fix(input): make sure a prompt sent through the API actually leaves the composer

[Ark0N]

What broke

Claude Code 2.1.277 (auto-installed on 2026-09-18) changed how a fresh pane treats input: the composer accepts typed text the moment it paints, but ignores Enter for roughly the first 30 to 50 seconds. Before that version the same key sequence had worked for weeks.

Codeman delivers a programmatic prompt as send-keys -l <text> followed 50 ms later by send-keys Enter, and every caller that then waits for the turn (send-and-wait on the input route, the agent skill's sendwait, cron, Ralph, the maintainer bot) assumes the Enter landed. Under 2.1.277 it does not: the prompt sits in the composer with 0 tokens, the transcript stays empty, and the wait burns its whole timeout on a turn that never started.

Measured on 2026-09-19 through the input route on a throwaway session: readiness at 8 s, text plus Enter at 8 s stranded, a bare Enter at 28 s still stranded, a bare Enter at 51 s submitted. The maintainer bot lost twelve reviews in a row overnight this way, each waiting out a 40-minute deadline "without a report".

What changes

Server, the general fix. src/session-submit-verifier.ts is a small pure module: promptStillInComposer(screen, prompt, glyph) reads the LAST line that starts with the CLI's own composer glyph (capabilities.workDetect.promptGlyph, Claude's ❯, Codex's ›) and says whether it still holds the head of the prompt, whitespace removed. It is the last glyph line on purpose, because Claude Code echoes a submitted prompt with the same glyph higher up in the transcript. SubmitVerifier runs that check on a 2 s to 60 s schedule after a write and presses Enter again at every check that still finds the prompt. It stops on an empty composer (taken), on other text (the user's or the CLI's, never ours), and does nothing at all for a pane with no composer line (a shell, a direct-PTY session, a CLI whose glyph is not declared), since a blind Enter from a path every programmatic sender shares could confirm a dialog nobody asked about. A newer write replaces the schedule, so an old prompt can never submit a new one, and teardown cancels it. Session.writeViaMux arms it for every mux write that carried a \r. A prompt that was taken costs exactly one pane capture at 2 s.

Skill, for servers that predate this. skills/codeman/preamble.sh gets _composer_text, which reads the rendered pane the way _trust_key already does and returns the last ❯ line with all whitespace removed, including the no-break space Claude Code draws after the glyph (stripped by its bytes, since BSD sed has no \xHH). sendwait now uses it after its first 20-second wait. The long re-wait (a tagged duplicate of the original frame, so the server re-waits without retyping) is registered first and kept open in the background for the rest of the call, and beside it the loop reads the composer and sends another \r about every ten seconds while the prompt's head is still there, up to twelve times. That shape matters: wait signals have no history, so a stop that fires while no wait is open (during a composer read between two short waits) is lost and the next wait runs its whole timeout on a turn that already ended. The first version of this loop did exactly that, twice, before the wait moved to the background. Prefix matching goes through "${c#"$head"}" so a * or ? in a prompt cannot act as a glob. CODEMAN_PREAMBLE moves from 1.22.0 to 1.30.1 (preamble header, SKILL.md's §0 loader and version checks, reference/recipes.md), so an agent whose cached preamble predates this falls back to the fresh copy. The §0 heredoc in SKILL.md and the plugins/codeman/ mirror are regenerated (both byte-pinned by tests).

Docs. The single-line-prompts gotcha in CLAUDE.md records the window, the verifier, and that the shift+tab footer means "painted", never "accepting Enter". A patch changeset.

Verified

• tsc, lint, format, test/session-submit-verifier.test.ts (the schedule, the cap, every stop condition, the glyph, ANSI, the no-break space, cancellation, a throwing Enter), test/agent-skill.test.ts and test/plugin-manifest.test.ts; full npm test green.
• Server side, end to end: a second Codeman instance started from this branch (CODEMAN_INSTANCE, own port and tmux socket), one throwaway Claude session, one POST /api/sessions/:id/input with text plus \r and no client-side retry of any kind. The server log shows the verifier re-sending Enter and the prompt being taken.
• Skill side, end to end against the production instance: the new preamble sourced into a shell, readiness at 1 s, one sendwait. The trace shows two composer reads (stranded, then empty), one extra Enter, and sendwait returning signal: stop at 36 s with the answer in the transcript.
• Same loop live in the maintainer bot since 03:44 UTC: the first review after its restart stranded at 20 s and 30 s and was taken on the second re-send.

Not in this PR

• writeViaMux still sends text and Enter 50 ms apart; with the verifier behind it that is fine once the CLI is warm and self-correcting when it is not.
• The DeepSeek Harness path gets the same treatment for free (its composer glyph is also ❯), but it was not re-measured here.