Synapse-Network-Growing is the Company Ops control plane for the Synapse system. Its purpose is to turn goals, signals, incidents, and operator intent into observable agent-first workflows that can plan, execute, verify, request approval, and deliver safely.
The current operating model is:
Operating Manifest
-> Planner Sweep
-> RequirementCandidate
-> Requirement
-> Declarative Workflow Run
-> State Machine Runner
-> Atom Invocations + Blackboard + Evidence
-> Human Gate / Release / Done
The key boundary is simple: Agent is the engine, declarative state machine is the rail, human operator is the high-risk gate.
The repository is no longer centered on legacy TaskBoard / CrewAI / fixed execution steps. New business flow must use:
- Operating Manifest: the single active strategy snapshot and budget/goal control surface.
- RequirementCandidate: the buffer between autonomous planning and the formal requirement pool.
- Requirement: the approved or materialized work item tied to a project, manifest, and goal lineage.
- Declarative Workflow Run: the source of truth for execution, validation, retry, evidence, and delivery.
- Blackboard + Artifacts: shared workflow context and immutable evidence.
- Human Gate: mandatory approval for production, funds, settlement, auth, external spend, and self-modifying deployment.
- Recurring Jobs + Watchers: observable background systems for bug sweep, notification dispatch, recurring workflow registry, and planning sweeps.
Legacy /api/tasks and /api/approvals public routes have been retired. Any remaining old task/execution internals are deletion candidates, not compatibility surfaces for new work.
Coding agents and humans should start here:
- ARCHITECTURE.md - short architecture constitution for coding agents.
- AGENTS.md - repo agent operating protocol and capability profile rules.
- System Design Principles - canonical engineering constraints.
- Declarative Agent Workflow Factory - state machine, blackboard, atoms, gates.
- Planning Lifecycle Operating Manifest - manifest, planner sweep, candidates, drop goal.
- Recurring Workflow Control Plane - scheduled jobs and dispatcher observability.
- Workflow Factory Frontend Control Plane - operator UI for workflow observability.
- Figma Driven Human-Agent Control Plane Frontend - Figma/Stitch-driven cockpit UI direction and implementation contract.
- Current Feature Inventory - implemented capability baseline.
- Ops Console Usage - local UI/API usage notes.
- Ops Console Page Map and E2E Validation - page-by-page function map, tests, and GoalB dogfood validation.
Historical evidence, old plans, and legacy docs are intentionally moved under docs/archive/. They are audit evidence, not current architecture sources.
/command-center: Human-Agent control plane overview, desired state, blockers, active runs, and background health./planning: active Operating Manifest, planner candidates, goal rollups, drop-goal safe halt, manifest history./intake: requirement discovery entry point for errors, watcher events, forum feedback, and planner signals./gates: Human Oracle queue for pending gates, failed runs, and high-risk decisions./requirements: requirement pool and workflow readiness./workflows: declarative workflow run operations center./workflows/:id: explainable workflow workbench with steps, evidence, gates, delivery, and debug views./factory/workflows: read-only workflow definition registry./factory/atoms: read-only atom registry./knowledge: artifacts and SOP Ledger entry point./schedules: recurring workflow registry, run history, scheduler/dispatcher health./system/notifications: notification queue, dispatcher health, suppression, and delivery observability./errors: error fingerprints and bug materialization state./projects: project registry and delivery policy configuration./releases: PR, delivery, deploy, and rollback governance.
The current product-level validation target is GoalB: the system must show that an Operating Manifest goal can become a RequirementCandidate, then a formal Requirement, then a completed declarative WorkflowRun with evidence. See Ops Console Page Map and E2E Validation.
The local backend expects Python 3.14+ through the shared workspace virtualenv:
/opt/homebrew/bin/python3.14 -m venv /Users/cliff/workspace/agent/.venv
/Users/cliff/workspace/agent/.venv/bin/python -m pip install -U pip setuptools wheel
/Users/cliff/workspace/agent/.venv/bin/pip install -r requirements.txtCommon local commands:
sh scripts/local/restart.sh
/Users/cliff/workspace/agent/.venv/bin/python -m pytest -q
cd dashboard && npm test
cd dashboard && npm run buildRecommended planner adapter:
PLANNER_BACKEND=codex_sdk
PLANNER_MODEL=gpt-5.4
PLANNER_TIMEOUT=300Supported planner adapters are:
openai | claude | copilot | codex_cli | codex_sdk
Runtime execution can still be configured independently. For local Codex CLI execution:
ORCHESTRATION_BACKEND=codex_cli
LLM_PROVIDER=codex
LLM_MODEL=gpt-5-codex
CODEX_CLI_MODEL=gpt-5-codex
CODEX_CLI_FULL_AUTO=trueFor OpenAI API usage:
LLM_PROVIDER=openai
LLM_MODEL=gpt-5.4
ORCHESTRATION_BACKEND=openai_responses
OPENAI_API_KEY=...Lark delivery is queued, deduplicated, and rate-limited before sending. Configure separate webhooks for alert, approval, workflow notification, and release channels:
OPS_LARK_BOT_WEBHOOK_URL=...
OPS_LARK_BOT_KEYWORD=告警
OPS_LARK_REQUIREMENT_WEBHOOK_URL=...
OPS_LARK_REQUIREMENT_KEYWORD=需求
OPS_LARK_ALERT_WEBHOOK_URL=...
OPS_LARK_ALERT_KEYWORD=告警
OPS_LARK_APPROVAL_WEBHOOK_URL=...
OPS_LARK_APPROVAL_KEYWORD=审批
OPS_LARK_RELEASE_WEBHOOK_URL=...
OPS_LARK_RELEASE_KEYWORD=发布
OPS_LARK_DELIVERY_DRY_RUN=false
OPS_NOTIFICATION_DISPATCHER_ENABLED=true
OPS_NOTIFICATION_ALERT_BURST_PER_MINUTE=5
OPS_NOTIFICATION_DEFAULT_COOLDOWN_SECONDS=900Set OPS_LARK_DELIVERY_DRY_RUN=true when running smoke tests against configured webhooks. Dry-run mode still persists notification deliveries and renders Lark cards, but it never calls the Lark webhook.
Routing rules:
- P0/P1 interruption -> alert.
- Human gate -> approval.
- PR / release / deploy -> release.
- Workflow lifecycle summary -> workflow notification.
- Ordinary state transitions, atom success, artifacts, and blackboard snapshots stay in Ops Console.
docs/design-docs/: current architecture and formal design docs.docs/ops/: operational governance and deployment runbooks.docs/plans/: temporary staging only for new plan bundles; implementation truth remains inARCHITECTURE.mdanddocs/design-docs/.docs/archive/: historical evidence, old plans, old design drafts, and retired docs.docs/assets/: architecture diagrams used by current docs.
When adding or deleting formal capabilities, update Current Feature Inventory with the same change.