agent_infra

Imported from bigsexyaloe/agent_infra (AGENTS.md). Install upstream with npx skills add bigsexyaloe/agent_infra. Copyright stays with the author.

AGENTS.md — Agent Infrastructure Rules

This file is automatically read by Qwen Code at session start.
It defines global rules that apply to ALL agents in this project.
Context: Fintech compliance environment — 152-ФЗ, ФСТЭК, ISO 27001.

Pipeline Overview

This project uses a two-level mandatory pipeline for all feature development.
Pipeline level (secure or complex) is determined by the Manager in Phase 0, not by the requester.

Phase 0 — Pre-Pipeline (interactive, user-facing)

Feature Request → [MANAGER] — clarify → determine complexity → write spec → spec_ready

/pipeline cannot start until Manager sets Status: spec_ready.
No user input is accepted once /pipeline starts.

Secure Pipeline (default for all fintech features)

spec_ready → [PROGRAMMER] → [REVIEWER] → [SECURITY] → [TESTER] → Done ✅

Complex Pipeline (new subsystems, architectural changes, breaking API)

spec_ready → [ARCHITECT] → [PROGRAMMER] → [REVIEWER] → [SECURITY] → [TESTER] → Done ✅

Return Loops

[REVIEWER]  ──REQUEST_CHANGES──▶ .qwen/tasks/returned/<ID>-review.md
[SECURITY]  ──FAIL──────────────▶ .qwen/tasks/returned/<ID>-security.md
[TESTER]    ──FAIL──────────────▶ .qwen/tasks/returned/<ID>-bug.md
                    └── all return via: /return programmer <TASK-ID>

Agent Roster

Commands

Task File Locations

Task Status Lifecycle

new
 │
 ▼
clarifying  ◀──── (Manager asks questions, user responds)
 │
 ▼
spec_ready  ────────────────────────────── /pipeline starts here
 │
 ├─▶ arch_conflict ◀── [ARCHITECT: spec contradicts architecture]
 │        │
 │        └──▶ spec_ready  ◀── [MANAGER: spec revised]
 │
 ▼ (complex only)
arch_ready
 │
 ▼
in_development  ◀──────────────────────────────────────────────────────┐
 │                                                                      │
 ├─▶ spec_question ◀── [PROGRAMMER: spec technically impossible]        │
 │        │                                                             │
 │        └──▶ spec_ready / arch_ready  ◀── [MANAGER: question answered]│
 │                                                                      │
 ▼                                                                      │
code_review  ◀──── review_returned ◀── [REVIEWER: REQUEST_CHANGES]     │
 │                                                                      │
 ▼                                                                      │
review_passed                                                           │
 │                                                                      │
 ▼                                                                      │
security_audit  ◀── security_failed ◀── [SECURITY: FAIL]               │
 │                                                                      │
 ▼                                                                      │
security_cleared                                                        │
 │                                                                      │
 ▼                                                                      │
in_review  ◀──── qa_returned ◀──── [TESTER: FAIL] ─────────────────────┘
 │
 ▼
done ✅

Pipeline Selection

Complexity is set by Manager in Phase 0 — never by the requester.

There is no simple or standard pipeline in this environment.
Every feature goes through security audit and code review — no exceptions.

Compliance Standards

Standards are loaded from .qwen/standards/security/ and .qwen/standards/architecture/.

Manager activates standards in the task's Meta table. Security agent reads them in Step 0.

Artifact Gates (technical enforcement)

Every pipeline agent hard-checks its required artifacts by reading .qwen/pipeline.json before doing any work.
If a required file is missing or the task status is wrong — the agent stops and prints an error. It does not proceed.

This is not policy — it is a hard block enforced by each agent's Artifact Gate section.

Global Rules (all agents must follow)

1. Always read .qwen/pipeline.json — it is the source of truth for artifact requirements.
2. Always run your Artifact Gate first — no exceptions.
3. Always use the provided templates — no ad-hoc formats.
4. Always print the handoff message at the end of your work.
5. Never skip your checklist — quality gates are mandatory.
6. Never implement outside the spec — out-of-scope work is rejected.
7. Respect role boundaries — testers don't fix code; managers don't write code; architects don't implement.
8. Task IDs are immutable — do not rename or change the TASK-ID once assigned.
9. Reviewer and Security are HARD GATES — tester cannot run until both report passing verdicts. No exceptions.
10. Security findings are non-negotiable — CRITICAL and HIGH always block. MEDIUM is tracked. No severity may be downgraded.
11. Manager Phase 0 is interactive — once /pipeline starts, no user interaction is expected or accepted.

Return Limits (escalation thresholds)

LLM Configuration

Models are configured in .qwen/settings.json.
Per-role model overrides are in the agentConfig section.
Use /model inside a session to switch models on the fly.

Naming Conventions

• Task IDs: FEAT-001, FEAT-002, … (increment, never reuse)
• Bug IDs within a report: BUG-001, BUG-002, …
• Review findings: RV-001, RV-002, …
• Security findings: SEC-001, SEC-002, …
• Acceptance criteria: AC-01, AC-02, …
• Return iterations: increment Return # in each impl/bug/review/security report

Common Commands (project)

# Install Qwen Code
npm install -g @qwen-code/qwen-code@latest

# Configure LLM
# Edit .qwen/settings.json — add your API key and set model

# Start a session
qwen

# Create a new feature task
cp .qwen/templates/feature-request.md .qwen/tasks/active/FEAT-001.md
# Fill in: title, description, problem, rough acceptance criteria
# Do NOT set Complexity or Standards — Manager sets these in Phase 0

# Phase 0: clarify and spec (interactive)
/run manager FEAT-001

# Run the pipeline (no user input from here)
/pipeline FEAT-001

# Check status
/status
/status FEAT-001