一键导入
using-control-json
Use when executing an existing v1 compiled run.control.json/runtime.control.json chain.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Use when executing an existing v1 compiled run.control.json/runtime.control.json chain.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when a formed task needs durable human-approved meaning, completion obligations, evidence claims, non-goals, or live/destructive authority boundaries.
Use when maintaining or resuming an existing v1 requirements.control.json run; new runs use forming-cybernetic-commitments.
Use when an existing v1 run must compile its reviewed control chain into runtime.control.json.
Use when optional local workflow guidance or legacy v1 workflow compatibility needs diagnosis.
Use when an existing v1 control chain requires design.control.json; v2 strategies remain agent-owned.
Use when cybernetic routing is requested but user input is pre-task intent or role-ambiguous context: confusion, dissatisfaction, risk sense, observed symptoms, failed attempts, method preference, process distrust, source material, declared current state, or unclear requested transformation.
| name | using-control-json |
| description | Use when executing an existing v1 compiled run.control.json/runtime.control.json chain. |
Legacy v1 compatibility only. New runs use $using-commitment-control-json.
This is not bounded_runtime. Bounded runtime goals use
.agents/skills/using-bounded-control-json.
It supports official generation-aware runs with requirements.control.json,
run.control.json, and the current gen-N/runtime.control.json.
Load all generation control_refs; the plan remains policy authority and
runtime is its validated projection.
Read references/runtime-control-json-protocol.md before executing, reporting status, or preparing completion evidence.
Gate result semantics follow ../references/transition-gate-protocol.md.
progress.jsonl, runtime-status.json, and final-report.json; evidence stays under authorized evidence paths.progress.jsonl. Runtime must use append_progress_event.py. Direct writes to progress.jsonl are invalid. Record observations; do not mutate approved JSON.final-report.json does not grant verifier permission to itself. Run the structural verifier before goal_achieved: true; Counterexample Gate supplies semantic review./goal is a pointer, not control fact. Resolve current_generation at execution; old concrete runtime paths are stale after a switch.required_single_manifest, bind completion/review/verification events
and final report to one content-addressed evidence subject. Manifest
run:/repo: references must resolve to actual bytes, every required
work-package component must be represented, and subject_id must equal the
canonical manifest-content hash. Reject mixed, unresolved, or stale subjects.action_alignment_ref, produced artifact hashes, and
required evidence paths. Source-file lists are not an authority gate:
implementation may discover any repository files needed to produce the
approved state change, subject to semantic non-goals and live/destructive
gates. Tests are not mandatory progress fields and do not
impose test-first ordering, a test quota, a package frontier, or a completion
condition. A later started/failed/blocked event revokes the earlier pass for
that step and invalidates downstream package credit until rerun.Structural gates are schema checks, control_chain_guard, validate_control_chain,
and verify_runtime_progress. They prove shape, hashes, coverage, and progress,
not quality approval.
Quality gate means counterexample-gate: an independent reviewer tries to
disprove the target decomposition, runtime strategy, blocked claim, or
completion claim. It must execute the requirements-approved contract, every
per-outcome gate, and for schema 1.3.0+ every required result path. Runtime
must not invent or weaken these gates.
Runtime may add or run tests when useful, but developer-authored tests do not replace Counterexample Gate. A discovered counterexample may be retained as a regression test when it protects a real failure, a named critical invariant, or materially lowers repeated observation cost; retention is optional.
When strategy cannot produce a blocking outcome, do not substitute readiness,
compatibility, or weaker evidence. Append a
control.amendment.proposed event with the current runtime_generation,
reason, triggering observation, affected strategy stages,
affected_source_requirements, anchor-change flags, and a patch_ref to the
next candidate strategy.
Anchor-preserving tactical amendments are reviewed into a new generation by the
orchestrator only when run.control.json.strategy_policy is
reviewed_replanning. Under frozen_strategy, an amendment proposal reports
that the approved strategy does not fit and requires a human decision before
continuing. The orchestrator must not synthesize its own approved review; it can
switch generations only after the complete generation plan and patch have an
approved independent amendment review, then it switches automatically.
Proposals that change requirements.control.json,
semantic_base, required outcomes, what counts as done, work coverage,
commitment authority, or forbidden actions are not amendments and return to
human commitment.
If validation, progress append, status write, or verifier execution cannot be completed, report goal_achieved: false with the missing JSON, missing evidence, or failing verifier condition. Component completion is only progress evidence until the verifier permits the final claim.