You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Long spawn launch line is corrupted mid-send on herdr and tmux (spliced env prefix), agent never starts #4983
Intermittently, fm-spawn.sh / fm-control.sh relaunch leave the pane holding a corrupted launch line and no agent starts. The corruption is always at the same spot: the export COMPACT_ADVISER_DISABLE=1; env ... prefix appears twice, spliced mid-token, with the first copy truncated. Seen on herdr (three times, two different tasks) and on tmux (once, explicit --backend tmux retry of the same task), so it does not look backend-specific.
Two COMPACT_ADVISER_DISABLE=1 fragments spliced together; the first is cut at =-8. In the other occurrences the pane sat on the launch echo with a lone spinner and fm-control relaunch reported the replacement agent ... did not come up within 90s (endpoint reads 'dead').
Conditions
Both affected tasks had long briefs; the launch line embeds the brief through "$('.../fm-operational-input.sh' encode launch-brief < '.../launch-brief.md')", so the one-line command is several KB.
The same brief launched fine on a fresh window a few minutes later (a re-filed task id, new window), so it is intermittent rather than deterministic on the bytes.
No work lost (clean worktree, no branch), but each occurrence costs a relaunch cycle and, on the third attempt, a re-filed task. A send-and-verify of the launch line (or writing the encoded brief to a file and passing its path instead of inlining it) would remove the race.
Verdict:ready-for-pr · contract-class: restore · Firstmate flag: no
Tip check (dd9b2ef):bin/fm-spawn.sh still builds a multi-KB one-line launch (inlined \"$(…/fm-operational-input.sh encode launch-brief < …/launch-brief.md)\") and delivers it via spawn_send_literal → backend pane send-text / tmux equivalent with no send-and-verify. COMPACT_ADVISER_DISABLE=1 is both exported as its own line and prefixed into LAUNCH, matching the spliced-prefix corruption in the report. Seen shape is intermittent on herdr and tmux. Not fixed on tip. Related: #4991.
VISION (per-rule):
Scripts own the mechanics: launch delivery must be exact and repeatable; a corrupted launch line is a scripted transport failure, not agent judgment.
Peace of mind: intermittent spawn death forces relaunch/re-file cycles and erodes confidence that dispatch sticks.
Leave open for a PR: verify the launch line after send (or pass the encoded brief by path instead of inlining) so a spliced prefix cannot leave a dead pane.
Fixed on main by #4994: the launch command is now staged to a private file under /tmp and only a short . <file> source line is typed into the pane, so the multi-KB launch line no longer crosses the pane input buffer that was splicing it mid-send.
Activity
kunchenguid commentedon Sep 20, 2026
Speaking as Kun's firstmate:
Verdict:
ready-for-pr· contract-class: restore · Firstmate flag: noTip check (dd9b2ef):
bin/fm-spawn.shstill builds a multi-KB one-line launch (inlined\"$(…/fm-operational-input.sh encode launch-brief < …/launch-brief.md)\") and delivers it viaspawn_send_literal→ backendpane send-text/ tmux equivalent with no send-and-verify.COMPACT_ADVISER_DISABLE=1is both exported as its own line and prefixed intoLAUNCH, matching the spliced-prefix corruption in the report. Seen shape is intermittent on herdr and tmux. Not fixed on tip. Related: #4991.VISION (per-rule):
Aligns. contract-class: restore (spawn's already-promised start path broken mid-send).
Leave open for a PR: verify the launch line after send (or pass the encoded brief by path instead of inlining) so a spliced prefix cannot leave a dead pane.
devin-ai-integration commentedon Sep 24, 2026
Fixed on main by #4994: the launch command is now staged to a private file under /tmp and only a short
. <file>source line is typed into the pane, so the multi-KB launch line no longer crosses the pane input buffer that was splicing it mid-send.mehulbhagwani commentedon Sep 25, 2026
Reporter close-out
Confirmed fixed by #4994 (launch staged to a private file; pane only sources a short path). No further work from me on this issue. Thanks.