Index of every page on this site: llms.txt
All skills
openclaw avatar

/autoreview

@0a5ba4b
by openclawopenclaw/agent-skills1.1k stars
126
Run it yourself

Structured code review when explicitly requested, preferring OpenAI/Codex before Claude.

Use this Skill: https://skilld.dev/gh/openclaw/agent-skills/autoreview

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

referencesengines.md

≈3k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Engine options

Read for model/effort overrides, provider projection, optional engines.

These options configure helper engines, not Copilot model access. When Codex CLI is unavailable, the skill permits a separate Copilot subagent review, preferably with an accessible Codex model or otherwise a comparable model. Select it before launch; it is not --engine copilot, does not inherit Codex CLI flags/defaults, and produces no helper-validated machine status or helper isolation guarantees.

  • Defaults + engine-selection policy: SKILL.md.

  • Explicit gpt-6.1-sol: same access-only gpt-6-sol retry. Explicit gpt-6-sol: access-only gpt-6-luna retry.

  • Other explicit models, including Luna/Astra: no model fallback.

  • Explicit gpt-5.6-sol: access-only gpt-5.6-terra retry.

  • GPT-6.1 Sol rejects none/minimal effort before review preparation; GPT-6 Sol/Luna reject minimal. Effort-only override keeps default model.

  • Honor explicit user engine/model choices. No automatic fallback between engines.

  • Override defaults: --engine, --model, --thinking.

  • Codex tier: Ultrafast when active model catalog lists it; otherwise Codex silently sends standard tier. Raw --codex-config service_tier=... keeps its value.

  • --codex-speed fast: priority service. --codex-speed default: standard tier.

  • --codex-no-fallback: keep the selected Codex model and initial authentication route. Disables all helper model retries and configured-provider rerouting, including the missing-login shortcut before the first attempt. An explicit initial provider remains selected; otherwise retain subscription authentication. Failure remains reviewer_unavailable with no validated report. Codex-only and opt-in; defaults, native recovery within the same provider, and the per-attempt deadline remain unchanged. This flag does not grant model or account access.

  • --codex-config app_server.max_input_text_chars=3145728 sets a startup text-input character limit only when the selected CLI supports that setting. Accepts decimal digits from 1 to 4194304; omission preserves the selected CLI's default. This does not change the model context window or output budget. Other app_server settings remain unsupported.

  • Only Claude accepts --fallback-model.

  • Per-engine environment overrides: AUTOREVIEW_<ENGINE>_*.

Account cannot access Sol/Luna: pin available model. Require GPT-6 Astra without model fallback: select explicitly:

"$AUTOREVIEW" --mode local --model gpt-6-astra --thinking high
  • GPT-6.1 Sol + GPT-6 Astra: low, medium, high, xhigh, max; neither supports none/minimal.
  • GPT-6 Sol + Luna: additionally support none, never minimal.
  • AutoReview default: medium; no fallback from explicit Luna/Astra selection.
  • Codex ultra: automatic delegation; outside helper's supported effort levels. Deepest supported review: max.
  • EU data residency: --codex-speed default; GPT-6 fast mode unavailable there.

See GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna model docs and Codex reasoning modes.

Default: ChatGPT subscription; preserve only user authentication settings. Live authentication failure → stop attempt fast; retry pass once on operator's validated configured named command-auth route, including context + catalogue. Definite credential failure → immediate stop; generic auth failure → stop after 3 lines within 60s. Single 401: allow Codex's reload/refresh recovery. Explicit provider / non-auth failure / no valid configured route: no auth fallback. Profile settings remain ignored. Force configured route before subscription attempt:

"$AUTOREVIEW" --mode local --codex-config 'model_provider="review_api"'
  • Selector must match model_provider in operator's external CODEX_HOME/config.toml. Accepts one bare or simply quoted identifier; overrides cannot supply provider definitions or other capabilities.

  • Projection requires Python 3.11 or tomli; default auth-only operation keeps existing fallback parser.

  • Selected route must use an absolute https base URL with a host and no userinfo, query, or fragment (including empty ?/#). OpenAI-compatible Responses gateways are supported; http is refused, including loopback. The configured URL is projected unchanged. Command authentication requires an absolute external executable. Fixed arguments belong in its wrapper; omitted/empty auth.args accepted.

  • Choose the endpoint in the named provider, not through OPENAI_BASE_URL: passing that environment variable through does not guarantee the Codex client uses it.

  • Omitted wire_api/requires_openai_auth: keep Codex responses/false defaults, respectively. Optional auth timing/context settings: native defaults/semantics.

  • POSIX: private launcher restores validated caller HOME only for selected authentication executable. Engine/reviewer tools keep isolated environment/filesystem access.

  • Caller HOME must be available absolute directory; no repository-owned path or symlink provenance. Windows keeps native executable route.

  • Command-auth runs: suppress raw provider diagnostics; report fixed failure categories; retain compact progress, usage, assistant report streaming.

  • Command-auth runs use a separate empty file credential store, including on subscription fallback; they never inherit or refresh subscription credentials.

  • Empty final report: fail without exposing captured stdout.

  • Catalogue + authentication working-directory paths: resolve relative to operator config directory; must stay outside reviewed repository.

  • Supplied catalogue: copy byte-for-byte into private client runtime; retries use same route/catalogue snapshot.

  • Dry runs: same ownership/route shape checks; no authentication execution.

  • Codex owns catalogue validation, model access, context clamping.

  • Projection selected: other custom provider forms/split context overrides unsupported.

API hostname selection

"$AUTOREVIEW" --mode local --codex-route api.example.com
"$AUTOREVIEW" --mode branch --base origin/main --codex-route api.example.com --codex-no-fallback
  • Opt-in, Codex-only, POSIX-only. Reads ${CODEX_HOME:-~/.codex}/config.toml, never repository configuration or profiles. Configuration and its directory must be user-owned, external to the reviewed repository, with no group/world writable ancestry (root-owned sticky temporary ancestors are allowed). Repository-owned symlink provenance and multiply linked config files are refused. Requires Python 3.11 or tomli. Windows hostname selection fails closed because this ownership check does not attest Windows ACLs; existing routes are unchanged.
  • Selects from all top-level named provider definitions, independently of the active model_provider. Exactly one definition must match. Case, IDNA DNS names and one trailing DNS root dot normalize; lossy IDNA2003 mappings are refused rather than conflated with another hostname. IP literals normalize by address. No wildcard/suffix matching. Supply only a host, without URL, port, path or userinfo; IPv6 uses a bare address. Multiple definitions on the same host are ambiguous even with different ports or paths. Malformed/missing matches fail.
  • Refuses repeated --codex-route and any model_provider override, including AUTOREVIEW_CODEX_CONFIG, even if it names the same provider. Tuning overrides remain available, except context-window/compaction overrides that split the configured route. Profile and top-level base URL overrides remain unsupported.
  • Preserves the selected full HTTPS base URL, including port/path/escaping, Responses protocol, WebSocket and standalone-search support booleans, request and stream retry counts, stream idle timeout, external model catalogue and context settings. The existing URL safety constraints above still apply. Unsupported fields, headers, query parameters, environment auth and OpenAI login auth fail explicitly; settings are never silently dropped.
  • Accepts the existing validated command-auth form or one nonempty experimental_bearer_token using bearer-token syntax. requires_openai_auth must be false or omitted. Combining auth forms is refused. The existing --codex-config model_provider=... contract stays unchanged and does not admit direct bearer tokens.
  • Direct bearer tokens use a temporary mode-0600 file and an isolated Python reader in the private client runtime, outside reviewer-readable paths. Only native command auth reads the token; it never enters process arguments, environment, prompt bundles or provider flags. Runtime artifacts are removed on normal exit, handled failure and interruption; abrupt process/host death retains the normal temporary-directory cleanup limitation.
  • No subscription/auth.json linkage, keyring policy, inherited API credential or endpoint variables, unrelated tools, hooks, MCP servers or permissions are imported. Provider diagnostics stay suppressed; known bearer values are also redacted from displayed assistant output and final report text.
  • Selection is explicit: failed validation/auth/transport never switches provider, endpoint or credential. Existing model-access retries stay on the same frozen route/token/catalogue snapshot; add --codex-no-fallback to disable those too. Dry-run performs selection and ownership checks without staging or executing authentication. Omitting the new option preserves existing behavior.
Optional engine Prerequisites
Claude CLI 2.1.169+; safe mode with web-only tools
Amp AMP_API_KEY for a plugin-free account; local POSIX execution, no custom endpoint or cloud/orb agent
Pi CLI 0.79.0+; configured model; no tools or project resources
  • --engine kimi: recognized, refused for reviews + --dry-run before any Kimi process, configuration read, or authentication setup.
  • Supported Kimi Code prompt mode accepts review content only as command-line argument; helper has no supported private prompt input channel.
  • Previous Kimi execution path intentionally retired; never expose bundle in process arguments. Existing --kimi-bin arguments accepted for same clear refusal.
  • Never silently select another engine. Custom agent file ≠ equivalent helper replacement: changes input into templated system prompt. The documented Copilot subagent route is a separate assessment with explicitly reported limitations.

OpenClaw

--engine openclaw uses restricted openclaw agent exec, directly on a trusted host or through the bound sandbox autoreview tool. Omitted model and effort use normal OpenClaw selection; explicit helper options and environment settings retain their normal precedence. Direct execution inherits native configuration and authentication; optional --openclaw-config selects an external operator config. The helper owns review restrictions, the empty workspace, private prompt input, optional timeout and cleanup. Bound integrations apply current caller permissions before execution and return the admitted model for terminal validation. There is no child reviewer profile, model fallback or output-schema CLI dependency. See execution and integration requirements before launch.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 alerttoday3 checks · Risk SAFE
  • Gen Agent Trust Hubtoday

    The autoreview skill is a specialized tool for performing automated code reviews using AI engines. It includes extensive testing fixtures and documentation focused on security, specifically addressing process isolation, environment sanitization, and the detection of credentials and vulnerabilities in code. All potentially suspicious patterns, such as dynamic code execution and hardcoded fake secrets, are confined to the test suite and are used to validate the tool's robustness and security boundaries.

  • Socket3d

    1 alert: gptSecurity

  • Snyktoday

    Risk: CRITICAL · 3 issues

Signed by skilld at 0a5ba4b. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated 2 weeks ago

README badge

README badge for openclaw/agent-skills/autoreview