dark-memory-mcp

Return server identity, version, and a liveness snapshot. No side effects.

LLM-backed one-shot judge evaluation. Runs the 6-step pipeline (ADR-007 §2), persists the verdict to sdd_evaluations + audit_log (INV-1), and returns the verdict.

List registered judge personas (8 legacy + 3 v4-new). Optionally include rich content (PromptTemplate + EvaluationLens + BiasControls).

Run text through the T5 normalization pipeline. Operator-facing primitive for inspecting what the validator sees. Deterministic and pure.

List the canonical 10 prompt-injection override patterns (OP-1..OP-10) with descriptions. Stable across versions.

Generate a fresh W3C Trace Context. Returns trace_id (32 hex), span_id (16 hex), trace_flags (01=sampled), and the canonical traceparent header.

Scan text for the 10 prompt-injection override patterns. Returns the list of hits (each is a pattern id, severity, description, and a 40-char normalized context snippet). Severity=block triggers EC-003 short-circuit in the judge pipeline.

Validate a W3C traceparent string against the grammar. Returns valid=true on match. The W3C spec is case-insensitive for hex.

Verify an Ed25519 signature over content. Returns valid=true on a clean match, valid=false otherwise. Malformed inputs return an error envelope.

Verify that the SHA-256 of content matches the supplied hex digest. Returns match=true on a clean match, match=false otherwise.

List recent judge verdicts persisted to sdd_evaluations. Read-only. Supports optional filters by eval_type, target_type, target_id.

Load a constitution by id+version. v4-alpha.1 returns the hard-coded default.

DB health snapshot: driver, schema version, per-table counts.

Compose a sub-agent system_prompt for the given vibe_case + task_description. Phase 6 alpha.18.1 full implementation: procedural composition with cache (agent_memory, kind=context, TTL via DARK_MINDSET_CACHE_TTL default 1h), judge validation via LLM-as-judge (eval_type=mindset_compose), retry loop up to DARK_MINDSET_MAX_ITERATIONS (default 3) if judge rejects.

Create or replay a project. Idempotent on project_id (first-writer wins; mutating display_name/description/default_agent_id requires Archive + re-Create). Reserved ids 'default' and 'dark' are rejected. Emits one INV-1 audit_log row on first insert.

Look up one project by id. Returns `project: null` when not found (no error). Includes archived projects; check `is_archived` to filter.

Recall prior research items by query string. Returns Evidence-shaped items the operator can re-use without re-dialing the backends.

Continue a multi-turn research thread. The executor stitches the new query onto the prior context and returns the merged evidence.

Run a research query across the 17 no-API-key backends. Tier-aware fan-out (T1 → T2 → T3), health-aware routing, per-type cache TTL with stale-while-revalidate, SSRF guard, prompt-injection gate, and entity resolution across corroborating backends.

Terminalize a session. Emits one INV-1 audit row.

Refresh session last_heartbeat_at. No-op if the session is closed.

Rebind to an existing session. Refreshes heartbeat; returns current state.

Open a new session and emit one INV-1 audit row. Returns the session_id plus a Loadout of the operator's startup context (pinned rows, open todos, recent audit writes, constitution, schema version, server time). PRE-1 C3: loadout is best-effort; per-field failures surface in loadout_warnings.

Read a session's current state. INV-7 enforces project_id match.

Record a skill load event in agent_memory (kind=observation, tags=skill_loaded:v1+skill:<name>). Emits one INV-1 audit row per call. The skill_loaded convention is the explicit form of PRE-1 C2 L2.

Return a markdown handoff document for a session. Includes metadata, pinned rows, open todos, recent audit_log writes, and skills loaded. Read-only; no audit emission.

Latest drift verdict for one artifact.

Publish one artifact under a spec. Triggers judge + drift_log; returns drift verdict.

Operator gate: accept (artifact correct as-is) or reject (artifact wrong).

Create a vibe-loop spec with a validated task graph. Returns the new spec_id.

Recent INV-1 audit_log rows (most-recent first).