Windows: add Z.AI Coding Plan usage

[stanthewizzard]

Summary

• add Windows support for Z.AI Coding Plan usage through existing ZCode, OpenCode, or manual glm.json credentials
• show Z.AI in the notch, tray, settings, diagnostics, and provider refresh flow
• migrate an existing customized notch selection once so the new provider can appear
• widen the horizontal Windows notch to fit six provider rings

Scope

Windows implementation only (windows/codenotch). macOS is unchanged.

Security

Credentials remain local and are not persisted by Codenotch. Encrypted ZCode tokens are skipped, and requests are limited to trusted Z.AI/BigModel hosts.

Validation

cargo test --release --locked -p codenotch: 99 passed, 0 failed, 2 ignored.

[vinzdg]

Thanks @stanthewizzard — the endpoint, the unit/number window identity, the ms nextResetTime handling, the console selection and the enc:v1: skip all match the Mac provider correctly. Two things block it, both where it diverges from the Mac:

1. The Bearer prefix will 401 for everyone. glm.rs:233 sends Authorization: Bearer <key>. The Mac deliberately does not — from GLMProvider.swift: "The monitor takes the key raw — no 'Bearer' scheme. Prefixing it is exactly what an auth failure looks like from here." Send the key raw.

2. Upstream prose reaches the UI and the log. glm.rs:354-355 puts Z.ai's msg into the note and into the app log. We removed exactly this on the Mac a couple of releases ago — a provider's own error text can carry account details, and it ends up in log files people paste into issues. Show the code only, and a fixed string for anything else.

Two more worth fixing while you're in there:

• No glm branch in ring_window (main.rs:639-648) or in headlineOf/weeklyOf in notch.html, so glm falls through to the Antigravity lane picker and the ring shows the tightest window instead of the session one. The Mac sets headlineID: "session", weeklyID: "weekly".
• It drops the Mac's first credential source — Claude Code's ~/.claude/settings.json (ANTHROPIC_AUTH_TOKEN plus a Z.ai ANTHROPIC_BASE_URL) — and invents a glm.json the Mac has no notion of. Worth matching the Mac's source order so the same machine works on both.

Nit: status "none" at glm.rs:361 isn't a status the UI knows.

Note that #276, #287, #288, #274, #210 and #257 have all landed since you opened this, so config.rs, main.rs and both HTML pages have moved — you'll want a rebase. Happy to look again after.

[vinzdg]

Merged as 174a5c7, with my fixes in cd3f2a1. Thanks @stanthewizzard — the endpoint, the unit/number window identity, the ms nextResetTime, the console selection, the enc:v1: skip and the one-time GLM migration were all correct, and I took them as written.

What I changed rather than sending back:

• The Bearer prefix. GLMProvider.swift:110 is explicit that the monitor takes the key raw and that prefixing it is exactly what an auth failure looks like from here — every read would have answered 401. Fixed the module doc too, which described it as Bearer-authenticated.
• The upstream prose. Z.ai's msg went into both the card and the log; now it's the code only.
• Added Claude Code as a source. ~/.claude/settings.json env.ANTHROPIC_AUTH_TOKEN with a Z.ai ANTHROPIC_BASE_URL, claimed only when the base URL names a Z.ai console. It's the Mac's first source and the documented way to point Claude Code at the plan — without it, a machine set up the normal way showed nothing on Windows while working on the Mac. Your glm.json stays as the first source.
• Ring routing. glm had no arm in ring_window, headlineOf or weeklyOf, so it fell through to Antigravity's lane picker and showed the tightest window rather than the session. Now session/weekly, matching the Mac's headlineID/weeklyID.
• Dropped a "none" status the UI has no case for.

Merge note: today's six Windows merges moved the notch to 44 px rings, so your six-ring width calculation resolved to 6 × 44 rather than 6 × 56. The flat notch needs 504 px and has 650.

CI green on all four workflows.