authority, introduced-regression, reasoning, and action-plan contracts.
   • Validate the real shipped mach6-review and mach6-implement skill bodies for the corresponding scope handoff and parent-owned delegation contracts.
   • Prefer focused semantic assertions over broad snapshots so harmless prose edits do not create unnecessary churn.

Acceptance criteria

• A finding is never genuine merely because it is technically correct or factually observable.
• The assessor reads the original issue, acceptance criteria, latest explicit plan, and later human-approved scope updates before classification.
• Review findings and automated assessments cannot expand scope by repetition.
• A factually valid out-of-scope finding passes the factual gate but fails the scope gate and is normally deferred.
• Optional improvements, speculative hardening, unrelated pre-existing defects, architecture preferences, and broader cleanup are not genuine unless explicitly authorized.
• Regressions and correctness, security, safety, or integrity failures introduced by the PR can still be genuine.
• Each genuine classification explains both factual validity and scope relevance.
• The classification table covers every finding, while the action plan contains only genuine findings necessary for the scoped PR to merge.
• mach6-review passes the full authoritative scope context and repeats the same rules.
• mach6-implement makes direct parent implementation generally acceptable in implement and fix modes.
• mach6-implement identifies feature-dev delegation primarily as optional context preservation for high-volume, repetitive grunt work with settled decision rules, not as a mandatory workflow or substitute for parent reasoning.
• Every delegated feature-dev task is clear, detailed, and specific enough to execute without inventing design decisions.
• Formal mach6-review can only begin from an explicit user invocation; agents offer the command with suggest_next and stop.
• Formal review refuses to run against uncommitted or unpushed work and redirects to mach6-push before launching review agents or changing PR review state.
• mach6-implement independently explains the accountability and work-preservation reason for committing and pushing before review, even if the agent never loads mach6-push or mach6-review.
• mach6-implement and mach6-push do not autonomously continue into formal review or a review-fix-review loop.
• Focused, one-off reviewer/checker subagents remain allowed for narrow correctness checks and second opinions during implementation.
• Public documentation and regression tests reflect these contracts.
• After the versioned behavior is implemented and validated, the obsolete global mach6_push_before_review.md memory and its index entry are removed.

Files to modify

• packages/coding-agent/agents/independent-assessor.md — scope sources, two-gate classification, reasoning requirements, and action-plan boundary.
• packages/coding-agent/skills/mach6-review/SKILL.md — authoritative scope gathering and assessor task contract.
• packages/coding-agent/skills/mach6-implement/SKILL.md — parent reasoning ownership, optional context-preserving delegation, detailed handoffs in both modes, and a hard stop before formal review.
• packages/coding-agent/skills/mach6-push/SKILL.md — durable save boundary followed by a suggested, never automatic, formal review.
• packages/coding-agent/docs/mach6.md — public review-assessment, implementation-delegation, and user-controlled review-checkpoint behavior.
• packages/coding-agent/docs/skills.md — identify mach6-review as an explicitly user-triggered built-in skill.
• packages/coding-agent/test/builtin-agents.test.ts — structural inclusion and semantic prompt assertions for independent-assessor.
• packages/coding-agent/test/skills.test.ts — semantic contract assertions for the two changed built-in skills.

Testing approach

• Extend packages/coding-agent/test/builtin-agents.test.ts to verify independent-assessor discovery/frontmatter and stable prompt concepts: both gates, authoritative scope, non-authoritative automated findings, introduced-regression handling, two-part genuine reasoning, and action-plan filtering.
• Extend packages/coding-agent/test/skills.test.ts to read the real shipped mach6-review, mach6-implement, and mach6-push bodies. Verify scope handoff/two-gate instructions; parent ownership, direct implementation, mechanically scoped high-volume delegation, and detailed handoffs in both implementation modes; the self-contained mach6-implement explanation of accountability, work preservation, mandatory push-before-review ordering, user-triggered formal review, and the one-off-checker exception; explicit user-only formal review; clean-and-pushed preflight; and stop-and-suggest boundaries.
• Assert that the shipped mach6-review skill has model invocation disabled while remaining user-invocable.
• Run Biome on changed test files.
• Run targeted Vitest for both modified test files.
• Run npm run build, npm run verify-workspace-links, and the full npm test suite; fix every failure.

Risks and open questions

• Prompt-content tests can become wording-fragile; assertions should protect stable semantic contracts and relationships rather than exact full prose.
• Scope wording must permit explicit later human approval without allowing automated comments to self-authorize.
• The scope gate must not become an excuse to ignore regressions, missing tests for changed behavior, or newly introduced security/integrity failures.
• mach6-plan intentionally remains high-level. The parent creates detailed execution instructions during mach6-implement, so no planning-skill change is needed.
• feature-dev.md already requires specific direction, staying in scope, and not guessing. Changing it would not add the missing parent-side guarantee and is therefore outside this plan.
• The review checkpoint must distinguish the formal mach6-review workflow from legitimate narrow checker subagents; overbroad wording that forbids all reviewer use would remove a useful implementation tool.
• The preflight must check both local cleanliness and remote PR-head parity. A clean but unpushed commit is not yet durable shared state.
• The memory deletion is intentionally post-implementation cleanup: remove it only after the equivalent versioned skill rules, docs, and tests pass.

---

Plan created by mach6

[github-actions]

Vitest coverage

Metric	Covered	Total	Coverage
Statements	25007	42119	59.37%
Branches	13380	25733	51.99%
Functions	4727	7643	61.84%
Lines	21440	36437	58.84%

View full coverage run

[m-aebrer]

Progress Update

Implemented the approved mach6 workflow changes:

• Added factual and authorized-scope gates to independent assessment, including explicit handling for deferred observations and PR-introduced regressions.
• Made formal mach6-review explicitly user-triggered and added clean-worktree/pushed-PR-head guards before review begins.
• Reworked mach6-implement so the parent owns design and implementation reasoning; feature-dev is optional support for mechanically scoped grunt work with detailed instructions.
• Added durable accountability checkpoints to implementation and push workflows while preserving focused one-off reviewer/checker use.
• Updated mach6 and skills documentation and added prompt-contract regression coverage.
• Removed the obsolete local push-before-review memory after the equivalent versioned rules were validated.

Validation passed: build, workspace-link verification, Biome, 47 targeted prompt-contract tests, and the full deterministic test suite with live API tests skipped. Commit hooks also passed 4,948 tests.

Commit: 2f9228b

---

Progress tracked by mach6

[m-aebrer]

Code Review

Critical

None.

Important

None.

Suggestions

1. Review-agent handoff drops authoritative plan and scope context — packages/coding-agent/skills/mach6-review/SKILL.md:87,114-117 (medium, confidence 85). The preparation step says every reviewer receives full PR context and authoritative scope, but the concrete launch checklist only names changed files, the PR description, and linked issue context. The latest plan, all PR comments, and later human-approved scope updates can therefore be omitted at the actual handoff. Make the checklist explicitly include the assembled authoritative-scope block and its non-authority rule for automated findings.

2. Assessor regression test does not protect classification relationships — packages/coding-agent/test/builtin-agents.test.ts:79-89 (medium, confidence 97). Isolated phrase assertions do not require a factually valid but out-of-scope finding to be deferred, do not tie PR-introduced correctness/security/safety/integrity failures to genuine eligibility, and do not specifically protect review findings as non-authoritative. Add section-scoped assertions for those relationships and genuine-only action-plan filtering.

3. Review-skill test does not protect the complete preflight or full scope handoff — packages/coding-agent/test/skills.test.ts:477-489 (medium, confidence 98). Whole-body token checks could pass after removing the second clean check, local-versus-PR-head comparison, ordering before review activity, or one of the required scope sources. Scope assertions to the preparation and assessor-handoff sections and verify the complete sequence and inputs.

4. Implement-skill tests do not enforce detailed handoffs and checkpoints in both modes — packages/coding-agent/test/skills.test.ts:491-509 (medium, confidence 98). Most assertions target the shared preamble, so either operational mode could lose its parent-first design step, detailed delegated-task requirements, or stop-after-verification boundary. Add separate implement-mode and fix-mode assertions for mechanically settled delegation, files/patterns/rules/tests/verification/result details, and the durable stop-and-suggest behavior.

5. Push-skill test does not protect the exact stop-and-suggest boundary — packages/coding-agent/test/skills.test.ts:511-518 (low, confidence 94). Generic phrase checks do not require commit → push → progress-comment ordering, exactly one suggestion, or the feature-PR-specific user-run review command. Add assertions scoped to the final post-comment section for that sequence and command selection.

6. Assessor content test performs redundant frontmatter parsing — packages/coding-agent/test/builtin-agents.test.ts:76-91 (low, confidence 85). Frontmatter validity is already tested immediately above, and none of the searched contract phrases can occur in frontmatter. Reading and asserting against the file content directly would remove unnecessary parse/null/body ceremony and match the skill-test helper pattern.

Strengths

• The implementation comprehensively delivers the linked issue and approved plan across agent instructions, workflow skills, public documentation, and regression tests.
• The assessor contract clearly separates factual validity from scope relevance and preserves PR-introduced regression, correctness, security, safety, and integrity concerns.
• The formal-review checkpoint is consistently documented across review, implementation, and push workflows while preserving focused one-off checks.
• Parent-owned implementation reasoning and optional, mechanically scoped delegation are explained consistently in both implement and fix modes.

Agents run: code-reviewer, test-reviewer, completeness-checker, simplifier

---

Reviewed by mach6

[m-aebrer]

Review Assessment

Review comment

Classifications

Finding	Classification	Reasoning
Finding 1: Review-agent handoff drops authoritative plan and scope context	genuine	Factual: Step 3 says to provide full PR context and authoritative scope to every review agent, but Step 4's concrete handoff list includes only changed files, PR description, linked-issue context, and source-reading instructions; it omits all comments, the latest plan, later approved updates, and the automated-findings non-authority rule. Scope: Full authoritative-scope handoff through mach6-review is an explicit plan deliverable.
Finding 2: Assessor regression test does not protect classification relationships	genuine	Factual: Isolated strings such as introduced by the PR and prior automated assessments are evidence only can remain while the deferred relationship, genuine eligibility for introduced failures, or review-findings non-authority is removed. Scope: The approved plan explicitly requires semantic prompt-contract coverage for these relationships and genuine-only action plans.
Finding 3: Review-skill test does not protect the complete preflight or full scope handoff	genuine	Factual: Generic whole-body tokens do not enforce both cleanliness checks, LOCAL_HEAD/PR_HEAD equality, ordering before review activity, every authoritative source, or the assessor handoff. Scope: The plan specifically requires regression coverage for the complete clean-and-pushed preflight and full authoritative-scope handoff.
Finding 4: Implement-skill tests do not enforce detailed handoffs and checkpoints in both modes	genuine	Factual: Most expectations can be satisfied by the shared preamble, allowing either mode to lose its parent-first design step, delegated-task fields, or mode-specific stop-and-suggest ending without failing. Scope: Separate semantic protection for implement and fix modes, including parent ownership, detailed delegation, and durable stopping boundaries, is explicitly required.
Finding 5: Push-skill test does not protect the exact stop-and-suggest boundary	genuine	Factual: Generic phrases also occur in global rules, so the final post-comment boundary could be removed or reordered while tests pass; commit → push → progress comment → one context-appropriate suggestion and the feature-PR review command are not protected. Scope: The plan requires prompt-contract coverage for mach6-push's durable checkpoint, user-triggered review, and stop-and-suggest boundary.
Finding 6: Assessor content test performs redundant frontmatter parsing	nitpick	Factual: Frontmatter validity is already covered by the parameterized test, making the additional parse and null check redundant for these content assertions. Scope: This is harmless cleanup with no effect on correctness or an authorized requirement.

Action Plan

1. Make the review-agent launch checklist explicitly pass all PR comments, the assembled authoritative-scope block, and the automated-findings non-authority rule.
2. Strengthen the assessor test to protect factual-pass/scope-fail deferral, introduced-failure genuine eligibility, review-finding non-authority, two-part reasoning, and genuine-only action plans.
3. Add section-scoped mach6-review assertions covering both cleanliness checks, pushed-head equality, pre-review ordering, every authoritative source, and the complete assessor handoff.
4. Add separate implement-mode and fix-mode assertions for parent-first decisions, complete delegation instructions, verification, and each mode's stop-and-suggest checkpoint.
5. Scope mach6-push assertions to the final workflow and protect commit → push → progress comment → exactly one context-appropriate suggestion, including the feature-PR review command.

---

Assessment by mach6

[m-aebrer]

Progress Update

Addressed genuine review findings 1–5:

• Completed the formal reviewer handoff with all PR comments, authoritative scope sources, human-approved updates, and the automated-findings non-authority rule.
• Strengthened assessor contract tests for deferred scope failures, introduced-failure eligibility, scope authority, two-part reasoning, and genuine-only action plans.
• Added section-scoped review preflight and scope-handoff coverage.
• Added separate implement-mode and fix-mode coverage for parent reasoning, detailed delegation, verification, and durable stop boundaries.
• Protected the push workflow's commit, push, progress-comment, hard-stop, and single contextual suggestion sequence.

Validation passed: Biome, 47 targeted prompt-contract tests, build, workspace-link verification, the full deterministic suite with live API tests skipped, git diff --check, and the commit hook's 4,948 tests.

Commit: 728891d

---

Progress tracked by mach6

[m-aebrer]

Code Review

Critical

None.

Important

1. Formal review is disabled for model invocation despite the latest human-approved scope — packages/coding-agent/skills/mach6-review/SKILL.md:5 (high, confidence 100). During this review, the user explicitly changed the invocation policy: model/agent invocation should remain available because users sometimes want to tell an agent to invoke the skill. Remove disable-model-invocation: true and update the related docs and contract tests while preserving the explicit user-controlled checkpoint: an agent may invoke formal review only in response to a direct user request, never autonomously.

Suggestions

2. Assessor scope-boundary exceptions are not fully regression-tested — packages/coding-agent/test/builtin-agents.test.ts:73-99 (medium, confidence 98). The semantic test does not protect reading acceptance criteria or locating the latest plan marker, the explicit list of normally deferred optional/out-of-scope categories, the rule that tests for PR-added or changed behavior remain in scope, classification of every supplied finding, or the requirement that genuine work be necessary for the authorized PR. Removing those relationships could leave the test green.

3. The review-skill test does not protect the actual two-gate decision rule — packages/coding-agent/test/skills.test.ts:504-512 (medium, confidence 97). The assessor-handoff assertions protect the gate labels but not that technical correctness alone is insufficient, both gates must pass, factual scope failures are deferred, PR-introduced correctness/security/safety/integrity failures remain eligible, and the action plan is limited to work necessary for the scoped PR to merge.

4. The first dirty-worktree stop condition is mentioned but not asserted — packages/coding-agent/test/skills.test.ts:484-495 (medium, confidence 94). The test confirms that the pre-checkout section mentions git status --porcelain, but it does not require dirty work to stop and suggest mach6-push before gh pr checkout. Deleting that protective outcome would still leave the current test green.

5. Push-workflow coverage does not enforce the actual durable sequence or single suggestion — packages/coding-agent/test/skills.test.ts:567-583 (medium, confidence 96). Header ordering and a prose substring do not require the actual git commit, git push, progress-comment operation, hard stop, and exactly one suggest_next directive to remain in sequence. Scope assertions to the commands and final workflow section.

Strengths

• The implementation cleanly establishes separate factual and scope gates, authoritative scope sources, deferred out-of-scope observations, and genuine-only action plans.
• Review, implementation, and push instructions consistently preserve a durable clean-and-pushed checkpoint and focused one-off checker use.
• Parent-owned reasoning and optional mechanically scoped delegation are applied in both implementation modes.
• Public documentation is concise and aligned with the shipped workflow.
• The targeted prompt-contract suite passes all 47 tests, and the completeness review found every original issue and approved-plan deliverable implemented.
• No simplification or idiomatic-code concerns were found.

Agents run: code-reviewer, test-reviewer, completeness-checker, simplifier

---

Reviewed by mach6

[m-aebrer]

Review Assessment

Review comment

Classifications

Finding	Classification	Reasoning
Finding 1: Formal review is disabled for model invocation despite the latest human-approved scope	genuine	Factual: mach6-review/SKILL.md sets disable-model-invocation: true; the loader therefore hides the skill from the model, while documentation and tests enforce the restriction. Scope: The latest direct human update explicitly overrides this part of the earlier plan: an agent must be able to invoke review when directly requested, while autonomous review cycles remain prohibited.
Finding 2: Assessor scope-boundary exceptions are not fully regression-tested	genuine	Factual: The current test does not protect reading acceptance criteria, locating the latest plan marker, the normally deferred category list, changed-behavior test coverage, classification of every finding, or the authorized-PR action-plan boundary. Those instructions could be removed while the test stayed green. Scope: These are explicit issue and approved-plan requirements, including required prompt-contract regression coverage.
Finding 3: The review-skill test does not protect the actual two-gate decision rule	genuine	Factual: The handoff test checks gate labels but would not detect removal of the rule that both gates must pass, factual-pass/scope-fail findings are deferred, or PR-introduced failures remain eligible. Scope: Repeating and regression-testing these relationships in mach6-review is central to issue 395 and the approved plan.
Finding 4: The first dirty-worktree stop condition is mentioned but not asserted	genuine	Factual: The initial status command is asserted, but the stop-and-suggest assertion can be satisfied by the later post-checkout failure path, so removing the initial protective outcome would pass. Scope: The approved preflight protects unsaved work before checkout, and the plan explicitly requires regression coverage for that integrity boundary.
Finding 5: Push-workflow coverage does not enforce the actual durable sequence or single suggestion	deferred	Factual: Stronger command-level mutation coverage and occurrence counting are possible, but the test already protects ordered workflow sections, the actual progress-comment operation, the hard stop, the one-command instruction, and the feature-PR review command. Scope: Additional command-level assertions are optional hardening of already protected behavior, not necessary for this scoped PR.

Action Plan

1. Remove disable-model-invocation: true; update the invocation contract test and affected mach6/skills documentation to permit agent invocation only following a direct user request. Retain the prohibition on autonomous formal review and review-fix-review loops.
2. Strengthen the mach6-review assessor-handoff test around both-gates-required, factual-pass/scope-fail deferral, introduced-failure eligibility, and genuine-only necessary action items.
3. Extend the independent-assessor contract test for explicit authoritative-scope sources, deferred categories, changed-behavior test rules, every-finding classification, and the authorized-PR action-plan boundary.
4. Assert that the initial dirty-worktree check stops before checkout and suggests mach6-push.

---

Assessment by mach6

[m-aebrer]

Progress Update

Addressed genuine review findings 1–4:

• Re-enabled model/agent invocation of mach6-review when directly requested by the user while preserving the prohibition on autonomous review cycles.
• Updated mach6 and skills documentation for slash-command and direct-agent invocation.
• Strengthened independent-assessor contract tests for authoritative scope sources, deferred categories, changed-behavior tests, comprehensive classification, and authorized-PR action plans.
• Strengthened review-skill tests for the complete two-gate relationship and the pre-checkout dirty-worktree stop.

Validation passed: Biome, 47 targeted prompt-contract tests, build, workspace-link verification, the full test suite with 5,026 passing tests, git diff --check, and the commit hook's 4,948 deterministic tests.

Commit: e3dc889

---

Progress tracked by mach6