Cf Review

Detect eligible structural and authoritative-requirement violations exposed by a bounded change set, record them with evidence, and route them to the skill that owns the next decision.

Do not edit repository files other than the batched pass's recap, and do not invoke another skill that edits them.

Scope

Resolve exactly one change set from the current request:

• Pending work, the default: staged, unstaged, and untracked files together, read from repository state.
• A named history range: the last few commits, a commit range, or a branch against its base. Take this path only when the current request names it. A clean working tree is not a reason to infer a range.

Then, for either one:

• Primary files are the selected, existing files. Each is in scope as a whole, not only at changed lines.
• Read primary files from the working tree as it stands, which is the state a maintainer inherits. When a range was named and uncommitted work also exists, say so: a finding may already be addressed on disk.
• Deleted entries are not primary files. Take their old names and paths from the change-set diff or range base, and use them only to find stranded references, incomplete transitions, and documentation drift.
• Audit surfaces are repository-controlled references or documentation outside the primary files that lenses 6 or 10 inspect because a selected change may have made them stale. Do not run structural lenses on an audit surface unless the change set also selected it as a primary file.
• Authoritative intent sources are explicit requirements from the current request or change description, repository-controlled product or domain documentation and acceptance criteria that name the affected behavior, and primary external contracts linked by the request, code, or repository docs. Lens 11 compares against them; implementation code is never its own intent source.
• Test files selected by the change set remain Primary files for structural, convention, dependency, transition, and documentation lenses. Do not classify or count them separately, and do not assess their assertion quality in this pass.
• A violation that predates the change set is still a finding. Clean new code inside a file that breaks a rule does not clear that file; the change set is the moment the violation resurfaces.
• Generated, vendored, and ignored paths stay out of scope.

Hard Gates

Report candidates; never confirm them. Every finding carries evidence and stays a candidate. Do not expand a file-level trigger into an inventory of subordinate hotspots in that file; continue running the other mandatory lenses, but leave deeper diagnosis, causes, remedies, and confirmation to the receiving skill. A file that breaks a file-level trigger is one finding. Express false-positive risk through status: candidate and confidence with its basis.

Report only violations whose remedy is confined to a nameable unit. A smell whose fix propagates to every site sharing a shape — and which therefore no single commit can clear — stays out, however real it is. references/sweep.md names the excluded families and the rationalization to refuse.

Complete the sweep before any routing. Do not name a destination skill, propose a fix, or start a deeper look while findings are still being collected. Routing happens once, after the last lens has run, and only then does references/handoff.md load. Stopping mid-sweep to act on the first hotspot is the failure this gate exists to prevent.

Flow

1. Resolve the change set from the current request. List primary files, deleted entries, and authoritative intent sources separately.
2. Read references/dynamic-agents.md. Run its installed-local context gate against every primary file and already-identified authoritative source before loading them in full. Follow its policy exactly; when it is not local, read references/delegated-execution.md and follow it.
3. Read references/sweep.md and references/navigation-cost.md. When the change is move-shaped, also read references/reference-audit.md; use its read-only audit rule and do not apply its editing rule. Run every lens against its declared surface. All lenses are mandatory; only a lens whose observable condition is absent may be not applicable.
4. For local, run the sweep sequentially. For subagent-1, subagent-2, or batched, read references/review-agent-brief.md and fill every placeholder. Every assignment applies all lenses. The shared reference owns assignment and completion; treat its merged ledger as the completed sweep before step 5 loads references/handoff.md.
5. When the sweep produced at least one finding, read references/handoff.md and build the routing output. With no findings, do not read it: report Findings: none, Handoff: none, and Result: clear for the completed sweep. Either way the output accounts for every lens.

If the current request names no range and nothing is pending, say so and stop; never widen the pass to the repository. If a named range reaches most of the repository, say so and ask for a narrower one: a change set that covers everything has stopped being a reference point, which is the only thing keeping these findings verifiable.

Routing

Each lens carries a fixed destination, listed in references/sweep.md. The destinations are cf-cognitive, cf-split, cf-cohesion, cf-scenario, cf-docs, and cf-start. A finding no destination owns goes to cf-mr-wolf.

Route, do not invoke: name the destination and let the user choose what to open next.

Artifacts

• Owns .cflow/cf-review-recap.md: the batched pass's recap, written under the Recap File contract in references/delegated-execution.md. It is the pass's one repository write.
• When it already exists and the request is about it, work through it under the same contract instead of running the Flow.
• Create or update no other repository file, including todo files and any other .cflow/* file.
• If the current request also asks to persist findings, finish this read-only pass and name cf-todo as the separate next action.

Output Format

Return only:

• Scope: which change set was resolved and how, primary file count and list, deleted entries, audit surfaces inspected, and authoritative intent sources checked.
• Context budget: measured files, LOC, estimated tokens, selected policy, consent source, and actual model or runtime default; use local when no agent ran.
• Lenses: every lens by number, each marked as reporting, silent, or not applicable. A not applicable mark must name the condition that was absent; a mark without one is a lens that was skipped, and skipping is what this slot exists to make visible.
• Findings: every candidate, grouped by file, in the shape references/handoff.md defines. Report all of them; do not cap, rank away, or soften a finding that fires a hard trigger.
• Handoff: with findings, destination skill mapped to the findings it receives and the recommended first one to open; otherwise none.
• Test assertion quality: not assessed in this pass; recommend running cf-test against the same pending work or named history range when that change set adds or changes executable tests. Do not classify tests, report a test count, or run cf-test from this pass.
• Result: with findings, the shipping recommendation from references/handoff.md; otherwise clear for the completed sweep, never unqualified clear or commit-ready. Always append exactly one business-alignment qualifier: business alignment checked against <sources>, business correctness not assessed: no authoritative requirement source identified, or business correctness not assessed for <case>: authoritative sources conflict or are ambiguous. State confirmation still needed, that no repository file was modified apart from the recap when the pass wrote one, and the next action.