Community编程与开发github.com

hava-idea/fy-flow

适用于 AI 多 Agent 与人类团队协同研发的多角色状态流转、质量审计、防错闭环与看板自动化工作流技能包。可通过 /fy-flow help、/fy-flow start、/fy-flow status、/fy-flow kanban、/fy-flow trace 快捷指令或自然语言唤醒。

fy-flow 是什么？

fy-flow is a Codex agent skill that 适用于 AI 多 Agent 与人类团队协同研发的多角色状态流转、质量审计、防错闭环与看板自动化工作流技能包。可通过 /fy-flow help、/fy-flow start、/fy-flow status、/fy-flow kanban、/fy-flow trace 快捷指令或自然语言唤醒。.

兼容平台~Claude Code✓Codex CLI~Cursor
Part ofAgent Workflows
npx skills add hava-idea/fy-flow

Installed? Explore more 编程与开发 skills: steipete/bluebubbles, steipete/eightctl, steipete/blucli · View all 6 →

查看原文→浏览所有技能

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

OpenAIChatGPTClaudeClaudeGeminiGeminiGrokGrokPerplexityPerplexityDeepSeekDeepSeek

文档

FY MultiAgent Flow Skill (FY-Flow)

Contract-driven Multi-Agent Collaborative Development Microkernel (FY-Flow) — Aligned with the AutoGen paradigm, effectively preventing unauthorized actions, fragmented outputs, and state suspension.

Unified Gateway Protocol

All calls to this Skill are uniformly converged to this gateway entry point, and are split and evaluated according to the following two-branch rule:

1. Two-Branch Diversion Matrix

Branch TypeTrigger Features & CommandsProcessing Mechanism & Response ContractTypical Scenarios
Branch A: System Console Operations/fy-flow help/fy-flow start/fy-flow status/fy-flow kanban/fy-flow traceDirect pass-through to the underlying Ops Engine (0 pre-level gating), execute the corresponding script directly and output system status.View help manual, initialize SOP, start kanban, dashboard health check, offline trace map
Branch B: Business R&D & Interaction IntentNatural language requirements, code modifications, feature development, bug fixes, troubleshooting, /fy-flow autoMandatory pre-execution of [Three-Level Classification and Single-Responsibility Decomposition], and explicitly output dual header identifier on the first line of the response:`[Task Level: L0 Instant Q&A / L1 Short-Chain Task / L2 Standard R&D][Single Responsibility: Compliant Single Card / Compound Requirements must be decomposed into N atomic tasks]`

2. Business Requirement [Three-Level Classification] Decision Flow

[Business Requirement Input]
   |
   +---> 1. Will it result in modifications to code/documentation/config/other repository files?
   |        +-- No ---> 2. Does the conclusion need to be traced for subsequent engineering basis?
   |        |            +-- No ----------------> [L0 Instant Q&A] (No card, answer directly, 0ms CLI latency, draft saved to drafts)
   |        |            +-- Yes ---------------> [L1 Lightweight Task] (Create card, stop and wait for start -> wait for human authorization, single role completes)
   |        |
   |        +-- Yes ---> 3. Does it require multi-role collaboration, modification of core business code, or changes to common base libraries?
   |                        +-- No ---------------> [L1 Lightweight Task]
   |                        +-- Yes ---------------> [L2 Standard R&D Task] (Create card, stop and wait for start -> wait for human authorization, multi-role collaboration)
  • Gateway Classification & CCP Elastic Classification Mapping:
    • [L0 Instant Q&A]: No card, answer directly, 0 state machine overhead;
    • [L1 Lightweight Task]: Mapped to CCP Tier-3 (Fast-Path quick release, no physical independent report, credentials written directly back to the kanban);
    • [L2 Standard R&D]: Automatically matches CCP Tier-2 (local optimization, effort ≤ 4.0h, no physical report) or CCP Tier-1 (core architecture & interface contract strong constraints) based on scale and contract.
  • Any business requirement falling under L1/L2 must undergo a mandatory single-responsibility review (effort ≤ 8.0h, single atomic deliverable). Cross-orthogonal domain requirements must be decomposed into separate cards.
  • Hard Red Line for Start Authorization: Card creation marks the end of the current planning session. The task card must strictly remain in the [To Do] status; without explicit authorization from the human user (e.g., "Start", "Begin development", "Confirm execution"), automatic assignment to [In Progress] is strictly prohibited.
  • The execution Task ID (e.g., T0001) directly penetrates the gateway into the state machine for progression; secondary card creation is prohibited.

Core CLI Command Quick Reference

Business ScenarioUnified CLI Facade (cli.py)Underlying Dedicated ScriptCore Function & Parameters
Create Task Cardpython3 scripts/cli.py task create ...python3 scripts/quick_task.py create ...--name "..." --assignee "..." --stage "Sprint 2.1 Menu Adjustments" --type "A" --target "..." --criteria "..." [--in-scope "file # reason"]
Stage Gate Admission/Closurepython3 scripts/check_stage_gate.py ...python3 scripts/check_stage_gate.py ...--action start/close --stage "Sprint 2.1 Menu Adjustments" (admission as needed, predecessor dependencies automatically verified)
Codified Dispatchpython3 scripts/cli.py dispatch --task-id T00xx --format codexpython3 scripts/dispatch_task.pyValidate dependencies and concurrency, move to [In Progress], output JSON parameters that can be directly consumed by Codex sub-agent tools
Advance Task Statepython3 scripts/transition_task.py ...python3 scripts/transition_task.py ...--task-id T00xx --role DEV --from-status In Progress --to-status In Review --token <token> --summary "Core change description" --remarks "..." (--summary is mandatory for submitting review, automatically injecting diff and credential snapshot)
Human Final Acceptancepython3 scripts/cli.py task accept ...python3 scripts/quick_task.py accept ...--task-id T00xx (Human user only, Agent counter-signature strictly prohibited)
Start Kanbanpython3 scripts/cli.py kanbanpython3 scripts/start_kanban_server.pyStarts by default at http://127.0.0.1:32886/
Health Checkpython3 scripts/cli.py statuspython3 scripts/heartbeat.pyOutputs overall health dashboard, blocked cards, and efficiency metrics (including sub-agent cognitive roaming and alerts > 600s)
Tech Stack Sync/Profile Recalculation-scripts/update_agent_tech_stacks.py--force-recalc forces recalculation of expert technical capability profile and refreshes exported Subagents, solving the problem of outdated profiles after tech stack migration
Continuity Checkpython3 scripts/cli.py ccp ...python3 scripts/cli.py ccp ...--task-id T00xx --stage In Review verifies context continuity
Trace Mappython3 scripts/cli.py tracepython3 scripts/generate_trace_html.pyOffline parsing of session traces, generates a full-chain trace map HTML in seconds (0 Token)
Smart Incremental Testpython3 scripts/cli.py testscripts/run_tests.pyAutomatically detects Git changes, only executes related unit tests (seconds), document changes are automatically exempted, --all for full regression

8 Expert Sub-Agent Write Permissions & State Transition Interval Matrix

[!IMPORTANT] Physical Isolation Iron Law & Dispatch Token Gating:

  1. Main Agent (PM Manager Yan) is strictly prohibited from self-role-playing as a developer to write business code in the main session! Codex hosts must use the actual sub-agent tool for dispatch; other hosts can use their respective sub-agent tools.
  2. Codified Dispatch & Token Gating: When dispatching a task, cli.py dispatch generates a unique dispatch_token. Professional R&D roles (DEV/FRONTEND/REVIEWER/QA) must carry --token <token> during state transitions. The state machine engine enforces strict interception at the execution layer, preventing self-role-playing that fakes green checks.

Codex Desktop Automatic Dispatch Contract

After the user explicitly authorizes the start, the Main Agent executes the following in the target project directory: python3 .agents/skills/fy-flow/scripts/dispatch_task.py --task-id <ID> --format codex. Standard output is a single JSON object containing tool, arguments, task_id, and role; diagnostic information goes to standard error. The Main Agent must immediately pass arguments as-is as the parameter to a single real collaboration.spawn_agent tool call, wait for the sub-agent to return, then continue dispatching the review and test experts according to the current task status. Do not manually rewrite, copy, or reduce the dispatch message; do not print the JSON as a dispatch confirmation, nor have the Main Agent perform the expert's work. If the current Codex host does not have a sub-agent tool, clearly report the host capability deficiency, keep the task in its current state, and do not forge an expert receipt. Dispatched experts should only execute exit_contract.required_cli in the message; they must not create new cards or start work. Since user authorization covers this execution, dispatch can continue in the same session after card creation; only cards that have not yet received authorization to start remain in [To Do].

Sub-Agent IDRole NameCore ResponsibilitiesLegal State Transition IntervalExit Contract
@flow-pmManager Yan (Project Manager)Requirement decomposition, classification decision, and card creationTo Do -> In Progress / Request AcceptanceFor card creation scenarios, stop at [To Do] and wait for authorization to start; request human acceptance
@flow-architectArchitect Qian (System Architect)Overall system architecture, interface contracts, and technology selectionTo Do -> In Progress -> DoneProduce ADR and architecture design documents
@flow-devDeveloper Li (Development Engineer)Backend/full-stack core coding and unit test implementationTo Do -> In Progress -> In ReviewMust run unit tests, generate development report, provide --summary when submitting for review
@flow-frontendFrontend Ma (Frontend Development Engineer)Web/UI component and frontend experience implementationTo Do -> In Progress -> In ReviewVerify component rendering and interactions, generate report, provide --summary when submitting for review
@flow-reviewerReviewer Zhou (Code Review Expert)Code standards, security scanning, and quality gatingIn Review -> In Test / In Review -> Done (Tier-3 credentials fresh and non-drifted, bypass QA direct pass) / In Review -> ReturnedStatic scanning passes without warnings, produce review comments
@flow-qaQA Zhang (Test Engineer)Targeted integration tests and release qualification (only execute tests related to the current task, full scripts strictly prohibited)In Test -> Done / In Test -> ReturnedWrite targeted test cases, record actual effort
@flow-docsDocs Li (Documentation Engineer)Deliverable document governance and user manualTo Do -> In Progress -> DoneFollow GB/T 7713 academic and engineering standards
@flow-devopsDevOps Lv (Operations Administrator)Branch merging, release builds, and Git history protectionTo Do -> In Progress -> DoneVerify pipeline passes, resolve PR blocking

Modular Specification Entity Index (Refer to Specific Specification Files as Needed)

If and only when entering a specific business scenario, the Agent MUST call view_file to consult the corresponding specification's entity Markdown file:

相关技能

steipete/bluebubbles

Send and manage iMessages via BlueBubbles, including attachments, tapbacks, edits, replies, and groups.

community

steipete/eightctl

Control Eight Sleep pods (status, temperature, alarms, schedules).

community

steipete/blucli

BluOS CLI (blu) for discovery, playback, grouping, and volume.

community

steipete/bear-notes

Create, search, and manage Bear notes via grizzly CLI.

community

steipete/camsnap

Capture frames or clips from RTSP/ONVIF cameras.

community

steipete/gifgrep

Search GIF providers with CLI/TUI, download results, and extract stills/sheets.

community
← More 编程与开发 skills