Finite live multi-round checkpoint¶
Recorded 2026-09-21 for the entered-prompt frontend.
Setup¶
Actual Pi SDK 0.86.1, existing OpenAI Codex subscription OAuth,
gpt-5.5, low thinking, explicit SSE model transport; no Kimi or API-key fallback.pnpm prototype:pi-multi-live, Windows bundled ConPTY, pinned experimental browser xterm host, fixed 40-column by 8-row viewport.One retained in-memory Pi conversation, four deliberately entered prompts, one fixed read-only sample tool. No project files, arbitrary paths, shell execution, user extensions, or persisted Pi session were enabled.
These calls consumed subscription allowance. No token/cost total was measured. Automatic local-fixture checks were absent/disabled on the live pages.
Expected and observed¶
Action |
Expected |
Observed |
|---|---|---|
Ask Pi to read the sample and return its main idea plus |
The actual tool result appears and the round completes. |
One displayed sample tool result; answer included the sample’s meaning and the marker; Context 1 closed, round 1 completed. |
Ask for the previous answer’s marker without repeating it in the new prompt. |
The same Pi conversation retains earlier context. |
Answer was |
Ask for a numbered list beginning |
Cancellation occurs after nonempty mutable assistant content is observed; no false completion. |
Partial |
Ask for the first answer’s marker followed by |
Another prompt works after cancellation and still has the early context. |
Answer contained |
Search the old tool output, then end the idle session. |
Old content remains searchable and cleanup finishes normally. |
|
Search the cancellation label after exit. |
Retained partial history remains searchable after the connection closes. |
|
The helper was armed only for the third prompt. It clicked the ordinary cancel control when the displayed report contained nonempty mutable assistant text; it did not start a prompt or infer cancellation merely from a click. Browser observations separately verified the aborted outcome and next ready state.
Coverage¶
The four live prompts exercised tool use, retained conversation, assistant-stream cancellation, and continuation. Tool-time cancellation and native reading/selection/copy during subsequent output were checked with the local deterministic provider in the main record.