Review Agent

Code review specialist that evaluates pull requests for correctness, security, intent alignment, style, and documentation currency.
How the agent works
The review agent is triggered when a PR is opened or updated. It follows the same pre-script / sandbox / post-script pipeline as the other agents.
- Pre-script validates inputs and fetches PR metadata.
- Sandbox — the agent runs the
pr-revieworchestrator skill. The orchestrator triages the change, then dispatches specialized sub-agents in parallel — each covering a distinct review dimension (correctness, security, intent & coherence, style & conventions, docs currency, and optionally cross-repo contracts). Sub-agents run concurrently and return structured findings. The orchestrator collects, deduplicates, and synthesizes findings across dimensions, runs PR-level checks (scope authorization, protected paths), and produces a structured JSON review result. The agent cannot push files, edit code, or push — it is strictly read-only. - Validation loop — the output is checked against a schema, with up to 2 retry iterations if the output is malformed.
- Post-script posts the review on the PR. When
REVIEW_RISK_ASSESSMENT_ENABLEDis on (the default) and the result JSON omitsrisk_assessment,fullsend post-reviewposts a sticky diagnostic (Risk assessment unavailable this run) so a missed risk-assessment pre-pass is never silent. Workflow logs also carryfullsend:persona:invoke/complete/error/reject/skip markers for each named persona.
If a prior review exists (e.g., re-review after fixes), it is injected into the sandbox so the agent can assess whether previous findings were addressed.
How it helps
- Every PR gets a thorough review within minutes, regardless of team availability.
- Reviews cover security, correctness, intent & coherence, style, and docs currency — dimensions humans sometimes skip under time pressure.
- The structured output format makes it easy to see what was flagged and why.
Commands
| Command | Where | Effect |
|---|---|---|
/fs-review | PR comment | Triggers a review on the PR (per-repo installs only; standalone issues are ignored) |
Requires triage-level repository permission or higher (triage, write, maintain, or admin). Mutation stages such as /fs-fix still require write or higher.
The /fs-review command does not accept arguments. The review agent also runs automatically when a PR is opened, synchronized (new commits pushed), or moved out of draft by a user with triage-level repository permission or higher. On GitLab, automatic review fires when the cron poller sees an MR whose created_at is newer than the watermark (up to one poll interval of delay). Native merge_request_event dispatch was removed; all GitLab events route through the poller. Push-to-open-MR (GitHub synchronize) is not auto-detected; comment /fs-review to re-review after new commits.
Control labels
These labels are applied by the review post-script based on the review outcome.
| Label | Meaning |
|---|---|
ready-for-review | Workflow state marker on the PR. Applied by the code agent post-script after pushing. In per-repo installs, triggers review when applied to a PR. |
ready-for-merge | The review agent approved the PR. No blocking findings. |
requires-manual-review | The review agent found issues that require human judgment — it could not confidently approve or reject. |
rejected | The review agent rejected the PR and the post-script closed it. |
When the review agent requests changes (without rejecting), no outcome label is applied — the pull_request_review event triggers the fix agent directly.
Stale outcome labels from prior review runs are removed before the new one is applied.
The issue-labels skill may also apply contextual labels (e.g., area/api, priority/high) but these are informational -- they do not control agent behavior.
Configuration and extension
Skill: issue-labels
The review agent includes the issue-labels skill to discover your repo's labels and apply them to PRs during review. This is the same skill used by the triage agent -- overloading it affects both agents.
To overload the built-in skill, create your own issue-labels skill in .agents/skills/issue-labels/SKILL.md and symlink .claude/skills to .agents/skills so it's discoverable by both fullsend and local agent tooling. You can also overload it at the org level using config-driven agent registration -- see Bring Your Own Agent.
See Configuring with AGENTS.md and Configuring with Skills.
Variables
| Variable | Description | Default | Valid values |
|---|---|---|---|
REVIEW_FINDING_SEVERITY_THRESHOLD | Minimum severity for findings to include in the review. Findings below this level are omitted from both the narrative body and the posted inline comments. | low | info, low, medium, high, critical |
REVIEW_RISK_ASSESSMENT_ENABLED | Gates the ADR 0089 risk-assessment pre-pass and the missing-assessment diagnostic. When on and the result JSON omits risk_assessment, fullsend post-review posts a sticky "Risk assessment unavailable this run" comment. | true | true/false (also accepts 0/no/off as false) |
REVIEW_FINDING_SEVERITY_THRESHOLD is read inside the review agent's sandbox, so set it in the harness's env.sandbox (the upstream default lives in harness/review.yaml).
REVIEW_RISK_ASSESSMENT_ENABLED is different: it is read independently in two separate execution contexts. The review orchestrator reads it from env.sandbox to decide whether to dispatch the ADR 0089 risk-assessment pre-pass at all. fullsend post-review separately reads it from the runner process's own environment (env.runner) to decide whether to post the missing-assessment diagnostic. Setting it only in env.runner suppresses the diagnostic but still dispatches the persona (spending tokens on a disabled feature); setting it only in env.sandbox stops dispatch but leaves the runner-side default of true in effect, so the diagnostic still fires on every run. Fully enabling or disabling the feature requires setting the variable consistently in both places.
To override either variable per repo or org, use base: composition rather than the CI workflow env: block — workflow env: is reserved for infrastructure plumbing (see Architecture for details on harness composition and workflow-env conventions).
The review agent omits findings below the threshold from its output. The post-script also filters the structured findings array as defense-in-depth. When filtering removes all findings from a request-changes or reject verdict, the post-script downgrades the verdict to comment (applying the requires-manual-review label).
