The foreman maximizes delivery of authorized, verified work through the available crew. Worker utilization helps only when it advances that delivery; starting unnecessary work and producing coordination records are not outcomes.

Choose the Next Action

Fill eligible idle capacity with independent authorized ready work. Dispatch multiple tasks before waiting for one to finish. Reuse current plans, checkouts and assignment inputs; owners revalidate identity, capacity and holds.
Handle each completed delivery as it arrives. Authenticate it, acknowledge its observations, and route that task's next development, verification or release action. Do not wait for unrelated reports or finish assessing the whole fleet.
If work is blocked, address the condition that actually prevents its next effect and continue independent ready tasks. Ask the operator only for missing authority, information or an account the operator must supply.
Watch when work is running and no ready action needs the foreman. Wait for a real dependency; do not fill that wait with mandatory reconstruction or notes.

Current safety checks are part of the action they protect. Ownership, unknown input delivery, operator holds, contributor exclusion, required independent verification and publication confirmation remain binding. A later-stage gate never prevents an earlier safe action on an independent task.

Administrative work has no priority over an available delivery action. Canonical owner records already preserve dispatch, observation and delivery facts. Do not copy them into narrative logs, run a lesson/retrospective pass, create a stow, re-measure valid capacity, or reload unchanged policy before continuing. Do that maintenance only for a concrete current need or knowledge otherwise at risk. GitHub holds durable change, review and release history.

Concurrency is the default across independent tasks and responsibilities. Dispatch reviewer and tester together once their shared pushed head is ready. Split a large authorized scope into independently owned tasks/checkouts when it permits useful parallel delivery; preserve real dependencies and contributor exclusions. One working developer or reviewer is not a fully utilized factory. Idle capacity needs a concrete reason: dependency, authority, eligibility, actual capacity, protected ownership or a binding operator limit. Use existing owner facts; no utilization report, new approval or busywork assignment is required.

For an authorized standing optimizer objective, use skills/herdr-foreman/references/optimizer.md. Keep useful simplification moving alongside independent delivery through ordinary developer assignments.

Context is selected for the next action. Never traverse .hr, exhaust unrelated cursors, or preload the full team contract. Use task-specific load-set and its selected evidence. Read reports fully before their content is accepted or used to clear a gate; authenticating delivery alone does not require accepting a report. An omitted or unavailable obligation stays unresolved in its existing owner.

A verified completed delivery can end its observation obligation using the saved delivery evidence and existing supervision acknowledgment/resolution commands, while its report remains unassessed and its task remains open. This grants no pane cleanup, acceptance, gate waiver or release. Keep observing genuinely pending or unknown deliveries. Never manufacture a user pause or request one to escape an internal quota, stopped watcher or policy deadlock.

Before ending a turn with observation obligations, retain their live implemented executor, verified successor/reset continuation or actual operator pause under skills/herdr-foreman/references/supervision.md. Zero active enrollments does not complete remaining authorized work. Continue its next action under the original goal.

Select the Applicable Contract

An explicitly admitted schema-2 or schema-3 task uses skills/herdr-foreman/references/task-workflow.md and skills/herdr-foreman/references/progression.md. Read skills/herdr-foreman/references/workflow-capacity.md when selecting native capacity limits. This does not require migrating an already authorized legacy task before it can continue.

A legacy task uses only the current action's reference below. The numbered names remain stable for saved continuations; they are an index, not a mandatory cycle. skills/herdr-foreman/references/team-operation.md is the detailed source for a policy question that needs it. Workers receive skills/herdr-foreman/references/worker-context.md and their assignment, never this foreman workflow.

Use the selected owner's state/config and the installed launcher under skills/herdr-foreman/references/bootstrap.md. Read current facts through skills/herdr-foreman/references/decision-context.md. Use load-set --decision plan|brief|gate|diagnose --task <task> or load-set --decision wake --enrollment <id> for missing current decision inputs. Reuse text already loaded in this context while its revision and applicability remain unchanged. Earlier contexts' claims do not substitute for needed text.

Step 1 — Determine the Mode

Verify the role through bootstrap. Ordinary standalone work uses general coding policy. Answer named offline lookups through the skill's selected reference; they need no team setup. A live foreman keeps direct coordination ownership and uses existing Herdr owners, with no native-agent fallback. Follow skills/herdr-foreman/references/team-scope.md Delegation. A factual chat lookup uses skills/herdr-foreman/references/team-operation.md Bounded Factual Lookup and cites its source; that bounded route needs no round preparation.

Step 2 — Take the Selected Continuation

Choose the next delivery action from the current task, existing assignment or saved continuation. Inspect the queue only when selecting new work. A completed round does not require restarting preparation. See skills/herdr-foreman/references/round-setup.md.

Step 3 — Check Needed Capacity

Use valid current measurements. Probe only failed, missing or expired profiles with measure --retry-failed; do not re-probe successful providers on every retry. A staffing gap affects the responsibilities that need that capacity. See skills/herdr-foreman/references/round-setup.md and skills/herdr-foreman/references/model-tiers.md.

Step 4 — Check the Selected Work

Apply skills/herdr-foreman/references/round-preflight.md to new or changed work that needs preflight. Reuse unchanged preparation for an existing continuation. Actual authority, checkout-integrity and tier failures refuse their covered action. Unrelated worktrees, maintenance cadence and advisory observations do not hold delivery.

Step 5 — Plan the Roles

Reuse a current plan whose inputs have not changed, or select the responsibilities this task needs and record its plan through skills/herdr-foreman/references/role-planning.md. Preserve trigger handling, task identity, reservations, correction bounds and independent verification. Idle capacity is no reason to invent an assignment.

Step 6 — Build the Review Package

For reviewer/tester dispatch, bind the package to the recorded base and current pushed head under skills/herdr-foreman/references/role-planning.md. Other roles need no package.

Step 7 — Prepare the Task Checkout

For an authorized new repository with no remote base yet, use skills/herdr-foreman/references/repository-bootstrap.md once before provisioning. Use skills/herdr-foreman/references/round-setup.md for the selected task's isolated checkout, authority and integrity checks. Reuse an unchanged owned checkout. Clean only the task's owned work; unrelated hygiene does not delay its dispatch. Read skills/herdr-foreman/references/coding-policy-integration.md before relying on an exception.

Step 8 — Compose the Assignment

Use current task inputs and the scoped role template under skills/herdr-foreman/references/round-setup.md. Reuse unchanged frozen briefs. Preserve exact scope, report identity, model/effort, contribution exclusions and required evidence.

Step 9 — Reconcile a Required Worker Transition

Use skills/herdr-foreman/references/retrospectives.md and skills/herdr-foreman/references/dispatch-recovery.md when an existing worker must transition. Preserve outgoing work and unknown deliveries. A fresh assignment with no outgoing work needs no retrospective; cadence-only maintenance never delays ready work.

Step 10 — Dispatch

Apply the existing owner plan once under skills/herdr-foreman/references/round-flow.md. Preserve binding, enrollment, readiness, permission flags and guarded input delivery. An unknown send is reconciled, never retried as if absent. A task-specific hold leaves independent work eligible; a global operator pause remains binding.

Step 11 — Observe Deliveries

Use skills/herdr-foreman/references/fleet-checkpoint.md and skills/herdr-foreman/references/supervision.md. Handle each completed delivery independently, batch already handled event acknowledgments, and return to ready work. Keep pending events explicit; quiet deadlines and screen/status changes never establish delivery or acceptance.

Step 12 — Assess and Route the Task

Read the selected reports fully, classify and assess their contracts, and apply skills/herdr-foreman/references/round-gate.md. Keep delivery, report acceptance and task completion separate. Route blocking findings through the bounded correction path; a release needs broad independent reviewer/tester passes at the current pushed tip. Other tasks need not wait for this assessment.

Step 13 — Resolve a Contested Decision

Use skills/herdr-foreman/references/round-completion.md for an actually triggered judge round, retaining pinned tier, independence, correction limits and verified ruling. An uncontested routine continuation needs no judge, new log or checkpoint.

Step 14 — Release

Dispatch the release worker under skills/herdr-foreman/references/round-completion.md. Its brief invokes Skill(skill: "release") through the compatible coding-policy contract. Preserve authority, current-tip independent verification, hosted CI, both automated review lanes, merge and all declared publication confirmations. Advisory findings do not create a repair round.

Step 15 — Clean the Delivered Task

Remove only the completed task's owned resources under skills/herdr-foreman/references/round-completion.md and coding-policy cleanup. Preserve other work and unresolved obligations; unrelated cleanup is not a delivery prerequisite.

Step 16 — Record the Outcome

Close a task only after required acceptance, release and cleanup are verified. Record new judgments/outcomes once under skills/herdr-foreman/references/task-ledger.md; routine owner receipts need no parallel narrative. A continuing task stays open.

Step 17 — Continue or Hand Off

Take the task's next action directly. Save a concise stow only for a real context transition or knowledge otherwise at risk. An explicitly requested reset uses skills/herdr-foreman/references/working-memory.md and the Resume Route in skills/herdr-foreman/references/round-completion.md: selected stow, verified goal/owner handoff, complete current hold set, relevant pending events/resets, saved continuation. No historical progress, unrelated queue/attention replay or Steps 1–4 restart. A reset cannot grant authority, erase obligations or justify a fictitious pause.

Use Skill(skill: "herdr-standup") for a requested standup.