multi-agent-tcp

star 1

Work on the current multi_agent_tcp direction: gulicode-bp Codex plugin productization, Blueprint web workbench, GraphRuntimeControlPlane, GraphRuntime scheduling, plugin-controlled start plans, AgentNode queues, workspace state, events, and CLIWorkerBackend adapters. Use for the gulicode-bp plugin, blueprint workbench/debug startup, runtime start/status/end, agent dispatch, workspace/archive flow, per-agent private workspaces, MCP workspace tools, and GuLiCode desktop compatibility work.

QHXRPG-A By QHXRPG-A schedule Updated 6/8/2026
play_arrow Run Skill in Manus View GitHub

name: multi-agent-tcp description: >- Work on the current multi_agent_tcp direction: gulicode-bp Codex plugin productization, Blueprint web workbench, GraphRuntimeControlPlane, GraphRuntime scheduling, plugin-controlled start plans, AgentNode queues, workspace state, events, and CLIWorkerBackend adapters. Use for the gulicode-bp plugin, blueprint workbench/debug startup, runtime start/status/end, agent dispatch, workspace/archive flow, per-agent private workspaces, MCP workspace tools, and GuLiCode desktop compatibility work.

multi_agent_tcp Project Skill

Use this skill when working on the multi_agent_tcp repository, especially the gulicode-bp Codex plugin, Blueprint web workbench, POPO-triggered Blueprint execution, blueprint runtime, plugin-controlled start plans, multi-agent communication, workspace isolation, MCP workspace tools, and CLI worker adapters. Temporarily treat GuLiCode desktop, mobile, console, and hosted/server surfaces as out of scope unless the user explicitly asks for them.

This file is intentionally short so Codex can discover the skill reliably. Detailed context lives in the files listed below. Load only the relevant files for the current task.

Most Important Local Command

When the user asks to reinstall, refresh, reload, or restart the installed gulicode-bp plugin after Python/runtime/framework changes, prefer this one-step command from the repo root:

F:\src\Package\Script\Python\multi_agent_tcp\restart-gulicode-bp-plugin.cmd -Install -SkipWebBuild -SyncFrameworkAssets -NoOpen -ProjectDir F:\src\Package\Script\Python\multi_agent_tcp

This command reinstalls/refreshes the personal plugin package, refreshes the plugin runtime Python package, syncs framework assets, and restarts the plugin service. If startup is slow or a previous run returned a health-check false negative, add a wider health window:

F:\src\Package\Script\Python\multi_agent_tcp\restart-gulicode-bp-plugin.cmd -Install -SkipWebBuild -SyncFrameworkAssets -NoOpen -ProjectDir F:\src\Package\Script\Python\multi_agent_tcp -HealthTimeoutSeconds 90

Use -SkipWebBuild only when F:\src\Package\Script\Python\multi_agent_tcp\GuLiCode\packages\app\dist\index.html exists and dist\assets\ is present. If the dist is missing or stale, rerun without -SkipWebBuild; the installer can run bun install and build the app. Never accept the fallback plugins\gulicode-bp\web\index.html page as the Workbench.

If the one-step restart command returns exit code 1, do not treat that as a failure by itself, but only ignore it after checking that the installed plugin service, POPO callback service, and MCP status are running and the Workbench URL passes these checks:

  • <origin>/config.js has projectDir equal to the requested project, normally F:\src\Package\Script\Python\multi_agent_tcp.
  • The Workbench HTML is the built GuLiCode app, not the fallback page with <main class="shell">.
  • POST <origin>/api/blueprint with the config token and blueprint.list(projectDir) returns the expected Blueprint list.

If the log says command not found: vite, GuLiCode app dist is missing, or fallback web UI, that exit code 1 is a real install/build failure: install or build the GuLiCode app first, then rerun the restart.

The script writes restart logs under:

F:\src\Package\Script\Python\multi_agent_tcp\logs\restart-gulicode-bp-plugin-*.log

After it finishes, verify the useful health endpoints first:

http://127.0.0.1:8787/api/health
http://127.0.0.1:3100/health

After a successful reinstall/restart, open or query the installed plugin Workbench through start_blueprint_workbench or the installed script:

C:\Users\qiuhaoxuan\plugins\gulicode-bp\.runtime\venv\Scripts\python.exe C:\Users\qiuhaoxuan\plugins\gulicode-bp\scripts\start_workbench.py --project-dir <project-dir> --blueprint-id <blueprint-id> --ready-file <ready-json>

For this machine's main Blueprint project, use F:\src\Package\Script\Python\multi_agent_tcp as <project-dir>. Do not give a Workbench URL whose decoded route or config.js points at a temporary Codex worktree such as C:\Users\qiuhaoxuan\.codex\worktrees\... unless the user explicitly asked to debug that worktree.

When the user asked for one-click reinstall/restart and did not ask for logs or diagnostics, the final response should default to only the plugin Workbench URL as a Markdown link. Do not include POPO callback URLs, health URLs, process details, or restart-log summaries unless the restart failed or the user asks for those details.

Current Center

The active product direction is:

gulicode-bp Codex plugin
  -> local Blueprint web workbench
  -> POPO callback and Blueprint session message entry
  -> GraphRuntimeControlPlane
  -> GraphRuntime
  -> AgentNode queues, outgoing batches, joins, workspace state, events
  -> CLIWorkerBackend
  -> AgentTCPClient / Broker / CLIAdapter / worker process

Interpretation rules:

  • Treat gulicode-bp as the default user-facing development/debug entrypoint.
  • Treat the Blueprint web workbench as a plugin-served surface backed by the existing GuLiCode app build and Python runtime service.
  • Focus current implementation on the plugin and POPO invocation chain.
  • Do not include /mobile, /console, GuLiCode desktop/Electron, or hosted server product work in the default scope. Use those paths only when the user explicitly asks for that surface.
  • Treat GraphRuntimeControlPlane and GraphRuntime as the framework-owned execution center.
  • Treat Codex desktop control as two explicit actions: set the saved start AgentNode, then execute a caller-provided plan against that saved start node.
  • Treat low-level TCP workers as a backend adapter path, not the product architecture center.
  • Prefer CLIWorkerBackend in new writing. Mention CodeMakerCluster only as a backward-compatible alias for the old cluster facade.
  • Keep Codex-first adapter work as the default for live blueprint smoke, streaming, and MCP workspace-tool validation.
  • Do not revive the retired Ryven/editor UI line unless the user explicitly asks for it.

Current common local root:

F:\src\Package\Script\Python\multi_agent_tcp

Historical paths such as D:\agents\multi_agent_tcp may appear in archive notes. Do not use them as current defaults unless they exist on this machine.

Start Here

Read these first based on the task:

  • Project architecture and ownership: knowledge_base/core_architecture.md
  • GuLiCode desktop startup, packaging, and shell/runtime layering for explicit desktop work: knowledge_base/gulicode_desktop.md
  • GuLiCode app UI, plugin-served blueprint workbench, mobile/console debug surfaces, and renderer conventions: knowledge_base/guli_desktop_ui.md
  • Runtime control-plane CLI/RPC flows: knowledge_base/dispatch_workflows.md
  • Plugin start-plan governance and multi-agent communication design: 多agents通信设计.md
  • Registry, skill selection, and skill injection: knowledge_base/registry_and_skills.md
  • CLI adapter/backend history and constraints: knowledge_base/multi_cli_workflow.md
  • Environment setup and machine-specific diagnostics: environment_setup.md
  • Current active priorities: tasks/current_goals.md
  • Historical plugin-first debug startup workflow: knowledge_base/debug_start.md

Recent Handoff

For the latest persistent POPO start-Agent session, Codex app-server thread reuse, in-flight turn/steer injection, callback timeout silence, and restart orphan worker cleanup from 2026-06-18:

  • Runtime/backend POPO-bound saved start full Agents now use a codex_app_server backend by default, keep one Codex thread for the active POPO sessionKey, inject transcript history only on activation/re-activation, steer in-flight POPO messages into the active Codex turn when possible, and fall back to the existing Agent queue when steer is rejected: archive/runtime-backend/blueprint_popo_persistent_codex_session_steer_cleanup_2026-06-18.md
  • The POPO callback no longer sends 蓝图运行仍在处理中 for a still-running persistent run, and restart-gulicode-bp-plugin.ps1 now kills plugin-owned multi_agent_tcp_cluster broker/agent process trees so child codex app-server --listen stdio:// processes cannot survive restart as orphans: archive/runtime-backend/blueprint_popo_persistent_codex_session_steer_cleanup_2026-06-18.md

Use this file when the user reports:

  • POPO messages within 10 minutes do not reach the same Agent/Codex thread.
  • later active-run POPO messages repeat [Recent BlueprintSession Messages].
  • a busy Agent ignores or fails to receive an in-flight POPO steer message.
  • POPO users receive the framework status text 蓝图运行仍在处理中.
  • plugin restart leaves multi_agent_tcp_cluster or codex app-server processes behind.

For the latest POPO file-only attachment delay, fileId download, and POPO file-send token refresh from 2026-06-17:

  • Runtime/backend POPO callback now downloads fileInfo.fileId attachments to absolute local paths, caches non-image file-only messages until the user's next ordinary text message, accumulates multiple files in order, keeps images immediate, clears pending files on /new and /stop, and fixes blueprint_send_popo_file / file_sender.send(path) upload failures by sending Open-Access-Token to uploadUrl and retrying once after token expiry: archive/runtime-backend/blueprint_popo_file_attachment_delay_token_refresh_2026-06-17.md

Use this file when the user reports:

  • a POPO file-only message reaches the Agent before the next text message.
  • a POPO file attachment is unresolved or lacks an absolute path.
  • a POPO callback has fileInfo.fileId but no direct file URL.
  • multiple uploaded POPO files should be delivered with one later text message.
  • /new, /stop, /help, or /excel-log consumes pending files incorrectly.
  • blueprint_send_popo_file or file_sender.send(path) fails with access token expired while the local file exists.

For the latest one-click reinstall/restart script hardening, Workbench ProjectDir binding, fallback web UI refusal, and exit-code-1 diagnosis from 2026-06-17:

  • Installer/runtime script changes ensure --skip-web-build cannot silently install the fallback Workbench, missing vite triggers bun install, restart scripts pass and verify the target ProjectDir, and one-click restart should default to the main project Workbench URL: archive/runtime-backend/gulicode_bp_reinstall_workbench_guardrail_2026-06-17.md

Use this file when the user reports:

  • a Workbench URL points at C:\Users\qiuhaoxuan\.codex\worktrees\... instead of F:\src\Package\Script\Python\multi_agent_tcp.
  • reinstall/restart returns exit code 1 and the cause may be a real build failure.
  • logs mention command not found: vite, missing GuLiCode app dist, or fallback web UI.
  • the installed Workbench shows the simple fallback page instead of the built GuLiCode app.

For the latest Blueprint slot termination hardening, POPO concurrent session write fix, framework-forwarded start-Agent POPO replies, and POPO /new confirmation behavior from 2026-06-10:

  • Runtime/backend blueprint.slots.terminate structure-level forced cleanup, session.json unique temp writes and Windows retry, ordinary start-Agent replies forwarded to POPO with messageId transcript fields, and /new returning 已开启新会话: archive/runtime-backend/blueprint_slot_termination_popo_direct_reply_new_session_2026-06-10.md

Use this file when the user reports:

  • 终止运行槽 only killed one instance or left queued/running sessions behind.
  • POPO callback fails with [WinError 5] 拒绝访问 while replacing session.json.
  • POPO receives no final reply even though Workbench shows an ordinary start Agent reply.
  • POPO /new clears the session but appears stuck because there is no confirmation.

For the latest Blueprint session-slot lifecycle, framework-managed session termination, POPO callback resilience, and Agent panel transcript timeline work from 2026-06-10:

  • Runtime/backend structure-level slot summaries, per-pool max three active sessions, queued sessions, removal of Agent-visible blueprint_terminate_session, framework idle auto termination, POPO callback local-route fallback, agent_info() service-lock fix, and POPO reply auto-task-status behavior: archive/runtime-backend/blueprint_session_slots_framework_auto_popo_resilience_2026-06-10.md
  • Workbench current-session vs Blueprint-slot status card, active/queued session counts, session terminate / slot terminate actions, transcript-backed Agent panel timeline, maintenance-event filtering, and stream text null-guard crash fix: archive/frontend/blueprint_session_slot_panel_timeline_2026-06-10.md

Use those files when the user reports:

  • A POPO message was sent but did not enter the Blueprint session transcript.
  • /popo/callback returns 500 because callback config lookup timed out.
  • The main runtime or callback config query is unresponsive while the Agent information panel is open.
  • An Agent can see or call blueprint_terminate_session, or a run slot is completed by an Agent instead of by UI/framework policy.
  • The runtime panel confuses session count, live run instance count, and structure-level Blueprint slot status.
  • The Agent panel shows task-status/summary/session-maintenance text as a normal Agent reply.

For readable POPO session keys from 2026-06-09:

  • Runtime/backend service-side receiver ownership via saved popoReplyTo, POPO send helper, transcript agent_reply / popo_reply_sent events, readable bps_popo_<user>_<hash> session keys, and legacy bps_<hash> migration: archive/runtime-backend/blueprint_popo_reply_tool_readable_sessions_2026-06-09.md
  • Workbench session dropdown sessionDisplayName support, session-key-first display, and POPO user-readable second-line label: archive/frontend/blueprint_popo_readable_session_list_2026-06-09.md

Use those files when the user reports:

  • A POPO-started Agent can see the message but its ordinary final reply is not forwarded to the POPO user.
  • POPO reply target selection is wrong for private or group conversations.
  • The session list shows only opaque bps_<hash> ids instead of including the POPO user such as qiuhaoxuan.
  • Existing old POPO sessions do not migrate to readable bps_popo_... keys.

For the latest POPO callback service global robot route configuration and Workbench control drawer work from 2026-06-09:

  • Runtime/backend plugin-level POPO callback robot allowlist popo_robot_routes.json, internal list/save/delete/toggle commands, blueprint.popo.callbackConfig, legacy /popo/callback unique-enabled auto resolution, multi-enabled conflict behavior, and MCP public/internal credential boundary: archive/runtime-backend/blueprint_popo_global_callback_routes_2026-06-09.md
  • Workbench Blueprint Control drawer, POPO Callback Service panel, service status/callback URL display, robot enable checkbox and credential editor, and collaboration panel relocation into the drawer: archive/frontend/blueprint_popo_global_callback_service_drawer_2026-06-09.md

Use those files when the user reports:

  • /popo/callback/<robot_app_key> accepts or rejects the wrong robot.
  • Legacy /popo/callback should auto resolve one enabled robot or reject multiple enabled robots.
  • blueprint.popo.robots or related Workbench API calls return UNKNOWN_COMMAND.
  • The Blueprint Control drawer or POPO Callback Service panel is missing from the Workbench.
  • A packaged plugin still uses old POPO callback credentials after rebuild or restart.
  • Public Codex MCP exposes or can mutate POPO callback robot credentials.

For the latest POPO Blueprint session identity, start-Agent ownership, idle termination, and agent-level POPO Inspector work from 2026-06-09:

  • Runtime/backend direct Workbench session key main+<blueprintId>, POPO session-key separation, agent-level popo_entry validation, /new session clear behavior, POPO idle termination checks, start-Agent-only blueprint_terminate_session MCP tool, plugin rebuild/restart, and smoke results: archive/runtime-backend/blueprint_popo_session_identity_termination_2026-06-09.md
  • Workbench full-Agent popo_entry model, legacy runtime POPO migration, Inspector disabled-state reasons, saved-start-only POPO editability, previous start-Agent POPO clearing, POPO Blueprint indicator, i18n, frontend tests, and page smoke: archive/frontend/blueprint_agent_level_popo_entry_ui_2026-06-09.md

Use those files when the user reports:

  • Direct Blueprint run-slot messages still create bps_... sessions instead of using main+<blueprintId>.
  • POPO forwarding appears editable on the wrong full Agent, multiple full Agents can enable POPO, or the disabled reason is unclear.
  • A POPO-enabled full Agent is not forced to be the saved start Agent.
  • A POPO Blueprint is missing its visual indicator.
  • /new does not clear the expected Blueprint session history.
  • POPO Blueprint sessions do not ask the start full Agent whether to terminate when all Agents, ScriptNodes, and resident-service calls are idle.
  • blueprint_terminate_session is missing for the POPO start Agent or is exposed to the wrong Agent.

For the latest Blueprint sessions, run slots, POPO entry, Codex start-agent setting, and MCP boundary work from 2026-06-05:

  • Runtime/backend session storage, run-slot assignment, POPO singleton-service routing, Codex desktop setStartAgent / executePlan, run-scoped monitor tools, public/internal MCP boundaries, Workbench internal slot bridge fix, and automated test results: archive/runtime-backend/blueprint_sessions_slots_popo_mcp_2026-06-05.md
  • Workbench sessions dropdown, Blueprint run-slot panel, single AgentNode start selector, POPO entry inspector fields, monitor-tools toggle, platform bridge methods, browser smoke, and UI test results: archive/frontend/blueprint_sessions_slots_popo_entry_ui_2026-06-05.md

Use those files when the user reports:

  • Blueprint sessions are unclear, missing from the toolbar, not scrolling, or cannot be deleted correctly
  • "Blueprint run slot" start/message UI returns UNKNOWN_COMMAND or does not show POPO entry errors
  • runtime.start_node_id is not saved, the start node selector is empty, or a plan starts from a node other than the saved AgentNode
  • POPO robot app key/config binding, session routing, or slot load balancing is behaving unexpectedly
  • Codex desktop should set the start Agent separately from executing a specified plan
  • public MCP tools should expose blueprint_set_start_agent / blueprint_execute_plan but not blueprint.slots.*
  • full AgentNode Blueprint monitor tools need run-scoped status/events/diff access without cross-run mutation

For the latest table_queue_service ScriptNode help/schema pattern and table-occupation resident service work from 2026-06-08:

  • table_queue_service as the reference design for ScriptNodes that expose multiple operations through action + arguments, targeted help/schema behavior, script-only resident service gating, valid table name validation, and refresh_valid_tables running SVN update before scanning .xlsx files: archive/runtime-backend/blueprint_table_queue_script_help_pattern_2026-06-08.md

Use that file when the user reports:

  • a future ScriptNode needs an Agent-facing help/schema contract
  • an Agent is unsure what structure to pass to a ScriptNode
  • a ScriptNode multiplexes operations through action + arguments
  • table_queue_service action formats, unknown-action behavior, or schema output size need clarification
  • table occupation requests should reject unknown .xlsx names
  • refresh_valid_tables should update SVN before reading table names

For the latest Blueprint resident service and Script Function Node boundary notes, plus a real AgentNode xltool resident-service write smoke from 2026-06-08:

  • Script Function Node project scope, resident service global plugin runtime scope, service description ownership, synchronous blueprint_service_call return behavior, xltool service shape, live Excel write smoke result, and known service-result/Windows archive follow-ups: archive/runtime-backend/blueprint_resident_services_script_nodes_xltool_smoke_2026-06-08.md

Use that file when the user reports:

  • an AgentNode cannot discover or call a resident service
  • resident service description text is unclear or stale
  • blueprint_service_call timing/result behavior is ambiguous
  • Script Function Nodes and resident services are being conflated
  • xltool resident-service Excel writes need a known-good smoke reference
  • a successful Blueprint run is marked failed during Windows archive/close

For the latest Blueprint resident services panel pagination/search/collapse work from 2026-06-05:

  • Resident services panel local search, fixed five-item pagination, header collapse/expand, dedicated refresh action, no-match empty-state fix, i18n, and browser smoke: archive/frontend/blueprint_resident_services_panel_search_pagination_collapse_2026-06-05.md

Use that file when the user reports:

  • the resident services panel should show only five services per page
  • resident service search, no-match copy, or page clamping behaves incorrectly
  • the resident services panel header should collapse/expand instead of looking like a close button
  • pagination buttons or localized resident service labels are missing
  • a packaged Workbench still shows the old resident services panel after plugin rebuild/restart

For the latest Blueprint Prompt editor and popout chrome polish from 2026-06-03:

  • Prompt node compact canvas editing, double-click draggable/resizable editor dialog, custom resize handle, Escape close behavior, hidden redundant Blueprint popout title bar, and browser smoke: archive/frontend/blueprint_prompt_editor_popout_chrome_2026-06-03.md

Use that file when the user reports:

  • Prompt node double-click editing should open a separate text dialog
  • the Prompt editor dialog cannot be dragged by its title bar
  • the Prompt editor dialog cannot be resized from the bottom-right handle
  • dragging/resizing Prompt dialogs behaves incorrectly after canvas pan/zoom
  • /blueprint-window/... still shows the redundant top title bar, dock button, or close button above the real Blueprint toolbar

For the latest Blueprint Prompt node and Agent port expansion work from 2026-06-02:

  • Prompt node model/UI, editable prompt text, once/always trigger controls, fixed Agent prompt: str input, expandable Agent port rows, and frontend smoke: archive/frontend/blueprint_prompt_node_agent_ports_ui_2026-06-02.md
  • Python PromptNode, GraphDefinition.prompt_nodes, fixed Agent prompt-port validation, once/always prompt injection, no-op dispatch behavior, and runtime tests: archive/runtime-backend/blueprint_prompt_node_runtime_2026-06-02.md

Use those files when the user reports:

  • Prompt nodes are missing from node search, do not expand into a textarea, or do not save trigger mode.
  • Agent/Worker Agent nodes should show fixed port rows like Script nodes, the prompt: str triangle pin is missing, or the old add/remove prompt menu appears.
  • PromptNode(out) cannot connect to Agent(prompt), creates an exec edge instead of a data edge, or non-Prompt sources can target Agent prompt.
  • Prompt text does not inject with legacy run_prompt, once/always behavior is wrong, or no-op dispatches consume once prompts.

For the latest gulicode-bp MCP direct-control start and live close-loop work from 2026-06-02:

  • Workbench planning inbox removal, Codex MCP direct start plan create/validate/start flow, live-start nonblocking return with startPending, terminal manifest persistence, Windows long-path workspace copy fixes, installed runtime wheel refresh, and direct smoke results: archive/runtime-backend/gulicode_bp_mcp_direct_start_live_close_loop_2026-06-02.md

Use that file when the user reports:

  • Workbench still exposes planning-request UI or task text for plan generation
  • blueprint_start hangs, times out, or creates a run but does not return
  • MCP/UI status disagree after a live start
  • cancelled runs leave run_manifest.status=running
  • long Windows paths under .bun / node_modules break run workspace snapshot
  • packaged plugin runtime still runs stale Python after source changes

For the latest Workbench external-run synchronization work from 2026-06-02:

  • Planning button/task textarea removal in the runtime panel, direct-run UI retention, periodic listBlueprintRuns sync, active-run preference, and the Ctrl+R workaround fix for already-open Workbench tabs: archive/frontend/blueprint_workbench_external_run_sync_2026-06-02.md

Use that file when the user reports:

  • Codex MCP starts a Blueprint run but the already-open Workbench still shows an old CANCELLED run
  • pressing Ctrl+R makes the correct running run appear
  • Workbench should observe external MCP starts without a browser refresh

For the latest gulicode-bp plugin singleton service and reload work from 2026-06-02:

  • Codex-plugin-only singleton service boundary, stateless stdio MCP proxies, service.lock / service.json / service.log.jsonl, Workbench attach wrapper, owner takeover, planning-request reassignment, duplicate active-run guard, canonical personal-plugin .runtime state, Windows venv launcher cleanup, and final clean reload process inventory: archive/runtime-backend/gulicode_bp_plugin_singleton_service_2026-06-02.md

Use that file when the user reports:

  • duplicate gulicode_bp_mcp.py --service, Workbench, live-run registry, or plugin-managed Collaboration Server ownership
  • uncertainty about which plugin processes must be unique versus allowed Codex-side bootstrap/proxy access processes
  • stale service.json, service.lock, or singleton health after killing or reloading the plugin
  • planning requests moving between Codex threads or owner takeover behavior
  • duplicate active blueprint starts for the same projectDir + blueprintId

For the latest gulicode-bp MCP transport, bootstrap lock, and status logging work from 2026-06-02:

  • Transport/bootstrap/Workbench lifecycle boundaries, dead-PID bootstrap lock cleanup, bootstrap/MCP logs, mcp_status.json heartbeat diagnostics, Codex stdio reconnect boundary after Transport closed, and plugin manifest default prompt limit: archive/runtime-backend/gulicode_bp_mcp_transport_bootstrap_logging_2026-06-02.md

Use that file when the user reports:

  • gulicode-bp MCP tools are visible but calls return Transport closed
  • Workbench is open but MCP tools cannot be called
  • MCP startup appears stuck behind bootstrap.lock
  • mcp_status.json says running while the MCP process is gone or stale
  • the user asks what transport, bootstrap, or Workbench means
  • Codex warns that interface.defaultPrompt has too many prompts

For the latest gulicode-bp first-run bootstrap and packaging work from 2026-06-01:

  • New-user standalone install path, .runtime/venv first-run bootstrap, bootstrap MCP entrypoint, Windows re-entry behavior, release package layout, fixed package-gulicode-bp-plugin.cmd packaging command, release smoke, and release-versus-trunk semantics: archive/runtime-backend/gulicode_bp_first_run_bootstrap_packaging_2026-06-01.md

Use that file when the user reports:

  • a new Codex user cannot run gulicode-bp without local multi_agent_tcp source
  • .runtime/venv is missing, stale, locked, or not loading from plugin-owned site-packages
  • .mcp.json still points at repo-local Python, PYTHONPATH, or GULICODE_BP_REPO_ROOT
  • release packages include .runtime, lack runtime/wheels, or need package shape verification
  • the user asks what release versus trunk/source means
  • the user says 鎵撳寘, 鎻掍欢鎵撳寘, or package

For the latest gulicode-bp standalone runtime and Codex fan-out smoke work from 2026-06-01:

  • Plugin-owned runtime venv install/cache sync, installer repair behavior, MCP event-loop asyncio.run() fix, queued outgoing-target reminders, reminder race guards, fanout-worker-smoke, and Codex live fan-out smoke: archive/runtime-backend/gulicode_bp_standalone_codex_fanout_runtime_2026-06-01.md

Use that file when the user reports:

  • installed gulicode-bp still depends on local repo source or stale Codex cache .mcp.json
  • blueprint_start fails inside MCP with asyncio.run() cannot be called from a running event loop
  • fan-out runs leave remaining_targets stuck or reminders create duplicate batches
  • a side-browser workbench URL shows ERR_CONNECTION_REFUSED after runtime restart
  • live worker smoke should use Codex rather than CodeMaker

For the latest gulicode-bp plugin direct-control and start-plan work from 2026-06-01:

  • Plugin CRUD MCP/API, deterministic blueprint.plan.create, blueprint.plan.validate, confirmed blueprint_start, Top Agent wording removal, two-step workbench run UI, and the current P0 standalone-distribution boundary: archive/runtime-backend/gulicode_bp_plugin_direct_control_start_plan_2026-06-01.md

Use that file when the user reports:

  • installing gulicode-bp without a local multi_agent_tcp checkout does not work
  • Blueprint CRUD, plan creation/validation, or confirmed start should be driven directly by plugin MCP tools
  • the UI still exposes Top Agent planning language or skips the generate-plan confirmation step

For the latest Agent and Script Function Node interaction work from 2026-05-31:

  • Agent -> Script multi-input fan-out, Script multi-output -> Agent fan-in, true sibling edges, hub rendering, and group hover/select/delete: archive/frontend/blueprint_agent_script_fan_edges_2026-05-31.md
  • Ordinary-Agent blueprint_script_call MCP tool, batch script_calls state, idle reminders, script execution events, and automatic downstream Agent delivery: archive/runtime-backend/blueprint_script_call_mcp_runtime_2026-05-31.md
  • GuLiCode desktop debug loopback issue where stale 127.0.0.1:5173 blocked renderer health while [::1]:5173 worked: archive/runtime-backend/gulicode_desktop_debug_loopback_renderer_2026-05-31.md

Use those files when the user reports:

  • Agent -> Script or Script -> Agent connections should fan across all script ports, delete as a group, or render with a hub.
  • Agents are not calling Script Function Nodes, required_script_calls is missing, blueprint_script_call is unavailable, or downstream Script output is not automatically delivered to Agents.
  • agent.dispatch incorrectly bypasses a Script Function Node path.
  • GuLiCode desktop appears to have an Electron process and bridge but the window is blank or the debug launcher reports desktop failure while [::1]:5173 is healthy.

For the latest built-in Blueprint Branch/Tick common node work from 2026-05-31:

  • Branch/Tick node search, common node graph model, port type validation, Inspector edge validation, Tick-only direct-run allowance, and Branch pin labels/triangle input: archive/frontend/blueprint_common_nodes_branch_tick_ui_type_ports_2026-05-31.md
  • Python CommonNode, common_nodes graph parsing, backend port type validation, agent.dispatch to Branch, Branch strict-bool routing, Tick scheduling/backpressure, and common-node status: archive/runtime-backend/blueprint_common_nodes_branch_tick_runtime_2026-05-31.md

Use those files when the user reports:

  • Branch/Tick nodes are missing from node search or serialize incorrectly
  • Branch pins should show condition: bool, true: message, or false: message
  • invalid tick -> bool or other typed-port connections are accepted
  • canvas drag connection and Inspector edge edits disagree
  • Tick-only start-plan generation incorrectly requires a selected start node
  • common_nodes JSON fails to load, Branch dispatch does not route, or Tick cadence/backpressure/status looks wrong

For the latest Blueprint Script Function Node work from 2026-05-31:

  • Right-click node search, Script Function Node creation dialog, IDE selector, compile button, dynamic script ports, and Python-style node visual: archive/frontend/blueprint_script_function_node_ui_ide_compile_2026-05-31.md
  • Python @blueprint_node discovery/execution, .multi_agent_workspace/scripts management, desktop bridge commands, editor discovery/opening, and runtime Agent -> Script -> Agent flow: archive/runtime-backend/blueprint_script_function_node_runtime_bridge_2026-05-31.md

Use those files when the user reports:

  • right-click node search or the Script Function Node + flow is wrong
  • the toolbar Script editor selector, "system default" fallback, or script double-click open behavior is wrong
  • the Blueprint Compile button does not refresh script signatures/ports
  • Script Function Node visual shape, collapse/expand, or port rendering is wrong
  • blueprint.scriptNodes / blueprint.createScriptNode fails, script catalog scanning imports user code, or Agent -> Script -> Agent dispatch fails

For the latest full CLI Agent / Worker Agent split from 2026-05-30:

  • Full CLI Agent node UI, access-policy inspector, i18n, and opaque light-green visual treatment: archive/frontend/blueprint_full_agent_node_ui_2026-05-30.md
  • Full Agent runtime launch semantics, message-only MCP scope, and real Codex end-to-end smoke: archive/runtime-backend/blueprint_full_agent_runtime_real_codex_smoke_2026-05-30.md

Use those files when the user reports:

  • the add-node menu, inspector, or visible name for Agent / Worker Agent is wrong
  • full Agent nodes are not opaque light green
  • a full Agent unexpectedly gets private workspace or workspace MCP tools
  • worker_agent can bypass sandbox or dangerous access policy
  • real Codex full-Agent dispatch through agent_dispatch needs to be checked

For the latest Blueprint Agent model/runtime cleanup from 2026-05-29:

  • Test Agent merge into normal Agent UI, global Agent panel JSON test-log switch, Agent 简介 / 提示词 split, and frontend tests: archive/frontend/blueprint_agent_prompt_log_cleanup_2026-05-29.md
  • Hidden Windows worker/adapter launches and per-run run_prompt runtime injection: archive/runtime-backend/blueprint_hidden_console_run_prompt_2026-05-29.md

Use those files when the user reports:

  • Test Agent nodes or yellow test styling still appear in the Blueprint UI
  • Agent panel test JSON snapshots are missing, unexpectedly produced, or need the global switch behavior checked
  • Agent inspector 简介 / 提示词 behavior is unclear
  • Blueprint startup opens visible terminal windows
  • run_prompt is injected repeatedly or not injected on first dispatch

For the latest Collaboration Server /mobile desktop-bridge work from 2026-05-29:

  • Account-level desktop bridge, mobile chat, composer mode sync, structured segments, and tick/scroll fixes: archive/future-server/collaboration_server_desktop_bridge_mobile_chat_2026-05-29.md

Use that file when the user reports:

  • mobile chat send reaches desktop but does not appear on /mobile
  • /mobile mode selector is missing desktop agents or 蓝图规划
  • reasoning/tool disclosures close, resize, or jump scroll during tick refresh
  • desktop session mirror or desktop.session.submit behavior is unclear

For the latest blueprint multi-document and direct-run work from 2026-05-30:

  • Project blueprint dropdown, new blueprint dialog, current blueprint id binding, and direct live run button: archive/frontend/blueprint_multi_document_direct_run_2026-05-30.md

Use that file when the user reports:

  • the Blueprint toolbar dropdown, project blueprint switching, or new blueprint creation is wrong
  • runtime Confirm run should start live only after a generated start plan
  • start-plan generation does not require or respect selected start nodes
  • plugin start-plan generation starts the wrong blueprint document

For earlier blueprint panel work from 2026-05-25:

  • Progress overlay and staged-plan sync: archive/frontend/blueprint_flow_progress_overlay_plan_sync_2026-05-25.md
  • Runtime visual flow and pending filter: archive/frontend/blueprint_runtime_visual_flow_pending_filter_2026-05-25.md

Use those files when the user reports:

  • blueprint submit progress is missing or stuck
  • the start-plan preview card does not appear
  • the blueprint panel overlay/mask behavior is wrong
  • runtime node glow, edge flow, or pending task filtering looks wrong

Archive Structure

Archive records are split by ownership:

  • archive/frontend/: GuLiCode renderer/app UI, blueprint side panel, session composer planning UX, CSS/i18n, popout windows, and UI tests.
  • archive/runtime-backend/: GraphRuntime, GraphRuntimeControlPlane, Blueprint MCP runtime, DesktopBlueprintService, AgentTCP, CLIWorkerBackend, queues, workspaces, and local desktop service behavior.
  • archive/future-server/: reserved for future hosted/server-side product work.

Archives are historical records. Prefer current knowledge_base/ and tasks/ files for implementation decisions unless an archive is explicitly called out.

Query Map

  • Local environment setup, Python/Bun/Codex paths, PowerShell script policy, localhost MCP 502 / 503, proxy, NO_PROXY, and repo skill_list: read environment_setup.md.
  • GuLiCode desktop startup, packaged bring-up, taskbar icon, and direct Electron fallback: read knowledge_base/gulicode_desktop.md; use it only for explicit desktop-shell work.
  • When the user says 调试启动: read knowledge_base/debug_start.md, then start the gulicode-bp plugin workbench, the Collaboration Server, the app dev server, the mock mobile /mobile client, and the /console server console. Do not start the GuLiCode Electron desktop unless explicitly asked.
  • Desktop UI ownership, branding, icon replacement, empty-state wording, blueprint entry placement, and blueprint workbench embedding: read knowledge_base/guli_desktop_ui.md, then tasks/guli_desktop_ui_tasks.md.
  • Agent information panel interactions, task status display, automatic run summary, long-press progress, Markdown reply rendering, context menu, move/resize behavior, Test Agent JSON snapshots, and clean desktop debug: read archive/frontend/blueprint_agent_task_panel_auto_top_summary_2026-05-22.md plus the relevant archive/frontend/agent_info_panel_* files.
  • GuLiCode plugin control, organization view, start plan, and status explanation: read 多agents通信设计.md, then tasks/multi_agent_communication_tasks.md.
  • Runtime start/status/end, organization, message batch, agent dispatch, join-create, and join-contribute: read knowledge_base/dispatch_workflows.md.
  • Current architecture or component ownership: read knowledge_base/core_architecture.md.
  • GraphDefinition.agent_cycle_groups() and cycle/loop handling: read knowledge_base/ring_structure_solution.md, then knowledge_base/core_architecture.md.
  • Legacy CodeMakerCluster, run_parallel, run_chain, broker, and TCP worker path: read knowledge_base/cluster_api.md.
  • Workspace API, changesets, archive, private checkout, conflict flow, and MCP workspace tools: read knowledge_base/core_architecture.md and tasks/multi_agent_communication_tasks.md.
  • Multi CLI adapters, Codex/CodeMaker process adapters, and CLIAdapter: read knowledge_base/multi_cli_workflow.md.

Working Rules

  • Prefer repo-local patterns and existing tests.
  • For frontend work, keep GuLiCode UI restrained, dense, and consistent with the current app. Avoid unrelated redesigns.
  • For runtime work, keep the control-plane API as the product boundary and keep CLI workers as replaceable adapters.
  • For ordinary agents, preserve private workspace isolation. Project mutations should flow through the framework workspace/MCP path, not direct writes to the shared project root.
  • For ScriptNodes that expose multiple operations, use the table_queue_service help/schema contract as the reference: targeted help returns one schema, empty help returns all schemas, unknown actions return allowed actions plus a hint, and business calls return only business results.
  • When changing skill knowledge, keep this SKILL.md short and move detail to knowledge_base/, tasks/, or archive/.
  • When copying into KM_docs/skills-snapshot, mirror the current installed skill tree.

Repository Link

Install via CLI
npx skills add https://github.com/QHXRPG-A/multi_agent_tcp --skill multi-agent-tcp
Repository Details
star Stars 1
call_split Forks 0
navigation Branch main
article Path SKILL.md
More from Creator
QHXRPG-A
QHXRPG-A Explore all skills →