Context budgets
Kiyo’s main performance concern is context: how much instruction text an agent must read before it can act. ADR-002 sets ceilings, and the static contract tests (group G14) enforce them. The framework states these are “not vendor-imposed context limits”, and they are not token measurements.
Method
Section titled “Method”The counts below were measured for this documentation at framework revision 897516e using the framework’s own definition:
- Lines: physical lines, including blank lines, headings, and frontmatter, after normalizing CRLF to LF.
- Words: whitespace-delimited tokens.
This is a static text measurement. It says nothing about model tokenization, host truncation, or actual read behavior, which the framework lists as untested.
Bootstrap chain
Section titled “Bootstrap chain”| File | Lines | Words |
|---|---|---|
src/kiyo/KIYO.md |
22 | 112 |
src/kiyo/framework/bootstrap.md |
59 | 469 |
| Combined | 81 of 120 | 581 of 600 |
The combined bootstrap is within its ceiling, but close on words: 19 words of headroom.
Skill entries
Section titled “Skill entries”Ceiling: 250 lines and 1,200 words per rendered SKILL.md, including native frontmatter. Packaged entries are the canonical text plus rewritten link paths and a short “native guidance” suffix. The Claude, Codex, and Copilot renderings measured identically.
| Skill | Canonical lines | Canonical words | Packaged lines | Packaged words |
|---|---|---|---|---|
| init | 73 | 566 | 78 | 584 |
| requirement | 69 | 567 | 74 | 585 |
| implement | 93 | 789 | 98 | 807 |
| review | 73 | 583 | 78 | 601 |
| test | 89 | 692 | 94 | 710 |
| security | 80 | 615 | 85 | 633 |
| architecture | 71 | 515 | 75 | 533 |
| memory | 74 | 609 | 79 | 627 |
The largest entry, Implement, uses 39% of the line ceiling and 67% of the word ceiling.
Project adapter block
Section titled “Project adapter block”Ceiling: 250 words for the Kiyo-owned managed block, excluding surrounding human instructions. The framework recorded sample renderings of 114 (Claude), 108 (Codex), and 116 (Copilot) words.
What happens after the bootstrap
Section titled “What happens after the bootstrap”Everything else loads progressively. The agent reads the selected procedure and only the references and single template the task needs. The 97-file shared snapshot in each Skill is installed size, not loaded context. Near a context limit, the agent produces a handoff (intent, facts, decisions, unknowns, checks, next step) rather than continuing on partial context.
Reproducing the measurement
Section titled “Reproducing the measurement”From the framework repository root:
python -c "from pathlib import Path; t=lambda p: Path(p).read_text(encoding='utf-8').replace('\r\n','\n'); f=lambda p: (t(p).count('\n') + (0 if t(p).endswith('\n') else 1), len(t(p).split())); print(f('src/kiyo/KIYO.md'), f('src/kiyo/framework/bootstrap.md'))"