Failure modes I kept hitting with supervised Orca workers — and the skill that names the counters

[whutlichao]

Dispatch is the easy part. The hard part of supervised multi-agent work on a real repo is the last 20%. Here are failure modes I kept hitting with Orca orchestration — and the counters I finally wrote down.

Failure	Counter
Orchestrator edits main — a new terminal is not a new top-level worktree	Force --worktree new-top-level per ticket; coordinator never edits or commits in a worker worktree or on the main working copy
state: ready ≠ turn started — the receipt only means the bytes were written	Require per-agent start evidence (or turnStart: observed) before waiting; if it did not start, recover — do not resend the whole prompt
Settled mail, nobody handled — an inbox that is only a pile of files drops messages	Persist batches atomically by delivery ID, dedupe, ack only after the whole batch is handled; auto-ack heartbeat-only batches
Worker suggests a scope change, the loop accepts — “also delete that unused component?” is a scope change	Classify first: answer details yourself; send scope / approach forks / credentials / payment / irreversible live-data ops to a human
Huge diff is formatter reflow — nearly lost the real work with a revert	Keep evidence before judging an unusual diff
Same gate fails, infinite retry	Bounded retries plus per-ticket and run-level breakers

I published that procedure as two installable Agent Skills (English + 简体中文, same semantics). Orca-specific — they drive orca orchestration … and are not a general orchestration framework.

The entrypoint keeps the routing table and hard rules; dispatch-loop / reply-protocol detail lives under references/ and loads on demand (less always-on activation context — no A/B claims).

npx skills add whutlichao/orchestra-auto --skill orchestra-auto
npx skills add whutlichao/orchestra-auto --skill orchestra-auto-zh

Repo: https://github.com/whutlichao/orchestra-auto

Happy to be corrected on any of the rules — I'd rather fix the procedure than keep rediscovering the same failures.

[whutlichao]

Follow-up note: the supervised / Agent Skills framing in this write-up was inspired by Matt Pocock’s ideas on that topic (mattpocock/skills); orchestra-auto is my port of those process judgements onto Orca’s orca orchestration verbs. Not an endorsement by him — any mistakes in the rules are mine. Acknowledgments are also in the repo README after PR #6.

[capitaharlock]

With ack deferred until the whole batch is handled, how do you recover if an action finishes but the process dies before durably recording completion? The batch is still unacknowledged, but that action has already happened.