| name | review |
| description | Multi-angle code review via sideclaw MCP tool (claude-sonnet-5, currently on Max per SIDECLAW_WORKER_BACKEND=max; IU unified endpoint fallback when unset). Add --deep to also run Anthropic's native correctness + security review, also on Max. |
Review
Default (currently on Max) — sideclaw multi-angle
mcp__sideclaw__review is asynchronous (background job on claude-sonnet-5, currently on Max per SIDECLAW_WORKER_BACKEND=max; falls back to the IU unified endpoint when unset):
- Call
mcp__sideclaw__review with cwd set to the repo root → returns { jobId }. Parse args for scope (default uncommitted): e.g. head (last commit only), HEAD~3 (a bare ref = the range up to HEAD, i.e. the last 3 commits — not the single commit), main..HEAD (explicit range), path/to/file.ts. Strip any leading flags (like --deep) before extracting the scope.
- Call
mcp__sideclaw__job_wait({ jobId }) to block until it finishes (loop while stillRunning: true); read result on status: "done". The submit call does not return the findings.
This runs the deterministic floor (architect, senior-dev, and file-type angles)
plus a triage router that adds content-driven angles — security, performance,
concurrency, data-migration, api-contract — when the diff warrants them. All on
claude-sonnet-5, currently on Max per SIDECLAW_WORKER_BACKEND=max (IU
unified endpoint fallback when unset). Use this by default.
--deep — add native correctness + security (also on Max)
When the args contain --deep, ALSO run Anthropic's native reviewers and merge
them with the sideclaw result. These run on the orchestrator's Max model and are
tuned for real correctness bugs — complementary to sideclaw's architecture /
framework / style angles. Reserve --deep for pre-ship gates or risky changes,
not routine reviews — it spends Max.
- Run the sideclaw review (as above).
- Invoke the native
code-review skill at high effort over the same scope
(Skill tool, args: "high"). It targets correctness bugs the angle reviewers
may miss.
- Conditionally invoke the native
security-review skill if the diff
touches security-sensitive surface — auth/authz, secrets or credentials,
crypto, SQL/command/path construction from input, file uploads, shelling out,
or env-var handling. Skip it otherwise.
- Merge all findings into one verdict, deduplicated against the sideclaw
findings. Map native Important → blocking, Nit → improvements. Keep
sideclaw's outcome classification (
clean / actionable / needs-human); if
the native pass surfaces a blocking bug, the merged outcome is at least
actionable.
Output
Present one consolidated verdict — outcome, blocking, improvements, discussions,
testGaps, summary — noting the catching reviewer per finding. Don't echo each
tool's raw output; synthesize.
Worker caveats when consuming findings:
- Line numbers are unreliable — the worker often cites lines that don't exist (e.g. L688 in a 600-line file). The substance is usually accurate; the location isn't. Before reporting or acting on a finding, grep for the symbol/code, not the cited line.
- Angle reviewers explore the live tree, not just the diff (they have Bash/Read/Grep). A finding may reference current code outside the reviewed scope — useful, but verify it's in-scope before treating it as introduced by the change.
- Before claiming a finding is wrong, verify it in the actual code — don't dismiss it on the bad line number alone.