Sam Valladares 80c823a3ca feat(blackbox): Agent Black Box + Receipts + risk-gated Memory PRs ... Watch the agent think. Watch memory change. Watch the receipt prove why. Make Vestige the first memory server where you can replay an agent run, audit every retrieval, and review changes to the agent's brain like code. Phase 0 — the trace-correlation spine. One runId threads, unbroken, through every layer: MCP tool output (runId + traceUri) -> SQLite agent_traces rows -> WebSocket TraceEvent -> dashboard pulse -> /api/traces/:runId -> vestige://trace/{runId} -> .vestige-trace.json export -> Cinema replay input. Proven end to end by a real JSON-RPC round-trip integration test. Core (vestige-core): - trace/ module: MemoryTraceEvent (7 variants incl. contradiction.detected), Receipt, and classify_write — the pure, DB-free immune-system logic. - Risk taxonomy: contradiction-vs-high-trust, supersede/forget/merge/protect, identity/preference/workflow/positioning, auth/security/money/legal, dream consolidation, decay resurrection, low-confidence batch, weak-provenance connector. Fast / Risk-Gated (default) / Paranoid modes. - V18 migration: agent_traces, agent_runs, memory_receipts, memory_prs. - trace_store.rs: CRUD following the established store idiom. MCP (vestige-mcp): - trace_recorder.rs: records mcp.call + downstream retrieve/suppress/write/ contradiction/veto/dream events; builds + persists receipts; risk-gates writes into Memory PRs. Args are hashed, never stored raw. - server.rs dispatch stamps runId/traceUri/receipt onto every tool result and routes risky writes to the PR queue; trace events broadcast over WebSocket. - vestige://trace/{runId} resource; /api/traces, /api/receipts, /api/memory-prs. Dashboard: - Black Box tab: live spine header + Proof Mode, run picker, timeline scrubber, per-event detail, memory pulse, full event log, .vestige-trace.json export. - Memory PRs tab: GitHub-style cognition diff, self-explaining risk signals, Promote/Merge/Supersede/Quarantine/Forget/Ask-Agent-Why, mode toggle. - ReceiptCard with "Open receipt in Cinema" (deep-links graph; Cinema untouched). Gates: 987 lib tests pass, clippy -D warnings clean, dashboard check + build clean. Live proof in blackbox-proof-2026-06-22/. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>		2026-06-22 17:06:35 -05:00
..
dashboard	feat(blackbox): Agent Black Box + Receipts + risk-gated Memory PRs	2026-06-22 17:06:35 -05:00