review-fix-loop
Autonomous review-and-fix loop. Runs any code analysis tool, dispatches parallel subagents to fix all findings, verifies with the project's own lint and test commands, and loops until zero findings remain.
Examples
CodeRabbit review of all changes:
/review-fix-loop coderabbit review --plain -t allDeepSource issues as JSON:
/review-fix-loop deepsource issues list --json --output-file deepsource-findings.jsonAuto-detect tool from project config:
/review-fix-loopStep 1 — Determine the analysis command
Everything the user types after /review-fix-loop is available as $ARGUMENTS.
If the user provided a command (from $ARGUMENTS), use it as-is.
If no command was provided, auto-detect by checking for config files in the project root:
| Config file | Tool | Command | Auth check |
|---|---|---|---|
.coderabbit.yaml | CodeRabbit | coderabbit review --plain -t all | coderabbit auth status |
.deepsource.toml | DeepSource | deepsource issues list --json --output-file deepsource-findings.json | deepsource auth status |
.eslintrc* or eslint.config.* | ESLint | eslint. --format json | — |
ruff.toml or [tool.ruff] in pyproject.toml | Ruff | ruff check --output-format json | — |
If multiple config files exist, list them and ask the user which tool to run.
If no config file is found, ask the user to provide the analysis command.
Step 2 — Prerequisite check
Extract the tool name (first word of the analysis command) and verify it is available:
command -v <tool-name> >/dev/null 2>&1If not found, tell the user the tool is not installed and suggest installation. Halt — do not proceed until the tool is available.
If the tool has an auth check listed in the Step 1 table, run it. For user-provided commands not in the table, skip the auth check. If unauthenticated, tell the user how to authenticate. Halt until resolved.
Step 3 — Run analysis
Execute the analysis command and capture full stdout and stderr:
<analysis-command> 2>&1If the command exits with a non-zero status and produces no parseable findings, report the error to the user and halt.
Store the complete output for parsing.
Step 4 — Parse findings
From the captured output, extract all findings. Each finding has:
- File path (relative to project root)
- Line number (if present in the output)
- Issue description — title, category, severity, or equivalent
- Fix instructions — the tool's suggested fix, remediation text, or "Prompt for AI Agent" block if present. If none, use the issue description.
Build a list: findings = [{file, line, description, fix_instructions},...]
Display a count:
Findings: <total> across <file_count> filesStep 5 — Zero findings check
If findings is empty:
- Print:
Zero findings. Codebase is clean. - Exit loop — skip to Step 10.
Step 6 — Group by file
Group findings by file path. All findings in the same file go to one subagent to avoid conflicting edits.
Step 7 — Fix with subagents (parallel)
For each file group, dispatch one subagent in parallel with this prompt:
Fix the following findings in `<file_path>`.
Current file content:
<content of the file>
Findings to fix:
<for each finding in the group>
- Line <line>: <description>
Fix instructions: <fix_instructions>
</for each>
Requirements:
- Edit only what is needed to fix the listed findings.
- Do not refactor unrelated code.
- Preserve all existing tests.
- After editing, confirm what was changed.Claude Code
Use the Agent tool with subagent_type: "general-purpose" for each file group. Launch all agents in a single message to run them in parallel. Wait for all to complete before proceeding.
Gemini CLI
Use the generalist_agent tool for each file group. Dispatch all calls to run in parallel. Wait for all to complete before proceeding.
Other agents
If parallel subagent dispatch is not available, fix each file group sequentially — read the file, apply the fixes directly, then move to the next file group.
Step 8 — Verify
Discover and run the project's lint and test commands.
Discovery order
Check these sources in order. Use the first lint command and first test command found:
1. Project files:
| File | Lint command | Test command |
|---|---|---|
package.json | scripts.lint (run via npm run lint) | scripts.test (run via npm test) |
pyproject.toml | Look for [tool.ruff] → ruff check. | Look for [tool.pytest] → pytest --tb=short -q |
Makefile | make lint (if target exists) | make test (if target exists) |
Justfile | just lint (if recipe exists) | just test (if recipe exists) |
2. CLAUDE.md or project docs:
Look for a "Commands" section listing lint/test commands.
3. Ask the user:
If no lint or test commands can be detected, ask: "What commands do you use to lint and test this project?"
Run verification
Run the lint command first, then the test command:
<lint-command>
<test-command>If either fails:
- Read the error output.
- Fix the regressions directly (do not spawn subagents for verification fixes).
- Re-run the failing command until it passes before continuing.
Step 9 — Loop
Go back to Step 3.
Maximum iterations: if the loop has completed 5 iterations, exit regardless of findings and proceed to Step 10.
Stuck detection: track the set of findings each iteration by (file + description). If the same findings appear unchanged for 2 consecutive loops:
- Print:
Stuck after <n> iterations. The following findings require manual review: - List each stuck finding with file, line, and description.
- Exit loop.
Step 10 — Report
After the loop exits (clean or stuck), print a summary:
+----------------------------------+
| Review Fix Loop — Summary |
+----------------+-----------------+
| Iterations | <n> |
| Fixed | <count> |
| Remaining | <count> |
+----------------+-----------------+If remaining > 0, list each stuck finding below the table with its file path, line number, and description.