Usage/rate-limit tracking for ZCode (Z.ai GLM Coding Plan)?

[sunganhao8-lgtm]

Follow-up from #21756 (ZCode as a supported agent).

Orca's usage providers cover Claude, Codex, Gemini, OpenCode Go, Kimi, Antigravity, MiniMax, and Grok. ZCode (Z.ai's GLM Coding Plan CLI) is missing from that list — for subscribers it's the same kind of subscription worth surfacing in the status bar.

What I found while looking into this

• The desktop bundle carries internal endpoints for plan billing (zcode.z.ai/api/v1/zcode-plan/billing/current and …/billing/balance), but no documented public usage API for third-party callers that I could find (docs.z.ai/devpack only points at the web billing console).
• The CLI stores coding-plan credentials locally (e.g. ~/.zcode/v2/credentials.json, structure varies by plan: zai-coding-plan vs bigmodel-coding-plan).

So the Kimi pattern (read the CLI's own credentials file → call the CLI's own managed-usage endpoint) doesn't transfer 1:1 yet, and I didn't want to ship a fetcher against an undocumented internal endpoint in a drive-by PR.

Possible paths

1. ZCode CLI exposes a stable usage/quota command or endpoint that Orca could call with the CLI's own credentials (cleanest — same shape as Kimi).
2. Z.ai documents a public usage endpoint for the coding plan.
3. As a last resort, the MiniMax-style cookie/console integration.

Happy to contribute the provider once there's agreement on which path is acceptable.

[AmethystLiang]

thanks for the feedback! we're looking into it.

[nwparker]

Now that Z.ai has open-sourced zai-org/ZCode, there's a fourth path that avoids the concern in the original post.

ZCode keeps usage locally. Its SQLite store at ~/.zcode/cli/db/db.sqlite has a model_usage table written per model request:

started_at, provider_id, model_id, variant, agent, mode,
input_tokens, output_tokens, reasoning_tokens,
cache_creation_input_tokens, cache_read_input_tokens,
computed_total_tokens, provider_total_tokens, status, duration_ms

with create index model_usage_started_model_idx on model_usage(started_at, provider_id, model_id). There are also turn_usage and tool_usage tables. I confirmed 93 real rows on a machine with ZCode installed.

So Orca can build a local usage provider — the same shape as the Muse Code local provider (#22379) — reading the CLI's own on-disk records. No undocumented zcode-plan/billing/* call, no cookie/console scraping, no credential read, and it keeps working offline.

The one thing it can't give you is the plan's remaining quota, only consumption. If surfacing remaining quota matters, that still needs one of the three paths in the original post, and #18506's observation (key in Authorization with no Bearer prefix) is the relevant detail.

Happy to take this as a follow-up to #22464, which adds the ZCode harness itself and deliberately leaves usage out.

[nwparker]

Please try the ZCode harness build and report back on #22464

#22464 (the ZCode agent harness) is green and has installable builds. Usage tracking — this issue — is not in it, but the harness is the foundation it lands on, so it's worth confirming the base works.

Test instructions: #22464 (comment)

Build: https://github.com/stablyai/orca-adhoc/releases/tag/v1.4.210-adhoc.20260923230858 — orca-macos-arm64.dmg, orca-macos-x64.dmg, or orca-windows-setup.exe (Windows is unsigned; SmartScreen → More info → Run anyway).

Please comment on #22464 with what you find — whether it works or breaks. Either is useful. Include your OS and zcode --version. Once a few people confirm, it gets approved and merged.

Most useful from this thread specifically: you have a GLM Coding Plan account, and I don't. My test machine has no model configured, so I proved the status pipeline with ZCode's real hook payloads rather than live turns. Nobody has yet watched a real ZCode session drive Orca end to end — that's the gap.

On usage itself: as noted above, the open-sourced CLI records usage locally in ~/.zcode/cli/db/db.sqlite (model_usage, per-request token columns), so a local provider is viable without the undocumented billing endpoint you were right to be wary of. That's the follow-up once this lands.

@sunganhao8-lgtm