glm-coding-router 0.1.1 → 0.2.0

package/README.md

@@ -1,248 +1,286 @@
-# GLM Coding Router
-
-GLM Coding Plan workers for Claude Code and Codex, on Windows.
-
-Claude Code and Codex stay your orchestrators — they keep responsibility for requirements,
-architecture, review, and integration. `glm-coding-router` delegates well-scoped
-implementation work (exploration, CRUD, boilerplate, tests, mechanical refactoring) to
-GLM workers via Z.ai's Anthropic-compatible endpoint.
-
-One global npm install replaces the manual `.cmd` shim setup:
-
-```text
-Claude / Codex → shell → glm-worker → claude.exe harness → Z.ai endpoint → GLM Coding Plan
-```
-
-## Architecture
-
-```text
-                         Developer
-                             │
-             ┌───────────────┴───────────────┐
-             ▼                               ▼
-       Claude Code                         Codex
-             │                               │
-             └───────────────┬───────────────┘
-                        shell command
-                             │
-             ┌───────────────┼───────────────┐
-             ▼               ▼               ▼
-         glm-chat        glm-worker      glm-review
-             │               │               │
-             └───────────────┼───────────────┘
-                        claude.exe
-              (injected environment only)
-                             │
-                             ▼
-              https://api.z.ai/api/anthropic
-                             │
-                             ▼
-                    GLM Coding Plan
-                    GLM-5.3 / GLM-5.3-Flash
-```
-
-## Requirements
-
-- Windows 10/11 (v0.1; Linux/macOS planned for v0.2)
-- Node.js >= 20
-- Claude Code (`claude.exe`) — the GLM commands run on the Claude Code harness
-- Codex (optional — Claude-only setups are fully supported)
-- A Z.ai Coding Plan API key
-
-No Anthropic pay-as-you-go, no OpenAI API, no LiteLLM, no proxy.
-
-## Installation
-
-```powershell
-npm install -g glm-coding-router
-glm-router init
-```
-
-`npx glm-coding-router init` also works for a one-off check, but the global install is
-what puts `glm-worker` on your PATH long-term.
-
-## Quick start
-
-After `glm-router init`:
-
-```powershell
-glm-chat
-glm-worker "Implement validation and add tests"
-glm-review "Analyze the auth module"
-```
-
-## glm-chat
-
-Interactive GLM-backed Claude Code session. Resolves the Z.ai key, locates `claude.exe`,
-injects the Z.ai environment **into the child process only**, and spawns it with
-pass-through arguments:
-
-```powershell
-glm-chat
-glm-chat --version
-glm-chat --any-claude-flag
-```
-
-Your normal `claude` command and its authentication are untouched.
-
-## glm-worker
-
-Headless implementation worker:
-
-```powershell
-glm-worker "Implement validation and add tests"
-```
-
-Or via stdin (a structured task packet):
-
-```powershell
-@"
-TASK:
-Implement refresh token validation.
-
-SCOPE:
-internal/auth/
-
-VALIDATION:
-go test ./internal/auth/...
-"@ | glm-worker
-```
-
-Input priority: **stdin → arguments → error**. The worker runs with
-`--max-turns 20 --permission-mode acceptEdits --tools Read,Glob,Grep,Edit,Write,Bash`.
-It never uses `--dangerously-skip-permissions`.
-
-## glm-review
-
-Read-only worker for repository exploration, call-graph discovery, duplicate detection,
-dependency inspection, and preliminary review:
-
-```powershell
-glm-review "Inspect this repository"
-```
-
-Runs with `--tools Read,Glob,Grep` — it cannot edit files or run commands.
-
-## CLI reference
-
-```text
-glm-router init              guided setup
-glm-router doctor [--network]  full runtime diagnosis
-glm-router status            quick offline overview
-glm-router key set           store ZAI_API_KEY (Windows User Environment)
-glm-router key check         key configured? from which source?
-glm-router config show
-glm-router config set models.main glm-5.3
-glm-router project init      CLAUDE.md / AGENTS.md managed blocks (--dry-run supported)
-glm-router project remove
-glm-router skill install     optional Codex delegation skill
-glm-router skill remove
-glm-router uninstall         guided removal (keeps ZAI_API_KEY by default)
-```
-
-Global flags: `--json --quiet --verbose --dry-run --force --yes`
-
-## Claude integration
-
-`glm-router project init` adds a **managed block** to `CLAUDE.md` at the project root
-(`git rev-parse --show-toplevel`, falling back to cwd):
-
-```text
-<!-- glm-coding-router:start -->
-... delegation policy ...
-<!-- glm-coding-router:end -->
-```
-
-- Everything outside the markers is preserved; existing blocks are replaced in place;
-  runs are idempotent and never duplicate.
-- Files are updated atomically (tmp file → fsync → rename).
-- On a malformed marker pair the file is left untouched with an actionable error.
-- CRLF/LF and UTF-8 are preserved.
-- `glm-router project remove` deletes only the managed block. A file the router
-  created entirely is deleted only when it would otherwise be empty.
-
-## Codex integration
-
-The same command updates `AGENTS.md` (Codex's repository instruction file) with an
-equivalent managed block. Additionally, `glm-router skill install` installs the optional
-`glm-delegation` skill to `~/.codex/skills/glm-delegation/SKILL.md`. If Codex is not
-detected, the skill step warns and skips — AGENTS.md integration and the core tool are
-unaffected.
-
-## Orca behavior (stale environments)
-
-Terminals embedded in Orca snapshot the Windows environment at startup. A key added
-after Orca starts is invisible to those terminals. Every GLM command therefore resolves
-the key in this order:
-
-1. `process.env.ZAI_API_KEY`
-2. Windows User Environment (via PowerShell)
-3. fail with an actionable error
-
-The key is never cached to disk.
-
-## Security model
-
-- The key lives only in the Windows User Environment; it is never written to
-  `config.json`, the repo, logs, or stack traces. Debug output redacts
-  `ZAI_API_KEY`, `ANTHROPIC_AUTH_TOKEN`, and Authorization headers.
-- Z.ai routing environment variables (`ANTHROPIC_AUTH_TOKEN`,
-  `ANTHROPIC_BASE_URL`, model overrides) are injected **only** into the spawned
-  `claude.exe` child process. `ANTHROPIC_API_KEY` is blanked in the child so your
-  normal Claude auth is never in play. `ANTHROPIC_BASE_URL` is never persisted globally.
-- Claude Code and Codex global authentication are never modified.
-- Child processes are spawned with argument arrays (`shell: false`) — prompts with
-  quotes, pipes, ampersands, or newlines are passed verbatim, never through a shell.
-- No telemetry, no automatic git commits.
-
-## Troubleshooting
-
-| Symptom | Fix |
-| --- | --- |
-| `ERROR [ZAI_KEY_MISSING]` | `glm-router key set`, then open a **new** terminal |
-| `ERROR [CLAUDE_NOT_FOUND]` | Install Claude Code, or `glm-router config set claudePath C:\path\to\claude.exe` |
-| Key works in a new terminal but not inside Orca | Expected — workers re-read the Windows User Environment automatically; run `glm-router doctor` to confirm |
-| `glm-*` not on PATH after install | Reopen the terminal; check `npm config get prefix` is on PATH |
-| `ERROR [MANAGED_BLOCK_CORRUPT]` | Fix the marker pair in the named file manually, then re-run |
-
-Run `glm-router doctor` (add `--network` to probe the Z.ai endpoint) for a full diagnosis.
-
-## Uninstall
-
-```powershell
-glm-router uninstall
-```
-
-The wizard removes the config, the Codex skill, and optionally the current project
-integration. `ZAI_API_KEY` is **kept** by default — removing credentials requires
-explicit consent. Finish with `npm uninstall -g glm-coding-router`.
-
-## Development
-
-```powershell
-npm install
-npm run build      # tsc → dist/
-npm test           # vitest run
-npm run lint       # eslint src tests
-npm run dev        # tsx src/cli.ts <args>
-```
-
-Integration tests spawn `tests/fixtures/fake-agent.mjs` (via `node.exe`) to verify
-argument passing, environment injection, and exit-code propagation without spending
-API quota. See the `GLM Coding Router — Technical Specification v0.1.md` for the full
-v0.1 contract (exit codes, managed-block test matrix, acceptance criteria).
-
-## Publishing
-
-```powershell
-npm run build
-npm test
-npm publish
-```
-
-`prepublishOnly` runs build + tests. The package ships only `dist/`; the four binaries
-(`glm-router`, `glm-chat`, `glm-worker`, `glm-review`) are declared in `bin`.
-
-## License
-
-MIT
+# GLM Coding Router
+
+GLM Coding Plan workers for Claude Code and Codex, on Windows.
+
+Claude Code and Codex stay your orchestrators — they keep responsibility for requirements,
+architecture, review, and integration. `glm-coding-router` delegates well-scoped
+implementation work (exploration, CRUD, boilerplate, tests, mechanical refactoring) to
+GLM workers via Z.ai's Anthropic-compatible endpoint.
+
+One global npm install replaces the manual `.cmd` shim setup:
+
+```text
+Claude / Codex → shell → glm-worker → claude.exe harness → Z.ai endpoint → GLM Coding Plan
+```
+
+## Architecture
+
+```text
+                         Developer
+                             │
+             ┌───────────────┴───────────────┐
+             ▼                               ▼
+       Claude Code                         Codex
+             │                               │
+             └───────────────┬───────────────┘
+                        shell command
+                             │
+             ┌───────────────┼───────────────┐
+             ▼               ▼               ▼
+    glm-chat  glm-fast  glm-worker  glm-review
+        │         │          │          │
+        └─────────┴──────────┴──────────┘
+                        claude.exe
+              (injected environment only)
+                             │
+                             ▼
+              https://api.z.ai/api/anthropic
+                             │
+                             ▼
+                    GLM Coding Plan
+                    GLM-5.3 / GLM-5.3-Flash
+```
+
+## Requirements
+
+- Windows 10/11 (v0.1; Linux/macOS planned for v0.2)
+- Node.js >= 20
+- Claude Code (`claude.exe`) — the GLM commands run on the Claude Code harness
+- Codex (optional — Claude-only setups are fully supported)
+- A Z.ai Coding Plan API key
+
+No Anthropic pay-as-you-go, no OpenAI API, no LiteLLM, no proxy.
+
+## Installation
+
+```powershell
+npm install -g glm-coding-router
+glm-router init
+```
+
+`npx glm-coding-router init` also works for a one-off check, but the global install is
+what puts `glm-worker` on your PATH long-term.
+
+## Quick start
+
+After `glm-router init`:
+
+```powershell
+glm-chat
+glm-fast
+glm-worker "Implement validation and add tests"
+glm-review "Analyze the auth module"
+```
+
+## glm-chat
+
+Interactive GLM-backed Claude Code session. Resolves the Z.ai key, locates `claude.exe`,
+injects the Z.ai environment **into the child process only**, and spawns it with
+pass-through arguments:
+
+```powershell
+glm-chat
+glm-chat --version
+glm-chat --any-claude-flag
+```
+
+Your normal `claude` command and its authentication are untouched.
+
+## glm-fast
+
+Interactive GLM-backed session pinned to the **fast model** (`models.fast`,
+`glm-5.3-flash` by default) — every model slot in the child environment maps to
+it, so whichever tier Claude Code picks, it gets the fast model. Same pass-through
+arguments as `glm-chat`:
+
+```powershell
+glm-fast
+glm-fast --profile air
+```
+
+## glm-worker
+
+Headless implementation worker:
+
+```powershell
+glm-worker "Implement validation and add tests"
+```
+
+Or via stdin (a structured task packet):
+
+```powershell
+@"
+TASK:
+Implement refresh token validation.
+
+SCOPE:
+internal/auth/
+
+VALIDATION:
+go test ./internal/auth/...
+"@ | glm-worker
+```
+
+Input priority: **stdin → arguments → error**. The worker runs with
+`--max-turns 20 --permission-mode acceptEdits --tools Read,Glob,Grep,Edit,Write,Bash`.
+It never uses `--dangerously-skip-permissions`.
+
+## glm-review
+
+Read-only worker for repository exploration, call-graph discovery, duplicate detection,
+dependency inspection, and preliminary review:
+
+```powershell
+glm-review "Inspect this repository"
+```
+
+Runs with `--tools Read,Glob,Grep` — it cannot edit files or run commands.
+
+## Profiles
+
+All four task binaries (`glm-chat`, `glm-fast`, `glm-worker`, `glm-review`)
+accept `--profile <name>` to overlay saved model/maxTurns settings. Profiles
+live in `config.json`:
+
+```json
+{
+  "profiles": {
+    "test":     { "workerMaxTurns": 10, "fast": "glm-5.3-flash" },
+    "frontend": { "main": "glm-5.3", "reviewMaxTurns": 30 }
+  }
+}
+```
+
+```powershell
+glm-worker --profile test "Add failing test then fix it"
+glm-review --profile frontend "Review the component tree"
+```
+
+Fields (all optional): `main`, `fast`, `workerMaxTurns`, `reviewMaxTurns`.
+Unknown profile names fail with `ERROR [11]` listing the available ones.
+Note: `--profile` belongs to these wrappers — it shadows Claude Code's own
+`--profile` flag inside them.
+
+## CLI reference
+
+```text
+glm-router init              guided setup
+glm-router doctor [--network]  full runtime diagnosis
+glm-router status            quick offline overview
+glm-router key set           store ZAI_API_KEY (Windows User Environment)
+glm-router key check         key configured? from which source?
+glm-router config show
+glm-router config set models.main glm-5.3
+glm-router project init      CLAUDE.md / AGENTS.md managed blocks (--dry-run supported)
+glm-router project remove
+glm-router skill install     optional Codex delegation skill
+glm-router skill remove
+glm-router uninstall         guided removal (keeps ZAI_API_KEY by default)
+```
+
+Global flags: `--json --quiet --verbose --dry-run --force --yes`
+
+## Claude integration
+
+`glm-router project init` adds a **managed block** to `CLAUDE.md` at the project root
+(`git rev-parse --show-toplevel`, falling back to cwd):
+
+```text
+<!-- glm-coding-router:start -->
+... delegation policy ...
+<!-- glm-coding-router:end -->
+```
+
+- Everything outside the markers is preserved; existing blocks are replaced in place;
+  runs are idempotent and never duplicate.
+- Files are updated atomically (tmp file → fsync → rename).
+- On a malformed marker pair the file is left untouched with an actionable error.
+- CRLF/LF and UTF-8 are preserved.
+- `glm-router project remove` deletes only the managed block. A file the router
+  created entirely is deleted only when it would otherwise be empty.
+
+## Codex integration
+
+The same command updates `AGENTS.md` (Codex's repository instruction file) with an
+equivalent managed block. Additionally, `glm-router skill install` installs the optional
+`glm-delegation` skill to `~/.codex/skills/glm-delegation/SKILL.md`. If Codex is not
+detected, the skill step warns and skips — AGENTS.md integration and the core tool are
+unaffected.
+
+## Orca behavior (stale environments)
+
+Terminals embedded in Orca snapshot the Windows environment at startup. A key added
+after Orca starts is invisible to those terminals. Every GLM command therefore resolves
+the key in this order:
+
+1. `process.env.ZAI_API_KEY`
+2. Windows User Environment (via PowerShell)
+3. fail with an actionable error
+
+The key is never cached to disk.
+
+## Security model
+
+- The key lives only in the Windows User Environment; it is never written to
+  `config.json`, the repo, logs, or stack traces. Debug output redacts
+  `ZAI_API_KEY`, `ANTHROPIC_AUTH_TOKEN`, and Authorization headers.
+- Z.ai routing environment variables (`ANTHROPIC_AUTH_TOKEN`,
+  `ANTHROPIC_BASE_URL`, model overrides) are injected **only** into the spawned
+  `claude.exe` child process. `ANTHROPIC_API_KEY` is blanked in the child so your
+  normal Claude auth is never in play. `ANTHROPIC_BASE_URL` is never persisted globally.
+- Claude Code and Codex global authentication are never modified.
+- Child processes are spawned with argument arrays (`shell: false`) — prompts with
+  quotes, pipes, ampersands, or newlines are passed verbatim, never through a shell.
+- No telemetry, no automatic git commits.
+
+## Troubleshooting
+
+| Symptom | Fix |
+| --- | --- |
+| `ERROR [ZAI_KEY_MISSING]` | `glm-router key set`, then open a **new** terminal |
+| `ERROR [CLAUDE_NOT_FOUND]` | Install Claude Code, or `glm-router config set claudePath C:\path\to\claude.exe` |
+| Key works in a new terminal but not inside Orca | Expected — workers re-read the Windows User Environment automatically; run `glm-router doctor` to confirm |
+| `glm-*` not on PATH after install | Reopen the terminal; check `npm config get prefix` is on PATH |
+| `ERROR [MANAGED_BLOCK_CORRUPT]` | Fix the marker pair in the named file manually, then re-run |
+
+Run `glm-router doctor` (add `--network` to probe the Z.ai endpoint) for a full diagnosis.
+
+## Uninstall
+
+```powershell
+glm-router uninstall
+```
+
+The wizard removes the config, the Codex skill, and optionally the current project
+integration. `ZAI_API_KEY` is **kept** by default — removing credentials requires
+explicit consent. Finish with `npm uninstall -g glm-coding-router`.
+
+## Development
+
+```powershell
+npm install
+npm run build      # tsc → dist/
+npm test           # vitest run
+npm run lint       # eslint src tests
+npm run dev        # tsx src/cli.ts <args>
+```
+
+Integration tests spawn `tests/fixtures/fake-agent.mjs` (via `node.exe`) to verify
+argument passing, environment injection, and exit-code propagation without spending
+API quota. See the `GLM Coding Router — Technical Specification v0.1.md` for the full
+v0.1 contract (exit codes, managed-block test matrix, acceptance criteria).
+
+## Publishing
+
+```powershell
+npm run build
+npm test
+npm publish
+```
+
+`prepublishOnly` runs build + tests. The package ships only `dist/`; the four binaries
+(`glm-router`, `glm-chat`, `glm-fast`, `glm-worker`, `glm-review`) are declared in `bin`.
+
+## License
+
+MIT

package/dist/bin/glm-chat.js

@@ -5,6 +5,7 @@ import { createGlmEnv } from "../core/env.js";
 import { Errors, formatGlmError, GlmRouterError } from "../core/errors.js";
 import { isMainModule } from "../core/main-guard.js";
 import { logger, redact } from "../core/logging.js";
+import { applyProfile, extractProfileFlag } from "../core/profile.js";
 import { spawnAgent } from "../core/process.js";
 import { resolveZaiApiKey } from "../core/zai-key.js";
 /**
@@ -12,7 +13,8 @@ import { resolveZaiApiKey } from "../core/zai-key.js";
  * spawn claude interactively with pass-through arguments.
  */
 export async function runChat(argv) {
-    const config = loadConfig();
+    const { rest, profile } = extractProfileFlag(argv);
+    const config = applyProfile(loadConfig(), profile);
     const resolved = resolveZaiApiKey();
     if (!resolved) {
         throw Errors.zaiKeyMissing();
@@ -22,7 +24,7 @@ export async function runChat(argv) {
     logger.debug(`spawning ${claudePath}`);
     logger.debug(redact(`env ANTHROPIC_BASE_URL=${env.ANTHROPIC_BASE_URL}`, [resolved.key]));
     return spawnAgent(claudePath, {
-        args: [...argv],
+        args: [...rest],
         cwd: process.cwd(),
         env,
         interactive: true,

package/dist/bin/glm-fast.js

@@ -0,0 +1,50 @@
+#!/usr/bin/env node
+import { loadConfig } from "../core/config.js";
+import { locateClaude } from "../core/claude.js";
+import { createGlmEnv } from "../core/env.js";
+import { Errors, formatGlmError, GlmRouterError } from "../core/errors.js";
+import { isMainModule } from "../core/main-guard.js";
+import { logger, redact } from "../core/logging.js";
+import { applyProfile, extractProfileFlag } from "../core/profile.js";
+import { spawnAgent } from "../core/process.js";
+import { resolveZaiApiKey } from "../core/zai-key.js";
+/**
+ * Effective config for glm-fast: every model slot pinned to the fast model
+ * (specs/glm-fast-profiles.md). Applied AFTER the profile so a profile's
+ * `fast` model flows into all slots; a profile's `main` is overridden by design.
+ */
+export function fastModelConfig(config) {
+    return { ...config, models: { main: config.models.fast, fast: config.models.fast } };
+}
+/**
+ * glm-fast (spec §54, specs/glm-fast-profiles.md): interactive chat pinned to
+ * the fast model. Pass-through args like glm-chat; supports --profile.
+ */
+export async function runFast(argv) {
+    const { rest, profile } = extractProfileFlag(argv);
+    const config = fastModelConfig(applyProfile(loadConfig(), profile));
+    const resolved = resolveZaiApiKey();
+    if (!resolved) {
+        throw Errors.zaiKeyMissing();
+    }
+    const claudePath = locateClaude(config);
+    const env = createGlmEnv(config, resolved.key);
+    logger.debug(`spawning ${claudePath}`);
+    logger.debug(redact(`env ANTHROPIC_BASE_URL=${env.ANTHROPIC_BASE_URL}`, [resolved.key]));
+    return spawnAgent(claudePath, {
+        args: [...rest],
+        cwd: process.cwd(),
+        env,
+        interactive: true,
+    });
+}
+if (isMainModule(import.meta.url)) {
+    runFast(process.argv.slice(2)).then((code) => process.exit(code), (error) => {
+        if (error instanceof GlmRouterError) {
+            process.stderr.write(formatGlmError(error) + "\n");
+            process.exit(error.exitCode);
+        }
+        process.stderr.write(String(error) + "\n");
+        process.exit(1);
+    });
+}

package/dist/bin/glm-review.js

@@ -5,6 +5,7 @@ import { createGlmEnv } from "../core/env.js";
 import { Errors, formatGlmError, GlmRouterError } from "../core/errors.js";
 import { isMainModule } from "../core/main-guard.js";
 import { logger, redact } from "../core/logging.js";
+import { applyProfile, extractProfileFlag } from "../core/profile.js";
 import { readStdin, resolvePrompt } from "../core/prompt.js";
 import { spawnAgent } from "../core/process.js";
 import { resolveZaiApiKey } from "../core/zai-key.js";
@@ -18,8 +19,9 @@ export function buildReviewArgs(prompt, config) {
  * discovery, duplicate detection, dependency inspection, and review.
  */
 export async function runReview(argv) {
-    const prompt = await resolvePrompt(argv, readStdin, "glm-review");
-    const config = loadConfig();
+    const { rest, profile } = extractProfileFlag(argv);
+    const prompt = await resolvePrompt(rest, readStdin, "glm-review");
+    const config = applyProfile(loadConfig(), profile);
     const resolved = resolveZaiApiKey();
     if (!resolved) {
         throw Errors.zaiKeyMissing();

package/dist/bin/glm-worker.js

@@ -5,6 +5,7 @@ import { createGlmEnv } from "../core/env.js";
 import { Errors, formatGlmError, GlmRouterError } from "../core/errors.js";
 import { isMainModule } from "../core/main-guard.js";
 import { logger, redact } from "../core/logging.js";
+import { applyProfile, extractProfileFlag } from "../core/profile.js";
 import { resolvePrompt } from "../core/prompt.js";
 import { spawnAgent } from "../core/process.js";
 import { resolveZaiApiKey } from "../core/zai-key.js";
@@ -28,8 +29,9 @@ export function buildWorkerArgs(prompt, config) {
  * --dangerously-skip-permissions.
  */
 export async function runWorker(argv) {
-    const prompt = await resolvePrompt(argv);
-    const config = loadConfig();
+    const { rest, profile } = extractProfileFlag(argv);
+    const prompt = await resolvePrompt(rest);
+    const config = applyProfile(loadConfig(), profile);
     const resolved = resolveZaiApiKey();
     if (!resolved) {
         throw Errors.zaiKeyMissing();

package/dist/core/config.js

@@ -6,6 +6,13 @@ import { configDir, configPath } from "./paths.js";
 export const DEFAULT_ANTHROPIC_BASE_URL = "https://api.z.ai/api/anthropic";
 export const DEFAULT_MAIN_MODEL = "glm-5.3";
 export const DEFAULT_FAST_MODEL = "glm-5.3-flash";
+/** Named model/maxTurns overlay selected via --profile (specs/glm-fast-profiles.md). */
+export const ProfileSchema = z.object({
+    main: z.string().min(1).optional(),
+    fast: z.string().min(1).optional(),
+    workerMaxTurns: z.number().int().positive().optional(),
+    reviewMaxTurns: z.number().int().positive().optional(),
+});
 export const ConfigSchema = z.object({
     schemaVersion: z.literal(1),
     provider: z.object({
@@ -30,6 +37,8 @@ export const ConfigSchema = z.object({
     // Optional executable overrides used by discovery (spec §33, §34).
     claudePath: z.string().min(1).optional(),
     codexPath: z.string().min(1).optional(),
+    // Named overlays selected via --profile (specs/glm-fast-profiles.md).
+    profiles: z.record(z.string(), ProfileSchema).default({}),
 });
 export function defaultConfig() {
     return {
@@ -49,6 +58,7 @@ export function defaultConfig() {
             codex: true,
             codexSkill: true,
         },
+        profiles: {},
     };
 }
 /**

package/dist/core/profile.js

@@ -0,0 +1,57 @@
+import { Errors } from "./errors.js";
+/**
+ * Remove the router's --profile flag from argv (specs/glm-fast-profiles.md).
+ * Accepts `--profile name` and `--profile=name`; the first occurrence wins,
+ * later ones are consumed too so they never leak into the forwarded args.
+ */
+export function extractProfileFlag(argv) {
+    const rest = [];
+    let profile;
+    for (let i = 0; i < argv.length; i++) {
+        const arg = argv[i];
+        if (arg === "--profile") {
+            const value = argv[i + 1];
+            if (value === undefined) {
+                throw Errors.configInvalid("--profile requires a profile name");
+            }
+            profile ??= value;
+            i++;
+            continue;
+        }
+        if (arg.startsWith("--profile=")) {
+            const value = arg.slice("--profile=".length);
+            if (value.length === 0) {
+                throw Errors.configInvalid("--profile requires a profile name");
+            }
+            profile ??= value;
+            continue;
+        }
+        rest.push(arg);
+    }
+    return { rest, profile };
+}
+/**
+ * Overlay a named profile onto the config (specs/glm-fast-profiles.md):
+ * main/fast models plus worker/review maxTurns. Unknown names fail with
+ * ERROR [11] listing the defined profiles.
+ */
+export function applyProfile(config, name) {
+    if (!name) {
+        return config;
+    }
+    const profile = config.profiles[name];
+    if (!profile) {
+        const known = Object.keys(config.profiles);
+        const listed = known.length > 0 ? known.sort().join(", ") : "(none defined)";
+        throw Errors.configInvalid(`unknown profile "${name}" — available profiles: ${listed}`);
+    }
+    return {
+        ...config,
+        models: {
+            main: profile.main ?? config.models.main,
+            fast: profile.fast ?? config.models.fast,
+        },
+        worker: { maxTurns: profile.workerMaxTurns ?? config.worker.maxTurns },
+        review: { maxTurns: profile.reviewMaxTurns ?? config.review.maxTurns },
+    };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "glm-coding-router",
-  "version": "0.1.1",
+  "version": "0.2.0",
   "description": "GLM Coding Plan workers for Claude Code and Codex",
   "type": "module",
   "license": "MIT",
@@ -13,7 +13,8 @@
     "glm-router": "./dist/cli.js",
     "glm-chat": "./dist/bin/glm-chat.js",
     "glm-worker": "./dist/bin/glm-worker.js",
-    "glm-review": "./dist/bin/glm-review.js"
+    "glm-review": "./dist/bin/glm-review.js",
+    "glm-fast": "./dist/bin/glm-fast.js"
   },
   "files": [
     "dist"