Scan assistant messages for decision language at Stop time

[baljanak]

Summary

• The plugin detected decision language in user messages (capture_nudge on UserPromptSubmit) but missed decisions stated by the assistant — there's no AssistantMessage hook event in Claude Code
• At Stop time, reads the session's JSONL transcript from disk and scans assistant text blocks for the same decision phrase patterns with identical corroboration requirements
• Integrates as the highest-priority secondary hint in stop_nudge, before impl/plan/staleness checks

Changes

• helpers.py: Add _discover_transcript() — locates ~/.claude/projects/{key}/{session-id}.jsonl
• constants.py: Add TRANSCRIPT_TAIL_BYTES (64KB) and TRANSCRIPT_MAX_BLOCKS (20)
• stop_nudge.py: Add _scan_assistant_decisions() (tail-read + pattern matching) and _assistant_decision_summary(), wire into _stop_nudge_condition() priority chain
• tests/test_stop_nudge.py: 8 new tests — detection, corroboration, dedup with capture_nudge, missing/corrupt transcript, dismissed nudges, end-to-end

Design choices

• Tail-read last 64KB — bounds scan time for large transcripts while capturing recent assistant turns
• Same corroboration bar as capture_nudge (phrase + tech signal OR reasoning signal OR 2+ phrases)
• Deduplicates against phrases already detected by capture_nudge from user input
• Extends stop_nudge rather than adding a new policy — avoids competing for the 2/session nudge budget

Test plan

• All 589 tests pass (make check)
• ruff, mypy, shellcheck clean
• Manual: run a session, have assistant state a decision, end session, verify nudge appears

🤖 Generated with Claude Code