Skip to content

Latest commit

 

History

History
54 lines (44 loc) · 2.74 KB

documentation.md

File metadata and controls

54 lines (44 loc) · 2.74 KB

Documentation ownership and maintenance

Current guides describe the checked-in runtime contract. Package availability, GitHub Actions state and PyPI approval are release-time facts; do not hardcode transient pipeline status into a user guide.

Content Canonical location
Product scope and first steps Paired README.md / README.zh-CN.md
Runtime/Skill installation INSTALL_AGENT.md
Agent decisions and integration boundary AGENT_INTEGRATION.md
CLI, Cases, Python and audit docs/user-guide.md
MCP configuration, all eight tools, paging and paths docs/mcp.md
Media consent, coordinates, OCR and video/decryption limits docs/media-perception.md
Capability contracts Registry; generated docs/capabilities.json; human capability map
Architecture/safety/security ARCHITECTURE.md, SAFETY.md, SECURITY.md
External dependencies DEPLOY-TOOLCHAIN.md
Tests and publishing docs/development.md, docs/agent-reliability.md, PUBLISHING.md
Investigation methodology and specialist references Separate DFTK-skill repository

Keep the two README languages equivalent. Detailed runtime guides are English-first; link rather than copying an entire secondary guide. Skill references must be self-contained after installation while linking to runtime documentation for installation details. Specialist guidance does not imply a matching parser exists.

Generated data

Regenerate the static manifest from the registry and overwrite the tracked file explicitly:

python -m dftk export-manifest --out docs/capabilities.json --force
python scripts/check_docs.py

Use the active development interpreter. Never commit runtime-specific readiness from --runtime-readiness as the static contract. Synchronize the companion manifest/catalog using its scripts/sync_capabilities.py; regenerate specialist descriptions with its scripts/generate_skill_catalog.py.

Review requirements

Validate command/parameter names against CLI help and registered contracts. Run representative examples against synthetic evidence with a separate workspace. Check local links, counts, safety classifications and referenced capability names. For documentation-only edits, use these focused checks instead of inventing tests that merely assert headings or duplicate prose.

Changelogs and versioned release notes are historical records; retain their original version-specific counts. Label old development plans as historical where they could be mistaken for current instructions. Update current capability limits, deployment assumptions and examples rather than mechanically replacing every old version string. Do not promise a security backport or an accuracy benchmark that the project does not actually maintain.