sliced-review

Version 2.0.1 — see CHANGELOG.md. Your installed version is the version: line in SKILL.md.

A Claude Code skill that reviews a whole codebase with a second model without the budget being spent on the model wandering around the repository.

It splits the codebase into scoped slices with explicit file lists, orders them into tiers by blast radius, and drives them one at a time with a cost ledger, an activity-aware watchdog, and an optional quota guard. Works with any headless backend — a metered subscription, a pay-as-you-go API, a local model, or a Claude subagent.

The one idea: the cost is in the wandering, not the thinking. A prompt that says "review this repository" makes the model glob, grep and read its way across everything before it reasons about anything. A prompt that says "read exactly these twelve files, never run a repo-wide grep" costs a fraction and returns findings.

Read FIELD-NOTES.md first — it is the measured evidence behind every rule in the skill (three parallel whole-repo lanes that burned 21% of a weekly plan for zero findings; a 20k-line codebase later reviewed for 4% of the same plan; a backend that hung silently for 19 hours).

What's in the box

Path	What it is
`SKILL.md`	The skill itself — invocation modes, then 12 steps from measuring the codebase to recording the findings
`CHANGELOG.md`	What changed in each version
`FIELD-NOTES.md`	Measured results from real runs. Why each rule exists
`references/runners.md`	How to wire a review backend: the four questions, plus worked examples
`references/templates/_context.md`	Shared context every slice prompt starts with
`references/templates/run-slice.sh`	Runs ONE slice: quota guard, idle + hard-cap watchdog, process-group kill, ledger line
`references/templates/chain.sh`	Works the queue unattended; retries on 429, never records a quota error as a report
`references/templates/ORDER.md`	Tier-order template: what is still covered if the budget dies after tier 2
`templates/findings.json`	Shape of the `docs/reviews/findings.json` record every review ends by writing (Step 12)

Install

git clone https://github.com/Mariomarquezt/sliced-review ~/.claude/skills/sliced-review

Then in Claude Code: "review the whole codebase with GLM" / "audit this project with a second model" / /sliced-review.

Modes

invocation	reviewers	applies fixes?
`/sliced-review` or `/sliced-review 3`	core three: GPT-5.6 Sol (xhigh), Grok 4.6 (high), Opus 5 (high)	no
`/sliced-review 5`	core three + GLM 5.3 Flash + Composer 2.5	no
`/sliced-review fast`	Composer 2.5 alone	no
`/sliced-review fix` / `fix 3` / `fix 5` / `fix high`	as above	yes — triage, fix, cross-check with a different model, commit
`/sliced-review fix fast` / `fix fast 3` / `fix fast 5`	as above	yes, pipelined per slice

Append no wrap to skip the closing wrap-up. The models are the author's measured stack; swap in whatever your runner.env can reach.

Wiring your own backend

The backend lives in one file, runner.env, generated per project in Step 6. The wrapper scripts named in references/runners.md (~/.hermes/glm/glm-run.sh, ~/.hermes/kimi/kimi-run.sh, ~/.hermes/cmd/cmd-run.sh) are the author's local one-shot launchers and are not included — they are examples of the shape, not dependencies. Any command that takes a prompt and prints a report works:

RUNNER_CMD='codex exec -m gpt-5.6-sol "$(cat {PROMPT_FILE})" < /dev/null'
QUOTA_PCT_CMD=''   # leave empty if the backend cannot report remaining quota
COST_CMD=''

references/runners.md lists the four questions to answer for any backend.

Honesty rules (from the skill)

Never present a cost estimate as measured until slice one has run.
Never report a slice as done when its output contains a quota error.
If the budget runs out partway, state exactly which tiers were covered.
Never report a fix as verified when nobody watched it fail.
Never report multi-model agreement when the models saw each other's answers first.

License

MIT — see LICENSE.