This skill is the track-level entry point. It composes the deterministic
coordinator with $curriculum-preparation and $track-completion; it does not
reimplement readiness, preparation, building, certification, review,
integration, or publication policy. Derive active tracks and order only from
curriculum/l2-uk-en/curriculum.yaml.
For a normal BIO request, this is one serial workflow: the coordinator owns
manifest order, prerequisites, acquisition, deterministic resume, and the
hash-bound receipt; the acquired module's $track-completion ledger is the
only bounded-completion authority.
-
Satisfy the repository issue, stream, worktree, research-classification,
pending-decision, and quota/health preflight before mutation. Learner changes
belong to the target track's stream epic, not the shared infrastructure epic.
-
Require an explicit operator-selected terminal goal for every new run. If it
is absent, ask once before starting; never infer it from scope or history.
Resume uses the immutable goal in the exact recorded run.
-
For a new run, map the requested selector to one of all, built,
unbuilt, stale, or one, then start exactly one manifest-backed run:
.venv/bin/python scripts/orchestration/curriculum_coordinator.py start \
--track <track> --owner <agent/task> --scope <scope> \
--terminal-goal <merge|certify|deploy> \
[--module <slug>] [--start <slug-or-position>] \
[--end <slug-or-position>] [--wave-size <n>]
Preserve the returned run_id and ledger path. Do not create a replacement
run to bypass a lease, health pause, or manifest drift.
The goal is immutable authority, not a progress label. A legacy run without
one must use migrate-terminal-goal on its exact run id; never infer intent
from complete or PBR_PASS_QG_PENDING history.
-
Resume only the exact recorded run:
.venv/bin/python scripts/orchestration/curriculum_coordinator.py resume \
--run-id <run-id> --owner <agent/task>
If manifest authority changed, stop and use the coordinator's explicit
adjudication path before starting a fresh run. Never migrate old state by
hand.
The coordinator status includes this exact resume_command; use it rather
than reconstructing or replacing the run.
-
Acquire the next prerequisite-safe module. The coordinator pauses without
mutation when its fresh health/capability gate is unavailable:
.venv/bin/python scripts/orchestration/curriculum_coordinator.py acquire-next \
--run-id <run-id> --owner <agent/task>
-
Read the canonical readiness routing snapshot from the latest matching
MODULE_ACQUIRED event in the returned ledger. Bind it to that event's
track, slug, and attempt. Route its exact next_action uniformly for CORE,
seminar, and BIO targets:
next_action | Exact next owner |
|---|
plan | $curriculum-preparation |
prepare | Requirement- and identity-sensitive; resolve from the full result |
build | $track-completion |
certify | $track-completion |
stop | State- and owner-sensitive; reviewed HOLD or partial recovery |
For plan or prepare, run the canonical evaluator once for the exact
acquired target and use its full typed result to resolve ownership. If any
current failed requirement is owned by plan or preparation, route those
cells through $curriculum-preparation; accept its returned typed result
without calling lifecycle recursively.
When every current requirement passes but a built result remains prepare
solely because of PREPARATION_IDENTITY_MISSING or
PREPARATION_IDENTITY_DRIFT, hand the exact target to $track-completion.
After that skill starts or resumes its authoritative ledger, it must run the
documented request-preparation-rebuild <track/slug> --run-id <id>
transition before any post-build review. That transition reruns readiness
with the consumed identity from BUILD_RECORDED, then either returns
BUILD_REQUIRED or records the exact preparation blocker. Do not reacquire,
repeat the evaluator, or loop through preparation. Fail closed on mixed or
unknown combinations.
For stop, run the canonical evaluator once to validate the full typed
result. PREPARATION_HOLD_ACTIVE without a partial-bundle finding is a
terminal reviewed HOLD: preserve its evidence, leave the acquisition in
place, report the hold, and do not record module completion or reacquire. A
result whose state is partial-bundle, contains PARTIAL_LEARNER_BUNDLE
owned by built_artifact, and has no active HOLD goes once to
$track-completion for its existing forensic recovery. Do not reacquire.
Reject a result containing both routes, or any unknown stop combination,
without mutation or an evaluation loop.
-
For build, certify, an identity-only prepare exception, or the exact
partial-bundle stop exception, resolve the acquired track's registered
semantic profile from the manifest. Put the typed context in a gitignored
runtime JSON file. Supply the schema-required track, slug, family,
phase, module_state, and evidence_identity from the acquired result and
authoritative module ledger; do not infer them or add free-form prompt prose
to a profile:
.venv/bin/python scripts/orchestration/prompt_contracts.py resolve-track \
--track <track> --context <runtime-context.json>
Preserve the returned prompt/source identity with the phase evidence. A
retired or unknown track, stale selector, family mismatch, missing fragment,
or unresolved input is a fail-closed configuration defect.
-
Follow $track-completion completely for the acquired track/slug. Reuse
its exact ledger on resume. It owns plan review, build/recovery, read-only
$post-build-review, repair routing, reviewer-instability handling,
independent review, PR/CI/merge, publication evidence, and cleanup.
-
Record the module result from its authoritative ledger. A publishable result
is accepted only after $track-completion reaches COMPLETE for the
coordinator's exact goal; a terminal bounded-budget blocker is recorded as
blocked rather than left stale:
.venv/bin/python scripts/orchestration/curriculum_coordinator.py record-module \
--run-id <run-id> --owner <agent/task> --module <slug> \
--track-ledger <track-completion-ledger> --track-run-id <module-run-id>
Do not supply a coordinator disposition or repeat PR/worktree/cleanup
policy. The coordinator validates the bounded run, rejects semantic evidence
outside its model-call count, hashes the module ledger, and derives a concise
result with elapsed time, model-call count, repair count, disposition,
blocker, and next action. An incomplete module ledger cannot complete the
coordinator. A migrated historical module is reacquired until its exact goal
is proven.
-
Repeat acquisition serially until the coordinator reports complete.
Module-at-a-time mutation remains mandatory; a wave groups health and review
capacity but does not authorize concurrent learner writes.
