# Context Governor

> 任务需要在生成前制定受限的上下文加载计划，或当前加载策略有上下文腐蚀风险时使用。

- Skill: `xucroyuri/context-governor` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add xucroyuri/context-governor`
- Raw SKILL.md: https://api.skillmd.com/api/skills/xucroyuri/context-governor/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: xucroyuri (https://skillmd.com/u/xucroyuri)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/xucroyuri/context-governor

---


# Context Governor

Use this skill when the user asks about agent-skill design, loading strategy, context balance, reference breadth, or when the route is at risk of being drowned by too much adjacent material.

## Workflow
1. Lock the primary route first.
2. Choose one loading mode from `route_kernel`, `focus_pack`, `compare_pack`, `teaching_pack`, or `survey_pack`.
3. If the request is broad and theory-facing, pick one declared research bundle before expanding.
4. If continuity pressure is the real bottleneck, test whether a `story_memory_checkpoint` should replace broader reload.
5. Define the mandatory bundle.
6. Define the optional expansion queue and triggers.
7. Define stop conditions and corrosion signals.
8. If voice is the real bottleneck, prefer one bounded expression lens over a broad style survey.

## Output Contract
- `context_loading_plan`: primary route, loading mode, mandatory bundle, optional expansion queue, stop rules, and context-corrosion warnings.
- Do not confuse "more related material" with "better generation conditions."
- Expand only when the next asset can plausibly change route quality, answer quality, or creative extension quality.
- For broad theory/background requests, mandatory bundle should include one declared background bundle rather than a grab bag of neighboring docs.
- If the real problem is resumable continuity, prefer a checkpoint artifact over a wider bundle.
- Treat expression guidance as an adjunct bundle, not as a reason to pre-load every nearby style example.

## Posture-Adaptive Guidance

**When `certainty = lost`:**
Skip the loading plan. Ask: "What is the one thing you actually need to know or produce right now?" Route to the narrowest possible answer first. A loading plan is only useful once the destination is clear.

**When `certainty = exploring`:**
Use `compare_pack` mode: lock the primary route, then load one rival route or one contrastive boundary layer. Do not expand to `survey_pack` unless the user explicitly asks for breadth.

**When `certainty = certain`:**
Use `focus_pack` or `route_kernel`. Stop expansion as soon as the next asset would not change the answer quality.

**When `source = discover`:**
Prioritise possibility-expanding assets. Suppress loading of hard constraints and evaluation rubrics until the route is stable.

**When `source = construct`:**
Apply full loading discipline: lock route → choose mode → define mandatory bundle → define expansion queue → state stop conditions.

## References
- `wp.context-loading-plan`
- `references/background-bundles.json`
- `rb.context-loading-plan`
- `ka.bounded-context-loading`
- `ka.context-corrosion-signals`
- `ka.cross-protocol-referral-edges`
- `ka.reference-expansion-balance`
- `ka.scenario-factorization`
- `ka.story-memory-checkpoint`