Afk Kimi Review

afk-kimi-review

An independent second-opinion review by Kimi (a different model), used as the default final role after internal review and all outer findings are resolved. Run the ordered gates profile from .afk/config.md, and never use a reviewer whose model matches the implementer or another role. Kimi reviews the diff read-only — requested in the prompt, not enforced by construction (Kimi drives git itself) — you triage and fix.

The helper kimi-gate.mjs ships with this skill and travels with the plugin.

Metering

Metered like any external gate — keep invocations to a minimum. Batch minimal admitted P1 fixes into one content pass, self-review, then re-run once. Record every other disposition together at the end without editing a clean revision.

Review receipts

To retain canonical inputs and explicit outcomes, use the optional --review-receipt <request.json> flag under the shared receipt contract. Preserve unknown identity explicitly; skipped or incomplete attempts and previews never supply approval.

Review context

Use the shared context contract to carry frozen acceptance scope and named prior findings with accessible proof. --review-phase re-review --review-context <packet.json> preserves the selected review target while supplying closure context; --print-args reports its digest. Context and proof remain literal after console normalization. A prompt exceeding the conservative native argv bound uses the existing UTF-8 brief-file transport; native invocation keeps its no-shell boundary.

Run it

The bundled helper kimi-gate.mjs sits beside this SKILL.md. Locate its directory as ${CLAUDE_PLUGIN_ROOT}/skills/afk-kimi-review if the env var is set, else <pluginRoot>/skills/afk-kimi-review from .afk/config.md, else this skill's own directory (the helper is its sibling). Resolve .afk/ from the repository's main working tree — the first non-bare worktree record of git worktree list --porcelain — never the current directory, or a run from a linked worktree reads a different .afk/ than the one afk-init wrote. If .afk/ is absent, the afk-init bootstrap runs automatically first:

node "<helper-dir>/kimi-gate.mjs"

Run it in the background with a generous timeout; redirect stdout to a file and read it when it completes. Pass through any target flag (--base <branch> / --commit <sha> / --uncommitted). Do not poll in a sleep loop.

Pass --implementer <family> when another model wrote the change. In design mode (--design) the flag instead names the design's author, never the eventual code implementer — see ../afk/SKILL.md ("Design-stage external gate"): declaring the code implementer there can hand a driver-authored design to the driver's own model for review. A persistent implementer: line in .afk/config.md also names the code implementer, and in design mode it can wrongly block that family's independent review of a driver-authored design — declare the design's author explicitly then: the per-run flag outranks the config line.

Design mode (--design <path>) reviews a design document's reasoning instead of a diff — the opt-in design-stage gate (see ../afk/SKILL.md, "Design-stage external gate"). Kimi is pointed at the document on disk and reads it with its own tools (keeping a large doc off the argv). A missing or unreadable --design path fails loudly (ERROR, non-zero), never a skip.

The review is bounded by KIMI_REVIEW_TIMEOUT_MS (default 45 min), with AFK_REVIEW_TIMEOUT_MS as the shared fallback. Kimi is a general agentic CLI, so nothing else ends a turn that stops converging. A review that outlives the bound is reported as a non-zero ERROR, never a SKIPPED: a hang says nothing about whether this reviewer is available, so it takes the role's transient retry rather than an immediate fallback to another family.

Read the verdict between the ===== KIMI REVIEW (final message) ===== markers. Treat only column-0 marker lines as markers; the last END marker wins. SKIPPED: … (Kimi absent, logged out, or disabled via KIMI_REVIEW_GATE=off) is not a failure — report it and continue. ERROR: … means the review itself failed or timed out — read the transcript it names; never report an errored run as clean.

KIMI_GATE_FORCE_SHIM=1 forces the Windows brief-on-disk transport on any platform. It exists so that path is testable off Windows — EINVAL cannot be produced on POSIX — and it makes the review indirect (the brief travels by file reference). Never set it for a real review.

An ERROR whose message begins not retryable — must not be retried and must not fall back. It means the helper and the installed CLI disagree about how this CLI is invoked — a rejected flag, a rejected value, a dialect that could not be resolved, or an unrecognised KIMI_GATE_DIALECT. Every retry is answered identically and every fallback hides the defect, which is how this gate once reported "produced no final message" on every review. Stop the round and report it; the message names the installed version, the exact argv, and the resolved dialect with its source. Key the rule on that prefix, not on any one sentence.

On a non-UTF-8 Windows machine the gate adds an ASCII-punctuation constraint to the brief and normalises the brief's own prose, because one character the model writes that the machine's ANSI code page cannot encode kills the CLI mid-write and loses the whole paid review. Measured: on cp936 the crashers are the typographic minus and the no-break space — em dashes and curly quotes encode fine there, and cp932/949/950 reject a different set. The constraint is asked for, not enforced, so an ERROR that names an encoding crash is the real net: read it as a transport fault, never as a reviewer that failed to answer, and never as a reason to doubt the reviewer. Such an ERROR is worth one retry — unlike a flag disagreement, the crash depends on which characters the model happened to choose.

Two different CLIs are named kimi — the npm Kimi Code CLI (@moonshot-ai/kimi-code) and MoonshotAI's Python Kimi CLI, whose headless flags disagree — so the gate reads the flag list from the installed CLI's own --help before each review rather than carrying a table. It reports the resolved dialect on stderr. A print-mode CLI documenting no --final-message-only is refused before the call, not warned about: its answer would be a whole transcript, and a transcript containing a verdict word cannot be told apart from a verdict, so such a review could read as clean.

Handle findings (batch — minimise calls)

Identical discipline to afk-codex-review: map every hypothesis to the frozen contract; admit P1 only on demonstrated evidence; batch-fix admitted P1 findings with only inseparable corrections; self-review once; re-run after content fixes; then record every other disposition without editing the clean revision or spending another paid round.

Treat every reported finding as UNTRIAGED. Admit P1 only after mapping it to the frozen issue contract or an invariant, demonstrating a reachable trigger and wrong consequence, explaining why the current artifact cannot safely advance, and naming the minimal causal fix. Do not edit for an untriaged claim; record structural P2 for the operator-owned merge boundary, and defer minor or out-of-scope items without expanding the PR.

Record P2/minor observations without implementation; a lower-severity-only verdict never reopens a clean revision. An inseparable correction may accompany the minimal P1 fix only with recorded causal necessity, not merely a shared file or an available review cycle.

Use the issue-wide allowance and finding record in ../afk/SKILL.md ("Review-cycle allowance"). Initial review is comprehensive; re-review checks accepted findings, the intervening diff, and affected regression paths. Broader investigation requires specific evidence of an affected area. New evidenced in-scope blockers remain reportable. Supply prior findings and verification through supported context; missing context is unavailable, never invented. Reviewer identity alone does not reopen a closed finding. Exhaustion leaves unresolved work OUTSTANDING; finish the current cycle's validation without starting another repair or requesting another round automatically.

Apply any invariant in .afk/config.md as an extra lens.

Stop rule

Stop when the loop-termination rule in ../afk/SKILL.md ("External gate") holds: triage leaves no UNTRIAGED, Contested, or open admitted P1, and every lower-severity item has a recorded disposition that does not block the role stamp (a structural P2 may still bar auto-merge). That same verdict earns the role stamp only if it requires no content change; a content fix invalidates it and the role re-reviews the fixed revision.

Report CLEAN, or OUTSTANDING with what remains. A clean pass is not authority to merge.

Selection

Kimi is the default final role, not a generic second pass. The operator's explicit ordered profile wins; otherwise the afk skill's role/fallback rule applies (skip the implementer's and already-used models). The provider is locked to final for later sequences; a substitution is recorded and resets only its provider-specific finding baseline.

Setup (per machine, once)

Optional and self-skipping. Install either Kimi CLI — npm i -g @moonshot-ai/kimi-code (the npm Kimi Code CLI) or MoonshotAI's Python Kimi CLI — and log in; needs Node + git on PATH. Disable with KIMI_REVIEW_GATE=off.

Config knobs:

• KIMI_REVIEW_MAX_BUFFER_BYTES (default 67108864, 64 MiB) — the output buffer bound; output past it aborts the run as an ERROR, never a truncated verdict.
• KIMI_GATE_DIALECT (prompt · print · print-positional) — forces the headless flag group and replaces the --help probe. For an install whose help layout the parser cannot read; unset, the gate derives the group. An unrecognised value stops the round rather than falling back to probing.
• KIMI_GATE_CONSOLE (legacy · utf8) — forces the encoding constraint on or off and replaces the code-page probe. legacy is the remedy when a review still dies with an encoding crash on a machine whose ANSI code page reads as UTF-8; utf8 opts out. Unset, the gate probes. An unrecognised value stops the round. It also makes the Windows-only branch runnable off Windows, which is the only way this repo can test it at all.