iso-review

name: iso-review description: Review the uncommitted working tree. By default spawns codex /review and claude /code-review in two visible herdr tabs; with --codex-only, spawns only codex. Merges and de-duplicates findings in the main session, applies every fix except the net-negative ones via a fix tab (codex by default, claude via --fix-agent, or an existing tab via --fix-term) that then runs the project's tests and type-check — leaving all changes uncommitted. Use when invoked as /iso-review [--codex-only] [--claude-review-effort high|max] [--fix-agent codex|claude] [--fix-term TERM] [--kill-tabs], or asked to review-and-fix the current uncommitted changes.

iso-review

Review the uncommitted working-tree diff, keep the fixes that help, apply them, verify, and stop — uncommitted — for your final read.

Invocation: /iso-review [--codex-only] [--claude-review-effort high|max] [--fix-agent codex|claude] [--fix-term TERM] [--kill-review-tabs] [--kill-fix-tab] [--kill-tabs].

By default, the two reviewers are codex /review + claude /code-review. Use --codex-only when Claude tokens are unavailable or you want a Codex-only lifecycle test. Only the claude reviewer's effort and the fixer are selectable (codex /review is pinned to the uncommitted preset, no effort knob). Teardown is opt-in: by default every tab is left alive for live inspection. The kill flags make cleanup systematic — each tab is killed only after its output has been persisted, so a kill reclaims the process without losing anything you read. Map the invocation flags onto the orchestrator calls: pass --codex-only / --claude-review-effort <level> / --kill-review-tabs (or --kill-tabs) to reviews, and --fix-agent <agent> or --fix-term <TERM> / --kill-fix-tab (or --kill-tabs) to apply.

Orchestrator: skills/iso-review/scripts/review.sh. Run it with its absolute path. Reviewer behavior lives in one Reviewer module, scripts/lib/reviewer.sh, with Agent kind (codex|claude) as the seam: reviewer_dispatch <kind> <pane> <level> owns per-kind dispatch (Codex /review preset-menu nav, Claude /code-review), and reviewer_normalize <kind> <raw> <out> shares one parser+writer, branching only the output shape. A third kind is one new case arm, not a new file. Use review.sh run ... for the full scripted path, or the lower-level subcommands below when the main session needs to inspect and decide each phase manually.

Flow (the main session drives this)

1. Pre-flight — review.sh preflight. If it exits non-zero, print its message and stop.
2. Reviews — review.sh reviews [--codex-only] [--claude-review-effort high|max] [--kill-review-tabs]. Wipes .iso/logs/review clean first (so no prior run's accepted-fixes.md/transcripts/.spawned-terms can leak in), then spawns the reviewer tabs, drives them, waits for them to truly finish, and writes .iso/logs/review/review-codex.txt and .iso/logs/review/review-claude.txt. With --codex-only, the Claude files are empty/[] placeholders so downstream merge logic stays stable. Read both files. With --kill-review-tabs, reviewer tabs are torn down right after those files are written (their findings are already on disk). (One review at a time per working tree; for parallel reviews use separate git worktrees — each gets its own cwd-local .iso/logs/review.)
3. Extract — pull every finding into { file, line, problem, fix, source }. Both reviewers emit JSON, so prefer parsing it; fall back to reading prose if a file isn't JSON.
   • codex: { "findings": [ { "title", "body", "priority", "code_location": { "absolute_file_path", "line_range" } } ] }
   • claude: [ { "file", "line", "summary", "failure_scenario" } ]
   • If a file is empty, that reviewer produced nothing — continue with the other. If both are empty, stop: "no findings".
4. Merge + dedup — fold findings that hit the same file + overlapping lines + same underlying problem into one (note both reviewers raised it).
5. Filter (keep almost everything) — accept every merged finding except ones that make the code worse or overcomplicated: unwarranted abstraction, over-engineering, speculative "consider…" notes, readability churn that fixes nothing, anything adding coupling/length without real gain. Conflicting fixes for one spot → take the simpler; if ambiguous, skip. Drop nothing else.
6. Ledger — print the accepted list and the dropped list (each drop with a one-line reason) so the decisions are visible.
7. Apply + self-verify — write the accepted fixes as an itemised instruction list to .iso/logs/review/accepted-fixes.md, then review.sh apply .iso/logs/review/accepted-fixes.md [--fix-agent codex|claude | --fix-term TERM] [--kill-fix-tab]. If the accepted list is empty (every finding was dropped), do not call apply — review.sh apply returns 3 and spawns no fix tab; report "no fixes to apply" and skip to close-out. When there are fixes, this either spawns the fix tab (codex by default, or claude via --fix-agent claude) or reuses the caller-provided --fix-term tab. The fix tab implements the fixes and then runs the repo's tests + type-check and reports. Read its report. With --kill-fix-tab, the tab is torn down right after the report is captured.
8. Close-out (no commit, no extra tab) — leave every change in the working tree. By default the review and fix tabs stay alive for inspection (kill them yourself, or pass the kill flags on Steps 2/7 for systematic teardown). Print the final summary (accepted/dropped ledger + the fix tab's test/type report) and stop. The user reviews and commits. Never commit, never open a PR, never spawn a re-review tab.

Scripted helper:

skills/iso-review/scripts/review.sh run --kill-review-tabs --fix-term "$TERM_IMPL"
# Codex-only:
skills/iso-review/scripts/review.sh run --codex-only --kill-review-tabs --fix-term "$TERM_IMPL"

This runs preflight, reviewer dispatch, normalized finding collection, accepted-fix file creation, and apply. The main-session /iso-review flow may still use the lower-level subcommands above when human judgment is needed for the merge/filter ledger.