muonroi-cli 1.8.5 → 1.9.0

package/dist/src/mcp/lsp-tools.js

@@ -8,7 +8,7 @@
  * deps are injected so unit tests never spawn a real language server.
  */
 import { z } from "zod";
-import { LSP_TOOL_OPERATIONS } from "../lsp/types.js";
+import { LSP_TOOL_OPERATIONS, } from "../lsp/types.js";
 function ok(data) {
     return { content: [{ type: "text", text: JSON.stringify(data) }] };
 }
@@ -18,6 +18,17 @@ function fail(error, message) {
         isError: true,
     };
 }
+/** Map a thrown error to the LspError tagged-union shape. */
+function mapLspError(e) {
+    if (e && typeof e === "object" && "kind" in e && "message" in e) {
+        const lspErr = e;
+        return { error: lspErr.kind, message: lspErr.message };
+    }
+    return {
+        error: "lsp_error",
+        message: e instanceof Error ? e.message : String(e),
+    };
+}
 async function defaultQuery(cwd, input) {
     const { queryLsp } = await import("../lsp/runtime.js");
     return queryLsp(cwd, input);
@@ -26,9 +37,29 @@ async function defaultEnabled(cwd) {
     const { isLspToolEnabled } = await import("../lsp/runtime.js");
     return isLspToolEnabled(cwd);
 }
+async function defaultWaitForDiagnostics(cwd, filePath, timeout = 1500) {
+    const { getOrCreateManager } = await import("../lsp/runtime.js");
+    return getOrCreateManager(cwd).waitForDiagnostics(filePath, timeout);
+}
+async function defaultImpactOfChange(cwd, filePath, query) {
+    const { getOrCreateManager } = await import("../lsp/runtime.js");
+    return getOrCreateManager(cwd).impactOfChange(filePath, query);
+}
+async function defaultMutationPreview(cwd, filePath, change) {
+    const { getOrCreateManager } = await import("../lsp/runtime.js");
+    return getOrCreateManager(cwd).lspMutationPreview(filePath, change);
+}
+async function defaultBeforeGrep(cwd, filePath, query) {
+    const { getOrCreateManager } = await import("../lsp/runtime.js");
+    return getOrCreateManager(cwd).lspBeforeGrep(filePath, query);
+}
 export function registerLspTools(server, deps = {}) {
     const query = deps.query ?? defaultQuery;
     const enabled = deps.enabled ?? defaultEnabled;
+    const waitForDiagnostics = deps.waitForDiagnostics ?? defaultWaitForDiagnostics;
+    const impactOfChange = deps.impactOfChange ?? defaultImpactOfChange;
+    const mutationPreview = deps.lspMutationPreview ?? defaultMutationPreview;
+    const lspBeforeGrep = deps.lspBeforeGrep ?? defaultBeforeGrep;
     server.registerTool("lsp_query", {
         description: "Semantic code intelligence via language servers. operation is one of: goToDefinition, findReferences, hover, documentSymbol, workspaceSymbol, goToImplementation, prepareCallHierarchy, incomingCalls, outgoingCalls. " +
             "filePath: absolute, or relative to the workspace root (cwd). line/character: 1-based (line 1 = first line, like an editor or file:line reference) — required for position-based ops; omit for documentSymbol; use query (not position) for workspaceSymbol. " +
@@ -48,7 +79,8 @@ export function registerLspTools(server, deps = {}) {
             isEnabled = await enabled(cwd);
         }
         catch (e) {
-            return fail("lsp_error", e instanceof Error ? e.message : String(e));
+            const m = mapLspError(e);
+            return fail(m.error, m.message);
         }
         if (!isEnabled) {
             return fail("lsp_disabled", "LSP tool is disabled in settings (lsp.enabled / lsp.tool)");
@@ -57,9 +89,68 @@ export function registerLspTools(server, deps = {}) {
             const resp = await query(cwd, args);
             return ok(resp);
         }
+        catch (e) {
+            const m = mapLspError(e);
+            return fail(m.error, m.message);
+        }
+    });
+    server.registerTool("lsp_waitForDiagnostics", {
+        description: "Wait for LSP diagnostics for a file. Returns { diagnostics, lspStatus, clean, metadata }. " +
+            "lspStatus: 'ok'|'partial'|'unavailable'. clean: true when zero error-level diagnostics. " +
+            "timeout defaults to 1500ms, max 5000ms. The lsp-before-grep policy allows grep fallback when lspStatus !== 'ok'.",
+        inputSchema: {
+            operation: z.literal("waitForDiagnostics"),
+            filePath: z.string().min(1).max(1000),
+            timeout: z.number().int().min(0).optional(),
+        },
+    }, async (args) => {
+        const cwd = process.cwd();
+        let isEnabled;
+        try {
+            isEnabled = await enabled(cwd);
+        }
+        catch (e) {
+            return fail("lsp_error", e instanceof Error ? e.message : String(e));
+        }
+        if (!isEnabled) {
+            return fail("lsp_disabled", "LSP tool is disabled in settings (lsp.enabled / lsp.tool)");
+        }
+        return ok(await waitForDiagnostics(cwd, args.filePath, args.timeout));
+    });
+    server.registerTool("lsp_impactOfChange", {
+        description: "Composite LSP analysis: returns { references, diagnostics, referencesComplete, safeToRename, clean, " +
+            "suggestedGuard, degraded, lspStatus, metadata }. Fans in diagnostics + references + rename safety over the " +
+            "frozen union (symbol file + reference files). Grep fallback allowed when lspStatus !== 'ok'.",
+        inputSchema: {
+            operation: z.literal("impactOfChange"),
+            filePath: z.string().min(1).max(1000),
+            query: z.string().max(1000).optional(),
+        },
+    }, async (args) => {
+        const cwd = process.cwd();
+        let isEnabled;
+        try {
+            isEnabled = await enabled(cwd);
+        }
         catch (e) {
             return fail("lsp_error", e instanceof Error ? e.message : String(e));
         }
+        if (!isEnabled) {
+            return fail("lsp_disabled", "LSP tool is disabled in settings (lsp.enabled / lsp.tool)");
+        }
+        return ok(await impactOfChange(cwd, args.filePath, args.query));
+    });
+    server.registerTool("lsp_mutationPreview", {
+        description: "Preview an LSP code mutation (stub). Returns { preview: [] }. " +
+            "No side-effects in slice 1. Registered in MUTATION_TOOLS set for routing through the mutation gate.",
+        inputSchema: {
+            operation: z.literal("mutationPreview"),
+            filePath: z.string().min(1).max(1000),
+            change: z.string().min(1).max(10000),
+        },
+    }, async (args) => {
+        const cwd = process.cwd();
+        return ok(await mutationPreview(cwd, args.filePath, args.change));
     });
 }
 //# sourceMappingURL=lsp-tools.js.map

package/dist/src/mcp/mcp-keychain.d.ts

@@ -1,11 +1,9 @@
 /**
  * src/mcp/mcp-keychain.ts
  *
- * Per-MCP keychain loader with env-var fallback.
- * Parallel module to src/providers/keychain.ts but keyed by McpKeyId
- * (MCP servers like Tavily) instead of ProviderId (LLM providers).
- *
- * Priority: OS keychain (keytar) > environment variable > null.
+ * Per-MCP key store, keyed by McpKeyId (MCP servers like Tavily). Backed by the
+ * env-store (`.env` file + process.env + Windows registry mirror) — the OS
+ * keychain (keytar) has been removed. Reads come straight from process.env.
  */
 export type McpKeyId = "tavily";
 export declare function setMcpKey(id: McpKeyId, key: string): Promise<boolean>;

package/dist/src/mcp/mcp-keychain.js

@@ -1,63 +1,24 @@
 /**
  * src/mcp/mcp-keychain.ts
  *
- * Per-MCP keychain loader with env-var fallback.
- * Parallel module to src/providers/keychain.ts but keyed by McpKeyId
- * (MCP servers like Tavily) instead of ProviderId (LLM providers).
- *
- * Priority: OS keychain (keytar) > environment variable > null.
+ * Per-MCP key store, keyed by McpKeyId (MCP servers like Tavily). Backed by the
+ * env-store (`.env` file + process.env + Windows registry mirror) — the OS
+ * keychain (keytar) has been removed. Reads come straight from process.env.
  */
+import { clearEnvVar, persistEnvVar } from "../providers/env-store.js";
 import { redactor } from "../utils/redactor.js";
-const KEYCHAIN_SERVICE = "muonroi-cli";
-const ACCOUNT_BY_MCP = {
-    tavily: "mcp-tavily",
-};
 const ENV_BY_MCP = {
     tavily: "TAVILY_API_KEY",
 };
 const MIN_KEY_LEN = 16;
-async function loadKeytar() {
-    try {
-        return (await import("keytar"));
-    }
-    catch {
-        return null;
-    }
-}
 export async function setMcpKey(id, key) {
     if (!key || key.length < MIN_KEY_LEN) {
         throw new Error(`Key for MCP '${id}' is too short (< ${MIN_KEY_LEN} chars).`);
     }
-    const kt = await loadKeytar();
-    if (!kt?.setPassword)
-        return false;
-    redactor.enrollSecret(key);
-    try {
-        await kt.setPassword(KEYCHAIN_SERVICE, ACCOUNT_BY_MCP[id], key);
-        return true;
-    }
-    catch (err) {
-        // Runtime backend failure (e.g. Linux without libsecret or no active secret service).
-        if (process.env.DEBUG || process.env.MUONROI_DEBUG_KEYCHAIN) {
-            console.error(`[mcp-keychain] setPassword backend error for ${id}:`, err?.message || err);
-        }
-        return false;
-    }
+    persistEnvVar(ENV_BY_MCP[id], key);
+    return true;
 }
 export async function getMcpKey(id) {
-    const kt = await loadKeytar();
-    if (kt) {
-        try {
-            const k = await kt.getPassword(KEYCHAIN_SERVICE, ACCOUNT_BY_MCP[id]);
-            if (k && k.length >= MIN_KEY_LEN) {
-                redactor.enrollSecret(k);
-                return k;
-            }
-        }
-        catch {
-            /* keytar backend failure → fall through to env */
-        }
-    }
     const envKey = process.env[ENV_BY_MCP[id]];
     if (envKey && envKey.length >= MIN_KEY_LEN) {
         redactor.enrollSecret(envKey);
@@ -66,9 +27,8 @@ export async function getMcpKey(id) {
     return null;
 }
 export async function deleteMcpKey(id) {
-    const kt = await loadKeytar();
-    if (!kt?.deletePassword)
-        return false;
-    return kt.deletePassword(KEYCHAIN_SERVICE, ACCOUNT_BY_MCP[id]);
+    const had = !!process.env[ENV_BY_MCP[id]];
+    clearEnvVar(ENV_BY_MCP[id]);
+    return had;
 }
 //# sourceMappingURL=mcp-keychain.js.map

package/dist/src/mcp/setup-guide-text.d.ts

@@ -6,4 +6,4 @@
  * and the muonroi-tools MCP server (src/mcp/tools-server.ts, for external agents).
  * Keeping it here avoids duplicating ~70 lines across the two.
  */
-export declare const SETUP_GUIDE_TEXT = "# muonroi-cli Setup Guide\n\n## Install (Recommended \u2014 NPM/Bun global install)\nNPM (requires Node >= 20):\n  npm install -g muonroi-cli\n\nBun (requires Bun >= 1.3):\n  bun add -g muonroi-cli\n\n## Install (Alternative \u2014 Prebuilt standalone binary, requires repo access)\nLinux / macOS:\n  curl -fsSL https://raw.githubusercontent.com/muonroi/muonroi-cli/master/install.sh | bash\n\nWindows PowerShell:\n  irm https://raw.githubusercontent.com/muonroi/muonroi-cli/master/install.ps1 | iex\n\nThe installers fetch a pre-compiled single binary from GitHub Releases.\n\n## First run\n- Wizard appears automatically.\n- Lists supported providers (DeepSeek + SiliconFlow ready; others via BYOK).\n- Four credential options: paste key, Bitwarden sync (B in /providers), keys export/import (encrypted bundle), or skip for later.\n- Keys land in OS keychain (keytar). Settings written to ~/.muonroi-cli/user-settings.json.\n- Role routing (leader/implement/verify/research) is configured for you.\n\nAfter setup: run `muonroi-cli doctor` to validate.\n\n## Essential commands\n- Interactive TUI: `muonroi-cli` (or `node dist/index.js` after build)\n- Headless one-shot: `muonroi-cli --prompt \"your task\" --max-tool-rounds 8`\n- Health + MCP nudge: `muonroi-cli doctor`\n- Update: `muonroi-cli update` (or set \"autoUpdate\": true in user-settings)\n- Keys move between machines: `muonroi-cli keys export file.json` then import on target\n- Native tools MCP (for external agents): `muonroi-cli tools-mcp` (stdio)\n- Harness driver MCP: `muonroi-cli mcp-driver`\n\n## MCP integration (for Claude Desktop, Cursor, other agents)\nAdd to your MCP client config:\n\n{\n  \"mcpServers\": {\n    \"muonroi-tools\": {\n      \"command\": \"bun\",\n      \"args\": [\"run\", \"/absolute/path/to/muonroi-cli/src/index.ts\", \"tools-mcp\"]\n    }\n  }\n}\n\n(Use absolute path. After `bun run build`: \"node\", \"dist/index.js\", \"tools-mcp\")\n\nThe CLI's OWN inner agent exposes these as NATIVE in-process tools (no MCP self-spawn):\n- setup_guide (this document)\n- ee_query / ee_health / ee_feedback \u2014 Experience Engine semantic recall + compaction checkpoints + feedback for learning\n- usage_forensics <id-prefix> \u2014 per-session cost/token forensics (peak input, cache hits, anomalies)\n- lsp_query \u2014 goToDefinition, findReferences, hover, symbols, call hierarchy etc.\n- selfverify_* \u2014 Tier-1 heuristic + Tier-2 agentic self-QA harness runs (start/poll/result/cancel/list)\n\nFor BB/.NET template recipes and package docs, also connect an external \"muonroi-docs\" MCP server if available (provides docs_search + setup_guide for the templates).\n\n## Development\ngit clone https://github.com/muonroi/muonroi-cli.git\ncd muonroi-cli && bun install\n\nbun run dev                 # run from source (TUI)\nbun run typecheck           # tsc --noEmit\nbun run test                # vitest (unit + headless)\nbunx vitest -c vitest.harness.config.ts run tests/harness/   # TUI E2E (named-pipes on Win, fd3/4 on POSIX)\nbun run build               # or build:binary for standalone exe\n\nSee AGENTS.md (quick ref + rules), CLAUDE.md (harness verification), README.md.\n\n## Verify\nmuonroi-cli doctor\n# Checks runtimes, catalog load, keychain, MCP servers enabled, council research MCP nudge, EE reachability, recent error rate.\n# Any \"warn\" entries tell you exactly what to enable (e.g. tavily for web research in council).\n\nFor BB-aware scaffolding (/ideal on a muonroi-building-block target): ensure dotnet SDK + the three Muonroi.*.Template packages are installed via NuGet; doctor surfaces missing feed/template cases.\n";
+export declare const SETUP_GUIDE_TEXT = "# muonroi-cli Setup Guide\n\n## Install (Recommended \u2014 NPM/Bun global install)\nNPM (requires Node >= 20):\n  npm install -g muonroi-cli\n\nBun (requires Bun >= 1.3):\n  bun add -g muonroi-cli\n\n## Install (Alternative \u2014 Prebuilt standalone binary, requires repo access)\nLinux / macOS:\n  curl -fsSL https://raw.githubusercontent.com/muonroi/muonroi-cli/master/install.sh | bash\n\nWindows PowerShell:\n  irm https://raw.githubusercontent.com/muonroi/muonroi-cli/master/install.ps1 | iex\n\nThe installers fetch a pre-compiled single binary from GitHub Releases.\n\n## First run\n- Wizard appears automatically.\n- Lists supported providers (DeepSeek + SiliconFlow ready; others via BYOK).\n- Four credential options: paste key, Bitwarden sync (bw in /providers), keys export/import (encrypted bundle), or skip for later.\n- Keys land in OS keychain (keytar). Settings written to ~/.muonroi-cli/user-settings.json.\n- Role routing (leader/implement/verify/research) is configured for you.\n\nAfter setup: run `muonroi-cli doctor` to validate.\n\n## Essential commands\n- Interactive TUI: `muonroi-cli` (or `node dist/index.js` after build)\n- Headless one-shot: `muonroi-cli --prompt \"your task\" --max-tool-rounds 8`\n- Health + MCP nudge: `muonroi-cli doctor`\n- Update: `muonroi-cli update` (or set \"autoUpdate\": true in user-settings)\n- Keys move between machines: `muonroi-cli keys export file.json` then import on target\n- Native tools MCP (for external agents): `muonroi-cli tools-mcp` (stdio)\n- Harness driver MCP: `muonroi-cli mcp-driver`\n\n## MCP integration (for Claude Desktop, Cursor, other agents)\nAdd to your MCP client config:\n\n{\n  \"mcpServers\": {\n    \"muonroi-tools\": {\n      \"command\": \"bun\",\n      \"args\": [\"run\", \"/absolute/path/to/muonroi-cli/src/index.ts\", \"tools-mcp\"]\n    }\n  }\n}\n\n(Use absolute path. After `bun run build`: \"node\", \"dist/index.js\", \"tools-mcp\")\n\nThe CLI's OWN inner agent exposes these as NATIVE in-process tools (no MCP self-spawn):\n- setup_guide (this document)\n- ee_query / ee_health / ee_feedback / ee_write \u2014 Experience Engine semantic recall + compaction checkpoints + feedback for learning\n- usage_forensics <id-prefix> \u2014 per-session cost/token forensics (peak input, cache hits, anomalies)\n- lsp_query \u2014 goToDefinition, findReferences, hover, symbols, call hierarchy etc.\n- selfverify_* \u2014 Tier-1 heuristic + Tier-2 agentic self-QA harness runs (start/poll/result/cancel/list)\n\n## Experience Engine over MCP (other agents)\n\nThe ee_* tools above are native to this CLI's agent. To give ANOTHER agent\n(Claude Code, Codex, Gemini CLI\u2026) the same brain, install the engine and register\nits own MCP server \u2014 no muonroi-cli needed:\n\n    npm i -g @muonroi/experience-engine\n    claude mcp add experience-engine -- exp-mcp\n\nOr in any mcpServers config:\n\n{\n  \"mcpServers\": {\n    \"experience-engine\": { \"command\": \"exp-mcp\" }\n  }\n}\n\nThe \"muonroi-tools\" server above deliberately does NOT serve ee_* \u2014 same brain,\none implementation, reachable without installing a CLI you have no use for.\n\nFor BB/.NET template recipes and package docs, also connect an external \"muonroi-docs\" MCP server if available (provides docs_search + setup_guide for the templates).\n\n## Development\ngit clone https://github.com/muonroi/muonroi-cli.git\ncd muonroi-cli && bun install\n\nbun run dev                 # run from source (TUI)\nbun run typecheck           # tsc --noEmit\nbun run test                # vitest (unit + headless)\nbunx vitest -c vitest.harness.config.ts run tests/harness/   # TUI E2E (named-pipes on Win, fd3/4 on POSIX)\nbun run build               # or build:binary for standalone exe\n\nSee AGENTS.md (quick ref + rules), CLAUDE.md (harness verification), README.md.\n\n## Verify\nmuonroi-cli doctor\n# Checks runtimes, catalog load, keychain, MCP servers enabled, council research MCP nudge, EE reachability, recent error rate.\n# Any \"warn\" entries tell you exactly what to enable (e.g. tavily for web research in council).\n\nFor BB-aware scaffolding (/ideal on a muonroi-building-block target): ensure dotnet SDK + the three Muonroi.*.Template packages are installed via NuGet; doctor surfaces missing feed/template cases.\n";

package/dist/src/mcp/setup-guide-text.js

@@ -27,7 +27,7 @@ The installers fetch a pre-compiled single binary from GitHub Releases.
 ## First run
 - Wizard appears automatically.
 - Lists supported providers (DeepSeek + SiliconFlow ready; others via BYOK).
-- Four credential options: paste key, Bitwarden sync (B in /providers), keys export/import (encrypted bundle), or skip for later.
+- Four credential options: paste key, Bitwarden sync (bw in /providers), keys export/import (encrypted bundle), or skip for later.
 - Keys land in OS keychain (keytar). Settings written to ~/.muonroi-cli/user-settings.json.
 - Role routing (leader/implement/verify/research) is configured for you.
 
@@ -58,11 +58,31 @@ Add to your MCP client config:
 
 The CLI's OWN inner agent exposes these as NATIVE in-process tools (no MCP self-spawn):
 - setup_guide (this document)
-- ee_query / ee_health / ee_feedback — Experience Engine semantic recall + compaction checkpoints + feedback for learning
+- ee_query / ee_health / ee_feedback / ee_write — Experience Engine semantic recall + compaction checkpoints + feedback for learning
 - usage_forensics <id-prefix> — per-session cost/token forensics (peak input, cache hits, anomalies)
 - lsp_query — goToDefinition, findReferences, hover, symbols, call hierarchy etc.
 - selfverify_* — Tier-1 heuristic + Tier-2 agentic self-QA harness runs (start/poll/result/cancel/list)
 
+## Experience Engine over MCP (other agents)
+
+The ee_* tools above are native to this CLI's agent. To give ANOTHER agent
+(Claude Code, Codex, Gemini CLI…) the same brain, install the engine and register
+its own MCP server — no muonroi-cli needed:
+
+    npm i -g @muonroi/experience-engine
+    claude mcp add experience-engine -- exp-mcp
+
+Or in any mcpServers config:
+
+{
+  "mcpServers": {
+    "experience-engine": { "command": "exp-mcp" }
+  }
+}
+
+The "muonroi-tools" server above deliberately does NOT serve ee_* — same brain,
+one implementation, reachable without installing a CLI you have no use for.
+
 For BB/.NET template recipes and package docs, also connect an external "muonroi-docs" MCP server if available (provides docs_search + setup_guide for the templates).
 
 ## Development

package/dist/src/mcp/tools-server.d.ts

@@ -12,6 +12,16 @@
  * (src/tools/native-tools.ts) — it no longer self-spawns this server. The two
  * surfaces share their cores (self-verify-runner.ts, setup-guide-text.ts, the
  * ee/forensics/lsp modules) so behaviour is identical.
+ *
+ * The ee_* tools are deliberately NOT here. The brain is the product and
+ * muonroi-cli is one consumer of it, so serving the brain to foreign agents is
+ * experience-engine's job, not ours: `npm i -g @muonroi/experience-engine` then
+ * `claude mcp add experience-engine -- exp-mcp`. Shipping them from here forced
+ * anyone who wanted the brain over MCP to install a whole CLI they had no use
+ * for, and left two implementations of the same four tools to drift apart. The
+ * CLI's own agent is unaffected — it calls the brain natively via
+ * src/ee/search.ts (no MCP hop), which is where ee_query/ee_write/ee_feedback
+ * live for us.
  */
 import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { JobManager, type Runner } from "./self-verify-jobs.js";

package/dist/src/mcp/tools-server.js

@@ -12,11 +12,20 @@
  * (src/tools/native-tools.ts) — it no longer self-spawns this server. The two
  * surfaces share their cores (self-verify-runner.ts, setup-guide-text.ts, the
  * ee/forensics/lsp modules) so behaviour is identical.
+ *
+ * The ee_* tools are deliberately NOT here. The brain is the product and
+ * muonroi-cli is one consumer of it, so serving the brain to foreign agents is
+ * experience-engine's job, not ours: `npm i -g @muonroi/experience-engine` then
+ * `claude mcp add experience-engine -- exp-mcp`. Shipping them from here forced
+ * anyone who wanted the brain over MCP to install a whole CLI they had no use
+ * for, and left two implementations of the same four tools to drift apart. The
+ * CLI's own agent is unaffected — it calls the brain natively via
+ * src/ee/search.ts (no MCP hop), which is where ee_query/ee_write/ee_feedback
+ * live for us.
  */
 import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 import { z } from "zod";
-import { registerEETools } from "./ee-tools.js";
 import { registerForensicsTools } from "./forensics-tools.js";
 import { registerLspTools } from "./lsp-tools.js";
 import { JobManager } from "./self-verify-jobs.js";
@@ -129,7 +138,6 @@ export function createToolsServer(runner = defaultRunner) {
     const server = new McpServer({ name: "muonroi-tools", version: "0.1.0" });
     const jm = new JobManager(runner);
     registerSelfVerifyTools(server, jm);
-    registerEETools(server);
     registerForensicsTools(server);
     registerLspTools(server);
     registerSetupGuideTool(server);

package/dist/src/models/catalog.json

@@ -1,6 +1,6 @@
 {
-  "version": "2.14",
-  "updated_at": "2026-07-09",
+  "version": "2.15",
+  "updated_at": "2026-07-15",
   "description": "Local fallback catalog: deepseek + zai + opencode-go + xai. routing.council multi-provider debate; routing.vision_proxy Z.ai/xai backend for text-only models (deepseek etc.). provider_policies.peak_hour from vendor docs. Pricing APPROXIMATE.",
   "models": [
     {
@@ -46,10 +46,10 @@
       "web_research_kind": null
     },
     {
-      "id": "gpt-5.4-mini",
-      "name": "GPT-5.4 mini (OpenAI ChatGPT OAuth)",
+      "id": "gpt-5.5",
+      "name": "GPT-5.5 (OpenAI ChatGPT OAuth)",
       "provider": "openai",
-      "tier": "fast",
+      "tier": "premium",
       "context_window": 256000,
       "max_output_tokens": 32000,
       "input_price_per_million": 0,
@@ -57,11 +57,11 @@
       "reasoning": true,
       "thinking_type": "enabled",
       "supports_effort": true,
-      "default_reasoning_effort": "low",
-      "description": "OpenAI GPT-5.4 mini via ChatGPT subscription OAuth (chatgpt.com/backend-api/codex, Responses API). Subscription-billed — per-token price N/A (0 placeholder). Added to local fallback catalog so OAuth-granted OpenAI models resolve when the CP catalog endpoint is unreachable.",
-      "aliases": ["gpt-5.4-mini-oauth"],
+      "default_reasoning_effort": "high",
+      "description": "OpenAI GPT-5.5 — frontier model for complex coding, research, and real-world work. Via ChatGPT subscription OAuth (chatgpt.com/backend-api/codex, Responses API). Subscription-billed — per-token price N/A (0 placeholder). Added to local fallback catalog so OAuth-granted OpenAI models resolve when the CP catalog endpoint is unreachable.",
+      "aliases": ["gpt-5.5-oauth"],
       "supports_vision": false,
-      "roles": ["research", "implement", "verify"],
+      "roles": ["leader", "implement", "verify", "research"],
       "native_web_research": true,
       "web_research_kind": "web-tool"
     },
@@ -69,7 +69,7 @@
       "id": "gpt-5.4",
       "name": "GPT-5.4 (OpenAI ChatGPT OAuth)",
       "provider": "openai",
-      "tier": "premium",
+      "tier": "balanced",
       "context_window": 256000,
       "max_output_tokens": 32000,
       "input_price_per_million": 0,
@@ -78,18 +78,18 @@
       "thinking_type": "enabled",
       "supports_effort": true,
       "default_reasoning_effort": "medium",
-      "description": "OpenAI GPT-5.4 via ChatGPT subscription OAuth (chatgpt.com/backend-api/codex, Responses API). Subscription-billed — per-token price N/A (0 placeholder). Added to local fallback catalog so OAuth-granted OpenAI models resolve when the CP catalog endpoint is unreachable.",
+      "description": "OpenAI GPT-5.4 — strong model for everyday coding. Via ChatGPT subscription OAuth (chatgpt.com/backend-api/codex, Responses API). Subscription-billed — per-token price N/A (0 placeholder). Added to local fallback catalog so OAuth-granted OpenAI models resolve when the CP catalog endpoint is unreachable.",
       "aliases": ["gpt-5.4-oauth"],
       "supports_vision": false,
-      "roles": ["leader", "implement", "verify"],
+      "roles": ["implement", "verify", "research"],
       "native_web_research": true,
       "web_research_kind": "web-tool"
     },
     {
-      "id": "gpt-5.3-codex",
-      "name": "GPT-5.3 Codex (OpenAI ChatGPT OAuth)",
+      "id": "gpt-5.4-mini",
+      "name": "GPT-5.4 mini (OpenAI ChatGPT OAuth)",
       "provider": "openai",
-      "tier": "balanced",
+      "tier": "fast",
       "context_window": 256000,
       "max_output_tokens": 32000,
       "input_price_per_million": 0,
@@ -97,11 +97,11 @@
       "reasoning": true,
       "thinking_type": "enabled",
       "supports_effort": true,
-      "default_reasoning_effort": "medium",
-      "description": "OpenAI GPT-5.3 Codex (coding-specialized) via ChatGPT subscription OAuth (chatgpt.com/backend-api/codex, Responses API). Subscription-billed — per-token price N/A (0 placeholder). Added to local fallback catalog so OAuth-granted OpenAI models resolve when the CP catalog endpoint is unreachable.",
-      "aliases": ["gpt-5.3-codex-oauth", "codex"],
+      "default_reasoning_effort": "low",
+      "description": "OpenAI GPT-5.4 mini — small, fast, and cost-efficient model for simpler coding tasks. Via ChatGPT subscription OAuth (chatgpt.com/backend-api/codex, Responses API). Subscription-billed — per-token price N/A (0 placeholder). Added to local fallback catalog so OAuth-granted OpenAI models resolve when the CP catalog endpoint is unreachable.",
+      "aliases": ["gpt-5.4-mini-oauth"],
       "supports_vision": false,
-      "roles": ["implement", "verify"],
+      "roles": ["research", "implement", "verify"],
       "native_web_research": true,
       "web_research_kind": "web-tool"
     },

package/dist/src/orchestrator/ask-user.d.ts

@@ -0,0 +1,61 @@
+/**
+ * src/orchestrator/ask-user.ts
+ *
+ * Types + pure helpers for the model-callable `ask_user` tool.
+ *
+ * The agent calls `ask_user` from its OWN intent — typically after a
+ * `convene_council` conclusion, when it is in an implementation discussion and
+ * wants the human's go/no-go before building. Per the hard no-hardcode
+ * directive, the CLI does NOT synthesise the option set or decide the branch:
+ * the question, options, and default ALL come from the agent's tool input. The
+ * CLI only renders the card and returns the human's answer AS the tool result
+ * string; the agent decides what to do next.
+ *
+ * This mirrors the blocking safety-override askcard plumbing
+ * (setSafetyOverrideHandler → resolver ref-map → answer/cancel drain) but the
+ * card content is agent-supplied, never CLI-built.
+ */
+import type { CouncilQuestionData } from "../types/index.js";
+/** One agent-supplied choice. `value` defaults to `label` when omitted. */
+export interface AskUserOption {
+    label: string;
+    description?: string;
+    value?: string;
+}
+/** The agent's request payload (validated from the tool input). */
+export interface AskUserAskInfo {
+    /** The question to put to the human. Required. */
+    question: string;
+    /** Optional short context/detail shown under the question. */
+    context?: string;
+    /**
+     * Agent-supplied choices. When empty/omitted the card is a free-text prompt.
+     * The CLI NEVER injects its own options.
+     */
+    options?: AskUserOption[];
+    /**
+     * Agent-supplied default selection index into `options`. Clamped into range;
+     * defaults to a neutral 0 (the agent's FIRST option) — index 0 is NOT a CLI
+     * recommendation, just first-listed.
+     */
+    defaultIndex?: number;
+}
+/** Sentinel returned to the agent when the human dismisses the card (Esc). */
+export declare const ASK_USER_DISMISSED = "(the user dismissed the question without answering)";
+/**
+ * Resolve the agent-facing answer string for a chosen option index (or free
+ * text). Returns the option's `value` (falling back to its `label`), the raw
+ * free-text, or the dismissed sentinel. This is the value the AI-SDK
+ * `execute()` returns — the agent reads it and decides the follow-up.
+ */
+export declare function resolveAskUserAnswer(info: AskUserAskInfo, choice: {
+    index?: number;
+    text?: string;
+    cancelled?: boolean;
+}): string;
+/**
+ * Build the CouncilQuestionData card from the agent's request. Options come
+ * ONLY from `info.options`; with none, the card is a single free-text field.
+ * `defaultIndex` is clamped into range (neutral 0 otherwise).
+ */
+export declare function buildAskUserQuestion(info: AskUserAskInfo, questionId: string): CouncilQuestionData;

package/dist/src/orchestrator/ask-user.js

@@ -0,0 +1,65 @@
+/**
+ * src/orchestrator/ask-user.ts
+ *
+ * Types + pure helpers for the model-callable `ask_user` tool.
+ *
+ * The agent calls `ask_user` from its OWN intent — typically after a
+ * `convene_council` conclusion, when it is in an implementation discussion and
+ * wants the human's go/no-go before building. Per the hard no-hardcode
+ * directive, the CLI does NOT synthesise the option set or decide the branch:
+ * the question, options, and default ALL come from the agent's tool input. The
+ * CLI only renders the card and returns the human's answer AS the tool result
+ * string; the agent decides what to do next.
+ *
+ * This mirrors the blocking safety-override askcard plumbing
+ * (setSafetyOverrideHandler → resolver ref-map → answer/cancel drain) but the
+ * card content is agent-supplied, never CLI-built.
+ */
+/** Sentinel returned to the agent when the human dismisses the card (Esc). */
+export const ASK_USER_DISMISSED = "(the user dismissed the question without answering)";
+/**
+ * Resolve the agent-facing answer string for a chosen option index (or free
+ * text). Returns the option's `value` (falling back to its `label`), the raw
+ * free-text, or the dismissed sentinel. This is the value the AI-SDK
+ * `execute()` returns — the agent reads it and decides the follow-up.
+ */
+export function resolveAskUserAnswer(info, choice) {
+    if (choice.cancelled)
+        return ASK_USER_DISMISSED;
+    if (typeof choice.text === "string" && choice.text.length > 0)
+        return choice.text;
+    const opts = info.options ?? [];
+    if (typeof choice.index === "number" && choice.index >= 0 && choice.index < opts.length) {
+        const opt = opts[choice.index];
+        return opt.value ?? opt.label;
+    }
+    return ASK_USER_DISMISSED;
+}
+/**
+ * Build the CouncilQuestionData card from the agent's request. Options come
+ * ONLY from `info.options`; with none, the card is a single free-text field.
+ * `defaultIndex` is clamped into range (neutral 0 otherwise).
+ */
+export function buildAskUserQuestion(info, questionId) {
+    const rawOptions = info.options ?? [];
+    const options = rawOptions.map((o) => ({
+        label: o.label,
+        value: o.value ?? o.label,
+        kind: "choice",
+        description: o.description,
+    }));
+    const hasOptions = options.length > 0;
+    const clampedDefault = hasOptions ? Math.min(Math.max(0, info.defaultIndex ?? 0), options.length - 1) : 0;
+    return {
+        questionId,
+        question: info.question,
+        context: info.context,
+        isRequired: true,
+        phase: "ask-user",
+        // No options → a free-text prompt (single freetext field); otherwise the
+        // agent-supplied choices, verbatim.
+        options: hasOptions ? options : [{ label: "Your answer", value: "", kind: "freetext" }],
+        defaultIndex: clampedDefault,
+    };
+}
+//# sourceMappingURL=ask-user.js.map

package/dist/src/orchestrator/compaction.d.ts

@@ -1,5 +1,4 @@
 import { type ModelMessage } from "ai";
-import type { ProviderFactory as LegacyProvider } from "../providers/runtime.js";
 /**
  * A single action the compaction proposer model wants to take on one message.
  */
@@ -21,7 +20,7 @@ export interface CompactionProposal {
  * The model sees the current messages and decides whether compaction is needed at all.
  * Returns null if the proposer call fails (caller should fall back gracefully).
  */
-export declare function proposeCompaction(provider: LegacyProvider, modelId: string, messages: ModelMessage[], signal?: AbortSignal): Promise<CompactionProposal | null>;
+export declare function proposeCompaction(modelId: string, messages: ModelMessage[], signal?: AbortSignal): Promise<CompactionProposal | null>;
 export interface CompactionSettings {
     reserveTokens: number;
     keepRecentTokens: number;
@@ -78,4 +77,4 @@ export interface CompactionSummaryResult {
     summary: string;
     usage: CompactionUsage;
 }
-export declare function generateCompactionSummary(provider: LegacyProvider, modelId: string, preparation: PreparedCompaction, customInstructions?: string, signal?: AbortSignal): Promise<CompactionSummaryResult>;
+export declare function generateCompactionSummary(modelId: string, preparation: PreparedCompaction, customInstructions?: string, signal?: AbortSignal): Promise<CompactionSummaryResult>;

package/dist/src/orchestrator/compaction.js

@@ -11,9 +11,9 @@ import { countTokens } from "./token-counter.js";
  * The model sees the current messages and decides whether compaction is needed at all.
  * Returns null if the proposer call fails (caller should fall back gracefully).
  */
-export async function proposeCompaction(provider, modelId, messages, signal) {
+export async function proposeCompaction(modelId, messages, signal) {
     try {
-        const runtime = resolveModelRuntime(provider, modelId);
+        const runtime = resolveModelRuntime(modelId);
         const compactCaps = getProviderCapabilities(requireRuntimeProvider(runtime));
         const serialized = serializeConversation(messages);
         const result = await generateText({
@@ -517,7 +517,7 @@ function readUsage(usage) {
             : 0;
     return { promptTokens: prompt, completionTokens: completion };
 }
-async function summarizeConversation(provider, modelId, messages, reserveTokens, customInstructions, previousSummary, promptOverride, signal) {
+async function summarizeConversation(modelId, messages, reserveTokens, customInstructions, previousSummary, promptOverride, signal) {
     const serialized = serializeConversation(messages);
     const promptParts = [serialized];
     // Improvement from session df2dbb878984 analysis: detect meta-analysis (self-evaluation of agent/PIL)
@@ -535,7 +535,7 @@ async function summarizeConversation(provider, modelId, messages, reserveTokens,
     if (customInstructions?.trim()) {
         promptParts.push(`Additional focus: ${customInstructions.trim()}`);
     }
-    const runtime = resolveModelRuntime(provider, modelId);
+    const runtime = resolveModelRuntime(modelId);
     const compactCaps = getProviderCapabilities(requireRuntimeProvider(runtime));
     const result = await generateText({
         model: runtime.model,
@@ -558,14 +558,14 @@ function addUsage(a, b) {
         completionTokens: a.completionTokens + b.completionTokens,
     };
 }
-export async function generateCompactionSummary(provider, modelId, preparation, customInstructions, signal) {
+export async function generateCompactionSummary(modelId, preparation, customInstructions, signal) {
     const { messagesToSummarize, turnPrefixMessages, isSplitTurn, previousSummary, settings } = preparation;
     if (isSplitTurn && turnPrefixMessages.length > 0) {
         const [historyResult, prefixResult] = await Promise.all([
             messagesToSummarize.length > 0
-                ? summarizeConversation(provider, modelId, messagesToSummarize, settings.reserveTokens, customInstructions, previousSummary, undefined, signal)
+                ? summarizeConversation(modelId, messagesToSummarize, settings.reserveTokens, customInstructions, previousSummary, undefined, signal)
                 : Promise.resolve({ summary: previousSummary?.trim() || "", usage: emptyUsage() }),
-            summarizeConversation(provider, modelId, turnPrefixMessages, settings.reserveTokens, undefined, undefined, TURN_PREFIX_SUMMARIZATION_PROMPT, signal),
+            summarizeConversation(modelId, turnPrefixMessages, settings.reserveTokens, undefined, undefined, TURN_PREFIX_SUMMARIZATION_PROMPT, signal),
         ]);
         const usage = addUsage(historyResult.usage, prefixResult.usage);
         if (historyResult.summary && prefixResult.summary) {
@@ -573,6 +573,6 @@ export async function generateCompactionSummary(provider, modelId, preparation,
         }
         return { summary: (historyResult.summary || prefixResult.summary).trim(), usage };
     }
-    return summarizeConversation(provider, modelId, messagesToSummarize, settings.reserveTokens, customInstructions, previousSummary, undefined, signal);
+    return summarizeConversation(modelId, messagesToSummarize, settings.reserveTokens, customInstructions, previousSummary, undefined, signal);
 }
 //# sourceMappingURL=compaction.js.map