Summary

• expose a compact workspace view over MCP and HTTP with registered media, active-index coverage, model readiness, capability roles, and identity semantics
• reject unsupported capability/media selections synchronously with structured alternatives and recovery guidance
• route index preflight through the shared job boundary so HTTP, MCP, CLI, frontend, and application composition cannot enqueue invalid work
• keep actor evidence queryable, inspectable, and renderable without presenting anonymous clusters as text-searchable identities

Why

Agents previously had to discover capability/index incompatibilities through trial and error, and some invalid requests could become durable jobs before failing. Preflight also differed by adapter: HTTP and MCP checked more than the CLI and frontend paths.

This change gives every client the same truthful state projection and moves index validation to the common job submission boundary.

User and developer impact

Agents can call get_workspace before planning work, select only roles supported by the active snapshot, and recover from validation failures using structured reason, available, indexed, and next_action fields. Invalid index requests do not reach the durable backend.

Validation

• python -m compileall -q src
• ruff check src tests
• full suite: 509 passed, 3 skipped, 61 subtests passed
• editable install: uv tool install --python 3.13 --editable ".[local-worker,mcp]"
• installed stdio handshake: 17 tools discovered and get_workspace returned the active two-video repository
• installed MCP preflight smoke: invalid actor search and unknown indexing capability returned structured errors while the 26-job list remained unchanged