[abdelnabiii]

Summary

• The inbox-nudge delivery (text, then a fixed 140ms wait, then \r) never checked that either write actually reached the TUI. A swallowed Enter left the text sitting unsubmitted in the input box, and the next wake beat typed a second full nudge on top of it -- the reported "stuck in the input box, retries duplicate it" bug.
• Shortened the nudge to one line (hiveNudge.ts): the full "read inbox, move to inbox/.done/, act autonomously" prose is already injected once into every agent's system prompt at spawn (hive.ts's HIVE PROTOCOL block), so restating it on every nudge only gave a slow/occluded PTY more text to swallow and more to duplicate when it did.
• Added nudgeDelivery.ts: wait 400-600ms for the PTY to echo the typed text (polling, not a fixed guess) before sending Enter; after Enter, confirm the PTY actually reacted within 500ms; if not, resend only Enter (never the text) with 300/600/1200ms backoff, up to 3 times, then give up. Every phase (typed/enter-sent/confirmed/stuck/enter-retry/gave-up) is logged via hive.appendLog to log.jsonl.
• Added in-flight tracking to WorkerWakeWatchdog (markNudging/clearNudging, new 'in-flight' hold reason) so a second delivery can never start for an agent whose first one hasn't resolved -- this was the actual mechanism behind the duplication, not just the fixed-delay timing.
• Hold-if-busy and coalesce were already implemented well in the existing watchdog (mid-turn/booting/cooldown/edge-triggered announce); this only adds the missing in-flight piece on top.

Caveat (stated plainly, not hidden)

"Check whether the prompt line still holds the text" is done via a heuristic -- did the PTY produce any new output -- not true terminal-buffer inspection; the main process has no headless xterm reading rendered screen state (xterm.js only runs in the renderer). It's strictly better than the old fixed-delay/no-feedback sequence, but not literal line inspection.

Scoped to the main-process watchdog's own nudge path (nudgeWorker/runWorkerWakeBeat) only. The renderer's separate submitToPty (useHive.ts, used for slash commands, seed prompts, work orders -- not just nudges) is untouched; flagging it as a known follow-up rather than silently leaving it inconsistent.

Before

Old delivery: a swallowed Enter left the nudge text sitting unsubmitted, and the next beat retyped it -- duplicated in the input box. Reproduced in evidence/demo-nudge-fix.cjs against the removed pre-fix nudgeWorker logic:

After

Fixed delivery (nudgeDelivery.ts, this PR): the same swallowed-Enter scenario now retries Enter alone, confirms the PTY reacted, and never retypes the text:

Test plan

• New test/nudge-delivery.test.cjs (6 tests, including a 3-concurrent-agents-with-different-retry-depths scenario)
• Updated test/worker-wake.test.cjs (added in-flight coalescing tests, incl. "3 agents nudged at once, one mid-task")
• Updated test/hive-nudge.test.cjs for the shortened text
• node --test test/*.test.cjs: 970 tests, 25 pre-existing failures (confirmed via git stash to fail identically without this change -- Windows symlink-permission and Electron-runtime tests, unrelated)
• npx tsc --noEmit -p tsconfig.node.json: clean

🤖 Generated with Claude Code

https://claude.ai/code/session_01MBNN17AULnHcT97qhTc4ja

[github-actions]

✅ Evidence received. Before and after are both attached. Thanks — this is what makes a PR reviewable in one pass.

[abdelnabiii]

@chaitanyagiri friendly ping on this one -- CI is green (before/after evidence check passed) and it's been open since Oct 13. Happy to address any review feedback. Thanks!