glm-plugin-cc

Use GLM-5.3 (z.ai) from inside Claude Code for code reviews or to delegate tasks.

This plugin is for Claude Code users who want an easy way to start using GLM-5.3 from the workflow they already have — without leaving Claude Code, and without pointing their primary Claude session away from Anthropic.

glm-plugin-cc is directly inspired by openai/codex-plugin-cc, which lets Claude Code delegate to OpenAI Codex. This plugin does the same thing, but delegates to GLM-5.3 via z.ai's Anthropic-compatible endpoint. The command surface (review, adversarial-review, rescue, transfer, status, result, cancel, setup), the review-gate hook, and the plugin layout all mirror the original. See Credit & inspiration.

What you get

Command	What it does
/glm:review	Read-only, non-destructive code-quality review of local git state.
/glm:adversarial-review	Steerable, skeptical design review that challenges the approach, tradeoffs, auth, and reliability.
/glm:rescue	Delegate active debugging, multi-file refactors, or implementation loops to GLM with full edit/test permissions.
/glm:transfer	Hand the current Claude Code session off to a resumable GLM thread.
/glm:status	Check the status of a GLM job (or list recent jobs).
/glm:result	Fetch the final result of a GLM job, including the resumable session id.
/glm:cancel	Abort a running GLM job.
/glm:setup	Check readiness and toggle the review gate.

You also get the glm:glm-rescue subagent in /agents, the glm-5-2-prompting skill, and an optional Stop-hook review gate.

Requirements

• A z.ai GLM Coding Plan (recommended) or a z.ai API key. Usage counts against your plan or key balance.
• Node.js 18.18 or later.
• The claude CLI installed and logged in normally. This plugin reuses that same binary, run headless, redirected at z.ai. (If you can run claude today, you're already set.)

Auth setup

GLM needs a z.ai API key. Provide it via one of:

• an environment variable:

  export ZAI_API_KEY=...

  (ZA_API_KEY is also accepted as a fallback), or

• a key file:

  mkdir -p ~/.config/zai && echo -n 'YOUR_KEY' > ~/.config/zai/api-key
  chmod 600 ~/.config/zai/api-key

Then run /glm:setup to confirm the key and claude binary are detected.

Install

/plugin marketplace add radityasurya/glm-plugin-cc
/plugin install glm@radityasurya-glm
/reload-plugins
/glm:setup

/glm:setup prints whether the key and claude binary were found, and offers a small readiness probe. If the plugin isn't picking up your key, re-export the env var in the shell that launched Claude Code, or write the key file.

A simple first run:

/glm:review --background
/glm:status
/glm:result

How it works

glm-plugin-cc does not call the z.ai HTTP API itself. Instead, the broker (glm-broker.mjs) spawns a headless claude -p subprocess whose environment is redirected to z.ai's Anthropic-compatible endpoint:

Variable	Value
ANTHROPIC_BASE_URL	https://api.z.ai/api/anthropic
ANTHROPIC_AUTH_TOKEN	your z.ai key
ANTHROPIC_MODEL	glm-5.3
ANTHROPIC_SMALL_FAST_MODEL	glm-5.3 (prevents a Claude-side haiku fallback)
ANTHROPIC_API_KEY	cleared (so the auth token wins)

Because z.ai speaks the Anthropic Messages wire protocol, the headless claude loop — including tool use, file edits, and bash — works against GLM unchanged. Your primary Claude session keeps running against Anthropic; only the delegated subprocess talks to z.ai.

• Reviews run in plan mode (read-only: Read, Grep, Glob, non-mutating Bash).
• /glm:rescue runs with full permissions so GLM can edit, run tests, and run lint.

For the full design, see docs/ARCHITECTURE.md.

Usage

/glm:review

A normal, read-only GLM review of your current work — the same kind of review you'd get by asking GLM directly.

Use it when you want:

• a review of your current uncommitted changes
• a review of your branch compared to a base branch like main

Use --base <ref> for branch review. It also supports --wait and --background. It is not steerable and does not take custom focus text — use /glm:adversarial-review when you want to challenge a specific decision.

Examples:

/glm:review
/glm:review --base main
/glm:review --background

This command is read-only and will not change anything. When run in the background, use /glm:status to check progress and /glm:cancel to abort.

/glm:adversarial-review

A steerable review that questions the chosen implementation and design.

Use it when you want:

• a pre-ship review that challenges the direction, not just the code details
• focus on design choices, tradeoffs, hidden assumptions, and alternative approaches
• pressure-testing of specific risk areas like auth, data loss, rollback, race conditions, or reliability

Same target selection as /glm:review (including --base <ref>), plus --wait / --background. Unlike /glm:review, it takes extra focus text after the flags.

Examples:

/glm:adversarial-review
/glm:adversarial-review --base main challenge whether this was the right caching and retry design
/glm:adversarial-review --background look for race conditions and question the chosen approach

Read-only. It does not fix code.

/glm:rescue

Hands a task to GLM through the glm:glm-rescue subagent / broker.

Use it when you want GLM to:

• investigate a bug
• try a fix
• continue a previous GLM rescue task
• take a cheaper/faster pass

Supports --background, --wait, --resume, and --fresh. If you omit both --resume and --fresh, the broker continues the most recent finished rescue session for this repo, if it ran on the same model and finished in the last 6 hours. Otherwise it starts a fresh session.

Examples:

/glm:rescue investigate why the tests started failing
/glm:rescue fix the failing test with the smallest safe patch
/glm:rescue --resume apply the top fix from the last run
/glm:rescue --background investigate the regression

Notes:

• /glm:rescue runs GLM with full file/bash permissions. It will edit your repo. Run it on a clean working tree or a branch you don't mind touching.
• Follow-up rescue requests continue the latest GLM rescue thread in the repo by default. A session that z.ai no longer has is replaced by a fresh one instead of failing the job.

/glm:transfer

Creates a GLM session seeded from the current Claude Code transcript and prints a /glm:rescue --resume <session-id> command you can use to continue.

Use it when you started a conversation in Claude Code and want to keep going on GLM.

/glm:transfer
/glm:transfer --source ~/.claude/projects/<proj>/<session>.jsonl

The SessionStart hook captures the current transcript path automatically; --source is a manual override. The source must be a Claude JSONL transcript.

/glm:status

Shows running and recent GLM jobs for the current repo.

/glm:status
/glm:status glm-<id>

/glm:result

Shows the final stored GLM output for a finished job. Includes the GLM session id so you can reopen that run with /glm:rescue --resume <session-id>.

/glm:result
/glm:result glm-<id>

/glm:cancel

Cancels an active background GLM job.

/glm:cancel
/glm:cancel glm-<id>

/glm:setup

Checks whether GLM is reachable and manages the optional review gate.

/glm:setup
/glm:setup --check          # non-interactive readiness probe
/glm:setup --enable-review-gate
/glm:setup --disable-review-gate

Typical flows

Review before shipping

/glm:review

Hand a problem to GLM

/glm:rescue investigate why the build is failing in CI

Start something long-running

/glm:adversarial-review --background
/glm:rescue --background investigate the flaky test

Then check in with:

/glm:status
/glm:result

Review gate

/glm:setup --enable-review-gate
/glm:setup --disable-review-gate

When enabled, a Stop hook runs a targeted GLM review based on Claude's response. If that review finds blocking issues, the stop is denied so Claude can address them first.

Warning: feedback loops. The review gate runs GLM at every stop. If GLM and Claude disagree, the session can iterate for many turns and rack up z.ai cost quickly. It's off by default. Enable it only when you plan to actively watch the session. This mirrors the same warning on codex-plugin-cc's review gate.

GLM integration

The plugin talks to GLM via the claude binary you already have, redirected to z.ai. Configuration is resolved in this order:

1. ~/.glm-plugin/settings.json (plugin-managed)
2. z.ai key from ZAI_API_KEY / ZA_API_KEY env var or ~/.config/zai/api-key
3. --model <id> flag per command (default glm-5.3)

Common configurations

To change the default model, set it via the --model flag, or edit ~/.glm-plugin/settings.json. To change where the plugin looks for the key, set ZAI_API_KEY in the shell that launches Claude Code.

Moving the work over to GLM

Delegated tasks and any review-gate run can be resumed directly on GLM by passing the session id you got from /glm:result or /glm:status to /glm:rescue --resume <session-id>.

Config reference

Setting	Default	Where
Model	glm-5.3	per-call --model, or ~/.glm-plugin/settings.json
z.ai endpoint	https://api.z.ai/api/anthropic	hardcoded (Anthropic Messages protocol)
Job state	~/.glm-plugin/jobs/	JSON per job + stream JSONL
Settings	~/.glm-plugin/settings.json	review gate toggle, etc.
Transcript capture	~/.glm-plugin/state/current-transcript.json	written by SessionStart hook

Troubleshooting

• No z.ai API key found — set ZAI_API_KEY in the shell that launches Claude Code, or write ~/.config/zai/api-key. Re-run /glm:setup.
• Reviews come back empty — make sure you have uncommitted changes (git status) or pass --base <ref> for a branch diff. Untracked files are included.
• claude exited 1 — usually means the headless subprocess couldn't reach z.ai. Run /glm:setup --check and confirm your key is valid at z.ai/manage-apikey/apikey-list.
• Background job shows (not responding) — the detached worker crashed. Run the same task with --wait to see the error inline.
• Review gate won't stop arguing — disable it with /glm:setup --disable-review-gate.

FAQ

Do I need a separate account?

No — if you already have a z.ai API key or GLM Coding Plan, that's all you need. This plugin uses your key directly.

Does the plugin use a separate runtime?

No. It delegates through your local claude CLI, run headless against z.ai. Same install, same machine, same repo checkout.

Will it use my existing Claude login?

Your primary Claude session is untouched. The broker constructs a separate environment for the subprocess that points only at z.ai; it never touches your Anthropic credentials.

Is this an official z.ai / Anthropic / OpenAI product?

No. glm-plugin-cc is an independent community plugin. GLM and z.ai are products of Z.ai. Claude Code is a product of Anthropic. Codex and codex-plugin-cc are products of OpenAI. This project is not endorsed by any of them.

How is this different from just setting ANTHROPIC_BASE_URL to z.ai?

Setting ANTHROPIC_BASE_URL globally makes GLM your primary model — Claude itself runs on GLM. This plugin keeps Claude on Anthropic and uses GLM only for delegated work (reviews, rescues). You get two models in one session.

Credit & inspiration

This project is closely modeled on openai/codex-plugin-cc by OpenAI (Apache-2.0). The following were adapted directly:

• The command surface and file layout (commands/, agents/, hooks/, prompts/, schemas/, skills/).
• The two-tier review concept (review vs adversarial-review).
• The rescue / transfer / status / result / cancel delegation model.
• The Stop-hook review gate.

The core difference is the broker. codex-plugin-cc wraps the Codex CLI / Codex app-server (an OpenAI-built agent runtime). GLM ships no equivalent CLI, so glm-plugin-cc's broker instead spawns the existing claude CLI, run headless, with its environment redirected to z.ai's Anthropic-compatible endpoint. z.ai speaking the Anthropic Messages protocol is what makes this work without a custom agent loop.

See NOTICE for attribution details.

Development

pnpm install
pnpm test          # vitest, no network
pnpm run setup     # glm-broker.mjs setup --check

Tests stub spawn/spawnSync and never hit the network. See docs/DEVELOPMENT.md.

License

Apache-2.0. See LICENSE.