Intelligence Endpoints (v2.1)

Intelligence endpoints expose the Prism Engine’s v2.1 scoring and insight output. All endpoints are scoped to the authenticated organization.

PRISM scores (Layer 2 + Layer 3)

v3.0 equivalent: GET /v1/score_v3/score, plus /v1/score_v3/speed, /v1/score_v3/tokens, and /v1/score_v3/skill for the pillars.

GET /prism/scores

Layer 3 composite scores — Speed, Skill, Efficiency — per developer and period.

GET /prism/scores?from=2024-01-01T00:00:00Z&to=2024-01-07T00:00:00Z
Authorization: Bearer gck_your_key

GET /prism/metrics

Layer 2 per-sub-session metrics (SSE, PES, IE, CRR, FC, QR) used as inputs to the Layer 3 scores.

GET /prism/metrics?from=2024-01-01T00:00:00Z
Authorization: Bearer gck_your_key

POST /prism/metrics/recompute

Trigger a Layer 2 recompute for a time window. Used by admin tooling after baseline changes.

GET /prism/trend

Trend series for Skill / Speed / Efficiency.

GET /prism/l1-hourly

Layer 1 hourly aggregates.

Insight (per-prompt rubric)

v3.0 equivalent: GET /v1/score_v3/insight/sessions — the v3 rubric is per-sub-session (not per-prompt) and uses seven booleans + judge status instead of the four PES dimensions.

GET /prism/insight

List scored prompts (intent + four-dimension Efficiency Rubric).

GET /prism/insight/:prompt_id

Single scored prompt by ID.

POST /prism/insight/score

Score a prompt on-demand using the same LLM rubric as the background worker.

POST /prism/insight/score
Authorization: Bearer gck_your_key
Content-Type: application/json

{
  "prompt_text": "Fix the bug in auth.ts",
  "org_id": "…",
  "session_id": "…",
  "turn_index": 0,
  "prompt_id": "…"
}

Response: a ScoreResponse object with camelCase fields:

{
  "detectedLanguage": "en",
  "intent": "fix",
  "shortCircuited": false,
  "rubric": { "contextLeverage": 7, "informationDensity": 8, "turnEconomy": 8, "ambiguityCost": 7 },
  "rubricScore": 7.5,
  "grade": "B0",
  "confidence": 0.82,
  "agentVersion": "…",
  "version": "1.0"
}

Conversational or low-value prompts are classified as question, meta, continuation, or system_callback; for those, shortCircuited is true and rubric / rubricScore / grade are omitted.

GET /prism/insight/efficiency-rubric

Averages for the four PES dimensions over the requested window.

GET /prism/insight/efficiency-rubric-trend

Trend for each PES dimension.

GET /prism/insight/session-efficiency

Per-session PES roll-up.

Drill-downs (no direct v3 equivalent yet)

These endpoints feed dedicated v2.1 insight pages (/insights/rlr, /insights/qr). The v3.0 page set doesn’t have a 1:1 replacement.

GET /prism/response-latency-hourly

Hourly breakdown of model response latency across the window.

GET /prism/response-latency-worst

Slowest individual turns in the window with turn identifiers for drill-in.

GET /prism/quality-retention-worst

Sessions with the lowest Quality Retention.

Sub-session efficiency

v3.0 equivalent: the v3 score itself is computed at the sub-session level — see /v1/score_v3/score and /v1/score_v3/speed/sessions. The v2.1 efficiency endpoints surface raw SSE breakdowns and are still used by the v2.1 Efficiency page.

GET /v1/efficiency/sub-sessions

List scored sub-sessions with SSE breakdown.

GET /v1/efficiency/summary

Rolled-up SSE for the date range.

GET /v1/efficiency/by-sessions

Sub-session efficiency grouped by parent session.

GET /v1/efficiency/sub-sessions-by-sessions

Sub-sessions grouped under their parent session.

Advisor

v3.0 equivalent: no direct replacement yet — the advisor still runs against the v2.1 scoring inputs.

POST /v1/advisor/sub-session

Run the advisor for a specific sub-session — produces coaching cards.

GET /v1/advisor/cards

List stored advisor cards for the org.

POST /v1/advisor/cards/:id/feedback

Submit thumbs-up / thumbs-down feedback on a coaching card.

GET /v1/advisor/stream

Server-sent events stream of advisor cards as they’re produced.

Worst / best prompts

v3.0 equivalent: see the Trivia page (/v1/score_v3/trivia) for filtered-out sessions and the Insight rubric for low-grade prompts. There’s no like-for-like “worst prompts” endpoint in v3 yet.

GET /v1/intelligence/worst-prompts

Lowest-scoring prompts with coaching notes.

GET /v1/intelligence/worst-prompts?from=2024-01-01T00:00:00Z&limit=20
Authorization: Bearer gck_your_key

Parameter	Type	Default	Description
`session_id`	string	—	Filter by session
`from`	ISO8601	—	Start time
`to`	ISO8601	—	End time
`limit`	integer	20	Max records (max 50)

Response: { "prompts": [...], "total": N }.

GET /v1/intelligence/best-prompts

Highest-scoring prompts — useful for celebrating patterns that work.

GET /v1/intelligence/longest-chains

Longest coherent prompt chains in the window.

Throttle & Rightsizing

v3.0 equivalent: no direct replacement — these are operational signals, not scoring outputs, and stay where they are.

GET /v1/intelligence/throttle

Rate-limiting events, latency spikes, and recurring drop-off detection.

GET /v1/intelligence/rightsizing

Model selection and tier recommendations based on usage patterns. Served directly from Parquet usage (no background worker).

Sub-session overrides

v3.0 equivalent: the v3 dashboard uses PATCH /v1/score_v3/sub-sessions/:id to rename sub-sessions. The v2.1 override endpoint below is for boundary-detection corrections.

POST /v1/sub-sessions/:id/override

Submit a human override on an automatically-detected sub-session boundary.

GET /v1/sub-sessions/accuracy

Accuracy metrics for sub-session boundary detection.

Satisfaction

The weekly satisfaction check-in is unchanged between v2.1 and v3.0.

GET /v1/satisfaction/status

Whether the weekly satisfaction check-in is pending for the current developer.

POST /v1/satisfaction

Submit a satisfaction response.

Removed endpoints

The following endpoints existed in earlier versions and have been removed. Clients calling them will get a 404:

Endpoint	Replacement
`GET /v1/intelligence/prism`	`GET /prism/scores` (v2.1) or `GET /v1/score_v3/score` (v3.0)
`POST /v1/intelligence/score-prompt`	`POST /prism/insight/score`
`GET /v1/intelligence/waste`	Surfaced in-app only via the report pages
`GET /v1/intelligence/coaching`	`GET /v1/advisor/cards`
`GET /v1/intelligence/recommendations`	`GET /v1/intelligence/rightsizing`
`GET /v1/telemetry/vibe-metrics`	Removed with the Vibe Metrics feature