Explicitly admitted contracts share the existing task authority owner. Schema 1 runs a fixed prepared assignment chain. Schema 2 executes a bounded or complex task through native implementation, independent verification, correction, release and local cleanup. Schema 3 adopts an existing source assignment into that executor without resetting its history. Read skills/herdr-foreman/references/task-workflow.md before using schema 2. Installing this package starts neither contract and migrates no live team. Comparative cost, throughput and defect evidence still requires the real pilot tracked by herdr-policy epic #1; source fixtures prove no ROI improvement.

Source, Tessl and realized ACR native installations can identify their actual policy bytes. Native ACR startup verifies the public lock/realization ledger and requires the installed owned bytes to match the checked target hashes. skills/herdr-foreman/policy-paths.py identity, invoked with explicit Python 3 from the selected skill, is read-only and grants no dispatch authority. A missing, changed or ambiguous installation refuses startup. Remote/attested task execution is not enabled by these adapters. The native reader supports per-adapter owned skill targets in ledger schemas 1 and 2. Coordinator-owned targets may coexist in the ledger; execution from the shared .agents/skills layout is not enabled by this adapter. ACR's full public realization check can fetch locked dependencies into temporary storage. Its metadata cache is process-local, so fresh native Stop readers also need verification access; a failed check refuses continuation. Source/Tessl identity checks use their local installed bytes.

Admit one task

Run skills/herdr-foreman/foreman.sh through the selected installation's resolved skill root with explicit Bash. Only the actual bound native foreman or verified remote owner may hand off a registered, open and settled task. Admission refuses pending sends, active enrollments, another service grant and operator holds. Schema-3 admission may transfer only the original authenticated completed source enrollment; other active enrollments and unknown effects still refuse admission. It never changes an existing binding, resumes a hold or starts a worker.

progression-admit --state OWNER --record FILE takes an object with task, expected_generation, authorization, service_id, absolute lease_path and absolute contract. Authorization carries source and quote. Read the actual generation from the task owner; an unadmitted legacy task starts at zero. Use a dedicated lease distinct from state files and transaction locks. The result names the new generation; admission alone authorizes no dispatch. The state owner's ordinary read upgrades older grant records without changing their generation or authority. An old service grant remains a prepared chain; its saved document cannot promote it to schema-2 task execution. Read-only previews leave old records untouched and name the required owner migration.

Prepare a schema-1 chain

The immutable contract contains schema_version: 1, its exact task and steps. Each step contains schema_version: 1, unique id, argv, after, gate and inputs. argv is one prepared apply invocation with that task, an explicit config, assignment plan, common/role briefs and fresh reports. inputs contains the owner's {path,sha256} receipts for every config, plan, common, role brief and correction-work input. Keep source inputs available for replay. No report or GitHub-history copy belongs in this contract.

after names earlier steps. gate is delivered or assessed; delivery uses the canonical native checkpoint and exact enrolled identity. Assessment also requires the existing enrollment's accepted resolution. A closed enrollment uses its captured ledger prefix, current canonical accepted assessment, report contract and native final source; it does not require the removed pane. Missing, changed or contradictory evidence wakes the foreman. Neither gate supplies acceptance of the task. Supported responsibilities, budgets, refusal predicates and parser contracts live in skills/herdr-foreman/foreman/progression_service.py. Closed dependencies require clear classifier gates. The actual owner apply rechecks these gates under their existing lock and keeps that lock through dispatch. Evidence failures return a bounded refusal naming the failed stage and the required recovery action. The admitted CLI clones the native transport with a supervision-owner boundary. Each invocation checks the live binding and unresumed holds under the same lock used by hold/rebind writers and keeps that lock through transport. Enrollment transactions run outside that short boundary. A hold or rebind that wins the lock prevents transport; an in-progress invocation prevents their commit.

The owner captures verified config/plan/work bytes for apply and checks the frozen brief bytes before worker effects. A mutable replacement cannot change the admitted assignment. This adapter refuses arbitrary commands and release roles; use the existing verified release owner for that boundary.

Prepare a schema-2 task

The immutable object contains exactly schema_version: 2, task, risk, repository, config, common, briefs, artifact_directory, effects, publication and acceptance. skills/herdr-foreman/foreman/workflow_contract.py owns validation and bounds; unsupported or missing facts fail before worker launch.

risk supplies every schema-1 fact accepted by skills/herdr-foreman/foreman/workflow_profiles.py, without defaults. Bounded work combines independent review and testing; complex work uses separate identities. Sensitive work and required specialist coverage currently create an explicit pre-implementation exception. Do not misclassify the work to avoid that boundary.

repository names the exact GitHub owner/repo, absolute additional worktree, branch and base_branch. Admission requires that clean writer at the task's registered base, its actual origin and a distinct base checkout. Existing implementation is not adopted as a new task.

config, common and each briefs entry are {path,sha256} receipts of nonempty UTF-8 files. The config supplies assignment-scoped worker kinds with tier tables. Use the co-shipped common and selected role inputs:

skills/herdr-foreman/templates/workflow-common.md
skills/herdr-foreman/templates/workflow-developer.md
skills/herdr-foreman/templates/workflow-reviewer.md
skills/herdr-foreman/templates/workflow-tester.md
skills/herdr-foreman/templates/workflow-release.md

A bounded task omits the tester role. The executor appends actual scope, policy paths, native report paths and current evidence; templates cannot substitute for those owner-derived facts.

artifact_directory is a fresh, owner-controlled absolute directory outside the writer, with an existing parent. Admission reserves it by exclusive creation; an existing directory, including an empty one, is preserved and refused. Keep frozen templates outside this new directory. It holds current plan/common inputs, role briefs and native reports plus narrowly named delivery/review artifacts. It is not a history directory. Workers consume the specific inputs named in their brief, never a recursive .hr or owner-state dump. GitHub remains the change and review history.

After verified acceptance, selected publications and Git teardown, delivery schema 2 retires the private directory through the existing cleanup intent. It moves the directory and each file into its exclusively owned quarantine before checking the captured inode and bytes, then deleting. The bounded manifest holds at most 256 identities; each batch holds at most 32 receipts. Unknown entries, changed evidence, links, open report gates and unresolved enrollments preserve the source or quarantine. Restart resumes the same intent. The .herdr-retire- namespace belongs exclusively to this lifecycle. Owner-proved unsent attempts retire only their byte-bound unused briefs, with no active enrollment or report; their canonical no-send evidence remains.

The primary owner migrates delivery 1 on read; run the owner state command before a read-only preview or restarted executor. Already completed history is marked legacy_preserved without a directory sweep. Older delivery-1 readers refuse the new lifecycle. Keep an owner that reads delivery 2 for in-flight cleanup. Canonical receipts and the GitHub closure reference retain completed history; the CLI returns a compact lifecycle view, never its file manifest.

effects is a unique list drawn from write, push, open_pr, merge and publish. All four source effects are required; publish matches the channels. publication contains tessl (null or {workspace,plugin}), acr (boolean) and workflow (null without publication, otherwise the exact merge-triggered publisher). acceptance contains kind (merge, deployment or operator) and one to sixteen criteria. Only merge acceptance is automated currently; other kinds remain explicit exceptions until their actual evidence owner acts.

Adopt existing delivered source

Prepare the ordinary schema-2 declaration with the existing task, original base, real risk facts and effects. Preserve complex and specialist profiles. progression-prepare-adoption --state OWNER --contract FILE --dispatch ID --output ABSOLUTE_FILE authenticates the latest applied developer enrollment, the original frozen prompt, exact native completed turn, report and current clean pushed Git head. It exclusively writes a new immutable schema-3 contract; it changes no task, gate, enrollment, launch or authorization. Missing native provenance, an unfinished source, another active enrollment, unpushed/changed head or unknown effect stays with its actual owner.

Pass that contract to the existing progression-admit handoff. It preserves the task's original base, source assignments, contributor exclusions, correction counter, approach/allowance and report gates. A reported source receives one versioned adopt intent for its original enrollment. The executor verifies its provenance again and closes only that delivery obligation as transferred_for_verification. It creates no developer WORKFLOW claim or accepted assessment. Independent native verification reads the original report in full and supplies the substantive assessment. Already accepted closed source keeps its original resolution. Specialist/production boundaries remain explicit task-local exceptions and do not turn into bounded profiles.

NEW_OWNER_WRITES_ENABLED in skills/herdr-foreman/foreman/release_activation.py remains false in the reader release. Publish and verify the compatible readers through Tessl and ACR first, then enable new admissions in the activation release. No waiting period or new operator approval is added. The reader release supports recovery of an existing schema-3 grant regardless of this admission flag. Older exact contract/grant readers refuse version 3; they cannot silently reinterpret it as fresh work.

Collect and advance a current page

collect-deliveries --state OWNER [--task TASK] [--limit N] [--offset N] collects bounded current enrolled native completions independently of report assessment. Each canonical once-check and original source proof runs outside the primary owner transaction. Changed observations commit through the existing supervision owner; duplicate completions append no event. The result reports collected, waiting and task-local decisions with explicit pagination coverage. Collection never accepts the report or infers completion from an idle pane or file alone. Existing report/classifier gates remain for their actual owners.

progression-batch --state OWNER [--task TASK] [--limit N] [--offset N] --once drives a bounded current page of admitted task executors through their next routine actions. Independent task preparation runs concurrently. Git inspection, planning, rendering and native startup/composer waits happen outside the shared primary transaction; short checkpoints reload canonical state before saving. Every task retains its original kernel lease, single writer and pending intent. One held decision or invalid preparation never discards another task's ready work. Busy owner transactions stay observational and retry after the concurrent preparation wave; no lock is removed or native send repeated. The current result names actual starts/waits/decisions and pagination. It creates no batch state or additional report archive. Omitting --once repeats current cycles through the command's retained execution handle; a page is never fleet-wide coverage. Use the existing individual progression-run lifetime when delegating continuous task supervision at Stop.

Run and observe

progression-run --state OWNER --task TASK --generation N --service-id ID claims the existing grant and dedicated kernel lease. --once performs one tick. The native adapter uses the actual inherited Herdr connection; never construct HERDR_ENV, a pane or a session identity from saved records. The actual HERDR_PANE_ID, live native foreman session and current working directory must match the bound foreman. Startup, task writes, liveness and native transport verify that caller; another managed pane has no authority from knowing the task, generation or service ID. Schema-2 admission also requires this native caller; verified remote admission remains schema 1 only. Retain the process's actual execution handle; a background command submission alone proves neither a running executor nor continuation at Stop.

The loop sleeps without a model relay, checks dependencies outside the shared transaction lock and writes only a current intent or cursor transition. Unchanged waiting polls write nothing. Results contain wake_agent and a bounded data packet. Waiting stays observational. Exceptions wake the foreman with the task, generation, current intent and refusal; no count or timeout permits self-approval. Current evidence retrieval follows skills/herdr-foreman/references/working-memory.md. The process returns this packet on its retained execution handle; it supplies no generic native asynchronous message service. Keep that handle available for the caller's completion/exception delivery rather than relying on a PID file. Heartbeat transaction contention retries within the existing liveness grace; authority failures still stop immediately. Persistent contention loses health instead of authorizing takeover or repeating an effect. Foreground transaction contention returns an observational waiting result; the next tick reads the original intent and canonical readback before effects.

Recover and return

An applied operation lost before cursor commit replays the same owner receipt without another send. Missing, partial or uncertain readback preserves the pending intent and stops progression. A dead process or free lease proves no effect outcome. Reconcile through the existing dispatch/native owners; the initial adapter does not automatically clear an ambiguous service intent.

After the effect owner supplies its operation-specific canonical readback and settles its enrollments or retained probes, the verified bound foreman may run progression-reconcile --state OWNER --record FILE. The record contains task, actual expected_generation, exact intent, explicit authorization and, for measurement, a schema-1 readback with kind and canonical sha256. The service lease must be free. This clears only that pending intent, leaves the schema-1 cursor unchanged and records one owner reconciliation receipt. It does not dispatch, resume a hold or accept work. Missing, uncertain or partly applied sets remain pending. The same completed reconciliation replays without a write. Associated transport-retry identities also refuse this initial reconciliation adapter. The original no-send row never proves a later attempt had no effect. Pre-reservation refusals without durable dispatch proof remain pending for owner recovery; process exit or a refusal message alone does not clear them.

For schema 2, reconciliation accepts only the exact pending launch after every named dispatch is owner-proved not_sent, all task enrollments are settled and no report exists. It retires the intent and current plan, preserving the task, generation, original base and correction counter. Subsequent execution selects a fresh native identity; the old no-send row remains canonical evidence.

New measurement reconciliation requires Shared Writer Activation; it is disabled in the 0.19.0 compatibility release. Existing reconciliation receipts remain readable and replayable without that activation.

A schema-2 measurement uses intent_not_started, snapshot_completed or probes_closed_incomplete. The first hashes the durable intent and requires effects_started: false. A completed snapshot must hash to exactly one state-owner snapshot and match the intent's saved --now, current native measurement context, captured config and complete configured worker set. A probe readback must hash to one canonical closed startup/cleanup row, match the captured config and saved timestamp, leave no pending or live disposable probe, and records only an incomplete measurement outcome. Missing processes, dispatches, reports and timeouts are not readback. Legacy started measurements without a saved timestamp can use only one unambiguous exact probe/config receipt; a later snapshot is never adopted. Both completed-snapshot and closed-probe readbacks require unambiguous live native inventory without disposable probes. Cleanup can fail before its recovery row is retained; a snapshot or absent row never proves that native owner is gone.

Use --readback-current on the existing progression-reconcile command to derive the measurement evidence from its owners. Its record needs only the task, expected generation, pending intent and existing authorization. The owner selects one exact snapshot or retained probe attempt and records its actual digest; callers need no ledger dump or handwritten hash. Ambiguous or missing evidence leaves the intent pending. Repeating a settled request reads its existing receipt. Policy-change refusals name the expected and observed version/digest in the current exception, without loading policy history.

progression-resume --state OWNER --record FILE retries a restored schema-2 observation or cleanup exception. The record contains task, actual expected_generation, the current packet's exception_digest as exception, and explicit authorization. The bound foreman, admitted policy/contract and binding must still match, the lease must be free and operator holds must be resumed through their own owner first. Resume restores the packet's exact saved phase; it does not reset jobs, counters, confirmed channels or pending effects. An already applied dispatch or closed enrollment is read back; an uncertain native effect refuses resume. Cleanup uses its fixed, idempotent readback. Disputes, changed scope, exhausted corrections, specialist work and missing acceptance are decisions, not generic retries. The exact retryable packet codes are owned by skills/herdr-foreman/foreman/workflow_recovery.py.

A healthy schema-2 executor can cover only its exact current task/member obligations at Stop. Its current supervision row must match the live process, kernel lock owner, admitted generation, contract, policy and binding. Pending unknown effects and exceptions cover nothing. Schema-1 chains never gain Stop authority from a heartbeat. No continuation check reconstructs a legacy round or reads its historical contracts to decide whether it can cover work.

The bound foreman may record current exception assessments through the normal owners without taking the service's dispatch lease. Returning a task requires settled dispatches/enrollments, no pending service intent and a free lease. progression-return --state OWNER --record FILE takes task, actual expected_generation and explicit authorization; it increments generation and clears service capabilities. Do not edit owner files, reset a cursor, delete a lease, duplicate dispatch or infer a successful outcome to recover. After return, a new schema-2 admission binds the actual loaded policy and a fresh artifact directory. It increments authority generation and preserves the existing task record, base, assignment/snapshot/event history and cumulative correction value; the original first-measurement counter remains zero rather than being reclassified as an attempt.