fix(engine): surface action errors to LLM with [ACTION FAILED] prefix

[serrrfirat]

Summary

• Prefix failed action results with [ACTION FAILED] <tool>: in the Python orchestrator so the LLM can't ignore tool failures and falsely claim success
• Add TODO(#2325) for consecutive action error counting, deferred pending unified progress-tracking design

Context

When tools fail (e.g., "No lease for action"), the orchestrator appended the raw error JSON as a generic ActionResult message — indistinguishable from success. The LLM frequently ignored these errors and told the user "Successfully sent!" despite every tool call failing.

The Rust executor already sets is_error: true on lease and policy failures. This change reads that flag and prefixes the message content so the error is unmissable.

Error counting (failing the thread after repeated action errors) is tracked in #2325 and depends on the progress-tracking design discussed in #2240.

Closes #2279

Test plan

• cargo test -p ironclaw_engine — 281 tests pass
• Manual: trigger a "No lease for action" error on staging and verify the bot reports failure instead of claiming success

🤖 Generated with Claude Code

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[henrypark133]

Review: Surface action errors to LLM with [ACTION FAILED] prefix (Risk: Low)

Clean fix. When is_error is true on an ActionResult, the output is now prefixed with [ACTION FAILED] action_name: so the LLM can clearly distinguish failures from successful outputs in its next turn.

Positives:

• Python orchestrator change is minimal (2 lines) and targeted
• Rust loop_engine test failed_action_result_is_explicit_in_next_llm_message verifies the prefix appears in the message content
• TODO comment for consecutive error tracking is well-scoped with issue reference (#2325)

LGTM.

[henrypark133]

Review: Surface action errors to LLM with [ACTION FAILED] prefix

Clean fix. The Python-side prefixing change is minimal, and the new loop_engine test verifies the next LLM turn now sees an explicit failure marker including the action name.

Residual risk is low: I only spot-checked the engine path and ran cargo test -p ironclaw_engine failed_action_result_is_explicit_in_next_llm_message locally.