radar-cc

Intent alignment checker for Claude Code. Radar watches Claude Code's OpenTelemetry stream and tells you — in a second terminal pane — whether your prompt was clear before Claude starts, and whether Claude stayed on target after it finishes.

Requirements

• Node.js >= 22
• An Anthropic API key

Install

npm install -g radar-cc
radar setup

radar setup writes the required OTel environment variables to ~/.claude/settings.json and walks you through storing your Anthropic API key — either on local disk or in 1Password. Restart Claude Code after running it.

You can also pass the key directly to skip the prompt:

radar setup --api-key <your-key>

Usage

Open a second terminal pane alongside Claude Code and run:

radar watch

Send prompts in your Claude Code pane as normal. Radar listens passively on localhost:4820.

Output

── Radar v0.1.0 ─────────────────────────────────────
Listening on localhost:4820
Waiting for Claude Code telemetry...
─────────────────────────────────────────────────────

── PRE ── 14:23:07 ── score: 0.34 ── ✓ Clear ────────

── PRE ── 14:25:12 ── score: 0.78 ───────────────────
⚠ "clean up this module" is ambiguous.
  Claude will likely restructure imports and rename functions.
  Did you mean: remove the 3 commented-out functions?
  → Try: "delete the dead code in auth.ts — the 3 commented
    functions at the bottom"
─────────────────────────────────────────────────────

── POST ── 14:25:38 ──────────────────────────────────
✓ Response aligned with intent.
  Tools: Edit (2 files) · 847 tokens · $0.003
─────────────────────────────────────────────────────

── POST ── 14:31:02 ──────────────────────────────────
✗ Scope exceeded likely intent.
  Claude ran Edit on 5 files, Bash (3 commands), 12k tokens, $0.08.
  Developer likely wanted: coverage for 2 new edge cases only.
  → "undo all changes. add test cases for the null input
    and timeout edge cases in processOrder — nothing else"
─────────────────────────────────────────────────────

PRE advisories fire within ~2 seconds of your prompt. POST advisories fire after Claude's turn ends, based on tool activity, cost, and a Haiku-generated summary of Claude's actual response text.

How it works

Setup: radar watch starts an HTTP server on port 4820. Claude Code streams OpenTelemetry events to it.

When you submit a prompt:

1. Claude Code emits a user_prompt OTel event → Radar opens a new turn context
2. Radar immediately calls Haiku to score ambiguity (0–1). Result cached on the turn context
3. Score < 0.6 → suppress. Score ≥ 0.6 → call Haiku again to generate a warning, print yellow pre-advisory box
4. Meanwhile, Claude Code is running — tool results and API costs stream in and accumulate on the turn context

When Claude finishes:

5. The Stop hook fires, reads Claude Code's JSONL transcript (~/.claude/projects/**/<session_id>.jsonl) to extract the last assistant response, POSTs { sessionId, lastAssistantMessage } to Radar
6. Radar waits 3.5s for straggling OTel events, then closes the turn using the cached ambiguity score from step 2
7. Score < 0.6 → skip post-advisory entirely (no further Haiku calls)
8. Score ≥ 0.6 → call Haiku to summarise the assistant response, then call Haiku to judge whether it aligned with the original intent
9. Aligned → dim one-liner (suppressed in alert-only mode). Misaligned → red box with a re-prompt suggestion
10. Full turn written to ~/.config/radar/history/YYYY-MM-DD.jsonl for history and future review

Next prompt: the classifier receives the last 3 turns as context, so references like "do the same for the tests" score correctly rather than looking vague in isolation.

Claude is never blocked or interrupted.

OTel variables (set by radar setup)

| Variable | Value | |---|---| | CLAUDE_CODE_ENABLE_TELEMETRY | 1 | | OTEL_LOGS_EXPORTER | otlp | | OTEL_EXPORTER_OTLP_PROTOCOL | http/json | | OTEL_EXPORTER_OTLP_LOGS_ENDPOINT | http://localhost:4820/v1/logs | | OTEL_LOG_USER_PROMPTS | 1 | | OTEL_LOG_TOOL_DETAILS | 1 | | OTEL_LOGS_EXPORT_INTERVAL | 2000 |

OTEL_LOG_USER_PROMPTS=1 is required for prompt content analysis. Without it, Radar can detect turn boundaries but cannot analyze intent.

These are written to the env block in ~/.claude/settings.json. If Radar is not running, Claude Code silently ignores them.

Options

radar setup [options]

  -k, --api-key <key>       API key to store (skips the interactive prompt)

radar watch [options]

  -p, --port <number>       OTLP listener port (default: 4820)
  -t, --timeout <ms>        Turn boundary silence window (default: 5000)
  -s, --threshold <score>   Ambiguity score threshold 0.0–1.0 (default: 0.6)
  -k, --api-key <key>       Anthropic API key (overrides all stored sources)

Developing locally

Switch from the global install to a local build

1. Uninstall the global package:

   npm uninstall -g radar-cc

2. Install dependencies and build:

   npm install
   npm run build

3. Link the local build as the global radar binary:

   npm link

   From now on, radar points to dist/cli/index.js in this repo. Re-run npm run build after any code change — no re-linking needed.

Run setup and tests

4. Run setup to (re-)install the hook scripts:

   radar setup

   Confirm you see both of these lines in the output:

   ✓ Hook script written to ~/.radar/hooks/stop.sh
   ✓ Extract script written to ~/.radar/hooks/extract-response.py

5. Restart Claude Code so the new Stop hook takes effect.
6. Run the test suite:

   npm test

Verify the pipeline end-to-end

7. Check the installed hook version:

   head -2 ~/.radar/hooks/stop.sh
   # should print: # radar-hook-v2

8. Confirm the extract script is present:

   ls ~/.radar/hooks/extract-response.py

9. Manually test the extract script against a real transcript:

   # Find a session transcript (created after Claude Code runs with OTel enabled)
   ls ~/.claude/projects/
   python3 ~/.radar/hooks/extract-response.py \
     ~/.claude/projects/<folder>/<session-id>.jsonl
   # Should print a JSON-encoded string of the last assistant response

10. With radar watch running, manually POST a stop payload to confirm the endpoint accepts the new field:

    curl -s -X POST http://localhost:4820/v1/hook/stop \
      -H "Content-Type: application/json" \
      -d '{"sessionId":"test-123","lastAssistantMessage":"I refactored the auth module."}' | cat
    # Should return: {"ok":true}

11. Send a prompt in Claude Code and watch the radar watch pane. The POST advisory should now reflect what Claude actually said, not just which tools it used.

License

MIT