thumbgate 1.14.1 → 1.16.0

package/config/model-candidates.json

@@ -0,0 +1,131 @@
+{
+  "version": 1,
+  "description": "Managed model candidates for ThumbGate workload benchmarking. Catalog only: no provider-specific runtime dependency is assumed here.",
+  "workloads": {
+    "pretool-gating": {
+      "label": "PreTool gating",
+      "summary": "Fast, reliable gate judgments for tool-use and agentic coding decisions before commands run.",
+      "desiredStrengths": ["agentic-coding", "tool-use", "reliability"],
+      "targetContextWindow": 64000,
+      "benchmarkCommands": [
+        "npx thumbgate eval --from-feedback --json --min-score=0",
+        "node scripts/gate-eval.js run",
+        "npx thumbgate bench --json --min-score=90"
+      ],
+      "metrics": [
+        "passRate",
+        "falsePositiveRate",
+        "falseNegativeRate",
+        "medianLatencyMs",
+        "costPer1kActionsUsd"
+      ]
+    },
+    "long-trace-review": {
+      "label": "Long trace review",
+      "summary": "Review long agent traces, multi-step failures, and large-context coding sessions without dropping important detail.",
+      "desiredStrengths": ["long-horizon-coding", "multi-agent", "reliability"],
+      "targetContextWindow": 128000,
+      "benchmarkCommands": [
+        "npx thumbgate eval --from-feedback --json --min-score=0",
+        "node scripts/gate-eval.js run",
+        "npx thumbgate bench --json --min-score=90"
+      ],
+      "metrics": [
+        "passRate",
+        "longContextReliability",
+        "traceCompressionLoss",
+        "medianLatencyMs",
+        "costPerTraceUsd"
+      ]
+    },
+    "cheap-fast-path": {
+      "label": "Cheap fast path",
+      "summary": "Low-cost first-pass model for cheap approval triage before escalating ambiguous work.",
+      "desiredStrengths": ["agentic-coding", "tool-use"],
+      "targetContextWindow": 32000,
+      "benchmarkCommands": [
+        "npx thumbgate eval --from-feedback --json --min-score=0",
+        "node scripts/gate-eval.js run",
+        "npx thumbgate bench --json --min-score=90"
+      ],
+      "metrics": [
+        "passRate",
+        "medianLatencyMs",
+        "costPer1kActionsUsd",
+        "escalationRate"
+      ]
+    }
+  },
+  "candidates": [
+    {
+      "id": "anthropic/claude-haiku-4-5",
+      "vendor": "Anthropic",
+      "family": "claude",
+      "provider": "anthropic",
+      "model": "claude-haiku-4-5-20251001",
+      "contextWindow": 200000,
+      "costClass": "low",
+      "strengths": ["tool-use", "reliability", "fast-inference"],
+      "notes": "Fast control candidate for cheap approval triage."
+    },
+    {
+      "id": "anthropic/claude-sonnet-4-6",
+      "vendor": "Anthropic",
+      "family": "claude",
+      "provider": "anthropic",
+      "model": "claude-sonnet-4-6",
+      "contextWindow": 200000,
+      "costClass": "medium",
+      "strengths": ["agentic-coding", "tool-use", "reliability", "long-horizon-coding"],
+      "notes": "Current stronger managed control candidate."
+    },
+    {
+      "id": "tinker/kimi-k2.6-32k",
+      "vendor": "Thinking Machines",
+      "family": "kimi",
+      "provider": "openai-compatible",
+      "gateway": "tinker",
+      "model": "kimi-k2.6-32k",
+      "contextWindow": 32000,
+      "costClass": "medium",
+      "strengths": ["long-horizon-coding", "multi-agent", "reliability"],
+      "notes": "Tinker April 23, 2026 release. Good candidate when long-horizon coding matters more than ultra-low latency."
+    },
+    {
+      "id": "tinker/kimi-k2.6-128k",
+      "vendor": "Thinking Machines",
+      "family": "kimi",
+      "provider": "openai-compatible",
+      "gateway": "tinker",
+      "model": "kimi-k2.6-128k",
+      "contextWindow": 128000,
+      "costClass": "medium",
+      "strengths": ["long-horizon-coding", "multi-agent", "reliability", "long-context"],
+      "notes": "Highest-ROI Kimi candidate for long traces and multi-step review."
+    },
+    {
+      "id": "tinker/qwen3.6-35b-a3b",
+      "vendor": "Thinking Machines",
+      "family": "qwen",
+      "provider": "openai-compatible",
+      "gateway": "tinker",
+      "model": "qwen3.6-35b-a3b",
+      "contextWindow": 64000,
+      "costClass": "low",
+      "strengths": ["agentic-coding", "tool-use", "reliability", "fast-inference"],
+      "notes": "Best first Tinker candidate for ThumbGate pre-action gating and tool-risk classification."
+    },
+    {
+      "id": "tinker/qwen3.6-27b",
+      "vendor": "Thinking Machines",
+      "family": "qwen",
+      "provider": "openai-compatible",
+      "gateway": "tinker",
+      "model": "qwen3.6-27b",
+      "contextWindow": 64000,
+      "costClass": "low",
+      "strengths": ["agentic-coding", "tool-use", "fast-inference"],
+      "notes": "Cheapest Tinker candidate for the fast gate path; use when latency/cost matter most."
+    }
+  ]
+}

package/openapi/openapi.yaml

@@ -751,6 +751,34 @@ paths:
                 $ref: '#/components/schemas/FunnelAnalyticsResponse'
         '401':
           description: Unauthorized
+  /v1/analytics/losses:
+    get:
+      operationId: getLossAnalytics
+      parameters:
+        - in: query
+          name: window
+          schema:
+            type: string
+            enum: [today, 7d, 30d, lifetime]
+        - in: query
+          name: timezone
+          schema:
+            type: string
+        - in: query
+          name: now
+          schema:
+            type: string
+            format: date-time
+      responses:
+        '200':
+          description: Ranked buyer-loss and revenue-opportunity analysis for the active analytics window
+          content:
+            application/json:
+              schema:
+                type: object
+                additionalProperties: true
+        '401':
+          description: Unauthorized
   /v1/dashboard:
     get:
       operationId: getDashboard
@@ -848,10 +876,79 @@ paths:
           application/json:
             schema:
               type: object
-              required: [toolName]
               properties:
                 toolName:
                   type: string
+                  description: Tool name is optional when provider-native tool call payload is supplied.
+                provider:
+                  type: string
+                model:
+                  type: string
+                providerToolCall:
+                  type: object
+                  additionalProperties: true
+                toolCall:
+                  type: object
+                  additionalProperties: true
+                toolUse:
+                  type: object
+                  additionalProperties: true
+                content:
+                  type: array
+                  items:
+                    type: object
+                    additionalProperties: true
+                input:
+                  type: object
+                  additionalProperties: true
+                arguments:
+                  type: object
+                  additionalProperties: true
+                method:
+                  type: string
+                params:
+                  type: object
+                  additionalProperties: true
+                mcp:
+                  type: object
+                  additionalProperties: true
+                mcpToolCall:
+                  type: object
+                  additionalProperties: true
+                usage:
+                  type: object
+                  additionalProperties: true
+                tokenEstimate:
+                  type: number
+                costUsd:
+                  type: number
+                budget:
+                  type: object
+                  additionalProperties: true
+                workflowPattern:
+                  type: string
+                  enum: [single_action, chaining, routing, parallelization, evaluator-optimizer, agent]
+                workflow:
+                  type: object
+                  additionalProperties: true
+                goal:
+                  type: string
+                tools:
+                  type: array
+                  items:
+                    type: string
+                branches:
+                  type: array
+                  items:
+                    type: string
+                steps:
+                  type: array
+                  items:
+                    type: string
+                routes:
+                  type: array
+                  items:
+                    type: string
                 command:
                   type: string
                 filePath:
@@ -868,6 +965,25 @@ paths:
                   type: boolean
                 requireVersionNotBehindBase:
                   type: boolean
+                workflowDispatch:
+                  type: object
+                  description: Evidence required before running `gh workflow run` or another environment-specific workflow dispatch.
+                  properties:
+                    environment:
+                      type: string
+                      description: Requested environment such as dev, staging, beta, or release.
+                    workflow:
+                      type: string
+                      description: Expected workflow file or workflow name.
+                    ref:
+                      type: string
+                      description: Expected branch or ref passed to the workflow dispatch command.
+                    sha:
+                      type: string
+                      description: Expected HEAD SHA to verify before and after dispatch.
+                    job:
+                      type: string
+                      description: Expected job name to verify before reporting the workflow URL.
       responses:
         '200':
           description: Persisted workflow-sentinel recommendation with decision-control metadata and actionId
@@ -1121,7 +1237,7 @@ paths:
           description: Comma-separated tags that must all be present on a lesson.
       responses:
         '200':
-          description: Searchable promoted lessons with linked corrective actions, prevention rules, and auto-gates
+          description: Searchable promoted lessons with linked corrective actions, prevention rules, and auto-promoted checks
         '401':
           description: Unauthorized
   /v1/search: