AGENTS.md · diff

git:20260602.16ca2bf to git:20260916.5146276

45 added, 58 removed. Audit A to A.

- # AGENTS.md
-
- This repository uses agents-stack v3 — a Goal-QA-Driven development harness with three verification checkpoints.
-
- ## Key Paths
-
- - `.agents-stack/workstream/<ws-id>_{YYYYMMDD}/` — workstream artifacts (goal.md, spec.md, plan.md, arch-report.md, tasks.md, report.md, handoff.md, qa-report.md, changelog.md, status.json)
- - `.agents-stack/tracked-work.json` — active/parked workstream registry
- - `.agents-stack/reference/architecture.md` — pipeline, checkpoint system, project architecture
- - `.agents/skills/using-agents-stack/` — orchestrator + phase SKILL.md files
-
- ## Core Invariants
-
- 1. **Files beat chat memory.** Durable state lives in `.agents-stack/`. Chat is not state.
- 2. **One active workstream.** Park or complete before starting another.
- 3. **Generator ≠ Auditor.** All three checkpoint workers (verify-architecture, analyze, qa) must be separate agent instances from the phases they verify. The agent that designs must not verify.
- 4. **Cold start must work.** A new agent recovers from files alone — no chat history needed.
- 5. **Iteration ≠ Retry.** Retry fixes execution within the same contract. Iteration questions the premise; go back to spec/plan.
- 6. **Three-checkpoint rhythm.** Every ~2 phases, a verification gate isolates risk before it compounds. Checkpoint #1: Architecture vs Goal. Checkpoint #2: SPEC×PLAN×TASKS consistency. Checkpoint #3: CODE vs REALITY. Skip a checkpoint = errors locked into the foundation. Backtrack cost: 1x vs 5-8x.
-
- ## Safety / Do Not Do
-
- - Do NOT skip blocking gates. If `status.json.blocking_gate` is set and unmet → STOP, do not advance.
- - Do NOT self-review. The agent that produces an artifact (spec, plan, code) must not verify that same artifact. Dispatched checkpoint workers enforce this — do not override.
- - Do NOT modify `.agents-stack/reference/` outside of the release phase or an explicit `/update-reference` command.
- - Do NOT change requirements by editing code. Update spec.md first, re-derive downstream artifacts.
- - Do NOT bypass plan when architecture issues surface. Update plan.md, then re-derive tasks.
- - Do NOT write implementation code as the orchestrator. Route, dispatch workers, verify results — never implement.
+ # Agent Workspace Index
- ## Quick Resume
+ Use this file as the first-stop index. Open only the section and linked file
+ needed by the current task; do not load every skill into context.
- 1. Read `CONSTITUTION.md`, `AGENTS.md`, `.agents-stack/tracked-work.json`
- 2. Read `.agents-stack/workstream/<ws-id>_{YYYYMMDD}/status.json` and strongest artifact
- 3. Load the phase SKILL.md from `.agents/skills/using-agents-stack/<phase>/SKILL.md`
- - If the loaded skill is `using-agents-stack` (orchestrator): route only, do NOT implement
- 4. Verify checkpoint matches disk state; continue from strongest valid checkpoint
+ ## Decide
- Success: a cold-start agent reads these files and resumes safely without chat history.
+ - [Meta Thinking Framework](meta-thinking-framework/SKILL.md) — six-lens analysis for complex decisions.
+ - [Interview to Build](interview-to-build/SKILL.md) — product discovery before implementation.
- ## Pipeline
+ ## Specify
- ```
- goal → spec → plan → [CHECK #1: Arch vs Goal] → tasks → [CHECK #2: ANALYZE] → implement → [CHECK #3: QA] → release
- ```
+ - [System Prompt SOP](system-prompt-sop/SKILL.md) — route prompt constraints to enforcement layers.
+ - [Design Context](design-context/SKILL.md) — define visual and UX context before generation.
+ - [Plain Sales Copy](plain-sales-copy/SKILL.md) — concise business copy with no puff.
+ - [Marketing Director](marketing-director/SKILL.md) — brief-first marketing and multimedia planning.
- Each phase has a SKILL.md defining its contract, output format, verification gates, and handoff protocol. **Load it before executing the phase.** Skipping the skill = working without the spec.
+ ## Build
- Full pipeline documentation including checkpoint mechanics, risk isolation principle, and cost table: `.agents-stack/reference/architecture.md`.
+ - [Agentic Dev E2E](agentic-dev-e2e/SKILL.md) — feature or bugfix from spec to evidence.
- ## Commit & PR
+ ## Meta Skills
- ### Commit Style
+ - [Skill Creator](skill-creator/SKILL.md) — create, evaluate, and package skills.
+ - [Skill Evolution](skill-evolution/SKILL.md) — evolve skills from real usage.
+ - [Conversation to Skills](conversation-to-skills/SKILL.md) — turn corrections into versioned skills.
+ - [Interview to Personas](interview-to-personas/SKILL.md) — interview users and maintain persona memory.
- Conventional commits: `type(scope): description`. Types: `feat`, `fix`, `refactor`, `docs`, `chore`. Scope is the phase or component (e.g., `qa`, `plan`, `AGENTS.md`). Use the commit body for rationale — what changed and why, not just what.
+ ## Learning
- ### Before Committing
+ - [Learn Anything About X](learn-anything-about-x/SKILL.md) — adaptive domain mentorship.
- - Inspect `git status` + `git diff` — stage only intended files, never commit secrets
- - Verify all checkpoint gates for the current workstream have passed
+ ## Agent Runtime
- ### Before PR
+ - [OpenRouter Adapter](agent-plugin/scripts/openrouter.py) — shared provider adapter for media and text skills.
+ - [Runtime Doctor](agent-plugin/scripts/doctor.py) — prerequisite check.
+ - [Generate Image](agent-plugin/generate-image/SKILL.md)
+ - [Generate Video](agent-plugin/generate-video/SKILL.md)
+ - [Generate Text](agent-plugin/generate-text/SKILL.md)
+ - [Generate Audio](agent-plugin/generate-audio/SKILL.md)
+ - [Read Image](agent-plugin/read-image/SKILL.md)
+ - [Read Video](agent-plugin/read-video/SKILL.md)
+ - [Read Audio](agent-plugin/read-audio/SKILL.md)
+ - [Agent Memory](agent-plugin/agent-memory/SKILL.md) — five-duty memory protocol.
- - Cold-start test: a fresh agent reading only `CONSTITUTION.md`, `AGENTS.md`, and `.agents-stack/` can resume the workstream
- - Generator ≠ Auditor: the reviewer must not be the agent that implemented the code
+ ## Memory Drawers
- ## Intent Routing
+ - [Working](agent-plugin/templates/memory/working/INDEX.md) — current task desk.
+ - [Semantic](agent-plugin/templates/memory/semantic/INDEX.md) — stable facts.
+ - [Episodic](agent-plugin/templates/memory/episodic/INDEX.md) — experience with lessons.
+ - [Procedural](agent-plugin/templates/memory/procedural/INDEX.md) — verified methods and skill pointers.
+ - [Personas](agent-plugin/templates/memory/personas/INDEX.md) — user collaboration protocol.
- Pipeline enforcement is binary, controlled by `status.json.state_machine`:
+ ## Project Operations
- | `state_machine` | Behavior |
- |-----------------|----------|
- | `"on"` (default) | Linear pipeline enforced. All requests route through orchestrator. Phases execute in strict order — no skipping, no reordering, no bypass. |
- | `"off"` | No pipeline interference. Execute ad-hoc. |
+ - [README](README.md) — repository overview, installation, and runtime commands.
+ - Tests live under `agent-plugin/tests`; run them with:
- Toggle via `/state-machine on|off` or natural language "turn off state machine" / "stop using pipeline".
+ ```bash
+ pytest agent-plugin/tests
+ ```
- For skill-specific intents (code review, frontend QA, design review, adversarial QA, complexity audit, reflect/learn), see the Contextual Skill Resolver in your platform's AGENTS.md.
+ Generated runtime state belongs under `.agents/`. Do not treat memory entries,
+ generated artifacts, or credentials as project source.