Activation and loading
What it is. The rules for how Kiyo’s instructions reach the agent, and how honestly that is reported. Defined in framework/activation-contract.md (KIYO-ACT-001) and framework/context-loading.md (KIYO-LOAD-001, KIYO-LOAD-002).
Why it exists. “The plugin is installed” is often mistaken for “the agent follows the rules”. Kiyo treats them as separate facts with separate evidence. It never adds a hook, runtime, or MCP server to simulate always-on loading.
Four activation capabilities
Section titled “Four activation capabilities”| Capability | What establishes it | What it does not establish |
|---|---|---|
| Explicit Skill invocation | You select an available native Skill entry with the host’s evidenced syntax | Automatic Core loading, successful execution, or approval of every action |
| Implicit Skill selection | An observed host selection for a particular prompt and configuration | Explicit invocation, or reliable selection for future prompts |
| Automatic Core loading | Target-specific evidence that a native facility loaded the Core | That all sessions or targets load it, or that the agent obeyed every control |
| Agent-directed Core read | An actual read of the packaged bootstrap after selection | Native automatic loading or permanent installation |
Evidence labels used throughout the documentation:
| Label | Meaning |
|---|---|
DOCUMENTED_ONLY |
Supported by checked native documentation, not a Kiyo test |
NOT_TESTED |
No live test of that capability on the stated target |
UNKNOWN |
Insufficient or conflicting evidence |
UNSUPPORTED |
An applicable source explicitly establishes the limit |
VERIFIED |
An actual recorded observation, limited to the conditions tested |
NOT_REVALIDATED |
A freshness qualifier for a source not rechecked |
The loading order
Section titled “The loading order”The logical order is bootstrap → selected Skill procedure → relevant references. Host discovery happens first. A host may open SKILL.md before the agent can follow its instruction to read the bootstrap; static Markdown cannot change that order.
sequenceDiagram accTitle: Kiyo loading sequence accDescr: The host discovers Skill metadata, the user selects a Skill, the agent reads SKILL.md, then KIYO.md and the bootstrap, then the selected procedure and only relevant references and one template, then produces output. participant U as User participant H as Host agent participant S as SKILL.md participant C as KIYO.md + bootstrap participant R as Relevant references H->>H: Discover name + description U->>H: Select Skill and state intent H->>S: Read entry (modes, access contract) S->>C: "Read KIYO.md and its bootstrap before workflow actions" C->>R: Selected procedure, then only needed references R-->>H: One template for the report H-->>U: Scoped work + chat report
Links are locators, not imports. The bootstrap does not recursively load every file it links. A missing or denied reference is reported with its path, and dependent actions stop. The agent never fetches a replacement Core from the network or runs a generator.
Context budgets
Section titled “Context budgets”These are Kiyo’s authoring ceilings (KIYO-LOAD-002), not vendor limits or token counts. Lines count blanks and frontmatter; words split on whitespace.
| Material | Ceiling | Measured at revision 897516e |
|---|---|---|
KIYO.md + framework/bootstrap.md |
120 lines, 600 words | 81 lines, 581 words |
Each rendered SKILL.md |
250 lines, 1,200 words | largest (Implement): 98 lines, 807 words |
| Kiyo-owned project instruction block | 250 words | samples recorded at 108–116 words |
Per-Skill measurements are on Context budgets.
Persistent project guidance
Section titled “Persistent project guidance”Installing a package exposes metadata; it does not make the Core always-on. If you want every session in a project to point at Kiyo state, an authorized Init can add a small managed block to the host’s own instruction file:
| Host | Project instruction file used for the block |
|---|---|
| Claude Code | Existing CLAUDE.md or .claude/CLAUDE.md (a plugin-root CLAUDE.md is not project context) |
| Codex CLI | The applicable AGENTS.md chain; AGENTS.override.md and global config are never edited |
| GitHub Copilot | .github/copilot-instructions.md, AGENTS.md, or path-specific instructions |
The block has <!-- KIYO:BEGIN project-context --> and <!-- KIYO:END project-context --> markers, an adapter revision (init-locator-1, a text-format revision, not a product version), and project-relative locators. It stays within 250 words and never copies the Core. After insertion it is your file content. Plugin updates and uninstall never modify or remove it.
Whether any host loads that block automatically in a fresh session is NOT_TESTED.
Related
Section titled “Related”- Architecture: Loading and activation
- Operations: Compatibility and evidence
- Security: self-check submode in Security