fix(afk): stop herdr escalation redelivery loop

[kunchenguid]

Intent

Fix the away-mode supervise daemon's escalation-redelivery loop on the herdr backend. Symptom: while state/.afk is set and the fleet runs on herdr, bin/fm-supervise-daemon.sh re-injected the SAME buffered escalation digest every cycle instead of clearing state/.subsuper-escalations once delivery landed, waking firstmate's LLM repeatedly for an already-handled escalation. Reproduced live against a real, isolated herdr session running a real claude process as the supervisor pane and the real fm-supervise-daemon.sh (not a synthetic script): observed 5+ identical digest deliveries in 40 seconds. Root cause confirmed empirically (not assumed): fm_backend_herdr_composer_state's structural composer-row classifier only recognized BORDERED composer rows (the grok shape it was originally built for). Real claude's and codex's live input rows are UNBORDERED - claude's is a bare '❯' with no surrounding border characters at all, flanked by plain horizontal-rule separators, and codex's is a bare '›'. Both harnesses' own startup banners ARE bordered, so immediately after launch the 'last bordered row wins' scan locks onto the banner's own blank interior row (a coincidental false 'empty'), and once that banner scrolls out of the 20-line capture window (true of any real, in-use supervisor pane) NO bordered row exists anywhere in view, so the classifier reports 'unknown' for a genuinely idle composer, forever. fm_backend_herdr_send_text_submit only treats 'empty' as a confirmed submit, so escalate_flush never clears the buffer even though the real Enter key actually submitted the digest - and because the pre-type pending-guard only defers on 'pending' (never 'unknown'), the next housekeeping tick retypes and resubmits the identical buffered digest, which is the redelivery loop. Fix: fm_backend_herdr_composer_state (bin/backends/herdr.sh) now recognizes TWO composer-row shapes in one forward scan - the existing bordered shape and a new bare/unbordered shape (a trimmed line starting with a known prompt glyph: claude's ❯, codex's ›, plus the existing >/$/%/# set) - keeping whichever match is bottom-most, since a decorative bordered box is always rendered above the live composer, never below it, in every harness observed. This is a deliberate decision to prefer 'last match of either shape' over 'bordered-only with a bare fallback', because the earlier design was empirically proven wrong (a bordered decorative banner coexisting with a live bare prompt caused a false-empty misread). Also empirically discovered but explicitly NOT fixed in this change (documented as a known, separate follow-up in docs/herdr-backend.md): codex's idle composer shows dynamic/rotating tip text rather than a blank row or a fixed placeholder string, so it cannot be told apart from real pending text by regex; this makes an idle codex composer classify as 'pending' rather than 'empty', which is a strictly SAFER failure mode (defer forever, buffer preserved, max-defer wedge alarm still fires) than the redelivery loop, so it was deliberately left as a documented gap rather than folded into this fix. Similarly, the task's suggested secondary scope item - a 'paused/awaiting-external' crew state so an intentional rate-limit wait isn't flagged as a wedge - was deliberately deferred and documented as a separate backlog-style follow-up note in docs/herdr-backend.md rather than implemented, per explicit instruction to keep this PR scoped to the redelivery-loop fix if the secondary item would balloon the change. Verification performed: re-ran the exact same real-herdr+real-claude E2E repro after the fix and confirmed the composer now reads 'empty' correctly, zero 'inject failed' log lines, the escalation buffer clears, and the digest is delivered exactly once (was unknown/redelivered before). Added 4 new regression tests to tests/fm-backend-herdr.test.sh using verbatim byte-for-byte fixtures captured from the real claude and codex panes (not invented approximations) - confirmed each fails against the pre-fix code (git stash proof) and passes post-fix. All pre-existing herdr backend and daemon test suites (fm-backend-herdr.test.sh, fm-daemon.test.sh, fm-backend-herdr-smoke.test.sh, fm-afk-inject-herdr-e2e.test.sh) still pass unchanged, confirming no regression to the grok/synthetic bordered-composer path or the existing tmux/orca dispatch. shellcheck bin/.sh bin/backends/.sh tests/*.sh (the exact CI invocation, all three globs in one call) is clean. Dated empirical evidence, the codex gap, and the deferred paused-state follow-up are recorded in docs/herdr-backend.md under a new 'Incident (2026-07-07)' section; the new FM_BACKEND_HERDR_BARE_PROMPT_RE env knob is documented in docs/configuration.md alongside the existing herdr composer knobs.

What Changed

• Fixed herdr AFK escalation delivery so unbordered Claude and Codex composer rows are recognized instead of leaving buffered digests stuck in a redelivery loop.
• Narrowed herdr bare-prompt matching to verified agent glyphs and updated cmux/herdr backend handling around composer-state detection.
• Documented the herdr incident, new bare-prompt configuration knob, and known follow-up gaps while expanding herdr AFK/backend regression coverage.

Risk Assessment

✅ Low: Captain, the change is narrowly scoped to herdr composer-row classification and the previous shell-prompt false-empty risk was addressed by narrowing the bare matcher to the verified glyphs.

Testing

Captain, the full configured baseline command was already reported successful; I re-ran the focused herdr backend and daemon tests, ran the real-herdr AFK injection E2E, and captured an isolated AFK/herdr housekeeping transcript proving a bare real-claude composer is classified as empty, the digest is sent once, and state/.subsuper-escalations remains cleared after repeated cycles.

Evidence: AFK/herdr redelivery evidence

composer_state_before_flush=empty buffer_bytes_after_three_housekeeping_cycles=0 send_text_calls=1 enter_key_calls=1 sent_digest_lines=1

composer_state_before_flush=empty
buffer_bytes_after_three_housekeeping_cycles=0
buffer_since_exists_after_flush=no
wedge_marker_exists=no
send_text_calls=1
enter_key_calls=1
pane_read_calls=4
sent_digest_lines=1
sent_digest_preview=<US>Supervisor escalate (       1 event(s)): stale persisted 241s (possible wedge): evidence:w1:p2 (pre-read; re-arm not needed — watcher daemon-managed)
case_dir=/var/folders/5x/4nqprlbx0518k3ybcb1sz6gr0000gn/T/no-mistakes-evidence/01KWYT44J5FYQFP39REC5BWD4D/herdr-afk-redelivery-case
herdr_call_log=/var/folders/5x/4nqprlbx0518k3ybcb1sz6gr0000gn/T/no-mistakes-evidence/01KWYT44J5FYQFP39REC5BWD4D/herdr-afk-redelivery-case/herdr-calls.log
typed_digest_log=/var/folders/5x/4nqprlbx0518k3ybcb1sz6gr0000gn/T/no-mistakes-evidence/01KWYT44J5FYQFP39REC5BWD4D/herdr-afk-redelivery-case/sent-digests.log

Evidence: Real-herdr AFK E2E result

real herdr AFK injection E2E passed scenarios A-D, including swallowed Enter and max-defer wedge preservation.

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

🔧 **Review** - 1 issue found → auto-fixed ✅

• 🚨 bin/backends/herdr.sh:657 - The new bare-prompt default treats plain shell prompts ($, %, #, >) as empty agent composers. If a herdr supervisor/task pane has fallen back to a no-agent shell, away-mode injection can type the digest into the shell, see the prompt as empty, and clear the escalation buffer even though no agent received it. Restrict the bare default to the verified agent glyphs (❯, ›) or add stronger structural checks around bare rows.

🔧 Fix: Narrow herdr bare prompt matcher
✅ Re-checked - no issues remain.

✅ **Test** - passed

✅ No issues found.

• command -v tmux >/dev/null || { echo "tmux is required for e2e tests" >&2; exit 1; }; tmux -V; rc=0; for t in tests/*.test.sh; do echo "== $t =="; bash "$t" || rc=1; done; exit "$rc"
• bash tests/fm-backend-herdr.test.sh
• bash tests/fm-daemon.test.sh
• bash tests/fm-afk-inject-herdr-e2e.test.sh
• Evidence-producing AFK/herdr housekeeping check using bin/fm-supervise-daemon.sh and bin/backends/herdr.sh with fake-herdr CLI responses, writing /var/folders/5x/4nqprlbx0518k3ybcb1sz6gr0000gn/T/no-mistakes-evidence/01KWYT44J5FYQFP39REC5BWD4D/herdr-afk-redelivery-evidence.txt

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.