Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/vgmakeev/katana/agents-mdgit clone --depth 1 https://github.com/vgmakeev/katanaWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00397 | $0.00397 |
| Opus 5 | $0.00198 | $0.00198 |
| Sonnet 5 | $0.00079 | $0.00079 |
| Haiku 4.5 | $0.00040 | $0.00040 |
Grade A, and why
katana AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
Agent instructions
Katana is an agent-first, deterministic polyglot refactoring executor.
Runtime
- Use Python 3.14.
- Install with
uv sync --all-groups. - Run the complete gate with
make check. - Run
make markdown-checkwhen the Node bridge changes.
Boundaries
- Core owns
DocumentIR, plans, conformance, apply orchestration, and the CLI. - Language behavior belongs in
katana.adapters.<stack>or an external package registered underkatana.adapters. - Formatter, lint-fix, and checker behavior belongs in
katana.toolchains.<stack>orkatana.toolchainsentry points. - Built-in adapters use the same public contracts as third-party adapters.
- Command adapter protocol v1 is analysis-only and must fail rewrite closed.
Compatibility
- Preserve split-plan v1 and existing
schema,outline,assess,plan,apply,large, andprojectJSON behavior. - Keep legacy adapter ids
node_typescript,dart_analyzer, andnode_postcss. - Existing public functions remain compatibility facades over adapters.
- Add conformance and regression tests for protocol changes.
Agent contract
- Never include source bodies in
DocumentIRor other agent maps. - Keep commands non-interactive and machine-readable.
- Prefer exact hashes and byte spans over parser-specific node objects.
- Analysis-only support is useful; do not claim rewrite capabilities without dependency closure, import handling, formatting, and validation.
- Preserve source bytes during mechanical moves. Semantic changes are separate.
Source budget
- Keep source files within 500 physical and language-aware code lines.
- The hard repository ceiling is 500 lines.
make size-checkenforces both measures across runtime, tests, and examples.- Split by cohesive responsibility before raising the ceiling or adding exclusions.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 46 lines · 397 tokens per session scan A d90329bc4940
katana AGENTS.md is an instructions file published in the GitHub repository vgmakeev/katana (12 stars, last pushed 12d ago), licensed Apache-2.0. It adds 397 tokens to every session, about $0.0020 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
jcodemunch-mcp CLAUDE.md
Claude Code instructions for jgravelle/jcodemunch-mcp, covering jcodemunch-mcp — project brief, current state, key files, cli subcommands and architecture notes.
ariadne AGENTS.md
Instructions for CRJFisher/ariadne, covering guidelines, project layout / intention tree, refactoring ethos, documentation style: canonical and self-contained and scope.
ariadne CLAUDE.md
Instructions for CRJFisher/ariadne, covering guidelines, goals, project layout / intention tree, refactoring ethos and processing pipeline.
jcodemunch-mcp AGENTS.md
AGENTS.md instructions for jgravelle/jcodemunch-mcp, covering code exploration policy, session-aware routing and model-driven tool tiering.
coldstart CLAUDE.md
Instructions for AkashGoenka/coldstart: Lightweight navigation layer for AI agents. Answers one question: which files are relevant to this task? Fast static index over file paths, symbol names, and exports — queried via two operations, find (ranked file paths matched against declared names: filenames, path segments…
codemap AGENTS.md
Instructions for stainless-code/codemap, a project described as: Cut AI-agent token waste 90% — query a local SQLite structural index of your JS/TS/CSS codebase with SQL in one round-trip instead of 3–5 file reads. Symbols, imports, calls, components, CSS tokens, coverage, markers. CLI, MCP (21 tools), HTTP, GitHub…