Add multi-backend fixer support (Claude / Codex / Cursor)

[Senna46]

Summary

• Abstract the fix backend behind a BugFixer interface so Fixooly can dispatch to Claude CLI, OpenAI Codex CLI, or Cursor CLI depending on the new required AUTOFIX_FIXER environment variable.
• Cursor is the recommended default for daemon use because the Cursor Pro Auto+Composer pool has no 5-hour rolling rate limit, unlike Codex on ChatGPT Plus/Pro. The plan in .plans/multi-backend-fixer.plan.md walks through the comparison.
• Centralise child-process plumbing (timeout, stdin, bounded stdout) in src/fixers/spawnRunner.ts and the COMMIT_MSG: / FIX_DETAIL: marker parsing in src/fixers/outputParser.ts. The parser now also understands Cursor's JSON output and Codex's JSONL event stream.
• FixGenerator keeps the cloning / commit / push responsibilities and receives the active BugFixer via constructor injection. main.ts instantiates the right fixer through createFixer() and delegates verifyPrerequisites() to the backend itself.

Files

• New: src/fixers/{types,factory,claudeFixer,codexFixer,cursorFixer,spawnRunner,outputParser}.ts
• Modified: src/types.ts, src/config.ts, src/fixGenerator.ts, src/main.ts
• Docs: README.md, .env.example, CLAUDE.md, AGENTS.md

Test plan

• npm run typecheck -- zero errors
• npm run build -- dist/ produced without errors
• Smoke test with AUTOFIX_FIXER=cursor against a real PR (deferred until the new Cursor Pro account is provisioned)
• Smoke test with AUTOFIX_FIXER=codex against a real PR (deferred until ChatGPT Plus is provisioned)
• Regression test with AUTOFIX_FIXER=claude against the existing setup (manual run on the daemon host)

Out of scope

• Live runtime testing of the Cursor / Codex backends (will be validated by the user after subscriptions are provisioned).
• Automatic fallback between backends when one is rate-limited.

Closes #11

---

Note

Medium Risk
Breaking config (mandatory AUTOFIX_FIXER) and new untested CLI integrations change how fixes are applied and committed to PR branches; misconfiguration or parser/CLI mismatches could fail silently or produce bad commits.

Overview
This PR replaces the hard-coded Claude claude -p path with a pluggable BugFixer layer so the daemon can run fixes through Cursor, Codex, or Claude CLIs.

Configuration: AUTOFIX_FIXER is now required (no default)—claude / codex / cursor—with optional per-backend model overrides (AUTOFIX_CURSOR_MODEL, AUTOFIX_CODEX_MODEL, AUTOFIX_CLAUDE_MODEL) and documented API keys in .env.example. Existing deployments must set AUTOFIX_FIXER explicitly (e.g. claude to keep prior behavior).

Code: New src/fixers/ modules (types, factory, three fixer implementations, spawnRunner, outputParser) own CLI invocation, timeouts, and parsing of COMMIT_MSG: / FIX_DETAIL: across plain text, JSON, and JSONL. FixGenerator only clones, builds the prompt, calls fixer.generateFix, then commits/pushes; main.ts wires createFixer(config) and delegates prerequisite checks to the selected backend.

Docs: README, AGENTS.md, CLAUDE.md, and a plan file describe backend selection (Cursor recommended for daemons) and the updated architecture.

^{Reviewed by Cursor Bugbot for commit 6729e01. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ff70199. Configure here.}

[senna-fixooly]

Fixooly committed fixes for 1 bug(s). (6729e01cb4)

• ✅ Fixed: Codex JSONL message not extracted
  • Added a branch in extractMessageFromJson that unwraps obj.item and returns itemObj.text when itemObj.type is "agent_message", so the current codex exec --json item.completed events yield the assistant reply for COMMIT_MSG/FIX_DETAIL marker parsing