fix(interactive): the trigger prompt was never actually submitted - #278

Two defects in the injection step, both found by watching a real claude TUI in its tmux pane. Together they meant interactive trigger mode had never submitted its own prompt — the pane sat with the text visible and unsent until the 2-hour timeout, which reads as a hung agent rather than a delivery bug.

1. The prompt was injected as a multiline payload

tmux send-keys -l sends text literally, newlines included. That does not submit each line — it leaves the TUI in multi-line input, and the Enter that follows adds another line instead of sending. Observed directly:

❯ add() subtracts instead of adding
  See .task.md for full spec.
  Follow the instructions in .agent.md.

…sitting in the composer indefinitely.

The default trigger prompt is "<title>\n\nSee .task.md for full spec.", so this was never an edge case.

flatten_for_injection collapses the prompt to one line. Nothing is lost — the detail lives in .task.md and .agent.md; the prompt is only a pointer.

Worth noting: the doc comment on append_sentinel_to_prompt already said the payload must stay newline-free, but the rule was only ever applied to the sentinel suffix, never to the prompt it was appended to. Its explanation of what a newline does was also wrong — it doesn't submit early — so that's corrected to match what the pane actually shows.

2. Enter was sent too soon after the text

Even flattened, the prompt stayed unsent. These TUIs ingest pasted text asynchronously and an immediate Enter is dropped. Confirmed by hand: the same pane submitted instantly when Enter was sent a moment later.

A 600ms settle now separates the two — generous on purpose, since the cost of waiting is a fraction of a second on a run measured in minutes, and the cost of being early is a trigger that hangs for two hours.

Verified end to end on a real column

That last row is correct behaviour, not a defect: interactive completion is advisory because the session stays alive for the human. The code says so itself, and the run confirmed the advisory path works.

Also confirmed on this path: git status in the worktree showed only math.js and agent-proof.txt — .task.md and .agent.md correctly excluded, so #276's fix holds here too.

Checks

cargo clippy -D warnings · cargo test --lib 566 (+2) · tsc · eslint · test:ipc · test:type-scale · vitest 444.