fleet/fleetd
3
0
Fork 0
Code Issues 73 Pull Requests Actions Packages Projects Releases 1 Wiki Activity

#480 lead rollover: /clear and bootstrapText concatenate into one line — the roll never bootstraps #489

New Issue
Open
opened 2026-09-12 02:18:36 +02:00 by ltms · 0 comments
ltms commented 2026-09-12 02:18:36 +02:00
Owner
Copy Link
Copy Source

Measured live on the Mac, 2026-09-12, first real rollover of the shipped feature.

What happened

fleet_handover{action:"confirm"} was accepted, the turn ended, and the pane received a single submitted line:

/clearFresh lead session. First run fleet_whoami; …

Claude Code answered Unknown command: /clearFresh. The context was not cleared, nothing was destroyed, and the handover file stayed intact. The roll failed safely, and did nothing.

Evidence

From the lead session transcript (d17cb4ad-….jsonl, lines 2352-2354) and fleetd/fleetd.out:

00:03:43.560  system turn_duration                     <- the calling turn ends
00:03:43.998  lead-rollover: rolled token=94d81e93-...  (fleetd.out)
00:03:44.309  warning: Unknown command: /clearFresh
00:03:44.310  warning: Args from unknown skill: lead session. First run fleet_whoami; ...

/clear + Fresh lead session. ... = /clearFresh lead session. ..., joined with no space, because a paste lands at the cursor. The whole four-step roll ran in 438 ms of a 20-second budget.

Cause

Two faults, in LeadRollover.runRollover (fleetd/src/main/java/dev/ltms/fleet/lead/LeadRollover.java:380-392):

  1. The second gate is a no-op. waitUntilAtTurnBoundary(lead, clearSettleSeconds) returns on its first poll, because /clear starts no turn and the pane never leaves IDLE. LeadRolloverTest:204 already says this out loud - "default agentStatus is idle - both waits settle immediately" - and nobody asked whether it should. The first gate (turnSettleSeconds) is fine; it is what consumed the 13 seconds before the failure.
  2. The submit keystroke raced the paste. AgentControl.submit's own javadoc records this: "The submit that accompanies a delivery can race the paste ... leaving the text unsubmitted; the injector nudges it with this until the worker actually picks up (CB-113)." LeadRollover bypasses Injector on purpose (routing /clear through it wedges the pane forever), so it inherits none of that nudging. The /clear Enter was lost, the text sat in the input box, and the next paste appended to it.

A live probe on a throwaway sonnet member confirms the second half: a /clear sent through Injector executed cleanly - a new transcript was born at 07:15:29 whose line 4 is literally <command-name>/clear</command-name>. The difference between that success and this failure is the nudge.

The fix

Copy the pattern Injector already ships for post-turn /clear housekeeping (fleetd #306, Injector.java:293-334): after sending /clear, re-send Enter with AgentControl.submit on each poll while the pane is still injectable and no pickup has been seen, bounded like PICKUP_GRACE_POLLS = 8. A second Enter on an empty Claude Code prompt is a no-op, so the nudge is safe to repeat. Only then send bootstrapText.

Related

  • fleetd #480 - acceptance criterion 6 ("the bootstrap prompt lands in the fresh session") is now answered negatively, not merely unproven. wiki/11-Features.md and .claude/skills/handover/SKILL.md both still say "not yet proven end-to-end" and need correcting to "measured failing".
  • fleetd #486 - the settle poll has no bound of its own.
  • fleetd #482 - a context reset changes the peer's session id.
Measured live on the Mac, 2026-09-12, first real rollover of the shipped feature. ## What happened `fleet_handover{action:"confirm"}` was accepted, the turn ended, and the pane received a single submitted line: ``` /clearFresh lead session. First run fleet_whoami; … ``` Claude Code answered `Unknown command: /clearFresh`. The context was **not** cleared, nothing was destroyed, and the handover file stayed intact. The roll failed safely, and did nothing. ## Evidence From the lead session transcript (`d17cb4ad-….jsonl`, lines 2352-2354) and `fleetd/fleetd.out`: ``` 00:03:43.560 system turn_duration <- the calling turn ends 00:03:43.998 lead-rollover: rolled token=94d81e93-... (fleetd.out) 00:03:44.309 warning: Unknown command: /clearFresh 00:03:44.310 warning: Args from unknown skill: lead session. First run fleet_whoami; ... ``` `/clear` + `Fresh lead session. ...` = `/clearFresh lead session. ...`, joined with no space, because a paste lands at the cursor. The whole four-step roll ran in 438 ms of a 20-second budget. ## Cause Two faults, in `LeadRollover.runRollover` (`fleetd/src/main/java/dev/ltms/fleet/lead/LeadRollover.java:380-392`): 1. **The second gate is a no-op.** `waitUntilAtTurnBoundary(lead, clearSettleSeconds)` returns on its first poll, because `/clear` starts no turn and the pane never leaves `IDLE`. `LeadRolloverTest:204` already says this out loud - *"default agentStatus is idle - both waits settle immediately"* - and nobody asked whether it should. The first gate (`turnSettleSeconds`) is fine; it is what consumed the 13 seconds before the failure. 2. **The submit keystroke raced the paste.** `AgentControl.submit`'s own javadoc records this: *"The submit that accompanies a delivery can race the paste ... leaving the text unsubmitted; the injector nudges it with this until the worker actually picks up (CB-113)."* `LeadRollover` bypasses `Injector` on purpose (routing `/clear` through it wedges the pane forever), so it inherits none of that nudging. The `/clear` Enter was lost, the text sat in the input box, and the next paste appended to it. A live probe on a throwaway sonnet member confirms the second half: a `/clear` sent through `Injector` executed cleanly - a new transcript was born at `07:15:29` whose line 4 is literally `<command-name>/clear</command-name>`. The difference between that success and this failure is the nudge. ## The fix Copy the pattern `Injector` already ships for post-turn `/clear` housekeeping (fleetd #306, `Injector.java:293-334`): after sending `/clear`, re-send Enter with `AgentControl.submit` on each poll while the pane is still injectable and no pickup has been seen, bounded like `PICKUP_GRACE_POLLS = 8`. A second Enter on an empty Claude Code prompt is a no-op, so the nudge is safe to repeat. Only then send `bootstrapText`. ## Related - fleetd #480 - acceptance criterion 6 ("the bootstrap prompt lands in the fresh session") is now answered **negatively**, not merely unproven. `wiki/11-Features.md` and `.claude/skills/handover/SKILL.md` both still say "not yet proven end-to-end" and need correcting to "measured failing". - fleetd #486 - the settle poll has no bound of its own. - fleetd #482 - a context reset changes the peer's session id.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
blocked
Cannot start until something else lands. The body says what.
 needs-live-proof
Merged and green, but never shown working on the running daemon. Not the same as done.
 ready-to-delegate
Scope, files and acceptance criteria are written. A worker can be briefed from the body alone.
 silent-default
A feature that compiles, passes tests, and ships turned off. Nine recurrences and counting.
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
agent kevin ltms
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: fleet/fleetd#489
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?