Skip to content

AGENTS.md

Latest commit

 

History

History
53 lines (37 loc) · 3.64 KB

AGENTS.md

File metadata and controls

53 lines (37 loc) · 3.64 KB

🔧 Start Here (for AI agents)

Before modifying code, load required context with node agents/scripts/load-context.mjs, then work inside the repo-native agent system under agents/. When a task spec exists, include it with node agents/scripts/load-context.mjs --task agents/specs/task-specs/<YYYY-MM-DD>-<slug>.md.

  • Memory Bank overview: agents/memory-bank.md
  • Workflows overview: agents/workflows.md
  • Tool catalog: agents/tools.md
  • Retrieval policy: agents/memory-bank.md#retrieval-policy is the canonical source for discovery tooling, one-pass note taking, and context tiers.

🧭 Mode Selection (Orchestrator vs One-off)

  • If the user does not specify a mode, ask them to choose orchestrator or one-off before proceeding.
  • Orchestrator: follow agents/workflows/orchestrator.workflow.md and the spec-first pipeline in agents/memory-bank/spec-orchestration.design.md.
  • Spec author: follow agents/workflows/spec-author.workflow.md when assigned a workstream spec.
  • Implementer: follow agents/workflows/implementer.workflow.md when executing an approved MasterSpec.
  • One-off: ask whether the request is one-off vibe or one-off spec; use agents/workflows/oneoff.workflow.md for the overview.
  • One-off vibe: no spec, small scope only; follow agents/workflows/oneoff-vibe.workflow.md and recommend switching modes if scope grows.
  • One-off spec: create a single spec using the required section schema, log approvals in the Decision & Work Log, then execute via agents/workflows/oneoff-spec.workflow.md.
  • Use user selection as the only threshold; do not infer mode from scope or complexity.
  • Reference agents/workflows/spec-author.workflow.md or agents/workflows/implementer.workflow.md when assigned a specific spec role.

🔑 Memory Bank

The Memory Bank provides durable, structured context for all tasks.

  • Overview and canonical policy: agents/memory-bank.md (source of truth for storage tiers and retrieval rules)
  • Canonical files live under: agents/memory-bank/ (PR-reviewed)
  • Operating model: agents/memory-bank/operating-model.md describes the default four-phase loop and expectations.
  • Task spec guide: agents/memory-bank/task-spec.guide.md shows the per-task spec structure.

Update Requirements (per task)

  • Update relevant canonical files under agents/memory-bank/ to reflect changes.
  • Validate and check quality:
    • npm run agent:finalize - format markdown, verify referenced paths exist across Memory/Workflow docs, and run repo quality checks.
  • Retrieval helpers: follow the Retrieval Policy in agents/memory-bank.md for the authoritative list of discovery commands (list-files-recursively.mjs, smart-file-query.mjs, read-files.mjs), numbered output expectations, and when --json is appropriate.

Convenience helpers

  • node agents/scripts/reset-active-context.mjs --slug <task-slug> [--title "..."] [--date YYYY-MM-DD] creates a new per-task spec.
  • node agents/scripts/append-memory-entry.mjs --requirements "..." --design "..." --implementation "..." --execution "..." is deprecated; it prints a reflection entry for manual copy into the task spec.

🧭 Workflow Process List

One LLM executes work by following process markdowns in agents/workflows/.

  • Phases: Requirements → Design → Implementation Planning → Execution (default)
  • Each process file defines checklists, inputs/outputs, and quality gates.
  • The LLM loads the process .md (+ linked partials), executes the current phase, writes artifacts, updates phase state, advances when gates pass, and logs a short reflection to the Memory Bank.

IT IS VERY IMPORTANT YOU STRICTLY FOLLOW AGENT WORKFLOWS

Start with: agents/workflows/oneoff.workflow.md