## Summary
Investigated OpenAI's new GPT-5-Codex model and Codex GitHub PR review integration capabilities.
## GPT-5-Codex Analysis
### Benchmark Performance (Good)
- SWE-bench Verified: 74.5% (vs GPT-5's 72.8%)
- Refactoring tasks: 51.3% (vs GPT-5's 33.9%)
- Code review: Higher developer ratings
### Real-World Issues (Concerning)
- Users report degraded coding performance
- Scripts that previously worked now fail
- Less consistent than GPT-4.5
- Longer response times (minutes vs instant)
- "Creatively and emotionally flat"
- Basic errors (e.g., counting letters incorrectly)
### Key Finding
Classic case of "optimizing for benchmarks vs real usability" - scores well on tests but performs poorly in practice.
## Codex GitHub PR Integration
### Setup Process
1. Enable MFA and connect GitHub account
2. Authorize Codex GitHub app for repos
3. Enable "Code review" in repository settings
### Usage Methods
- **Manual**: Comment '@codex review' in PR
- **Automatic**: Triggers when PR moves from draft to ready
### Current Limitations
- One-way communication (doesn't respond to review comments)
- Prefers creating new PRs over updating existing ones
- Better for single-pass reviews than iterative feedback
## 'codex resume' Feature
New session management capability:
- Resume previous codex exec sessions
- Useful for continuing long tasks across days
- Maintains context from interrupted work
🐱 The investigation reveals that while GPT-5-Codex shows benchmark improvements, practical developer experience has declined - a reminder that metrics don't always reflect real-world utility\!
Changes to resolver.py:
- Improved PHI value tracking in _value_at_end_i64() (lines 268-285)
- Added trace logging for snap hits with PHI detection
- Fixed PHI placeholder reuse logic to preserve dominance
- PHI values now returned directly from snapshots when valid
Changes to llvm_builder.py:
- Fixed externcall instruction parsing (line 522: 'func' instead of 'name')
- Improved block snapshot tracing (line 439)
- Added PHI incoming metadata tracking (lines 316-376)
- Enhanced definition tracking for lifetime hints
This should help debug the string carry=0 issue in esc_dirname_smoke where
PHI values were being incorrectly coerced instead of preserved.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
✅ Print and FileBox paths now working correctly
✅ Resolver simplified by removing overly aggressive fast-path optimization
✅ Both OFF/ON in compare_harness_on_off.sh now use Python version
✅ String handle propagation issues resolved
Key changes:
- Removed instruction reordering in llvm_builder.py (respecting MIR order)
- Resolver now more conservative but reliable
- compare_harness_on_off.sh updated to use Python backend for both paths
This marks a major milestone towards Phase 15 self-hosting with Python/llvmlite!
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Resolver-only reads across BBs; remove vmap fallbacks
- Create PHIs at block start; insert casts in preds before terminators
- Re-materialize int in preds to satisfy dominance (add/zext/trunc)
- Use constant GEP for method strings to avoid order dependency
- Order non-PHI lowering to preserve producer→consumer dominance
- Update docs: RESOLVER_API.md, LLVM_HARNESS.md
- compare_harness_on_off: ON/OFF exits match; linking green
- Created llvmlite-based LLVM backend in src/llvm_py/
- Implemented all MIR14 instructions (const, binop, jump, branch, ret, compare, phi, call, boxcall, externcall, typeop, newbox, safepoint, barrier)
- Experimental LoopForm support
- ~2000 lines of clean Python code vs complex Rust/inkwell
- Useful for PHI/SSA validation and rapid prototyping
- Added documentation to CLAUDE.md
This was created while waiting for ChatGPT's investigation of BuilderCursor issues.
