Multi-agent astrophotography assistant — plan, watch, operate. Solar-first. Astronomy planning and framing capability server with read-only and idempotent calculation tools.
Nadir demonstrates solid naming conventions (verb_noun pattern: compute_*, resolve_*, compose_*, rank_*, etc.) and comprehensive parameter descriptions. All 29 tools have descriptions and input schemas with typed parameters. However, output schemas are partially documented, some tools return complex nested objects (FrameDescription, SessionPlanModel) without inline field documentation visible in the server code. Parameter descriptions are generally good (60-150 chars) but lack explicit constraints (ranges, enums, format patterns) that would prevent LLM hallucination. Tool annotations are present and correctly applied (readOnlyHint, destructiveHint, idempotentHint). Error handling is minimal, no recovery guidance or actionable error messages visible. The 'object' type for location, setup, and components parameters accepts 'shorthand' but the schema does not formally define what shorthand means, forcing LLMs to guess.
Build a full night-session plan: night overview, ranked targets, observability curves, and fallbacks.
Check whether a setup is under- / well- / over-sampled for a given seeing.
Compose a target framing description (FOV-on-target) for the Sun and deep-sky objects, with optional SVG previews.
Compose group framing for folk-name multi-object groups (e.g. 'Leo Triplet') with bounding box and per-member positions.
Compute the field of view for a telescope + sensor combination.
Compute the image scale (arcseconds per pixel) for a given focal length and pixel pitch.
Polymorphic 'object' parameters (location, setup, components) lack formal schema definition. Descriptions mention 'shorthand' and 'optional' but do not specify valid keys, types, or constraints. LLMs cannot validate input without explicit schema.
Output schemas for complex tools (compose_frame, build_session_plan, rank_targets) are defined as Pydantic models (FrameDescription, SessionPlanModel, RankedTargetsResponse) but field-level documentation is not visible in the server.py excerpt. LLMs cannot infer which fields are required or what they contain.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Compute planning-grade local solar-eclipse circumstances and rank sites. Geometric contacts, observability at maximum, and candidate-site ranking. Verify path-edge calls against an authoritative Besselian-elements source and check the real horizon manually.
Get night overview: twilight times, moon phase, and observability summary.
Snapshot of current solar activity from NOAA SWPC (X-ray, regions, Kp).
Rank observable targets for a night session, paginated. Returns ranked targets with observability curves and fallbacks.
Pre-fill a payload for a missing profile entry (location auto-geocoded) or skeleton for user completion.
Remove a camera from the user profile.
Remove an eyepiece from the user profile.
Remove a filter from the user profile.
Remove a location from the user profile.
Remove an imaging or visual setup from the user profile.
Remove a telescope from the user profile.
Geocode a place name to a Location with elevation.
Resolve a target name (catalog ID, Messier, folk-name aliases, 'Sun') to a discriminated single | group payload.
Save a camera to the user profile.
Save an eyepiece to the user profile.
Save a filter to the user profile.
Save a location to the user profile.
Save an imaging or visual setup to the user profile.
Save a telescope to the user profile.
Set the default imaging setup in the user profile.
Set the default location in the user profile.
Set the default visual (eyepiece) setup in the user profile.
Compute observability curve for a target over a time window.
No error handling or recovery guidance visible. Tools that call external services (resolve_place, get_sun_now, eclipse_circumstances) do not document failure modes, timeouts, or what the LLM should do if a call fails.
Numeric parameters lack explicit min/max constraints. E.g., focal_mm, pixel_um, expected_seeing_arcsec have no documented ranges. LLMs may pass negative, zero, or absurdly large values.
Pagination parameters (offset, limit) in rank_targets lack documented bounds. No indication of max page size, default limit, or total count in response.