Port reflection-3 to Claude Code as Stop-hook plugin

[dzianisv]

Context

Parent: #131 (portability verdict — Claude Code adapter declared viable).

In long CC sessions, the agent forgets project-level skill guidance (e.g. /take-ownership) and back-delegates to the user prematurely. reflection-3.ts already solves the equivalent problem in OpenCode via session.idle hook. This issue tracks porting that pattern to Claude Code.

Goal

A CC plugin that, on the Stop event:

1. Grades whether the assistant actually finished the task (re-uses reflection-3 evaluator).
2. If incomplete, injects a follow-up prompt via decision: "block" + additionalContext to keep the agent driving.
3. Honors stop_hook_active to prevent infinite loops.
4. Detects planning/action loops and back-delegation to the user as failure modes.

Success Metric

Manual smoke: open a fresh CC session in this repo, give a multi-step coding task, observe at least one Stop→inject→continue cycle where the plugin successfully re-prompts the agent. Plugin must NOT loop more than 3 times on same failure mode, and must not fire when stop_hook_active=true.

Scope

• New: claude/ directory containing the plugin (manifest + hook + bin script)
• Reuse: import core grading/loop-detection from a shared module extracted from reflection-3.ts (or, speed path, copy + adapt)
• Out of scope: Codex adapter (Assess portability of OpenCode plugins to Codex and Claude Code #131 says App Server sidecar — separate issue)

Acceptance Criteria

• .claude-plugin/plugin.json valid, installable via claude --plugin-dir
• Stop hook injects follow-up prompt for incomplete tasks
• stop_hook_active loop guard implemented
• Real-feature smoke test passes (documented in .tasks/.../test-report.md)
• Atomic commits per phase

[dzianisv]

Shipped in #139 as 28d1aa0.

Post-merge prod verify: PASS

Plugin installed globally on dev machine via ~/.claude/settings.json (settings-based path is authoritative for CC v2.1.150 — --plugin-dir does not enable Stop hooks in -p mode; see follow-up note in #139).

Real CC session (claude -p headless, fresh /tmp dir) → Stop hook fired → classifier ran (Haiku 4.5 via OAuth) → category=complete confidence=0.99 → no inject (correct, trivial Q&A). Hook latency ~870ms.

Evidence: .tasks/131-cc-reflection/smoke-evidence/prod-verify/

Follow-ups

• Refine cc-stop classifier prompt — over-assigns working, under-assigns tool_available_punt #138 — refine classifier prompt (working over-assigned 374/907, tool_available_punt under-assigned 0/907).
• New: provoke drift with Sonnet 4.6 in E2E for stronger inject-cycle proof (Haiku resists drift on small synthetic prompts).
• Investigate --plugin-dir not enabling Stop hooks in -p mode (likely upstream CC behavior).
• Codex adapter (App Server sidecar — per Assess portability of OpenCode plugins to Codex and Claude Code #131 verdict).