Autoreview

Use the bundled helper as an advisory closeout gate. It freezes and scans the review bundle, isolates the reviewer from repository instructions and host files, validates structured findings, and stops after one clean result.

Pre-PR gate

When an agent is about to push or publish a PR, finish verification and commit the intended change, then run:

export AUTOREVIEW="${AGENTS_HOME:-$HOME/.agents}/skills/autoreview/scripts/autoreview"
"$AUTOREVIEW" --gate pre-pr --mode auto

The default cadence is exactly this pre-PR checkpoint. It reviews the complete branch diff and exits successfully without a model call when the PR changes only documentation, configuration, metadata, generated files, or locks. A dirty worktree fails the pre-PR gate because it would make the branch review incomplete.

If an accepted finding changes code, verify and commit the fix, then rerun the gate before publishing the PR.

A clean pre-PR review stores a private attestation for 24 hours. The key covers the exact substantive diff, base commit, reviewer contract, threshold, prompt, datasets, and helper revision. A later proof or documentation commit reuses the attestation when the substantive diff has not changed. Secret scanning and prompt validation still run before reuse. Pass --no-review-cache to require a fresh model review.

Configured checkpoints

Global and repository config can move the automatic cadence to item, task, phase, or step. At that plan boundary, run the matching gate:

"$AUTOREVIEW" --gate phase --mode auto

The helper skips mismatched checkpoints and non-code-only automatic checkpoints. A manual request always runs regardless of cadence:

"$AUTOREVIEW" --mode local
"$AUTOREVIEW" --mode branch --base origin/main
"$AUTOREVIEW" --mode commit --commit HEAD

Read CONFIG.md when adding a global, repository, or profile configuration.

Reviewer selection

With no explicit reviewer, the auto profile scores configured candidates, keeps only installed isolation-safe harnesses, and avoids recursively invoking the current host agent. The built-in profiles are:

"$AUTOREVIEW" --profile sol
"$AUTOREVIEW" --profile opus
"$AUTOREVIEW" --profile cross-lab
"$AUTOREVIEW" --profile value
"$AUTOREVIEW" --profile budget

Opus 5.5 high is the default Claude-side code reviewer. The built-in value profile selects Sol high. Shared model-selection policy version 2 checks every reviewer and fallback, including explicit requests.

Fable only coordinates work. Astra only handles focused specialist tasks with high complexity. Neither can review through this helper. Policy always prohibits Haiku.

Built-in Claude policy caps effort at high. A configured higher cap must still use an approved model and effort pair. Config schema version = 1 remains valid. Install the sibling model-selection skill with autoreview.

Config can register supported models and harnesses as candidates. Automatic selection checks the shared frontier before score ranking. External adapters require an explicit concrete model name for third-provider catalog models.

The helper also supports Kimi Code, OpenCode, and Cursor Agent when their CLIs exist:

"$AUTOREVIEW" --engine kimi --model kimi-k2
"$AUTOREVIEW" --engine opencode --model opencode/kimi-k3 --thinking max
"$AUTOREVIEW" --engine cursor --model grok-4.5
"$AUTOREVIEW" --reviewers cursor:grok-4.5,opencode:opencode/glm-5.2:max
"$AUTOREVIEW" --list-harnesses

Kimi model names use its configured model catalog. OpenCode model names use its runtime provider/model catalog. Cursor model names use the Cursor account catalog, including subscription-hosted models. A compatible Pi CLI enables Pi as an explicit or configured candidate.

Desktop applications and headless review harnesses are separate capabilities. For example, Cursor Desktop does not satisfy the cursor-agent requirement, and Codex Desktop does not satisfy the codex requirement. A missing CLI is never automatically eligible. Inspect or explicitly install harnesses with:

"$AUTOREVIEW" --list-harnesses
"$AUTOREVIEW" --install-harness cursor
"$AUTOREVIEW" --engine cursor --model grok-4.5 --install-if-missing

Installation is always explicit. Automatic scored selection skips missing harnesses. It never downloads software unless the caller passes --install-if-missing. With that opt-in, a profile lacking quorum installs the best missing isolation-safe harness. An unavailable harness fails with a machine-readable harness_unavailable or profile_unavailable marker, desktop-versus-CLI diagnosis, canonical installation command, and recovery options.

The helper does not install Kimi. Install Kimi Code separately before you select it.

Read MODEL_SELECTION.md when changing score axes, candidate defaults, effort policy, or benchmark inputs.

Findings

• Verify each advisory finding against the real code path.
• Accept concrete defects introduced by the reviewed change.
• Reject speculative edge cases, unrelated cleanup, and broad redesigns.
• Keep fixes inside the original task, changed owner boundary, and public contract.
• Stop on a clean helper exit. Add another reviewer only when the selected profile requires one.

The default threshold is P0. The helper sends the selected threshold to the reviewer and asks it to omit lower-priority findings. Deterministic post-processing applies the same threshold as a final guard. Widen the threshold only when requested:

"$AUTOREVIEW" --max-priority P1

Pause and report when a finding requires a new protocol, schema, storage layout, public API, release process, or owner boundary. Also pause when two review-triggered patch cycles have not converged.

Large review inputs

The helper permits eight bounded passes by default. Prefer a coherent change split when the complete diff exceeds that budget. Set an explicit --max-passes N budget for a replacement that must remain intact. Values range from 1 through 64. Record the reason and actual pass count in the owning plan. This option changes only the number of passes. It does not increase the per-pass byte limit, permit truncation, bypass secret scanning, or change reviewer isolation.

Safety

• TruffleHog scans the exact changed content before model invocation and fails closed on verified or unknown credentials.
• Codex receives the bundle in an empty read-only workspace with user config and repository rules disabled.
• Claude runs in safe mode with project skills, hooks, plugins, MCP servers, memory, filesystem, and shell access disabled.
• OpenCode runs from an empty workspace with project instructions, Claude compatibility, plugins, MCP servers, filesystem, shell, and edit tools disabled.
• Cursor Agent runs from an empty workspace with an isolated permission config.
• It uses Ask mode when supported and denies filesystem, shell, write, fetch, force-write, and MCP permissions.
• Pi runs without repository context files, extensions, skills, sessions, or tools.
• Kimi runs from an empty workspace with a staged configuration, no tools or subagents, and an empty skill directory. It requires Kimi Code 0.30.0 or newer.
• Each engine runs in an owned process group. An interrupt terminates the engine and its child processes.
• Explicit prompt, dataset, and repository config inputs remain inside the reviewed repository unless the user passes a trusted config path.
• The helper never pushes, commits, or mutates the reviewed repository.

Report the selected profile/reviewer, focused proof, accepted and rejected findings, and the final clean result or consciously rejected remaining finding.