mirror of
https://github.com/stablyai/orca.git
synced 2026-10-07 08:02:21 +00:00
990b0ed5cd
* fix(orchestration): keep pressing Enter until an OpenCode worker's task starts OpenCode 2 draws its input box before it has loaded its agent list from its server, and silently drops an Enter that arrives in that gap. On a loaded machine the gap outlasts the 0.5 s Orca waits after pasting, so a worker's task sat unsent in the box while worker-start reported it ready. OpenCode's turn start is now observed through its status hooks, like Codex and Claude. While no turn has started, the prompt verifier presses Enter again on an interval (per the agent row, opencode and opencode2) for as long as the pane's hooks are reporting, never past an approval or question prompt or a replaced pane, and at most a fixed number of times. A pane whose hooks never report keeps today's single Enter and unobserved receipt. Co-Authored-By: Claude <noreply@anthropic.com> * fix(opencode): give a worker its task only once OpenCode can submit it OpenCode 2 draws its input box, and shows its cursor there, before its agent list has loaded from its server, and it silently drops an Enter that arrives in that gap. Orca pasted a worker's task on the box's cursor and pressed Enter half a second later, so on a loaded machine the task stayed unsent in the box while the worker read as started. OpenCode's paste readiness now also waits for the row under the input box (`<agent> · <model>`), which OpenCode paints only once the agent list exists. The rule keys on the row's separator inside OpenCode's alternate screen, not on names or colours. Both the opencode and opencode2 rows use it, worker start and the renderer's draft paste alike. OpenCode 1 paints the row in the box's own frame, so it is ready on the same read as before. Recorded OpenCode 2.0.21 boots (cold, hidden pane, four at once, attached to a running server) are the evidence; two more show an Enter after the row submitting and one at the box being dropped. The earlier resubmit loop and the settlement-provider changes are removed. Co-Authored-By: Claude <noreply@anthropic.com> * fix(opencode): prove the agent-row signal with natural captures and never paste a draft blind Replaces three recordings made by launching four OpenCode boots at once with natural single-run evidence: one cold start that pasted on the agent row and pressed Enter on Orca's real 500 ms timing (submitted), and one cold start recorded under heavy load from other sessions, where the paste on the box's cursor and Orca's Enter came before the agent row and the brief stayed unsent. Adds a 2.0.14 cold start as a cross-version check. The renderer's draft paste no longer falls back to a blind paste when the OpenCode agent-row signal never arrives: OpenCode would drop the Enter and leave the draft in its box, so the caller gets its timeout notice instead. Other agents keep the fallback. Co-Authored-By: Claude <noreply@anthropic.com> * fix(opencode): type the single-signal predicate against the public signals Co-Authored-By: Claude <noreply@anthropic.com> * fix(opencode): read the agent row by its place on screen and wait for it only before Enter - The agent-row signal now keys on structure: the `·` must sit on the line directly above the input box's bottom-left corner, so a `·` in the footer path or a session tab title no longer counts. Leaving the alternate screen or turning bracketed paste off withdraws what each part established. - Only deliveries that press Enter wait for the row (new `submitPasteReadySignal`); drafts keep the box-cursor rule, as on main. Worker start always submits. - When the box cursor has shown but the row never comes (OpenCode 2 drops the agent below 44 columns), both waiters proceed after a 5 s grace, as main did. A natural 40-column capture proves that start submits. - The renderer's blind-paste fallback is back to main's behaviour. Co-Authored-By: Claude <noreply@anthropic.com> * fix(opencode): take the box at the readiness deadline while the agent-row grace is pending - Both paste waiters resolve ready, not timed out, when the hard deadline lands while the 5 s grace is pending: the box was seen, so this is the verdict the box rule gave before. - Tab-bound draft paste is tested to wait on the agent row only when Enter follows. - Scanner: a separator pairs with its corner only inside one synchronized frame, row 0 clamps to 1, and the grace comment names the transparent-theme case. The draft signal type no longer admits the agent-row signal, whose grace only the paste waiters honour. Co-Authored-By: Claude <noreply@anthropic.com> * fix(opencode): withdraw the agent-row grace when OpenCode leaves the alternate screen If OpenCode exited after drawing its box but before its agent row, the grace still ran out and the task was pasted into the shell. Leaving the alternate screen now clears the box cursor, and once an app has left it a cursor the shell shows outside it no longer starts a grace. A terminal that never forwards the alternate screen still gets one. Co-Authored-By: Claude <noreply@anthropic.com> * test(serialize): register the new OpenCode captures' pre-existing grid-restore counts The six OpenCode 2 captures this PR adds hit the serializer limitations the 1.18.32 and 2.0.18 captures are already registered for: the live pen's background (and once its bold bit) on restored cells, rows kept below the source after a shrink, and the cursor one column short after a shrink. No I1 or I3 failure; the serializer is unchanged. Co-Authored-By: Claude <noreply@anthropic.com> * fix(opencode): bound the readiness deadline's last settle check With a grace pending, the deadline runs one more settle check, and over SSH its foreground read could hold a past-due wait for the channel's 30 s timeout. It now gives up after 2 s. Tests cover a hand-off, a failure and an abort during that check, a refused grace settled at the deadline, the bound, and the grace restarting at a hand-off. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: m4air <m4air@Mac.localdomain> Co-authored-by: Claude <noreply@anthropic.com>