Claude Code reflection plugin + 907-stop classification baseline

.gitignore

@@ -3,6 +3,8 @@
 !.agents/**
 !.github
 !.github/**
+!claude/.claude-plugin
+!claude/.claude-plugin/**
 .tts
 .reflection
 .opencode/
@@ -28,9 +30,18 @@ test/mocks/
 evals/results/
 evals/evals/
 
+# Raw CC stop datasets (contain real user session data — never commit)
+evals/datasets/cc-stop-candidates-raw.jsonl
+evals/datasets/cc-stop-candidates-filtered.jsonl
+evals/datasets/cc-stop-classified.jsonl
+# Allow committing the redacted gold set only
+!evals/datasets/cc-stop-labeled-gold-redacted.jsonl
+
 # E2E eval reports
 .eval/
 .eval-tmp/
 
 # Working notes
 plan.md
+# Added by code-review-graph
+.code-review-graph/

README.md

@@ -4,6 +4,9 @@
 ## @reflection-3.ts - push opencode agent to reflect on the task, useful for continuous interrupted runs
 <img width="1472" height="972" alt="image" src="https://github.com/user-attachments/assets/40f3a752-4b84-4151-93f4-2614330ac653" />
 
+## @claude/ - Claude Code reflection plugin (experimental)
+Port of the reflection idea to Claude Code as a `Stop` hook. Classifies the last assistant turn into one of six categories (complete, working, waiting_for_user_legitimate, tool_available_punt, summary_drift_stop, genuinely_stuck) with Claude Haiku 4.5, and re-prompts the agent when it punted to the user, drifted into "summary + next step + stop", or halted mid-thought. Honors `stop_hook_active` loop guard and caps at 3 inject cycles per session. Install with `claude --plugin-dir ./claude` for dev or via the marketplace once published. See [`claude/README.md`](claude/README.md). Baseline classifier accuracy and dataset are tracked in [`evals/datasets/README.md`](evals/datasets/README.md) and follow-up [#138](https://github.com/dzianisv/opencode-plugins/issues/138).
+
 ## @telegram.ts - integrates with Telegram over [t.me/OpencodeMgrBot](@OpenCodeMgrBot) bot
 <img width="1019" height="734" alt="image" src="https://github.com/user-attachments/assets/6f120c14-dba5-431b-a458-0f51f360f561" />
 @tts.ts - uses coqui TTS to read the opencode agent response. Useful to run a few agents on macOS and be notified when one finishes a task. 

claude/.claude-plugin/plugin.json

@@ -0,0 +1,13 @@
+{
+  "name": "reflection-cc",
+  "version": "0.1.0",
+  "description": "Re-prompts Claude Code when it stops prematurely due to failure modes like summary-drift-stop or tool-available-punt",
+  "author": "dzianisv",
+  "license": "MIT",
+  "hooks": {
+    "stop": {
+      "command": "${CLAUDE_PLUGIN_ROOT}/bin/reflect.mjs",
+      "timeout": 30000
+    }
+  }
+}

claude/README.md

@@ -0,0 +1,80 @@
+# reflection-cc
+
+Re-prompts Claude Code when it stops prematurely due to failure modes like summary-drift-stop or tool-available-punt. This plugin intercepts the Stop hook, analyzes the session transcript using Claude Haiku, classifies the failure reason, and decides whether to re-prompt with recovery instructions or accept the stop.
+
+## Install
+
+**Recommended (works today, CC v2.x):** add the Stop hook directly to `~/.claude/settings.json`:
+
+```json
+{
+  "hooks": {
+    "Stop": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "/absolute/path/to/opencode-plugins/claude/bin/reflect.mjs",
+            "timeout": 30
+          }
+        ]
+      }
+    ]
+  }
+}
+```
+
+The plugin manifest under `.claude-plugin/` is included for future marketplace publication, but in CC v2.1.150 `--plugin-dir` and the `enabledPlugins` config path do NOT activate `Stop` hooks for headless `-p` sessions. The settings-based install above is the authoritative path until that gap closes.
+
+**One-session try:** `claude --settings '<json above>'` ... or write the JSON to a file and pass `--settings ./reflect-settings.json`.
+
+## Failure Categories
+
+The classifier maps each Stop into one of these categories (only the first three trigger an inject):
+
+| Category | Inject? | Description |
+|----------|---------|-------------|
+| `summary_drift_stop` | **yes** | Agent wrote a plan with a "next step" then stopped before doing it |
+| `tool_available_punt` | **yes** | Agent asked the user about something an available tool could resolve |
+| `genuinely_stuck` | **yes** | Agent halted mid-thought, no question, no plan |
+| `complete` | no | Task finished |
+| `working` | no | Mid-action narration (rare at Stop) |
+| `waiting_for_user_legitimate` | no | Agent legitimately needs user input |
+
+## How it works
+
+1. **Stop Hook**: Claude Code invokes the Stop hook when the agent terminates
+2. **Transcript Analysis**: Haiku classifies the session transcript into failure categories
+3. **Verdict**: Judge decides to re-prompt with recovery instructions or accept the stop
+4. **Session Guards**: Loop prevention via attempt counter (max 3 cycles per session)
+
+## Testing
+
+`node claude/test/e2e-cc.mjs` runs 4 real E2E scenarios:
+
+- `explicit_wait_negative` — user asked "wait" → plugin must not inject.
+- `complete_negative` — trivial Q&A → plugin must not inject.
+- `attempt_cap_respected` — multi-file task → cap honored.
+- `direct_pipe_summary_drift` — synthetic drift transcript piped to `reflect.mjs` → verifies the inject pathway end-to-end including schema-correct stdout.
+
+Real `claude -p` headless sessions + real Anthropic API. No stubs. Costs roughly $0.05–0.20 per scenario via Haiku 4.5 over your Max-subscription OAuth. Out of CI (auth + cost). Run before any change to the hook payload format.
+
+## Configuration
+
+Environment variables:
+- `REFLECTION_CC_DEBUG=1` — Enable debug output
+- `REFLECTION_CC_MODEL` — Model for classification (default: `haiku-4-5`)
+- `REFLECTION_CC_MAX_ATTEMPTS=3` — Max re-prompt cycles per session
+
+## Disk Artifacts
+
+Plugin stores local transcripts and verdicts:
+- `.reflection/verdict_<sid>.json` — Haiku verdict + recovery instructions
+- `.reflection/<sid>_<ts>.json` — Full transcript snapshot
+- `.reflection/<sid>_attempts.json` — Attempt counter for session
+
+**Privacy**: Transcripts stored locally only; never sent externally except to Haiku during classification.
+
+## Status
+
+Experimental. Baseline accuracy numbers pending PR evaluation.