Run Claude Code Review

Previous value: -"Base branch/tag for scope=branch (default: auto-detect the default branch)"New value: +"Base branch/tag for scope=branch and for scope=auto's branch step (default: auto-detect the default branch). A validated ref always reaches the reviewer — it is never silently dropped."

Previous value: -"high"New value: +"max"

Previous value: -"Review effort / reasoning level, same scale as Claude Code (--effort): low..max. max/xhigh = exhaustive deep review, slower and pricier."New value: +"Review effort / reasoning level, same scale as Claude Code (--effort): low..max. Default max = exhaustive deep review (slower and pricier). Omit for the deepest default; pass low/medium for a fast, cheap pass."

Previous value: -trueNew value: +false

Previous value: -"Auto-install the bundled code-review skill into ~/.claude/skills when missing (default true)"New value: +"Only meaningful with pipeline=skill: when the bundled skill is MISSING, install it into the Claude config dir (default false — a missing skill degrades to the inline fleet pipeline instead; a MANAGED existing copy is refreshed whenever it is about to be used; a user-managed copy is never touched). env CLAUDE_REVIEW_AUTO_INSTALL_SKILL=1 sets the default"

Previous value: -"API spend cap for this run via --max-budget-usd (default 5); 0 disables. On third-party gateways this depends on usage data the gateway returns."New value: +"API spend cap via --max-budget-usd. DEFAULT 0 = NO CAP at any effort level: a cap stops the parent before it synthesizes, and every finished subagent report is already captured incrementally (and recovered from output_file), so a 429/token exhaustion no longer wastes it. The wall-clock timeout (timeoutMinutes) is the guard. Set a number only if you want a hard spend stop; env CLAUDE_REVIEW_MAX_BUDGET_USD sets the DEFAULT for this argument (an explicit value here always wins). On third-party gateways this depends on usage data the gateway returns."

Added value: +{
+  "default": "auto",
+  "description": "Which review engine runs. auto (default): claude's built-in /code-review command for levels low..high — it needs no file on disk; fix runs ride its own --fix flag. max (including effort=xhigh), scope=last-commit with no parent commit, and a shadowing code-review skill use the fleet pipeline. Only max uses the MCP's own fleet pipeline, because the built-in hands max to subagents whose exploration commands the locked-down permission mode denies. native: force the built-in command (refused when a same-named skill in the Claude config dir shadows it). fleet: force the fleet pipeline (its instructions travel in the prompt; nothing is installed). skill: legacy — use/optionally install the bundled skill (the only mode that may write to disk).",
+  "enum": [
+    "auto",
+    "native",
+    "fleet",
+    "skill"
+  ],
+  "type": "string"
+}

Previous value: -"Also post the summary comment to the GitHub PR via gh (default false — report only)"New value: +"Also post the summary comment to a GitHub PR via gh. Requires a fully-qualified pr — \"owner/repo#123\" or a GitHub PR URL; a bare number is refused because gh would resolve it against the reviewed repository's remote (default false — report only)."

Added value: +{
+  "description": "The REVIEW's own depth level — the same argument claude's native /code-review command takes. The fleet has the same three-phase structure as the native playbook (finder -> one verifier per surviving candidate -> sweep), with these numbers (rendered from FLEET_NUMBERS in src/prompt.js): low = one pass, no subagents (<= 4 findings); medium/high = 8 finders (3 extra finders split briefs on top of the 5 briefs) x up to 6 candidates, findings capped at 8; max = 10 finders (every brief split) x up to 8 candidates + one verifier per surviving candidate + 1 sweep finder, findings capped at 15; at most 20 candidates reach verification. Defaults to a mapping of effort (max/xhigh->max, high->high, medium->medium, low->low). Separate knob from the CLI reasoning effort and from ultracode.",
+  "enum": [
+    "low",
+    "medium",
+    "high",
+    "max"
+  ],
+  "type": "string"
+}

Previous value: -"auto: PR if pr given / branch ahead of default / working-dir changes / last commit"New value: +"auto: PR if pr given / branch ahead of baseRef (when given) or the default branch / working-dir changes / last commit"

Previous value: -20New value: +30

Previous value: -"Hard wall-clock timeout for the run"New value: +"Hard wall-clock timeout for the run — with no spend cap by default this is the guard; every finished subagent report is kept even when it fires"

Added value: +{
+  "default": false,
+  "description": "Enable claude's Ultracode mode for this run (the `ultracode` settings key): standing dynamic-workflow orchestration — every substantive step goes through the native Workflow tool, multi-agent and token-cost-no-object. Orthogonal to effort (any effort level). Requires a claude CLI that supports the key; older builds ignore it (see workflowRuns in the result).",
+  "type": "boolean"
+}

Previous value: -"Block until the review finishes and return the report synchronously (2-15 min). Default false: returns a jobId at once — poll claude_review_status(jobId). Only set true if your MCP client has no ~30s tool-call timeout."New value: +"Block until the review finishes and return the report synchronously (2-15 min). Default false: returns a jobId after a fast preflight — poll claude_review_status(jobId). Only set true if your MCP client has no ~30s tool-call timeout."

Added value: +{
+  "default": false,
+  "description": "Block until the review finishes and return the report synchronously (2-15 min). Default false: returns a jobId at once — poll claude_review_status(jobId). Only set true if your MCP client has no ~30s tool-call timeout.",
+  "type": "boolean"
+}