Skip to content

Latest commit

 

History

History
123 lines (104 loc) · 7.35 KB

bot-continuity.md

File metadata and controls

123 lines (104 loc) · 7.35 KB

Bot setup, model scope, and file continuity

pnpm exec vitest run server/setup-mode.test.ts server/bot-setup.e2e.test.ts server/bot-continuity.e2e.test.ts server/independent-threads-api.test.ts
OMB_UI_E2E=1 pnpm exec vitest run scripts/testing/control-omb-ui.e2e.test.ts
pnpm exec electron scripts/smoke-approval-modes.cjs --model-ui-only

The server recipe launches the real app in a temporary home with only the offline CLI. It sends actual messages through control-omb, waits for settled turns, and checks the prompt and model delivered to the provider boundary:

  • A named bot with no description or SOUL can receive a normal work request without the setup interview or its "Wait for a yes" instruction.
  • /setup Help me track garden watering explicitly enables coaching; the next ordinary request disables it, even if no profile card was confirmed.
  • A pinned thread's model change with updateBotDefault: true updates that thread and the bot default used by groups and future threads. It does not change a selected sibling, other saved model choices, or approval levels.
  • Without that flag, only the selected thread changes. Group dispatch still uses the bot default. Busy group changes and invalid requests are rejected.
  • Direct and room prompts carry the same saved identity and standing rules, plus explicit current/shared/sibling file locations. Old files are retained; working-directory pins and concurrent-thread isolation do not change.

The test prints a retained .bot-continuity.json evidence path with commands, wait results, bounded transcripts, and provider-input receipts. Launch environments and MCP tokens are not retained. The renderer recipe checks both scope buttons in the real model picker, saves a screenshot, changes models, checks server persistence, and sends a message through the composer.

The header starts at Only this thread; Thread + bot default explicitly includes group turns and future threads. The desktop model-switch recipe uses the real picker and private approval channel against offline providers. It checks Cancel, the 390px confirmation layout, switching a Custom Codex thread to Claude with Ask in one confirmed operation, then updating a mismatched bot default without changing another existing Custom thread. HTTP cannot bypass Custom; ordinary Full switches may use the atomic HTTP downgrade. A fresh thread adopts the new default, and a sample engineering-handoff request sent through the composer completes with the fake provider reply. Screenshots stay in .omb-scratch/verify-evidence/model-switch/. This proves settings and turn dispatch, not the quality of a real model's engineering output. The store test also simulates a failed disk write and confirms neither scope changes.

The picker smoke also includes an installed but signed-out Claude account and a missing Codex installation. The signed-out account stays in the picker's Account list; choosing it shows its sign-in card and Use a local model, never its cloud models. The missing Codex, which no bot uses, stays in Settings only, and the picker keeps an Engines and accounts shortcut. A missing engine that a bot already runs on stays in that bot's picker with its install card, so the model never silently vanishes. src/lib/engine-rail.test.ts covers these rules for each Claude account, empty catalogs and leaving the source catalog unchanged. src/components/ModelPicker.interaction.test.ts opens the real picker component and checks the sign-in and install cards, that no cloud model can be picked on a signed-out account while its local models still can, and that opening local models re-probes local servers behind a Looking for local models… status that ends after five seconds at most. The smoke selects a configured local Codex model, browses Claude, closes and reopens the picker, and checks that the selected local model is visible again.

Signed-out Claude in the picker — 2026-09-26

After a Discord report that Claude could no longer be selected, the picker was checked in the headless renderer (control-omb ui launch, a Vite preview and a headless browser on a disposable fixture; no Electron, no real account):

FAKE_CLAUDE_AUTH=out node --experimental-strip-types scripts/control-omb.ts ui launch --mode not-logged-in

The fake CLI answers auth status with loggedIn: false and every prompt with Claude's own "Not logged in" frame. Pepper is on that Claude.

  • Opening the picker shows Claude on the rail with Sign-in required, the Sign in to Claude card and Use a local model; no Claude model row is offered. Before this fix the filter returned no engine at all for this fixture; that reproduction is now a test in src/lib/engine-rail.test.ts.
  • Use a local model sent one POST /api/instances/claude/refresh-models and showed Looking for local models… (held for three seconds by delaying that request in the page), then No local models found.
  • Sending hello on the signed-out Claude ends in the existing setup error with the Sign in to Claude card, not a silent failure.
  • Two more accounts were added through POST /api/instances/claude-accounts. Spare and Pepper's own account were then pointed at a missing CLI with PATCH /api/instances/:id. The picker kept Pepper's account with its install card and the CLI-not-found reason, kept the signed-out, unused Work account (choosing it showed its sign-in card and did not change Pepper's model), and left Spare out.
  • The browser console had no errors.

This proves the renderer and the fixture server's instance catalog. It does not prove a real Claude Code sign-in, a real Ollama or LM Studio server, or the Electron shell, and it is not a production qualification. The Electron picker smoke's assertions were updated for these rules but were not run for this change.

Pepper's signed-out Claude in the picker

Opening local models re-probes local servers

Pepper's Claude with a missing CLI keeps its install card

Live Claude smoke — 2026-09-12

Separately, the exact system prompts captured from the disposable app were replayed to the signed-in Claude CLI (Sonnet 5). These were synthetic requests, not live user conversations. --safe-mode, --restricted, --strict-mcp-config and --no-session-persistence disabled customizations and saved sessions. Tools were disabled except for the final test, which allowed only Read inside the disposable working directories.

Request Observed response
What is 17 + 25? Answer with the number. 42
Explicit setup for tracking garden watering Asked about the job, timing, apps and folder
Find garden-plan.txt from earlier work, outside the current directory Read the shared-bot file and returned its hidden verification code, MOSS-42

An arithmetic control using the old automatic-setup block also answered 42. This does not reproduce every reported refusal, establish malicious intent, or prove every provider behaves identically. The verified regression is the unrequested coaching sent by OMB; live results confirm ordinary work, explicit setup and cross-folder reading on the tested Claude model. Group model routing and UI persistence use the offline provider. No real bots or files were moved.