feat: add auto-fix service — auto-lint and test after AI file edits

[gnanam1990]

Summary

Adds a built-in auto-fix service that automatically runs user-configured lint and test commands after AI file edits (FileEdit, FileWrite). When errors are detected, they are fed back to the AI as additional context so it can self-repair — closing the manual feedback loop that currently requires users to run lint/test, copy errors, and paste them back.

Inspired by Aider's auto-lint-and-test loop, which is their #1 user satisfaction driver.

What's included

• Config schema (src/services/autoFix/autoFixConfig.ts) — Zod-validated settings with enabled, lint, test, maxRetries (0–10, default 3), and timeout (1–300s, default 30s). Rejects enabled configs with no lint or test command.

• Runner (src/services/autoFix/autoFixRunner.ts) — Executes lint/test commands via child_process.spawn, captures stdout/stderr, enforces timeouts, skips tests if lint fails first, and builds an AI-readable error summary.

• Hook integration (src/services/autoFix/autoFixHook.ts) — shouldRunAutoFix() gates execution to file_edit/file_write tools only. buildAutoFixContext() wraps errors in <auto_fix_feedback> tags that the AI can act on.

• PostToolUse wiring (src/services/tools/toolHooks.ts) — After the existing PostToolUse hook loop, runs the auto-fix check and yields errors as hook_additional_context attachment messages.

• /auto-fix slash command (src/commands/auto-fix.ts) — Guided setup wizard that helps users configure their lint/test commands in settings.

• Settings schema (src/utils/settings/types.ts) — Added optional autoFix field to SettingsSchema.

Configuration

Add to .claude/settings.json or .claude/settings.local.json:

{
  "autoFix": {
    "enabled": true,
    "lint": "eslint . --fix",
    "test": "bun test",
    "maxRetries": 3,
    "timeout": 30000
  }
}

Or run /auto-fix for guided setup.

How it works

AI edits file → lint runs → errors? → feed back to AI → AI fixes → re-lint
                         ↓ no errors
                    tests run → failures? → feed back to AI → AI fixes → re-test
                             ↓ no failures
                        clean result delivered

Test coverage

• 29 tests across 4 test files (config, runner, hook, integration)
• 541/541 full suite tests pass
• Zero new type errors
• Build succeeds

Files changed

File	Change
src/services/autoFix/autoFixConfig.ts	New — config schema and reader
src/services/autoFix/autoFixConfig.test.ts	New — 11 tests
src/services/autoFix/autoFixRunner.ts	New — command execution engine
src/services/autoFix/autoFixRunner.test.ts	New — 8 tests
src/services/autoFix/autoFixHook.ts	New — hook helpers
src/services/autoFix/autoFixHook.test.ts	New — 8 tests
src/services/autoFix/autoFixIntegration.test.ts	New — 2 e2e tests
src/commands/auto-fix.ts	New — slash command
src/commands.ts	Register /auto-fix command
src/services/tools/toolHooks.ts	Wire auto-fix into PostToolUse
src/utils/settings/types.ts	Add autoFix to settings schema

[kevincodex1]

this looks great a much needed feature indeed! thanks @gnanam1990

[FluxLuFFy]

@auriti @Vasanthdev2004 @anandh8x see this

[Vasanthdev2004]

1. maxRetries is exposed as part of the new user-facing config (src/services/autoFix/autoFixConfig.ts), but the feature never reads or enforces it in the actual auto-fix path (src/services/tools/toolHooks.ts, src/services/autoFix/autoFixRunner.ts). In practice that means the new self-repair loop has no feature-specific retry cap: every failing file_edit or file_write can emit another hook_additional_context and keep the model cycling until some unrelated outer limit stops it. If this setting is documented and shown in the example config, the loop needs to honor it or the docs/config should be removed.

2. Cancellation is threaded into the API but not into the subprocess. toolHooks.ts passes toolUseContext.abortController.signal into runAutoFixCheck(), and runAutoFixCheck() passes it down to runCommand(), but runCommand() never subscribes to the signal or kills the child on abort. The practical failure mode is a long or mutating command like eslint --fix, bun test --watch, etc.: if the user interrupts the turn, the auto-fix process keeps running until timeout or natural exit, still consuming resources and potentially still editing files after the user thinks execution stopped.

3. runCommand() hardcodes spawn('bash', ['-c', command], ...) in src/services/autoFix/autoFixRunner.ts. That bypasses the existing shell abstraction (src/utils/Shell.ts, src/utils/hooks.ts) and will fail on Windows setups where the app works via a resolved shell path or Git Bash helper but plain bash is not on PATH. Since this feature is exposed as a general cross-platform config option, it should execute commands through the same shell/provider resolution the rest of the app already uses rather than assuming a bash executable exists.

[gnanam1990]

@Vasanthdev2004 Addressed all three items in ff7eccc:

1. maxRetries now enforced — Added per-query-chain retry tracking in toolHooks.ts. Each time auto-fix feeds errors back to the AI, the counter increments. Once maxRetries is reached, auto-fix stops and emits a message telling the AI to let the user review manually. Counter resets when lint/test passes clean.

2. Abort signal wired to subprocess — runCommand() now subscribes to AbortController.signal and kills the process on abort. Uses detached: true with process.kill(-pid) on Unix to terminate the entire process group (so child processes like eslint --fix or bun test are also killed, not just the shell). Early-exits if signal is already aborted before spawning.

3. Cross-platform shell resolution — Replaced spawn('bash', ['-c', command]) with spawn(command, [], { shell: true, windowsHide: true }). Node's shell: true uses /bin/sh on Unix and cmd.exe on Windows — the standard cross-platform approach. Also added detached: !isWindows for proper process group management on each platform.

All 29 auto-fix tests pass, build succeeds, zero new type errors.

[Vasanthdev2004]

Rechecked the current head ff7eccc36c6997b94de769b4ce14f2c154c4b255.

Scope

This is a targeted current-head re-review of the auto-fix feature after the earlier blocker-level feedback.

Verdict

Approve-ready

What I rechecked on the current head:

• maxRetries is now actually enforced in the auto-fix loop in src/services/tools/toolHooks.ts, instead of only existing in the user-facing config schema
• cancellation is now wired through to the spawned subprocess path in src/services/autoFix/autoFixRunner.ts, including process-tree termination behavior on Unix
• the hardcoded bash -c path was replaced with a cross-platform shell: true spawn flow, which removes the earlier bash-on-PATH assumption
• current test surface for config parsing, hook behavior, integration, and runner coverage is in place
• current checks are green

This looks like a coherent feature now, and the earlier blocker-level concerns appear addressed on the current head.

[FluxLuFFy]

@kevincodex1 @auriti @anandh8x take a look and I will be helping community as much as possible as I can to improve it

[kevincodex1]

looks great to me!

[kevincodex1]

@FluxLuFFy looking forward to it. thank you