Comprehensive Code Review

Review the requested scope using concrete scenarios, consequences, precise locations and evidence. Report in the user's language. Discover project conventions at runtime; no other skill is required.

Boundaries

Reviewing permits inspection and necessary temporary artifacts within existing permissions. Preserve the main checkout/index and user resources. Fixing product code, committing, pushing, remote description edits, publishing, approval and merge each need their own authorization. Respect read-only/no-execution requests. Remove only verified run-owned resources; retain the report first.

Static decisive flow can prove a defect; missing tests or unavailable tools cannot. Distinguish regressions, preexisting defects, description discrepancies and unresolved requirements. Passing checks cover their actual scenarios, not the entire change.

Load by role

Assigned discovery/verification workers read worker-packets.md, their stage's section and the pinned brief with binding project instructions. Return the compact packet; the coordinator workflow below handles profiles, presentation and resources. Missing facts trigger a targeted coordinator refresh.

The coordinator loads references by phase and embeds relevant rules in briefs. Use reading-strategy.md to plan bounded access/checkpoints. Context identities bind exact inputs; routing preserves project gates.

Before the first artifact write, load artifacts.md. Use the archive helper when executable within permissions, preserve its returned run_dir and CR-<run_id> review identity, and keep temporary evidence in an owned registered session. Full durable runs require the trace; compact runs preserve the observed execution sources described in artifacts.md. Historical runs remain unchanged. Verify retained evidence before cleanup; a failed archive gate prevents claiming durable completion.

Resolve archive harness identity once under artifacts.md: preserve the original name and use the bundled catalog for harness_id. A generic family does not establish its interface; unknown/conflicting identity stays visible without changing the code verdict. Reuse the prepared identity throughout the run, without detection probes or context rediscovery. For each new review, capture supplied model metadata or explicit unknown for coordination and each actually dispatched discovery/verification/gate participant using setup model-usage under artifacts.md. Invoke at observed dispatch/result and preserve relevant phase/model changes; configured is not observed. Select context/ledger before retention, keep missing capture visible, and never query histories/APIs or create another dispatch to fill metadata.

Request options

Treat --profile focused|standard|deep|extended, --delivery full|brief and --output <delivery-root> as explicit request conventions, not shell commands or npx options. Equivalent clear natural-language requests apply. No profile means automatic selection below; an explicit profile stays fixed. Delivery is separate from review depth and only runs for an explicitly selected full/brief mode. Without a delivery request, respond normally and keep the internal archive. --output only chooses the delivery root: by itself it neither authorizes delivery nor redirects the internal archive. Ask for a missing/contradictory delivery choice without blocking independent review work. Read delivery.md only when delivering.

Profiles

Canonical strategies are focused, standard, deep and extended. Normalize explicit request aliases economy → focused and balanced → standard before selection; aliases remain explicit fixed choices. Use canonical names in new schema 13 results; preserve historical record names/versions without migration.

Keep an explicit profile; never escalate it silently. Without one, select extended only for justified independent discovery beyond deep's capacity, deep for observable material mechanisms, focused for demonstrated bounded eligibility, otherwise standard under profiles.md. Reuse discovered context and its ABCDE/flow assignment map; no extra selection agent/audit. Automatic selection can escalate with new facts. Project gates and execution isolation always apply.

Every discovery brief states assigned ABCDE areas, flows/interfaces, questions/invariants and expected evidence/coverage under reviewers.md. Group shared flows; split only for a distinct need. Invoking CCR requests the delegations required by the explicit or automatically selected profile; no separate request for subagents is necessary. Use native tools within actual permissions; workers do not delegate. Fresh contexts exclude other findings and the full coordinator transcript. Shared input context is not reviewer identity. Few slots require batches. Real restrictions, unavailable tools, failed dispatch or exhausted quota permit documented degraded execution, never claimed independence. Project-required independent gates still apply.

Workflow and conditional references

1. Scope/context. Read scopes.md, capabilities.md, review-areas.md and artifacts.md. Pin inputs, discover common facts once, account for all 27 dimension IDs (A01–D05 and E01–E07), map applicable dimensions to assigned flows/check owners, and load only the supporting dimension references needed for those subaspects. Separate MR/PR responsibility from version-bound change authors. Choose review kind and functional subject once for the canonical presentation. Select the profile before work. Before preparing a run, perform bounded lookup for the same logical scope. For ordinary MR/PR review requests, first apply effective-change admission: capture the current full diff and materials, then run artifacts assess-reuse. new_review starts a fresh review automatically using current inputs, without asking; offer_choice alone offers a new review or the latest verified result; blocked_current_scope requires acquiring the missing current scope before work. Prior existence or HEAD equality alone never grants reuse. Other scope modes keep exact-input applicability: changed/unknown inputs start a fresh bounded review, while a verified applicable result permits the same choice. An explicit rerun proceeds as a linked new run without asking again. “Show latest/last” uses lookup --select latest and presents only the returned verified candidate after checking its applicability; it never starts a review or silently chooses among timestamp ties, incomplete scans or changed inputs. Never select by folder order, mtime, run ID or report date. Recover only the exact durable operation and reserved bytes; changed or unknown inputs require a fresh run under the review request rather than claiming the old snapshot is current; they never permit resuming a different snapshot. Otherwise, automatically select compact or full persistence from observed facts under artifacts.md, after profile selection and before preparation or discovery; never ask the user to choose. Prepare the selected persistent run and preserve its initial profile. If new facts require full, promote before dependent work using the reserved operation and unchanged inputs. Execution worktrees remain separate. For full runs, register the active required assignments and run preflight --phase discovery before the first dispatch. For compact runs, retain the actual observed selection, assignments, starts and results in selected sources; do not synthesize milestones. In full runs record discovery/started, or record agent/started after observing dispatch, reusing an equivalent existing milestone. An independent agent/started records the observed executor.name; a static coordinator discovery needs no external executor. If an older open run already has an executor-less start, do not dispatch again just to fill identity. A single dispatch-identity/unchanged annotation may bind that original event to a selected local dispatch-metadata object; the same payload replays without another event, while missing evidence or conflicts preserve the run for correction. Observed review work moves new runs from prepared to processing; selection/planning and identity annotations alone do not. If final-candidate preflight reports only a candidate construction/projection issue and pinned inputs still apply, rebuild that unretained candidate in the same run and rerun retain preflight without dispatching discovery or verification again. Re-review/complement uses re-review.md.
2. Discovery/verification. Assign packets under worker-packets.md; read reviewers.md when delegating/grouping. After independent discovery, contrast pertinent external notes before grouped verification. Preserve results, map provisional IDs under identifiers.md, group evidenced shared causes and verify candidates/material questions in the final batch or the justified extended partition. Reopen only for new evidence or an uncompleted gate. Mark uncertainty explicitly. For candidates, load priority-calibration.md for the existing verification pass, including focused and fallback; supply its relevant rules to the verifier without adding a new round.
3. Report/close. Recheck code/description/attribution freshness. For new confirmed findings, complete the consequence-based priority justification in existing impact under priority-calibration.md, even when verifier deltas omit it. Complete schema 13 with a purpose for every check, all 27 dimension decisions grouped by shared status/evidence, classified limitations, profile history and observed obligations under result-contract.md. Use report-format.md for canonical projections; the private report contains one derived ABCDE summary and material gaps, without a duplicate dimension table. Every check declares what behavior/condition it verifies before execution. Show purpose and result separately in Validación; findings explain scenario/impact and link evidence by real check ID. Evidence and coverage use Markdown sublists in reports/comments. Public check purposes and every flow require source-bound editorial summaries under publication.md; new schema 13 drafts use schema 3. Public limitations communicate relevant impact, never review mechanics. For full runs, run preflight --phase retain against the candidate and selected evidence before retaining with --expected-schema 13; verify retained report, record, trace and evidence, observe cleanup, close, and pass validate --require-complete. For compact runs, retain selected observed execution sources and the schema 13 result with retain-compact, then verify with validate --require-complete. Report any persistence failure; never claim a partial artifact is complete. Distinguish review judgment, retained result and pending/complete cleanup. Preserve the exact run/resources when a gate fails; never bypass it with a legacy record or native write. Explicit full/brief delivery derives from the retained result under delivery.md; keep the archive immutable and omit raw evidence from delivery.

Read workspaces.md before builds/tests/installs/reproductions or other writes. Those executors need owned Git worktrees, preferably the project's established location (normally .worktrees); read-only workers may share pinned inputs or immutable Git objects. Blocked creation is an explicit validation limitation, never an automatic switch to scratch/copies. Load external-cli.md only for configured external execution and publication.md only for remote publication.

Helpers

Use actual source timestamps for occurred_at; unknown stays null, while recorded_at remains recording time. Preserve observed actor/executor IDs and optional explicit links from existing results; infer neither identities nor parallel execution. Processing means work began, not that a process is currently alive. The trace contract supplies input examples and recovery rules; no heartbeat, extra agents or timing/counter probes are required.

Invoke a helper, when needed, through python <absolute-skill-path>/scripts <area> ...; consult its --help first. The artifacts area prepares/registers/validates/retains/closes durable runs without deleting resources; its native fallback and mandatory gates are in artifacts.md. The history area is limited to verifying/showing a historical consultation package or porting a hash-verified old report; it cannot resume that source. The packets area validates/merges internal packets; complete the final record under result-contract.md. workspace, runner, setup, checks and reporting areas retain their bounded roles. See script migration for all areas and for the intentional removal of old CLI paths/imports. Measurement collection is opt-in for an explicitly requested cost evaluation: only then load measurements.md and evaluation.md. Normal reviews do not initialize measurement files, probe counters or estimate consumption. Helpers do not prove truth, select models or grant permissions.