Skip to content

AGENTS.md

Latest commit

 

History

History
97 lines (83 loc) · 3.38 KB

AGENTS.md

File metadata and controls

97 lines (83 loc) · 3.38 KB
Raw

CompText V7 Agent Instructions

Repo-local skill docs

  • Before starting work, read the relevant docs under docs/codex_skills/.
  • Use docs/codex_skills/mcp_context_layer.md for MCP context layer, CLI, prompt rendering, validation, and adapter work.
  • Use docs/codex_skills/artifact_validation.md for deterministic artifacts and regeneration checks.
  • Use docs/codex_skills/git_pr_workflow.md for branch sync, commits, pushes, and PR preparation.
  • Use docs/codex_skills/docs_positioning.md for README/docs positioning and scope boundaries.

Default safety rules

  • Repository state is the source of truth.
  • Inspect before editing.
  • Make the smallest safe patch.
  • Keep behavior deterministic and fixture-bound.
  • Maintain local sanitization and privacy boundaries (GDPR/DSGVO Art. 25) at all times.
  • Do not introduce semantic scoring, embeddings, vector DBs, external APIs, autonomous orchestration, or runtime tool execution unless explicitly scoped.
  • Prefer focused validation over broad checks unless the task requires broader validation.
  • Do not commit, push, create PRs, or merge unless explicitly requested.

Project focus

  • CompText V7 is the deterministic replay-integrity layer for compressed operational agent traces.
  • Current focus: core foundation, deterministic replay artifacts, CI artifacts, docs, and conservative positioning.
  • Showcase work belongs in the separate showcase repository and must not be reintroduced here unless explicitly approved.
  • Chilli/Hatch/Pet assets must not be touched unless explicitly requested.

Non-goals

  • No LLM judging.
  • No embeddings.
  • No vector DBs.
  • No external APIs.
  • No graph stores.
  • No runtime autonomous agent execution.
  • No benchmark logic changes unless explicitly requested.
  • No production-ready / clinical-grade / solved-memory claims.

PR discipline

  • Keep PRs small and focused.
  • Prefer docs-only PRs for positioning work.
  • Core logic PRs require tests.
  • Do not mix showcase, docs, and core refactors in one PR.
  • Use Draft PRs until CI and review threads are clean.
  • Do not mark ready until Actions are green and review threads are resolved/outdated.

Task shape

  • Use one concrete goal per task.
  • Name exact files or narrow paths that may change.
  • Keep instructions short and testable.
  • Split unrelated changes into separate PRs.
  • For test-only work, change only the target test file unless another file is required.
  • For fixture work, keep generation and validation in separate PRs.
  • If the task becomes unclear, leave a blocker comment instead of expanding scope.

Task template

Task:

  • One concrete change.

Allowed files:

  • Exact files or narrow paths.

Steps:

  • 3-6 explicit steps.

Tests:

  • Exact commands.

Done when:

  • Objective acceptance criteria.

Validation

  • Prefer cloud CI as source of truth.
  • Do not claim local validation unless commands actually ran.
  • Standard commands:
    • npm run layout
    • npm run typecheck
    • npm test
    • npm run check
  • For targeted TS/Python bridge tests, run the specific pytest file too.

Output format

Every agent PR summary must use: Summary: Changed files: Testing: Risks: Next:

Visibility requirement

  • Work incrementally.
  • Push after each meaningful change.
  • Comment after each push with: Progress:
    • changed files
    • what was fixed
    • tests actually run
    • remaining issues
    • current blocker, if any
  • If stuck, comment with current blocker instead of staying silent.