Same failure modes here: agents editing past the ticket, reviewers LGTM-ing mocked tests, and "one more retry" loops. What worked for me was one supervised chain (spec → tickets → implement in isolated worktrees) with merge only after gates, not after debate. Rough notes: https://t.co/ZLNzgG7APo
@saadkamal The "who's stuck" problem is real. One thing that helped me: treat ready ≠ turn started, and only advance when a gate says so (tests/diff/scope)—otherwise park and ask the human. Wrapped that as a small Orca skill if useful: https://t.co/ujoqQpzBg8
Yeah — isolation is the easy half. What kept biting me was "ready" looking like progress when the turn never started, plus scope creeping past the ticket. I've been running Matt's skill chain under Orca with one worktree per step and hard gates before merge; happy to compare notes with Campfire's vote/replay model. https://t.co/ZLNzgG7APo
Same failure modes here: agents editing past the ticket, reviewers LGTM-ing mocked tests, and "one more retry" loops. What worked for me was one supervised chain (spec → tickets → implement in isolated worktrees) with merge only after gates, not after debate. Rough notes: https://t.co/ZLNzgG7APo
Yeah — last-writer-wins isn't a merge strategy. What helped me: each agent in its own worktree with a hard path scope, and a parent that owns merge and only starts the next turn after preflight (not on main, scope still held).
Wrote the failure modes + gates here if useful: https://t.co/r4sqKFxOnT
@shugarDadddy@luccacerf Same worry. Overnight only works if they can't touch main, can't widen scope, and can't start the next turn just because they're "ready". Gates first, sleep second.
Short checklist: https://t.co/XVCL1sSjIs
@jaklimoff 3 is a solid default. Same loop here — reviewer keeps finding "one more" that was already known. Cap the rounds, then force a rethink instead of another retry. Also don't let "ready" auto-start the next turn.
Patterns we use: https://t.co/XVCL1sSjIs
@froessell Hash-before-copy is the right complement — ownership stops the race, the hash catches the silent overwrite. Glad the project’s in git now; that alone would have made the second wipe obvious in the diff.
@ojas_dev Agree — that preflight is the cheap version of the same rule. I treat "session up" as incomplete until `git rev-parse --show-toplevel` matches the ticket's worktree path; otherwise the coordinator should refuse edits and re-dispatch. Worth baking into the gate before any write.
Worktrees only isolate writes. Once two sessions land on the same target, you still need a merge gate: freeze merge to main when paths overlap or review can't hand back a replayable check — don't let the main agent "summarize then merge" itself.
I wrote that pattern up as an Orca skill (worktree per ticket + gates): https://t.co/r4sqKFxOnT
What's worked for me: declare each ticket's file/area scope up front and refuse to dispatch two tickets that overlap. Shared config/lockfile changes go to a single serial ticket instead of being done in parallel. And if a worker asks to widen its scope mid-run, the orchestrator escalates to a human rather than auto-accepting. Notes here: https://t.co/r4sqKFxOnT
Dispatch is easy; the last 20% is the hard part — supervised Orca orchestration skills.
• new terminal ≠ worktree → `--worktree new-top-level`
• ready ≠ turn started → wait on turnStart evidence
• scope / credentials / irreversible → human
Video: https://t.co/eHx2N98Kgq
`npx skills add whutlichao/orchestra-auto --skill orchestra-auto`
https://t.co/r4sqKFxOnT