A self-improving RLM agent for coding workflows and long-running autonomous tasks.
Rust
add authenticated subagent model selection (#453)
kt6c883334cb32
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.mdindex f17798b174..12a0cc62bd 100644--- a/packages/coding-agent/CHANGELOG.md+++ b/packages/coding-agent/CHANGELOG.md@@ -5,6 +5,7 @@- Changed top-level CLI help to show concise common options and commands without loading runtime resources ([ENG-4688](https://linear.app/primeintellect/issue/ENG-4688/help-command-is-obscenely-verbose)).- Fixed completed subagents cancelling their RLM heartbeats before the first run ([ENG-4652](https://linear.app/primeintellect/issue/ENG-4652/subagent-heartbeats-dont-work)).- Changed the fullscreen follow shortcut from `Alt+Down` to `Ctrl+Shift+Down` for more reliable terminal input ([ENG-4684](https://linear.app/primeintellect/issue/ENG-4684/altdown-doesnt-work)).+- Added user-requested model selection for subagents with bounded account-authorized discovery and explicit parent-model fallback warnings ([ENG-4649](https://linear.app/primeintellect/issue/ENG-4649/allow-subagents-to-use-a-different-model-than-the-parent-agent)).- Added subtle feature hints to longer-running agent turns ([ENG-4521](https://linear.app/primeintellect/issue/ENG-4521/add-subtle-hints-for-new-prime-agent-features)).- Fixed active heartbeats not resuming after Prime Agent updates ([ENG-4657](https://linear.app/primeintellect/issue/ENG-4657/heartbeats-dont-survive-updatesdaemon-reboots)).- Fixed the Agents View reordering sessions whenever prompts or heartbeats updated their activity timestamps ([ENG-4650](https://linear.app/primeintellect/issue/ENG-4650/agents-view-shifts-session-list-constantly)).diff --git a/packages/coding-agent/docs/kernel-and-rlm-recursion.md b/packages/coding-agent/docs/kernel-and-rlm-recursion.mdindex de5e88497b..f05e831eaf 100644--- a/packages/coding-agent/docs/kernel-and-rlm-recursion.md+++ b/packages/coding-agent/docs/kernel-and-rlm-recursion.md@@ -69,13 +69,15 @@ packages/coding-agent/src/core/agent-session.tsParent session owns recursion settings and implements runRlmChild().packages/coding-agent/src/core/rlm-runtime.ts- TypeScript request/result types for rlm.run, rlm.list_subagents, and- rlm.delete_subagent plus their typed host request handlers.+ TypeScript request/result types for rlm.run, rlm.find_models,+ rlm.list_subagents, and rlm.delete_subagent plus their typed host request+ handlers.prime-agent-runtime/src/rlm/__init__.pyPython shim installed into ~/.prime/agent/kernel-venv. Exposes rlm, rlm.run(),- rlm.list_subagents(), rlm.delete_subagent(), host_request(), RLMResult,- RLMSubagent, TokenUsage, and the session-backed harness state helper.+ rlm.find_models(), rlm.list_subagents(), rlm.delete_subagent(),+ host_request(), RLMResult, RLMModel, RLMSubagent, TokenUsage, and the+ session-backed harness state helper.prime-agent-runtime/src/rlm/harness.pySession-backed JSON store for reset-free harness refinement notes: prompt@@ -382,9 +384,11 @@ It exports:```pythonrlmrun(prompt: str, **kwargs)+find_models(query: str = "", limit: int = 8)list_subagents()host_request(request_type: str, payload: dict | None = None)RLMResult+RLMModelRLMSubagentTokenUsage```@@ -396,6 +400,8 @@ result.answer # strresult.usage # TokenUsageresult.turns # intresult.session_dir # pathlib.Path | None+result.model # exact provider/model selector+result.warning # fallback notice to surface to the user, or None```The shim also makes the module callable so `import rlm; await rlm("...")` can work, but the ipython tool normally injects the callable object directly as `rlm`.@@ -442,7 +448,7 @@ awaits the FutureThe callback strips the reply's `status` field and converts the payload into `RLMResult`, or raises `RuntimeError` when the host reports an error. The comm-open/await machinery is the public `host_request()`; `run()` is a typed wrapper around it, and other kernel-side skills reuse `host_request()` for their own request types.-The shim accepts `**kwargs` and includes them in the comm payload. The TypeScript host supports `name="..."` for an orchestrator-chosen child session name; all other kwargs are rejected with a clear error instead of being ignored.+The shim accepts `**kwargs` and includes them in the comm payload. The TypeScript host supports `name="..."` for an orchestrator-chosen child session name and `model="provider/model"` for an exact model selection; all other kwargs are rejected with a clear error instead of being ignored. `find_models()` searches a bounded catalog containing only models backed by active, non-expired credentials and returns `provider`, `id`, `name`, and exact `selector` fields.### Why Control-Channel Comm Replies Exist@@ -510,8 +516,7 @@ options.hostHandlers[data.type](data)sendCommMessage(comm_id, { status: "ok", ...result })```-The RLM handlers include `rlm.run` for child creation and-`rlm.list_subagents` for the current parent session's automatic child registry.+The RLM handlers include `rlm.run` for child creation, `rlm.find_models` for bounded authenticated model discovery, and `rlm.list_subagents` for the current parent session's automatic child registry.Unknown types fail with `host request type "<type>" is not available in this session`.@@ -544,10 +549,12 @@ hostHandlers: {Sequence:```text-runRlmChild(prompt)+runRlmChild(prompt, kwargs)|| check current depth < max depth- | require selected model+ | resolve an exact optional model selector against authenticated models+ | fall back to the parent model with a user-facing warning if unavailable+ | otherwise inherit the parent modelvcreate child session dir|@@ -555,7 +562,7 @@ create child session dirvcreate child SessionManager|- | append model change+ | append selected model change| append thinking level changevcreate child Agent@@ -592,6 +599,8 @@ The child returns:};turns: number;session_dir: string | null;+ model: string;+ warning?: string;}```@@ -606,6 +615,30 @@ Turns are counted as assistant messages in the child transcript.The answer is the child's final assistant text. This matches the RLM-1 training surface: the model stops calling tools and the final assistant text is the answer.+By default, a child inherits the parent's current model. When a user or skill+requests a different model, the orchestrator first searches the bounded+authenticated catalog, chooses a candidate, and passes its exact selector:++```python+matches = await rlm.find_models("DeepSeek V4 Flash")+result = await rlm(+ "Check the API",+ model=matches[0].selector,+)+```++`find_models()` returns up to eight matches by default and never more than 20.+Each match exposes `provider`, `id`, `name`, and the exact `provider/model`+`selector`; only models backed by active, non-expired credentials are included.+The full catalog is never injected into the system prompt.++`rlm.run(model=...)` intentionally accepts only an exact selector. If that+selector is unknown, unauthenticated, expired, unavailable, or fails auth+preflight, the child uses the parent model and the result returns the actual+`.model` plus a `.warning` that the parent must surface to the user. A selected+model is persisted with the child and remains active on later turns; changing+the parent model does not update existing children.+### Parent-Scoped Subagent RegistryThe host exposes the current parent session's direct children without relying on@@ -823,6 +856,9 @@ no selected modelunsupported rlm.run kwargs-> TS runRlmChild throws "Unsupported rlm.run kwargs: ..."+unknown, unauthenticated, expired, or unavailable rlm.run model+ -> TS runRlmChild uses the parent model and returns RLMResult.warning+child provider error-> child agent final error behavior is reflected in its final assistant statediff --git a/packages/coding-agent/src/core/agent-session.ts b/packages/coding-agent/src/core/agent-session.tsindex c5c990338e..85e834ff5d 100644--- a/packages/coding-agent/src/core/agent-session.ts+++ b/packages/coding-agent/src/core/agent-session.ts@@ -30,6 +30,7 @@ import {type ThinkingLevel,} from "@earendil-works/pi-agent-core";import type {+ Api,AssistantMessage,ImageContent,Model,@@ -201,10 +202,14 @@ import {type CreateRlmSubagentRuntimeOptions,createDefaultRlmSubagentSessionName,createRlmDeleteSubagentHostHandler,+ createRlmFindModelsHostHandler,createRlmListSubagentsHostHandler,createRlmRunHostHandler,+ findRlmModelMatches,+ normalizeRequestedRlmSubagentModel,normalizeRequestedRlmSubagentSessionName,type RlmDeleteSubagentResult,+ type RlmFindModelsResult,type RlmInternalRunResult,type RlmListSubagentsResult,type RlmRunResult,@@ -251,6 +256,8 @@ export interface RlmChildAgentSnapshot {activeSessionId?: string;/** Stable daemon-visible session name for addressing/displaying the child. */sessionName?: string;+ /** Exact provider/model selector used by the child. */+ model?: string;label: string;status: RlmChildAgentStatus;durationMs?: number;@@ -688,6 +695,11 @@ interface RlmChildRun {detachedDeletion?: RlmSubagentRegistryEntry;}+interface RlmSubagentModelSelection {+ model: Model<Api>;+ warning?: string;+}+// ============================================================================// Constants// ============================================================================@@ -889,6 +901,7 @@ export class AgentSession {private _rlmParentNodeId?: string;private _subagentRuntimeHost?: SubagentRuntimeHost;private _activeRlmChildRuns = new Map<string, RlmChildRun>();+ private _pendingRlmSubagentSessionNames = new Set<string>();// Inline mode keeps finished child sessions so the inspector can still read them;// the daemon does the same by leaving the child session resident in its registry.private _retainedRlmChildSessions = new Map<string, AgentSession>();@@ -6005,6 +6018,7 @@ export class AgentSession {"rlm.run": createRlmRunHostHandler(({ prompt, kwargs, cellSourceCode }) =>this.runRlmChild(prompt, kwargs, cellSourceCode),),+ "rlm.find_models": createRlmFindModelsHostHandler((query, limit) => this.findRlmModels(query, limit)),"rlm.list_subagents": createRlmListSubagentsHostHandler(() => this.listRlmSubagents()),"rlm.delete_subagent": createRlmDeleteSubagentHostHandler((target) => this.deleteRlmSubagent(target)),"model.info": async () => ({@@ -6263,8 +6277,9 @@ export class AgentSession {spawnCode: options.spawnCode,sessionDir: options.sessionDir,model: options.model,- thinkingLevel: this.thinkingLevel,- serviceTier: this.serviceTier,+ thinkingLevel: clampThinkingLevel(options.model, this.thinkingLevel) as ThinkingLevel,+ serviceTier:+ this.serviceTier === "priority" && !supportsFastMode(options.model) ? "default" : this.serviceTier,scopedModels: [...this._scopedModels],activeToolNames: this.getActiveToolNames(),allowedToolNames: this._allowedToolNames ? [...this._allowedToolNames] : undefined,@@ -6315,16 +6330,14 @@ export class AgentSession {}childSessionManager.appendModelChange(options.model.provider, options.model.id);childSessionManager.appendThinkingLevelChange(options.thinkingLevel);- const serviceTier =- options.serviceTier === "priority" && !supportsFastMode(options.model) ? "default" : options.serviceTier;- childSessionManager.appendServiceTierChange(serviceTier);+ childSessionManager.appendServiceTierChange(options.serviceTier);const childAgent = new Agent({initialState: {systemPrompt: "",model: options.model,thinkingLevel: options.thinkingLevel,- serviceTier,+ serviceTier: options.serviceTier,tools: [],},convertToLlm: this.agent.convertToLlm,@@ -6696,6 +6709,9 @@ export class AgentSession {}private _assertRlmSubagentSessionNameAvailable(name: string): void {+ if (this._pendingRlmSubagentSessionNames.has(name)) {+ throw new Error(`RLM subagent session name "${name}" is already in use`);+ }const conflictsWithSelector = (endpoint: {activeSessionId: string;sessionId: string;@@ -6736,25 +6752,86 @@ export class AgentSession {}}- private _startRlmChildRun(prompt: string, kwargs: Record<string, unknown> = {}, spawnCode?: string): RlmChildRun {- const { name: rawName, ...unsupported } = kwargs;+ private async _authenticatedRlmModels(): Promise<Model<Api>[]> {+ return (await this._modelRegistry.getExecutableModels()).filter((model) => {+ const status = this._modelRegistry.getProviderAuthStatus(model.provider);+ return status.source !== "stale" && status.label !== "expired";+ });+ }++ async findRlmModels(query: string, limit: number): Promise<RlmFindModelsResult> {+ return { models: findRlmModelMatches(query, await this._authenticatedRlmModels(), limit) };+ }++ private _rlmModelFallback(reference: string, parentModel: Model<Api>, reason: string): RlmSubagentModelSelection {+ const parentSelector = `${parentModel.provider}/${parentModel.id}`;+ return {+ model: parentModel,+ warning: `Requested subagent model "${reference}" ${reason}. Used the parent model "${parentSelector}" instead. Tell the user the subagent ran with "${parentSelector}", not "${reference}".`,+ };+ }++ private async _resolveRlmSubagentModel(reference: string | undefined): Promise<RlmSubagentModelSelection> {+ const parentModel = this.model;+ if (!parentModel) {+ throw new Error(formatNoModelSelectedMessage());+ }+ if (!reference) {+ return { model: parentModel };+ }++ const normalizedReference = reference.toLowerCase();+ if (`${parentModel.provider}/${parentModel.id}`.toLowerCase() === normalizedReference) {+ return { model: parentModel };+ }+ const model = (await this._authenticatedRlmModels()).find(+ (candidate) => `${candidate.provider}/${candidate.id}`.toLowerCase() === normalizedReference,+ );+ if (!model) {+ return this._rlmModelFallback(reference, parentModel, "is unavailable, unauthenticated, or expired");+ }++ const auth = await this._modelRegistry.getApiKeyAndHeaders(model);+ if (!auth.ok) {+ return this._rlmModelFallback(reference, parentModel, "failed authentication preflight");+ }+ return { model };+ }++ private async _startRlmChildRun(+ prompt: string,+ kwargs: Record<string, unknown> = {},+ spawnCode?: string,+ ): Promise<RlmChildRun> {+ const { name: rawName, model: rawModel, ...unsupported } = kwargs;const unsupportedKwargs = Object.keys(unsupported);if (unsupportedKwargs.length > 0) {throw new Error(`Unsupported rlm.run kwargs: ${unsupportedKwargs.sort().join(", ")}`);}const requestedSessionName = normalizeRequestedRlmSubagentSessionName(rawName);+ const requestedModel = normalizeRequestedRlmSubagentModel(rawModel);if (requestedSessionName) {assertDirectAgentMessageTarget(requestedSessionName);- this._assertRlmSubagentSessionNameAvailable(requestedSessionName);}if (this._rlmDepth >= this._rlmMaxDepth) {throw new Error(`RLM recursion depth limit reached (RLM_DEPTH=${this._rlmDepth}, RLM_MAX_DEPTH=${this._rlmMaxDepth})`,);}- const model = this.model;- if (!model) {- throw new Error(formatNoModelSelectedMessage());+ if (requestedSessionName) {+ this._assertRlmSubagentSessionNameAvailable(requestedSessionName);+ this._pendingRlmSubagentSessionNames.add(requestedSessionName);+ }+ let modelSelection: RlmSubagentModelSelection;+ try {+ modelSelection = await this._resolveRlmSubagentModel(requestedModel);+ } finally {+ if (requestedSessionName) {+ this._pendingRlmSubagentSessionNames.delete(requestedSessionName);+ }+ }+ if (this._disposed || this._disposing) {+ throw new Error("Cannot start a subagent after the parent session has been disposed");}const childSessionDir = this._createChildRlmSessionDir();@@ -6785,12 +6862,14 @@ export class AgentSession {this._activeRlmChildRuns.set(run.id, run);// Status-only relay; the conversation is read from the child's own session.const emitChildUpdate = () => {+ const childModel = childSession?.model ?? modelSelection.model;this._emit({type: "rlm_child_update",child: {id: childNodeId,parentId: this._rlmParentNodeId,sessionName: childSession?.sessionName ?? sessionName,+ model: `${childModel.provider}/${childModel.id}`,label,status: run.status,durationMs,@@ -6826,7 +6905,7 @@ export class AgentSession {sessionName,spawnCode,sessionDir: childSessionDir,- model,+ model: modelSelection.model,}),onSessionPublished: publishChildSession,};@@ -6923,6 +7002,8 @@ export class AgentSession {usage,turns: child._assistantTurnCount(),session_dir: childSessionDir,+ model: `${modelSelection.model.provider}/${modelSelection.model.id}`,+ ...(modelSelection.warning ? { warning: modelSelection.warning } : {}),};run.result = result;return { ...result, assistantUsage };@@ -7010,7 +7091,7 @@ export class AgentSession {}async runRlmChild(prompt: string, kwargs: Record<string, unknown> = {}, spawnCode?: string): Promise<RlmRunResult> {- const run = this._startRlmChildRun(prompt, kwargs, spawnCode);+ const run = await this._startRlmChildRun(prompt, kwargs, spawnCode);if (!run.task) {throw new Error("RLM child failed to start");}@@ -7020,6 +7101,8 @@ export class AgentSession {usage: result.usage,turns: result.turns,session_dir: result.session_dir,+ model: result.model,+ ...(result.warning ? { warning: result.warning } : {}),};}diff --git a/packages/coding-agent/src/core/kernel/bootstrap.ts b/packages/coding-agent/src/core/kernel/bootstrap.tsindex 68a1cf6612..9b12b4b413 100644--- a/packages/coding-agent/src/core/kernel/bootstrap.ts+++ b/packages/coding-agent/src/core/kernel/bootstrap.ts@@ -51,7 +51,7 @@ const REQUIRED_HARNESS_METHODS = ["delete_prompt_note","record_refinement",];-const RUNTIME_READY_CHECK = `import inspect; import rlm; from rlm import McpIntegration; from rlm.harness import HarnessEntry; _harness_methods = ${JSON.stringify(REQUIRED_HARNESS_METHODS)}; assert hasattr(rlm, 'run'); assert callable(rlm); assert hasattr(rlm, 'rlm'); assert callable(rlm.rlm); assert callable(rlm.host_request); assert hasattr(rlm, 'harness'); assert hasattr(rlm, 'get_harness_state'); assert hasattr(rlm.rlm, 'harness'); assert hasattr(rlm.rlm, 'get_harness_state'); assert all(callable(getattr(_harness, _method, None)) for _harness in (rlm.harness, rlm.rlm.harness) for _method in _harness_methods); assert 'reference' in HarnessEntry.__dataclass_fields__; assert 'scope' in HarnessEntry.__dataclass_fields__; assert 'reference' in inspect.signature(rlm.harness.create_skill).parameters; assert 'reference' in inspect.signature(rlm.harness.update_skill).parameters; assert 'global_' in inspect.signature(rlm.harness.create_memory).parameters; assert 'global_' in inspect.signature(rlm.get_harness_state).parameters; assert not hasattr(rlm, 'background'); assert not hasattr(rlm.rlm, 'background')`;+const RUNTIME_READY_CHECK = `import inspect; import rlm; from rlm import McpIntegration; from rlm.harness import HarnessEntry; _harness_methods = ${JSON.stringify(REQUIRED_HARNESS_METHODS)}; assert hasattr(rlm, 'run'); assert callable(rlm); assert hasattr(rlm, 'rlm'); assert callable(rlm.rlm); assert callable(rlm.host_request); assert callable(rlm.find_models); assert callable(rlm.rlm.find_models); assert hasattr(rlm, 'harness'); assert hasattr(rlm, 'get_harness_state'); assert hasattr(rlm.rlm, 'harness'); assert hasattr(rlm.rlm, 'get_harness_state'); assert all(callable(getattr(_harness, _method, None)) for _harness in (rlm.harness, rlm.rlm.harness) for _method in _harness_methods); assert 'reference' in HarnessEntry.__dataclass_fields__; assert 'scope' in HarnessEntry.__dataclass_fields__; assert 'reference' in inspect.signature(rlm.harness.create_skill).parameters; assert 'reference' in inspect.signature(rlm.harness.update_skill).parameters; assert 'global_' in inspect.signature(rlm.harness.create_memory).parameters; assert 'global_' in inspect.signature(rlm.get_harness_state).parameters; assert not hasattr(rlm, 'background'); assert not hasattr(rlm.rlm, 'background')`;const BOOTSTRAP_VERSION_FILE = ".bootstrap-version";const BOOTSTRAP_LOCK_NAME = ".bootstrap.lock";const BOOTSTRAP_LOCK_RETRY_MS = 100;diff --git a/packages/coding-agent/src/core/model-registry.ts b/packages/coding-agent/src/core/model-registry.tsindex eb41d60c16..c9e78a4248 100644--- a/packages/coding-agent/src/core/model-registry.ts+++ b/packages/coding-agent/src/core/model-registry.ts@@ -2,6 +2,7 @@* Model registry - manages built-in and custom models, provides API key resolution.*/+import { Buffer } from "node:buffer";import { createHash } from "node:crypto";import {type AnthropicMessagesCompat,@@ -26,7 +27,7 @@ import { join } from "path";import { type Static, type TProperties, Type } from "typebox";import type { Validator } from "typebox/compile";import type { TLocalizedValidationError } from "typebox/error";-import { getAgentDir } from "../config.js";+import { getAgentDir, VERSION } from "../config.js";import type { AuthSourceToken, AuthStatus, AuthStorage } from "./auth-storage.js";import { PRIME_INFERENCE_PROVIDER_ID } from "./prime-inference-auth.js";import {@@ -354,6 +355,47 @@ function applyModelOverride(model: Model<Api>, override: ModelOverride): Model<A/** Clear the config value command cache. Exported for testing. */export const clearApiKeyCache = clearConfigValueCache;+function readOpenAICodexAccountId(token: string): string | undefined {+ try {+ const payload = JSON.parse(Buffer.from(token.split(".")[1] ?? "", "base64url").toString("utf8")) as {+ "https://api.openai.com/auth"?: { chatgpt_account_id?: unknown };+ };+ const accountId = payload["https://api.openai.com/auth"]?.chatgpt_account_id;+ return typeof accountId === "string" && accountId.length > 0 ? accountId : undefined;+ } catch {+ return undefined;+ }+}++function openAICodexModelsUrl(baseUrl: string): string {+ const normalized = baseUrl.replace(/\/+$/, "");+ let path: string;+ if (normalized.endsWith("/codex/responses")) {+ path = `${normalized.slice(0, -"/responses".length)}/models`;+ } else if (normalized.endsWith("/codex")) {+ path = `${normalized}/models`;+ } else {+ path = `${normalized}/codex/models`;+ }+ const url = new URL(path);+ url.searchParams.set("client_version", VERSION);+ return url.toString();+}++function readOpenAICodexModelIds(value: unknown): Set<string> {+ if (!value || typeof value !== "object" || !("models" in value) || !Array.isArray(value.models)) {+ throw new Error("Invalid OpenAI Codex model catalog");+ }+ return new Set(+ value.models.flatMap((model) => {+ if (!model || typeof model !== "object" || !("slug" in model) || typeof model.slug !== "string") {+ return [];+ }+ return [model.slug];+ }),+ );+}+/*** Model registry - loads and manages models, resolves API keys via AuthStorage.*/@@ -367,6 +409,7 @@ export class ModelRegistry {private authorizedPrivatePrimeInferenceModelIds = new Set<string>();private authorizedPrivatePrimeInferenceTeamId: string | undefined;private explicitPrivatePrimeInferenceModelIds = new Set<string>();+ private openAICodexModelsCache: { authFingerprint: string; modelIds: Set<string>; refreshedAt: number } | undefined;private loadError: string | undefined = undefined;/** Re-register dynamic OAuth providers (e.g. user MCP servers) after refresh() resets the registry. */@@ -713,11 +756,21 @@ export class ModelRegistry {const previousPrivateModelIds = new Set(this.authorizedPrivatePrimeInferenceModelIds);const previousTeamId = this.authorizedPrivatePrimeInferenceTeamId;this.refresh();+ await this.refreshPrivatePrimeInferenceAuthorization(previousPrivateModelIds, previousTeamId);+ return this.getAvailable();+ }++ private async refreshPrivatePrimeInferenceAuthorization(+ previousPrivateModelIds = new Set(this.authorizedPrivatePrimeInferenceModelIds),+ previousTeamId = this.authorizedPrivatePrimeInferenceTeamId,+ ): Promise<void> {const apiKey = await this.authStorage.getApiKey(PRIME_INFERENCE_PROVIDER_ID);const teamHeaders = this.authStorage.getProviderHeaders(PRIME_INFERENCE_PROVIDER_ID);const teamId = teamHeaders?.["X-Prime-Team-ID"];if (!apiKey || !teamHeaders || !teamId) {- return this.getAvailable();+ this.authorizedPrivatePrimeInferenceModelIds.clear();+ this.authorizedPrivatePrimeInferenceTeamId = undefined;+ return;}try {@@ -730,9 +783,11 @@ export class ModelRegistry {if (teamId === previousTeamId) {this.authorizedPrivatePrimeInferenceModelIds = previousPrivateModelIds;this.authorizedPrivatePrimeInferenceTeamId = teamId;+ } else {+ this.authorizedPrivatePrimeInferenceModelIds.clear();+ this.authorizedPrivatePrimeInferenceTeamId = undefined;}}- return this.getAvailable();}async canUseModel(model: Model<Api>): Promise<boolean> {@@ -747,6 +802,54 @@ export class ModelRegistry {return availableModels.some((candidate) => candidate.provider === model.provider && candidate.id === model.id);}+ async getExecutableModels(): Promise<Model<Api>[]> {+ await this.refreshPrivatePrimeInferenceAuthorization();+ const availableModels = this.getAvailable();+ const codexModels = availableModels.filter((model) => model.provider === "openai-codex");+ if (codexModels.length === 0) {+ return availableModels;+ }++ const auth = await this.getApiKeyAndHeaders(codexModels[0]!);+ if (!auth.ok || !auth.apiKey) {+ return availableModels.filter((model) => model.provider !== "openai-codex");+ }+ const authFingerprint = createHash("sha256").update(auth.apiKey).digest("hex");+ const cached = this.openAICodexModelsCache;+ if (cached?.authFingerprint === authFingerprint && Date.now() - cached.refreshedAt < 300_000) {+ return availableModels.filter((model) => model.provider !== "openai-codex" || cached.modelIds.has(model.id));+ }++ const accountId = readOpenAICodexAccountId(auth.apiKey);+ if (!accountId) {+ return availableModels.filter((model) => model.provider !== "openai-codex");+ }+ try {+ const response = await fetch(openAICodexModelsUrl(codexModels[0]!.baseUrl), {+ headers: {+ ...auth.headers,+ Authorization: `Bearer ${auth.apiKey}`,+ "chatgpt-account-id": accountId,+ originator: "pi",+ },+ signal: AbortSignal.timeout(5_000),+ });+ if (!response.ok) {+ throw new Error(`OpenAI Codex model discovery failed with HTTP ${response.status}`);+ }+ const modelIds = readOpenAICodexModelIds(await response.json());+ this.openAICodexModelsCache = { authFingerprint, modelIds, refreshedAt: Date.now() };+ return availableModels.filter((model) => model.provider !== "openai-codex" || modelIds.has(model.id));+ } catch {+ if (cached?.authFingerprint === authFingerprint && Date.now() - cached.refreshedAt < 300_000) {+ return availableModels.filter(+ (model) => model.provider !== "openai-codex" || cached.modelIds.has(model.id),+ );+ }+ return availableModels.filter((model) => model.provider !== "openai-codex");+ }+ }+/*** Find a model by provider and ID.*/diff --git a/packages/coding-agent/src/core/prompts/rlm.ts b/packages/coding-agent/src/core/prompts/rlm.tsindex d6d98087ff..4bfd21331a 100644--- a/packages/coding-agent/src/core/prompts/rlm.ts+++ b/packages/coding-agent/src/core/prompts/rlm.ts@@ -83,8 +83,10 @@ export function buildRlmPrompt(options: RlmPromptOptions): string {if (allowRecursion && hasIpython) {parts.push("",- "A callable `rlm` is already in your global namespace. It returns an `RLMResult` with `.answer` (string), `.usage`, `.turns`, and `.session_dir`. A direct `await rlm('sub-task')` is valid only when the result is immediately required.",+ "A callable `rlm` is already in your global namespace. It returns an `RLMResult` with `.answer` (string), `.usage`, `.turns`, `.session_dir`, `.model`, and optional `.warning`. A direct `await rlm('sub-task')` is valid only when the result is immediately required.","Choose a stable child name with `await rlm('sub-task', name='api-reviewer')`; names must be non-empty and unique among addressable sessions. If omitted, the host generates a readable unique name.",+ "A child inherits your current model by default. When the user or an applicable skill requests a different model, search the bounded authenticated catalog with `matches = await rlm.find_models('requested model')`, choose from each match's `provider`, `id`, `name`, and `selector`, then pass the exact `provider/model` selector with `model=matches[0].selector`. Do not choose a different model on your own.",+ "If an `RLMResult.warning` is set, the requested model could not be used and the child fell back to `.model`; follow the warning and tell the user which model actually ran.","Sub-agents should not block Prime Agent by default: start them with `task = asyncio.create_task(rlm('sub-task'))`, keep the task handle, continue any independent work, and await the task only when you need its result.","For long-running fan-out, do not rely only on in-memory `asyncio.Task` handles: they can be lost if the kernel restarts or state is restored. Recover the current parent session's automatic child registry with `children = await rlm.list_subagents()`; each entry exposes `rlm_child_id`, `active_session_id`, `session_id`, `session_name`, `session_dir`, and `status`.","Delete a running or retained direct child with `await rlm.delete_subagent(child)` (or pass its name/ID); deletion cancels running work, closes the child runtime, and removes it from the registry and daemon addressability.",@@ -118,6 +120,8 @@ export function buildSubagentGuidance(): string {"The host automatically keeps a parent-scoped subagent registry across kernel restarts, state restore, and compaction. Recover it with `children = await rlm.list_subagents()` instead of maintaining a separate registry file or relying on lost `asyncio.Task` handles.","Successful subagent sessions remain in that registry after their initial `rlm()` call finishes, but only while the current parent session remains open. Failed or cancelled children are removed, and retained children close when their parent session closes.","Choose a child name at spawn time with `rlm('task', name='api-reviewer')`, or let the host generate a readable, unique default `session_name`. If the `agent_observe` skill is installed and a registry entry has `active_session_id`, inspect it by name with `await agent_observe.get_agent(child.session_name)` or read bounded previews with `await agent_observe.recent_messages(child.session_name, limit=...)`.",+ "Subagents inherit the parent model. Use `model=...` only when the user or an applicable skill requests another model; call `rlm.find_models()` for a bounded authenticated shortlist and pass one returned exact `selector`. The selected model remains attached to that child across later turns.",+ "When an `RLMResult` has a `.warning`, tell the user which `.model` actually ran instead of the requested model.","If the `agent_message` skill is installed and a registry entry has `active_session_id`, continue that same child by name with `await agent_message.send(child.session_name, message, mode='auto')`; use `mode='steer'` only when you intend to interrupt current work.","Delete a direct child by registry entry or selector with `await rlm.delete_subagent(child)` or `await rlm.delete_subagent('api-reviewer')`. Deleting a running child cancels it first; deleting any child closes its runtime, removes it from the parent registry, and makes it unavailable to messaging and observation.","",diff --git a/packages/coding-agent/src/core/rlm-runtime.ts b/packages/coding-agent/src/core/rlm-runtime.tsindex 5f00f818b2..eb273950ce 100644--- a/packages/coding-agent/src/core/rlm-runtime.ts+++ b/packages/coding-agent/src/core/rlm-runtime.ts@@ -1,5 +1,5 @@import type { ThinkingLevel } from "@earendil-works/pi-agent-core";-import type { Model, ServiceTier, Usage } from "@earendil-works/pi-ai";+import type { Api, Model, ServiceTier, Usage } from "@earendil-works/pi-ai";import type { AgentSession } from "./agent-session.js";import type { ToolDefinition } from "./extensions/index.js";import type { HostRequestHandler } from "./kernel/index.js";@@ -21,6 +21,8 @@ export interface RlmRunResult {usage: RlmUsage;turns: number;session_dir: string | null;+ model: string;+ warning?: string;}export interface RlmInternalRunResult extends RlmRunResult {@@ -46,11 +48,25 @@ export interface RlmDeleteSubagentResult {subagent: RlmSubagentRegistryEntry;}+export interface RlmModelMatch {+ provider: string;+ id: string;+ name: string;+ selector: string;+}++export interface RlmFindModelsResult {+ models: RlmModelMatch[];+}+export type RlmRunHandler = (request: RlmRunRequest) => Promise<RlmRunResult>;export type RlmListSubagentsHandler = () => RlmListSubagentsResult;export type RlmDeleteSubagentHandler = (target: string) => Promise<RlmDeleteSubagentResult>;+export type RlmFindModelsHandler = (query: string, limit: number) => RlmFindModelsResult | Promise<RlmFindModelsResult>;const RLM_SUBAGENT_SESSION_NAME_MAX_LENGTH = 64;+export const DEFAULT_RLM_MODEL_SEARCH_LIMIT = 8;+export const MAX_RLM_MODEL_SEARCH_LIMIT = 20;/** Validate and normalize an orchestrator-supplied subagent session name. */export function normalizeRequestedRlmSubagentSessionName(value: unknown): string | undefined {@@ -70,6 +86,21 @@ export function normalizeRequestedRlmSubagentSessionName(value: unknown): stringreturn name;}+/** Validate and normalize an orchestrator-supplied subagent model reference. */+export function normalizeRequestedRlmSubagentModel(value: unknown): string | undefined {+ if (value === undefined) {+ return undefined;+ }+ if (typeof value !== "string") {+ throw new Error("rlm.run model must be a string");+ }+ const model = value.trim();+ if (!model) {+ throw new Error("rlm.run model must not be empty");+ }+ return model;+}+/** Create a readable, collision-resistant default name usable as an agent-message selector. */export function createDefaultRlmSubagentSessionName(prompt: string, childId: string): string {const promptSlug = prompt@@ -94,6 +125,39 @@ function isRecord(value: unknown): value is Record<string, unknown> {return typeof value === "object" && value !== null && !Array.isArray(value);}+function normalizeModelSearchText(value: string): string {+ return value.toLowerCase().replace(/[^a-z0-9]+/g, "");+}++export function findRlmModelMatches(query: string, models: Model<Api>[], limit: number): RlmModelMatch[] {+ const normalizedQuery = normalizeModelSearchText(query.trim());+ return models+ .map((model) => {+ const selector = `${model.provider}/${model.id}`;+ const fields = [selector, model.id, model.name || model.id];+ const normalizedFields = fields.map(normalizeModelSearchText);+ let score = normalizedQuery ? Number.POSITIVE_INFINITY : 0;+ if (normalizedQuery) {+ const exactIndex = normalizedFields.indexOf(normalizedQuery);+ const prefixIndex = normalizedFields.findIndex((field) => field.startsWith(normalizedQuery));+ const partialIndex = normalizedFields.findIndex((field) => field.includes(normalizedQuery));+ if (exactIndex >= 0) score = exactIndex;+ else if (prefixIndex >= 0) score = 3 + prefixIndex;+ else if (partialIndex >= 0) score = 6 + partialIndex;+ }+ return { model, selector, score };+ })+ .filter((candidate) => Number.isFinite(candidate.score))+ .sort((a, b) => a.score - b.score || a.selector.localeCompare(b.selector))+ .slice(0, limit)+ .map(({ model, selector }) => ({+ provider: model.provider,+ id: model.id,+ name: model.name || model.id,+ selector,+ }));+}+/** Adapt an RlmRunHandler into the typed "rlm.run" handler for the kernel host bridge. */export function createRlmRunHostHandler(handler: RlmRunHandler): HostRequestHandler {return async (payload) => {@@ -107,6 +171,20 @@ export function createRlmRunHostHandler(handler: RlmRunHandler): HostRequestHand};}+/** Search a bounded authenticated model catalog without adding it to the system prompt. */+export function createRlmFindModelsHostHandler(handler: RlmFindModelsHandler): HostRequestHandler {+ return async (payload) => {+ if (typeof payload.query !== "string") {+ throw new Error("rlm.find_models query must be a string");+ }+ const limit = payload.limit === undefined ? DEFAULT_RLM_MODEL_SEARCH_LIMIT : payload.limit;+ if (!Number.isInteger(limit) || (limit as number) < 1 || (limit as number) > MAX_RLM_MODEL_SEARCH_LIMIT) {+ throw new Error(`rlm.find_models limit must be an integer from 1 to ${MAX_RLM_MODEL_SEARCH_LIMIT}`);+ }+ return { models: (await handler(payload.query, limit as number)).models };+ };+}+/** Expose the current parent session's RLM child registry to its kernel. */export function createRlmListSubagentsHostHandler(handler: RlmListSubagentsHandler): HostRequestHandler {return async () => {diff --git a/packages/coding-agent/src/core/tools/ipython.ts b/packages/coding-agent/src/core/tools/ipython.tsindex 3fc7a6cefe..7641fb74f6 100644--- a/packages/coding-agent/src/core/tools/ipython.ts+++ b/packages/coding-agent/src/core/tools/ipython.ts@@ -51,6 +51,9 @@ except Exception as _prime_agent_rlm_error:async def run(self, prompt, **kwargs):self._raise_missing()+ async def find_models(self, query="", limit=8):+ self._raise_missing()+async def list_subagents(self):self._raise_missing()diff --git a/packages/coding-agent/src/modes/agent-connection/types.ts b/packages/coding-agent/src/modes/agent-connection/types.tsindex f67b0d14df..e1a9ec3e8a 100644--- a/packages/coding-agent/src/modes/agent-connection/types.ts+++ b/packages/coding-agent/src/modes/agent-connection/types.ts@@ -479,6 +479,8 @@ export interface AgentConnectionRlmChildAgentSnapshot {activeSessionId?: string;/** Stable daemon-visible session name for addressing/displaying the child. */sessionName?: string;+ /** Exact provider/model selector used by the child. */+ model?: string;label: string;status: AgentConnectionRlmChildAgentStatus;durationMs?: number;diff --git a/packages/coding-agent/src/modes/daemon/daemon-session-list.ts b/packages/coding-agent/src/modes/daemon/daemon-session-list.tsindex 946e4bb6e5..71eec1d9e6 100644--- a/packages/coding-agent/src/modes/daemon/daemon-session-list.ts+++ b/packages/coding-agent/src/modes/daemon/daemon-session-list.ts@@ -310,6 +310,7 @@ function rlmChildSnapshotForActiveSession(parentId: parentNodeId,activeSessionId: activeSession.activeSessionId,sessionName: session.sessionName,+ model: session.model ? `${session.model.provider}/${session.model.id}` : undefined,label: rlmChildLabel(metadata.prompt ?? ""),status,answerPreview,diff --git a/packages/coding-agent/src/modes/interactive/components/child-agent-inspector.ts b/packages/coding-agent/src/modes/interactive/components/child-agent-inspector.tsindex c4b89e7c91..1627a212ad 100644--- a/packages/coding-agent/src/modes/interactive/components/child-agent-inspector.ts+++ b/packages/coding-agent/src/modes/interactive/components/child-agent-inspector.ts@@ -26,6 +26,8 @@ export interface ChildAgentInspectorNode {activeSessionId?: string;/** Stable daemon-visible session name for addressing/displaying the child. */sessionName?: string;+ /** Exact provider/model selector used by the child. */+ model?: string;label: string;status: ChildAgentStatus;durationMs?: number;@@ -191,6 +193,25 @@ function padTableCell(value: string, width: number, ellipsis = ""): string {return truncated + " ".repeat(Math.max(0, width - visibleWidth(truncated)));}+function truncateModelSelector(value: string, width: number): string {+ if (visibleWidth(value) <= width) {+ return value;+ }+ if (width <= 1) {+ return truncateToWidth(value, width, "");+ }+ const prefixWidth = Math.max(1, Math.floor((width - 1) * 0.35));+ const suffixWidth = width - prefixWidth - 1;+ let suffix = "";+ for (const character of [...value].reverse()) {+ if (visibleWidth(character + suffix) > suffixWidth) {+ break;+ }+ suffix = character + suffix;+ }+ return `${truncateToWidth(value, prefixWidth, "")}…${suffix}`;+}+// Rows shown at once in the inline list below the prompt; extras scroll.const SUMMARY_VISIBLE_ROWS = 5;const SUMMARY_LIST_INDENT = 0;@@ -198,12 +219,12 @@ const SHARED_PREFIX_MIN = 12;const SUMMARY_COLUMN_GAP = 2;const SUMMARY_PROMPT_MAX_WIDTH = 24;const SUMMARY_RECAP_MIN_WIDTH = 12;-const SUMMARY_TOOLS_WIDTH = 7;+const SUMMARY_MODEL_WIDTH = 20;const SUMMARY_TOKENS_WIDTH = 8;const SUMMARY_DURATION_WIDTH = 4;const SUMMARY_TOKEN_TIME_GAP = 2;const SUMMARY_METRICS_WIDTH =- SUMMARY_TOOLS_WIDTH + 1 + SUMMARY_TOKENS_WIDTH + SUMMARY_TOKEN_TIME_GAP + SUMMARY_DURATION_WIDTH;+ SUMMARY_MODEL_WIDTH + 1 + SUMMARY_TOKENS_WIDTH + SUMMARY_TOKEN_TIME_GAP + SUMMARY_DURATION_WIDTH;// Opening chars of the prompt kept for context before eliding a shared prefix.const PROMPT_LEADING_CONTEXT = 6;// Words of context kept on each side of the divergence.@@ -449,16 +470,13 @@ export class ChildAgentSummaryComponent implements Component, Focusable {const agentCell = theme.fg("muted", padTableCell(agentLabel, agentWidth));const promptCell = this.elidePrompt(this.entryLabel(entry), sharedPrefix, sharedSuffix, promptWidth);const recapCell = theme.fg("muted", padTableCell(childAgentRecap(entry.node), recapWidth, "…"));- const tools =- entry.node.toolUseCount === undefined- ? ""- : `${entry.node.toolUseCount} ${entry.node.toolUseCount === 1 ? "tool" : "tools"}`;+ const model = entry.node.model ?? "";const tokens = entry.node.tokenCount === undefined ? "" : `${formatTokenCount(entry.node.tokenCount)} tok`;const duration = formatChildAgentDuration(entry.node.durationMs);- const toolsCell = padTableCell(tools, SUMMARY_TOOLS_WIDTH, "…");+ const modelCell = padTableCell(truncateModelSelector(model, SUMMARY_MODEL_WIDTH), SUMMARY_MODEL_WIDTH);const tokensCell = padTableCell(tokens, SUMMARY_TOKENS_WIDTH, "…");const durationCell = padTableCell(duration, SUMMARY_DURATION_WIDTH, "…");- const metrics = `${toolsCell} ${tokensCell}${" ".repeat(SUMMARY_TOKEN_TIME_GAP)}${durationCell}`;+ const metrics = `${modelCell} ${tokensCell}${" ".repeat(SUMMARY_TOKEN_TIME_GAP)}${durationCell}`;const gap = " ".repeat(SUMMARY_COLUMN_GAP);return `${indent}${icon} ${agentCell}${gap}${promptCell}${gap}${recapCell}${gap}${theme.fg("muted", metrics)}`;}@@ -804,6 +822,7 @@ export class ChildAgentDetailComponent implements Component, Focusable {}private detailHintLine(width: number): string {+ const backAction = keyAction("app.agents.back", this.backHintLabel, { primaryOnly: true });const expandAction = keyAction("app.tools.expand", this.toolsExpanded ? "to collapse" : "to expand");const killable = this.node !== undefined && isKillableChildAgentStatus(this.node.status);const stopAction = !killable@@ -811,10 +830,14 @@ export class ChildAgentDetailComponent implements Component, Focusable {: this.isKillConfirmationVisible()? theme.fg("error", `${keyText("app.agents.delete", { primaryOnly: true }).trim()} again to stop`): keyAction("app.agents.delete", "stop", { primaryOnly: true });- return hintLine(- [keyAction("app.agents.back", this.backHintLabel, { primaryOnly: true }), expandAction, stopAction],- width,- );+ const actions = [backAction, expandAction, stopAction];+ const separatorWidth = visibleWidth(joinHints(["", ""]));+ const modelWidth = Math.max(0, width - visibleWidth(joinHints(actions)) - separatorWidth);+ const model =+ this.node?.model && modelWidth >= 4+ ? theme.fg("muted", truncateToWidth(this.node.model, modelWidth, "…"))+ : undefined;+ return hintLine([backAction, model, expandAction, stopAction], width);}private panelLine(line: string, width: number): string {diff --git a/packages/coding-agent/src/modes/interactive/interactive-mode.ts b/packages/coding-agent/src/modes/interactive/interactive-mode.tsindex 967a743a87..ca36b71c8a 100644--- a/packages/coding-agent/src/modes/interactive/interactive-mode.ts+++ b/packages/coding-agent/src/modes/interactive/interactive-mode.ts@@ -317,6 +317,7 @@ export function mergeChildAgentSnapshots(parentId: incoming.parentId ?? previous.parentId,activeSessionId: incoming.activeSessionId ?? previous.activeSessionId,sessionName: incoming.sessionName ?? previous.sessionName,+ model: incoming.model ?? previous.model,durationMs: incoming.durationMs ?? previous.durationMs,answerPreview: incoming.answerPreview ?? previous.answerPreview,toolUseCount:@@ -339,6 +340,7 @@ function childAgentSummaryChanged(return (previous.parentId !== next.parentId ||previous.sessionName !== next.sessionName ||+ previous.model !== next.model ||previous.label !== next.label ||previous.status !== next.status ||previous.durationMs !== next.durationMs ||@@ -5663,6 +5665,7 @@ export class InteractiveMode {id: child.id,activeSessionId: child.activeSessionId,sessionName: child.sessionName,+ model: child.model,label: child.label,status: child.status,durationMs: child.durationMs,diff --git a/packages/coding-agent/test/agent-session-recursion.test.ts b/packages/coding-agent/test/agent-session-recursion.test.tsindex 0e2657160a..be2223c70e 100644--- a/packages/coding-agent/test/agent-session-recursion.test.ts+++ b/packages/coding-agent/test/agent-session-recursion.test.ts@@ -723,8 +723,8 @@ describe("AgentSession rlm recursion", () => {it("rejects unsupported rlm.run kwargs loudly", async () => {const root = createSession();- await expect(root.runRlmChild("nested", { model: "other-model" })).rejects.toThrow(- "Unsupported rlm.run kwargs: model",+ await expect(root.runRlmChild("nested", { temperature: 0 })).rejects.toThrow(+ "Unsupported rlm.run kwargs: temperature",);});@@ -1322,6 +1322,7 @@ describe("AgentSession rlm recursion", () => {usage: { prompt_tokens: 1, completion_tokens: 1 },turns: 1,session_dir: null,+ model: "test/model",};}),},@@ -1349,6 +1350,7 @@ describe("AgentSession rlm recursion", () => {usage: { prompt_tokens: 1, completion_tokens: 1 },turns: 1,session_dir: null,+ model: "test/model",});expect(byCommId.get("comm-b")).toEqual({status: "ok",@@ -1356,6 +1358,7 @@ describe("AgentSession rlm recursion", () => {usage: { prompt_tokens: 1, completion_tokens: 1 },turns: 1,session_dir: null,+ model: "test/model",});} finally {await manager.dispose();@@ -1375,6 +1378,7 @@ describe("AgentSession rlm recursion", () => {usage: { prompt_tokens: 1, completion_tokens: 1 },turns: 1,session_dir: null,+ model: "test/model",};}),},@@ -1400,6 +1404,7 @@ describe("AgentSession rlm recursion", () => {usage: { prompt_tokens: 1, completion_tokens: 1 },turns: 1,session_dir: null,+ model: "test/model",},});} finally {@@ -1419,6 +1424,7 @@ describe("AgentSession rlm recursion", () => {usage: { prompt_tokens: 1, completion_tokens: 1 },turns: 1,session_dir: null,+ model: "test/model",};}),},@@ -1496,6 +1502,7 @@ print(_result.answer)usage: { prompt_tokens: 1, completion_tokens: 1 },turns: 1,session_dir: null,+ model: "test/model",})),},});diff --git a/packages/coding-agent/test/child-agent-summary-list.test.ts b/packages/coding-agent/test/child-agent-summary-list.test.tsindex ee8bb3cbae..699af06685 100644--- a/packages/coding-agent/test/child-agent-summary-list.test.ts+++ b/packages/coding-agent/test/child-agent-summary-list.test.ts@@ -1,6 +1,8 @@-import { visibleWidth } from "@earendil-works/pi-tui";+import { setKeybindings, visibleWidth } from "@earendil-works/pi-tui";import { beforeAll, describe, expect, it } from "vitest";+import { KeybindingsManager } from "../src/core/keybindings.js";import {+ ChildAgentDetailComponent,type ChildAgentInspectorNode,ChildAgentSummaryComponent,} from "../src/modes/interactive/components/child-agent-inspector.js";@@ -18,6 +20,7 @@ function node(id: string, status: ChildAgentInspectorNode["status"] = "running")describe("ChildAgentSummaryComponent inline list", () => {beforeAll(() => {initTheme("dark");+ setKeybindings(new KeybindingsManager());});it("renders one row per subagent when within the cap", () => {@@ -50,12 +53,21 @@ describe("ChildAgentSummaryComponent inline list", () => {expect(out).not.toContain(" a ");});- it("shows tool and token usage beside each subagent", () => {+ it("shows the model and token usage without a tool column", () => {const summary = new ChildAgentSummaryComponent();- summary.setNodes([{ ...node("a"), toolUseCount: 3, tokenCount: 41_000 }]);+ summary.setNodes([{ ...node("a"), model: "openai/gpt-5.4", toolUseCount: 3, tokenCount: 41_000 }]);const out = stripAnsi(summary.render(80).join("\n"));- expect(out).toContain("3 tools");+ expect(out).toContain("openai/gpt-5.4");expect(out).toContain("41k tok");+ expect(out).not.toContain("tools");+ });++ it("preserves the model suffix when its selector is truncated", () => {+ const summary = new ChildAgentSummaryComponent();+ summary.setNodes([{ ...node("a"), model: "prime-inference/internal/glm-5.2-fast" }]);+ const out = stripAnsi(summary.render(80).join("\n"));+ expect(out).toContain("prime-…/glm-5.2-fast");+ expect(out).not.toContain("prime-inference/int…");});it("shows recaps in a fixed column and falls back to activity", () => {@@ -182,6 +194,28 @@ describe("ChildAgentSummaryComponent inline list", () => {expect(out).toMatch(/41k tok\s{3}5s/);});+ it("shows the model to the right of the back action in detail view", () => {+ const detail = new ChildAgentDetailComponent();+ detail.setNode({ ...node("a"), model: "anthropic/claude-opus-4-7" });+ const hint = stripAnsi(detail.render(100).at(-1) ?? "");+ expect(hint).toContain("back to chat");+ expect(hint).toContain("anthropic/claude-opus-4-7");+ expect(hint.indexOf("anthropic/claude-opus-4-7")).toBeGreaterThan(hint.indexOf("back to chat"));+ });++ it("truncates long models before hiding detail actions", () => {+ const detail = new ChildAgentDetailComponent();+ detail.setNode({+ ...node("a"),+ model: "provider-with-a-long-name/model-with-an-even-longer-name",+ });+ const hint = stripAnsi(detail.render(60).at(-1) ?? "");+ expect(hint).toContain("back to chat");+ expect(hint).toContain("provider-…");+ expect(hint).toContain("to expand");+ expect(hint).toContain("stop");+ });+it("keeps the starting prompt compact so the recap gets more room", () => {const summary = new ChildAgentSummaryComponent();const prompt = "Investigate every scheduler dispatch path and identify all possible starvation conditions";diff --git a/packages/coding-agent/test/daemon-session-list.test.ts b/packages/coding-agent/test/daemon-session-list.test.tsindex 393bf95f38..065de503f5 100644--- a/packages/coding-agent/test/daemon-session-list.test.ts+++ b/packages/coding-agent/test/daemon-session-list.test.ts@@ -372,6 +372,7 @@ describe("buildRlmChildSnapshots", () => {const parent = makeState({ activeSessionId: "parent", sessionFile: "/tmp/parent.jsonl" });const child = makeState({activeSessionId: "child",+ model: { provider: "anthropic", id: "claude-opus-4-7" },isStreaming: true,metadata: {kind: "subagent",@@ -423,6 +424,7 @@ describe("buildRlmChildSnapshots", () => {["sub-bbb", "sub-aaa", "done"],]);expect(snapshots[0]).toMatchObject({+ model: "anthropic/claude-opus-4-7",label: "Summarize the repo layout",answerPreview: "The repo is an npm workspace.",toolUseCount: 1,@@ -529,6 +531,7 @@ describe("resolveAttachModelFallbackMessage", () => {interface StateOptions {activeSessionId: string;+ model?: { provider: string; id: string };sessionFile?: string;sessionId?: string;isStreaming?: boolean;@@ -570,7 +573,7 @@ function makeState(options: StateOptions): ActiveSessionState {metadata: options.metadata ?? { kind: "top-level", createdAt: 1 },diagnostics: [],session: {- model: undefined,+ model: options.model,thinkingLevel: "off",isStreaming: options.isStreaming ?? false,isCompacting: false,diff --git a/packages/coding-agent/test/interactive-mode-status.test.ts b/packages/coding-agent/test/interactive-mode-status.test.tsindex d54a3a8d06..1b8c0f7a17 100644--- a/packages/coding-agent/test/interactive-mode-status.test.ts+++ b/packages/coding-agent/test/interactive-mode-status.test.ts@@ -113,6 +113,7 @@ describe("mergeChildAgentSnapshots", () => {const rich: AgentConnectionRlmChildAgentSnapshot = {id: "child-1",activeSessionId: "active-child-1",+ model: "openai/gpt-5.4",label: "Inspect the scheduler",status: "running",durationMs: 5_000,@@ -134,6 +135,7 @@ describe("mergeChildAgentSnapshots", () => {expect(merged).toMatchObject({activeSessionId: "active-child-1",+ model: "openai/gpt-5.4",durationMs: 5_000,answerPreview: "Reading the queue",toolUseCount: 4,diff --git a/packages/coding-agent/test/ipython-bootstrap.test.ts b/packages/coding-agent/test/ipython-bootstrap.test.tsindex cad537f639..8aba99b463 100644--- a/packages/coding-agent/test/ipython-bootstrap.test.ts+++ b/packages/coding-agent/test/ipython-bootstrap.test.ts@@ -13,6 +13,7 @@ describe("IPython RLM bootstrap", () => {it("gives subagent registry operations the actionable missing-runtime fallback", () => {const code = buildRlmBootstrapCode();+ expect(code).toContain('async def find_models(self, query="", limit=8)');expect(code).toContain("async def list_subagents(self)");expect(code).toContain("async def delete_subagent(self, target)");expect(code).toContain("self._raise_missing()");diff --git a/packages/coding-agent/test/marquee-components.test.ts b/packages/coding-agent/test/marquee-components.test.tsindex b2e374defc..103f77f77e 100644--- a/packages/coding-agent/test/marquee-components.test.ts+++ b/packages/coding-agent/test/marquee-components.test.ts@@ -436,12 +436,14 @@ describe("marquee TUI components", () => {const node: ChildAgentInspectorNode = {id: "sub-a",label: "inspect training logs",+ model: "openai/gpt-5.4",status: "running",sessionDir: "/tmp/session/sub-a",children: [{id: "sub-b",label: "check shard 2",+ model: "anthropic/claude-sonnet-4-5",status: "done",sessionDir: "/tmp/session/sub-b",},@@ -456,7 +458,8 @@ describe("marquee TUI components", () => {expect(summaryText).toContain("agents-sidebar");expect(summaryText).toContain("37% context left");expect(summaryText).toContain("S1");- expect(summaryText).toContain("inspect training l…");+ expect(summaryText).toContain("inspect train…");+ expect(summaryText).toContain("openai/gpt-5.4");const infoRow = summary.render(90)[0] ?? "";expect(visibleWidth(infoRow)).toBe(90);@@ -485,6 +488,7 @@ describe("marquee TUI components", () => {expect(detail).toContain("inspect training logs");expect(detail).toContain("reading shard metrics");expect(detail).toContain("← back to chat");+ expect(stripAnsi(detailComponent.render(80).at(-1) ?? "")).toContain("openai/gpt-5.4");expect(detail).not.toContain("user: inspect training logs");expect(detail).not.toContain("assistant: reading shard metrics");});diff --git a/packages/coding-agent/test/suite/regressions/3885-subagent-runtime-host.test.ts b/packages/coding-agent/test/suite/regressions/3885-subagent-runtime-host.test.tsindex eec8f415c4..49bd315a8e 100644--- a/packages/coding-agent/test/suite/regressions/3885-subagent-runtime-host.test.ts+++ b/packages/coding-agent/test/suite/regressions/3885-subagent-runtime-host.test.ts@@ -41,7 +41,7 @@ describe("ENG-3885 subagent runtime host", () => {vi.unstubAllEnvs();});- it("creates running RLM children as tracked AgentSessionRuntime instances", async () => {+ it("creates explicitly modeled RLM children as tracked AgentSessionRuntime instances", async () => {const tempDir = join(tmpdir(), `pi-3885-subagent-runtime-${Date.now()}-${Math.random().toString(36).slice(2)}`);mkdirSync(tempDir, { recursive: true });vi.stubEnv(SESSION_LEASES_ENABLED_ENV, "1");@@ -129,8 +129,6 @@ describe("ENG-3885 subagent runtime host", () => {}});- await runtime.session.setModel(faux.getModel("faux-child")!);-let releaseChild: () => void = () => {};const release = new Promise<void>((resolve) => {releaseChild = resolve;@@ -146,7 +144,9 @@ describe("ENG-3885 subagent runtime host", () => {},]);- const resultPromise = runtime.session.runRlmChild("inspect child runtime");+ const resultPromise = runtime.session.runRlmChild("inspect child runtime", {+ model: `${faux.getModel("faux-child")!.provider}/faux-child`,+ });await waitFor(() => childStarted && runtime.listSubagentRuntimes().length === 1);const [childRuntime] = runtime.listSubagentRuntimes();diff --git a/packages/coding-agent/test/suite/regressions/4649-subagent-model-selection.test.ts b/packages/coding-agent/test/suite/regressions/4649-subagent-model-selection.test.tsnew file mode 100644index 0000000000..af0baf1dd7--- /dev/null+++ b/packages/coding-agent/test/suite/regressions/4649-subagent-model-selection.test.ts@@ -0,0 +1,403 @@+import { fauxAssistantMessage } from "@earendil-works/pi-ai";+import { describe, expect, it, vi } from "vitest";+import type { HostRequestHandlers } from "../../../src/core/kernel/index.js";+import { SessionManager } from "../../../src/core/session-manager.js";+import { createHarness } from "../harness.js";++const provider = "faux-eng-4649";++function openAICodexToken(accountId: string): string {+ const payload = Buffer.from(+ JSON.stringify({ "https://api.openai.com/auth": { chatgpt_account_id: accountId } }),+ ).toString("base64url");+ return `header.${payload}.signature`;+}++describe("ENG-4649 subagent model selection", () => {+ it("searches a bounded authenticated catalog without advertising it", async () => {+ const harness = await createHarness({+ provider,+ models: Array.from({ length: 320 }, (_, index) => ({ id: `model-${index}` })),+ });+ try {+ const prompt = harness.session.agent.state.systemPrompt;+ expect(prompt).not.toContain(`${provider}/model-319`);+ const handlers = (+ harness.session as unknown as { _createKernelHostHandlers(): HostRequestHandlers }+ )._createKernelHostHandlers();+ const findModels = handlers["rlm.find_models"];+ if (!findModels) throw new Error("Missing rlm.find_models host handler");+ await expect(findModels({ query: "model 319", limit: 5 })).resolves.toEqual({+ models: [+ {+ provider,+ id: "model-319",+ name: "model-319",+ selector: `${provider}/model-319`,+ },+ ],+ });+ await expect(findModels({ query: "model", limit: 21 })).rejects.toThrow("integer from 1 to 20");+ harness.setResponses([fauxAssistantMessage("resolved child answer")]);++ const result = await harness.session.runRlmChild("use the requested model", {+ model: `${provider}/model-319`,+ });++ const childEntry = harness.session.listRlmSubagents().subagents[0];+ const child = harness.session.getRlmChildSession(childEntry!.rlm_child_id);+ expect(child?.model?.id).toBe("model-319");+ expect(result.model).toBe(`${provider}/model-319`);+ expect(result.warning).toBeUndefined();+ } finally {+ harness.cleanup();+ }+ });++ it("omits providers whose credentials are marked expired", async () => {+ const harness = await createHarness({ provider, models: [{ id: "parent-model" }, { id: "child-model" }] });+ try {+ expect((await harness.session.findRlmModels("child", 8)).models).toHaveLength(1);+ expect(harness.session.modelRegistry.markProviderAuthStale(provider)).toBe(true);+ expect(harness.session.modelRegistry.markProviderAuthStale(provider)).toBe(true);+ expect(harness.session.modelRegistry.getProviderAuthStatus(provider)).toMatchObject({+ source: "stale",+ label: "expired",+ });+ await expect(harness.session.findRlmModels("", 8)).resolves.toEqual({ models: [] });+ } finally {+ harness.cleanup();+ }+ });++ it("limits ChatGPT discovery and execution to the account model catalog", async () => {+ const codexProvider = "openai-codex";+ const harness = await createHarness({+ provider: codexProvider,+ models: [{ id: "parent-model" }, { id: "unsupported-model" }],+ });+ const fetchModels = vi.fn(+ async () =>+ new Response(JSON.stringify({ models: [{ slug: "parent-model" }] }), {+ status: 200,+ headers: { "content-type": "application/json" },+ }),+ );+ vi.stubGlobal("fetch", fetchModels);+ try {+ harness.authStorage.setRuntimeApiKey(codexProvider, openAICodexToken("account-1"));+ const discovered = await harness.session.findRlmModels("", 20);+ expect(discovered.models.map((model) => model.selector)).toEqual([`${codexProvider}/parent-model`]);+ expect(fetchModels).toHaveBeenCalledWith(+ expect.stringMatching(/\/codex\/models\?client_version=/),+ expect.objectContaining({+ headers: expect.objectContaining({ "chatgpt-account-id": "account-1" }),+ }),+ );++ harness.setResponses([fauxAssistantMessage("parent fallback answer")]);+ const result = await harness.session.runRlmChild("reject unsupported account model", {+ model: `${codexProvider}/unsupported-model`,+ });+ expect(result.model).toBe(`${codexProvider}/parent-model`);+ expect(result.warning).toContain(`not "${codexProvider}/unsupported-model"`);+ } finally {+ vi.unstubAllGlobals();+ harness.cleanup();+ }+ });++ it("includes private Prime models authorized for the selected team", async () => {+ const harness = await createHarness({ provider, models: [{ id: "parent-model" }] });+ const fetchModels = vi.fn(+ async () =>+ new Response(JSON.stringify({ data: [{ id: "internal/glm-5.2-fast" }] }), {+ status: 200,+ headers: { "content-type": "application/json" },+ }),+ );+ vi.stubGlobal("fetch", fetchModels);+ try {+ harness.authStorage.set("prime-inference", {+ type: "api_key",+ key: "prime-key",+ primeTeam: { teamId: "engineering-team", name: "Prime Engineering" },+ });++ const discovered = await harness.session.findRlmModels("glm 5.2", 8);+ expect(discovered.models.map((model) => model.selector)).toContain("prime-inference/internal/glm-5.2-fast");+ expect(fetchModels).toHaveBeenCalledOnce();+ } finally {+ vi.unstubAllGlobals();+ harness.cleanup();+ }+ });++ it("does not reuse an expired ChatGPT model catalog after a refresh failure", async () => {+ const codexProvider = "openai-codex";+ const harness = await createHarness({ provider: codexProvider, models: [{ id: "parent-model" }] });+ const fetchModels = vi+ .fn()+ .mockResolvedValueOnce(+ new Response(JSON.stringify({ models: [{ slug: "parent-model" }] }), {+ status: 200,+ headers: { "content-type": "application/json" },+ }),+ )+ .mockRejectedValueOnce(new Error("offline"));+ vi.stubGlobal("fetch", fetchModels);+ let now = Date.now();+ const dateNow = vi.spyOn(Date, "now").mockImplementation(() => now);+ try {+ harness.authStorage.setRuntimeApiKey(codexProvider, openAICodexToken("account-1"));+ await expect(harness.session.findRlmModels("parent", 8)).resolves.toMatchObject({+ models: [{ selector: `${codexProvider}/parent-model` }],+ });++ now += 300_001;+ await expect(harness.session.findRlmModels("parent", 8)).resolves.toEqual({ models: [] });+ expect(fetchModels).toHaveBeenCalledTimes(2);+ } finally {+ dateNow.mockRestore();+ vi.unstubAllGlobals();+ harness.cleanup();+ }+ });++ it("does not warn when an unavailable selector is already the parent model", async () => {+ const codexProvider = "openai-codex";+ const harness = await createHarness({ provider: codexProvider, models: [{ id: "parent-model" }] });+ const fetchModels = vi.fn().mockRejectedValue(new Error("offline"));+ vi.stubGlobal("fetch", fetchModels);+ try {+ harness.authStorage.setRuntimeApiKey(codexProvider, openAICodexToken("account-1"));+ await expect(harness.session.findRlmModels("parent", 8)).resolves.toEqual({ models: [] });+ harness.setResponses([fauxAssistantMessage("same parent answer")]);++ const result = await harness.session.runRlmChild("keep the parent model", {+ model: `${codexProvider}/parent-model`,+ });+ expect(result).toMatchObject({+ answer: "same parent answer",+ model: `${codexProvider}/parent-model`,+ });+ expect(result.warning).toBeUndefined();+ } finally {+ vi.unstubAllGlobals();+ harness.cleanup();+ }+ });++ it("does not start a child after its parent is disposed during preflight", async () => {+ const harness = await createHarness({+ provider,+ models: [{ id: "parent-model" }, { id: "child-model" }],+ });+ let releasePreflight!: () => void;+ const preflightGate = new Promise<void>((resolve) => {+ releasePreflight = resolve;+ });+ let providerCalls = 0;+ try {+ harness.setResponses([+ () => {+ providerCalls++;+ return fauxAssistantMessage("late child ran");+ },+ ]);+ const authPreflight = vi+ .spyOn(harness.session.modelRegistry, "getApiKeyAndHeaders")+ .mockImplementationOnce(async () => {+ await preflightGate;+ return { ok: true, apiKey: "faux-key" };+ });+ const run = harness.session.runRlmChild("do not run after disposal", {+ model: `${provider}/child-model`,+ });+ await vi.waitFor(() => expect(authPreflight).toHaveBeenCalledOnce());+ harness.session.dispose();+ releasePreflight();++ await expect(run).rejects.toThrow("parent session has been disposed");+ expect(providerCalls).toBe(0);+ expect(harness.session.listRlmSubagents().subagents).toEqual([]);+ } finally {+ releasePreflight();+ harness.cleanup();+ }+ });++ it("reserves an explicit child name while model validation is pending", async () => {+ const harness = await createHarness({+ provider,+ models: [{ id: "parent-model" }],+ });+ try {+ harness.setResponses([fauxAssistantMessage("first child answer")]);++ const first = harness.session.runRlmChild("first task", { name: "shared-reviewer" });+ await expect(harness.session.runRlmChild("second task", { name: "shared-reviewer" })).rejects.toThrow(+ 'RLM subagent session name "shared-reviewer" is already in use',+ );+ await expect(first).resolves.toMatchObject({ answer: "first child answer" });+ } finally {+ harness.cleanup();+ }+ });++ it("runs and retains a child on an explicitly selected model", async () => {+ const harness = await createHarness({+ provider,+ models: [+ { id: "parent-model", reasoning: true },+ { id: "child-model", name: "Child Model", reasoning: false },+ { id: "later-parent-model", reasoning: true },+ ],+ persistSession: true,+ });+ try {+ harness.session.setThinkingLevel("high");+ const seenModels: string[] = [];+ harness.setResponses([+ (_context, _options, _state, model) => {+ seenModels.push(model.id);+ return fauxAssistantMessage("initial child answer");+ },+ (_context, _options, _state, model) => {+ seenModels.push(model.id);+ return fauxAssistantMessage("follow-up child answer");+ },+ ]);++ const result = await harness.session.runRlmChild("inspect the API", {+ name: "api-reviewer",+ model: `${provider}/child-model`,+ });+ const childEntry = harness.session.listRlmSubagents().subagents[0];+ expect(childEntry?.status).toBe("completed");+ const child = harness.session.getRlmChildSession(childEntry!.rlm_child_id);+ expect(child?.model?.id).toBe("child-model");+ expect(child?.thinkingLevel).toBe("off");++ await harness.session.setModel(harness.getModel("later-parent-model")!);+ await child!.prompt("check the follow-up", { expandPromptTemplates: false, source: "extension" });+ await child!.agent.waitForIdle();++ expect(result.answer).toBe("initial child answer");+ expect(seenModels).toEqual(["child-model", "child-model"]);+ expect(child?.model?.id).toBe("child-model");+ expect(result.session_dir).not.toBeNull();+ const childSessions = await SessionManager.list(harness.tempDir, result.session_dir!);+ const persisted = SessionManager.open(childSessions[0]!.path, result.session_dir!);+ expect(persisted.buildSessionContext().model).toEqual({+ provider,+ modelId: "child-model",+ });+ } finally {+ harness.cleanup();+ }+ });++ it("inherits the parent model when no override is supplied", async () => {+ const harness = await createHarness({+ provider,+ models: [{ id: "parent-model" }, { id: "child-model" }],+ });+ try {+ let seenModel: string | undefined;+ harness.setResponses([+ (_context, _options, _state, model) => {+ seenModel = model.id;+ return fauxAssistantMessage("inherited child answer");+ },+ ]);++ await harness.session.runRlmChild("inherit the model");+ expect(seenModel).toBe("parent-model");+ } finally {+ harness.cleanup();+ }+ });++ it("accepts a selected model authenticated by headers", async () => {+ const harness = await createHarness({+ provider,+ models: [{ id: "parent-model" }, { id: "child-model" }],+ });+ try {+ harness.setResponses([fauxAssistantMessage("header-auth child answer")]);+ const authPreflight = vi+ .spyOn(harness.session.modelRegistry, "getApiKeyAndHeaders")+ .mockResolvedValueOnce({ ok: true, headers: { Authorization: "Bearer header-token" } });++ const result = await harness.session.runRlmChild("use header auth", {+ model: `${provider}/child-model`,+ });+ authPreflight.mockRestore();++ expect(result.model).toBe(`${provider}/child-model`);+ expect(result.warning).toBeUndefined();+ } finally {+ harness.cleanup();+ }+ });++ it("falls back to the parent model and returns a user-facing warning", async () => {+ const harness = await createHarness({+ provider,+ models: [{ id: "parent-model" }, { id: "child-model" }],+ });+ try {+ await expect(harness.session.runRlmChild("bad type", { model: 42 })).rejects.toThrow(+ "rlm.run model must be a string",+ );+ harness.setResponses([+ fauxAssistantMessage("unknown fallback"),+ fauxAssistantMessage("unauthenticated fallback"),+ fauxAssistantMessage("unavailable fallback"),+ fauxAssistantMessage("preflight fallback"),+ ]);++ const unknown = await harness.session.runRlmChild("unknown model", {+ model: `${provider}/missing-model`,+ });+ expect(unknown.model).toBe(`${provider}/parent-model`);+ expect(unknown.warning).toContain(`not "${provider}/missing-model"`);+ expect(unknown.warning).toContain("Tell the user");++ const unauthenticated = await harness.session.runRlmChild("unauthenticated provider", {+ model: "not-authed/missing-model",+ });+ expect(unauthenticated.model).toBe(`${provider}/parent-model`);+ expect(unauthenticated.warning).toContain("unauthenticated");++ const availability = vi+ .spyOn(harness.session.modelRegistry, "getAvailable")+ .mockReturnValue([harness.getModel("parent-model")!]);+ const unavailable = await harness.session.runRlmChild("unavailable model", {+ model: `${provider}/child-model`,+ });+ availability.mockRestore();++ expect(unavailable.model).toBe(`${provider}/parent-model`);+ expect(unavailable.warning).toContain(`not "${provider}/child-model"`);++ const authPreflight = vi+ .spyOn(harness.session.modelRegistry, "getApiKeyAndHeaders")+ .mockResolvedValueOnce({ ok: false, error: "token expired" });+ const failedPreflight = await harness.session.runRlmChild("failed auth preflight", {+ model: `${provider}/child-model`,+ });+ authPreflight.mockRestore();+ expect(failedPreflight.model).toBe(`${provider}/parent-model`);+ expect(failedPreflight.warning).toContain("failed authentication preflight");++ expect(harness.session.listRlmSubagents().subagents).toHaveLength(4);+ for (const child of harness.session.listRlmSubagents().subagents) {+ expect(harness.session.getRlmChildSession(child.rlm_child_id)?.model?.id).toBe("parent-model");+ }+ } finally {+ harness.cleanup();+ }+ });+});diff --git a/packages/coding-agent/test/system-prompt.test.ts b/packages/coding-agent/test/system-prompt.test.tsindex 6ecc2d9470..8792282ee6 100644--- a/packages/coding-agent/test/system-prompt.test.ts+++ b/packages/coding-agent/test/system-prompt.test.ts@@ -99,6 +99,22 @@ describe("buildRlmPrompt", () => {expect(prompt).toContain("Each `%%bash` cell runs in a throw-away subshell");});+ test("discovers requested models through a bounded authenticated host search", () => {+ const prompt = buildRlmPrompt({+ cwd: "/repo",+ messagesPath: "/repo/.pi/sessions/session.jsonl",+ activeTools: ["ipython"],+ });++ expect(prompt).toContain("await rlm.find_models('requested model')");+ expect(prompt).toContain("bounded authenticated catalog");+ expect(prompt).toContain("exact `provider/model` selector");+ expect(prompt).toContain("If an `RLMResult.warning` is set");+ expect(prompt).toContain("tell the user which model actually ran");+ expect(prompt).toContain("Do not choose a different model on your own");+ expect(prompt).not.toContain("model choices for subagents");+ });+test("only documents ipython shell prefixes when ipython is active", () => {const prompt = buildRlmPrompt({cwd: "/repo",diff --git a/prime-agent-runtime/src/rlm/__init__.py b/prime-agent-runtime/src/rlm/__init__.pyindex b3f2a8fe6d..937ca6a185 100644--- a/prime-agent-runtime/src/rlm/__init__.py+++ b/prime-agent-runtime/src/rlm/__init__.py@@ -41,6 +41,16 @@ class RLMResult:session_dir: Path | None = Noneusage: TokenUsage = field(default_factory=TokenUsage)turns: int = 0+ model: str | None = None+ warning: str | None = None+++@dataclass(frozen=True)+class RLMModel:+ provider: str+ id: str+ name: str+ selector: str@dataclass(frozen=True)@@ -103,6 +113,8 @@ def _result_from_payload(payload: dict[str, Any]) -> RLMResult:usage=usage,turns=int(payload.get("turns", 0)),session_dir=session_dir,+ model=payload.get("model") if isinstance(payload.get("model"), str) else None,+ warning=payload.get("warning") if isinstance(payload.get("warning"), str) else None,)@@ -166,7 +178,10 @@ async def host_request(request_type: str, payload: dict[str, Any] | None = None)async def run(prompt: str, **kwargs: Any) -> RLMResult:- """Run a recursive Prime Agent child through the TypeScript host."""+ """Run a recursive Prime Agent child through the TypeScript host.++ ``model`` selects a child with an exact ``provider/model`` selector.+ """if not isinstance(prompt, str):raise TypeError(f"prompt must be str, got {type(prompt).__name__}")_ensure_recursion_allowed()@@ -174,6 +189,31 @@ async def run(prompt: str, **kwargs: Any) -> RLMResult:return _result_from_payload(payload)+def _model_from_payload(payload: Any) -> RLMModel:+ if not isinstance(payload, dict):+ raise RuntimeError("rlm.find_models returned an invalid model entry")+ provider = payload.get("provider")+ model_id = payload.get("id")+ name = payload.get("name")+ selector = payload.get("selector")+ if not all(isinstance(value, str) and value for value in (provider, model_id, name, selector)):+ raise RuntimeError("rlm.find_models returned an invalid model entry")+ return RLMModel(provider=provider, id=model_id, name=name, selector=selector)+++async def find_models(query: str = "", limit: int = 8) -> list[RLMModel]:+ """Search a bounded list of models backed by active user credentials."""+ if not isinstance(query, str):+ raise TypeError(f"query must be str, got {type(query).__name__}")+ if not isinstance(limit, int):+ raise TypeError(f"limit must be int, got {type(limit).__name__}")+ payload = await host_request("rlm.find_models", {"query": query, "limit": limit})+ models = payload.get("models")+ if not isinstance(models, list):+ raise RuntimeError("rlm.find_models returned an invalid models list")+ return [_model_from_payload(model) for model in models]++def _subagent_from_payload(payload: Any, operation: str = "rlm.list_subagents") -> RLMSubagent:if not isinstance(payload, dict):raise RuntimeError(f"{operation} returned an invalid subagent entry")@@ -286,6 +326,9 @@ class _RLMCallable:async def run(self, prompt: str, **kwargs: Any) -> RLMResult:return await run(prompt, **kwargs)+ async def find_models(self, query: str = "", limit: int = 8) -> list[RLMModel]:+ return await find_models(query, limit)+async def list_subagents(self) -> list[RLMSubagent]:return await list_subagents()@@ -314,11 +357,13 @@ __all__ = ["McpIntegration","McpToolError","NotEnabled",+ "RLMModel","RLMResult","RLMSubagent","RefinementEvent","TokenUsage","delete_subagent",+ "find_models","get_harness_state","harness","host_request",diff --git a/prime-agent-runtime/test/test_subagent_registry.py b/prime-agent-runtime/test/test_subagent_registry.pyindex dff6eadc9b..3a01c405a1 100644--- a/prime-agent-runtime/test/test_subagent_registry.py+++ b/prime-agent-runtime/test/test_subagent_registry.py@@ -39,23 +39,76 @@ class RlmSubagentRegistryTest(unittest.TestCase):self.assertEqual(subagents[0].status, "completed")host_request.assert_awaited_once_with("rlm.list_subagents")- def test_forwards_orchestrator_chosen_name_to_host(self) -> None:+ def test_forwards_orchestrator_chosen_name_and_model_to_host(self) -> None:host_request = AsyncMock(return_value={"answer": "done","usage": {"prompt_tokens": 1, "completion_tokens": 1},"turns": 1,"session_dir": "/tmp/parent/sub-a1b2c3d4",+ "model": "openai/gpt-5.4",+ "warning": "Requested model unavailable; tell the user about the fallback.",})with patch.object(rlm_module, "host_request", host_request):- asyncio.run(rlm_module.rlm("check the API", name="api-reviewer"))+ result = asyncio.run(+ rlm_module.rlm(+ "check the API",+ name="api-reviewer",+ model="deepseek/deepseek-v4-flash",+ )+ )host_request.assert_awaited_once_with("rlm.run",- {"prompt": "check the API", "kwargs": {"name": "api-reviewer"}},+ {+ "prompt": "check the API",+ "kwargs": {+ "name": "api-reviewer",+ "model": "deepseek/deepseek-v4-flash",+ },+ },)+ self.assertEqual(result.model, "openai/gpt-5.4")+ self.assertIn("tell the user", result.warning)++ def test_finds_authenticated_models_through_host(self) -> None:+ host_request = AsyncMock(+ return_value={+ "models": [+ {+ "provider": "anthropic",+ "id": "claude-opus-4-7",+ "name": "Claude Opus 4.7",+ "selector": "anthropic/claude-opus-4-7",+ }+ ]+ }+ )++ with patch.object(rlm_module, "host_request", host_request):+ models = asyncio.run(rlm_module.rlm.find_models("opus", limit=3))++ self.assertEqual(models[0].provider, "anthropic")+ self.assertEqual(models[0].id, "claude-opus-4-7")+ self.assertEqual(models[0].name, "Claude Opus 4.7")+ self.assertEqual(models[0].selector, "anthropic/claude-opus-4-7")+ host_request.assert_awaited_once_with(+ "rlm.find_models",+ {"query": "opus", "limit": 3},+ )++ def test_rejects_invalid_model_search_input_and_response(self) -> None:+ with self.assertRaisesRegex(TypeError, "query must be str"):+ asyncio.run(rlm_module.find_models(123))+ with self.assertRaisesRegex(TypeError, "limit must be int"):+ asyncio.run(rlm_module.find_models("opus", limit="3"))++ host_request = AsyncMock(return_value={"models": [{"provider": "anthropic"}]})+ with patch.object(rlm_module, "host_request", host_request):+ with self.assertRaisesRegex(RuntimeError, "invalid model entry"):+ asyncio.run(rlm_module.find_models("opus"))def test_deletes_subagent_by_name_through_host(self) -> None:deleted_payload = {