v1.1.0 — complete the Orca delegation substrate

[choiyounggi]

Closes the gaps that kept the Orca substrate behind the tmux one. Planned as a four-task orchestration; the run itself then surfaced more defects than the plan did, so those are here too — each with the command and error that proved it.

Wave 1

orca-wait.sh — three defects (134 → 153 tests)

1. Escalation was checked once, before the blocking window. guardrails writes its record from inside the worker whether or not the worker survives to send the matching Orca message, so a worker that died while blocked cost the entire timeout. The wait is now sliced into ORCA_WAIT_RECHECK_MS windows (default 15000, matching watch-status.sh) with the dir re-read between them. The slices sum to exactly the caller's <timeout-ms> — the deadline is neither extended nor truncated.
2. One call consumed one Delivery, so a Wave of N workers could cost N coordinator turns. --until-all <ids> consumes successful batches until every listed id is completed. 3/5/6 still return immediately without acking.
3. exit 0 meant "a batch arrived", not "a completion arrived". --types asks for worker_done/escalation/question but heartbeats are delivered anyway, and so are completions from earlier Waves or unrelated Tasks — all fell through to "everything in this batch was a successful completion", acked, and returned 0. Measured during this run: a batch of one previous-Wave plan completion + one probe completion + a heartbeat returned exit 0 while all three current tasks were still running; only the completed=0/3 line revealed it. Exit 0 now requires a successful worker_done for a task id the caller passed, using the same exact-match index semantics as the progress count.

orca-worker-start.sh — no second agent on one checkout (153 → 166)

Re-entry restarts a "dead" worker with --worktree/--agent, which for a dev-loop worker always lands in worker mode and used to run terminal create unconditionally. If the liveness read was wrong — and this run proved a worker can be ALIVE yet wedged — that puts two agents on one working tree with the Dispatch bound to one of them. Worker mode now joins dispatch-show against terminal list and rebinds the live assignee. Unreadable probe → exit 6, creates nothing; only ok:true + dispatch:null counts as a first start. Field names (orphaned, connected, assignee_handle) verified against the live runtime.

session-prompt.md — §O1–§O4

The file was four tmux one-liners, so the Orca protocol was spliced into every --spec by hand. Two of the four also ended with "wait for an approval message", correct only for a live pane. Adds the Orca set plus a shared protocol block; the four tmux prompts are byte-identical.

Wave 2

orca-worker-stalled.sh — the third worker state (166 → 183)

ALIVE/DEAD cannot express what stopped this orchestration: three workers wedged on an interactive usage-limit prompt held live PTYs for 75 minutes with byte-identical diffs while orca-worktree-alive.sh said ALIVE, Tasks read dispatched, worker-read said worker=ready terminal=running, and orca-wait.sh checkpointed forever. Only reading a worker's raw screen found it.

The signal was chosen by measurement and one candidate was rejected: lastOutputAt tracks terminal writes, and a TUI agent repaints its spinner continuously, so a wedged worker reads as output 0 seconds ago — against the three real worktrees it gave −1s, −1s, 340s. agents[].updatedAt discriminated cleanly (12s / 64s / 340s). A test pins the rejected case so the trap cannot come back.

SKILL.md — five instructions this run proved wrong

	Was	Actually
--deps	"chain with --deps"	--deps task_x errors (wants a JSON array); even the array form gave task_not_startable with the dep already completed — isolated by dispatching an identical no-deps Task to the same terminal, which started fine
failed start	"do not auto-retry"	the Task is spent (status=failed); a retry needs a new Task
send --type	unspecified	note/guidance/message/info are all invalid_argument; omit --type
--to dispatch:	"attempt-specific coordinator guidance"	a Claude worker never polls its mailbox and a new Dispatch cannot land while one runs — there is no mid-flight steering; the rework round is the path
Wave task ids	progress reporting	correctness — it is what separates your completions from heartbeats and other Waves

Verification

• 183/183 pass, 0 failures
• All shell scripts pass sh -n; new tests invoke the scripts with sh, not bash, so the #!/bin/sh shebang is actually exercised (a bash "$script" suite cannot detect the bashisms that shebang exists to prevent)
• Version bumped in both .claude-plugin/plugin.json and .claude-plugin/marketplace.json

Tag v1.1.0 after merge.

🤖 Generated with Claude Code