bacoco GLM Companion
Resolve ../../scripts/glm_companion.py relative to this file and use it as the
only runtime entry point. Pass --host codex when running inside Codex and
--host claude when running inside Claude Code. Never invoke Z.ai or
claude -p directly.
Route $glm or /glm
| Intent | Runtime command |
|---|---|
| task, ask, implement, fix, investigate | task |
| continue, resume | task --resume-last |
| setup, key, enable/disable gate | setup |
| review | review |
| adversarial review, challenge | adversarial-review |
| transfer this conversation | transfer |
| jobs, progress, status | status |
| result, output | result |
| cancel, stop | cancel |
If the user writes only $glm, ask what GLM should do.
Delegate
- Fresh tasks automatically include visible user/host-agent messages when the current transcript can be identified. Hidden instructions, private reasoning, tool traces, attachments, and secrets are excluded.
- Add
--writeonly for explicit implementation or mutation requests. Diagnosis, planning, research, and reviews remain read-only. - Use foreground for bounded work and
--backgroundfor long or explicitly asynchronous work. Never combine--backgroundand--wait. - Use
--resume-lastfor an obvious follow-up. Use--no-contextonly when the user requests isolation. - Leave model, effort, budget, and timeout unset unless requested. The runtime
tries
glm-5.2first and transparently falls back toglm-4.7only when Z.ai rate-limits the preferred model. The result always reports a fallback.
python3 <runtime> task --host <codex|claude> --wait -- "Investigate and report evidence."
python3 <runtime> task --host <codex|claude> --background --write -- "Implement and verify the fix."
python3 <runtime> task --host <codex|claude> --resume-last --write -- "Continue the implementation."
Use stdin or a safely quoted argument for arbitrary user text.
Transfer and review
transfer seeds a resumable GLM session with the visible host conversation.
review and adversarial-review are read-only and return structured findings.
python3 <runtime> transfer --host <codex|claude> --wait
python3 <runtime> review --host <codex|claude> --scope auto --wait
python3 <runtime> adversarial-review --host <codex|claude> --base main --wait "Focus on data loss"
Present findings without applying them until the user authorizes fixes.
Control and setup
python3 <runtime> setup --host <codex|claude>
python3 <runtime> status --host <codex|claude>
python3 <runtime> status --host <codex|claude> <job-id> --wait
python3 <runtime> result --host <codex|claude> <job-id>
python3 <runtime> cancel --host <codex|claude> <job-id>
python3 <runtime> setup --host <codex|claude> --enable-review-gate
The runtime looks first in the user-owned ~/.glm.env secret file (mode 0600),
then for ZAI_API_KEY, GLM_API_KEY, and the same keys in the workspace .env.
Never print the key. Leave the stop gate disabled unless explicitly requested.
After write-capable work, inspect the real diff and run proportionate checks. GLM output is evidence, not automatic proof.