refactor(runtime): enforce canonical model execution boundary

.github/workflows/ci.yml

@@ -81,21 +81,21 @@ jobs:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           repository: tangle-network/agent-sdk
-          ref: fc00613985c86f917a679c9bf957e6af8cbedb7c # @tangle-network/agent-interface@0.43.0
+          ref: 36b8e3ecc57e9a62130eb2b0fdc1c3b5be01b338 # @tangle-network/agent-interface@0.43.1
           path: .cohort/agent-sdk
           persist-credentials: false
 
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           repository: tangle-network/agent-eval
-          ref: 42a044f29725f0587d6ab0d9783f4e889d5fb8f7 # v0.143.0
+          ref: cd4c3195fec79c79976ce2fbe43f307107544798 # v0.144.4
           path: .cohort/agent-eval
           persist-credentials: false
 
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           repository: tangle-network/agent-knowledge
-          ref: 3a41d557a5549c324fad2b101d02f0bf30657cac # v7.0.8
+          ref: b5c82eb9a508b13a32e14353a867c197cda517b8 # v7.0.11
           path: .cohort/agent-knowledge
           persist-credentials: false
 
@@ -132,6 +132,8 @@ jobs:
         run: pnpm run build
 
       - name: Verify agent-bench against current agent-runtime
+        env:
+          AGENT_BENCH_PACKAGE_TEST_TIMEOUT_MS: 900000
         run: pnpm run verify:bench
 
   official-optimizers:

.github/workflows/publish.yml

@@ -135,23 +135,23 @@ jobs:
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           repository: tangle-network/agent-sdk
-          ref: fc00613985c86f917a679c9bf957e6af8cbedb7c # @tangle-network/agent-interface@0.43.0
+          ref: 36b8e3ecc57e9a62130eb2b0fdc1c3b5be01b338 # @tangle-network/agent-interface@0.43.1
           path: .cohort/agent-sdk
           persist-credentials: false
 
       - name: Check out exact Eval release
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           repository: tangle-network/agent-eval
-          ref: 42a044f29725f0587d6ab0d9783f4e889d5fb8f7 # v0.143.0
+          ref: cd4c3195fec79c79976ce2fbe43f307107544798 # v0.144.4
           path: .cohort/agent-eval
           persist-credentials: false
 
       - name: Check out exact Knowledge release
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           repository: tangle-network/agent-knowledge
-          ref: 3a41d557a5549c324fad2b101d02f0bf30657cac # v7.0.8
+          ref: b5c82eb9a508b13a32e14353a867c197cda517b8 # v7.0.11
           path: .cohort/agent-knowledge
           persist-credentials: false
 
@@ -169,6 +169,8 @@ jobs:
         run: pnpm run verify:primeintellect
 
       - name: Verify agent-bench against this release
+        env:
+          AGENT_BENCH_PACKAGE_TEST_TIMEOUT_MS: 900000
         run: pnpm run verify:bench
 
       - name: Package and verify the exact Runtime archive
@@ -296,6 +298,8 @@ jobs:
         run: pnpm --filter @tangle-network/agent-bench run wait:published-dependencies
 
       - name: Verify packed agent-bench against published dependencies
+        env:
+          AGENT_BENCH_PACKAGE_TEST_TIMEOUT_MS: 900000
         run: pnpm run verify:bench:published
 
       - name: Verify tag/version lock

CHANGELOG.md

@@ -1,5 +1,34 @@
 # Changelog
 
+## 0.129.0
+
+- Require Agent Eval 0.144.4, Agent Interface 0.43.1, Agent Knowledge 7.0.11, and Sandbox 0.19.1 as one dependency set, and route the official-optimizer callback through Runtime's exact `AgentProfile` execution path.
+- Reject model, provider, reasoning, prompt, tool, resource, harness, and generation-setting conflicts before transport; consumers must declare those fields in the profile.
+- Require `defineLeaderboard` callers to supply an exact `baseProfile`; remove its `modelBackend` override so each matrix cell's profile remains the only model authority.
+- Require generic coder, researcher, and supervised-knowledge paths to receive complete profiles; remove harness/model overlays and MCP environment alias ladders.
+- Resolve Sandbox execution only from `AgentProfile.harness`; a backend type may confirm that choice but cannot replace it.
+- Parse, detach, and deeply freeze a complete `AgentProfile` before Scope, registry, nested-driver, or personified execution can honor any built-in or caller-supplied executor.
+- Require `driverChild(profile, ...)` and `worktreeLoopRunner({ rootProfile, ... })`; remove name-only driver and personified-root shortcuts.
+- Remove the public `runLocalHarness` process shortcut; use `createWorktreeCliExecutor({ profile, ... })` directly or `createExecutor({ backend: 'cli-worktree', ... })` inside a supervised run so the CLI and model derive from one exact `AgentProfile`.
+- Keep missing token usage and billed cost unknown, and report reasoning-token usage when the provider supplies it.
+
+### Removed public execution shortcuts
+
+- Remove `driverLoopGenerator`, `DriverLoopGeneratorOptions`, `buildDriverSystem`, and `researchDriverNote`; use `improve({ surface: 'code', code: { profile, executorForWorktree, buildPrompt } })` or `agenticGenerator` so every authoring call uses the exact declared profile through Runtime.
+- Remove `AGENTIC_PROFILE_RESOURCE_ROOT`; profile resources now travel inside the exact `AgentProfile` and Runtime materializes them at execution.
+- Remove `AgentBackendKind`, `ResolveAgentBackendOptions`, `resolveAgentBackend`, `createOpenAICompatibleBackend`, and `BackendRetryPolicy`; select a built-in transport with `ExecutorConfig` and `createExecutor(...)`, bind it to an exact profile with `createProfileExecutionBackend(...)` for `runAgentTaskStream` or conversation APIs, or pass both directly to `streamAgentTurn`.
+- Move retry configuration into `AgentProfile.model.metadata.retry`, including attempt count, per-attempt timeout, retryable status codes, exponential-backoff bounds, and jitter, so deleting the parallel backend API does not delete caller control.
+- Remove the public Router chat and tool-loop API (`RouterConfig`, `RouterChatResult`, `RouterChatToolsResult`, `RouterToolCall`, `RouterToolLoopResult`, `routerChatWithUsage`, `routerChatWithTools`, `streamRouterChatWithTools`, `routerToolLoop`, and `routerBrain`); use `createExecutor(...)` plus `streamAgentTurn` and `collectAgentTurn`, use `ToolLoopToolCall` for the provider-neutral tool-call record, use `profileChatClient` for Eval integrations, or use `supervise(...)` for a profiled supervisor.
+- Replace `ShotPersona` with `ShotSpec.profile`, which carries a complete exact profile instead of a prompt/model-only override.
+- Remove `canonicalizeAuthoredProfile`; inputs must already satisfy `agentProfileSchema`, with no flat legacy spelling that Runtime silently repairs.
+- Remove `authoredWorker`; use `workerFromBackend(...)`, which admits each complete profile through the same executor path as every other supervised worker.
+- Remove the public `chatCompletionsTransport` constructor; inject an offline completion function through `ChatTransportExecutorOptions.complete`, or use an exact profile with `createExecutor(...)` for real execution.
+- Remove `createPrimeIntellectBackend`; use `primeIntellectExecutorConfig(context)` with `createExecutor(...)`, then execute an exact profile through `streamAgentTurn` or bind it with `createProfileExecutionBackend(...)` when an `AgentExecutionBackend` is required.
+- Remove `uiAuditorProfile`, `createInProcessUiAuditClient`, and their browser/judge option types; author a normal `AgentProfile`, execute it through Runtime, and keep using `UiAuditTask`, `encodeAuditTaskEnvelope`, `formatAuditorPrompt`, `parseAuditorEvents`, and `createUiAuditorValidator` from `/profiles`.
+- Remove the bundled UI-audit example and the Playwright peer dependency because browser placement belongs to the caller-owned execution environment, not a profile-specific Runtime client.
+- Remove direct model injection from `SuperviseOptions`, `RunGraphOptions`, and `SupervisorAgentDeps`, and remove `driverAgent`, `DriverAgentOptions`, and `ToolLoopChat` from `/kernel`; production supervisor calls now derive model execution only from the exact root `AgentProfile`, while deterministic scripted calls live under the explicit `/testing` entrypoint.
+- Bind every root Router turn to one Runtime-authored cancellation signal, call id, and correlation id; preserve provider-observed model, prompt-cache, retry-attempt, reasoning-token, and billed-cost evidence; meter a mismatched model response before refusing its output; and reuse one nonempty idempotency key across all physical retries of the logical call.
+
 ## 0.128.0
 
 ### chat-transport executor: workers on a bare chat-completions transport
@@ -151,7 +180,7 @@ The fix therefore lives in agent-eval, but a consumer only receives it if this p
 
 BREAKING. `LocalHarness` was a private three-member vocabulary (`'claude' | 'codex' | 'opencode'`) that spelled one harness differently from every other layer in the stack. It is now a narrowing of the shared `HarnessType`: **`'claude'` is renamed to `'claude-code'`**. `claude` remains the EXECUTABLE name and lives only in the harness table's `command` field.
 
-Callers to update: `runLocalHarness({ harness })`, `harnessInvocation(harness, …)`, `runWorktreeHarness({ harness })`, `agenticGenerator({ harness })`, `driverLoopGenerator({ harness })`, `createInProcessExecutor({ harnesses })`, `AuthoredHarness.harness`, and the `AGENT_RUNTIME_LOCAL_HARNESSES` env list. Anything that passed `'claude'` passes `'claude-code'`; `codex` and `opencode` are unchanged.
+Callers to update: `runLocalHarness({ harness })`, `harnessInvocation(harness, …)`, `runWorktreeHarness({ harness })`, `agenticGenerator({ harness })`, `createInProcessExecutor({ harnesses })`, `AuthoredHarness.harness`, and the `AGENT_RUNTIME_LOCAL_HARNESSES` env list. Anything that passed `'claude'` passes `'claude-code'`; `codex` and `opencode` are unchanged.
 
 Deleting the alias removed `materializerHarness()` outright — a `LocalHarness` is now handed straight to the profile materializer with no translation.
 

README.md

@@ -31,8 +31,15 @@ That file defines the scripted `worker`, `output`, and `validator` used below so
 Replace the scripted worker with a sandbox, CLI bridge, or router backend without changing the driver.
 
 ```ts
+import type { AgentProfile } from '@tangle-network/agent-interface'
 import { inProcessSandboxClient, runAgentRounds } from '@tangle-network/agent-runtime/kernel'
 
+const noteWriterProfile = {
+  name: 'note-writer',
+  harness: 'cli-base',
+  model: { provider: 'scripted', default: 'scripted/note-writer' },
+} satisfies AgentProfile
+
 const result = await runAgentRounds<Task, Note, 'refine' | 'pick-winner' | 'fail'>({
   task: { prompt: 'Write a one-line release note for one-click restore.' },
   driver: {
@@ -47,7 +54,7 @@ const result = await runAgentRounds<Task, Note, 'refine' | 'pick-winner' | 'fail
     decide: (history) =>
       history.some((shot) => shot.verdict?.valid) ? 'pick-winner' : history.length < 3 ? 'refine' : 'fail',
   },
-  agentRun: { profile: { name: 'note-writer' } as AgentProfile, taskToPrompt: (t) => t.prompt },
+  agentRun: { profile: noteWriterProfile, taskToPrompt: (t) => t.prompt },
   output,    // parses the worker's event stream into { note }
   validator, // pass/fail check: does the note mention "rollback"?
   ctx: { sandboxClient: worker },
@@ -60,9 +67,9 @@ Run it from a clone of this repo and you get exactly this:
 ```bash
 $ pnpm i && pnpm build
 $ pnpm tsx examples/quickstart/quickstart.ts
-shot 0: reject: "Shipped one-click restore."
-shot 1: PASS: "Shipped one-click restore with an instant rollback path."
-decision: pick-winner: winner: shot 1
+shot 0: reject — "Shipped one-click restore."
+shot 1: PASS — "Shipped one-click restore with an instant rollback path."
+decision: pick-winner — winner: shot 1
 ```
 
 The annotated version is [`examples/driver-loop`](./examples/driver-loop).
@@ -119,6 +126,7 @@ const result = await supervise(
   {
     name: 'supervisor',
     harness: 'cli-base',
+    model: { provider: 'tangle-router', default: process.env.TANGLE_MODEL! },
     prompt: {
       systemPrompt: 'Delegate to workers; do not solve the task yourself.',
     },
@@ -137,28 +145,55 @@ The profile is never changed.
 
 ```ts
 import { improve, officialGepa } from '@tangle-network/agent-runtime'
-import { canonicalCandidateDigest } from '@tangle-network/agent-interface'
+import { profileOptimizerModelCall } from '@tangle-network/agent-runtime/kernel'
+import {
+  type AgentProfile,
+  canonicalAgentProfileDigest,
+  canonicalCandidateDigest,
+} from '@tangle-network/agent-interface'
 
 const executionRef = canonicalCandidateDigest({
   deployment: process.env.AGENT_DEPLOYMENT_SHA!,
   model: process.env.AGENT_MODEL!,
   tools: process.env.AGENT_TOOLSET_SHA!,
 })
 
+const optimizerProfile = {
+  name: 'support-prompt-optimizer',
+  harness: 'cli-base',
+  model: {
+    provider: 'tangle-router',
+    default: process.env.OPTIMIZER_MODEL!,
+    metadata: { maxTokens: 16_384 },
+  },
+} satisfies AgentProfile
+const optimizerPricing = {
+  inputUsdPerMillion: Number(process.env.OPTIMIZER_INPUT_USD_PER_MILLION),
+  outputUsdPerMillion: Number(process.env.OPTIMIZER_OUTPUT_USD_PER_MILLION),
+}
 const optimizer = {
-  model: process.env.OPTIMIZER_MODEL!,
-  baseUrl: process.env.OPTIMIZER_BASE_URL!,
-  apiKey: process.env.OPTIMIZER_API_KEY!,
+  model: optimizerProfile.model.default,
+  call: profileOptimizerModelCall({
+    profile: optimizerProfile,
+    context: 'support-prompt optimizer',
+    executor: {
+      backend: 'router',
+      routerBaseUrl: process.env.OPTIMIZER_BASE_URL!,
+      routerKey: process.env.OPTIMIZER_API_KEY!,
+    },
+    pricing: optimizerPricing,
+  }),
+  callRef: canonicalCandidateDigest({
+    profile: canonicalAgentProfileDigest(optimizerProfile),
+    deployment: process.env.OPTIMIZER_DEPLOYMENT_SHA!,
+  }),
   budget: {
     maxCostUsd: 10,
     maxRequests: 50,
     maxRequestBytes: 2_000_000,
     maxResponseBytes: 2_000_000,
     maxOutputTokensPerRequest: 16_384,
-    pricing: {
-      inputUsdPerMillion: Number(process.env.OPTIMIZER_INPUT_USD_PER_MILLION),
-      outputUsdPerMillion: Number(process.env.OPTIMIZER_OUTPUT_USD_PER_MILLION),
-    },
+    pricing: optimizerPricing,
   },
 }
 
@@ -207,7 +242,7 @@ There is no local fallback.
 Install its optional Python process before using it:
 
 ```bash
-python -m pip install "agent-eval-rpc==0.143.0"
+python -m pip install "agent-eval-rpc==0.144.4"
 python -m pip install "gepa[full]==0.1.4"
 ```
 
@@ -221,14 +256,14 @@ python -m pip install "gepa[full] @ git+https://github.com/gepa-ai/gepa.git@f919
 Use `officialSkillOpt(...)` for Microsoft's SkillOpt:
 
 ```bash
-python -m pip install "agent-eval-rpc==0.143.0"
+python -m pip install "agent-eval-rpc==0.144.4"
 python -m pip install "skillopt @ git+https://github.com/microsoft/SkillOpt.git@61735e3922efc2b90c6d6cab561e62e98452ca90"
 ```
 
 SkillOpt 0.2.0's published wheel omits prompt files required by `ReflACTTrainer`, so the tested SkillOpt source revision remains necessary.
-SkillOpt and GEPA's standard reflection engine require `optimizer: { model, baseUrl, apiKey, budget }`.
+SkillOpt and GEPA's standard reflection engine require `optimizer: { model, call, callRef, budget }`.
 Agent-based GEPA engines may own their model connection instead.
-Agent Eval proxies those model calls, enforces the nested budget, and records their cost.
+Runtime owns those model calls through one exact `AgentProfile`; Agent Eval enforces the nested budget and records their measured cost and execution evidence without receiving provider credentials.
 `costCeiling` is the total limit for optimizer calls, candidate runs, judges, and final scoring.
 Runtime returns `hold` when any part of that cost is unknown.
 Runtime rejects a reported total above the limit.
@@ -269,7 +304,7 @@ It uses Runtime's isolated git worktrees and coding-agent candidate execution:
 ```ts
 const result = await improve({
   surface: 'code',
-  code: { repoRoot, baseRef, generator },
+  code: { repoRoot, baseRef, profile, generator },
   scenarios,
   judge,
   agent,
@@ -414,13 +449,27 @@ Here, `runProductAgent` is the application's existing entry point, not another l
 
 ```ts
 import {
-  createPrimeIntellectBackend,
+  primeIntellectExecutorConfig,
   runPrimeIntellectProgram,
 } from '@tangle-network/agent-runtime/primeintellect'
+import {
+  collectAgentTurn,
+  createExecutor,
+  streamAgentTurn,
+} from '@tangle-network/agent-runtime/kernel'
 
 await runPrimeIntellectProgram(async (episode) => {
-  const backend = createPrimeIntellectBackend(episode)
-  return runProductAgent({ task: episode.task, backend })
+  const profile = makeProductProfile({ model: episode.model.name })
+  return collectAgentTurn(
+    streamAgentTurn(
+      {
+        kind: 'executor',
+        profile,
+        factory: createExecutor(primeIntellectExecutorConfig(episode)),
+      },
+      episode.task.prompt,
+    ),
+  )
 })
 ```
 

bench/CHANGELOG.md

@@ -1,5 +1,9 @@
 # Changelog
 
+## 0.7.2
+
+- Consume Runtime 0.129.0, Eval 0.144.4, Interface 0.43.1, Knowledge 7.0.11, and Sandbox 0.19.1 so benchmark model calls use the exact-profile execution boundary and the released optimizer callback contract without loading duplicate agent contracts.
+
 ## 0.7.1
 
 - Consume Runtime 0.126.0 with Eval 0.143.0 and Knowledge 7.0.8, so campaign cost remains observed, estimated, or explicitly uncaptured across the complete benchmark dependency tree.

bench/gen3-config.json

@@ -41,7 +41,8 @@
   "repsPerInstance": 2,
   "premeasuredBaselinePath": "/tmp/claude-1000/-home-drew-code-supervisor-lab/f06fd156-042a-4ef9-bd88-f2ec7f52b90c/scratchpad/hh/gen3/premeasured-baseline.json",
   "maxShots": 3,
-  "proposerHarness": "claude-code",
+  "proposerHarness": "pi",
+  "proposerProfile": "default-author.profile.json",
   "proposerTimeoutMs": 2400000,
   "analystModels": [
     "glm-5.2",
@@ -77,19 +78,22 @@
     {
       "name": "default-author",
       "profile": "default-author.profile.json",
-      "harness": "claude"
+      "harness": "pi",
+      "model": "glm-5.2"
     },
     {
       "name": "mechanics-author",
       "profile": "default-author.profile.json",
-      "harness": "claude",
+      "harness": "pi",
+      "model": "glm-5.2",
       "diagnosisSlice": "mechanics",
       "lens": "Focus on MECHANICS: worker lifecycle, sandbox/clone contracts, settlement and delivery paths. Prefer code-path fixes over prompt wording."
     },
     {
       "name": "prompts-author",
       "profile": "default-author.profile.json",
-      "harness": "claude",
+      "harness": "pi",
+      "model": "glm-5.2",
       "diagnosisSlice": "prompts",
       "lens": "Focus on PROMPTS: worker/brain instruction wording, placement guidance, self-check discipline. Prefer prompt/instruction changes over code-path rewrites."
     }

bench/gen4-config.json

@@ -37,11 +37,12 @@
     "tangle-router.env"
   ],
   "generations": 1,
-  "populationSize": 4,
+  "populationSize": 3,
   "repsPerInstance": 2,
   "premeasuredBaselinePath": "/tmp/claude-1000/-home-drew-code-supervisor-lab/f06fd156-042a-4ef9-bd88-f2ec7f52b90c/scratchpad/hh/gen4/premeasured-baseline.json",
   "maxShots": 3,
-  "proposerHarness": "claude-code",
+  "proposerHarness": "pi",
+  "proposerProfile": "default-author.profile.json",
   "proposerTimeoutMs": 2400000,
   "analystModels": [
     "glm-5.2",
@@ -74,24 +75,23 @@
   "costGuardRatio": 1.2,
   "dispatchTimeoutMs": 7200000,
   "proposers": [
-    {
-      "name": "claude-author",
-      "profile": "default-author.profile.json",
-      "harness": "claude"
-    },
     {
       "name": "glm-author",
-      "harness": "opencode",
-      "model": "zai-coding-plan/glm-5.2"
+      "profile": "default-author.profile.json",
+      "harness": "pi",
+      "model": "glm-5.2"
     },
     {
-      "name": "codex-author",
-      "harness": "codex"
+      "name": "deepseek-author",
+      "profile": "deepseek-author.profile.json",
+      "harness": "pi",
+      "model": "deepseek-v4-flash"
     },
     {
       "name": "merge-author",
       "profile": "default-author.profile.json",
-      "harness": "claude",
+      "harness": "pi",
+      "model": "glm-5.2",
       "merge": true
     }
   ],