Dorveille

dorveille — the trust layer for agent self-improvement

dorveille makes Claude Code compound (cross-session memory + self-forged skills), but unlike Hermes and every other self-improving agent it verifies and evaluates what it learns so you can trust it didn't learn something wrong. It runs natively in your repo, on your own subscription, via the real claude binary in hooks. Every lesson (a memory entry or a forged skill) flows through:

1. Write + provenance (Stop / SessionEnd → review worker): a claude -p worker reads the finished transcript, writes lessons, and registers each with provenance (session, trigger, web-influence) + a base trust score.
2. Independent verification (chained verify.sh): a separate agent challenges each lesson — durable vs env-artifact, contradiction, web-poisoning, regression — and quarantines failures (pulling them from the live store). The writer can't verify its own drift; this agent can.
3. Trust-gated injection (SessionStart): only lessons at/above the trust threshold are injected. Low-trust / quarantined lessons are held back.
4. Probe replay (eval.sh, scheduled/on-demand): each lesson carries a tiny yes/no probe; replay re-checks it still holds, dropping trust on drift.
5. Curiosity (scheduled): researches one weak area, proposes an improvement (often with a one-command apply block).

Alongside the trust loop, dorveille has full parity with the Hermes mechanics:

• Episodic recall — FTS5 search over every past session transcript (dorveille recall), so the agent can pull back a decision or config from weeks ago instead of asking you to repeat yourself. See the dorveille-recall skill.
• The curator — a background pass that consolidates forged skills into class-level umbrellas and archives stale ones (never deletes), so the skill library stays sharp instead of sprawling (dorveille curator).
• Skill usage telemetry + pinning — forged skills track uses/last-used; dorveille skill pin <name> protects one from the curator.

The dorveille CLI

dorveille status                      # dashboard: memory, recall, skills, trust, probes, compounding
dorveille trust                       # every lesson with provenance + trust score
dorveille quarantine list             # lessons the verifier held back, with reasons
dorveille quarantine release <id>     # vouch for one → restore it to the live store
dorveille quarantine discard <id>     # permanently drop one
dorveille eval                        # replay probes now (did learning still hold?)
dorveille mem show user|memory        # inspect the bounded stores
dorveille recall search "<query>"     # search past sessions · scroll/browse/index
dorveille skill list                  # forged skills with usage/age · pin/unpin/stale
dorveille sweep                       # review idle, unreviewed sessions (reliable trigger)
dorveille backfill [N]                # one-time: learn from the N most-recent past sessions
dorveille curator [dry-run]           # consolidate the forged-skill library
dorveille curiosity                   # run the curiosity loop now
dorveille on | off                    # resume / pause the whole loop

(If dorveille isn't on PATH: python3 ${CLAUDE_PLUGIN_ROOT}/bin/dorveille ....) Schedule the proactive loops with bash ${CLAUDE_PLUGIN_ROOT}/scripts/install-cron.sh.

Where things live (all git-trackable, auditable, revertable)

• ~/.dorveille/USER.md (1375 chars) · MEMORY.md (2200) · SOUL.md — the stores.
• ~/.dorveille/lessons.json — the trust registry (provenance, trust, verification, probe).
• ~/.dorveille/quarantine/ — held lessons + stashed copies (release-able).
• ~/.dorveille/eval-history.jsonl — probe pass-rate over time.
• ~/.dorveille/sessions.db — FTS5 episodic-recall index over past transcripts.
• ~/.dorveille/pruned/ — skills the curator archived (recoverable; never deleted).
• ~/.dorveille/corrections.jsonl · curiosity/ · logs/ (review/verify/eval/curiosity/curator).
• Forged skills → ~/.claude/skills/<name>/ with forged-by: dorveille.

Trust scoring

Base from provenance: correction 85, technique 65, web-research 45 (−20 if the session used the web). The verifier may raise it (clear correction) or quarantine (poisoning / contradiction). A failed probe drops it 25. Below DORVEILLE_TRUST_THRESHOLD (default 50) → held from injection.

Safety invariants

• dorveille only writes its own store + skills it forged; never edits human skills, CLAUDE.md, or settings. The curator only ever touches forged-by: dorveille skills and archives (never deletes).
• Verification is independent (separate claude -p, DORVEILLE_REVIEWING-guarded so it can't recurse). The writer never signs off its own work.
• The memory store is hardened: a deterministic injection/exfil/invisible-unicode scan blocks promptware at write- and inject-time, an fcntl lock makes concurrent sessions safe, and a drift guard backs up rather than clobbers out-of-band edits.
• Memory is bounded; quarantine + release keep self-modification reversible.
• Pause: touch ~/.dorveille/OFF (or dorveille off). Tune: DORVEILLE_TRUST_THRESHOLD, DORVEILLE_NUDGE_INTERVAL, DORVEILLE_REVIEW_MODEL, DORVEILLE_VERIFY_MODEL, DORVEILLE_EVAL_MODEL.