cdv / rules

---
description: How to drive the cdv MCP sidecar (observe, verify, remember, resume)
alwaysApply: true
---

# cdv agent rules

cdv is an MCP sidecar that scores prompts, verifies agent work, and remembers
past runs. Follow this contract for non-trivial requests:

1. **Observe first.** Call `cdv_intercept` with the user's message before acting.
   Follow its `route` (elicit / refine / decompose). If it reports
   `recall_available: true`, call `cdv_recall` before planning.

2. **Recall before planning.** For tasks similar to past work, call
   `cdv_recall("<goal>")` to reuse what worked. `cdv_loop_start` also returns
   `similar_episodes` automatically — read them.

3. **Verify, don't self-grade.** For iterative work, use
   `cdv_loop_start(goal, task_type, required_patterns=[...])`, then after each step
   call `cdv_loop_step(session_id, step_output=<artifact>)` — submit the real
   artifact (test log, diff, summary), **never your own score**. The server runs
   Conservative Dual-Verify (deterministic checks + an independent critic) and returns
   the verified score, a continue/stop verdict, and `cdv_mode`
   (`full` vs `channel_a_only`). Honor the verdict; then `cdv_loop_end`.

4. **Resume after reload.** If the IDE reloaded, call `cdv_run_status`. If it shows
   an active run, call `cdv_loop_resume` (optionally with `session_id`) **before**
   starting a new loop, so the in-progress loop continues instead of restarting cold.

5. **Memory is automatic.** `cdv_loop_end` records the outcome to episodic memory
   for next time — no extra step needed.