Skip to content

Latest commit

 

History

History
85 lines (62 loc) · 2.86 KB

File metadata and controls

85 lines (62 loc) · 2.86 KB

AGENTS_FOCUS.md

This document defines the current agent-first workflow for this repository.

It does not replace AGENTS.md. AGENTS.md remains the policy file for worktrees, merge rules, deployment verification, and cleanup.

Default Workflow

The default loop is:

  1. Read strategy_spec.toml.
  2. Read .focus/context.json.
  3. Stay inside the editable surface unless there is a concrete reason to expand.
  4. Refresh generated focus assets when workflow intent, pack scope, or shared navigation changes.
  5. Run verification in this order:
    • focus.smoke
    • focus.full only when the smoke profile is insufficient
    • runtime or backtest workflows only when the task truly needs execution evidence

The repository no longer relies on an installable command-line wrapper. Treat the focus manifest and generated assets as the source for module-level runtime and verification entrypoints.

Source Of Truth

Use these assets in order:

  1. strategy_spec.toml
  2. .focus/context.json
  3. .focus/SYSTEM_MAP.md
  4. .focus/TASK_BRIEF.md
  5. .focus/WORKFLOWS.md
  6. .focus/TASK_ROUTER.md
  7. .focus/TEST_MATRIX.md
  8. tests/TEST.md
  9. artifacts/validation/latest.json

If generated assets drift from the spec, update the generator/source layer and refresh the agent assets instead of hand-maintaining the generated files.

Editing Boundaries

  • editable
    • default edit surface
  • reference
    • context only unless the task truly requires expansion
  • frozen
    • do not edit unless the task is explicitly about repository policy or asset generation

When you must expand scope:

  1. Explain why the editable surface is insufficient.
  2. Expand by the smallest possible step.
  3. Record the expansion in the delivery summary.

Verification Profiles

  • focus.smoke
    • default verification profile
    • runs the current focus selectors with the smoke keyword filter
  • focus.full
    • full runnable selector set for the current focus
  • runtime
    • use only for runtime lifecycle or monitoring work
  • backtest
    • use only when behavior or parameter effects need execution evidence

Structured evidence is written to artifacts/validation/latest.json and other refreshed asset files. Do not claim completion from code inspection alone.

Anti-Patterns

Do not:

  • bypass strategy_spec.toml when changing intent or pack scope
  • hand-edit generated .focus/* files unless the task is about the generator
  • widen scope without a concrete reason
  • reintroduce command-wrapper abstractions around runtime or verification flows
  • move domain or infrastructure responsibilities into generic coordinators

Delivery Checklist

Every delivery should state:

  1. which source-of-truth assets were consulted
  2. which surface was edited
  3. which verification profiles or execution workflows ran
  4. whether tests/TEST.md or artifacts/*/latest.json changed
  5. any remaining risks or skipped checks