@tangle-network/agent-runtime 0.103.1 → 0.105.0

package/dist/{types-DK11_O4L.d.ts → types-BCoemcXU.d.ts}

@@ -443,7 +443,7 @@ interface Driver<Task, Output, Decision> {
     /**
      * Optional: the driver AUTHORS the winner instead of the kernel's argmax. The
      * kernel consults this at finalize ONLY when the caller did not pass an explicit
-     * `selectWinner` to runLoop. Return the driver-declared winner (e.g. from a
+     * `selectWinner` to runAgentRounds. Return the driver-declared winner (e.g. from a
      * `select` topology move) or `undefined` to fall through to the default
      * (best-valid-score, earliest index). This is the SELECTOR role made
      * agent-authorable — the planner runs the selection, not the kernel.
@@ -522,7 +522,7 @@ interface SandboxClient {
     }>;
 }
 /**
- * Opt-in box-lineage controls for `runLoop`. Default OFF — with both flags
+ * Opt-in box-lineage controls for `runAgentRounds`. Default OFF — with both flags
  * unset the kernel's per-iteration behavior is byte-identical to acquiring a
  * fresh box, streaming once, and tearing it down. The independence of N fresh
  * boxes (e.g. `random@k`) is a compute-control invariant; these flags must
@@ -746,7 +746,7 @@ interface LoopTeardownFailedPayload {
     reason: string;
 }
 /**
- * Execution context for `runLoop`: the sandbox client the kernel creates boxes through, plus optional runtime hooks.
+ * Execution context for `runAgentRounds`: the sandbox client the kernel creates boxes through, plus optional runtime hooks.
  *
  * @experimental
  */
@@ -799,4 +799,4 @@ interface ExecCtx {
     parentSpanId?: string;
 }
 
-export { type AgentRunSpec as A, type LoopIterationEndedPayload as B, type LoopIterationStartedPayload as C, type Driver as D, type ExecCtx as E, type LoopPlanDescription as F, type LoopPlanPayload as G, type LoopStartedPayload as H, type Iteration as I, type LoopTeardownFailedPayload as J, type MountManifestEntry as K, type LoopTokenUsage as L, type MountRecorder as M, type RunProvenance as N, type OutputAdapter as O, type ValidationCtx as P, type RuntimeHooks as R, type SandboxClient as S, type Validator as V, type SelectionReceipt as a, type LoopSandboxPlacement as b, type LoopTraceEvent as c, type RuntimeDecisionEvidenceRef as d, type RuntimeDecisionKind as e, type RuntimeDecisionPoint as f, type RuntimeHookContext as g, type RuntimeHookErrorContext as h, type RuntimeHookEvent as i, type RuntimeHookPhase as j, type RuntimeHookTarget as k, type RuntimeRunHandle as l, type RuntimeRunPersistenceAdapter as m, type RuntimeRunRow as n, composeRuntimeHooks as o, defineRuntimeHooks as p, notifyRuntimeDecisionPoint as q, notifyRuntimeHookEvent as r, startRuntimeRun as s, type LoopTraceEmitter as t, type LoopWinner as u, type LoopLineageOptions as v, type LoopResult as w, type LoopDecisionPayload as x, type LoopEndedPayload as y, type LoopIterationDispatchPayload as z };
+export { type AgentRunSpec as A, type LoopIterationEndedPayload as B, type LoopIterationStartedPayload as C, type Driver as D, type ExecCtx as E, type LoopPlanDescription as F, type LoopPlanPayload as G, type LoopStartedPayload as H, type Iteration as I, type LoopTeardownFailedPayload as J, type MountManifestEntry as K, type LoopTokenUsage as L, type MountRecorder as M, type RunProvenance as N, type OutputAdapter as O, type ValidationCtx as P, type RuntimeHooks as R, type SelectionReceipt as S, type Validator as V, type SandboxClient as a, type LoopSandboxPlacement as b, type LoopTraceEvent as c, type RuntimeDecisionEvidenceRef as d, type RuntimeDecisionKind as e, type RuntimeDecisionPoint as f, type RuntimeHookContext as g, type RuntimeHookErrorContext as h, type RuntimeHookEvent as i, type RuntimeHookPhase as j, type RuntimeHookTarget as k, type RuntimeRunHandle as l, type RuntimeRunPersistenceAdapter as m, type RuntimeRunRow as n, composeRuntimeHooks as o, defineRuntimeHooks as p, notifyRuntimeDecisionPoint as q, notifyRuntimeHookEvent as r, startRuntimeRun as s, type LoopTraceEmitter as t, type LoopWinner as u, type LoopLineageOptions as v, type LoopResult as w, type LoopDecisionPayload as x, type LoopEndedPayload as y, type LoopIterationDispatchPayload as z };

package/dist/{types-lZTE_LBd.d.ts → types-CvNwMzQt.d.ts}

@@ -1,7 +1,7 @@
 import { ToolSpan, DefaultVerdict } from '@tangle-network/agent-eval';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { BackendType } from '@tangle-network/sandbox';
-import { L as LoopTokenUsage, R as RuntimeHooks } from './types-DK11_O4L.js';
+import { L as LoopTokenUsage, R as RuntimeHooks } from './types-BCoemcXU.js';
 
 /**
  *
@@ -373,7 +373,7 @@ interface Agent<Task, Out> {
  * meters every runtime identically.
  *
  * Built-in implementations (in `runtime.ts`, NOT variants here): router/inline (a direct
- * Router/HTTP inference call, no box), sandbox (COMPOSES `runLoop` as a leaf, forwarding
+ * Router/HTTP inference call, no box), sandbox (COMPOSES `runAgentRounds` as a leaf, forwarding
  * PR #150's optional `lineage` passthrough — does NOT reinvent checkpoint/fork), cli
  * (Halo/RLM subprocess; `budgetExempt`, excluded from equal-k by construction). A user's
  * own agent (mastra/agno/raw HTTP/anything) is first-class by implementing this interface.
@@ -482,7 +482,7 @@ type Runtime = 'router' | 'inline' | 'sandbox' | 'cli' | (string & {});
  * Resolution (in `runtime.ts`):
  *  - `executor` present        → BYO: use it verbatim (a user's own `Executor`).
  *  - `harness === null`        → router/inline: a direct Router call, no box.
- *  - `harness` is a `BackendType` → sandbox: compose `runLoop` against `profile` on that backend.
+ *  - `harness` is a `BackendType` → sandbox: compose `runAgentRounds` against `profile` on that backend.
  * Fail loud on an unresolvable spec (no executor and an unknown harness).
  */
 interface AgentSpec {
@@ -888,7 +888,7 @@ interface SupervisorOpts {
     readonly now?: () => number;
     readonly signal?: AbortSignal;
     /** Lifecycle stream sink, threaded into the root `Scope` so every `spawn`/settle emits on the
-     *  same `agent.spawn`/`agent.child` stream `runLoop` feeds — one observable recursive tree. */
+     *  same `agent.spawn`/`agent.child` stream `runAgentRounds` feeds — one observable recursive tree. */
     readonly hooks?: RuntimeHooks;
 }
 /** Typed terminal result (M2) — a no-winner is NEVER coerced to a best-effort output. */

package/dist/{worktree-fanout-DiiJWjJO.d.ts → worktree-fanout-DxhAWr5Z.d.ts}

@@ -1,10 +1,10 @@
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { AnalystFinding, DefaultVerdict } from '@tangle-network/agent-eval';
-import { N as NodeId, a as Settled, g as Spend, h as AgentSpec, S as Scope, c as SpawnJournal, b as ResultBlobStore, i as ExecutorRegistry, B as Budget, A as Agent, j as RootHandle, e as SupervisedResult, f as Executor } from './types-lZTE_LBd.js';
-import { I as Iteration, R as RuntimeHooks } from './types-DK11_O4L.js';
+import { N as NodeId, a as Settled, g as Spend, h as AgentSpec, S as Scope, c as SpawnJournal, b as ResultBlobStore, i as ExecutorRegistry, B as Budget, A as Agent, j as RootHandle, e as SupervisedResult, f as Executor } from './types-CvNwMzQt.js';
+import { I as Iteration, R as RuntimeHooks } from './types-BCoemcXU.js';
 import { BackendType } from '@tangle-network/sandbox';
-import { W as WorktreeHarnessResult, G as GitRunner, a as WorktreeCheckRunner, D as DeliverableSpec } from './completion-gate-DLINnrkM.js';
-import { L as LocalHarness, r as runLocalHarness } from './local-harness-CtK9dSny.js';
+import { W as WorktreeHarnessResult, G as GitRunner, a as WorktreeCheckRunner, D as DeliverableSpec } from './completion-gate-BMy5LGoP.js';
+import { L as LocalHarness, r as runLocalHarness } from './local-harness-BDNcl6jI.js';
 
 /**
  *
@@ -806,11 +806,11 @@ interface CoderCheckConstraints {
  * coding-harness CLI (claude / codex / opencode) on its OWN git worktree, surfaced as the open
  * `Executor<Out>` port (`./types`). It is a LEAF executor: it plugs straight into the
  * `Scope`/`Supervisor` recursion and `gateOnDeliverable`, so it IS the canonical recursive path
- * (no `runLoop`/virtual-SandboxInstance shim in between).
+ * (no `runAgentRounds`/virtual-SandboxInstance shim in between).
  *
  * This is a THIN adapter: the physical act (worktree → profile-aware harness invocation → diff →
  * checks → cleanup) lives ONCE in `runWorktreeHarness` (`../../mcp/worktree-harness`), shared with
- * the `runLoop`/coder-delegate `createInProcessExecutor`. This executor only projects that core's
+ * the `runAgentRounds`/coder-delegate `createInProcessExecutor`. This executor only projects that core's
  * result onto the `Executor` port (artifact + spend) and owns the teardown point. The complete
  * profile delivery — direct prompt/model plus materialized file-backed resources — lives there.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-runtime",
-  "version": "0.103.1",
+  "version": "0.105.0",
   "description": "Shared task-lifecycle skeleton for agents: a recursive loop kernel for chat turns, one-shot tasks, and multi-attempt loops, with trace capture and eval-gated self-improvement. Domain behavior lives in adapters; scoring and ship-gates in @tangle-network/agent-eval.",
   "homepage": "https://github.com/tangle-network/agent-runtime#readme",
   "repository": {
@@ -114,7 +114,8 @@
     "verify:bench": "pnpm build && pnpm --filter @tangle-network/agent-bench run typecheck:public && pnpm --filter @tangle-network/agent-bench test && pnpm --filter @tangle-network/agent-bench run verify:package:local-runtime",
     "verify:bench:published": "pnpm build && pnpm --filter @tangle-network/agent-bench run typecheck:public && pnpm --filter @tangle-network/agent-bench test && pnpm --filter @tangle-network/agent-bench run verify:package",
     "verify:package": "pnpm run check:testing-fixture && pnpm run check:skills && node scripts/verify-package-exports.mjs",
-    "verify:primeintellect": "pnpm build && node scripts/verify-primeintellect-v1.mjs",
+    "verify:official-optimizers": "node scripts/verify-official-optimizers.mjs",
+    "verify:primeintellect": "pnpm build && node scripts/verify-primeintellect.mjs",
     "verify:primeintellect:live": "pnpm build && node scripts/verify-primeintellect-live.mjs",
     "docs:api": "pnpm run build && typedoc && node scripts/gen-primitive-catalog.mjs",
     "docs:freshness": "node scripts/check-docs-freshness.mjs",
@@ -122,9 +123,9 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.15",
-    "@tangle-network/agent-eval": "0.125.0",
+    "@tangle-network/agent-eval": "0.126.6",
     "@tangle-network/agent-interface": "0.32.0",
-    "@tangle-network/sandbox": "^0.11.1",
+    "@tangle-network/sandbox": "^0.12.0",
     "@types/node": "^25.9.3",
     "@types/tar-stream": "3.1.4",
     "playwright": "^1.61.0",
@@ -154,9 +155,9 @@
   "license": "MIT",
   "packageManager": "pnpm@10.28.0",
   "peerDependencies": {
-    "@tangle-network/agent-eval": ">=0.122.8 <0.126.0",
+    "@tangle-network/agent-eval": ">=0.126.6 <0.127.0",
     "@tangle-network/agent-interface": ">=0.32.0 <0.33.0",
-    "@tangle-network/sandbox": ">=0.11.1 <1.0.0",
+    "@tangle-network/sandbox": ">=0.12.0 <1.0.0",
     "playwright": "^1.40.0"
   },
   "peerDependenciesMeta": {
@@ -168,7 +169,7 @@
     }
   },
   "dependencies": {
-    "@tangle-network/agent-knowledge": "^4.1.0",
+    "@tangle-network/agent-knowledge": "5.0.1",
     "@tangle-network/agent-profile-materialize": "0.5.1",
     "tar-stream": "3.2.0"
   }

package/skills/build-with-agent-runtime/SKILL.md

@@ -34,7 +34,7 @@ Do not move product storage transactions into a provider-neutral package.
 |---|---|
 | One product chat turn | `handleChatTurn(...)` |
 | One normalized streamed agent turn | `streamAgentTurn(...)` and `collectAgentTurn(...)` |
-| One task or multi-turn loop | `runAgentTask(...)`, `runAgentTaskStream(...)`, or `runLoop(...)` |
+| One task or multi-turn loop | `runAgentTask(...)`, `runAgentTaskStream(...)`, or `runAgentRounds(...)` |
 | Supervisor and workers | `supervise(...)` or `superviseSurface(...)` |
 | Parallel work with a shared budget | `fanout(...)` |
 | Fixed composition | `pipeline(...)`, `panel(...)`, or `verify(...)` |
@@ -50,11 +50,16 @@ Do not move product storage transactions into a provider-neutral package.
 
 ## Improvement flow
 
-`improve(profile, findings, options)` searches one surface and returns a detached winner.
+`improve(profile, options)` searches one surface and returns a detached winner.
 It never changes a profile, document, repository, memory store, or knowledge base.
 
-Supported profile surfaces are prompt, one named inline skill, tools, MCP, hooks, subagents, whole profile, and curated memory in `profile.resources.instructions`.
-Code uses isolated worktrees and returns a sealed patch candidate.
+For a profile field, pass one complete agent-eval `OptimizationMethod`, explicit train, selection, and final-test partitions, judges, and the candidate execution function.
+Use `officialGepa(...)` with an explicit recipe when upstream GEPA should own search.
+Use `officialSkillOpt(...)` when Microsoft's SkillOpt should own search.
+Both require `evaluationId`; change it whenever dispatch, judges, models, or scoring behavior changes.
+Resumable runs accept `never`, `if-compatible`, or `required` and reuse state only when agent-eval derives the same run identity.
+Runtime has no local prompt, skill, memory, or profile optimizer fallback.
+Code uses Runtime's isolated worktrees and returns a sealed patch candidate.
 Knowledge uses `runKnowledgeImprovementJob(...)` and returns paired snapshots.
 
 Use `proposeAgentImprovement(...)` for a production proposal.
@@ -80,12 +85,12 @@ Never treat a lost response as a failed write without reconciling it.
 
 ## Surface rules
 
-- Prompt changes `profile.prompt` only.
-- Skill optimization selects one inline skill by `skills.resourceName`; profile resources must fail closed.
+- Prompt changes `profile.prompt` only and requires a complete method.
+- Skill optimization selects one inline skill by `skills.resourceName`, requires a complete method, and requires profile resources to fail closed.
 - Curated memory changes `profile.resources.instructions`; retrieval stores and memory databases belong in the knowledge flow.
-- Tools, MCP, hooks, subagents, and whole-profile changes require an explicit proposer because Runtime cannot invent domain capabilities safely.
+- Tools, MCP, hooks, subagents, curated memory, rollout policy, and whole-profile changes require a complete method.
 - Code candidates must come from the Runtime worktree path so patch identity and cleanup stay intact.
-- Workflow and policy files are code surfaces; parameter sweeps come from agent-eval.
+- Workflow files are code surfaces. Parameter sweeps belong in a complete agent-eval method.
 - Knowledge candidates remain detached until the shared activation path applies or restores their frozen snapshots.
 
 ## Product integration
@@ -104,7 +109,8 @@ The product must not recreate candidate hashing, paired comparison, confidence i
 ## Do not duplicate
 
 - Do not write a provider-specific profile wrapper; extend `AgentProfile` and its materializer.
-- Do not write a second optimizer loop; compose Eval proposers through `improve(...)`.
+- Do not write a second optimizer loop; pass a complete agent-eval method to `improve(...)`.
+- Do not use Runtime's code generator to approximate GEPA, SkillOpt, or another upstream profile optimizer.
 - Do not write a second candidate catalog; persist the immutable proposal records.
 - Do not let an analyst or adapter commit, push, open a pull request, or edit a live store.
 - Do not hand-roll SSE parsing, usage totals, profile matrices, bootstrap statistics, sandbox acquisition, or worktree cleanup.

package/skills/loop-writer/SKILL.md

@@ -21,7 +21,7 @@ Do not copy signatures from this skill.
 | Parallel repository workers with isolated branches | `worktreeFanout(...)` |
 | Repeated work in a graded tool environment | `runAgentic(...)` |
 | Equal-budget comparison over that environment | `runBenchmark(...)` |
-| Low-level round policy with custom planning and stopping | `runLoop(...)` |
+| Low-level round policy with custom planning and stopping | `runAgentRounds(...)` |
 
 If an existing row fits, use it and stop.
 Do not create another wrapper solely to rename inputs or results.