Total Review

A local, bounded pre-PR gauntlet in increasing cost order. Compose installed skills, use explicit manual review where the host lacks a suitable reviewer, and keep one authoritative ledger of scope, fixes, gate evidence, and findings. A command's absence or an old clean result is never clearance.

Use before a non-trivial PR or after a substantial refactor. For a trivial change use clean-code; for unfinished work use verify-task; for someone else's PR use review-pr. This skill does not commit, push, publish PR feedback, or change the installed client configuration.

Usage

/total-review                         # Branch changes plus selected local work
/total-review --uncommitted           # Local changes; initial HEAD is the fixed base
/total-review --pr <N> --repo <owner/repo>
/total-review --skip-external         # Explicit local-only coverage
/total-review --no-iterate            # One pass; no fix/review loop
/total-review --inline-findings       # No Beads writes

Flags compose; reject conflicting scopes, missing values, duplicate or unknown flags. --pr requires an explicit repository or a verified current-checkout repository identity; never resolve a bare PR number across repositories. No resume or halt-override flag exists.

0. Resolve capabilities and scope

Portable composition

Read the installed skill by name and follow its instructions with the ledger's exact scope packet. Use the native skill loader when exposed; otherwise read ~/.agents/skills/<name>/SKILL.md and execute its instructions using the current harness's tools. Do not type a slash command into a shell.

Never infer availability from the client name, a familiar command name, or catalog prose. Check the exposed tools, installed skill files, and documented project commands first. Host-native correctness or security reviewers are usable only if their read-only behavior, scope input, and cost policy are known. Otherwise choose the explicit manual route before launch. No new reviewer skills or client plugins are required. Do not let a composed skill replace the ledger scope with its default branch or remote PR diff. Read and apply its review procedure to the supplied packet; if the route cannot accept that scope, mark it unavailable rather than reviewing a different change. G5a is explicitly a repository-wide publication audit alongside the selected review scope, not a substitute diff review. Reading a skill does not grant its tools. G6/G7 require authorized composition of second-opinion with its tools and consent policy; if the harness cannot provide that, record unavailable. The read-file fallback never authorizes running provider CLIs directly from this skill. Discovered cleanup/test commands require existing per-command permission or a current permission request; this composer does not restore blanket build-tool grants removed by its delegates.

Determine whether G7 is requested at preflight, not after seeing review results; this scope choice never authorizes metered routes. Freeze the expected gate set before starting: G1–G5 plus G5a when applicable, G6 by default (excluded only by --skip-external), and G7 only when explicitly requested/accepted. An unrequested G7 is recorded skipped with reason "not requested", never passed. Cost consent remains separate: a requested review whose route is missing or whose metered consent is declined remains expected but incomplete. Report intentional exclusions even when the requested coverage is clear.

clean-code owns mechanical cleanup; pedantic-review already covers reuse. There is no separate assumed auto-fix command. Manual routes are explicitly labeled self-review, never independent coverage. Read evidence and manual gates before starting; it defines the ledger, snapshot recipe, state meanings, and concrete manual review checks. Git is required; gh is required only for PR scope. Use verify-task's repository-native gate discovery for execution obligations; do not maintain a second runner catalog here or guess npx downloads. Missing dependencies degrade as above. Do not install or reconfigure tools during a run.

Honor the runtime's configured model-tier/effort routing. If reduced capability is known, disclose it and ask to continue or stop, unless the user explicitly selected that model. Never invent a model ID.

Local scope

Record repository root, current branch, initial HEAD, task requirements, and included/excluded local paths. Use a fixed comparison base for the whole run:

• Branch mode: resolve the actual default branch from origin/HEAD or documented repository policy

(local main/master only when verified). Resolve its merge-base with initial HEAD to a full SHA. Missing/ambiguous base or failed Git commands stop scope collection; never fall back to a literal ref.

• Uncommitted mode: fix the base to initial HEAD, not whichever HEAD exists after a later fix commit.
• Unborn HEAD: ask the user to make the initial commit first; this gauntlet needs a comparison commit.

The initial scope includes committed branch work (branch mode), staged, unstaged, and selected untracked files. If unrelated local work exists, ask which paths belong before mutation. clean-code is repository-wide: if its writes cannot be separated safely, stop or choose a nonmutating partial review rather than touching unrelated work. Never discard/stash/reset another change for this skill. A valid empty scope yields NO CHANGES, not a passed gauntlet.

PR scope

Collect qualified metadata and diff using read-only commands:

gh pr view {N} --repo {owner}/{repo} --json url,headRefOid,baseRefOid,headRefName,baseRefName,isCrossRepository
gh pr diff {N} --repo {owner}/{repo}

Bind repository, PR number, head SHA, and base SHA. Re-read identities after diff collection; a moving head/base invalidates the packet. For a local full run, prove that the current repository and checkout match the selected PR, the tree is clean, and local HEAD equals the PR head. Resolve the local merge-base from those exact commits before mutation. If any proof or commit is unavailable, use diff-only mode, not the current checkout's code. Do not auto-checkout, fetch into a worktree, or switch branches. Offer a separate user-controlled checkout and fresh run if full local coverage is wanted.

In diff-only mode G1 and executable tests in G2 are unavailable. Requirements can still be reviewed, but that does not imply test execution. Run the read-only gates against the qualified packet; return PARTIAL even if those gates are clean. Do not apply local fixes in diff-only mode. Diff-only PR: G5a is unavailable; never run the helper against an unrelated local checkout or pass it the downloaded diff. Its repository argument is a worktree path, not a PR identity or base ref.

Once a matching local PR checkout receives accepted fixes, the result is a local candidate derived from that PR, not clearance of the published PR: local fixes are not evidence for the remote PR head. Do not switch back to a remote-only diff for external review. At the final checkpoint, re-read the qualified PR head and base; unexpected remote movement invalidates the run and requires fresh scope.

1. Run the gates in cost order

Initialize pass_count=1. Every gate receives the current scope revision plus requirements and writes its method, result, evidence, and finding IDs back to the ledger. Missing output, nonzero commands, partial reads, or an unavailable route are not a clean result. Recheck the scope after every gate.

G1 — Cleanup and scope refresh

Run clean-code when its project target is available. Otherwise use an existing documented equivalent; never invent commands or silently install dependencies. If no applicable cleanup exists, record na with repository evidence. A missing expected linter is unavailable; a failing one halts.

Only mechanical cleanup is automatic. Behavioral fixes need a concrete user choice first. Record all accepted edits, including files outside the starting diff. Rebuild the scope after every accepted fix, including formatter changes, before verification or review. Repeat the cleanup check if its first invocation changed files; only a successful stable run earns pass for the new revision.

G2 — Requirements and tests

Apply verify-task to the stated requirements, not an inferred nearby task. Check requirements for all changes, including documentation and config. Docs-only may make executable tests na, not the whole gate. Record the reason; file extensions alone do not establish that config or skill behavior needs no tests. Use the project's actual test command. Diff-only analysis cannot borrow local tests.

Failing tests or unmet requirements halt. Partial coverage becomes a finding with its concrete gap; missing required execution evidence is unavailable, so the verdict cannot be CLEAR.

G3 — Craft and reuse

Apply pedantic-review read-only to the scope packet and nearby repository patterns. In diff-only mode use head-pinned neighboring context from the same qualified repository; if unavailable, mark G3 incomplete rather than reading an unrelated checkout. Its Must (or Must Fix) tier halts, Should becomes a P1 candidate, and Consider a P2/P3 candidate. For pure prose with no meaningful craft dimension, record na and why. Do not auto-apply suggestions. A craft Owner handoff naming a new requirements/coverage gap must invalidate G2 even when the revision is unchanged. Record it under its actual owner, not as a second craft score. A material handoff halts this run with G2 incomplete; there is no in-pass jump back to G2. Revalidation belongs to a fresh run from G1 under the sticky-halt rule below, never a hidden extra pass or CLEAR.

G4 — Correctness

Use the verified native route or the manual correctness checklist. Validate findings against actual code and requirements. Blocking correctness issues halt; other proven bugs become P1 candidates. Missing context needed to assess correctness is unavailable, not pass.

G5 — Security

Use the verified native route or the manual security checklist. Any validated security finding halts. A speculative concern remains a finding or evidence gap until validated; do not label it a clean audit. Never send secrets to reviewers or reproduce secret values in evidence, prompts, output, or Beads. Sanitize context before independent review; if sanitizing removes material review context, mark that coverage incomplete rather than pretending to review the whole diff.

G5a — Artifact hygiene

In the proven local repository/worktree for this revision, run the authoritative, local-only, read-only helper (Python 3.10+, Git and Gitleaks required):

~/.agents/skills/artifact-hygiene/scripts/artifact_hygiene.py --pretty

Capture stdout and exit status, including nonzero exits. Render coverage before findings using artifact-hygiene's Report contract: exact status/verdict, each source's status and safe error codes, normalized redacted findings, and suppressed counts with clone-local allowances disclosed. Use only normalized evidence; partial is never clean. Never recover raw evidence from reported files, commits, scanner output, configuration, or the scope packet. Remediation is a separate explicitly approved task. Do not edit artifacts, rewrite history, change allowances, install tooling, or copy detectors during this gate.

Bind the report to the ledger revision, audited worktree, target.head, generatedAt, provenance, target.policy, and coverage. Verify target.head matches the revision's HEAD and recheck the scope after collection. The helper audits the full publishable working tree and its own locally resolved unpublished history, not the selected diff or fixed review base. Record that distinct coverage, including its local history base and gaps; never filter out findings merely because they are outside selected review paths. It does not audit remote PR text. A local PR run covers the proven local candidate only.

Exit 0 alone is not clearance; apply these gate results after validating artifact-hygiene/v2 JSON. Check the verdict against each finding's policy.grade; inconsistencies are unusable reports. Passing coverage requires every source complete with no errors or limits:

Excluded from the fix loop and Beads writes: G5a findings remain a redacted advisory report, never an invitation to inspect matches or mutate tracked work. Do not forward matched values to G6/G7. This gate does not replace G5's broader security review, and --skip-external does not skip it.

G6 — One independent peer

With --skip-external, record skipped and continue to the final checkpoint (or the fix decision if there are findings). Otherwise load second-opinion and use its ask mode with the current sanitized scope packet, never a separately fetched PR diff:

second-opinion ask "Review the attached scope revision and requirements. Focus on correctness, contradictions, unsafe commands, and silent failures; cite evidence." --agent peer

The skill owns provider independence, read-only tools, bounded timeouts, and metered-route consent. Include the actual packet, not just the template above. Report incomplete or failed responses faithfully. Validate material claims before severity assignment; repeated agreement does not prove correctness.

Fix decision — at most two passes

Halt is sticky: any halt stops later gates and goes to the final report. After the user fixes the halt, start a fresh run from G1; no hidden resume bypass exists.

Offer the fix option only when iteration is still available. For nonblocking findings within that budget, offer Apply selected fixes and re-review or Keep findings and finish; otherwise report findings and finish without offering in-run edits. Apply nothing without a concrete selection. If selected and iteration is allowed, record the accepted fix paths, increment pass_count before returning to G1, refresh the scope, and rerun G1–G6 including G5a. The cap is 2 total passes, including the first, and never resets inside this run. --no-iterate caps it at one. No fixes are applied inside this run once the cap is reached; changes made anyway invalidate earlier evidence and produce PARTIAL pending a new run. Never report stale reviews as completed on new code.

G7 — Premium quorum panel

Run once, after the final local pass and peer pass, only with explicit approval (or an explicit current-run request for that panel). --skip-external always skips it. Decline means declined, not a completed panel. This uses the configured premium profile; an absent profile is unavailable, not a decline or a built-in default panel. Never substitute another profile silently. Load second-opinion and let it enforce configured routes, quorum, independence reporting, and separate metered-route consent:

second-opinion ask "Review the attached final scope revision for material issues missed by the prior gates; cite repository evidence." --agent quorum --panel premium

Use the same final packet and record prior findings without instructing reviewers to agree. Incomplete quorum is unavailable coverage even if some routes succeed. Panel findings do not start another fix loop; record them and require a fresh run if fixes change the revision. Never invoke model CLIs directly here.

2. Final checkpoint and verdict

Recapture scope immediately before reporting. Any unexpected branch/HEAD/content change, unknown file provenance, or moved PR identity makes affected evidence stale; stop and ask for a fresh scope rather than silently expanding it. Expected accepted fixes require their own revision and gate reruns. The final ledger—not an early snapshot, phase counter, or saved "looks good"—is authoritative. For a local run with no prior stop, rerun G5a after this recapture and before reporting, then recheck scope stability and bind its result to the final revision. This also refreshes publication coverage outside selected paths and local ref/policy changes. A prior clean audit never clears changed content. Keep the previous report as provenance, not current clearance. G5a's disposition applies here too. Do not rerun after a halt or use this checkpoint to bypass a required fresh run.

Derive the outcome in this order:

1. HALTED — validated halt finding, G5a publication-risk stop, or failed required command;

record later gates as not-run. G5a's partial-coverage exit maps to PARTIAL, not command failure.

1. PARTIAL — missing expected gate evidence (unavailable, stale, pending, not-run, failed,

declined, or skipped), or diff-only execution. If core gates pass but an expected independent route is incomplete, say "core gates clear; external coverage incomplete"—never full-gauntlet clearance.

1. FINDINGS — complete current evidence with nonblocking unresolved findings.
2. CLEAR — every expected applicable gate passed on the final revision, no unresolved findings,

and no expected coverage missing. Evidence-backed na is permitted only for an inapplicable check.

CLEAR is scoped to the expected gate set: label intentional exclusions explicitly, e.g. "CLEAR — local-only; external skipped" or "CLEAR — core + peer; premium panel not requested". Neither claims the omitted reviewer ran; only all eight passed gates permit full-gauntlet clearance.

PARTIAL can include actionable findings; neither hides the other. Preserve successful individual routes without promoting a partial panel to complete. Never claim merge readiness or clearance of unpublished changes on the remote PR.

Render one compact report from the ledger:

## Total Review — {scope / local candidate derived from PR}
**Outcome:** CLEAR | FINDINGS | PARTIAL | HALTED | NO CHANGES
**Scope:** {repository, fixed base, final HEAD, revision, included/excluded paths}
**Passes:** {1 or 2}; **Coverage:** {local/manual/independent/panel limitations}

| Gate | Method | Result | Revision | Evidence / finding IDs |
|---|---|---|---|---|
| ...all eight gates (including G5a), with skipped and unavailable ones... |

**Accepted fixes:** {paths and reasons, or none}
**Findings:** {validated severity, source, file:line, open/fixed, bead or inline ID}
**Next:** {fix blocker, fill evidence gap, address findings, or create PR}

Findings and tracking

G5a is report-only: its audit findings are excluded from the validation-by-inspection and tracking writes below. Preserve only normalized redacted evidence and the separate-remediation handoff.

The ledger owns finding identity and status; Beads are downstream tracking, not a second verdict. Validate and deduplicate findings across gates/passes before writes. Recheck a previously reported finding on the final revision; a stale finding cannot silently become fixed. Prioritize from validated impact, not a blanket P0 for every halt. Do not create duplicate beads on the second pass.

With --inline-findings, absent Beads, or unproven ownership, keep numbered findings inline. Otherwise load the Beads baseline, prove the outcome's owning store with next-select stores, inspect candidate existing items in that store, and use bd -C <directory> for every read/write. Reuse an existing item when it owns the same issue; resolve it with next-select resolve <id> before updating. New items name source gate, final revision, severity rationale, and redaction-safe file references. Do not synchronize Dolt, initialize a store, or close tracking merely because a report says the finding was fixed.

Failures

• Missing capabilities are recorded before execution. An actual route failure is not permission to switch execution modes.

Preserve the exact failure and state, stop, and request a same-route retry or explicit user decision. In particular, never silently replace a failed subagent with a foreground/CLI process.

• Authentication, timeout, malformed output, or denied cost consent never count as completed review.

An unavailable optional route may leave the run PARTIAL; no silent retry, cost expansion, or substitution.

• If lint/tests/review generate new source changes, refresh and revalidate them like any other fix.

Unapproved or unrelated generated changes are a scope blocker; do not hide or revert them.

• No automatic remote mutations, branch switches, commits, broad staging, or history rewrites.