What's next

The backlog. No vision copy.

Shipped, needs finishing

Graph admin UI approval queue — src/daemon/graph-panel.ts lists pending classifications but the approve/reject buttons aren't wired. CLI does the same job via agentx graph review. ~1 day.

Classifier timeout resilience — bumped 20s → 60s in 2eba778; still aborts on cold-start. Next: retry-with-backoff, or bypass POST /task and call the LLM SDK directly.

A/B / online-eval harness — wiki ab-test works interactively; no standing run, no production shadow mode. Promoted from "waiting on traffic" after a 4-layer-stack proposal review confirmed this is the only real production-grade gap (rest of the Google ADK/MCP/Vertex/A2A surface is already covered by src/agents/, src/mcp/index.ts, src/a2a/, src/business/kpi.ts + token-tracker.ts). Build: shadow-mode flag in src/agents/registry.ts that, for a sampled % of incoming tasks, runs the same input through a B-version (alt model / alt prompt / alt agent) in parallel, persists both outputs alongside the existing TaskRecord, and surfaces a side-by-side diff in the dashboard. ~2 days. Owned ~$0; alternative was Vertex AI Agent Engine which is vendor lock-in + ongoing bill for capabilities already built.

Scoped, partial

Procedure-delta extraction — data model + agentx procedure list/add/show shipped. Deferred: the delta flow itself (agents emit one-line deltas when they run a procedure, cheap O(runs) instead of O(entries × articles)). ~1 week for the delta pipeline.

Cross-mesh graph sync — v1 shipped: GET /graph/{schema,nodes,classifications} on the daemon + agentx graph pull --from <peer> (leader-follower, local wins on conflict, fingerprint cache populated from peer's approved classifications). Bidirectional push + schema reconciliation still parked.

MCP tool surface beyond wiki_query — agentx_wiki_patch, agentx_wiki_interview, agentx_graph_review now exposed over MCP. Claude Code / Cursor / Windsurf can call them as tools. Remaining CLI-only: wiki edit, wiki quiz, skill sync, procedure add/show.

Parking lot

Auto-prune rejected classifications · retrieval-weights audit (verify graph.retrievalWeights is actually wired) · classifier content-hash short-circuit for trivial messages ("thanks", "ok") · MCP client compatibility sweep (Cursor/Windsurf/Zed) · MCP-client wiring per agent (centralize each agent's .mcp.json via an mcp: block in agentx.json, install at boot via the installAgentMemorySurface() pattern in src/daemon/index.ts; ~half a day) · public procedure registry (after Procedures land) · graph taxonomy tree-viz.

Explicitly not doing

RBAC / multi-operator auth — parked until a second operator shares a deployment. Hosted / SaaS — we're self-hosted for SMB by design. Voice / Canvas UI — OpenClaw's territory.