Aoa checkpoint closeout bridge

aoa-checkpoint-closeout-bridge

Intent

Preserve the existing checkpoint-closeout capability while its durable scenario remains an aoa-playbooks playbook without an executable MCP owner. Checkpoint hints remain provisional; the bridge makes every downstream skill selection, evidence reread, state transition, and stop explicit.

Trigger boundary

Use one mode:

• collect: record bounded checkpoint hints for later review without emitting harvest, progression, quest, or owner verdicts
• execute: carry one closed, explicitly reviewed artifact through the ordered closeout stages and return a session-local execution report

Do not use when the evidence is live, unreviewed, unbounded, or cross-session identity cannot be separated; when the request is only a recap; or when a mechanical checkpoint packet is being treated as proof, memory, progression, quest, playbook, or owner truth.

Inputs and outputs

Read references/contract.yaml first.

• collect consumes checkpoint-focus-input and returns provisional-checkpoint-focus
• execute consumes reviewed-checkpoint-closeout-input and returns checkpoint-closeout-execution-report

All outputs remain session-local. A caller may request an explicit local note or report path, but this skill never writes an owner repository, durable memory, quest state, progression state, proof surface, playbook, KAG, or stats surface.

Procedure

1. Select exactly one mode.
2. For collect, read references/collect.md to EOF and stop after the provisional focus packet.
3. For execute, read references/execute.md to EOF.
4. Before starting execute, require one closed and explicitly reviewed artifact plus its current-session boundary. Otherwise return blocked_unreviewed_evidence without loading downstream skills.
5. Load the exact available aoa-session-harvest and aoa-session-progression-lift bundles before their first nodes. Missing, ambiguous, or owner-unresolvable dependencies return blocked_missing_capability; do not replace them with generic reasoning or retired skill names.
6. Preflight each required owner dependency serially. In particular, obey the progression bundle's same-bundle source-handle -> manifest -> owner-model source-return gate with one completed tool turn per gate. Never substitute the profile root, .system, or a sibling skill directory for its exact handle. A manifest read batched with an owner document is a failed preflight, not a successful load.
7. Execute one node completely before selecting the next node. For each node: use one tool turn to load only that node's exact procedure; await it; use a later tool turn to reopen only the primary reviewed artifact; await it; then start, produce, and verify the typed output. Never preload several mode procedures or batch evidence rereads for several nodes.
8. Record the procedure-load and primary-evidence action references on that node. Earlier reads, cached context, checkpoint hints, another node's read, or a final assertion do not satisfy this requirement. Shared tool batches across nodes make every affected node unverified_nonserial_node_execution.
9. Record selected, loaded, started, produced, verified, and completed, or the exact blocked, failed, skipped, deferred, or handed-off terminal state. A node without its observable primary-evidence reread must terminate as unverified_missing_evidence_reread, never verified or completed.
10. Stop after the session-local execution report and explicit owner handoffs. Do not continue automatically into owner acceptance or writes.

Contracts

• preserve the dependency order: donor extraction/classification before progression; progression before quest-promotion judgment
• reread the primary reviewed artifact before every executed stage; checkpoint notes, generated packets, previous node outputs, and remembered content only narrow attention
• treat every classify[*] instance and closeout.report as its own stage; each executed instance requires its own post-selection, pre-start reread
• treat each downstream skill mode as a separate selection and typed node, not as an implicit multi-mode invocation
• do not parallelize DAG nodes merely because their source artifact is the same; the observable order is part of the checkpoint capability
• allow skipped, deferred, stopped, and no-candidate outcomes; the DAG shape never forces a positive verdict
• keep the aoa-playbooks checkpoint-closeout route as scenario authority; this transitional skill does not claim playbook runtime execution
• keep techniques as optional provenance only, never runtime dependencies

Verification

• verify review/closure and current-session identity before downstream load
• disposition every checkpoint hint as accepted, rejected, stale, cross-session, contaminated, or unresolved
• verify every produced node output against its loaded skill contract
• confirm stage order, observable evidence-read action refs, effects, terminal states, and owner handoffs are visible in the report
• compare every report claim with the actual node action record; never report a reread, load, tool call, effect, or verification that did not occur
• reject a run whose dependency gate or node reads were parallelized across phases, even if the final prose and action refs look plausible
• confirm no downstream owner surface changed and no mechanical artifact was narrated as reviewed judgment