Fresh-eyes review

A context that produced a change reads its intent, not its text, so same-context review misses what a fresh reader would catch. The fix is procedural: a reviewer whose context holds only the artifacts.

Workflow

1. Resolve the inputs. The changeset: whatever the invocation names — a branch, a commit, a diff range, a draft vs its original. Given none, infer it from the session — usually the work just finished, committed or not; no VCS required. With no session context to draw on, fall back to git: the working tree against HEAD (git diff HEAD plus untracked files) and the branch against its merge base with the fetched default branch of its tracking remote (git ls-remote --symref <remote> HEAD). One non-empty → take it; both → offer both in step 2's proposal; neither, or no target found → ask what to review. Pin the changeset as concretely as the environment allows — a diff or commit range where one exists, otherwise the touched files, with their prior state when reconstructable. Alongside it, a short statement of what the change is supposed to achieve, when one exists (the task as stated, a PR or ticket description); when this session authored the change, never include the session's own reasoning, plan, or messages — leaked rationale recreates the blindness the fresh context exists to remove. Strip what the change deliberately leaves for later, whatever its source: naming it walls off the omissions lens below. Done when changeset and intent are pinned down and free of authoring context.
2. Confirm the prompt. Assemble the reviewer prompt — changeset, intent, the mandate and exclusions below, and any further reviewer instructions the invocation supplies (e.g. what to report back). When the invocation supplied changeset, intent, and mandate explicitly (e.g. a driving skill), nothing was inferred, so skip the confirmation and proceed. Otherwise show the prompt to the user verbatim and wait for approval; fold any doubt about an inferred changeset into the proposal rather than asking separately. Text emitted before a tool call may not be displayed, so never show the prompt and then ask via a question tool in the same turn — end the turn with the prompt and a plain-text ask, or embed the prompt in the question tool. Done when the user has approved the prompt, as shown or amended — or the explicit-inputs skip applied.
3. Spawn one fresh-context reviewer (a subagent or equivalent isolated session, e.g. a non-interactive run of the agent's own CLI with the prompt) with the prompt, free to read any surrounding project material — except the paths the prompt lists as excluded: any exclusions the invocation supplies, plus, when this session authored the change, session-authored files that are not part of it (plans, notes, scratch), since a fresh context cannot tell them apart. It writes nothing into the project: every run it makes (tests, probes) disables or redirects to scratch space its side outputs (coverage, reports, snapshots), which tooling writes in-tree by default. Before spawning, create a marker file in scratch space and save there the file list (find <dir> -type f -not -path '*/.git/*') of each repo or directory the reviewer may run in; afterwards, a listed file now gone was deleted, an unlisted one created, and a listed one newer than the marker (-newer <marker>) modified. Its mandate, unless the invocation redirects it (e.g. security only): regressions and correctness, including contradictions with surrounding code, rules, or docs — though matching surrounding code is not correctness: verify any pattern the change extends or mirrors is itself sound, since completing a broken rollout inherits its breakage; ambiguities a reader without context would trip on; when an intent statement was given, whether the change does what it says; and omissions — what the change should have touched and didn't: an altered contract (file format, payload, schema, config or CLI surface) has counterparts that must move with it — templates, samples, fixtures, seed data, docs. Find those by content, not location: they often sit outside the changed tree and its stack, and a literal a changed file copies from a shipped artifact (a header row, a sample payload) names one. A deferral the submission itself states — a PR description or commit subject carried verbatim — is a claim to test, not scope conceded. Tough but grounded, aimed at mistakes that matter: every finding names its location and a concrete failure scenario; style nits, speculation, and padding are out of scope, and zero findings is a valid outcome. Besides what the invocation asks, it returns its findings numbered and what in scope it did not read, by name (or that it read everything). Out of scope bounds what is reported, never what is investigated: a pre-existing anomaly in the mechanism the change touches is a reason to audit it. If the harness cannot isolate a context, fall back to an adversarial pass over the same inputs in the main session. Done when an isolated reviewer has returned its findings, or the fallback pass ran and its result is flagged as same-context (weaker) — and each repo or directory the reviewer may run in, compared against that baseline, holds nothing the review wrote: created files removed, or reported when something else may have written them (another session, a watcher); modified and deleted ones reported.
4. Report back. Relay every finding intact — location and failure scenario included — plus whatever else the reviewer was instructed to return; add the session's own assessment when useful, but never silently drop or soften a finding. What to do with the findings is the caller's decision, not this skill's. Done when every reviewer finding appears in the report.