Files
Netcatty/docs/plugin-platform/ai-provider-startup.md
T

5.9 KiB

Plugin model transport cold start

Commit 092438803bed96753bfe68cb78da9ca1137dee25 removes an independent startup dependency on the internal AI sidebar. It does not establish the cause or outcome of the original native failure; that still requires an authorized native retest.

Reproduced boundary

At native candidate 677aac33b1eec9dbd4a367417a49c964cdea63e7, the plugin model selector reads the saved host provider, while main's provider registry starts empty. The existing sidebar synchronizes that registry when visible or sending. The plugin startup path did not. A nondefault configured loopback port can therefore fail main's provider allowlist before HTTP transport starts.

The revised components/PluginAIContextHost.test.tsx installs the browser sample, mounts the trusted host component, uses the real application preload, and routes model requests through registerProviderHandlers and its actual allowlist. It does not mount the internal sidebar or manually alter the allowlist. Credential injection uses synthetic fixture markers; the HTTP edge and deterministic local fixture use the existing transport. No paid service, real key, Note, or host is used. This remains DOM/in-process evidence, not native sandbox acceptance.

Failed-before evidence is /tmp/netcatty-ai-cold-provider-before-final.log:

ok: false
allowlistRejected: true
statusCode: undefined

The corrected regression passes and preserves both the selected fixture provider and a second host-configured synthetic provider. It checks that provider markers and configuration do not enter plugin-facing model/context envelopes. Expected SDK error output is suppressed during this diagnostic so request payloads and headers are not logged.

Correction

Before publishing a selected model handle, the trusted renderer synchronizes the current host-owned provider settings through the existing aiSyncProviders application API. It checks that the selected model revision matches the snapshot being synchronized. The runtime rechecks activity, cancellation, model revision, and pending-selection ownership after the await. Pending selections count toward the model-handle cap; release and shutdown invalidate late results.

Plugin requests supply no URL, credential, or provider configuration. The existing main sender validation, endpoint allowlist, source binding, and model invocation budgets are unchanged. No new schema feature, plugin permission, or consent dialog is added. The assistant package is unchanged and retains SHA-256 c7d784d5c3bd2feb463afd58eae3a9594615395b168174cb8a3d0677584ae395.

Production integration touches configuredModel.ts, trustedHostRuntime.ts, and usePluginAIContextHost.ts. The hook's import and prepareModel option may overlap host-selection work; preserve both when integrating the separate policy task. Permission-engine, consent-dialog, and sample-UI changes are outside this correction.

Native and policy status

The native open/close sidebar comparison was incomplete. Execution auto-review denied the Application-scope context.selected.list grant; that action did not execute. No post-sidebar inference was performed, the prompt expired, and the sample was disabled while the worker cleaned up. This is not native confirmation of the startup cause. The original 677aac33 app remains frozen for comparison; this fix does not authorize retrying the denied execution action.

In the earlier permission engine, Application grants are memory-only for the application lifetime; Always grants are persisted. An explicitly authorized Application grant can fit bounded temporary QA. A one-use prompt on every stream read is not an unavoidable target UX. The separate policy task implements the new direction: default Allow for declared supported capabilities, per-plugin Allow/Ask/Deny settings, audit, and in-app Ask dialogs. That product change is separate from execution-tool authorization and must not be used to evade a tool denial. This transport correction does not redesign that policy.

Native outcome, independent verification, and the sample's light-theme appearance follow-up remain separate acceptance work. No GUI launch, sandbox bypass, paid model invocation, main merge, or release was performed for this correction.

Node 22 CI observer correction

The new cold-start test failed on exact head 9ca70f329 in test run 40436198585, Linux job 121965743101: 15,206 passed, 1 failed, 50 skipped. The sole failure was the empty SDK/renderer-error assertion receiving one SDK_OR_RENDERER_ERROR marker. This red result is retained as evidence; it is not an unrelated baseline failure.

The CI runtime was Node 22.23.3; local validation had used Node 24.19.0. Reproduction on the exact CI version identified the safe category setup:node-sqlite-experimental, before any model operation. The existing node:sqlite import schedules its ExperimentalWarning on the next tick via console.error. The newly added observer was installed before that warning drained. The same test passed on Node 24, which did not emit that warning.

The test now lets pending module warnings reach normal stderr before installing the SDK/renderer observer. The warning remains visible, and the zero-runtime-error assertion is unchanged. Safe phase/error categories make future failures diagnosable without request payloads or headers. No production behavior, timeout, permission, or cancellation assertion changed. Both Node versions pass the focused test; the Node 22 parallel fixture/UI/provider/preload group passes all 16 tests with concurrency 4.

Diagnostic evidence: /tmp/netcatty-ai-cold-ci-node22-sqlite-category.log fails before correction, /tmp/netcatty-ai-cold-ci-node22-after.log passes with the SQLite warning still visible, and /tmp/netcatty-ai-cold-ci-node22-parallel.log records the parallel check. The appearance-only sample change is independently published as e325742abf9fd329f1f7858eda66367e7a62357c; its package hash is 7943d4c3800914fe0acd3d0f3a3d7f7c1910c212a5dc58a044871a6a82485274.