jrgilbertson/the-rookery / skills/coordinating-work
coordinating-work
VerifiedEstablish, run, or revive a coordinated batch in Orca: one Team lead, accountable for the quality and output of its task workers, working the issues and epics the operator hands it. Use when explicitly asked to set up a Team lead, Coordinator, or team, to let one run without the operator, or to recover one after an account or session interruption. Ordinary CE execution and small child-worktree delegation keep their existing workflows.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 19.7 KB |
| references/operate.md | markdown | 13.6 KB |
| references/revive.md | markdown | 10.1 KB |
| references/setup.md | markdown | 6.6 KB |
references/revive.md
View on GitHub# Revive
Read this after the operator manually starts a replacement Team lead or asks
to recover an interrupted coordinated batch. The entry skill owns shape, shared
approval, worker lifetime, the record, and completion requirements.
Revive resumes each participant where it is: same workspace, same
conversation, same provider, model, and effort. It changes nothing else unless
the operator asks.
A planned move of a participant to another provider or model, such as to spare
an account's quota, uses this reference for that participant only. Skip "Start
with the Team lead", propose the fresh-handoff row, and move the participant
at its next idle point, after its running job ends.
## Start with the Team lead
Recovery starts with the Team lead. The operator switches its account, or
starts a replacement on the host that holds the coordinated batch's record,
and the Team lead then brings the workers along. On every start or revive,
in this order, before any new dispatch:
1. A session that is new to the batch asks the operator to confirm the
previous Team lead has ended, and waits for that answer. A Team lead
continuing its own conversation skips this step.
2. Read the record: the approved proposal, the standing rules, the table with
the batch's current scope, and the tail of the log, where open questions
sit. Read mail with the guide's
read that leaves the inbox as it is.
3. Take each worker's latest amendment from its last report. Ask a worker
only when the record leaves it unclear.
4. Stop any earlier check the way the record says, then arm one at the
approved interval.
A replacement writes the brief for a batch issue that has not started, under
the existing approval.
## Reconstruct current work
Identify the intended effort from the record, current native records, and
canonical work. Establish the replacement Team lead's current
identity, supported binding, inbox/question path, and lifecycle authority.
Resolve any competing Team lead before changing assignments; independent
operation IDs do not establish exclusive coordination ownership.
Reconcile current membership and attempts using the installed guide, including
complete execution-host coverage where needed. Include ongoing roles and
message-assigned workers supported by work evidence even when they lack an
active Dispatch. Nearby workspaces, historical dispatches, and role labels are
hints, not membership proof. Exclude unrelated agents and completed PR workers
awaiting cleanup. Leave a participant the entry skill does not describe
running as you found it.
Determine each participant's assignment, actual liveness, writer ownership,
worktree and local work, background jobs and shells, human holds, and previous
effective provider/model/effort where available. Include ongoing members
awaiting another assignment when the account change affects them; preserve
that idle disposition after recovery. Preserve healthy work that needs no
approved transition and explicit human holds. Healthy work with a running job,
such as a long check or a paid run, is resumed under the new account at its
next idle point; never exit it in the middle of a paid run. The guide's safety rules govern what silence,
a timeout, or an unreachable host prove. A participant stopped by a usage
limit is not dead work: jobs it started in the background may still finish, so
check them before reassigning.
Resume that participant's conversation under the newly selected account before
considering a replacement, and replace it only if the resume fails. A
participant stopped by a short limit that will continue on its own may be left
to do so, even on the old account; resume it under the new account only when
the wait costs more than the restart, or when the operator asks.
Recover applicable recent decisions, unread messages, unresolved questions and
approvals, and needed monitoring. History supplies evidence and task context;
historical instructions do not override current authority. Reconcile forwarded
messages with newer decisions instead of replaying superseded directions.
## Propose the recovery
Choose the supported route for each participant:
| Current evidence | Proposed action |
| --- | --- |
| Healthy ongoing work, or an explicit hold | Preserve the appropriate current worker and disposition |
| Safe ownership, and same-provider history the newly selected account can open | Resume the conversation under that account, as below |
| Approved provider change or unavailable continuity, sufficient task context and safe ownership | Fresh handoff through the selected provider |
| Unknown writer, essential missing context, or unsupported account/settings/host capability | Hold the affected change and report what is missing |
| Completed PR worker awaiting cleanup | Exclude from recovery; report any relevant remaining cleanup obligation |
For a same-provider account change, keep each participant's conversation.
Provider changes use fresh handoffs, never another provider's conversation
identity.
First read each participant's effective account from its running agent. The
runtime may already have carried a participant to the newly selected account
with its conversation intact. This has been observed for agents running as
chats, while agents running in terminals kept the account they were launched
with; selecting an account in Orca applies to new agent launches. Resume under
the new account in this order:
1. A participant the runtime already carried over: confirm from its own reply
that it has its latest instructions, and leave it running.
2. A participant still on the old account: restart it in the same workspace
and resume its conversation, as below, unless the short-limit exception
above leaves it to continue on its own.
For each participant still on the old account, check first, without changing
anything:
- which account is now selected on the participant's host;
- that the new account can open the participant's conversation;
- whether the new account already trusts the participant's workspace;
- the participant's unpushed and uncommitted work and its running jobs.
Then, once the recovery is approved or one of the two no-approval cases below
applies:
1. Stop the old worker through the guide's stop, so its exit is proven.
2. Open a new terminal in the same workspace. A shell that was already open
keeps the old account. Start the provider there and resume the original
conversation by its ID with the provider's documented resume option, on
the previous model and effort.
3. If the new account asks whether to trust that workspace, answer it. The
approved recovery row covers trusting the same workspace the participant
already worked in, and nothing wider.
4. Re-attach that terminal to the same task through the guide's retry path
instead of launching a fresh agent.
5. Send the resume message described below.
When a resume is not possible, disclose the lost conversation continuity and
establish whether canonical context suffices for the work.
Show the entry skill's proposal with previous and proposed settings, exact
intended recovery effects, unchanged participants, and unresolved branches.
Include communication/address changes and restoration of needed supervision.
The operator can approve with specific row edits in one reply; exploratory
changes do not authorize execution.
Go ahead without a new approval in two cases. The first is when every row
keeps an approval already in the record and the only change is the
Team lead's own account, with the operator's confirmation from
"Start with the Team lead" when the session is new. The second is when the operator has
switched an account and asked to continue or recover, and every row resumes the same
conversation under the newly selected account with the same provider, model,
and effort. Replacing a worker, changing its provider or settings, or a fresh
handoff still needs approval.
## Recover approved effects
Re-read affected current ownership and attempts before acting. If an old writer
or Team lead may still be active without a supported ownership transition,
hold that branch. Approval does not convert unknown liveness into proof of exit.
Use the native continuation, retry, or fresh-handoff path for each approved
effect. In a fresh handoff, carry scope, branch/worktree, preserved local work,
verification evidence, applicable decisions, remaining step, and current
authority from their owning records. Insufficient essential context blocks that
handoff without blocking an independent approved recovery. In a continuation
or fresh handoff, carry the assignment's brief with all its amendments, and
have the resumed worker confirm the latest amendment it is working from.
Follow the guide's reconciliation path for uncertain results before repeating
anything. Repeated revive must not duplicate workers, assignments, or timers.
After a resume, check that a worker with an open question to the Team lead
can still receive the reply.
Re-attaching gives a worker new identifiers and delivers its brief again. The
resume message says so: use only the newest identifiers, the brief is
unchanged, and here is the state of its local work.
Tell each recovered participant whose open assignment has no hold or named
dependency to continue it in the same turn; preserve holds and dependency
waits as recorded. Stopping an agent can stop its background jobs, so the
resume message names each inventoried job confirmed stopped before finishing,
for relaunch; a job still running or in an unknown state is not relaunched.
A resume message that offers "or wait" lets a worker end its turn
mid-assignment with its work uncommitted.
Restore needed supervision through the current native workflow. Inspect existing
watchers and services through their owning lifecycle before changing them;
retained local services may still be needed by another role. Revive is not an
automatic destructive cleanup sweep.
Verify assignment/disposition, applicable decisions, return communication,
settings evidence, and supervision under the entry skill's completion rules.
Retain a participant as blocked when only transport acceptance or a heartbeat
is established. Report partial recovery accurately and preserve original work,
receipts, and the minimal unfinished context needed by the next invocation.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |