Workflow Automation: Agents, Approvals, and Push¶

Status: decommissioned 2026-07-06. The langgraph-agents fleet this page describes — its /inbox and /approval endpoints, the Zulip/ntfy/Pushover approval loop, and 16 of the Windmill workflows referenced below — was removed entirely (kubernetes/apps/ai/langgraph-agents/, kubernetes/apps/ai/sync-receiver/, the postgres-langgraph-checkpoints CNPG cluster, the langgraph-vault PVCs, and the hai.${SECRET_DOMAIN} / hai-web.${SECRET_DOMAIN} HTTPRoutes are all gone). See Current state below for what's actually running today; the rest of the page is kept as a historical record of the original design.

Current state (2026-07-06)¶

No agent fleet. langgraph-agents (the FastAPI service that used to live in the ai namespace) was deleted along with its Postgres checkpoints database, vault PVCs, and public routes. There is no /inbox, no /approval, no /admin/tasks.
Windmill is still deployed and still runs 8 workflows — none of them agent- or approval-related. What's left, all under kubernetes/apps/home/windmill/workflows/:
paperless-rag-fanout.ts — unified Paperless → Qdrant and LightRAG ingest from a single pull (one watermark, one 15m cron). Supersedes the two split ingests below as of 2026-07-20.
paperless-rag-ingest.ts / lightrag-rag-ingest.ts — the original split Qdrant / graph ingests, now superseded by the fan-out; retained on disk with schedules paused for rollback.
paperless-rag-tombstone.ts / lightrag-rag-tombstone.ts — Qdrant + LightRAG tombstone sweeps (still separate)
smart-home-intent-drift.ts — HA-native, unrelated to the fleet
windmill-failure-watcher.ts — Windmill self-introspection
workaround-watcher.ts — GitHub workaround-labeled issue discovery, backs the upstream-watcher convention
The approval flow no longer exists. There's no Class A–D task taxonomy, no Zulip #approvals stream, no ntfy/Pushover tap-to-approve loop — because there's no automated task pipeline producing anything to approve.
Known gap, not yet fixed: the HA voice "inbox …" intent (a rest_command defined in the separate home-assistant-config repo) still exists and will silently fail — it POSTs to a Windmill webhook that used to forward to langgraph-agents /inbox, which no longer exists. Anyone using the old "hold power button, say inbox …" gesture gets nothing. This is a known follow-up, not resolved here.
Critical AlertManager alerts still page Pushover directly. That part of the pipeline predates and is unrelated to this decommission — the AI investigation step in front of it (HolmesGPT) was removed in an earlier, separate pass, also landing 2026-07-06.
memory-mcp is unaffected — the cross-agent knowledge graph MCP server backed by postgres-langgraph-memory is still live, still used by Claude Code and Open WebUI. It simply lost langgraph-agents as a consumer.

If a replacement task-intake mechanism gets built, it belongs on this page — this file stays the canonical spot for "how does work get from a human to an agent" once that's true again.

Historical design (as originally built)¶

The section below documents the system as it existed before the 2026-07-06 decommission. Kept for design-rationale reference — none of the flows, endpoints, or Windmill scripts named past this point are live.

The system at a glance (historical)¶

flowchart LR
    subgraph "Triggers (you)"
        Voice[🎙️ Voice<br/>Wyoming → Whisper]
        ZulipApp[📱 Zulip app<br/>#inbox / @-mentions]
        HAApp[🏠 HA companion<br/>REST webhook]
    end

    subgraph "Workflow runtime"
        Windmill[💨 Windmill<br/>6 TypeScript flows]
    end

    subgraph "Agent fleet (ai ns)"
        LG[🧠 langgraph-agents<br/>FastAPI · :8765]
        Ollama[(Local Ollama<br/>P40 + Spark)]
        Claude[(Anthropic API<br/>opt-in per agent)]
    end

    subgraph "Surfaces (notifications)"
        Push[📲 ntfy / Pushover<br/>action buttons]
        ZulipStream[💬 Zulip streams<br/>#approvals · #digests]
        Vault[📓 Obsidian vault<br/>reports/]
    end

    subgraph "Alert path"
        AM[🚨 AlertManager<br/>severity=critical]
    end

    Voice & ZulipApp & HAApp --> Windmill
    AM --> Push
    Windmill --> LG
    LG --> Ollama
    LG -. opt-in .-> Claude
    Windmill --> Push
    Windmill --> ZulipStream
    LG --> ZulipStream
    LG --> Vault

Each block in Workflow runtime was a single TypeScript file checked into kubernetes/apps/home/windmill/workflows/, run in Deno sandboxes inside the Windmill worker pods, with secrets injected via the windmill-workflows-secret k8s Secret (sourced from 1Password).

The agent fleet was langgraph-agents — a FastAPI service in the ai namespace. It exposed /inbox (submit a task), /approval (answer a paused task), and /admin/* for housekeeping. Internally it routed each agent to either a local Ollama backend (default) or the Anthropic API (opt-in per agent, gated by a daily cost cap).

Triggering jobs from Android (historical)¶

Three first-class paths all ended up POSTing to /inbox on langgraph-agents — the difference was just the front door.

Option 1 — Zulip mobile app. Direct message to the triager-bot in the Zulip app. Triager listened on a webhook → forwarded to /inbox; the triager LLM classified the task and routed to a specialist agent (coder, errand-runner, homelab-engineer, smart-home-operator, etc.).

Option 2 — Voice intake (Wyoming + Whisper). The HA companion app's "Assist" button recorded audio → posted to HA → routed to Whisper (wyoming-services) → forwarded transcribed text to Windmill's langgraph-inbox webhook → same /inbox as the Zulip path.

Option 3 — HA companion app (templated tasks). A dashboard button calling the windmill_inbox script (an HA rest_command) for recurring asks.

Approval flow (historical)¶

The fleet used a four-class taxonomy:

Class	Examples	Default behavior
A	Read-only queries; vault writes; status reports	Run autonomously
B	Idempotent ops with trivial undo (toggle a HA light, set a thermostat)	Run autonomously
C	Stateful changes with non-trivial undo (reschedule a recurring HA automation, edit Frigate zones, schedule a CronJob, push a draft message)	Pause — needs approval
D	Irreversible or high-blast-radius (delete data, send a real email/SMS, ship to PR, restart a service)	Pause — needs approval; escalate to D only if undo_path is empty

When an agent proposed a Class C or D action, it paused and emitted an approval request over two channels (Zulip #approvals + ntfy/Pushover tap-to-approve buttons):

sequenceDiagram
    participant Agent as agent (langgraph)
    participant Windmill
    participant ZulipStream as #approvals
    participant Push as ntfy / Pushover
    participant You

    Agent->>Windmill: POST /webhook/approval-post<br/>{task_id, action_class, target, ...}
    Windmill->>ZulipStream: 🔔 New topic:<br/>"task-id — Class C: target"
    Windmill->>Push: Tier-1 push<br/>(tap-to-approve buttons)
    Note over You: 📲 You see push +<br/>Zulip notification

    alt approve via ntfy tap
        Push->>+Agent: POST /approval<br/>(pre-signed HMAC token)
        Agent-->>Agent: validate token,<br/>resume task
    else approve via Zulip @-mention
        You->>ZulipStream: @**Approval Receiver** approve
        ZulipStream->>Windmill: outgoing-webhook<br/>(message body + topic)
        Windmill->>Agent: POST /approval<br/>(HMAC-signed token)
        Agent-->>Agent: validate token,<br/>resume task
    end

    Agent->>ZulipStream: 📝 Task continues<br/>(status update in same topic)

Escalation timeline if you didn't respond: second push at 30 min; marked "cold" (no further pushes, still resumable) at 4 h; auto-cancelled at 7 d (with per-agent exceptions — health-tracker never auto-cancelled). Driven by the langgraph-awaiting-user-sweep Windmill flow, running every 5 minutes.

Where results landed (historical)¶

Long-form deliverables → the Obsidian vault, under ~/vaults/claude/reports/. The reporter agent wrote these.
Status updates + acknowledgements → the same Zulip topic as the approval request.
Daily roll-up → #digests stream, auto-posted at 22:00 ET by the langgraph-daily-digest cron flow.
Critical alerts → pushed straight to Pushover — this part is unchanged today.

The Windmill workflows that were removed¶

Workflow	Trigger	What it did
`langgraph-inbox`	Webhook (Zulip bot, voice, HA companion)	Forwarded to `/inbox`; if the task paused, fanned out an approval-post
`langgraph-approval-post`	Webhook (langgraph pause)	Posted approval request to Zulip `#approvals` + tier-1 push
`langgraph-approval-receive`	Outgoing-webhook (Zulip `@Approval Receiver`)	Verified actor + emoji/keyword; HMAC-signed token; POSTed `/approval`
`langgraph-awaiting-user-sweep`	Cron (every 5 min)	Tier-1 push at 30m; mark cold at 4h; auto-cancel at 7d
`langgraph-cost-cap-watcher`	Cron (every 4 h)	Push if today's Anthropic spend ≥ 80% (warn) or 100% (cap-hit)
`langgraph-daily-digest`	Cron (22:00 ET)	Triggered the reporter agent to write today's digest + posted summary to `#digests`
`langgraph-completion-post`	Webhook (langgraph task completion)	Posted a completed-task summary back to Zulip
`langgraph-dlq-watcher`, `langgraph-ml-weekly`, `langgraph-network-weekly`, `langgraph-observability-weekly`, `langgraph-renovate-triage`, `langgraph-reviewer-weekly`, `langgraph-storage-weekly`	Cron (various)	Weekly operator drift sweeps + DLQ retry, all fleet-specific
`smoke-approval-flow`	Manual	Smoke test for the approval round-trip above
`zulip-triager-webhook`	Outgoing-webhook	Forwarded Zulip DMs to langgraph's `/inbox`

All 16 files above are deleted from kubernetes/apps/home/windmill/workflows/ as of 2026-07-06. See Current state at the top of this page for the 7 that remain.