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.

referencesdiagnostics-and-results.md

≈2k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Diagnostic and result guidance

Exit codes, result interpretation, finding decisions: SKILL.md. Read for output-path edge cases, machine status, usage accounting, local stage observation.

Local stage diagnostic

  • Observe first streamed Codex process: --engine-stage-dir /outside/repo/fresh-private-directory

    • --stream-engine-output.
  • Create/exclusively hold fresh mode-0700 directory on responsive local POSIX storage. Helper neither creates it nor overwrites engine-stage.json.

  • No observation for other engines, buffered output, dry runs; no environment variable activates it.

  • Sidecar: only closed booleans/enums, ≤4096 bytes; local launch, read, parser, display, first-process terminal observations. Retries/passes never update it; terminal state can differ from review outcome.

  • spoofable always true; provider_receipt/acceptance always false.

  • Missing/false = unknown, not provider activity, native initialization, acceptance. Redaction wrapper leaves flushed-display observations unknown.

  • Ordinary observation/persistence failures preserve primary result. Persistence rejects repository paths, symlinks, unsafe ownership/permissions, existing names.

  • Failed cleanup may leave private .engine-stage.partial or unresolved closure; retain custody when clearing it. Never retry ambiguous closes.

  • Cancellation propagates; cleanup cancellation can replace pending exception after remaining attempts.

  • Assumes ordinary descriptor acquisition/assignment, trusted caller custody, responsive storage; not hostile same-user interference, concurrent descriptor reuse, arbitrary interruption, network/stalled filesystems, crash durability.

  • No change to report, usage, deadline, exit, isolation rules.

Results

  • All output paths must be outside repo as both final entries (resolve parent symlinks) and resolved referents.

  • Without --status-output: human/JSON outputs must name distinct final entries; distinct symlinks/hardlinks may share existing referent (publication replaces entries).

  • Status enabled: all resolved referents must differ; refuse hardlink aliases.

  • Conservatively refuse case-only/Unicode-normalization aliases on every platform.

  • Tilde/relative paths expand once, anchored to invocation directory.

  • Atomic writes replace final symlink; never modify its target.

  • After path and alias checks, probe report parents for file creation, writing, and atomic replacement before target selection or reviewer startup. Existing reports remain untouched; missing parents are not created. This checks current storage readiness, not capacity or permissions for a later report write. Existing read-only Windows report entries are refused before the probe.

  • Event streams: non-null structured_output must contain report object; only absent/null permits result. Last terminal event authoritative.

  • --status-output /outside/repo/status.json: versioned outcome; validated --json-output report format unchanged. Completed status values: scoped-clean, findings, filtered, incorrect, incomplete.

  • Launched reviewer fails or returns invalid output: reviewer_unavailable (exit 1).

  • A handled signal while configuring or running reviewers publishes interrupted, with reason: interrupted, report_produced: false, signal_number, and exit code 128 + signal_number. reviewer_exit_code stays null: the signal was received by the helper, not an observed reviewer exit. Owned-child cleanup still runs; this is no review verdict.

  • Failed later pass never publishes partial report.

{
  "schema_version": 1,
  "status": "reviewer_unavailable",
  "exit_code": 1,
  "engine": "codex",
  "report_produced": false,
  "reason": "engine_failed",
  "reviewer_exit_code": 124,
  "timed_out": true
}
  • Unavailable reasons: engine_failed, invalid_report, runtime_validation_failed; completed reviews use null.

  • Runtime validation: Amp post-launch isolation attestation/private-result check refused output, not transient-provider failure.

  • reviewer_exit_code: retained last process code (possibly zero for rejected output), otherwise null. timed_out: helper's deadline, not merely process exit 124.

  • Inspect status independently of --expect-findings exit changes.

  • Unfinished assessment: keep validated observations with incomplete, exit 2, report_produced: true, even with findings. Never publish private completion field.

  • Missing/invalid completion: invalid report, not unfinished.

  • Status contains no logs, prompts or findings. OpenClaw status may include effective_models, the unique provider/model identities observed in validated terminal responses; rejected or missing identities are not recorded. Other engines' status fields are unchanged.

  • Bounded display-safe stderr diagnostics + command-auth suppression stay enabled.

  • Use a fresh status path for real reviews: after argument/path validation, remove old status before target selection.

  • Dry runs validate output paths but do not create, replace, or remove output files. Any existing status belongs to a prior invocation, not the dry run.

  • Preflight/pre-launch isolation refusals, source mutations, interruptions outside reviewer execution, and output failures may produce no new status. Absence never means a clean verdict.

Preflight summary

--preflight-output PATH --preflight-repo owner/repository writes one raw openclaw-preflight:v1 comment, at most 4096 UTF-8 bytes. The explicit repository is a caller claim about the target, never inferred from a fork's remote. Only a clean, nonempty branch range is supported, including auto when it selects a branch. Base and head identify the actual captured merge-base and commit; local/index/working-tree mixtures are not relabelled as commits.

The deterministic summary includes the selected priority, native status, and known accepted/filtered/incomplete counts. Priorities below the selected threshold are not assessed. Validated incomplete maps to partial; other validated outcomes map to complete, including findings and incorrect verdicts. Neither means clean. The descriptive source-<12 hex> version fingerprints the helper file; it is not a released package version or trusted execution evidence.

The output follows the same external-path/atomic-write rules as native reports, with distinct destinations outside the engine-stage directory. A prior block is removed before review. No safe validated outcome, source drift, or unavailable reviewer means no new block. An export failure after native publication warns on stderr and leaves the native report, status, and exit unchanged. Check that the new file exists before sharing. The helper does not attach, upload, fetch a report, or bypass hosted review. Summary JSON carries no model prose, source text, report hash, roles, or usage.

Usage

  • Codex collects usage with live display on/off. Reports, status, terminal summary: process attempts; reported, unknown, partial counts; complete; observed token totals.

  • Each attempt contributes last valid cumulative snapshot once, including access retries + failed passes. Retry policy: engine options.

  • Cached input/reasoning output: subsets, not extra totals.

  • Counts: observed tokens, not billing estimates or cache-hit promises.

  • Missing telemetry (including all-zero defaults without sample): unknown, never measured zero. tokens: null: no usable totals; incomplete totals: lower bounds.

  • Interrupted runs retain usage in the metadata-only status when interrupted during reviewer execution, and print it when output remains available. They publish no review report; interruptions outside that phase may leave no status.

  • Other engines do not aggregate usage.

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