Stateless MCP server for AI film/video production state: keep characters, locations, and shots consistent across a project (orient / get_shot_status / update_shot).
Three tools with clear domain purpose (production state tracking). All have descriptions and basic schemas. However, parameter descriptions are sparse or missing entirely, output schemas are undocumented, and error handling lacks recovery guidance. Naming is verb-forward and appropriate. The server implements stateless reads and writes to sidecars, which is sound architecture, but the tool interface does not fully guide LLM usage.
Per-shot status for one scene, reconciling the IR plan, the canonical scene lock state, and the shim's working sidecar. Returns each shot's status (planned/pending/fired/locked/needs_rework/dropped) plus a summary.
Where am I right now: project, phase, active episode, budget/fire posture, open blockers, next actions, and the scene index. Replaces the manual session-start orientation paste. Pure read of STATE.json.
Mark a shot planned/pending/fired/locked/needs_rework/dropped and attach a generation job_id + still path. Writes ONLY to the scene's shim-owned shot_status.json sidecar; never mutates the canonical render_manifest. Refuses unknown statuses and anchors not in the scene's IR plan.
Output schemas are not documented. LLMs cannot predict what fields orient(), get_shot_status(), or update_shot() return, forcing them to guess downstream field names and risk broken chains.
Parameter descriptions are minimal or absent. 'scene' in get_shot_status has a 20-char description; 'shot_anchor' in update_shot has 40 chars. LLMs lack context on what values are valid or expected.
No enum constraints on 'status' parameter in update_shot(). The description lists valid values ('planned/pending/fired/locked/needs_rework/dropped') as free text, not as a formal enum. LLMs may hallucinate invalid statuses.
Error handling returns generic {'error': 'ClassName: message'} with no recovery guidance. If a scene is not found or a shot_anchor is invalid, the LLM has no hint on what to do next (e.g., 'Call orient() to list available scenes').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
No idempotency or confirmation pattern for update_shot(). This is a write operation that modifies state (shot_status.json). Agents retrying on ambiguous failures could mark the same shot multiple times. No dry-run or confirmation step exists.