tencentmusic/modus

Modus Development and Writing

This is a composite skill serving the Modus framework itself. It provides a single entry point connecting global cognition, module development, testing & release, and article writing; module guides are loaded according to the task, without stuffing the entire reference directory into the context at once.

Determine the Current Project First

The source location is determined by the user-specified checkout; if none is specified, the current Git repository root is used. This skill's directory is determined by the location of the loaded SKILL.md. On first entry into a checkout, set SKILL_DIR to this skill's directory and run a read-only diagnostic from the Modus repository root:

python3.11 "$SKILL_DIR/scripts/inspect_repository.py" --root .

It checks branch, HEAD, upstream, local modifications, package version, and knowledge baseline drift. It does not connect to the network, does not write to the repository, and does not call an installed Modus. Use any Python 3.11+ interpreter; replace python3.11 in the example with the actual environment.

Knowledge snapshot comes from modus-open / 49c51621f378a58d6724eedde4657d9505346d45, distribution package openmodus 0.0.4. Different branches, versions, or file changes mean relevant guides need to be cross-checked against the current implementation; do not automatically switch branches, upgrade a personal CLI, or generate project knowledge.

First read the Global Overview, then select a guide from the table below. If you are already familiar with the project and the task is clear, you can go directly to the relevant module.

Task Routing

Current Task	Read
Feature development, bug fix, refactoring, source code review	Development Process + corresponding module + Verification & Release
Help, parameters, exit codes, Chinese language experience, personal package maintenance	CLI & Maintenance Entry
Integration, upgrade, uninstall, manifest, asset conflicts, Git protection	Project Installation & Ownership
CodeBuddy / Claude / Cursor / Codex integration, activation hooks	Platforms & Hooks
Source discovery, freezing, parsing, compiler, language extensions, cache hits	N1–N2: Source & Observation
Identity resolution, call relationships, topology, entry to effect, capability traversal	N3–N4: Resolution & Capability
Scope candidates, clustering suggestions, evidence packets, context, cursor, summary	N5–N6: Projection & Delivery
Current project context, knowledge routing, ranking, source fallback, budget	Retrieval & Orchestration
Markdown identity, CAS, knowledge updates, anchoring, completeness & publication verification	Knowledge Writing & Verification
init/reinit/update/sync behavior, generated body text, scope confirmation, templates	Knowledge Authorship Workflow
Multi-repo workspace, member addressing, outer navigation, isolation	Workspace
wheel, source distribution, dependencies, CI, test selection	Verification & Release
Articles, reports, technical introductions, presentation outlines, architecture diagrams	Source Evidence & Writing; cross-reference corresponding modules for related mechanisms
Verifying iwiki, old materials, or this skill's basis	Sources & Differences

Most Commonly Misunderstood Boundaries

• Modus installation/integration, initialization evidence, and long-term Markdown knowledge have different owners. modus init installs assets; /modus-init is investigated and authored by the assistant.
• Ordinary retrieval reads current source code and formal Markdown. The initialization analysis cache does not become the factual source for daily operations; new capabilities should not import machine into the normal CLI path.
• N1–N6 are deterministic evidence processing; K1–K6 are the assistant's semantic authorship. Capability composition does not call an LLM; candidate and representative paths are not equivalent to formal domains or complete execution chains.
• Installation/update transactions and knowledge updates have different failure semantics. Knowledge updates commit per file and report partial success; they cannot promise full rollback.
• .modus/ and host copies are managed integration assets; modus/knowledge/ is long-term knowledge; manual constraint/ and references/business/ have independent protection.
• clean only proves the completeness or publication promises covered by the checker; it does not prove that the body text is true, complete, or clear, nor does it prove runtime success.
• Commands, JSON fields, status codes, stable IDs, and Chinese display are maintained separately. Templates are product behavioral contracts shipped with the package.

Completing the Task

Development results should state specific behavior, modification owners, actual verification, and unverified boundaries. Writing results should separate current implementation, design goals, historical materials, and evidence of effect. Referencing this skill can accelerate location, but implementation conclusions must be cross-checked against the current source code.

This skill supports subsequent development but does not grant permission to publish, commit, push, write to external documentation, or delegate authority; those actions are governed by the current user's task authorization. Do not automatically rebuild knowledge, update the skill, or modify unrelated assets as a result of using this skill.