First-party GLM terminal CLI + agent heuristic: ask user-only facts before deep investigation

[HotDogAmbulance]

Feature request / 功能建议

A first-party GLM terminal CLI, plus one trained agent heuristic: ask for user-only facts before deep investigation.

Motivation / 动机

1. GLM is tuned for agentic coding, but Z.ai ships no first-party CLI

Today GLM reaches coding agents through competitors' harnesses (Claude Code, opencode, Cline) and aggregators (OpenRouter). Consequences:

• Z.ai does not control the agent UX philosophy on the surface where coding agents actually live (the terminal) — e.g. when the model should ask the user vs. keep working (see the case study below for why this matters).
• Agent-behavior feedback and telemetry arrive second-hand, filtered through other companies' harnesses.
• A first-party CLI (even minimal) with the coding-plan endpoint as default auth would close the loop and reduce dependency on other harnesses' roadmaps.

2. Concrete case study: the missing "ask vs. investigate" routing

A real session (GLM agent in a terminal harness, diagnosing a personal macOS status HUD that live-probes Claude quota using Claude Code's keychain OAuth token):

• Symptom shown to the agent: the HUD warned "probe: Claude authentication expired" even though the meter looked healthy and the user had just re-checked usage.
• The agent then ran a long forensic chain — repeated keychain reads, cache-file mtime checks, process inspection, log tails — across multiple turns to reconstruct why the token had expired and why manual probes didn't clear the warning.
• The decisive fact was in the user's head the whole time: the user had only checked usage in the web console and had not run a Claude Code conversation since the token's last refresh — so the credential owner (Claude Code) had lazily not refreshed the keychain entry, every probe failed at the keychain step, and one sentence ("did you check the web, or Claude Code?") in the agent's first reply would have replaced roughly ten tool calls and several minutes of investigation.

Suggested heuristic to train into agentic GLM: before deep autonomous investigation, split unknowns into (a) machine-verifiable facts — verify directly, it's cheap and reliable; and (b) human-context facts (what the user did, saw, clicked) — ask one short question first. The question should fire only when it is strictly cheaper than the investigation it replaces — fewer wasted turns, not more interruptions.

Your contribution / 您的贡献

Happy to provide fuller session traces of the case study above, test early CLI builds, and report agent-behavior comparisons across harnesses.

[Dovis01]

Hi @HotDogAmbulance,

Thank you for taking the time to provide such a detailed case study and your deep insights into Agent Heuristics. This is incredibly helpful and inspiring for us!

We've shared your heuristic strategy and your suggestion for a first-party terminal CLI with our internal engineering and model development teams for review.

To address the pain points you highlighted regarding external harness experiences and control, we highly recommend trying our official harness product: Zcode. It features much stronger native support for agent interactions and underlying orchestration, which should provide you with a noticeably smoother development experience.

Thanks again for your continued support and contributions to GLM! If you have any further cross-harness behavioral comparisons or feedback in the future, please feel free to share them here.