AgentsMesh Generation Contract

NEVER edit generated files (CLAUDE.md, AGENTS.md, .claude/, .cursor/, .github/copilot-instructions.md, and other tool outputs) — agentsmesh generate overwrites them. Edit the canonical source in .agentsmesh instead, or agentsmesh.yaml to change targets and features, then run agentsmesh generate.

Lessons (BLOCKING)

Graph .agentsmesh/lessons/lessons.json is canonical; never hand-edit it. Manual: lessons skill.

Recall: before every file edit or state-changing command, MUST run agentsmesh lessons query --file <path> --cmd <command> --session auto and obey matches; at task start, ALSO run agentsmesh lessons query --keyword "<task terms>" --always --session auto for conceptual + universal rules no path/command names. Pure-read commands and recall itself are exempt.

Capture: after any failure, user correction, regression, wrong assumption, useful surprise, repeated friction, or non-obvious fix, MUST self-critique and run agentsmesh lessons add "<imperative rule>" --topic <id> --trigger-file <glob> --evidence <sha|lesson-id>.

Before final: report Lesson: captured <id> or Lesson: none. No recall/capture gate = task incomplete. No shell: use lessons_query / lessons_add.

Operational Guidelines

Workflow

Planning

• Enter plan mode for any non-trivial task (3+ steps or architectural decisions). Write plan to tasks/todo.md.
• If something goes sideways, STOP and re-plan — don't keep pushing.
• Check in before starting implementation. Mark items complete as you go.

Subagents

• Use subagents liberally for research, exploration, and parallel analysis. One tack per subagent.

Verification

• Never mark a task complete without proving it works — run tests, check logs, demonstrate correctness.
• After every feature/story: use post-feature-qa skill (.claude/skills/post-feature-qa/).
• Ask yourself: "Would a staff engineer approve this?"

Elegance

• For non-trivial changes: pause and ask "is there a more elegant way?"
• Skip this for simple, obvious fixes — don't over-engineer.

Bug Fixing

• When given a bug report: just fix it. Find root cause, point at logs/errors, resolve. Zero hand-holding.

Skills

• post-feature-qa — Apply after every feature/story. Senior QA: verify edge-case coverage and story alignment.
• add-agent-target — Use when adding a new AI agent target. Covers research, implementation, fixtures, full test coverage, docs.
• add-global-mode-target — Use when adding or extending global mode (--global, .agentsmesh) for an existing target. Covers descriptor.global wiring, import/generate paths, reference rewriting, tests, and matrix docs.

Core Principles

• Simplicity: Make every change as simple as possible. Only touch what's necessary.
• No Laziness: Find root causes. No temporary fixes. Senior developer standards.

Project Rules

• Architecture baseline: Read docs/architecture/review.md before architectural or multi-file changes.
• Core flow to preserve: canonical .agentsmesh content -> descriptor-driven generation/import (src/targets/<id>/index.ts) -> shared reference rewrite/lock checks -> strict artifact verification.
• Scale limitations to account for: avoid target-name hardcoding in shared/core code, avoid duplicated per-target helper logic, and keep capability variance expressed in descriptors (not ad-hoc generator branches).
• Global mode discipline: treat global support as one cohesive contract (layout + capabilities + detection + scope extras), not scattered one-off hooks.
• Plugins are first-class targets: third-party plugins register full TargetDescriptors via registerTargetDescriptor(), and shared/engine code resolves targets with getBuiltinTargetDefinition(id) ?? getDescriptor(id) (never builtin-only). When editing a target, a descriptor hook contract (e.g. emitScopedSettings, capability/dispatch resolvers), or anything iterating targets, the change MUST hold for plugin descriptors too — update the tests/fixtures/plugins/rich-plugin reference fixture to the new contract and add a registered-plugin-descriptor test, not just builtin coverage.
• TDD mandatory: Write failing tests FIRST, then implement.
• Max file size: 200 lines. Split by responsibility if larger.
• No classes unless stateful: Prefer pure functions + types.
• No any: Use unknown + narrowing.
• Config source of truth: .agentsmesh directory. Generated files are artifacts.
• Test naming: {module}.test.ts colocated with source. Integration tests in tests/integration/.
• Generated artifact tests must be strict: Assert exact file paths, exact file counts, and exact referenced wrapper/script sets. No loose checks (some(...), prefix-only, "at least one").
• Commit format: conventional commits — feat|fix|test|refactor(scope): message
• Docs must stay current: Any change to CLI commands, flags, config schema, supported targets, or canonical file formats must be reflected in both README.md and the website (website/src/content/docs/) before the task is marked complete.
• Target data single source of truth — do not hardcode target lists or support levels outside this chain:
  1. src/targets/catalog/target-ids.ts (TARGET_IDS) = canonical target ID list. Each target's capabilities in src/targets/<name>/index.ts = feature support levels.
  2. src/core/matrix/data.ts (SUPPORT_MATRIX) = built dynamically from descriptors. Never hardcode.
  3. website/src/content/docs/reference/supported-tools.mdx = single docs page for per-target support. All other pages link here — no duplicate tables.
  4. README.md matrix must stay in sync with code capabilities.
  5. No hardcoded target counts or enumerations on homepage, CLI overview, or other pages. Use generic language and link to the matrix.
  6. Adding/changing a target: update target-ids.ts + descriptor → supported-tools.mdx → README.md matrix. No other docs pages should need changes.
• CLI display paths must use forward slashes: Any CLI output displaying file paths to users must normalize with .replaceAll('\\', '/') before printing. Tests assert forward-slash paths unconditionally; native node:path separators leak platform differences into output.
• Load-bearing contracts — verify before claiming any non-trivial change complete:
  • Plugins: every install path picks up registerTargetDescriptor() non-.md mappers via getAllRegisteredDescriptorIds() (tests/unit/install/importers/target-native-commands-plugin.test.ts).
  • Reference rewriting: (SPEC.md or equivalent) prose stays verbatim; real Markdown inline-link destinations rewrite cross-format (tests/unit/core/link-token-classifier-prose-vs-md-link.test.ts).
  • Install: third-party dir-readers route through readEntityDirWithMappers; the 79-repo sweep in docs/testing/install-compatibility-repos.md stays green.