AGENT SKILL

workflow-transition-admission

Admit or hold a source-bound same-chat correction, candidate transition, known-cause retry, finalized Skill-dependent action, external action or decision-window wait from one identity-bound receipt. Preserve independent refutation, actual allocation-result consumption and lightweight read-only work. This is structural admission only, not semantic or outcome proof. Read legacy v1 only for explicit old split-chat leases.

License 许可
Apache-2.0
最近更新
2026-11-02
Stars
★ 23
Forks
1

01 前言 Front Matter

name
workflow-transition-admission
description
Admit or hold a source-bound same-chat correction, candidate transition, known-cause retry, finalized Skill-dependent action, external action or decision-window wait from one identity-bound receipt. Preserve independent refutation, actual allocation-result consumption and lightweight read-only work. This is structural admission only, not semantic or outcome proof. Read legacy v1 only for explicit old split-chat leases.

02 资源文件 4

03 SKILL.md 正文

Workflow Transition Admission

Use this Skill only at a decision-bearing workflow boundary. Do not invoke it for ordinary trivial read-only exploration.

  1. Build one receipt matching transition-admission-contract.md. Source-wide implementation uses v3 with the same scope/state binding as the owner and child dispatch. V2 remains for ordinary non-product boundaries and old same-chat inputs; v1 is historical split-chat compatibility. Neither older schema proves source-wide selection.
  2. For a binding material correction, commit/push, CI dispatch, or external apply, bind exact source authority or an independent refutation, closed FCR, frozen S0 plus interaction IDs, planned U0, omission consequence, expected objective delta, cost/rework counterfactual and exact return step. A proposed fix without these inputs is DIAGNOSE or ADVISORY, not an admitted material transition.
  3. The same-chat owner consumes the source-bound decision without claiming that its own control pass is independent review. A higher role never supplies missing source or closure. Preserve evidence on any dependent hold; independent refutation remains a separate key where required.
  4. Before diagnosis or repair, search both current masters for the current signature and disposition any known family. Retry a known family only with a changed premise and discriminating prediction.
  5. Before a finalized action, bind selected bytes, script when applicable, planned application and outcome capture. Append actual application/effect after execution. A future effect is not pre-action authorization. Selection/read alone is not use or benefit. A repeated mechanical family requires the deterministic script or another constrained representation.
  6. Immediately before a released external action, bind current action-relevant state. STALE or UNKNOWN returns REPLAN only for that action.
  7. Monitor at decision windows. Reuse unchanged cursors and evidence. Fixed polling or narration without a decision delta is not admitted. Slow work with objective evidence progress may return WAIT.
  8. Bind COORDINATED-WORK task/chat/lease/resource or BOUNDED-SUBAGENT parent lease/resource ownership. An independent reviewer cannot implement its reviewed candidate. Reconcile old leases and unknown effects before migration; v1 formal-worker restrictions do not block the new same-chat owner.
  9. Preserve scenario recomposition and the consumer target. A material edit binds its semantic lock, relational claims, recomposition witness, unresolved relations, consumer claim IDs and final oracle. The script checks only reference/set presence; it never claims those semantics are correct.
  10. Reassess allocation only when job shape changes. Bind and open actual view/result files, hashes, admitted decision and configuration; references or caller labels alone are insufficient. Auxiliary routes must name the mandatory outcome left unproven if omitted, time saved and retirement condition.
  11. After a correction episode, record induced/avoided rework, status-only cost, known-cause rediscovery, Skill recurrence, scenario omission, role violation, consumer result and objective delta as observed/unavailable/not-applicable. Feed a lifecycle revise/merge/narrow/retire/no-change candidate; the episode never self-activates a semantic policy change.

Run:

python -B scripts/transition_admission.py --input <receipt.json>

The script validates presence, identities, enumerated states and bounded cross-field relations only. It cannot prove source fidelity, FCR correctness, scenario or interaction completeness, impact safety, Skill semantics, external freshness truth, monitoring value, or objective outcome.

V3 consumes the sibling lifecycle work_phase.py decision before any lightweight bypass. EDIT must be implementation or grouped findings repair. Construction feedback cannot reclassify external apply, push, CI or an existing regression correction as an inexpensive local probe. Hold only the dependent transition; retain the already observed effects and continue independent safe work. The action gate and phase selection are separate necessary checks, not permission substitutes.

04 同仓库更多 Skills

Chisiki1
Chisiki1/objective-integrity-framework

scenario-interaction-closure

Freeze a source-bound, blind-first harm-scenario and interaction closure before a material correction, then preserve its meaning through evidence decomposition and recomposition.

★ 23
Chisiki1
Chisiki1/objective-integrity-framework

powershell-exact-action

Screen an exact finalized PowerShell command or batch member before execution for confirmed mechanical recurrence families, currently the invalid direct pipeline from a foreach statement and unsafe literal-dollar interpolation in metadata arguments. Use for Windows PowerShell actions after wrappers, defaults, JSON construction, member expansion, and quoting are final. It is a mechanical preflight, not semantic safety or runtime outcome proof.

★ 23
Chisiki1
Chisiki1/objective-integrity-framework

objective-supervisor-control

Turn a same-chat or legacy decision-bearing event into a bounded source-bound objective-control disposition without fixed polling or status narration. Use when consuming a result, correcting objective drift or deciding whether a wait has value; never treat this control pass as independent review or action authority.

★ 23
Chisiki1
Chisiki1/objective-integrity-framework

master-guided-skill-resolver

Compile provenance-bound source and finalized action facts, then resolve exact registered workflow skills with deterministic path- and hash-bound selected/rejected receipts. Use before material work when the two masters may contain reusable skill guidance, or when diagnosing missing, near, stale, conflicting, retired, or ambiguous selection. This skill does not replace the workflow, semantic judgment, independent review, or final-consumer evidence.

★ 23
Chisiki1
Chisiki1/objective-integrity-framework

master-guided-skill-lifecycle

Carry worthwhile workflow learning from a real result to an owned candidate, next use and measured disposition; retrieve source-bound master history, validate Skill effects and steward operating conditions or exact user-approved amendments. Use at material results, recurrence, useful reuse, avoidable cost, capability changes or a concrete policy-condition blocker. Do not turn ordinary reads into extra phases or treat records and invocations as benefit.

★ 23
Chisiki1
Chisiki1/objective-integrity-framework

durable-supervisor-status

Recover an explicitly existing legacy split-chat Supervisor/Worker lease using Worker-local status, hash-chain history and a later bound Supervisor message. Do not use for the default same-chat workflow, which uses chat-objective-continuity. Never use as external-action authorization or prevention proof.

★ 23