---
name: ozone-manager
description: Mandatory lightweight bootstrap when a request mentions OZM, OZoneManager, OZoneMaster legacy alias, ozone-manager, ozm-* skills, or when complex project work needs governance across intake, dispatch, writing, repair, review, acceptance, closeout, truth ownership, recurring failure control, or OZM skill hardening.
---

# OZoneManager

Repo-level control skill for complex governed work. It routes work through the minimum OZM stage skills, reuses existing governance skills, and keeps planning, write-set control, verification, and closeout honest.

## Mandatory Bootstrap

When a request names OZM, OZoneManager, OZoneMaster legacy alias, `ozone-manager`, any `ozm-*` skill, or asks for governed workflow behavior that touches intake, dispatch, writer admission, truth boundaries, claim ceilings, review, acceptance, closeout, recurring failures, or OZM hardening, load this umbrella before attempting the work.

This bootstrap is a routing gate only. It must classify the active question and current phase, then load the smallest necessary child skill set. It does not authorize preloading the full OZM family.

The current-phase child is the primary owner, not an exclusive lock. Phrases such as `load only the current-phase OZM child skill` prevent family preload and optional reference chasing; they do not suppress mandatory companion children when a T0 stop, control-surface write, audit/subagent tool event, post-compression reentry, closeout, next-packet admission, or positive claim makes another owner executable.

## Actual Child Skill Hydration Gate

Candidate routing is not activation. A route-graph result, guard script pass, prompt inventory, project instruction, remembered rule, or child skill name in chat does not count as loading the child skill.

Before any OZM-governed tool execution, file mutation, control-surface write, subagent launch/result consumption, guard/closeout run, post-compression continuation, next-packet admission, or positive claim, the currently owning child `SKILL.md` and all mandatory companion child `SKILL.md` files must be actually opened in the current turn. If this cannot be confirmed, stop at `child_skill_hydration_missing`, report the missing child ids, keep the claim ceiling at diagnostic/planned, and do not treat downstream evidence as acceptance proof.

Context compression starts a new hydration epoch. Any `loaded_child_skills` receipt, assistant activation statement, route output, or child `SKILL.md` read that happened before the latest `compacted` / `context_compacted` event is historical-only for resumed actions. After compression, the next tool execution, file mutation, subagent launch/result consumption, guard/closeout run, next-packet admission, or positive claim must reopen the owning child `SKILL.md` files and emit a new receipt tied to the post-compression turn.

The receipt must be epoch-bound when a compaction was observed: include `hydration_epoch=post_latest_compaction`, `latest_compaction_event`, `post_compaction_skill_loads`, and `expired_pre_compaction_skill_loads`. A final answer that lists `loaded_child_skills` without actual post-compaction `SKILL.md` reads is a false receipt and keeps the ceiling at `child_skill_hydration_missing`.

Every governed continuation should leave a short hydration receipt: `loaded_child_skills=[...]`, `mandatory_companions=[...]`, `missing_child_skills=[]`, and the epoch fields above when applicable. If the receipt is non-empty for missing children, the next action is route repair, not implementation or closeout.

For skill-effectiveness audits, hydration is necessary but not sufficient. Require the owning child to show an activation effect: owner question answered, blocker or admission posture, required artifact/ledger/receipt/anchor binding, downstream consumer, proof/script posture, and claim-ceiling effect. If a skill is only routed, mentioned, or opened without changing these fields, record `route_only_activation` or `loaded_without_effect_contract` and keep claims at diagnostic.

## Activation Effect Contract

The umbrella activation effect is route gating, not implementation authority. It must answer owner child, T0 stop, hydration epoch, activation-effect posture, and claim ceiling before governed action. Required bootstrap receipt fields live in `references/bootstrap-minimal.json` and `references/bootstrap-activation-effect.json`; detailed budget posture lives in `references/bootstrap-budget.md`.

## Canonical Absorbed Defaults

Canonical absorbed defaults are enforced by the owning OZM child stage; absorbed donor ids are historical inputs, not normal runtime dependencies. Load `references/bootstrap-routing-details.md#canonical-absorbed-defaults` only for donor-boundary audits, package-scope checks, or old-id trigger repair.

## Use When

Use this bootstrap for OZM/OZoneManager/OZoneMaster/`ozm-*` requests and for governed intake, dispatch, writing, repair, review, closeout, truth, record, context, claim, skill-hardening, capability-evolution/evo, reference-method, repo-graph, document drafting, UX/UI, image2, text I/O, agent-runtime, or long-loop work. Extended trigger inventory lives in `references/bootstrap-routing-details.md#use-when`; load it only when route coverage or missed-trigger behavior is being audited.

## Load Order

1. Read the active repo instructions when they exist, starting with the nearest applicable `AGENTS.md` or equivalent project guidance.
   - If no repo-local `AGENTS.md`, `CLAUDE.md`, or equivalent instruction file exists, record `repo_instruction_surface=absent` and continue with OZM plus user/system/developer instructions. Absence is not a warning, blocker, or request to create a file unless the user asks for repository instruction bootstrapping.
2. Read repo-local workflow governance docs when they exist and are relevant to the task.
3. If OZM, OZoneManager, OZoneMaster legacy alias, `ozone-manager`, or any `ozm-*` skill was named, treat this umbrella as already required and route from it before child-skill selection.
4. Read `references/module-routing.md` only when routing or hardening is non-trivial, a hard stop may apply, multiple child stages compete, or the user asks about OZM structure.
5. For ambiguous skill selection, large skill-library questions, OZM structure hardening, composite reference-plus-domain requests, or any request that names UI/UX/visual/browser/screenshot work alongside OZM/reference governance, `scripts/ozm_skill_graph.py query "<request>"` must be used as candidate routing aid after this umbrella is loaded; run it through a resolved Python interpreter, not bare `python` on Windows.
6. Activate only the child OZM skills and preserved specialists needed for the current phase. Mandatory support children required by T0 stops are part of the smallest necessary set, not optional preload. When the route includes a preserved specialist, OZM still owns request role, write-set, truth owner, verification target, and claim ceiling.
   - Activation means the child `SKILL.md` was actually opened in this turn, not merely named by route graph, guard output, a prompt, or a compressed summary.
   - If the next action is execution, mutation, subagent/audit carrier use, post-compression evidence consumption, closeout, next-packet admission, or positive wording, emit the hydration receipt before acting.
   - If a compaction/resume event occurred after the last child load, discard the old receipt and reload the current owner and mandatory companion child skills before acting.
7. Open exactly one second-level routing reference only when the decision needs it: `routing/stage-absorption-matrix.md`, `routing/failure-mode-routing.md`, or `routing/specialist-preserve-quarantine.md`.
8. Each child skill should keep high-frequency absorbed rules inline and load donor skills only for the remaining reusable workflow.

## Context Budget Rule

Default loading is bounded: classify phase, load the owning child plus mandatory companions, and stop reference chasing once the executable gate is known. Detailed budget, hydration, model/runtime, and route-noise rules live in `references/bootstrap-routing-details.md#context-budget-rule`; load them for context-pressure, missed-route, or post-compression audits.

Use `references/bootstrap-budget.md` when auditing progressive disclosure, default-load word budget, or archive-reference leakage.

## Top-Level Rule Weights

OZM weighs hard stops above convenience routing: hydration, truth ownership, write admission, acceptance proof, and claim ceiling override momentum. Detailed weight posture lives in `references/bootstrap-routing-details.md#top-level-rule-weights`.

## Core Workflow

1. Classify the latest visible user request, not a compressed summary: role (`plan_only`, `read_only_plan`, `execution_requested`, `audit_only`, etc.), phase, and control-plane weight (`control_dominant`, `hybrid`, `domain_dominant`, `evidence_closeout`).
2. Route to the minimum child set for the phase. Add companions only when the next action crosses ownership: record writes, truth changes, dispatch/writer admission, audit/subagent launch or consumption, review evidence, closeout, positive wording, post-compression reentry, reference-method claims, UI/runtime proof, or text I/O.
3. Apply the umbrella stop table before momentum. If a T0 stop fires, load the owner child, lower the request/claim posture, and do not treat route output, guard pass, summary, or remembered policy as proof.
4. For plan-only/read-only requests, stop at requirement/document planning surfaces. Dispatch, code writing, tests/builds, runtime probes, subagent execution, and product/source edits stay closed until a later explicit execution request.
5. For reference, paper, source-level rewrite, method-grounding, UI/runtime, repo-graph, document-drafting, or agent-runtime work, freeze the relevant method/map/contract owner before dispatch and require downstream consumption at writing, review, closeout, and claim ceiling.
6. For long-running loops or standing autonomy, keep mission authorization separate from bounded execution units: refresh queue, constraints, claim ceiling, carrier, evaluator, stop condition, context budget, and next packet before each admitted unit.
7. After compression, handoff, wait, replay, replacement, role switch, or audit-result consumption, treat previous child loads and receipts as historical. Rehydrate the current owner and mandatory companions, then reread owner truth before acting.
8. Before positive wording, reconcile stale summaries, competing truth surfaces, ledgers, receipts, cleanup posture, proof surfaces, and next gate through closeout plus claim ceiling.

Detailed route/background rules live in `references/bootstrap-routing-details.md`; this section is only the default execution index.

## Umbrella Stop Index

T0 stops stay in this umbrella because they must fire before child-stage detail is loaded. The table is the default execution surface: match the trigger, load only the owner child skill, and apply the claim/request effect before any further action.

| Stop ID | Trigger | Owner child | Claim/request effect |
| --- | --- | --- | --- |
| T0-ROLE | Latest request is plan-only, read-only, audit-only, or says not to execute, modify, repair, implement, run, test, or write. | `ozm-requirement-load`, `ozm-claim-ceiling` | Keep dispatch, code writing, tests/builds, subagents, runtime probes, and product/source edits closed until a later explicit execution request. |
| T0-SCOPE | Ambiguity changes scope, owner, behavior, verification, irreversible product posture, cost, secrets, destructive action, or legal/security/privacy posture. | `ozm-requirement-load` | Stop or lower to diagnostic/planned until owner evidence or a minimal human-owned answer resolves the ambiguity. |
| T0-OBJECTIVE | Slice, MVP, fallback, proof floor, recent packet, or local success tries to replace the final product/thread objective. | `ozm-requirement-load`, `ozm-claim-ceiling` | Restore final-objective trace and lower any completion wording that outruns it. |
| T0-LOCAL-FIRST | MVP-first, demo-first, real-environment-first, or live-first sequencing would skip locally realizable master-plan/reference-guided work. | `ozm-requirement-load`, `ozm-dispatch-freeze`, `ozm-external-prerequisite-gate` | Default to local-complete-first unless current owner evidence makes live work current. |
| T0-WRITER | Writer admission is attempted without requirement load, dispatch freeze, write-set, artifact placement, truth owner, or claim ceiling. | `ozm-dispatch-freeze` | Keep writer lane closed. |
| T0-CONTROLLER | Writer tries to mutate Plan, Goal, master-plan, roadmap, requirements, acceptance, schema, API contract, ADR, or truth calibration to match output. | `ozm-dispatch-freeze`, `ozm-truth-boundary-management` | Require explicit controller-update packet; ordinary writer may only record candidate deltas. |
| T0-EVIDENCE | Overview, label, tag, summary, score, screenshot, generated matrix, artifact, placeholder, stale receipt, or historical proof is treated as fresh evidence. | `ozm-truth-boundary-management`, `ozm-claim-ceiling` | Resolve to owner evidence or downgrade to navigation/historical-only. |
| T0-REPAIR | A bug/failure/fix request may be stale, already fixed, working as intended, unreproduced, diagnostic-only, or no-op. | `ozm-error-repair-debug` | Reproduce/classify before patching; no-op close needs diagnostic proof. |
| T0-OUTCOME | Essential outcome skeleton, must-observe state transitions, negative/recovery behavior, or owner evidence pointers are missing. | `ozm-requirement-load`, `ozm-review-diffgate-acceptance` | Cannot make acceptance-grade claims. |
| T0-TEST | Tests, CI, assertions, snapshots, mocks, timeouts, coverage, or workflow filters are weakened without owner/admitted packet approval. | `ozm-review-diffgate-acceptance`, `ozm-claim-ceiling` | Treat as blocker or ceiling downgrade. |
| T0-SHALLOW | Simple implementation, top-layer path, shortcut glue, facade-only wiring, weak tests, self-certified completion, or low-runtime-substance claim is used as acceptance. | `ozm-review-diffgate-acceptance`, `ozm-role-stack-coordination` | Require neutral/independent audit or lower claim. |
| T0-PLAN-MATRIX | Plan/Goal/spec/API/schema/status/waiver/deviation handoff lacks contract matrix, canonical field owner, listed endpoint completeness, escape-hatch binding, enum consistency, or plan-to-dev readiness. | `ozm-requirement-load`, `ozm-review-diffgate-acceptance` | Not dev-ready or auditable. |
| T0-DOC-STRENGTH | Governed planning/control docs are summary-only or too thin to drive future agentic coding. | `ozm-requirement-load`, `ozm-record-surface-management` | Add durable skeleton or indexed child docs before implementation-ready wording. |
| T0-DRAFT | Governed text drafting starts from prose or outline without draft research gate, audience/action contract, concept/unknown ledger when needed, claim-evidence-argument matrix, section depth floor, reader/editor issue registry, or draft closed-loop receipt. | `ozm-document-drafting`, `ozm-requirement-load`, `ozm-review-diffgate-acceptance`, `ozm-closeout-handoff`, `ozm-claim-ceiling` | Keep wording at `draft_candidate`, `shallow_summary_only`, `evidence_incomplete`, or `review_pending` until evidence rows, issue closure, and verdicts exist. |
| T0-METHOD | Implementation method, version plan, roadmap, MVP ladder, script plan, CLI/MCP command matrix, or per-version claim ceiling is too thin. | `ozm-requirement-load`, `ozm-role-stack-coordination` | Stop at planning; add method/version/script matrix before dispatch. |
| T0-REFERENCE | Reference project/paper/engine/mature-system work lacks source-first pre-analysis, runtime capability map, cross-reference synthesis, target truth runtime map, paper method card when papers govern execution, adoption decisions, or anti-transplant constraints. | `ozm-requirement-load`, `ozm-reference-method-grounding` | No reference-grade, paper-level, or mature-runtime dispatch. |
| T0-METHOD-MAP | Full restoration, same technical approach, same stack, source-level rewrite/rebuild, source-backed reconstruction, paper methodology landing, or reference method grounding lacks reference method map, Paper Method Card, method adoption contract, or source-backed gap ledger. | `ozm-requirement-load`, `ozm-reference-method-grounding` | No same-method or paper-method writer admission. |
| T0-METHOD-ANCHOR | Reference-guided packet lacks execution anchor ids, source-backed gap, proof surface, target-owner requirement link, or claim ceiling if the anchor is not consumed. | `ozm-reference-method-grounding`, `ozm-dispatch-freeze` | No reference-guided code writing or mainline reference progress. |
| T0-REPO-GRAPH | Repo graph, CodeGraph/MCP/CLI, `.repo_analysis`, source-level reconstruction, graph-first exploration, or impact-radius-before-write work lacks a fresh graph receipt, graph artifact contract, reconstruction bundle, impact posture, mechanism-fidelity row, or lowered claim ceiling for stale/absent backend evidence. | `ozm-repo-graph-reconstruction`, `ozm-reference-method-grounding`, `ozm-claim-ceiling` | Graph output is navigation-only; no source-level reconstruction, reference progress, or broad-impact claim, and no writer admission based on stale graph context. |
| T0-WRONG-DIR | Packet continues a technical path that conflicts with an adopted/adapted reference method node without owner-approved divergence. | `ozm-reference-method-grounding`, `ozm-dispatch-freeze` | Block or return to requirement load. |
| T0-REF-VALUE | Reference-guided progress claim proves only local truth, support work, guard work, record sync, or proof reduction without reducing a source-backed reference gap. | `ozm-reference-method-grounding`, `ozm-review-diffgate-acceptance` | Do not count as mainline reference progress. |
| T0-UI | UI/UX/frontend/browser/screenshot/visual-fidelity/map/globe work is in scope but no OZM UI owner handoff or not-needed reason is recorded. | `ozm-ux-ui-expert-suite`, `ozm-review-diffgate-acceptance`, `ozm-claim-ceiling` | Keep visual/reference-parity claims at candidate or pending-controller-gate. |
| T0-AUDIT-CARRIER | Subagent, independent-audit, neutral-audit, Codex-review, second-model review, review-helper, `NO_BLOCKING_FINDINGS`, `spawn_agent`, `wait_agent`, `send_input`, `resume_agent`, `close_agent`, or similar audit-pass/tool-event wording is launched or used as proof while the role-stack carrier, runtime authority, fork/model/tool contract, or result receipt is unavailable, stale, rejected, or unclassified. | `ozm-role-stack-coordination`, `ozm-review-diffgate-acceptance`, `ozm-closeout-handoff`, `ozm-claim-ceiling` | Load role-stack before tool events, count rejected spawns as tooling noise, and treat unsupported results as candidate/same-thread evidence only; final PASS still needs closeout and claim ceiling before progress wording. |
| T0-RUNTIME-PROOF | UI/browser/runtime/map/globe/product proof comes only from a harness, fixture, demo page, screenshot helper, generated artifact, test-only route, or smoke page while the claimed product/runtime entrypoint, owner API, console/error state, negative state, or integrated public seam is unchecked. | `ozm-review-diffgate-acceptance`, `ozm-ux-ui-expert-suite` when the proof is visual/UI, `ozm-closeout-handoff`, `ozm-claim-ceiling` | Lower to `harness-only-proof`, `runtime-entrypoint-unproven`, or candidate wording until the actual entrypoint is verified. |
| T0-COMPANION | A prompt, activation anchor, or route note says only the current-phase child should load, but the next action writes control surfaces, launches or consumes audit/subagent evidence, resumes after compression, closes a packet, admits the next packet, or makes positive wording. | Primary child plus the owning support child: record, truth, role, review, closeout, dispatch, or claim. | Treat companion loading as mandatory minimal routing; do not use current-phase-only wording as a waiver. |
| T0-HYDRATION | OZM route graph, guard output, prompt text, metadata, compressed summary, pre-compaction receipt, or remembered habit names a child skill, but the actual child `SKILL.md` was not opened in the current post-compression turn before execution, audit/subagent use, closeout, next-packet admission, or positive wording. | Matched owner child plus mandatory companions. | Stop at `child_skill_hydration_missing`; emit `loaded_child_skills`, `mandatory_companions`, `missing_child_skills`, and `hydration_epoch=post_latest_compaction`, and keep evidence navigation-only until the missing child skills are loaded. |
| T0-POST-AUDIT-MUTATION | A review/subagent/audit PASS is consumed, then queue/current-state/Plan/Goal/MTL/GL/report/manifest or other controller/control surfaces change before final closeout. | `ozm-review-diffgate-acceptance`, `ozm-record-surface-management`, `ozm-closeout-handoff`, `ozm-claim-ceiling` | Treat the PASS as stale for final-state wording until a final control-surface review reconciles the mutated surfaces. If the only later change is an append-only review/audit receipt explicitly marked `record_sync_only`, keep only that lowered claim and do not treat it as material implementation proof. |
| T0-CONTEXT | Context compression, stale summaries, lost-in-middle, context poisoning, filesystem scratch notes, large-output routing, or context-mode output may be treated as current truth. | `ozm-context-engineering`, `ozm-record-surface-management`, `ozm-truth-boundary-management` | Keep summaries and scratch files navigation-only; require owner reread and a context receipt before dispatch, review, closeout, or positive continuity claims. |
| T0-HISTORY | Old packet/version logs, screenshots, receipts, final wording, or `verified`/`passed` strings are bulk-read as current truth. | `ozm-record-surface-management`, `ozm-truth-boundary-management` | Load active window, truth calibration, continuation queue, and packet-history index first. |
| T0-MEMORY | Master-plan/current-state/ledger/gap-register/packet corpus is too large for reliable default reread and no compact memory index exists. | `ozm-record-surface-management` | Require navigation-only compact memory index; it cannot replace product proof. |
| T0-CONTINUITY | Long-running file-driven loop jumps to writer admission from chat momentum, stale queue, unsplit candidate, or revisionless priority. | `ozm-requirement-load`, `ozm-record-surface-management` | Run planning-continuity tick and dispatch one bounded packet only. |
| T0-GOAL | Native `/GOAL`, Codex Goal, explicit `create_goal`, an already-active Goal, heartbeat, scheduler, automation, auxiliary lane, or runtime bridge is implied without current carrier, evaluator, budget, stop condition, latest request role, and queue revision. Generic `持续推进` / `自动推进` / run-until-done wording is OZM standing autonomy, not native Goal creation. | `ozm-requirement-load`, `ozm-record-surface-management`, `ozm-closeout-handoff` | Treat unavailable or implicit carrier claims as text control only; do not create native Goal unless the user explicitly requested it or one already exists. |
| T0-STANDING-AUTONOMY | Agentic coding loop, `持续推进`, `自动推进`, continue-until-done, run-until-done, or `无限执行许可` is in force, but no contract separates mission-level standing authorization from bounded execution units, hard stops, latest-request overrides, and runtime carrier limits. | `ozm-requirement-load`, `ozm-record-surface-management`, `ozm-closeout-handoff` | Freeze Standing Autonomy Contract: default continue until hard stop, one bounded execution unit per evaluator pass, explicit checkpoint cadence, and no background/native-Goal claim unless an explicit carrier exists. |
| T0-CONTROL-WEIGHT | Task phase or type changed, or domain execution is active, but OZM/control-plane documents still dominate instead of handing domain judgment to the current owner skill, source evidence, UI/runtime specialist, or reference method map. | `ozm-requirement-load`, `ozm-dispatch-freeze`, preserved specialists when applicable | Reclassify `control_dominant`, `hybrid`, `domain_dominant`, or `evidence_closeout`; keep OZM as a thin guard when domain work should lead. |
| T0-CONTROL-NOISE | Control-surface reads, rewrites, summaries, ledgers, graph/routing checks, or record sync keep diluting attention even after the correct owner phase is known. | `ozm-record-surface-management`, `ozm-recurring-failure-governance` | Freeze a control-noise budget and batch cadence; stop ordinary feature looping if the next action is really record sync, control tooling, or method reset. |
| T0-AUTO-METHOD | `自动推进`, continue, run-until-done, or self-driving wording is treated as an interface, background capability, or blanket execution authorization rather than a bounded evaluator method. Explicit native Goal wording remains governed by T0-GOAL. | `ozm-requirement-load`, `ozm-dispatch-freeze`, `ozm-closeout-handoff` | Require evaluator method, one bounded packet, stop authority, retry budget, correction handling, and no cross-gate role escalation. |
| T0-THROUGHPUT | Repeated full gates, subagent waits, hot-surface rereads, hash fanout, audit recursion, environment preflight failures, or context pressure appear. | `ozm-dispatch-freeze`, `ozm-record-surface-management`, `ozm-recurring-failure-governance` | Freeze loop throughput posture; possibly switch to control-tooling, record-sync, environment-preflight, or semantic-freeze audit. |
| T0-REENTRY | Context compression, handoff, resume, long wait, replay, replacement, role switch, post-compression audit, or subagent-result consumption lacks latest-request binding, prompt reload, owner reread, forbidden actions, and claim ceiling. | `ozm-truth-boundary-management`, `ozm-record-surface-management` | Execution and audit consumption stay closed at `reentry-unbound` or `planned`. Post-compaction audit/subagent result consumption requires both truth-boundary and record-surface child hydration in the current epoch. |
| T0-ANCHOR | Governed recovery, fresh-thread prompt, continuation, goal runtime, auxiliary thread, scheduler, or handoff omits the literal OZM activation anchor. | `ozm-record-surface-management`, `ozm-closeout-handoff` | Add anchor; it reactivates umbrella only and does not authorize execution. |
| T0-PLACEMENT | Created/moved/renamed/generated/archived/deleted files lack owner, authority, allowed root, naming basis, lifecycle, cleanup trigger, or index/map impact. | `ozm-record-surface-management`, `ozm-dispatch-freeze` | Block or lower file/artifact claims. |
| T0-NAMING | Active project/source/config/data/UI/deployment/maintenance surfaces contain dates, versions, work-unit ids, packet ids, run ids, local absolute paths, or historical receipts outside planning/control/archive posture. | `ozm-dispatch-freeze`, `ozm-review-diffgate-acceptance` | Fix or lower deployment/maintenance/clean-baseline claims. |
| T0-CLOSEOUT | Clean-baseline, release, deployment-safe, maintainer-safe, broad hygiene, final-objective, packet closeout, controller-consumption, next-packet admission, final review PASS, final subagent PASS, or `pre-closeout` guard-pass wording is used without the closeout owner reconciling active surfaces, inherited proof, next gate, and claim scope. | `ozm-closeout-handoff` | Run closeout owner gate; final PASS and guard pass are inputs, not substitutes. Under standing autonomy, a dispatchable next gate is a continue signal unless a hard stop is named. |
| T0-CLAIM | OZM-governed closeout, progress report, commit summary, final PASS consumption, or positive wording such as completed, done, submitted, verified, ready, passed, 已完成, 完成, 已提交, or 已通过 lacks an explicit current claim ceiling. | `ozm-claim-ceiling` | Add exact ceiling and proof gap; do not let positive wording imply acceptance. |
| T0-DONOR | Archived donor ids such as `writing-plans`, `controller-truth-guard`, or `verification-before-completion` are invoked as live skills. | `ozm-skill-hardening`, owning child | Route by OZM owner; donor names remain archive/history only. |
| T0-SKILL-AUDIT | Session/thread skill-call audit counts ids from `session_meta`, global skill lists, prompt inventories, embedded skill bodies, route tables, or quoted text as real activation, or judges subagent review from notification prose without tool-event receipt. | `ozm-skill-hardening`, `ozm-role-stack-coordination`, `ozm-review-diffgate-acceptance` | Reconstruct actual activation from SKILL.md reads, assistant activation plus opened file, route/guard outputs, and subagent tool events; metadata mentions are candidate only. |
| T0-RUNTIME | Text records imply native `/goal`, background execution, scheduler, heartbeat, browser broker, external harness, model-switching audit, or subagent lane that current runtime lacks or user has not authorized. | `ozm-external-prerequisite-gate`, `ozm-claim-ceiling` | Record unavailable/degraded carrier posture and lower dependent claims. |
| T0-MODEL-TOOLS | GPT-5.5, GPT-5.5 pro, xhigh/extra-high reasoning, hosted API Skills, local Codex Skills, `tool_search`, `apply_patch`, hosted shell, computer use, web search, MCP, or any model/tool profile is used as an execution or audit assumption without exact runtime support. | `ozm-dispatch-freeze`, `ozm-role-stack-coordination`, `ozm-skill-hardening`, `ozm-claim-ceiling` | Freeze exact model id/variant, context/output cap, reasoning budget, skill carrier, discovery budget, and tool support matrix; lower dependent claims when support is absent or unknown. |
| T0-EVO | Capability evolution, evo/self-evolving agent, self-improving coding-agent, skill mutation, promotion, rollback, benchmark-first self-improvement, or LLM evaluator API is requested or implied without baseline, eval, rollback, promotion owner, runtime/API posture, and claim ceiling. | `ozm-capability-evolution-governance`, `ozm-skill-hardening`, `ozm-recurring-failure-governance`, `ozm-claim-ceiling` | Treat as an evolution candidate only; no background self-modification, remote hub, or API-execution authority, and no promotion until eval/review/rollback gates pass. |
| T0-EXPERIENCE | A single retrospective, one success, one failure, or candidate experience is treated as Training-Free GRPO, proof, or active instruction. | `ozm-recurring-failure-governance`, `ozm-skill-hardening` | Require comparable trajectories, verifier/reward basis, holdout/regression, and promotion gate. |
| T0-RECURRING | Same severe governance failure, review failure, verification failure, zero-write/nonstart, method failure, or large-skill-library failure family appears a second time. | `ozm-recurring-failure-governance`, `ozm-skill-hardening` when OZM itself changes | Downgrade method, link or create a recurring-failure registry family, add route/eval coverage, and verify the prevention mechanism before replay. |
| T0-FENCE | A long copyable prompt contains triple-backtick examples but is wrapped in a triple-backtick fence. | `ozm-closeout-handoff` | Use a longer outer fence such as four backticks. |
| T0-MAP | Active source-map, package-boundary, route registry, harness registry, or debug-navigation truth relies on stale archive/release maps. | `ozm-code-writing`, `ozm-record-surface-management` | Regenerate or lower source-map posture. |

## Umbrella Enforcement Rules

- If any T0 stop fires, do not proceed by model confidence; route to the owner child and record the lower claim/request posture.
- A route candidate, graph score, hook pass, eval pass, generated matrix, or subagent output cannot bypass the umbrella or raise a claim ceiling.
- Do not preload the full OZM family to resolve a stop; load the owner child plus at most one directly required support child.
- Do not move a T0 stop out of this umbrella. Move only executable detail, field dictionaries, history, and low-frequency examples to child skills or references.
- Do not treat a child skill, preserved specialist, donor archive, or project doc as authority to widen the latest visible user request.
- Do not convert plan-only/read-only requests into dispatch, code writing, tests/builds, subagent execution, product/source edits, or artifact-producing work except for an explicitly requested plan artifact.
- Do not accept proof weaker than the claim wording; lower the wording, keep the gap open, or run the owner gate.
- Do not hide an unresolved stop in a summary. Output the stop ID, owner child, current ceiling, and next gate.

## Load Additional References Only When Needed

- `references/semantic-outcome-gate.md` for semantic bootstrap schema, route trace, eval-run manifest, and activation-effect validation anchors.
- `references/route-v2-architecture.md` only when changing route retrieval, rerank, selected subunit evidence, domain execution packs, automated skill optimization, or benchmark audit records.
- `references/bootstrap-activation-effect.json` for the machine-readable umbrella activation effect, route latency budget, black-hole score budget, and post-compaction reread requirement.

## Output

Leave a compact route receipt: active phase, child loads, mandatory companions, missing children, activation effects, claim ceiling, next authorized action, and downgraded proof. Detailed output shapes live in `references/bootstrap-routing-details.md#output`.