@oh-my-pi/pi-coding-agent 16.2.5 → 16.2.7

package/dist/types/cli/bench-cli.d.ts

@@ -1,5 +1,5 @@
 import type { ResolvedThinkingLevel } from "@oh-my-pi/pi-agent-core";
-import type { Api, ApiKeyResolver, AssistantMessageEventStream, Context, Model, SimpleStreamOptions } from "@oh-my-pi/pi-ai";
+import type { Api, ApiKeyResolver, AssistantMessageEventStream, Context, Model, ServiceTierByFamily, SimpleStreamOptions } from "@oh-my-pi/pi-ai";
 import { type CanonicalModelVariant } from "@oh-my-pi/pi-catalog/identity";
 import type { ApiKeyResolverModel } from "../config/api-key-resolver";
 import { type CanonicalModelQueryOptions } from "../config/model-registry";
@@ -10,7 +10,10 @@ export interface BenchCommandArgs {
         runs?: number;
         maxTokens?: number;
         prompt?: string;
+        /** Service-tier setting value (`none` omits); overrides the configured `serviceTier` setting. */
+        serviceTier?: string;
         json?: boolean;
+        par?: number;
     };
 }
 export interface BenchModelRegistry {
@@ -62,6 +65,8 @@ export interface BenchSummary {
     maxTokens: number;
     models: BenchModelReport[];
     failures: number;
+    /** Requested per-family service tiers, resolved per model before reaching the wire. */
+    serviceTierByFamily?: ServiceTierByFamily;
 }
 type BenchStreamSimple = (model: Model<Api>, context: Context, options?: SimpleStreamOptions) => AssistantMessageEventStream;
 export interface BenchDependencies {

package/dist/types/commands/bench.d.ts

@@ -20,9 +20,17 @@ export default class Bench extends Command {
         prompt: import("@oh-my-pi/pi-utils/cli").FlagDescriptor<"string"> & {
             description: string;
         };
+        "service-tier": import("@oh-my-pi/pi-utils/cli").FlagDescriptor<"string"> & {
+            description: string;
+            options: readonly ["none", "auto", "default", "flex", "scale", "priority"];
+        };
         json: import("@oh-my-pi/pi-utils/cli").FlagDescriptor<"boolean"> & {
             description: string;
         };
+        par: import("@oh-my-pi/pi-utils/cli").FlagDescriptor<"integer"> & {
+            description: string;
+            default: number;
+        };
     };
     static examples: string[];
     run(): Promise<void>;

package/dist/types/config/model-discovery.d.ts

@@ -37,10 +37,14 @@ export interface DiscoveryContext {
      */
     getBearerApiKeyResolver(provider: string): Promise<ApiKey | undefined>;
 }
+type LlamaCppDiscoveredModelRuntimeMetadata = {
+    contextWindow: number;
+    maxTokens: number;
+};
 export declare function discoverModelsByProviderType(providerConfig: DiscoveryProviderConfig, ctx: DiscoveryContext): Promise<Model<Api>[]>;
 export declare function discoverOllamaModels(providerConfig: DiscoveryProviderConfig, ctx: DiscoveryContext): Promise<Model<Api>[]>;
 export declare function discoverLlamaCppModels(providerConfig: DiscoveryProviderConfig, ctx: DiscoveryContext): Promise<Model<Api>[]>;
-export declare function discoverLlamaCppModelContextWindow(model: Pick<Model<Api>, "provider" | "id" | "baseUrl" | "headers">, ctx: DiscoveryContext): Promise<number | undefined>;
+export declare function discoverLlamaCppModelRuntimeMetadata(model: Pick<Model<Api>, "provider" | "id" | "baseUrl" | "headers">, ctx: DiscoveryContext): Promise<LlamaCppDiscoveredModelRuntimeMetadata | undefined>;
 export declare function discoverOpenAIModelsList(providerConfig: DiscoveryProviderConfig, ctx: DiscoveryContext): Promise<Model<Api>[]>;
 export declare function discoverLiteLLMModels(providerConfig: DiscoveryProviderConfig, ctx: DiscoveryContext): Promise<Model<Api>[]>;
 /**
@@ -61,3 +65,4 @@ export declare function discoverLiteLLMModels(providerConfig: DiscoveryProviderC
 export declare function discoverProxyModels(providerConfig: DiscoveryProviderConfig, ctx: DiscoveryContext): Promise<Model<Api>[]>;
 export declare function normalizeLiteLLMDiscoveryBaseUrl(baseUrl?: string): string;
 export declare function normalizeOpenAIModelsListBaseUrl(baseUrl?: string): string;
+export {};

package/dist/types/config/service-tier.d.ts

@@ -1,34 +1,49 @@
-import type { ServiceTier } from "@oh-my-pi/pi-ai";
+import type { ServiceTier, ServiceTierByFamily } from "@oh-my-pi/pi-ai";
 import type { SubmenuOption } from "./settings-schema";
 /**
- * Service-tier setting values shared by every "Service Tier" setting. `"none"`
- * is the omit-the-parameter sentinel; the remaining values mirror
- * {@link ServiceTier}.
+ * Per-family service-tier setting values. `"none"` is the omit-the-parameter
+ * sentinel; the rest mirror the wire {@link ServiceTier} values each provider
+ * family actually realizes. OpenAI accepts the full set; Anthropic realizes
+ * only `priority` (fast mode); Google (Gemini API + Vertex) realizes
+ * `flex`/`priority`.
  */
-export declare const SERVICE_TIER_SETTING_VALUES: readonly ["none", "auto", "default", "flex", "scale", "priority", "openai-only", "claude-only"];
-export type ServiceTierSettingValue = (typeof SERVICE_TIER_SETTING_VALUES)[number];
-/** Variant value set for scoped service-tier settings (subagent/advisor) that can defer to the main agent. */
-export declare const SERVICE_TIER_INHERIT_SETTING_VALUES: readonly ["inherit", "none", "auto", "default", "flex", "scale", "priority", "openai-only", "claude-only"];
+export declare const SERVICE_TIER_OPENAI_VALUES: readonly ["none", "auto", "default", "flex", "scale", "priority"];
+export declare const SERVICE_TIER_ANTHROPIC_VALUES: readonly ["none", "priority"];
+export declare const SERVICE_TIER_GOOGLE_VALUES: readonly ["none", "flex", "priority"];
+export type ServiceTierOpenAISettingValue = (typeof SERVICE_TIER_OPENAI_VALUES)[number];
+export type ServiceTierAnthropicSettingValue = (typeof SERVICE_TIER_ANTHROPIC_VALUES)[number];
+export type ServiceTierGoogleSettingValue = (typeof SERVICE_TIER_GOOGLE_VALUES)[number];
+/**
+ * Inherit-capable single value for the subagent/advisor tiers. The chosen tier
+ * is broadcast across families and applied to whichever family the spawned
+ * model belongs to (clamped to what that family realizes); `"inherit"` defers
+ * to the main agent's live per-family selection.
+ */
+export declare const SERVICE_TIER_INHERIT_SETTING_VALUES: readonly ["inherit", "none", "auto", "default", "flex", "scale", "priority"];
 export type ServiceTierInheritSettingValue = (typeof SERVICE_TIER_INHERIT_SETTING_VALUES)[number];
-/** Submenu descriptions shared by the base `serviceTier` setting. */
-export declare const SERVICE_TIER_OPTIONS: ReadonlyArray<SubmenuOption<ServiceTierSettingValue>>;
-/** Submenu descriptions for inherit-capable service-tier settings. */
+export declare const SERVICE_TIER_OPENAI_OPTIONS: ReadonlyArray<SubmenuOption<ServiceTierOpenAISettingValue>>;
+export declare const SERVICE_TIER_ANTHROPIC_OPTIONS: ReadonlyArray<SubmenuOption<ServiceTierAnthropicSettingValue>>;
+export declare const SERVICE_TIER_GOOGLE_OPTIONS: ReadonlyArray<SubmenuOption<ServiceTierGoogleSettingValue>>;
 export declare const SERVICE_TIER_INHERIT_OPTIONS: ReadonlyArray<SubmenuOption<ServiceTierInheritSettingValue>>;
+/** Map a per-family setting value to a wire {@link ServiceTier}, or `undefined` to omit. */
+export declare function serviceTierSettingToTier(value: string): ServiceTier | undefined;
+/** Assemble the live per-family tier map from the three `tier.*` setting values. */
+export declare function buildServiceTierByFamily(openai: string, anthropic: string, google: string): ServiceTierByFamily;
 /**
- * Resolve a service-tier setting value to the wire {@link ServiceTier} (or
- * `undefined` to omit). `"inherit"` defers to `inherited`; `"none"` omits.
+ * Broadcast a single chosen tier across families, clamped to what each family
+ * realizes: OpenAI takes any tier, Anthropic only `priority`, Google only
+ * `flex`/`priority`. Used by the subagent/advisor single-value settings and the
+ * `omp bench --service-tier` flag, which apply one tier to whatever family the
+ * target model belongs to.
  */
-export declare function resolveServiceTierSetting(value: string, inherited: ServiceTier | undefined): ServiceTier | undefined;
+export declare function serviceTierForAllFamilies(tier: ServiceTier | undefined): ServiceTierByFamily;
 /**
- * Resolve the `serviceTier` *setting value* to stamp onto a subagent's settings
- * snapshot.
+ * Resolve a subagent/advisor service-tier setting to a per-family map.
  *
- * - A concrete `subagentSetting` (`"none"` or a tier) wins outright.
- * - `"inherit"` defers to the parent's live effective tier when the caller has a
- *   live session (`inherited` passed as `ServiceTier | null`, where `null` means
- *   the parent explicitly has no tier — e.g. `/fast off`). When no live session
- *   is available (`inherited === undefined`, e.g. cold subagent revive) it falls
- *   back to the parent's configured `serviceTier` setting so behavior matches a
- *   plain settings snapshot.
+ * - A concrete tier is broadcast across families (see
+ *   {@link serviceTierForAllFamilies}).
+ * - `"none"` yields an empty map.
+ * - `"inherit"` defers to `inherited` — the parent's live per-family tiers when
+ *   a live session supplied them, else the empty map.
  */
-export declare function resolveSubagentServiceTier(subagentSetting: string, configuredTier: ServiceTierSettingValue, inherited: ServiceTier | null | undefined): ServiceTierSettingValue;
+export declare function resolveSubagentServiceTier(setting: string, inherited: ServiceTierByFamily): ServiceTierByFamily;

package/dist/types/config/settings-schema.d.ts

@@ -1232,65 +1232,65 @@ export declare const SETTINGS_SCHEMA: {
             }];
         };
     };
-    readonly serviceTier: {
+    readonly "tier.openai": {
         readonly type: "enum";
-        readonly values: readonly ["none", "auto", "default", "flex", "scale", "priority", "openai-only", "claude-only"];
+        readonly values: readonly ["none", "auto", "default", "flex", "scale", "priority"];
         readonly default: "none";
         readonly ui: {
             readonly tab: "model";
             readonly group: "Sampling";
-            readonly label: "Service Tier";
-            readonly description: 'Processing priority hint (none = omit). OpenAI accepts the tier values directly; Anthropic realizes `priority` as `speed: "fast"` on supported Opus models. Scoped values target one family.';
-            readonly options: readonly SubmenuOption<"auto" | "claude-only" | "default" | "flex" | "none" | "openai-only" | "priority" | "scale">[];
+            readonly label: "Service Tier — OpenAI";
+            readonly description: "Processing tier for OpenAI / OpenAI-Codex requests, and OpenAI-family models routed via OpenRouter (none = omit). Sent as `service_tier`.";
+            readonly options: readonly SubmenuOption<"auto" | "default" | "flex" | "none" | "priority" | "scale">[];
         };
     };
-    readonly serviceTierSubagent: {
+    readonly "tier.anthropic": {
         readonly type: "enum";
-        readonly values: readonly ["inherit", "none", "auto", "default", "flex", "scale", "priority", "openai-only", "claude-only"];
-        readonly default: "inherit";
+        readonly values: readonly ["none", "priority"];
+        readonly default: "none";
         readonly ui: {
             readonly tab: "model";
             readonly group: "Sampling";
-            readonly label: "Service Tier - Subagent";
-            readonly description: "Service Tier for spawned task/eval subagents. Inherit = match the main agent's live tier (tracks /fast); pick a value to scope subagents independently.";
-            readonly options: readonly SubmenuOption<"auto" | "claude-only" | "default" | "flex" | "inherit" | "none" | "openai-only" | "priority" | "scale">[];
+            readonly label: "Service Tier — Anthropic";
+            readonly description: 'Processing tier for Claude requests. `priority` realizes fast mode (`speed: "fast"`) on supported direct Anthropic models; ignored on Bedrock/Vertex Claude and via OpenRouter.';
+            readonly options: readonly SubmenuOption<"none" | "priority">[];
         };
     };
-    readonly serviceTierAdvisor: {
+    readonly "tier.google": {
         readonly type: "enum";
-        readonly values: readonly ["inherit", "none", "auto", "default", "flex", "scale", "priority", "openai-only", "claude-only"];
+        readonly values: readonly ["none", "flex", "priority"];
         readonly default: "none";
         readonly ui: {
             readonly tab: "model";
             readonly group: "Sampling";
-            readonly label: "Service Tier - Advisor";
-            readonly description: "Service Tier for the advisor model. None = standard processing; Inherit = match the main agent's live tier; pick a value (e.g. Priority) to run the advisor on a faster serving path.";
-            readonly options: readonly SubmenuOption<"auto" | "claude-only" | "default" | "flex" | "inherit" | "none" | "openai-only" | "priority" | "scale">[];
-            readonly condition: "advisorEnabled";
+            readonly label: "Service Tier — Google";
+            readonly description: "Processing tier for Gemini (Google AI Studio + Vertex) requests, and Google-family models routed via OpenRouter (none = omit). Sent as the top-level `serviceTier` field.";
+            readonly options: readonly SubmenuOption<"flex" | "none" | "priority">[];
         };
     };
-    readonly fastModeScope: {
+    readonly "tier.subagent": {
         readonly type: "enum";
-        readonly values: readonly ["both", "openai", "claude"];
-        readonly default: "both";
+        readonly values: readonly ["inherit", "none", "auto", "default", "flex", "scale", "priority"];
+        readonly default: "inherit";
         readonly ui: {
             readonly tab: "model";
             readonly group: "Sampling";
-            readonly label: "Fast Mode Scope";
-            readonly description: 'Which providers `/fast on` (and the fast-mode toggle) target. "both" = priority on every supported provider; "openai"/"claude" scope it to one family (mirrors serviceTier openai-only/claude-only).';
-            readonly options: readonly [{
-                readonly value: "both";
-                readonly label: "Both";
-                readonly description: "Priority on every supported provider";
-            }, {
-                readonly value: "openai";
-                readonly label: "OpenAI only";
-                readonly description: "Priority on OpenAI/OpenAI-Codex requests; ignored elsewhere";
-            }, {
-                readonly value: "claude";
-                readonly label: "Claude only";
-                readonly description: "Anthropic fast mode on direct Claude requests; ignored elsewhere";
-            }];
+            readonly label: "Service Tier — Subagent";
+            readonly description: "Service Tier for spawned task/eval subagents. Inherit = match the main agent's live per-family tiers (tracks /fast); pick a value to apply it to whichever family the subagent's model belongs to.";
+            readonly options: readonly SubmenuOption<"auto" | "default" | "flex" | "inherit" | "none" | "priority" | "scale">[];
+        };
+    };
+    readonly "tier.advisor": {
+        readonly type: "enum";
+        readonly values: readonly ["inherit", "none", "auto", "default", "flex", "scale", "priority"];
+        readonly default: "none";
+        readonly ui: {
+            readonly tab: "model";
+            readonly group: "Sampling";
+            readonly label: "Service Tier — Advisor";
+            readonly description: "Service Tier for the advisor model. None = standard processing; Inherit = match the main agent's live per-family tiers; pick a value to apply it to the advisor model's family.";
+            readonly options: readonly SubmenuOption<"auto" | "default" | "flex" | "inherit" | "none" | "priority" | "scale">[];
+            readonly condition: "advisorEnabled";
         };
     };
     readonly "retry.enabled": {
@@ -2228,7 +2228,7 @@ export declare const SETTINGS_SCHEMA: {
     };
     readonly "snapcompact.shape": {
         readonly type: "enum";
-        readonly values: readonly ["auto", ...("11on16-bw" | "5x8-bw" | "5x8-sent" | "6x12-dim" | "6x6u-bw" | "6x6u-sent" | "8on16-bw" | "8on22-bw" | "8x13-bw" | "8x8r-bw" | "8x8r-sent" | "8x8u-bw" | "8x8u-sent" | "doc-8on16-bw" | "doc-8on16-sent" | "doc-8on16-sent-dim")[]];
+        readonly values: readonly ["auto", ...("11on16-bw" | "5x8-bw" | "5x8-sent" | "6x12-dim" | "6x6u-bw" | "6x6u-sent" | "8on16-bw" | "8on22-bw" | "8x13-bw" | "8x8r-bw" | "8x8r-sent" | "8x8u-bw" | "8x8u-sent" | "doc-8on16-bw" | "doc-8on16-sent" | "doc-8on16-sent-dim" | "silver16-bw")[]];
         readonly default: "auto";
         readonly ui: {
             readonly tab: "context";
@@ -4494,7 +4494,7 @@ export declare const SETTINGS_SCHEMA: {
             }, {
                 readonly value: "duckduckgo";
                 readonly label: "DuckDuckGo";
-                readonly description: "Uses DuckDuckGo Instant Answer API (no API key)";
+                readonly description: "Scrapes the DuckDuckGo HTML frontend (no API key)";
             }];
         };
     };

package/dist/types/edit/index.d.ts

@@ -18,6 +18,7 @@ export * from "./modes/patch";
 export * from "./modes/replace";
 export * from "./normalize";
 export * from "./renderer";
+export * from "./snapshot-details";
 export * from "./streaming";
 type TInput = typeof replaceEditSchema | typeof patchEditSchema | typeof hashlineEditParamsSchema | typeof applyPatchSchema;
 type HashlineParams = typeof hashlineEditParamsSchema.infer;

package/dist/types/edit/renderer.d.ts

@@ -29,6 +29,8 @@ export interface EditToolPerFileResult {
     oldText?: string;
     /** Source-of-truth content after the edit; `undefined` for delete operations. */
     newText?: string;
+    /** True when {@link pruneOversizedEditSnapshots} dropped `oldText`/`newText` from this entry. Aggregators check this to suppress misleading combined snapshots when at least one entry of a multi-entry single-path edit was pruned. */
+    snapshotsPruned?: boolean;
     /** Pre-move source path; set only when the edit moved/renamed the file. The header renders `sourcePath → path`. */
     sourcePath?: string;
 }
@@ -53,6 +55,8 @@ export interface EditToolDetails {
     oldText?: string;
     /** Source-of-truth content after the edit; `undefined` for delete operations. */
     newText?: string;
+    /** True when {@link pruneOversizedEditSnapshots} dropped `oldText`/`newText` from this entry. Aggregators check this to suppress misleading combined snapshots when at least one entry of a multi-entry single-path edit was pruned. */
+    snapshotsPruned?: boolean;
     /** Pre-move source path; set only when the edit moved/renamed the file. The header renders `sourcePath → path`. */
     sourcePath?: string;
 }

package/dist/types/edit/snapshot-details.d.ts

@@ -0,0 +1,33 @@
+/**
+ * Bound the size of the `oldText` / `newText` snapshots that edit-tool results
+ * carry in `details`. These fields hold the full pre/post file content; for
+ * large files they balloon the per-turn JSONL line and the session file
+ * (300 KB+ each on the cases reported in #3786) without paying for any LLM
+ * context (provider serializers send only `content`, never `details`).
+ *
+ * Only consumer of the raw snapshots is the ACP event mapper, which builds a
+ * `diff` ToolCallContent for ACP clients (Zed). When the snapshots are pruned
+ * the mapper returns `undefined` for that file and the text content still
+ * flows — diff visualization degrades gracefully for over-threshold edits.
+ */
+import type { EditToolDetails, EditToolPerFileResult } from "./renderer";
+/**
+ * Combined `oldText` + `newText` character budget for a single edit-tool
+ * result. Applies both per-entry (one file at a time) and as an aggregate
+ * across `perFileResults` (so a many-small-files batch can't accumulate
+ * unbounded snapshot bytes — see #3787 review).
+ *
+ * Picked so typical code-file edits keep ACP diff visualization while
+ * pathological cases (large generated files, full-file rewrites, or
+ * many-file batches) drop the raw snapshots before they hit the
+ * session JSONL.
+ */
+export declare const MAX_EDIT_SNAPSHOT_TEXT_CHARS = 32768;
+/**
+ * Prune oversized `oldText` / `newText` from an edit-tool details payload,
+ * recursing into `perFileResults` when present. Per-file overload comes first
+ * so the more specific shape (required `path`) wins overload resolution at
+ * the per-file call sites.
+ */
+export declare function pruneOversizedEditSnapshots(details: EditToolPerFileResult): EditToolPerFileResult;
+export declare function pruneOversizedEditSnapshots(details: EditToolDetails): EditToolDetails;

package/dist/types/mcp/config-writer.d.ts

@@ -51,3 +51,51 @@ export declare function readDisabledServers(filePath: string): Promise<string[]>
  * Add or remove a server name from the disabled servers list.
  */
 export declare function setServerDisabled(filePath: string, name: string, disabled: boolean): Promise<void>;
+/**
+ * Read the user-level force-enable list (allowlist that overrides a
+ * non-writable source config's `enabled: false`).
+ */
+export declare function readEnabledServers(filePath: string): Promise<string[]>;
+/**
+ * Add or remove a server name from the user-level force-enable list.
+ * The list overrides a discovered server's `enabled: false` flag but does
+ * NOT override the `disabledServers` denylist.
+ */
+export declare function setServerForceEnabled(filePath: string, name: string, force: boolean): Promise<void>;
+/** Paths and target state for toggling one MCP server across known config files. */
+export interface SetMcpServerEnabledOptions {
+    userPath: string;
+    projectPath: string;
+    /**
+     * Absolute path to the loaded row's source mcp.json. Provide ONLY for
+     * formats this codebase owns (native `.omp/mcp.json` and `mcp-json`
+     * `mcp.json`/`.mcp.json`). Tool-owned configs (opencode.json, claude.json,
+     * settings.json …) MUST be omitted; we never mutate another tool's file.
+     */
+    sourcePath?: string;
+    name: string;
+    enabled: boolean;
+}
+/**
+ * Flip a server's enabled/disabled state regardless of where it lives.
+ *
+ * Resolution order, mirroring `/mcp enable` / `/mcp disable` plus the dashboard
+ * fix for non-writable source configs:
+ *
+ * - Server found in `sourcePath` (writable) → write `enabled` on that entry.
+ * - Else server in project mcp.json → write `enabled` there.
+ * - Else server in user mcp.json → write `enabled` there.
+ * - Else (server defined in a tool-owned source like opencode.json, OR a
+ *   purely discovered server):
+ *   - Disable → add to the user-level `disabledServers` denylist.
+ *   - Enable → add to the user-level `enabledServers` allowlist so the
+ *     dashboard / runtime override the non-writable source's
+ *     `enabled: false` flag.
+ *
+ * Cleanup invariants — on every call:
+ * - Re-enable clears any stale denylist entry so a server disabled via
+ *   `/mcp disable` and re-enabled here doesn't stay suppressed.
+ * - Disable clears any stale allowlist entry so re-disabling a
+ *   force-enabled server actually takes effect.
+ */
+export declare function setMcpServerEnabled(options: SetMcpServerEnabledOptions): Promise<void>;

package/dist/types/mcp/oauth-flow.d.ts

@@ -59,12 +59,10 @@ export interface MCPOAuthConfig {
     /** OAuth scopes (space-separated) */
     scopes?: string;
     /**
-     * `prompt` parameter for the authorization request. Defaults to `"consent"`
-     * so the provider always shows its authorize screen instead of silently
-     * re-approving the browser's current session — without it, reauthorizing to
-     * switch accounts/workspaces is impossible once a session cookie exists
-     * (RFC 6749 §3.1 requires servers to ignore the param when unsupported).
-     * Set to `""` to omit the parameter entirely.
+     * `prompt` parameter for the authorization request. By default the parameter
+     * is omitted, matching the reference MCP SDK, except for `offline_access`
+     * requests where OIDC Core requires `prompt=consent` to issue refresh-token
+     * access. Set to `""` to omit the parameter entirely.
      */
     prompt?: string;
     /** Exact redirect URI to advertise to the provider */

package/dist/types/mcp/types.d.ts

@@ -88,7 +88,10 @@ export declare const MCP_CONFIG_SCHEMA_URL = "https://raw.githubusercontent.com/
 export interface MCPConfigFile {
     $schema?: string;
     mcpServers?: Record<string, MCPServerConfig>;
+    /** Names to hide regardless of any source `enabled` flag. Highest precedence. */
     disabledServers?: string[];
+    /** Names to force-enable when a non-writable source reports `enabled: false`. */
+    enabledServers?: string[];
 }
 /** MCP implementation info */
 export interface MCPImplementation {

package/dist/types/modes/controllers/tool-args-reveal.d.ts

@@ -6,13 +6,20 @@ type ToolArgsRevealControllerOptions = {
     getSmoothStreaming(): boolean;
     requestRender(): void;
 };
+type ToolArgsRevealTarget = {
+    rawInput: boolean;
+    exposeRawPartialJson: boolean;
+    fullArgs: Record<string, unknown>;
+};
 /**
  * Paces streamed tool-call arguments the same way StreamingRevealController
  * paces assistant text: providers that deliver `partialJson` in large batches
  * (or throttle their partial parses) would otherwise make write/edit/bash
  * streaming previews jump in chunks. Each pending tool call reveals its raw
  * argument stream at the shared 30fps cadence with the same adaptive
- * catch-up step, re-parsing the revealed prefix per frame.
+ * catch-up step. JSON prefixes are parsed only when enough new bytes arrive to
+ * change renderer-visible fields, while raw-prefix consumers still receive
+ * fresh `__partialJson` on every reveal frame.
  *
  * Reveal units are UTF-16 code units of the raw stream, not graphemes —
  * the prefix goes through a JSON parser rather than straight to the screen,
@@ -26,7 +33,7 @@ export declare class ToolArgsRevealController {
      * args to render right now. With smoothing disabled the full target passes
      * through in the caller's legacy shape (`{ ...args, __partialJson }`).
      */
-    setTarget(id: string, partialJson: string, rawInput: boolean, fullArgs: Record<string, unknown>): Record<string, unknown>;
+    setTarget(id: string, partialJson: string, target: ToolArgsRevealTarget): Record<string, unknown>;
     /** Attach the component future ticks push frames into. */
     bind(id: string, component: ToolArgsRevealComponent): void;
     /** Final arguments arrived (the JSON closed): drop the reveal so the

package/dist/types/session/agent-session.d.ts

@@ -15,7 +15,7 @@
 import type { InMemorySnapshotStore } from "@oh-my-pi/hashline";
 import { Agent, type AgentEvent, type AgentMessage, type AgentState, type AgentTool, type StreamFn, ThinkingLevel, type ToolChoiceDirective } from "@oh-my-pi/pi-agent-core";
 import { type CompactionResult, type ShakeConfig } from "@oh-my-pi/pi-agent-core/compaction";
-import type { AssistantMessage, Context, ImageContent, Message, MessageAttribution, Model, ProviderSessionState, ResetCreditAccountStatus, ResetCreditRedeemOutcome, ResetCreditTarget, ServiceTier, SimpleStreamOptions, TextContent, ToolChoice, UsageReport } from "@oh-my-pi/pi-ai";
+import type { AssistantMessage, Context, ImageContent, Message, MessageAttribution, Model, ProviderSessionState, ResetCreditAccountStatus, ResetCreditRedeemOutcome, ResetCreditTarget, ServiceTier, ServiceTierByFamily, ServiceTierFamily, SimpleStreamOptions, TextContent, ToolChoice, UsageReport } from "@oh-my-pi/pi-ai";
 import { Effort } from "@oh-my-pi/pi-ai";
 import { type AdvisorConfig } from "../advisor";
 import { type AsyncJob, type AsyncJobDeliveryState, AsyncJobManager } from "../async";
@@ -141,6 +141,8 @@ export interface AgentSessionConfig {
     }>;
     /** Initial session thinking selector. */
     thinkingLevel?: ConfiguredThinkingLevel;
+    /** Initial per-family service tiers (OpenAI / Anthropic / Google) for the live session. */
+    serviceTierByFamily?: ServiceTierByFamily;
     /** Prompt templates for expansion */
     promptTemplates?: PromptTemplate[];
     /** File-based slash commands for expansion */
@@ -184,6 +186,8 @@ export interface AgentSessionConfig {
      * main agent. Defaults to plain `streamSimple` when omitted.
      */
     advisorStreamFn?: StreamFn;
+    /** Hint that OpenAI Codex requests should prefer websocket transport when supported. */
+    preferWebsockets?: boolean;
     /** Provider payload hook used by the active session request path */
     onPayload?: SimpleStreamOptions["onPayload"];
     /** Provider response hook used by the active session request path */
@@ -470,6 +474,8 @@ export declare class AgentSession {
     setSessionSwitchReconciler(reconciler: (() => Promise<void>) | null): void;
     /** Provider-scoped mutable state store for transport/session caches. */
     get providerSessionState(): Map<string, ProviderSessionState>;
+    /** Hint forwarded to provider calls that support websocket transport. */
+    get preferWebsockets(): boolean | undefined;
     getHindsightSessionState(): HindsightSessionState | undefined;
     setHindsightSessionState(state: HindsightSessionState | undefined): HindsightSessionState | undefined;
     getMnemopiSessionState(): MnemopiSessionState | undefined;
@@ -542,7 +548,8 @@ export declare class AgentSession {
     get isAutoThinking(): boolean;
     /** The level `auto` resolved to for the current turn (undefined until classified). */
     autoResolvedThinkingLevel(): Effort | undefined;
-    get serviceTier(): ServiceTier | undefined;
+    /** Live per-family service tiers (OpenAI / Anthropic / Google). */
+    get serviceTierByFamily(): ServiceTierByFamily;
     /** Whether agent is currently streaming a response */
     get isStreaming(): boolean;
     get isAborting(): boolean;
@@ -915,23 +922,29 @@ export declare class AgentSession {
      */
     cycleThinkingLevel(): ConfiguredThinkingLevel | undefined;
     /**
-     * True when *any* fast-mode-granting service tier is configured, regardless
-     * of whether the active model's provider actually realizes it. Used by the
-     * toggle (`/fast on|off`) so re-toggling a scoped tier (`openai-only`,
-     * `claude-only`) doesn't silently broaden it to unscoped `priority`.
+     * True when the currently selected model's family is set to `priority` — the
+     * `/fast` on/off state for the active model. Returns false when no model is
+     * selected or the model exposes no service-tier family (e.g. Fireworks, which
+     * has its own Providers › Fireworks Tier toggle).
      *
-     * For "is fast mode actually applied to the next request?" use
-     * {@link isFastModeActive} instead — that one respects the model's provider.
+     * For "is priority actually applied to the next request?" use
+     * {@link isFastModeActive} instead.
      */
     isFastModeEnabled(): boolean;
     /**
-     * True when the configured `serviceTier` resolves to `"priority"` for the
-     * *currently selected model's provider*. Returns false for scoped tiers
-     * that don't match (e.g. `"openai-only"` on an anthropic model) and when
-     * no model is selected.
+     * True when `priority` is actually realized on the wire for the currently
+     * selected model (OpenAI/Google `service_tier`, direct Anthropic fast mode,
+     * or Fireworks priority). Returns false for tiers the active model can't
+     * realize and when no model is selected.
      */
     isFastModeActive(): boolean;
-    setServiceTier(serviceTier: ServiceTier | undefined): void;
+    /** Set one family's tier (or clear it with `undefined`); persists the change. */
+    setServiceTierFamily(family: ServiceTierFamily, tier: ServiceTier | undefined): void;
+    /**
+     * `/fast on|off` targets the family of the currently selected model: it sets
+     * (or clears) that family's `priority` tier. Models without a service-tier
+     * family (Fireworks, or providers with no tier knob) have nothing to toggle.
+     */
     setFastMode(enabled: boolean): void;
     toggleFastMode(): boolean;
     /**
@@ -1098,6 +1111,17 @@ export declare class AgentSession {
     get isEvalRunning(): boolean;
     /** Whether there are pending Python messages waiting to be flushed */
     get hasPendingPythonMessages(): boolean;
+    /**
+     * Surfaces (and consumes) IRC incoming asides that have reached this running
+     * session but have not yet been folded into the next model step.
+     *
+     * The inbox tool injects the formatted body into the tool result, so the
+     * model sees it once via the result. Leaving the record in
+     * {@link #pendingIrcAsides} would let the aside provider deliver it a second
+     * time at the next step boundary — including on `peek`, which is why peek
+     * also drains here.
+     */
+    drainPendingIrcInboxMessages(agentId: string): IrcMessage[];
     /**
      * Deliver an IRC message into this session (recipient side; called by the
      * IrcBus). Emits the `irc_message` session event for UI cards and injects

package/dist/types/session/messages.d.ts

@@ -184,7 +184,7 @@ export interface FileMentionMessage {
         /** File size in bytes, if known. */
         byteSize?: number;
         /** Why the file contents were omitted from auto-read. */
-        skippedReason?: "tooLarge";
+        skippedReason?: "tooLarge" | "binary";
         image?: ImageContent;
     }>;
     timestamp: number;

package/dist/types/session/session-context.d.ts

@@ -1,12 +1,12 @@
 import type { AgentMessage } from "@oh-my-pi/pi-agent-core";
-import type { ServiceTier } from "@oh-my-pi/pi-ai";
+import { type ServiceTierByFamily } from "@oh-my-pi/pi-ai";
 import { type CompactionEntry, type SessionEntry } from "./session-entries";
 export interface SessionContext {
     messages: AgentMessage[];
     thinkingLevel?: string;
     /** Configured thinking selector (`"auto"` or a concrete level) from the latest change. */
     configuredThinkingLevel?: string;
-    serviceTier?: ServiceTier;
+    serviceTier?: ServiceTierByFamily;
     /** Model roles: { default: "provider/modelId", small: "provider/modelId", ... } */
     models: Record<string, string>;
     /** Names of TTSR rules that have been injected this session */

package/dist/types/session/session-entries.d.ts

@@ -1,5 +1,5 @@
 import type { AgentMessage } from "@oh-my-pi/pi-agent-core";
-import type { ImageContent, MessageAttribution, ServiceTier, TextContent } from "@oh-my-pi/pi-ai";
+import type { ImageContent, MessageAttribution, ServiceTierByFamily, TextContent } from "@oh-my-pi/pi-ai";
 export declare const CURRENT_SESSION_VERSION = 3;
 export declare const SESSION_TITLE_SLOT_BYTES = 256;
 export declare const SESSION_TITLE_SLOT_ENTRY_TYPE = "title";
@@ -59,7 +59,7 @@ export interface ModelChangeEntry extends SessionEntryBase {
 }
 export interface ServiceTierChangeEntry extends SessionEntryBase {
     type: "service_tier_change";
-    serviceTier: ServiceTier | null;
+    serviceTier: ServiceTierByFamily | null;
 }
 export interface CompactionEntry<T = unknown> extends SessionEntryBase {
     type: "compaction";