@gajae-code/coding-agent 0.4.5 → 0.5.1

package/CHANGELOG.md

@@ -2,6 +2,67 @@
 
 ## [Unreleased]
 
+## [0.5.1] - 2026-06-14
+
+### Added
+
+- `gjc --mode rpc` registers each live session in a cross-process registry (`<agent-dir>/rpc-sessions/<id>.json`) on start and removes it on shutdown, so other processes can enumerate running RPC sessions. The Python `gjc_rpc` client exposes `list_sessions()` / `RpcClient.list_sessions()` returning typed `SessionHandle`s and reaps records whose owning process is gone (issue 10; foundation for reattach/issue 09).
+- `gjc --mode rpc --listen <socket-path>` runs a persistent Unix-domain-socket RPC server: the `AgentSession` outlives client disconnects (no stdin-EOF teardown) and a client can disconnect and reconnect to the same live session over the socket. The session is registered with `transport: "socket"` and the socket path as its `endpoint`, so it is discoverable/attachable via the registry. The stdio path is unchanged (frame output routes through a swappable sink shared by both transports) (issue 09).
+
+### Fixed
+
+- Subagent/job status panels now freeze a job's elapsed timer the moment it stops running (`completed`/`failed`/`cancelled`/`paused`) instead of counting up against `Date.now()` forever. `AsyncJob` records an `endTime` on the first terminal/pause transition and a shared `jobElapsedMs` helper backs the `subagent` panel, the `job` tool, `/jobs`, and the command-controller job line, so a finished subagent stops at its real duration while siblings keep ticking.
+- RPC control-plane hardening (from dogfooding `gjc --mode rpc`): `dispatchRpcCommand` now wraps the command switch so failures return a correlated response carrying the request `id` and the real command name, instead of dropping the id and mislabeling handler exceptions as `parse`; `set_thinking_level`/`set_steering_mode`/`set_follow_up_mode`/`set_interrupt_mode` validate their inputs and reject out-of-contract values instead of silently corrupting session state; `negotiate_unattended` rejects unknown scopes/action classes with `invalid_unattended_declaration` and merges the mandatory `prompt` scope plus its `command.prompt` action floor into the accepted grant (so prompt/`workflow_gate_response` are never locked out); and read-only/control RPC commands no longer consume the unattended `max_tool_calls` budget while wall-time enforcement is preserved. `docs/rpc.md`'s first `workflow_gate` example now matches the canonical `RpcWorkflowGate` shape.
+- RPC mode no longer head-of-line-blocks control/cancellation commands behind a long-running command: the stdin loop now dispatches ordered commands through a serial chain (so causal order is preserved — e.g. `get_state` after `bash` still observes the bash result) without blocking the reader, and routes `abort`/`abort_bash`/`abort_retry` on an immediate fast lane so they reach in-flight work. `abort_bash` can now cancel a running `bash`, and a slow `compact`/`handoff`/`login` no longer freezes the whole control plane. Shutdown drains in-flight commands (bounded) so their responses are still emitted on stdin EOF.
+- Counted active prompts and agent-initiated custom messages in pre-prompt context maintenance so background task notifications trigger compaction before they can overflow the next model request.
+- Bounded monitor task-notification payloads to a compact tail window while preserving full background job output for job inspection.
+
+### Changed
+
+- Expanded coordinator MCP coordination status into a canonical polling snapshot for sessions, session states, turns, questions, reports, and bounded event summaries, and documented that Hermes/coordinator consumption is polling/await rather than push subscription.
+
+## [0.5.0] - 2026-06-13
+
+### Fixed
+
+- Fixed forced `tool_choice` 400s ("tool_choice forces tool use is not compatible with this model") looping after `ast_edit` previews: named queue directives (resolve protocol, eager `todo_write` enforcement, subagent `yield` reminders) now enqueue only when the model supports exact named forcing; otherwise they degrade silently to the existing steer reminder without a forced `tool_choice`, and a runtime-discovered incapability drops the in-flight directive instead of requeueing it.
+- `models.yml` compat blocks now accept the `toolChoiceSupport` enum (`none`/`auto`/`required`/`named`) alongside the legacy `supportsToolChoice`/`supportsForcedToolChoice` booleans, mirrored in the generated JSON schema.
+### Added
+
+- Made `/model` open to a preset-first landing view: provider-grouped presets with live auth checkmarks, highlight-to-expand tiers, a full clamped role→model preview before applying, and a session/default apply scope choice; typing still jumps straight to model search, "Browse all models" opens the classic tabbed selector, and temporary-only quick-switch bypasses the landing entirely.
+- Rebuilt the builtin model profile catalog as 25 profiles: `codex-{eco,medium,pro}` on `gpt-5.5` effort spreads, a single `opencodego` preset, `claude-opus`, `{glm,kimi-coding-plan,mimo,grok,cursor,minimax}-{eco,medium,pro}` trios with thinking levels clamped to provider support, and `opus-codex`/`codex-opencodego` combos. Legacy profile names (including the `*-standard` family and retired Fable presets) were removed clean-break and now fail with the available-profile listing.
+- Added a post-`/login` smart preset recommendation: when login succeeds and no profile is active, prompts "Apply <preset> now?" (session-only on confirm); when a profile is active, prints a one-line hint instead. The active profile is tracked in-memory on the session with rollback-safe activation.
+- Bundled `kimi-code/kimi-k2.7-code` and `minimax-code/minimax-v3` model entries; MiniMax presets use the canonical `minimax-code` provider id throughout.
+- Added a harness receipt JSONL spool exporter for gajae receipt-runtime interop: configured `gjc harness --receipt-spool-dir <dir>` / `GJC_RECEIPT_SPOOL_DIR` now appends persisted native `ReceiptEnvelope` records as `{cursor,envelope}` lines to `spool.jsonl`, with restart-safe 12-digit cursors and installed-package smoke coverage (#545).
+- Added Gajae Trinity compatibility golden fixtures and tests that pin ReceiptEnvelope hash basis, validator compatibility, and replayable RPC exchange shape for downstream receipt-runtime interop.
+- Optimization Suite v3 Lane 1 (RSS): large resident text in persisted sessions is now backed by an ephemeral session-scoped disk cache (`EphemeralBlobStore`) instead of being pinned in JS heap for the whole session lifetime; canonical JSONL persistence, reload, and export semantics are byte-identical (resident refs never persist). Missing resident text cache blobs now surface a typed `ResidentBlobMissingError` instead of silently leaking `blob:sha256:` refs into provider payloads, UI, or exports. `getEntries()`/`buildSessionContext()` are served from revision-keyed WeakRef caches below the public ownership boundary (callers still receive caller-owned copies). Fixture retained heap −82%, RSS −55%, warm `getEntries()` p95 −80% on 10k-entry sessions; one-shot `exportFromFile()` now closes its session manager.
+- Added process-isolated deterministic TUI render-golden capture and fixtures for interactive editor overlays, rich-text resizing, multiplexer viewport repaint, sixel image line preservation, Termux height diffs, and transcript shrink/clear regressions.
+
+### Removed
+
+- Removed the hardcoded OpenAI Codex role-preset action from the model selector; builtin model profiles are now the only preset concept.
+- Removed retired Fable model profiles (`claude-fable`, `fable-codex`) after `claude-fable-5` was removed upstream.
+
+### Changed
+
+- Optimization Suite v3 Lane 3 (serialization): session-switch message comparison now uses per-message cached source strings + xxHash64 as an accelerator (source-string compare remains the authority; collision fallback tested) — unchanged-session compares −95% median. The secret obfuscator precomputes a longest-first combined regex (single-pass replace, −70% median/−77% p95 on 100 secrets × 1MiB) with a conservative sequential fallback whenever secrets overlap each other or any replacement/placeholder contains a secret — output bytes are identical in all cases. Intra-line diff rendering gains byte-identical fast paths for identical lines and whitespace-token-aligned prefix/suffix spans (identical −67%, single-token −60%; long lines skip the scan). Mental-model LCS keeps legacy dense-DP tie-break semantics (a Hunt-Szymanski variant was rejected for changing rendered bytes). Provider-visible fork-context seeds use JSON-semantic cloning instead of structuredClone.
+- Tightened tool-block rendering to remove vertical padding and rely on Spacer-only separation, reducing transcript noise while preserving stable render-golden output.
+- Improved the Bun runtime version guard diagnostic: when the Bun running `gjc` is older than the required version, the error now names the exact detected Bun runtime path and prints a platform-specific upgrade and PATH fix (Windows gets the `irm bun.sh/install.ps1|iex` reinstall plus a `%USERPROFILE%\.bun\bin` PATH hint) instead of a bare `bun upgrade` (#525).
+- Aligned the `codex-standard` and `codex-pro` model profiles on the `openai-codex/gpt-5.5` baseline so they no longer default to stale mixed model generations (`gpt-5.4`, `gpt-5.2-codex`, `gpt-5.1-codex-max`, `gpt-5.3-codex-spark`); the profiles now differentiate purely by per-role reasoning effort (#532).
+- Reduced the default RPC `get_state` payload by omitting static `dumpTools` and `systemPrompt` fields unless requested via `include: ["tools", "systemPrompt"]` (#539).
+- Updated `/model` documentation and generated docs index for the rebuilt preset catalog and preset-first selector.
+
+### Fixed
+
+- Tightened the Windows/psmux tmux provider boundary: `gjc team` now honors `GJC_TMUX_COMMAND` (not just `GJC_TEAM_TMUX_COMMAND`) so the team leader resolves the same multiplexer as `gjc session`/`gjc --tmux`; and when a multiplexer lists a session that lacks GJC's `@gjc-profile` ownership tag, `gjc session status` now returns `gjc_tmux_session_untagged` with a `detail` hint and `gjc team` reports the same cause, instead of a bare `gjc_tmux_session_not_found` / `unmanaged_tmux_session`. Documented that alternative multiplexers such as psmux on Windows are not fully supported because they do not round-trip tmux user options (#531).
+- Hardened RPC stdio lifecycle behavior: `gjc --mode rpc` now reports malformed JSONL frames as parse-error responses without killing the session, flushes durable session state before exiting on EOF/shutdown, and has red-team coverage for attached persistence, reload, malformed-frame recovery, and concurrent child-session isolation.
+- Hardened the harness RPC submit/router contract so `submit` is no longer advertised or accepted during finalizing/non-idle lifecycle windows, non-idle RPC state reports `submitted:false` with a retryable gate, and degraded owner endpoints fall back to `owner-not-live` without false acceptance (#544).
+- Ran estimated context maintenance before sending a new prompt, including tool-output pruning and threshold compaction, so large tool results appended after the last assistant turn cannot push the next model request over the context window.
+- `gjc team` now self-heals a missing `@gjc-profile` ownership tag when the current leader pane was genuinely launched by `gjc --tmux` (detected via `GJC_TMUX_LAUNCHED=1`): the session is re-tagged with `set-option` and startup proceeds, instead of hard-failing with `unmanaged_tmux_session` after a mid-startup attach failure or registry race stripped the tag. Sessions without the GJC launch marker are still rejected unchanged, so foreign tmux sessions cannot be hijacked.
+- Subagent task receipts and live render output now warn when requested role-agent models are substituted by auth fallback or provider-reported assistant model mismatch, including session model-change annotations for server-side substitutions (#559).
+- Converted Cursor wire shell timeouts from millisecond values to bash-tool seconds so delegated Cursor-native shell calls honor the expected timeout units.
+- Fixed pi-shell bash fixups on multibyte UTF-8 commands by converting parser source indexes to byte offsets before stripping `head`/`tail` pipelines.
+
 ## [0.4.5] - 2026-06-12
 
 ### Added
@@ -17,6 +78,7 @@
 - Reduced compiled CLI startup and native bundle pressure with default-small grammar loading, tokenizer tiering, and compiled fast-help paths.
 - Preserved dev/main release metadata and changelog consistency for the 0.4.5 lockstep release.
 
+- Added native `gjc ultragoal steer --kind` support for documented steering mutations beyond `add_subgoal`, including split, reorder, wording revision, ledger annotation, and blocked-goal supersession contracts with structured audit expectations.
 ### Fixed
 
 - Kept the unified `goal` tool registered and active by default whenever `goal.enabled` is true, including explicit tool subsets and `gjc ultragoal create-goals` arming flows.

package/dist/types/async/job-manager.d.ts

@@ -4,6 +4,13 @@ export interface AsyncJob {
     type: "bash" | "task";
     status: "running" | "completed" | "failed" | "cancelled" | "paused";
     startTime: number;
+    /**
+     * Wall-clock ms when the job left the `running` state (completed, failed,
+     * cancelled, or paused). Undefined while running. Frozen on the first
+     * terminal/pause transition so elapsed-time renderers stop counting once a
+     * job is no longer active instead of growing forever against `Date.now()`.
+     */
+    endTime?: number;
     label: string;
     abortController: AbortController;
     promise: Promise<void>;
@@ -18,6 +25,13 @@ export interface AsyncJob {
      */
     ownerId?: string;
 }
+/**
+ * Elapsed wall-clock ms for a job, frozen once it stops running. While the job
+ * is active (`endTime` undefined) this counts against `now`; after it stops it
+ * returns the fixed `endTime - startTime` span so status renderers do not keep
+ * incrementing a completed job's timer.
+ */
+export declare function jobElapsedMs(job: Pick<AsyncJob, "startTime" | "endTime">, now?: number): number;
 export interface AsyncJobMetadata {
     subagent?: {
         id: string;
@@ -77,6 +91,12 @@ export interface SubagentRecord {
         message?: string;
         createdAt: number;
     };
+    /** Resolved model the subagent was asked to use, e.g. "openai-codex/gpt-5.5". */
+    requestedModel?: string;
+    /** Model actually used after auth fallback (#985); equals requestedModel when no fallback. */
+    effectiveModel?: string;
+    /** True when the requested model lacked credentials and the subagent fell back to the parent model. */
+    modelFellBack?: boolean;
 }
 /** Lightweight, manager-owned resume payload. The async layer treats `data` as opaque. */
 export interface ResumeDescriptor {
@@ -184,6 +204,12 @@ export declare class AsyncJobManager {
     };
     /** Register or replace the canonical record for a subagent. */
     registerSubagentRecord(record: SubagentRecord): void;
+    /** Patch model metadata onto an existing subagent record (best-effort; no-op if unknown). */
+    updateSubagentModel(subagentId: string, model: {
+        requestedModel?: string;
+        effectiveModel?: string;
+        modelFellBack?: boolean;
+    }): void;
     getSubagentRecord(subagentId: string, filter?: AsyncJobFilter): SubagentRecord | undefined;
     getSubagentRecords(filter?: AsyncJobFilter): SubagentRecord[];
     registerLiveHandle(subagentId: string, handle: SubagentLiveHandle): void;

package/dist/types/cli/args.d.ts

@@ -24,6 +24,7 @@ export interface Args {
     mode?: Mode;
     noSession?: boolean;
     sessionDir?: string;
+    rpcListen?: string;
     providerSessionId?: string;
     fork?: string;
     models?: string[];

package/dist/types/cli/list-models.d.ts

@@ -1,4 +1,5 @@
 import type { ModelRegistry } from "../config/model-registry";
+import { type ExtensionFactory } from "../extensibility/extensions";
 /**
  * List available models, optionally filtered by search pattern
  */
@@ -11,6 +12,11 @@ export interface RunListModelsOptions {
     cwd: string;
     /** CLI-supplied extension paths (e.g. from `-e <path>`). */
     additionalExtensionPaths?: string[];
+    /** In-process extension factories to load without filesystem discovery. */
+    extensionFactories?: Array<{
+        factory: ExtensionFactory;
+        name: string;
+    }>;
     /** Extension paths configured under `extensions:` in user settings. */
     settingsExtensions?: string[];
     /** Disabled extension ids from settings (`disabledExtensions`). */

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

@@ -0,0 +1,26 @@
+import { Command } from "@gajae-code/utils/cli";
+export default class Gc extends Command {
+    static description: string;
+    static strict: boolean;
+    static flags: {
+        json: import("@gajae-code/utils/cli").FlagDescriptor<"boolean"> & {
+            char: string;
+            description: string;
+            default: boolean;
+        };
+        prune: import("@gajae-code/utils/cli").FlagDescriptor<"boolean"> & {
+            description: string;
+            default: boolean;
+        };
+        force: import("@gajae-code/utils/cli").FlagDescriptor<"boolean"> & {
+            description: string;
+            default: boolean;
+        };
+        "dry-run": import("@gajae-code/utils/cli").FlagDescriptor<"boolean"> & {
+            description: string;
+            default: boolean;
+        };
+    };
+    static examples: string[];
+    run(): Promise<void>;
+}

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

@@ -34,6 +34,9 @@ export default class Harness extends Command {
             description: string;
             default: boolean;
         };
+        "receipt-spool-dir": import("@gajae-code/utils/cli").FlagDescriptor<"string"> & {
+            description: string;
+        };
     };
     static examples: string[];
     run(): Promise<void>;

package/dist/types/config/file-lock-gc.d.ts

@@ -0,0 +1,5 @@
+/**
+ * GC adapter for config file-locks (`<file>.lock` dirs holding `{pid, timestamp}`).
+ */
+import type { GcStoreAdapter } from "../gjc-runtime/gc-runtime";
+export declare const fileLocksGcAdapter: GcStoreAdapter;

package/dist/types/config/file-lock.d.ts

@@ -3,4 +3,11 @@ export interface FileLockOptions {
     retries?: number;
     retryDelayMs?: number;
 }
+/** @internal */
+export declare function readFileLockInfoForGc(lockDir: string): Promise<{
+    pid: number;
+    timestamp: number;
+} | null>;
+/** @internal */
+export declare function removeFileLockDirForGc(lockDir: string): Promise<void>;
 export declare function withFileLock<T>(filePath: string, fn: () => Promise<T>, options?: FileLockOptions): Promise<T>;

package/dist/types/config/model-profile-activation.d.ts

@@ -3,15 +3,22 @@ import type { Api, Model } from "@gajae-code/ai";
 import type { AgentSession } from "../session/agent-session";
 import { type ModelRegistry } from "./model-registry";
 import type { Settings } from "./settings";
+type ModelProfileActivationSession = Pick<AgentSession, "model" | "thinkingLevel" | "sessionId"> & {
+    setModelTemporary?: AgentSession["setModelTemporary"];
+    setActiveModelProfile?: (name: string | undefined) => void;
+    getActiveModelProfile?: () => string | undefined;
+};
 export interface PrepareModelProfileActivationOptions {
-    session: Pick<AgentSession, "model" | "thinkingLevel" | "sessionId">;
+    session: ModelProfileActivationSession;
     modelRegistry: Pick<ModelRegistry, "getModelProfile" | "getModelProfiles" | "getAvailableModelProfileNames" | "getApiKeyForProvider" | "getAll" | "resolveCanonicalModel" | "getCanonicalVariants" | "getCanonicalId">;
     settings: Pick<Settings, "get">;
     profileName: string;
 }
 export interface PreparedModelProfileActivation {
     profileName: string;
-    session: Pick<AgentSession, "model" | "thinkingLevel" | "sessionId" | "setModelTemporary">;
+    session: ModelProfileActivationSession & {
+        setModelTemporary: AgentSession["setModelTemporary"];
+    };
     settings: Pick<Settings, "get" | "override" | "set" | "flush">;
     previousModel: Model<Api> | undefined;
     previousThinkingLevel: ThinkingLevel | undefined;
@@ -19,6 +26,7 @@ export interface PreparedModelProfileActivation {
     defaultModel: Model<Api> | undefined;
     defaultThinkingLevel: ThinkingLevel | undefined;
     agentModelOverrides: Record<string, string>;
+    previousActiveModelProfile: string | undefined;
 }
 export declare function formatModelProfileCredentialError(profileName: string, providers: readonly string[]): string;
 export declare function prepareModelProfileActivation(options: PrepareModelProfileActivationOptions): Promise<PreparedModelProfileActivation>;
@@ -28,3 +36,4 @@ export declare function applyPreparedModelProfileActivation(prepared: PreparedMo
 export declare function activateModelProfile(options: PrepareModelProfileActivationOptions, applyOptions?: {
     persistDefault?: boolean;
 }): Promise<void>;
+export {};

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

@@ -14,6 +14,13 @@ export interface ResolvedProfileBinding {
 export declare function deriveModelProfileMappedProviders(definition: Pick<ModelProfileDefinition, "modelMapping">): string[];
 export declare function aggregateModelProfileRequiredProviders(requiredProviders: readonly string[], definition: Pick<ModelProfileDefinition, "modelMapping">): string[];
 export declare const BUILTIN_MODEL_PROFILES: readonly ModelProfileDefinition[];
+export interface ModelProfilePresentation {
+    displayName: string;
+    providerGroup: string;
+}
+export declare function getModelProfilePresentation(name: string): ModelProfilePresentation;
+export declare function groupModelProfilesForPresetLanding(profiles: ReadonlyMap<string, ModelProfileDefinition>): Map<string, ModelProfileDefinition[]>;
+export declare function recommendModelProfileForProvider(providerId: string, profiles: ReadonlyMap<string, ModelProfileDefinition>): ModelProfileDefinition | undefined;
 export declare function mergeModelProfiles(userProfiles?: ModelsConfig["profiles"]): Map<string, ModelProfileDefinition>;
 export declare function resolveProfileBindings(definition: ModelProfileDefinition): ResolvedProfileBinding;
 export declare function formatAvailableProfileNames(profiles: ReadonlyMap<string, ModelProfileDefinition>): string;

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

@@ -71,6 +71,7 @@ export declare const ModelsConfigFile: ConfigFile<{
             requiresAssistantContentForToolCalls?: boolean | undefined;
             supportsToolChoice?: boolean | undefined;
             supportsForcedToolChoice?: boolean | undefined;
+            toolChoiceSupport?: "auto" | "named" | "none" | "required" | undefined;
             disableReasoningOnForcedToolChoice?: boolean | undefined;
             disableReasoningOnToolChoice?: boolean | undefined;
             thinkingFormat?: "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;
@@ -147,6 +148,7 @@ export declare const ModelsConfigFile: ConfigFile<{
                 requiresAssistantContentForToolCalls?: boolean | undefined;
                 supportsToolChoice?: boolean | undefined;
                 supportsForcedToolChoice?: boolean | undefined;
+                toolChoiceSupport?: "auto" | "named" | "none" | "required" | undefined;
                 disableReasoningOnForcedToolChoice?: boolean | undefined;
                 disableReasoningOnToolChoice?: boolean | undefined;
                 thinkingFormat?: "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;
@@ -219,6 +221,7 @@ export declare const ModelsConfigFile: ConfigFile<{
                 requiresAssistantContentForToolCalls?: boolean | undefined;
                 supportsToolChoice?: boolean | undefined;
                 supportsForcedToolChoice?: boolean | undefined;
+                toolChoiceSupport?: "auto" | "named" | "none" | "required" | undefined;
                 disableReasoningOnForcedToolChoice?: boolean | undefined;
                 disableReasoningOnToolChoice?: boolean | undefined;
                 thinkingFormat?: "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;

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

@@ -133,6 +133,8 @@ export declare function resolveModelOverrideWithAuthFallback(modelPatterns: stri
     thinkingLevel?: ThinkingLevel;
     explicitThinkingLevel: boolean;
     authFallbackUsed: boolean;
+    requestedModel?: Model<Api>;
+    fallbackReason?: "auth_unavailable";
 }>;
 /**
  * Resolve a list of role patterns to the first matching model.

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

@@ -31,6 +31,12 @@ export declare const OpenAICompatSchema: z.ZodObject<{
     requiresAssistantContentForToolCalls: z.ZodOptional<z.ZodBoolean>;
     supportsToolChoice: z.ZodOptional<z.ZodBoolean>;
     supportsForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
+    toolChoiceSupport: z.ZodOptional<z.ZodEnum<{
+        auto: "auto";
+        named: "named";
+        none: "none";
+        required: "required";
+    }>>;
     disableReasoningOnForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
     disableReasoningOnToolChoice: z.ZodOptional<z.ZodBoolean>;
     thinkingFormat: z.ZodOptional<z.ZodEnum<{
@@ -184,6 +190,12 @@ export declare const ModelOverrideSchema: z.ZodObject<{
         requiresAssistantContentForToolCalls: z.ZodOptional<z.ZodBoolean>;
         supportsToolChoice: z.ZodOptional<z.ZodBoolean>;
         supportsForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
+        toolChoiceSupport: z.ZodOptional<z.ZodEnum<{
+            auto: "auto";
+            named: "named";
+            none: "none";
+            required: "required";
+        }>>;
         disableReasoningOnForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
         disableReasoningOnToolChoice: z.ZodOptional<z.ZodBoolean>;
         thinkingFormat: z.ZodOptional<z.ZodEnum<{
@@ -291,6 +303,12 @@ export declare const ModelsConfigSchema: z.ZodObject<{
             requiresAssistantContentForToolCalls: z.ZodOptional<z.ZodBoolean>;
             supportsToolChoice: z.ZodOptional<z.ZodBoolean>;
             supportsForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
+            toolChoiceSupport: z.ZodOptional<z.ZodEnum<{
+                auto: "auto";
+                named: "named";
+                none: "none";
+                required: "required";
+            }>>;
             disableReasoningOnForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
             disableReasoningOnToolChoice: z.ZodOptional<z.ZodBoolean>;
             thinkingFormat: z.ZodOptional<z.ZodEnum<{
@@ -446,6 +464,12 @@ export declare const ModelsConfigSchema: z.ZodObject<{
                 requiresAssistantContentForToolCalls: z.ZodOptional<z.ZodBoolean>;
                 supportsToolChoice: z.ZodOptional<z.ZodBoolean>;
                 supportsForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
+                toolChoiceSupport: z.ZodOptional<z.ZodEnum<{
+                    auto: "auto";
+                    named: "named";
+                    none: "none";
+                    required: "required";
+                }>>;
                 disableReasoningOnForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
                 disableReasoningOnToolChoice: z.ZodOptional<z.ZodBoolean>;
                 thinkingFormat: z.ZodOptional<z.ZodEnum<{
@@ -580,6 +604,12 @@ export declare const ModelsConfigSchema: z.ZodObject<{
                 requiresAssistantContentForToolCalls: z.ZodOptional<z.ZodBoolean>;
                 supportsToolChoice: z.ZodOptional<z.ZodBoolean>;
                 supportsForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
+                toolChoiceSupport: z.ZodOptional<z.ZodEnum<{
+                    auto: "auto";
+                    named: "named";
+                    none: "none";
+                    required: "required";
+                }>>;
                 disableReasoningOnForcedToolChoice: z.ZodOptional<z.ZodBoolean>;
                 disableReasoningOnToolChoice: z.ZodOptional<z.ZodBoolean>;
                 thinkingFormat: z.ZodOptional<z.ZodEnum<{

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

@@ -1,3 +1,4 @@
+import type { Effort } from "@gajae-code/ai/model-thinking";
 import { type SkillDiscoverySettings } from "./skill-settings-defaults";
 /** Unified settings schema - single source of truth for all settings.
  *
@@ -697,13 +698,13 @@ export declare const SETTINGS_SCHEMA: {
     };
     readonly defaultThinkingLevel: {
         readonly type: "enum";
-        readonly values: readonly import("@gajae-code/ai").Effort[];
-        readonly default: "high";
+        readonly values: readonly Effort[];
+        readonly default: Effort;
         readonly ui: {
             readonly tab: "model";
             readonly label: "Thinking Level";
             readonly description: "Reasoning depth for thinking-capable models";
-            readonly options: readonly import("../thinking").ThinkingLevelMetadata[];
+            readonly options: readonly import("../thinking-metadata").ThinkingLevelMetadata[];
         };
     };
     readonly hideThinkingBlock: {

package/dist/types/coordinator/contract.d.ts

@@ -1,4 +1,4 @@
 export declare const COORDINATOR_MCP_PROTOCOL_VERSION = "2024-11-05";
 export declare const COORDINATOR_MCP_SERVER_NAME = "gjc-coordinator-mcp";
-export declare const COORDINATOR_MCP_TOOL_NAMES: readonly ["gjc_coordinator_list_sessions", "gjc_coordinator_read_status", "gjc_coordinator_read_tail", "gjc_coordinator_list_questions", "gjc_coordinator_list_artifacts", "gjc_coordinator_read_artifact", "gjc_coordinator_read_coordination_status", "gjc_coordinator_register_session", "gjc_coordinator_start_session", "gjc_coordinator_send_prompt", "gjc_coordinator_submit_question_answer", "gjc_coordinator_read_turn", "gjc_coordinator_await_turn", "gjc_coordinator_report_status"];
+export declare const COORDINATOR_MCP_TOOL_NAMES: readonly ["gjc_coordinator_list_sessions", "gjc_coordinator_read_status", "gjc_coordinator_read_tail", "gjc_coordinator_list_questions", "gjc_coordinator_list_artifacts", "gjc_coordinator_read_artifact", "gjc_coordinator_read_coordination_status", "gjc_coordinator_watch_events", "gjc_coordinator_register_session", "gjc_coordinator_start_session", "gjc_coordinator_send_prompt", "gjc_coordinator_submit_question_answer", "gjc_coordinator_read_turn", "gjc_coordinator_await_turn", "gjc_coordinator_report_status"];
 export type CoordinatorToolName = (typeof COORDINATOR_MCP_TOOL_NAMES)[number];

package/dist/types/defaults/gjc/extensions/grok-build/index.d.ts

@@ -0,0 +1 @@
+export { default } from "../grok-cli-vendor/src/index.js";

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/index.d.ts

@@ -0,0 +1 @@
+export { default } from './provider/register.js';

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/models/catalog.d.ts

@@ -0,0 +1,25 @@
+/**
+ * Model definitions for Grok CLI's API.
+ */
+export interface GrokCliModelConfig {
+    id: string;
+    name: string;
+    reasoning: boolean;
+    input: ('text' | 'image')[];
+    cost: {
+        input: number;
+        output: number;
+        cacheRead: number;
+        cacheWrite: number;
+    };
+    contextWindow: number;
+    maxTokens: number;
+    /** Models that don't support reasoning.effort get a thinkingLevelMap. */
+    thinkingLevelMap?: Record<string, string | null>;
+}
+export declare function supportsReasoningEffort(modelId: string): boolean;
+/**
+ * Resolve the active model list. If `GJC_GROK_CLI_MODELS` is set,
+ * it filters/reorders the fallback list; unknown IDs get sensible defaults.
+ */
+export declare function resolveModels(): GrokCliModelConfig[];

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/payload/sanitize.d.ts

@@ -0,0 +1,27 @@
+/**
+ * Payload sanitization for xAI's Responses API via cli-chat-proxy.grok.com.
+ *
+ * xAI's endpoint has quirks compared to stock OpenAI:
+ *   - Replayed or encrypted `reasoning` items in input cause 400 errors.
+ *   - `reasoning.effort` is only supported on a subset of models.
+ *   - Empty-string content items cause validation failures.
+ *   - `function_call_output.output` cannot contain image arrays.
+ *   - `image_url` parts must be normalized to `input_image` with data URIs.
+ *   - Local image paths must be resolved to base64 data URIs.
+ *   - xAI rejects `role: "developer"` and `role: "system"` in the input
+ *     array; these must be moved to top-level `instructions`.
+ *   - xAI uses `text.format` instead of OpenAI's `response_format`.
+ *   - xAI uses `prompt_cache_key` for conversation caching.
+ *   - xAI doesn't support `prompt_cache_retention`.
+ *
+ * Additional Grok CLI-specific behavior:
+ *   - Adds x-grok-* headers for client identification
+ *   - Uses prompt_cache_key for session affinity
+ */
+/**
+ * Sanitize a provider request payload for xAI's Responses API via
+ * cli-chat-proxy.grok.com.
+ *
+ * Returns the modified payload.  Mutates the input in place for efficiency.
+ */
+export declare function sanitizePayload(params: Record<string, unknown>, modelId: string, sessionId: string | undefined, cwd: string): Record<string, unknown>;

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/provider/billing.d.ts

@@ -0,0 +1,8 @@
+interface BillingUsage {
+    monthlyLimit: number;
+    used: number;
+    billingPeriodEnd: string;
+}
+export declare function fetchBillingUsage(token: string): Promise<BillingUsage>;
+export declare function formatQuota(usage: BillingUsage | undefined): string[];
+export {};

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/provider/register.d.ts

@@ -0,0 +1,5 @@
+/**
+ * GJC Grok Build provider — SuperGrok OAuth + cli-chat-proxy models.
+ */
+import type { ExtensionAPI } from '@gajae-code/coding-agent';
+export default function registerGrokCli(api: ExtensionAPI): void;

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/provider/stream.d.ts

@@ -0,0 +1,10 @@
+import type { Api, AssistantMessageEventStream, Context, Model, SimpleStreamOptions } from '@gajae-code/ai';
+/**
+ * Stream function that adds Grok CLI-specific headers to requests.
+ *
+ * GJC Grok Build extension sends cli-chat-proxy headers (see agent.models.grok-cli.yml):
+ *   - x-grok-conv-id: <session/conversation ID>
+ *   - x-grok-model-override: <model ID>
+ *   - x-xai-token-auth: xai-grok-cli
+ */
+export declare function streamGrokCli(model: Model<Api>, context: Context, options?: SimpleStreamOptions): AssistantMessageEventStream;

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/provider/usage.d.ts

@@ -0,0 +1,2 @@
+import type { ExtensionAPI } from '@gajae-code/coding-agent';
+export declare function registerUsageCommand(api: Pick<ExtensionAPI, 'registerCommand'>): void;

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/shared/base-url.d.ts

@@ -0,0 +1,2 @@
+export declare function getBaseUrl(): string;
+export declare function isGrokBuildBaseUrlOverrideIgnored(): boolean;

package/dist/types/defaults/gjc/extensions/grok-cli-vendor/src/shared/errors.d.ts

@@ -0,0 +1,38 @@
+/**
+ * Typed error for xAI OAuth failures.
+ *
+ * Codes allow the login flow and stream handlers to distinguish
+ * retryable failures (network) from fatal ones (revoked refresh token).
+ */
+export declare class XaiOAuthError extends Error {
+    readonly code: string;
+    readonly reloginRequired: boolean;
+    constructor(message: string, code: string, reloginRequired?: boolean);
+}
+/** Well-known error codes. */
+export declare const XaiErrorCode: {
+    /** OIDC discovery failed (network, invalid response). */
+    readonly DISCOVERY_FAILED: 'discovery_failed';
+    /** Discovery endpoint returned a non-xAI origin. */
+    readonly DISCOVERY_INVALID_ORIGIN: 'discovery_invalid_origin';
+    /** Authorization was denied or errored in the browser. */
+    readonly AUTHORIZATION_FAILED: 'authorization_failed';
+    /** CSRF state mismatch between request and callback. */
+    readonly STATE_MISMATCH: 'state_mismatch';
+    /** Callback did not include an authorization code. */
+    readonly CODE_MISSING: 'code_missing';
+    /** Token exchange failed (network, invalid response). */
+    readonly TOKEN_EXCHANGE_FAILED: 'token_exchange_failed';
+    /** Token exchange returned an invalid payload. */
+    readonly TOKEN_EXCHANGE_INVALID: 'token_exchange_invalid';
+    /** Refresh token is missing or empty. */
+    readonly REFRESH_MISSING: 'refresh_missing';
+    /** Token refresh failed (expired, revoked). */
+    readonly REFRESH_FAILED: 'refresh_failed';
+    /** No credentials stored. */
+    readonly AUTH_MISSING: 'auth_missing';
+    /** Loopback callback server could not bind. */
+    readonly CALLBACK_BIND_FAILED: 'callback_bind_failed';
+    /** Loopback callback timed out. */
+    readonly CALLBACK_TIMEOUT: 'callback_timeout';
+};

package/dist/types/defaults/gjc-grok-cli.d.ts

@@ -0,0 +1,5 @@
+import type { ExtensionFactory } from "../extensibility/extensions/types";
+export declare const BUNDLED_GROK_BUILD_EXTENSION_ID = "bundled:grok-build";
+export declare function getBundledGrokBuildExtensionFactory(): ExtensionFactory;
+export declare function getBundledGrokCliModelDefaults(): string;
+export declare function assertBundledGrokCliDefaults(): Promise<void>;

package/dist/types/extensibility/extensions/index.d.ts

@@ -3,6 +3,7 @@
  */
 export type { SlashCommandInfo, SlashCommandLocation, SlashCommandSource } from "../slash-commands";
 export { discoverAndLoadExtensions, ExtensionRuntimeNotInitializedError, loadExtensionFromFactory, loadExtensions, } from "./loader";
+export * from "./prefix-command-bridge";
 export * from "./runner";
 export * from "./types";
 export * from "./wrapper";

package/dist/types/extensibility/extensions/prefix-command-bridge.d.ts

@@ -0,0 +1,35 @@
+import { type ExecResult } from "../../exec/exec";
+import type { ExtensionContext, InputEvent, InputEventResult } from "./types";
+export declare const OOO_BRIDGE_RECURSION_ENV = "_OUROBOROS_GJC_BRIDGE_DEPTH";
+export declare const OOO_BRIDGE_CONTINUE_EXIT_CODE = 78;
+export declare const OOO_BRIDGE_TIMEOUT_ENV = "OUROBOROS_GJC_BRIDGE_TIMEOUT_MS";
+export interface ExactPrefixCommandBridgeOptions {
+    /** Bare command prefix to intercept, without trailing whitespace. */
+    prefix: string;
+    /** Command executable to run when the prefix matches. */
+    command: string;
+    /** Arguments inserted before the intercepted input text. */
+    args?: string[];
+    /** Environment variable used as the recursion-depth guard. */
+    recursionEnv?: string;
+    /** Exit code that maps to extension pass-through instead of handled input. */
+    continueExitCode?: number;
+    /** Optional dispatch timeout in milliseconds. */
+    timeout?: number;
+    /** Dispatch implementation. Defaults to the shared command executor in the extension context cwd. */
+    dispatch?: (command: string, args: string[], ctx: ExtensionContext, options: {
+        timeout?: number;
+    }) => Promise<ExecResult>;
+}
+/**
+ * Build an extension `input` handler for an exact-prefix command bridge.
+ *
+ * Matching input is passed to `command` as `args + [event.text]`. A zero exit code
+ * handles the input, `continueExitCode` returns pass-through, and any other
+ * non-zero exit code surfaces an error and handles the input so the failed
+ * command is not forwarded to the model. The recursion
+ * guard prevents extension-originated or nested dispatch from re-entering the
+ * bridge while the child command runs.
+ */
+export declare function createExactPrefixCommandBridge(options: ExactPrefixCommandBridgeOptions): (event: InputEvent, ctx: ExtensionContext) => Promise<InputEventResult>;
+export declare function createOuroborosOooBridge(): (event: InputEvent, ctx: ExtensionContext) => Promise<InputEventResult>;