Fix workflow stuck in refining/evaluation loops

[neuromechanist]

Summary

• Make evaluation informational-only when run_assessment=False (default), preventing the evaluation-driven refinement loop that caused the "stuck" behavior
• Add 15s LLM call timeout via request_timeout on ChatLiteLLM to prevent hanging on slow providers
• Default evaluation parsing to ACCEPT when ambiguous instead of REFINE
• Derive max_total_iterations from max_validation_attempts + 1 (was hardcoded 10)
• Add per-node timing (time.monotonic()) to all workflow nodes for diagnosing slowness
• Switch default evaluation model to openai/gpt-oss-120b on groq
• Lower LangGraph recursion_limit from 100 to 50
• Update default max_validation_attempts from 5 to 3

Behavior Change

Setting	Before	After
run_assessment=False	Evaluation could loop back to refinement	Evaluation is informational only, always ends
run_assessment=True	Same as above	Evaluation can trigger refinement (capped)
Default max iterations	10 (hardcoded)	max_validation_attempts + 1 (4 by default)
LLM call timeout	None (infinite)	15 seconds
Eval model	qwen/qwen3-235b on Cerebras	openai/gpt-oss-120b on groq

All settings remain tunable via the frontend (Max Validation Attempts dropdown, Run Assessment checkbox).

Test plan

• 415 unit tests pass, 0 failures
• Manual test: simple description completes in 1-2 iterations
• Manual test: run_assessment=False never triggers refinement from evaluation
• Manual test: run_assessment=True allows refinement but caps at max_validation_attempts + 1
• Verify per-node timing appears in server logs

Closes #119

[cloudflare-workers-and-pages]

Deploying hedit with    Cloudflare Pages

Latest commit:	6cf0986
Status:	✅  Deploy successful!
Preview URL:	https://228724b8.hedit.pages.dev
Branch Preview URL:	https://119-fix-workflow-stuck-in-re.hedit.pages.dev

View logs

[codecov]

Codecov Report

❌ Patch coverage is 26.53061% with 72 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/agents/workflow.py	13.33%	25 Missing and 1 partial ⚠️
src/api/main.py	31.57%	26 Missing ⚠️
src/agents/evaluation_agent.py	23.07%	10 Missing ⚠️
src/agents/assessment_agent.py	28.57%	5 Missing ⚠️
src/agents/feedback_summarizer.py	28.57%	5 Missing ⚠️

📢 Thoughts on this report? Let us know!

[neuromechanist]

PR Review Summary

Three specialized review agents analyzed this PR. Here is a summary of all findings and how each was addressed.

Silent Failure Hunter

CRITICAL - No timeout error handling in LLM calls (evaluation, assessment, feedback agents)

• Fixed in commit 6cf0986: Added try/except with logging around llm.ainvoke() calls in evaluation_agent.py, assessment_agent.py, and feedback_summarizer.py. Errors are logged with full traceback and re-raised.

HIGH - API endpoints return generic 500 for all errors

• Fixed in commit 6cf0986: Added specific exception handlers for APITimeoutError (maps to HTTP 504) and RateLimitError (maps to HTTP 429) in all four annotation endpoints. Streaming SSE error events now include an error_type field (timeout, rate_limit, internal). Raw exception strings no longer leak to clients.

MEDIUM - Silent ACCEPT fallback in evaluation parsing

• Fixed in commit 6cf0986: Added logger.debug() when _parse_decision() falls through to the default ACCEPT, so ambiguous responses are traceable in logs.

Code Reviewer

Stale defaults in create_initial_state() (max_validation_attempts=5, max_total_iterations=10)

• Fixed in commit c89dcac: Updated to max_validation_attempts=3, max_total_iterations=None (auto-derived as max_validation_attempts + 1).

import re inside method body instead of module-level

• Fixed in commit c89dcac: Moved to module-level import.

Missing "Entering assess node" log

• Fixed in commit c89dcac: Added print statement consistent with other nodes.

Test assertion using old default (5 instead of 3)

• Fixed in commit c89dcac: Updated test_create_initial_state assertion.

Code Simplifier

max_total_iterations derivation duplicated 3x in main.py

• Fixed in commit c89dcac: Centralized derivation in create_initial_state() and workflow.run() with None default and auto-derivation. All call sites in main.py simplified.

Pre-existing Issues (not introduced by this PR)

• test_cli_integration.py::test_annotate_complex_description fails on develop (server error, unrelated to this PR)
• Dependabot vulnerability on default branch (pre-existing)