scaffold: fleet-dispatch-fixes — 6 findings from marketing-waves telemetry#189
Merged
NagyVikt merged 4 commits intoMay 18, 2026
Conversation
…c change
Captures 6 dispatch-path defects surfaced by real telemetry from the
2026-05-18 marketing-content-waves fleet against recodee:
F1 — stale dead panes lingering silently in overview chrome
F2 — cap-probe cache outliving quota recovery (5/6 healthy → 8/8 on
fresh probe 5min later)
F3 — wake-prompt window blank on bringup; workers idle at default Codex
placeholders ("Implement {feature}", "Find and fix a bug in @filename")
F4 — plan-watcher re-validates without --allow-waves, silently falls back
to next-priority plan when our plan has depends_on (observed:
trading-edge-foundations dispatched while our priority plan skipped)
F5 — force-claim "not in a mode" on non-idle Codex panes, drops dispatch
with no retry/backoff
F6 — Codex auto-submit not firing on send-keys: context drops (text
arrived) but no Colony claim recorded — likely needs different
terminator key sequence
Plan workspace at openspec/plans/fleet-dispatch-fixes-2026-05-18/ has 6
parallel-ready subtasks (disjoint file_scope, no depends_on so plan-
watcher accepts without --allow-waves until F4 lands). Each subtask
ships with a focused test under scripts/codex-fleet/test/.
This PR is the scaffold; implementation comes from the fleet itself
(separate per-subtask PRs from the fleet, then one squashed integration
PR per the OpenSpec change tasks.md verification gates).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Live FLEET_ID=3 bringup surfaced bug #7: per-account CODEX_HOMEs under /tmp/codex-fleet/<acct> trigger Codex CLI first-launch prompts ("Do you trust …", "External agent config detected", "Press enter to continue") that block worker bootstrap before the input box exists. All 8 workers stalled; operator had to click each pane. This commit: - Seeds scripts/codex-fleet/codex-first-launch-supervisor.sh that polls each worker pane and auto-answers the three prompts. Verified working live (6/8 panes drained automatically; remaining 2 need slight backoff tuning). - Expands plan subtask sub-2 (F3) to wire BOTH auto-wake and auto-bypass into full-bringup.sh tail, gated by CODEX_FLEET_AUTO_WAKE / CODEX_FLEET_AUTO_BYPASS env (default 1). Order: auto-bypass runs before auto-wake. - Narrows sub-6 (F7) to ship a smoke test that asserts no panes remain stuck on first-launch prompts within 30s of DONE. - Adds matching acceptance criterion + proposal narrative for F7. The two-pane operator layout (one kitty with the agents, one empty operator shell) is now the default — see PR description. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Live run showed 6/8 panes drained at 1.5s; remaining 2 needed ~9-15s total but only got 7.5s of attempts (5 rounds × 1.5s). 2.5s × 5 = 12.5s window catches slow Codex bootstraps without making fast cases noticeably slower. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Workers' Codex CLI echoes the prior menu text into tool-call history once they reach the worker loop. The supervisor's grep against the full scrollback (-S -100) saw the menu in history and falsely flagged a fully-drained pane as still stuck. Switching to bare `capture-pane -p` (live screen only) eliminates the false positive while still catching live menus. Also add `1` + Enter combo for the External-agent menu — some Codex builds advance on bare digit, others need Enter to confirm; sending both is harmless on already-advanced panes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
5 tasks
NagyVikt
added a commit
that referenced
this pull request
May 18, 2026
Implements all 6+1 dispatch-path fixes from PR #189 (scaffold) with live evidence from the 2026-05-18 fleet runs. F1 dead-pane surfacing · F2 cap-probe TTL hardening · F3 auto-wake at bringup · F4 plan-watcher --allow-waves · F5 force-claim worker-ready gate · F6 auto-submit smoke test · F7 first-launch supervisor wired in (smoke test PASSES live).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Scaffolds the OpenSpec change + 7-subtask plan workspace for
fleet-dispatch-fixes-2026-05-18. Captures dispatch defects surfaced by real telemetry from the 2026-05-18 fleet bringups (recodee + codex-fleetui) plus a NEW finding F7 surfaced during the live FLEET_ID=3 bringup of this very plan.The 7 findings
Pane is dead (signal 15)on 5+ panes; never surfacedcap-probecache outlives quota recoverywake-promptwindow blank on bringupplan-watcherrejectsdepends_onwithout--allow-wavestrading-edge-foundationsdispatched in place of priority planforce-claim"not in a mode" on non-idle panescodex-first-launch-supervisor.sh(already seeded in this branch, auto-drains all three prompts in parallel; 6/8 cleared in 30s)Subtask split (disjoint file_scope, all parallel-ready)
show-fleet.sh,docs/fleet-telemetry-cases.mddoc_workcap-probe.sh,cap-probe-cache.shtest_workfull-bringup.sh(both auto-wake AND auto-bypass at tail)api_workplan-watcher.shfrontend_workforce-claim.shfrontend_worktest/codex-auto-submit-test.shtest_worktest/first-launch-bypass-test.shtest_workThe F7 supervisor script itself (
scripts/codex-fleet/codex-first-launch-supervisor.sh) is already in this branch — proven working live on the FLEET_ID=3 workers. Sub-2 just wires it into the bringup tail.Acceptance gate
After all 7 subtasks land, a clean re-run of
full-bringup.sh --plan-slug fleet-dispatch-fixes-2026-05-18 --n 4 --auto-fleet-id --no-cap-cachemust show:Operator layout
Three kitty windows on operator desktop:
fleet-dispatch-fixes (codex-fleet-3)— the 8 worker overviewfleet-dispatch-fixes ticker (fleet-ticker-3)— force-claim, plan-watcher, cap-swap, wake-promptoperator shell (codex-fleetui)— empty shell at the repo root for ad-hoc commandsTest plan
openspec validate <change-id> --type change --strictpasseslib/plan-validator.sh openspec/plans/fleet-dispatch-fixes-2026-05-18/plan.jsonreturnsok:truescripts/codex-fleet/test/Draft until the fleet has produced its work.
🤖 Generated with Claude Code