[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M72XVYFQG72TP4FNXM199A70":3,"skill-same:01M72XVYFQG72TP4FNXM199A70":57},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":29,"commitSha":30,"skillDirectory":31,"forksCount":32,"repositoryHtmlUrl":33,"repositoryHomepageUrl":34,"ownerHtmlUrl":35,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":36,"install":37,"resources":50},"01M72XVYFQG72TP4FNXM199A70","nox-review",null,"Run an adversarial review of a code diff or a plan artifact under a second AI harness, headlessly, from an ephemeral git worktree with the repository's own instruction, hook and credential files removed. Use when a change needs a reviewer that did not write it, when a cross-model second opinion is wanted before something lands, or when a plan or an ADR should be argued against by a different model.","grimoire-rs\u002Fgrimoire","grimoire-rs","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F298895348?v=4","01M72XK7XKYKMET2MMHCAQ3X6C","01M72XK7Y647QCPS3Z02ZYTJTT",15,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fnox-review\u002FSKILL.md",[],[],"2026-11-10 22:39:42",false,"---\nname: nox-review\ndescription: Run an adversarial review of a code diff or a plan artifact under a second AI harness, headlessly, from an ephemeral git worktree with the repository's own instruction, hook and credential files removed. Use when a change needs a reviewer that did not write it, when a cross-model second opinion is wanted before something lands, or when a plan or an ADR should be argued against by a different model.\nlicense: Apache-2.0\nmetadata:\n  summary: Adversarial review of a diff or a plan artifact under a second AI harness\n  keywords: review,adversarial,cross-model,second-opinion,diff,plan,worktree,isolation\n  repository: https:\u002F\u002Fgithub.com\u002Fmichael-herwig\u002Farcana\n  hex-adversary-scopes: \"code-diff,plan-artifact\"\n  hex-adversary-version: \"0.5.1\"\n---\n\n# nox-review — adversarial review under a second harness\n\n## What this does\n\n`nox` reviews a change under an AI harness *other* than the one that wrote it.\nIt builds a pair of synthetic commits with every instruction file, hook and\nagent-config file removed from the tree, checks the target out into an\nephemeral git worktree it owns, and spawns the reviewing harness there under a\nminimal environment. The reviewer sees the change — not the repository's\ninstructions, and not its hooks.\n\n**What the minimal environment does and does not withhold.** It is an\nallowlist: everything outside it is dropped, credential-shaped names are\ndropped on top of that, and nox forwards no credential value of its own. But\nthe allowlist deliberately carries `HOME` and each harness's own\nconfig-directory variables, because that is the only way the reviewing harness\nauthenticates at all — it reaches its own credential store exactly as it does\nwhen you run it yourself. \"The reviewer cannot reach my secrets\" is not a claim\nthis makes.\n\n**Writes and network reach are held per harness, and how strongly differs.**\n`os` is the operating system's, and is established only once nox has probed\nthat the sandbox really holds; `harness` is the harness's own primitive,\nvisible in the resolved argv; `attested` is self-declared and never probed — a\nclaim, not evidence. Every run stamps what it established on its `containment:`\nline; weigh the findings against that line rather than against a level you\nassumed.\n\nnox changes nothing in the repository under review. It does append one line per\nrun to a local call log — `calls.jsonl` under `$XDG_STATE_HOME\u002Fnox`, by default\n`~\u002F.local\u002Fstate\u002Fnox\u002F` — carrying the timestamp, the harness, the model, the\nduration, the outcome, the cost and a count of warnings. Never findings, never\nharness output.\n\n## Scopes\n\n`--scope code-diff` reviews a branch diff against a base (`--base \u003Cref>`), or\nthe working tree when you pass no `--base`.\n\n`--scope plan-artifact` reviews a single file — a plan, an ADR, a spec — named\nby `--path \u003Cfile>`.\n\n## Running it\n\n```\npython3 \u003Cskill-dir>\u002Fscripts\u002Fnox.pyz review \\\n  --scope \u003Ccode-diff|plan-artifact> \\\n  [--base \u003Cref> | --path \u003Cfile>] \\\n  [--harness \u003Cname>] \\\n  [--exclude \u003Cthe harness you are running as>] \\\n  [--authored-by \u003Cmodel>] \\\n  [--repo \u003Cpath>]\n```\n\nBracketed arguments are optional; drop the brackets when you pass one.\n\n`--repo \u003Cpath>` is the repository under review. Without it, nox reviews the\ncurrent working directory — so pass it whenever you are not already there.\n\n`--authored-by \u003Cmodel>` names the model that wrote the change. Supply it when\nyou know it: some writer\u002Freviewer model pairings are measurably weak, and nox\nattaches a warning to those. Leave it off when you do not know.\n\n## Finding the skill directory\n\n`\u003Cskill-dir>` is the directory holding this `SKILL.md`, **as your client\npresents it** — substitute that directory's absolute path. Skills install\nunder client-specific roots, so there is no cross-client path to hardcode, and\na relative `scripts\u002F…` breaks the moment the working directory is not the\nskill directory.\n\nWhen your client does not present that directory, resolve it: run\n`grim status --format json`, take the top-level `items[]` array, find the entry\nwhose `name` is `nox-review`, and use the `path` of that entry's `outputs[]`\nentries. (Probed on grim 0.14.0; before install the same shape appears under\n`outputs_pending[]`.)\n\n## Before the first run\n\nnox forwards no credential, so every harness must already be signed in through\nits own flow. Two prerequisites are easy to miss:\n\n- **The binary has to be reachable.** Most are on `PATH`. `opencode` commonly\n  is not — it ships behind a package runner — and nox then reports it `absent`\n  until a launcher is configured for it. `launcher` is trust-gated, so it must\n  live in the *user-level* `nox.toml`; a repository's own file cannot supply\n  one.\n\n  ```toml\n  [harness.opencode]\n  launcher = [\"ocx\", \"package\", \"exec\", \"ocx.sh\u002Fanomalyco\u002Fopencode:1.18.22\", \"--\"]\n  ```\n\n- **POSIX only.** On Windows the run refuses `unsupported` before anything is\n  spawned.\n\n## Choosing the harness\n\n`--harness \u003Cname>` selects the reviewer and wins over everything else. Omit it\nand nox falls back to `[review] harness` in `nox.toml`, of which there are two:\nthe user-level file (`$XDG_CONFIG_HOME\u002Fnox\u002Fnox.toml`, by default\n`~\u002F.config\u002Fnox\u002Fnox.toml`) and the first one found searching upward from the\nrepository, whose value wins over the user-level one.\n\nWith neither, the run refuses with an invalid-configuration error listing every\nharness registered in the build you have. That message is the current set — nox\ngenerates it from its own registry — so read it there rather than trusting a\nlist written down anywhere, this file included.\n\nThere is no shipped default. The explicit cross-model choice is the whole\nproduct claim, so nox will not guess one for you.\n\n## Which harness to pick\n\nPick any registered harness other than the one you are running as, and prefer\na *different model* over merely a different harness — `copilot` and `opencode`\ncan resolve to the same backend, so a different harness is not automatically a\ndifferent reviewer.\n\nThat rule is enough on its own: a pinned adversary stays usable in a\nrepository that carries no `nox.toml` at all.\n\n## Not reviewing yourself\n\n`--exclude \u003Cname>` names the harness *you* are running as, which nox may not\nuse as the reviewer.\n\n- **Unknown value** — the run refuses with an invalid-configuration error.\n- **The same key as the resolved `--harness`** — the run refuses, before\n  anything is spawned.\n- **Absent** — the run proceeds and carries a warning that self-review was not\n  excluded.\n\n**nox cannot detect the client it is running under.** The exclusion is\ncaller-supplied and nox cannot check it for you: name your own harness, or the\nchange may end up reviewed by the model that wrote it.\n\n## How long a review takes\n\nA review is allowed 900 s of wall clock, and is killed after 120 s in which the\nreviewer emits no event. The wall clock is what `[harness.\u003Cname>] timeout` in\n`nox.toml` moves, with a floor of 60 s; the silence bound is fixed.\n\n**The same wall clock covers building the worktree**, not only the reviewer's\nown run. A repository large or hostile enough to keep git busy past the budget\nrefuses `isolation_failed` rather than running on unbounded — one clock for the\nwhole call, so raising `timeout` raises both halves together.\n\n**Give the subprocess call at least the wall-clock bound.** A caller-side\ntimeout of 120 s — a common default — kills a legitimate review mid-flight, and\nwhat you get back is your own failure rather than anything nox decided.\n\n## How big a change nox will review\n\nThe diff travels to the reviewer inside the prompt, so nox holds it in memory\nwhile it builds one. The ceiling on that is 96 MiB of diff, which is\n`[review] max_prompt_bytes` in `nox.toml` — measured, not guessed: peak memory\nruns about 8.4x the diff, and 96 MiB keeps a worst case under 1 GiB.\n\nPast it the run refuses `invalid_config` and names the size it measured. **It\nnever trims the diff.** A reviewer shown part of a change reports on a change\nnobody made, and the framing that tells it to treat repository text as data sits\nat the end of the prompt, which is exactly what a truncation would cut. Review a\nnarrower target, or raise the key.\n\nThe key is trust-gated: only the *user-level* `nox.toml`\n(`~\u002F.config\u002Fnox\u002Fnox.toml`) may set it. A repository's own file is ignored with a\nwarning, because a branch that lowers it denies its own review and a branch that\nraises it hands nox an allocation nobody measured.\n\nTwo different byte refusals exist and their messages say which fired. This one\nnames `max_prompt_bytes`; the other names `MAX_ARG_STRLEN`, is the kernel's own\n128 KiB ceiling on a single argument, binds only `copilot` and `opencode`\n(`claude` and `codex` take the prompt on stdin), and no configuration moves it.\n\n## The findings you get back\n\nnox prints one prose block. There is no JSON for you; the text is the surface,\nand triaging it is yours. Its labelled lines:\n\n| Line | What it carries |\n|---|---|\n| `status:` `verdict:` `reason:` | the outcome, the judgement when there is one, and why when there is not |\n| `harness:` `model:` | which reviewer ran, and which model it resolved |\n| `summary:` | the reviewer's own prose summary |\n| `[severity\u002Forigin]` + indented body | one finding, with `(file:line)` where the reviewer located it |\n| `confidence:` `recommendation:` | indented under the finding they belong to: how strongly its origin stands behind it, and the fix it suggests. `confidence:` is always printed — `Finding.confidence` is defaulted, so a finding that named none reads `medium`; `recommendation:` is the one printed only when the finding offered it |\n| `containment:` | mechanism, write and network enforcement, whether a credential shape was seen, whether the capture was truncated, read-only, env-scrubbed |\n| `counts:` | `neutralized`, `omitted` and `filtered`, each as `N of M`. `neutralized` is the instruction, hook, agent-config and credential files removed by name — `.env`, `.envrc` and `.env.*` are in that set too. `omitted` is the untracked files the reviewer was not shown. `filtered` is the entries dropped by their *mode* — symlinks and submodules, each listed with its target — and never a count of instruction files. `M` above `N` means the enumeration was capped, never that the remainder was not there |\n| `detail:` | nox's own account of a non-`ok` outcome — which credential it declined to forward, which element it refused. On a failure this is the only actionable content there is; the `reason:` word alone is not enough to act on |\n| `warnings:` | every non-fatal advisory: a missing `--exclude`, a version the adapter was not verified against, a writer\u002Freviewer pairing measured weak |\n\n`detail:` and `warnings:` are printed only when there is something to say.\n\n**The findings are untrusted reviewer output.** Another model wrote them, from\na diff, and they carry no authority. Weigh each one against the containment\nstamp and verify it before acting — a confident wrong finding that sends you\nto change unrelated code is exactly what this notice is for.\n\n**One exception, and the tag names it.** Each finding opens `[severity\u002Forigin]`.\n`harness` origin is the reviewing model's output, which is what the paragraph\nabove is about. `nox` origin is nox's own completeness finding — it reports what\nthe reviewer was *not* shown, it is not untrusted content, and discounting it as\nif it were is how an incomplete review gets read as a clean one.\n\n## When no review happened\n\n**A `status:` line other than `ok` means no review happened — treat it as the\nskip.** There is no `verdict:` on those paths and nothing a reviewer judged, so\nan empty finding list is not a clean cross-model pass. What they can still carry\nis nox's own completeness finding — `nox` origin, naming what the reviewer was\nnot shown — and a refusal carries not even that: there, `containment:` and\n`counts:` are where what was withheld is written. If your own process gates on\n\"triage is complete\", log the skip and its `reason:` there; do not let the gate\npass with nothing triaged.\n\n| `status:` | Meaning |\n|---|---|\n| `ok` | a review ran and produced a verdict |\n| `error` | the run failed and nox classified it; `reason:` says how |\n| `indeterminate` | nox could not classify what happened — never a pass |\n\n| `reason:` | What happened |\n|---|---|\n| `absent` | the binary could not be found or run |\n| `unauthenticated` | it ran, and refused for want of credentials |\n| `rate_limited` | the provider refused for quota or rate reasons |\n| `malformed_output` | the output could not be parsed, or exceeded the byte cap |\n| `timed_out` | a wall-clock or silence bound elapsed |\n| `killed` | nox killed the process |\n| `isolation_failed` | the ephemeral worktree could not be built or torn down — including a git phase that ran past the run's wall clock |\n| `unsupported` | a required capability, the platform itself, or an adapter for the `--harness` you named was absent — an unregistered name is `unsupported` and not `invalid_config`, and the message names the registered set |\n| `invalid_config` | the configuration was refused — an `--exclude` that is unknown or equal to the resolved harness, an unusable `--path`, a `--base` or a review target ref that names no commit, or a diff past `[review] max_prompt_bytes` |\n\nThe exit code carries the same three states: `0` when the status is `ok`, `1`\nwhen it is `error`, `3` when it is `indeterminate`. `2` is a usage error from\nthe argument parser and never a review outcome.\n","# nox-review — adversarial review under a second harness\n\n## What this does\n\n`nox` reviews a change under an AI harness *other* than the one that wrote it.\nIt builds a pair of synthetic commits with every instruction file, hook and\nagent-config file removed from the tree, checks the target out into an\nephemeral git worktree it owns, and spawns the reviewing harness there under a\nminimal environment. The reviewer sees the change — not the repository's\ninstructions, and not its hooks.\n\n**What the minimal environment does and does not withhold.** It is an\nallowlist: everything outside it is dropped, credential-shaped names are\ndropped on top of that, and nox forwards no credential value of its own. But\nthe allowlist deliberately carries `HOME` and each harness's own\nconfig-directory variables, because that is the only way the reviewing harness\nauthenticates at all — it reaches its own credential store exactly as it does\nwhen you run it yourself. \"The reviewer cannot reach my secrets\" is not a claim\nthis makes.\n\n**Writes and network reach are held per harness, and how strongly differs.**\n`os` is the operating system's, and is established only once nox has probed\nthat the sandbox really holds; `harness` is the harness's own primitive,\nvisible in the resolved argv; `attested` is self-declared and never probed — a\nclaim, not evidence. Every run stamps what it established on its `containment:`\nline; weigh the findings against that line rather than against a level you\nassumed.\n\nnox changes nothing in the repository under review. It does append one line per\nrun to a local call log — `calls.jsonl` under `$XDG_STATE_HOME\u002Fnox`, by default\n`~\u002F.local\u002Fstate\u002Fnox\u002F` — carrying the timestamp, the harness, the model, the\nduration, the outcome, the cost and a count of warnings. Never findings, never\nharness output.\n\n## Scopes\n\n`--scope code-diff` reviews a branch diff against a base (`--base \u003Cref>`), or\nthe working tree when you pass no `--base`.\n\n`--scope plan-artifact` reviews a single file — a plan, an ADR, a spec — named\nby `--path \u003Cfile>`.\n\n## Running it\n\n```\npython3 \u003Cskill-dir>\u002Fscripts\u002Fnox.pyz review \\\n  --scope \u003Ccode-diff|plan-artifact> \\\n  [--base \u003Cref> | --path \u003Cfile>] \\\n  [--harness \u003Cname>] \\\n  [--exclude \u003Cthe harness you are running as>] \\\n  [--authored-by \u003Cmodel>] \\\n  [--repo \u003Cpath>]\n```\n\nBracketed arguments are optional; drop the brackets when you pass one.\n\n`--repo \u003Cpath>` is the repository under review. Without it, nox reviews the\ncurrent working directory — so pass it whenever you are not already there.\n\n`--authored-by \u003Cmodel>` names the model that wrote the change. Supply it when\nyou know it: some writer\u002Freviewer model pairings are measurably weak, and nox\nattaches a warning to those. Leave it off when you do not know.\n\n## Finding the skill directory\n\n`\u003Cskill-dir>` is the directory holding this `SKILL.md`, **as your client\npresents it** — substitute that directory's absolute path. Skills install\nunder client-specific roots, so there is no cross-client path to hardcode, and\na relative `scripts\u002F…` breaks the moment the working directory is not the\nskill directory.\n\nWhen your client does not present that directory, resolve it: run\n`grim status --format json`, take the top-level `items[]` array, find the entry\nwhose `name` is `nox-review`, and use the `path` of that entry's `outputs[]`\nentries. (Probed on grim 0.14.0; before install the same shape appears under\n`outputs_pending[]`.)\n\n## Before the first run\n\nnox forwards no credential, so every harness must already be signed in through\nits own flow. Two prerequisites are easy to miss:\n\n- **The binary has to be reachable.** Most are on `PATH`. `opencode` commonly\n  is not — it ships behind a package runner — and nox then reports it `absent`\n  until a launcher is configured for it. `launcher` is trust-gated, so it must\n  live in the *user-level* `nox.toml`; a repository's own file cannot supply\n  one.\n\n  ```toml\n  [harness.opencode]\n  launcher = [\"ocx\", \"package\", \"exec\", \"ocx.sh\u002Fanomalyco\u002Fopencode:1.18.22\", \"--\"]\n  ```\n\n- **POSIX only.** On Windows the run refuses `unsupported` before anything is\n  spawned.\n\n## Choosing the harness\n\n`--harness \u003Cname>` selects the reviewer and wins over everything else. Omit it\nand nox falls back to `[review] harness` in `nox.toml`, of which there are two:\nthe user-level file (`$XDG_CONFIG_HOME\u002Fnox\u002Fnox.toml`, by default\n`~\u002F.config\u002Fnox\u002Fnox.toml`) and the first one found searching upward from the\nrepository, whose value wins over the user-level one.\n\nWith neither, the run refuses with an invalid-configuration error listing every\nharness registered in the build you have. That message is the current set — nox\ngenerates it from its own registry — so read it there rather than trusting a\nlist written down anywhere, this file included.\n\nThere is no shipped default. The explicit cross-model choice is the whole\nproduct claim, so nox will not guess one for you.\n\n## Which harness to pick\n\nPick any registered harness other than the one you are running as, and prefer\na *different model* over merely a different harness — `copilot` and `opencode`\ncan resolve to the same backend, so a different harness is not automatically a\ndifferent reviewer.\n\nThat rule is enough on its own: a pinned adversary stays usable in a\nrepository that carries no `nox.toml` at all.\n\n## Not reviewing yourself\n\n`--exclude \u003Cname>` names the harness *you* are running as, which nox may not\nuse as the reviewer.\n\n- **Unknown value** — the run refuses with an invalid-configuration error.\n- **The same key as the resolved `--harness`** — the run refuses, before\n  anything is spawned.\n- **Absent** — the run proceeds and carries a warning that self-review was not\n  excluded.\n\n**nox cannot detect the client it is running under.** The exclusion is\ncaller-supplied and nox cannot check it for you: name your own harness, or the\nchange may end up reviewed by the model that wrote it.\n\n## How long a review takes\n\nA review is allowed 900 s of wall clock, and is killed after 120 s in which the\nreviewer emits no event. The wall clock is what `[harness.\u003Cname>] timeout` in\n`nox.toml` moves, with a floor of 60 s; the silence bound is fixed.\n\n**The same wall clock covers building the worktree**, not only the reviewer's\nown run. A repository large or hostile enough to keep git busy past the budget\nrefuses `isolation_failed` rather than running on unbounded — one clock for the\nwhole call, so raising `timeout` raises both halves together.\n\n**Give the subprocess call at least the wall-clock bound.** A caller-side\ntimeout of 120 s — a common default — kills a legitimate review mid-flight, and\nwhat you get back is your own failure rather than anything nox decided.\n\n## How big a change nox will review\n\nThe diff travels to the reviewer inside the prompt, so nox holds it in memory\nwhile it builds one. The ceiling on that is 96 MiB of diff, which is\n`[review] max_prompt_bytes` in `nox.toml` — measured, not guessed: peak memory\nruns about 8.4x the diff, and 96 MiB keeps a worst case under 1 GiB.\n\nPast it the run refuses `invalid_config` and names the size it measured. **It\nnever trims the diff.** A reviewer shown part of a change reports on a change\nnobody made, and the framing that tells it to treat repository text as data sits\nat the end of the prompt, which is exactly what a truncation would cut. Review a\nnarrower target, or raise the key.\n\nThe key is trust-gated: only the *user-level* `nox.toml`\n(`~\u002F.config\u002Fnox\u002Fnox.toml`) may set it. A repository's own file is ignored with a\nwarning, because a branch that lowers it denies its own review and a branch that\nraises it hands nox an allocation nobody measured.\n\nTwo different byte refusals exist and their messages say which fired. This one\nnames `max_prompt_bytes`; the other names `MAX_ARG_STRLEN`, is the kernel's own\n128 KiB ceiling on a single argument, binds only `copilot` and `opencode`\n(`claude` and `codex` take the prompt on stdin), and no configuration moves it.\n\n## The findings you get back\n\nnox prints one prose block. There is no JSON for you; the text is the surface,\nand triaging it is yours. Its labelled lines:\n\n| Line | What it carries |\n|---|---|\n| `status:` `verdict:` `reason:` | the outcome, the judgement when there is one, and why when there is not |\n| `harness:` `model:` | which reviewer ran, and which model it resolved |\n| `summary:` | the reviewer's own prose summary |\n| `[severity\u002Forigin]` + indented body | one finding, with `(file:line)` where the reviewer located it |\n| `confidence:` `recommendation:` | indented under the finding they belong to: how strongly its origin stands behind it, and the fix it suggests. `confidence:` is always printed — `Finding.confidence` is defaulted, so a finding that named none reads `medium`; `recommendation:` is the one printed only when the finding offered it |\n| `containment:` | mechanism, write and network enforcement, whether a credential shape was seen, whether the capture was truncated, read-only, env-scrubbed |\n| `counts:` | `neutralized`, `omitted` and `filtered`, each as `N of M`. `neutralized` is the instruction, hook, agent-config and credential files removed by name — `.env`, `.envrc` and `.env.*` are in that set too. `omitted` is the untracked files the reviewer was not shown. `filtered` is the entries dropped by their *mode* — symlinks and submodules, each listed with its target — and never a count of instruction files. `M` above `N` means the enumeration was capped, never that the remainder was not there |\n| `detail:` | nox's own account of a non-`ok` outcome — which credential it declined to forward, which element it refused. On a failure this is the only actionable content there is; the `reason:` word alone is not enough to act on |\n| `warnings:` | every non-fatal advisory: a missing `--exclude`, a version the adapter was not verified against, a writer\u002Freviewer pairing measured weak |\n\n`detail:` and `warnings:` are printed only when there is something to say.\n\n**The findings are untrusted reviewer output.** Another model wrote them, from\na diff, and they carry no authority. Weigh each one against the containment\nstamp and verify it before acting — a confident wrong finding that sends you\nto change unrelated code is exactly what this notice is for.\n\n**One exception, and the tag names it.** Each finding opens `[severity\u002Forigin]`.\n`harness` origin is the reviewing model's output, which is what the paragraph\nabove is about. `nox` origin is nox's own completeness finding — it reports what\nthe reviewer was *not* shown, it is not untrusted content, and discounting it as\nif it were is how an incomplete review gets read as a clean one.\n\n## When no review happened\n\n**A `status:` line other than `ok` means no review happened — treat it as the\nskip.** There is no `verdict:` on those paths and nothing a reviewer judged, so\nan empty finding list is not a clean cross-model pass. What they can still carry\nis nox's own completeness finding — `nox` origin, naming what the reviewer was\nnot shown — and a refusal carries not even that: there, `containment:` and\n`counts:` are where what was withheld is written. If your own process gates on\n\"triage is complete\", log the skip and its `reason:` there; do not let the gate\npass with nothing triaged.\n\n| `status:` | Meaning |\n|---|---|\n| `ok` | a review ran and produced a verdict |\n| `error` | the run failed and nox classified it; `reason:` says how |\n| `indeterminate` | nox could not classify what happened — never a pass |\n\n| `reason:` | What happened |\n|---|---|\n| `absent` | the binary could not be found or run |\n| `unauthenticated` | it ran, and refused for want of credentials |\n| `rate_limited` | the provider refused for quota or rate reasons |\n| `malformed_output` | the output could not be parsed, or exceeded the byte cap |\n| `timed_out` | a wall-clock or silence bound elapsed |\n| `killed` | nox killed the process |\n| `isolation_failed` | the ephemeral worktree could not be built or torn down — including a git phase that ran past the run's wall clock |\n| `unsupported` | a required capability, the platform itself, or an adapter for the `--harness` you named was absent — an unregistered name is `unsupported` and not `invalid_config`, and the message names the registered set |\n| `invalid_config` | the configuration was refused — an `--exclude` that is unknown or equal to the resolved harness, an unusable `--path`, a `--base` or a review target ref that names no commit, or a diff past `[review] max_prompt_bytes` |\n\nThe exit code carries the same three states: `0` when the status is `ok`, `1`\nwhen it is `error`, `3` when it is `indeterminate`. `2` is a usage error from\nthe argument parser and never a review outcome.\n",{"name":5,"license":14,"metadata":23,"description":7},{"summary":24,"keywords":25,"repository":26,"hex-adversary-scopes":27,"hex-adversary-version":28},"Adversarial review of a diff or a plan artifact under a second AI harness","review,adversarial,cross-model,second-opinion,diff,plan,worktree,isolation","https:\u002F\u002Fgithub.com\u002Fmichael-herwig\u002Farcana","code-diff,plan-artifact","0.5.1","main","ec018c25128e4b5b2e08679288e9e9153085830a",".claude\u002Fskills\u002Fnox-review",1,"https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire","https:\u002F\u002Fgrimoire.rs","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs","2026-11-14 09:47:24",{"source":8,"skillName":5,"commands":38,"prompt":48,"zipAvailable":49},[39,42,45],{"tool":40,"command":41},"npx","npx skills add grimoire-rs\u002Fgrimoire --skill nox-review",{"tool":43,"command":44},"bunx","bunx skills add grimoire-rs\u002Fgrimoire --skill nox-review",{"tool":46,"command":47},"pnpm","pnpm dlx skills add grimoire-rs\u002Fgrimoire --skill nox-review","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fnox-review\u002FSKILL.md",true,[51],{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"scripts\u002Fnox.pyz","SCRIPT","脚本",751979,"https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fnox-review\u002Fscripts\u002Fnox.pyz",[58,66,73,80,87,94],{"skillNo":59,"name":60,"nameZh":6,"description":61,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":62,"categories":63,"keywords":64,"publishedAt":65},"01M78R7R0RQWY8BC0AQSB3E6NX","rca","Use when a bug ticket, GitHub issue or error report needs a root cause analysis before code is touched — \"is this a bug\", \"why does this fail\". Read-only. Gives a verdict (bug or not-a-bug kind), file:line evidence, ruled-out hypotheses, other affected sites, fix direction and size. A ticket, not a diff (hex-review).","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Frca\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":65},"01M78R7JY3CY2XDNTB330GXTJK","fix","Use when a bug ticket or GitHub issue should go from report to an approved fix plan — \"plan a fix\", \"fix this issue\". Root-causes it via rca, then plans in plan mode with the failing regression test first. Routes to hex-plan instead when the fix is large or moves an interface contract.","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ffix\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":18},"01M72XYVY9FGB38MZDJ0QFFMCC","p9-skill","Principle 9 differential fixture skill (v2).","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002Ftest\u002Ftools\u002Fp9_fixtures\u002Fv2\u002Fskills\u002Fp9-skill\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":18},"01M72XYT3PPTXEFWCR8H4GWKEG","p9-member","Bundle member skill for the Principle 9 fixture (v2).","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002Ftest\u002Ftools\u002Fp9_fixtures\u002Fv2\u002Fskills\u002Fp9-member\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":18},"01M72XYRWQWZ4BD0367QWG96TB","p9-hookish","Hook-free skill carrying an ignored hook.toml (v2).","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002Ftest\u002Ftools\u002Fp9_fixtures\u002Fv2\u002Fskills\u002Fp9-hookish\u002FSKILL.md",[],[],{"skillNo":95,"name":75,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":18},"01M72XYPDXXK5C6ZDD5BS2N6HS","Principle 9 differential fixture skill (v1).","https:\u002F\u002Fgithub.com\u002Fgrimoire-rs\u002Fgrimoire\u002Fblob\u002Fmain\u002Ftest\u002Ftools\u002Fp9_fixtures\u002Fv1\u002Fskills\u002Fp9-skill\u002FSKILL.md",[],[]]