[Bug] pi agent catalog fails: 'model must be provider/modelId' when account default model is a bare id; SDK missing (ERR_MODULE_NOT_FOUND) until managed runtime ready

[baoyu0]

Summary

The pi agent catalog fails to load when the active account's default model is a bare model id (e.g. deepseek-v4-flash), because the pi adapter requires provider/modelId format and throws, which fails the whole catalog. Separately, at startup the daemon reports @earendil-works/pi-coding-agent as missing (ERR_MODULE_NOT_FOUND) until the managed runtime asset is ready, producing repeated "Failed to load agent catalog" noise.

Environment

• LinkCode desktop 0.20.0
• Windows 11 x64
• Account created from the "opencode-go" service catalog: { "model": "deepseek-v4-flash", "endpoint": { "baseUrl": "https://opencode.ai/zen/go/v1", "protocol": "openai-chat" } }

Symptom

%APPDATA%\LinkCode\logs\main.log:

{"subsystem":"agent","operation":"agent.catalog","err":{"type":"Error",
 "message":"pi: model must be 'provider/modelId' (got 'deepseek-v4-flash')",
 "stack":"Error: pi: model must be 'provider/modelId' (got 'deepseek-v4-flash')
   at createConfiguredRegistry (.../out/daemon/index.mjs:97941:11)
   at PiAdapter.startCatalog (.../out/daemon/index.mjs:97993:17)"},
 "msg":"Failed to load agent catalog"}

and, during the first ~20 s after startup:

"message":"pi: SDK '@earendil-works/pi-coding-agent' is unavailable (Error: Cannot find package '@earendil-works/pi-coding-agent' imported from .../out/daemon/index.mjs)"
"cause":{"code":"ERR_MODULE_NOT_FOUND"}

Root cause 1 — model format

PiAdapter.startCatalog → createConfiguredRegistry does:

const ref = opts.model ? parseModel(opts.model) : null;
if (!ref && opts.model) {
  throw new Error(`pi: model must be 'provider/modelId' (got '${opts.model}')`);
}

opts.model is the account's default model, which service-catalog accounts store as a bare id (the correct format for the claude-code/codex/opencode CLIs that share the same account). A bare id makes pi's catalog throw, so the pi agent surface is entirely unavailable until the user manually sets the model to provider/modelId (e.g. opencode-go/deepseek-v4-flash).

Root cause 2 — SDK not bundled

PiAdapter.importSdk() first tries agentRuntimeProber.resolveEntry("pi") (the managed runtime asset) and only falls back to import("@earendil-works/pi-coding-agent"), which is not bundled in app.asar. Catalog retries fail with ERR_MODULE_NOT_FOUND until asset.ensure finishes ("Managed agent runtime ready", ~20 s after startup). If the managed runtime is GC'd by asset.gc (we observed "Removed superseded managed assets"), the errors return until the asset is re-ensured.

Repro

1. Create an account from a service catalog with a bare default model (e.g. deepseek-v4-flash).
2. Enable the pi provider.
3. Restart the app; watch agent.catalog errors in main.log.

Workaround

Set the account's model to provider/modelId format (e.g. opencode-go/deepseek-v4-flash) in the app settings or ~/.linkcode/config.json.

Suggested fix

• pi adapter: tolerate bare model ids by defaulting the provider to the account's service/endpoint name instead of throwing (or normalize the account model when configuring pi), so a shared account with a bare model id still produces a working catalog.
• Packaging: bundle @earendil-works/pi-coding-agent in the app, or gate catalog errors on the managed-runtime readiness state to avoid the misleading ERR_MODULE_NOT_FOUND noise.

[pullfrog]

  ^{｜ Build this ➔ ｜ Make a plan ➔}

[linear-code]

CODE-579