[Bug] Model picker empty in WSL remote projects after server restart (account/plan state never re-synced)[Bug]

提交前确认 · Pre-submission checklist

• 我已搜索过现有 issue，确认这不是重复 / I searched existing issues and confirmed this isn't a duplicate.
• 我已阅读 CONTRIBUTING.md / I've read CONTRIBUTING.md.

问题类别 · Category

账号 / 登录 · Account / Login

涉及的 Agent 框架 · Agent framework

ZCode Agent（自研）

严重程度 · Severity

阻塞使用 · Blocking (无法使用核心功能 / core function unusable)

复现频率 · Reproducibility

必现 · Always

问题描述 · Description

Environment

• OS: Windows 11 (build 26200), ZCode desktop 3.12.2
• Remote: WSL2, distro Ubuntu-20.04, project opened as remote workspace (remote:wsl:Ubuntu-20.04:/home/marcelo/...)
• Network: USB 4G modem (relevant only to the timeline below; the picker bug reproduces with fully working WSL networking)
• No proxy/VPN active during the repro steps below (Tailscale installed but stopped).

Summary

When opening a WSL project, the model picker shows only "Manage" — no models are listed and none can be selected. Local (Windows) projects in the same app instance work perfectly with the same account (GLM-5.3, zai individual coding plan).

The key observation: the picker worked at one point while a long-lived WSL protocol server process (started before I logged in on the Windows side) held the account/plan state in memory. After every wsl --shutdown or app restart, freshly started WSL servers never receive the account/plan state again, so the picker stays empty even though:

• WSL has internet access (verified: https://api.z.ai returns 200 in ~1s from inside WSL)
• ~/.zcode/v2/credentials.json inside WSL exists and is fresh (written by the app itself during the working window)
• ~/.zcode/v2/coding-plan-cache.json inside WSL reports builtin:zai-coding-plan: available
• ~/.zcode/v2/setting.json inside WSL has the correct selection (modelProviderFamilySelectedKeys.zai = "coding-plan:builtin:zai-coding-plan", identical to the Windows side)

复现步骤 · Steps to reproduce

Steps to reproduce

1. Windows app 3.12.2, logged in with a zai coding plan. Open a WSL remote project and confirm the model picker works.
2. Run wsl --shutdown (or fully restart the ZCode app).
3. Reopen the same WSL project in the app.
4. Model picker shows only "Manage"; no models selectable.

Evidence

1. Model selection persistence fails with a database FK error (Windows-side log, ~/.zcode/cli/log/zcode-2026-09-16.jsonl):

{"level":"warn","event":"session.model_selection.persist_failed",
 "message":"Session model selection persistence failed",
 "context":{"error":"FOREIGN KEY constraint failed",
            "modelId":"GLM-5.3","providerId":"account:zai-individual-coding-plan",
            "thoughtLevel":"low"}}

The failing sessions (sess_3b8c40f2…, sess_ba775482…) do not exist as rows in the session table of either ~/.zcode/cli/db/db.sqlite (Windows) or the WSL-side DB — the session row was never persisted (session_create logged with "persistence":"deferred"), so dependent inserts fail the FK.

2. Provider registry divergence between Windows and WSL. The zcode_protocol.provider_registry.ready event carries a different accountRevision hash on each side:

• Windows: zcode-builtin:28:9cd1cc2f4a86…
• WSL: zcode-builtin:28:ac6ecccf22e18… (all account providers entitled:false; plan entitlement from the coding-plan cache is apparently never merged in)

3. No account/plan/subscription fetch attempts at all on the WSL side after server startup. Scanning the WSL-side log after a fresh server start shows zero events matching plan/subscription/entitlement/account — the server never even tries to refresh the plan state.

4. Server redeploy is byte-identical. Deleting ~/.zcode/server/ and letting the app re-deploy produces the exact same zcode-server.cjs (same size and timestamp), ruling out a stale-server-version theory.

What did NOT fix it

• Full app restart / killing all zcode processes
• Removing and letting the app re-deploy the WSL server bundle
• Removing and restoring credentials.json in WSL (the app does not re-push it on its own after a fresh server start; the only time it was written was during the original login flow while the then-running server was alive)
• Switching WSL to networkingMode=mirrored (fixed a separate NAT issue with the 4G modem, but the picker remains empty)
• Uninstalling an old standalone ZCode .deb (3.7.7) that was previously installed inside WSL

期望表现 · Expected behavior

Workaround / expected behavior

Expected: after any WSL server restart, the app re-syncs the account/plan state to the WSL server (or the server fetches it using the stored credentials), and the model picker populates.

Actual: the picker only ever worked on a server process that happened to be alive across the Windows-side login event.

Timeline (2026-09-16, local UTC-3)

• 11:59 — WSL protocol server starts (pre-login state)
• ~13:28 — Login on Windows side; credentials.json in WSL written; account state reaches the running WSL server (in memory)
• 15:22 — Model picker works in WSL project; a message sent to GLM-5.3 fails with network errors (unrelated: WSL2 NAT broken with USB 4G modem — fixed separately)
• 15:33+ — wsl --shutdown (multiple times, for the NAT fix), app restarts
• Since then — every fresh WSL server start leaves the model picker empty; evidence above collected across these restarts

Happy to provide full sanitized logs, setting.json dumps, or DB schema details if useful.

实际表现 · Actual behavior

When opening a WSL remote project, the model picker shows only "Manage" — the model list is completely empty and no model can be selected. Local Windows projects in the same app instance, with the same account, list and use models normally (GLM-5.3, zai individual coding plan).

This happens on every freshly started WSL zcode-server, even though all state on the WSL side is valid:

• internet access from WSL verified (api.z.ai returns 200 in ~1s)
• ~/.zcode/v2/credentials.json present and fresh (written by the app itself)
• ~/.zcode/v2/coding-plan-cache.json reports builtin:zai-coding-plan: "available"
• ~/.zcode/v2/setting.json has the same modelProviderFamilySelectedKeys value as the Windows side

The only time the picker worked was on a WSL server process that had been started BEFORE I logged in on the Windows side, i.e. it received the account/plan state in memory during the login flow. After any wsl --shutdown or app restart, freshly started WSL servers never receive the plan state again: the WSL-side log shows zero plan/subscription/entitlement fetch attempts after startup, and the provider registry accountRevision hash differs between Windows (zcode-builtin:28:9cd1cc2f...) and WSL (zcode-builtin:28:ac6ecccf...).

Additionally, when a model selection is attempted, persistence fails with:
event: session.model_selection.persist_failed, error: "FOREIGN KEY constraint failed" (modelId GLM-5.3, providerId account:zai-individual-coding-plan) — the referenced session row was never inserted into ~/.zcode/cli/db/db.sqlite on either side (session_create logged with "persistence": "deferred").

ZCode 版本 · ZCode version

3.12.2

设备 / 系统 / 浏览器 · Device / OS / Browser

Windows / WSL

截图 / 录屏 / 日志 · Screenshots / Recordings / Logs

No response

Start with fresh WSL server startup and the WSL-side protocol log, then compare provider_registry.ready accountRevision values with the Windows side. Inspect ~/.zcode/server/zcode-server.cjs, credentials.json, coding-plan-cache.json, and the CLI db.sqlite evidence around account/plan synchronization and deferred session creation. Done means a restarted WSL server repopulates the model picker and model selection no longer fails with a foreign-key error.