chore: promote staging to staging-promote/66ccafb9-24321596712 (2026-04-13 03:02 UTC)

[ironclaw-ci]

Auto-promotion from staging CI

Batch range: a53eac5c2dec6b6cd5c08189086093fde64aa9cb..4529f009d4e239e06590f083662ac8804bcfaf22
Promotion branch: staging-promote/4529f009-24323637408
Base: staging-promote/66ccafb9-24321596712
Triggered by: Staging CI batch at 2026-04-13 03:02 UTC

Commits in this batch (13):

• a7401ec fix(gateway): scope chat approvals to the active thread (fix(gateway): scope chat approvals to the active thread #2267)
• 4032f6d Fix paired Telegram owner scope routine visibility (Fix paired Telegram owner scope routine visibility #2258)
• 7d66d83 fix(engine): always append ActionResult for every tool call (fix(engine): always append ActionResult for every tool call #2322)
• 764e586 feat(engine): LLM council via per-call model override in CodeAct (feat(engine): LLM council via per-call model override in CodeAct #2320)
• 70862ed feat(config): default CLI_MODE to TUI instead of REPL (feat(config): default CLI_MODE to TUI #2329)
• 207c4d4 ci: build docker image in release process (ci: build docker image in release process #2321)
• cd9b60c fix: re-apply Telegram UTF-16 splitting and DB MIGRATION label (fix: re-apply Telegram UTF-16 splitting and DB MIGRATION label #2304)
• 88b87c0 feat: user-facing temperature setting (feat: user-facing temperature setting #2275)
• fdb0a13 chore: sync staging and main (chore: sync staging and main #2337)
• 3cb77fe fix: resolve cargo-deny failures (wildcard deps + rand advisory) (fix: resolve cargo-deny failures (wildcard deps + rand advisory) #2370)
• ed2d6dc fix web chat refresh active thread ([codex] Fix web chat refresh active thread #2330)
• 66ccafb chore(engine): update monty to v0.0.11 (chore(engine): update monty to v0.0.11 #2364)
• 4529f00 fix(engine): track consecutive action errors in orchestrator Tier 0 path (Orchestrator: add action error counting to Python execution loop #2325) (fix(engine): track consecutive action errors in orchestrator (#2325) #2340)

Current commits in this promotion (1)

Current base: staging-promote/66ccafb9-24321596712
Current head: staging-promote/4529f009-24323637408
Current range: origin/staging-promote/66ccafb9-24321596712..origin/staging-promote/4529f009-24323637408

• 4529f00 fix(engine): track consecutive action errors in orchestrator Tier 0 path (Orchestrator: add action error counting to Python execution loop #2325) (fix(engine): track consecutive action errors in orchestrator (#2325) #2340)

Auto-updated by staging promotion metadata workflow

Waiting for gates:

• Tests: pending
• E2E: pending
• Claude Code review: pending (will post comments on this PR)

---

Auto-created by staging-ci workflow

[claude]

Code review

Found 12 issues across security, architecture, bugs, and performance:

CRITICAL (1)

1. [CRITICAL:88] Race condition: consecutive_action_errors counter not atomic — If orchestrator ever runs multiple concurrent Monty VMs with shared state (e.g., parallel sub-agents loading state from DB), the consecutive_action_errors counter will have lost-update bugs. Checkpoint persistence makes inconsistencies durable. Current single-VM-per-thread architecture is safe, but this becomes a data corruption hazard if concurrency is added.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/orchestrator/default.py#L473-L474

---

HIGH (3)

2. [HIGH:92] String concatenation for error messages instead of templates — Error messages at lines 874 and 880-885 use inline string concatenation instead of template files. CLAUDE.md requires multi-line prompt strings to live in crates/ironclaw_engine/prompts/*.md and be loaded via include_str!(). This violates the stated architecture and makes localization/A/B testing difficult.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/orchestrator/default.py#L874-L885

3. [HIGH:88] Magic number fragmentation: max_consecutive_errors + 2 threshold lacks documentation — Failure at >= max_consecutive_errors + 2 vs nudge at >= max_consecutive_errors introduces unexplained logic. The + 2 offset is unmotivated, unparameterized, and inconsistent with Tier 1 error handling. This diverges from design goal of "Consistency with Rust loop" (issue Orchestrator: add action error counting to Python execution loop #2325).

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/orchestrator/default.py#L869-L876

4. [HIGH:78] Missing timeout on Python VM execution — run_python_final() (refactored test helper) has memory limits (500_000 allocations) but no wall-clock timeout. If orchestrator Python code ever includes user-influenced loops, it can hang indefinitely. Current code is safe only because orchestrator has no dynamic loops.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/src/executor/orchestrator.rs#L2481-L2483

---

MEDIUM (7)

5. [MEDIUM:85] Unbounded string concatenation in hot path — Nudge message constructed via string concatenation at every failed batch crossing threshold. Repeated append_message() calls accumulate context weight unnecessarily.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/orchestrator/default.py#L880-L885

6. [MEDIUM:78] Test extraction refactoring introduces untyped Monty bridge — run_python_final() loses type safety by returning MontyObject and forcing callers to pattern-match with panics. Better approach: generic eval_python_value<T: From<MontyObject>> with a proper conversion trait.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/src/executor/orchestrator.rs#L2475-L2516

7. [MEDIUM:75] Test refactoring introduces code duplication — eval_python_bool() and eval_python_int() both duplicate the helpers extraction logic (searching for \ndef run_loop(). Should extract extract_helpers() function.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/src/executor/orchestrator.rs#L2518-L2525

8. [MEDIUM:72] DRY violation: checkpoint persistence repeated 7 times — __save_checkpoint__() dict is copy-pasted at lines 615, 635, 671, 769, 790, 807, 890. Future counter additions require hunting down all sites. Should extract _save_action_error_checkpoint() helper.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/orchestrator/default.py#L612-L617

9. [MEDIUM:72] Repeated JSON parsing in checkpoint serialization — monty_to_json() called twice per checkpoint; then .as_u64() chained 4 times. Low practical impact (checkpoints are infrequent), but indicates opportunity for optimization.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/src/executor/orchestrator.rs#L1741-L1744

10. [MEDIUM:70] Batch error/success counting assumes parallel results alignment — No assertion verifies len(results) == len(executable_calls). If Rust batch handler ever reorders results (e.g., by completion time), errors will pair with wrong calls silently.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/orchestrator/default.py#L734-L757

11. [MEDIUM:68] Partial batch success logic lacks explicit documentation — Reset on ANY success, increment only on ALL-fail design is sound but not explained. Code comment should justify why.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/orchestrator/default.py#L864-L867

---

LOW (1)

12. [LOW:52] Docstring synchronization debt — Old docstring at line 2470 describes bool-only behavior; new one describes generic behavior. Orphaned first line should be removed.

https://github.com/anthropics/ironclaw/blob/3f727c41afda86b7a1c74f354cefb27d9d192dce/crates/ironclaw_engine/src/executor/orchestrator.rs#L2470-L2474

---

Recommendations (by priority)

IMMEDIATE (production safety):

• Add timeout wrapper to run_python_final() to prevent infinite-loop hangs
• Add assertion for len(results) == len(executable_calls) to catch result alignment bugs

BEFORE MERGE (architectural consistency):

• Move error messages to crates/ironclaw_engine/prompts/error_messages.md
• Document or parameterize the + 2 threshold offset
• Extract _save_action_error_checkpoint() helper to reduce DRY violations

NICE-TO-HAVE (cleanup):

• Fix docstring at line 2470
• Extract extract_helpers() to reduce test code duplication
• Add generic eval_python_value<T> test helper for type safety

---

Overall: Feature is sound and tests are comprehensive. These are maintainability and clarity improvements, not correctness blockers. Most LOW/MEDIUM issues are maintenance debt rather than production bugs.