hotfix: notification visibility (Reports tab filter) + LLM skill-hallucination retry

[AVADSA25]

Summary

Two related issues from your first real anchor-example run.

Issue 1 — notifications counted but invisible

"I keep getting notifications but I don't see what it is — doesn't show in history or report."

Root cause: codec_tasks.html:913 was filtering the Reports tab to ONLY n.type === 'task_report'. All Phase 1+2+3+3.5 notification types got silently dropped.

Fix: introduce REPORT_NOTIF_TYPES whitelist with all 8 CODEC types:

• task_report (existing crews)
• question (Phase 1 Step 3 ask_user)
• shift_report (Phase 2 Step 7)
• agent_update, agent_blocked, agent_question, agent_done, agent_aborted (Phase 3 Step 10)
• proactive_suggestion (Phase 3.5)

Reports tab now shows everything. Sorted newest-first.

Issue 2 — LLM skill-hallucination → unrecoverable block

Your forex-briefing agent (agent_07b2b6eeab02) failed at checkpoint 1 because Qwen-3.6 chose to call skill fetch_url (doesn't exist) instead of web_fetch (which IS in the manifest). The permission gate correctly blocked — but there was no recovery, so the agent stalled in blocked_on_permission waiting for a useless /grant for a non-existent skill.

Fix in codec_agent_runner._execute_checkpoint:

• On PermissionViolation(reason="skill_not_authorized"), append a correction nudge to history (lists allowed skills) and re-call _qwen_next_action ONCE
• If the second action passes the gate → execution continues
• If the second action also violates → falls through to the original block path
• Path/domain violations still block immediately (those are real scoped gaps the user has to grant)

Naming drift now self-corrects in one retry. Most LLM hallucinations are surface naming — this gives Qwen a chance to course-correct without needing user intervention.

Tests

+1 test: test_skill_hallucination_retries_with_corrected_skill_list

• Fakes Qwen returning fetch_url (hallucinated) → weather (real) → checkpoint_done
• Asserts only weather actually ran via _run_skill
• Asserts correction nudge appears in history
• No exception raised

test_agent_runner.py: 42 → 43. No regressions.

What to do after merge

cd ~/codec-repo
git pull
pm2 restart codec-dashboard codec-agent-runner

Then:

1. Abort the stalled agent (in /chat Project mode → click [abort] on the status pill, or curl -X POST http://localhost:8090/api/agents/agent_07b2b6eeab02/abort)
2. Re-drop the forex briefing prompt — should run further this time (will retry on hallucinations)
3. Click the bell at top → navigate to /tasks#reports → you'll now see all the agent_update / agent_blocked / agent_done events from this run + previous

🤖 Generated with Claude Code