René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"Orchestration mail should inject a pointer, not the payload (finishes #12536, fixes #7337 for worker panes)","articleBody":"Split out from #12584, which fixed the first half of #12536 (mail to an already-idle recipient never fired). Two related problems remain, and they share one root cause.\n\n## What's still broken\n\n**1. A bound coordinator receives mail but never acts on it (#12536's second half)**\n\n`deliverPendingMessages` writes the banner, then returns early for the active `coordinator_handle` without the delayed Enter, so the text sits at the prompt unsubmitted. The row is still stamped `delivered_at`, so push never retries it.\n\nNote this cannot be fixed by making the coordinator auto-submit. `worker_done`/`heartbeat` are readdressed to `run:\u003cid>` in `rpc/methods/orchestration.ts` whenever the sender has a run and a dispatch, and `run:` is not a terminal handle, so it is rejected before delivery. Lifecycle mail reaches a coordinator only through `check --wait`. The fix belongs at the `run:` consumer layer, not the push path.\n\n**2. Mail still clobbers a half-typed prompt (#7337 is only partly fixed)**\n\nThe original report was a user mid-sentence in the orchestrator whose partial text was submitted together with an incoming worker message. #7391 removed the synthesized Enter for the coordinator pane, but left the write in place — so the banner is still pasted into whatever you were typing, and `delivered_at` is stamped for mail nobody acted on.\n\nWorse, the guard barely applies. `getActiveCoordinatorRun()` is an unscoped global `SELECT ... WHERE status='running' ORDER BY created_at DESC LIMIT 1` over `coordinator_runs`, a table written only by `orchestration.run` and `Coordinator.run()`. No CLI command writes it — the CLI's `run-create` writes the modern `runs` table. So in CLI-driven workflows the branch effectively never fires, and **a human typing into a worker pane gets the banner plus a synthesized Enter today**.\n\n#12584 makes this more likely to bite, not less: push previously fired only on a busy->idle transition, so a pane you were merely typing in was rarely targeted. It now fires the moment mail is sent.\n\n## Root cause\n\nThe terminal write carries the **payload**. That makes a collision destructive (a real instruction is corrupted and lost) rather than cosmetic.\n\n## Proposed direction\n\n**Inject a pointer, not the payload.** Write a short line — \"you have N orchestration messages, run `orca orchestration check`\" — and keep the content in the store the agent reads on demand. Then:\n\n- A collision garbles a short machine-generated string instead of losing a task.\n- Delivery stops being an event that can be missed. The row is durable, so any later trigger drains it; already-idle stops being a special case.\n- `delivered_at` regains an honest meaning: the pointer is not the delivery, the `check` is.\n\n**Then wire `check --inject` into the `UserPromptSubmit` hooks Orca already installs** (~18 agents under `src/main/*/hook-service.ts`), so mail arrives at turn boundaries and the pointer write becomes the rare fallback for runtimes without hooks.\n\nTwo caveats worth carrying:\n- Hooks are reliable for *delivery at turn boundaries* but not as *lifecycle state authority* — they miss transitions (cancellations, interrupts). Keep title/process detection as the status source.\n- `UserPromptSubmit` does not fire for an idle agent nobody is talking to, so a wake path is still needed. That is what the pointer write is for.\n\n**Also worth adopting:** invert the trigger. Send, then use the state change only to *confirm* delivery landed, against a monotonic sequence counter rather than a status comparison — a fast idle->working->idle turn produces no status difference but does bump a counter — with a bounded timeout and a distinct \"no observed effect\" error, so an ineffective injection surfaces instead of failing silently.\n\n## Rejected\n\nTracking unsubmitted keystrokes per pane to detect a draft. Implemented and tested, then discarded: the counter tracks magnitude, which is unknowable. Word-deletion (Ctrl+W, Ctrl+K, Alt+Backspace) empties the box without decrementing, leaving the pane permanently \"dirty\" so mail defers forever — and there is currently no UI surface for pending mail, so that stranding is silent. Ctrl+W cannot be treated as a clear either, since it deletes one word and the box may still hold text.\n\nPrompt-emptiness detection from rendered output was also assessed as unreliable: the composer always renders something, and stripping ANSI makes a dim placeholder byte-identical to typed text. It may be feasible if cell attributes are retained (the emulator is xterm-backed, so the data exists even though `getVisibleLines()` discards it) — unverified, and unnecessary if the terminal only ever carries a pointer.\n","author":{"url":"https://github.com/Jinwoo-H","@type":"Person","name":"Jinwoo-H"},"datePublished":"2026-08-06T22:06:50Z","url":"https://github.com/stablyai/orca/issues/12953"}