- Q3: prod void DB role NOSUPERUSER (vector marked trusted; deploy/README documents it)
- Q4: buildChildEnv allow-list for the claude subprocess (no OWNER_TOKEN/DATABASE_URL/secrets leak)
- Q5: pending-change approve claims-before-applying + reopens on failure (no re-approvable dup)
- Q6: /capture/upload validates space_id (UUID+existence); pg pool statement_timeout 30s
- Q9: disabled failing syncoid-donatello timer on Z
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three more read-only tools on securityRegistry:
- pending_review: agent-proposed changes awaiting approval (injection surface)
- resource_exposure: host/url/status attack-surface inventory (resources.listExposure,
scalar cols only — no monitoring/metadata/credentials)
- token_audit: token label/last_used/revoked, never the hash
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- pool.js: add pool.on('error') handler — an idle-client error (DB restart /
.215 failover) previously crashed the process (no 'error' listener → throw)
- context tool: project a SAFE_COLUMNS allow-list for resources (never the
monitoring/metadata JSON blobs); also add 'resource' to TABLE (was unhandled)
- applyPendingChange: guard the 'upsert' arm so a non-upsertable entity_type
fails with a clear ValidationError instead of a bare TypeError
Tests: pool_error, context (resource case), pending_extended_actions (guard).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New securityRegistry (separate from companionRegistry) with two read-only,
secret-free tools for the Yerin security agent:
- audit_log: query the redacted audit trail by actor_kind/actor_id
- agent_inventory: list agents + capabilities/scopes (explicit projection,
never SELECT *, no token material)
Follows the existing createRegistry() pattern. Design + wiring roadmap in
docs/yerin-security-agent.md. Not yet seeded/exposed over MCP (left for review).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- system prompt = Dross (Ozriel's construct fragment, per Void 1.0), with tool guidance
- migration 008 renames the seeded agent 'companion' → display name 'Dross'
- removed lib/ai/anthropic.js + lib/ai/agent/runtime.js + tests + @anthropic-ai/sdk dep (companion now runs via the claude CLI; kept lib/ai/secret.js for the Vaultwarden roadmap)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- driver: tool_results arrive as type:'user' content blocks (not bare); parse them
- route: tool_result content is a JSON string; parse it for pending_change_id → draft event
- propose_change: inject ctx.space_id into create payloads (model can't know the uuid; tables require it)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
claude resolves the MCP server command against the child env (no PATH), so a
bare 'node' failed to spawn (status:failed). Use process.execPath. Also pass
--tools to drop claude's built-ins (Bash/Read/Write/…) — companion gets only
the four mcp__void__* tools.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Implements runClaudeTurn() — spawns the claude CLI for a single companion
turn using subscription/OAuth auth (strips ANTHROPIC_API_KEY +
ANTHROPIC_AUTH_TOKEN from child env), streaming normalised events (delta,
tool, tool_result, result, error) via onEvent callback.
Includes hermetic test + fake-claude.js fixture that mimics real 2.1.159
stream-json output; zero network/CLI calls in the test suite.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
