thumbgate 1.27.16 → 1.27.18

package/config/model-candidates.json

@@ -4,8 +4,8 @@
   "workloads": {
     "pretool-gating": {
       "label": "PreTool gating",
-      "summary": "Fast, reliable gate judgments for tool-use and agentic coding decisions before commands run. Hybrid local-cloud candidates (e.g. perplexity/hybrid-local) excel here for privacy + low latency on sensitive paths.",
-      "desiredStrengths": ["agentic-coding", "tool-use", "reliability", "privacy", "fast-inference"],
+      "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",
@@ -43,8 +43,8 @@
     },
     "cheap-fast-path": {
       "label": "Cheap fast path",
-      "summary": "Low-cost first-pass model for cheap approval triage before escalating ambiguous work. Perplexity hybrid-local is ideal: on-device for speed/privacy, escalate only when needed via orchestrator.",
-      "desiredStrengths": ["agentic-coding", "tool-use", "fast-inference", "privacy", "cost-efficiency"],
+      "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",
@@ -60,8 +60,8 @@
     },
     "dashboard-analysis": {
       "label": "Dashboard and dataset analysis",
-      "summary": "Evaluate frontier models for dataset analysis, chart generation, dashboard planning, and proof-backed insight quality before routing expensive analytical work. Perplexity hybrid excels for sensitive lessons/feedback data (local for privacy, cloud for depth).",
-      "desiredStrengths": ["data-analysis", "dashboard-creation", "charting", "long-context", "reliability", "privacy"],
+      "summary": "Evaluate frontier models for dataset analysis, chart generation, dashboard planning, and proof-backed insight quality before routing expensive analytical work.",
+      "desiredStrengths": ["data-analysis", "dashboard-creation", "charting", "long-context", "reliability"],
       "targetContextWindow": 200000,
       "benchmarkCommands": [
         "npx thumbgate eval --from-feedback --json --min-score=0",
@@ -77,27 +77,6 @@
         "costPerAnalysisUsd"
       ]
     },
-    "claw-style-enterprise-agent": {
-      "label": "Claw-style enterprise agent governance",
-      "summary": "Governance, gating, and feedback for autonomous 'claw-style' agents (Automation Anywhere EnterpriseClaw, Nvidia OpenShell-inspired) that have device file system access, runtime dynamic tool creation, screen/UI interaction, and multi-platform orchestration. Especially relevant for on-prem/air-gapped/hybrid enterprise data realities.",
-      "desiredStrengths": ["agentic-coding", "tool-use", "reliability", "security", "orchestration", "audit-trail", "privacy"],
-      "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",
-        "falsePositiveRate",
-        "agentIdentitySeparation",
-        "dynamicToolSafety",
-        "screenInteractionAudit",
-        "orchestrationCompliance",
-        "medianLatencyMs",
-        "costPer1kActionsUsd"
-      ]
-    },
     "tokenizer-brittleness": {
       "label": "Tokenizer brittleness and byte-level robustness",
       "summary": "Evaluate models for malformed JSONL, Unicode confusables, stack traces, secrets, SQL snippets, file paths, and code-symbol-heavy inputs before routing log, code, or security workloads.",
@@ -116,23 +95,6 @@
         "medianLatencyMs",
         "memoryBandwidthEstimate"
       ]
-    },
-    "self-improving-agent-skill-synthesis": {
-      "label": "Self-improving agent skill synthesis",
-      "summary": "Evaluate models on their ability to self-improve, generate safe Markdown skills, and correctly check synthesized skills against existing gate/prevention rules.",
-      "desiredStrengths": ["agentic-coding", "tool-use", "self-evolving", "skill-synthesis", "reliability"],
-      "targetContextWindow": 128000,
-      "benchmarkCommands": [
-        "npx thumbgate eval --from-feedback --json --min-score=0",
-        "node scripts/gate-eval.js run"
-      ],
-      "metrics": [
-        "passRate",
-        "skillSafetyCheckAccuracy",
-        "ruleBypassDetectionRate",
-        "medianLatencyMs",
-        "costPerEvolutionUsd"
-      ]
     }
   },
   "candidates": [
@@ -252,83 +214,6 @@
       "costClass": "low",
       "strengths": ["agentic-coding", "tool-use", "fast-inference"],
       "notes": "Cheapest Tinker candidate for the fast gate path; use when latency/cost matter most."
-    },
-    {
-      "id": "perplexity/hybrid-local-cloud",
-      "vendor": "Perplexity",
-      "family": "hybrid",
-      "provider": "perplexity",
-      "model": "hybrid-local-cloud-orchestrator",
-      "contextWindow": 200000,
-      "costClass": "variable",
-      "strengths": ["agentic-coding", "tool-use", "privacy", "cost-efficiency", "fast-inference", "long-context", "reliability"],
-      "notes": "Perplexity hybrid local-cloud inference orchestrator (announced Computex 2026, part of Personal Computer). Autonomously routes: sensitive/privacy work to local on-device models, complex reasoning to frontier cloud. High-ROI for pretool-gating (local fast/privacy path), cheap-fast-path, and dashboard-analysis with sensitive data/lessons. Pair with ThumbGate hybrid-routing gates (see adapters/perplexity/HYBRID.md). Coming July 2026 for local inference."
-    },
-    {
-      "id": "perplexity/hybrid-local",
-      "vendor": "Perplexity",
-      "family": "hybrid",
-      "provider": "perplexity",
-      "model": "local-inference",
-      "contextWindow": 128000,
-      "costClass": "low",
-      "strengths": ["fast-inference", "privacy", "tool-use", "reliability"],
-      "notes": "Local-only mode of Perplexity hybrid for on-device pre-action gating, sensitivity classification, and low-latency checks on AI PCs (Intel, NVIDIA). Escalate via orchestrator for full capability. Use for cheap-fast-path and pretool-gating workloads."
-    },
-    {
-      "id": "automation-anywhere/enterprise-claw",
-      "vendor": "Automation Anywhere",
-      "family": "claw-style",
-      "provider": "automation-anywhere",
-      "model": "enterprise-claw",
-      "contextWindow": 200000,
-      "costClass": "variable",
-      "strengths": ["agentic-coding", "tool-use", "orchestration", "audit-trail", "security", "on-prem", "airgap", "dynamic-tool-creation", "screen-interaction"],
-      "notes": "Claw-style autonomous enterprise agents (EnterpriseClaw, inspired by Nvidia OpenShell). Device-level access, runtime tool creation, screen/UI interaction, multi-platform orchestration. Governance infrastructure (ThumbGate) is explicitly called out as catching up. High-ROI for enterprise on-prem/hybrid use cases. Pair with perplexity/hybrid for inference routing. See adapters/claw/CLAW.md and new gate templates."
-    },
-    {
-      "id": "nvidia/openshell-claw",
-      "vendor": "NVIDIA",
-      "family": "claw-style",
-      "provider": "nvidia",
-      "model": "openshell",
-      "contextWindow": 128000,
-      "costClass": "medium",
-      "strengths": ["agentic-coding", "tool-use", "dynamic-tool-creation", "screen-interaction", "on-prem", "self-evolving"],
-      "notes": "Nvidia OpenShell runtime for autonomous self-evolving claw-style agents (basis for Automation Anywhere EnterpriseClaw). Run locally/on-prem. ThumbGate provides the missing governance layer (gates, feedback, rules). Use with hybrid local-cloud for full enterprise deployment."
-    },
-    {
-      "id": "nousresearch/hermes-3-llama-3.1-70b",
-      "vendor": "Nous Research",
-      "family": "hermes",
-      "provider": "openrouter",
-      "model": "hermes-3-llama-3.1-70b",
-      "contextWindow": 128000,
-      "costClass": "medium",
-      "strengths": ["agentic-coding", "tool-use", "self-evolving", "skill-synthesis", "multi-platform", "reasoning"],
-      "notes": "Nous Research's Hermes 3 model family, optimized for autonomous reasoning, tool-use, and self-improvement loops. Benchmark KV caching, trace-length constraints, and dynamic skill synthesis validation rules before deploying."
-    },
-    {
-      "id": "anthropic/claude-opus-4-8",
-      "vendor": "Anthropic",
-      "family": "claude",
-      "provider": "anthropic",
-      "model": "claude-opus-4.8",
-      "contextWindow": 200000,
-      "costClass": "high",
-      "strengths": ["agentic-coding", "tool-use", "reliability", "long-context", "long-horizon-coding", "multi-agent"],
-      "notes": "Frontier coding model with 56.7 Coding Index. Grew 140% WoW on OpenRouter to 1.33T tokens. Ideal for long trace reviews and complex reasoning steps, despite higher latency and cost."
-    },
-    {
-      "id": "google/gemini-3.1-pro-preview",
-      "vendor": "Google",
-      "family": "gemini",
-      "provider": "google",
-      "model": "gemini-3.1-pro-preview",
-      "contextWindow": 2000000,
-      "costClass": "medium",
-      "strengths": ["agentic-coding", "tool-use", "reliability", "long-context", "data-analysis", "cost-efficiency"],
-      "notes": "Google's 3.1 Pro Preview model with 55.5 Coding Index and massive 2M token context window. Strong choice for complex RAG tasks, repository-scale auditing, and long-context trace analysis."
     }
   ]
 }

package/openapi/openapi.yaml

@@ -72,11 +72,6 @@ components:
           description: Optional domain tags. If omitted, ThumbGate infers one from the feedback text before promotion.
         skill:
           type: string
-        source:
-          type: string
-          enum: [chatgpt_gpt]
-          default: chatgpt_gpt
-          description: Attribution marker for ThumbGate analytics. The published ThumbGate GPT should send `chatgpt_gpt` so owner dashboards can distinguish GPT Action calls from local API calls.
     IntentPlanRequest:
       type: object
       required: [intentId]
@@ -94,8 +89,6 @@ components:
         delegationMode:
           type: string
           enum: [off, auto, sequential]
-        enforcePlanQuality:
-          type: boolean
         approved:
           type: boolean
         repoPath:
@@ -887,11 +880,6 @@ paths:
                 toolName:
                   type: string
                   description: Tool name is optional when provider-native tool call payload is supplied.
-                source:
-                  type: string
-                  enum: [chatgpt_gpt]
-                  default: chatgpt_gpt
-                  description: Attribution marker for ThumbGate analytics. The published ThumbGate GPT should send `chatgpt_gpt` so owner dashboards can distinguish GPT Action calls from local API calls.
                 provider:
                   type: string
                 model: