Skip to content

Latest commit

 

History

History
84 lines (71 loc) · 4.95 KB

custom-acp-models.md

File metadata and controls

84 lines (71 loc) · 4.95 KB

Custom ACP model and thinking choices

pnpm exec vitest run server/custom-acp-models.e2e.test.ts
pnpm exec vitest run src/components/CustomAcpModelPicker.test.ts src/components/ModelPicker.test.ts src/components/ModelPicker.interaction.test.ts src/components/ModelPicker.simple.test.ts

The end-to-end test uses launchVerificationServer from the shared control surface. It creates a disposable home, data directory and port pair, stops only the child it owns, adds the repository's independent fake-custom-acp-models.ts engine while stopped, then restarts the same isolated server. Model calls use only that offline ACP fixture. No installed ZCode, CodeBuddy or Qoder binary, real account, credential, or user data is used.

The fixture advertises two models using ACP configOptions: glm-5.3 offers low and max; glm-5.3-flash offers medium and high. These values are synthetic protocol cases, not a statement of real provider capabilities. Thought levels use the thought_level category and the deliberately nonstandard config ID thinking, so a driver hardcoded to the ID effort cannot pass.

The test requires catalog refresh to expose the model labels, model-specific thinking options and capability without sending a prompt. It saves independent defaults on two bots and an explicit override on a sibling thread, then sends through control-omb send, wait and messages with the fixture URL explicitly pinned. The fake ACP process records only model settings, session IDs and synthetic request markers. Its reply includes the effective model and thought level, so an auxiliary call or merely successful settings save cannot satisfy the main-turn assertion.

After alternating both bots and the sibling, the test changes one bot's model and thought level, checks the other bot and explicit sibling remain unchanged, then restarts the real server and sends all three again. The settings are read through the HTTP API after both the change and restart. A fresh instances read after restart must retain the complete cached catalog and per-model variants without another discovery call. Native session receipts must show that each saved thought level was applied using session/set_config_option before its prompt.

Evidence is retained alongside the server log as *.custom-acp-models.json, including control command outcomes, read-back selections, the discovered catalog and sanitized ACP receipts. The test prints both paths. Cleanup stops the exact owned child and removes only the fixture home. The persistent evidence does not contain bearer tokens, full prompts, provider credentials or real account identifiers.

The renderer tests use the production catalog resolver against the offline ACP fixture, then pass its unmodified catalog into the complete ModelPicker with synthetic ZCode, CodeBuddy, Qoder and QoderCN instance identities. They open the picker, require the advertised model names to be visible, invoke its model button, reopen it on the saved model, choose a supported thinking option and verify the saved choice survives a fresh render. They also verify model-specific options, preserving the selected bot/thread scope when saving, discarding an incompatible choice after a model switch, ignoring runtime options belonging to another instance, and hiding unsupported thinking controls. The existing ModelPicker suites also run to cover common selection behavior.

For the before screenshot, the unmodified base commit 9596671 was exported into a separate directory and launched with its own isolated server and the same synthetic bot/provider setup. Its custom engines advertised only agent-default, with no variant capability; the picker displayed "No models to show here yet." No feature code was used for this baseline.

Before: custom ACP picker has no native model choices

The real renderer was also opened in an isolated browser preview backed by the same disposable server and fake ACP engine. The operator opened the synthetic novel-zcode-flash bot profile, opened its model picker, selected GLM-5.3-Flash, selected high, and reloaded the page. An API read-back after reload retained that bot's model and thinking choice while the second synthetic bot still used agent-default. All four synthetic provider identities were visible in the picker. These screenshots show the implemented feature against the offline fixture; they are not screenshots of installed provider accounts.

Custom ACP providers and model choices

Selected Flash model and its supported thinking choices

These checks do not establish a packaged desktop launch. The isolated E2E establishes OMB persistence and actual ACP transport against a fake agent; acceptance by an installed provider must be reported separately, with the exact model ID and supported options advertised by that provider. A failing or skipped provider check is not evidence of a successful model switch.