Observer

The Observer is the component that produces patterns from signals. Nothing else in Tapestry does interpretation; the Observer is the layer.

What it is

A component with two runtime modes:

• Scheduled cron — runs every 6 hours, scans your repos for category drift, writes synthesis memos + candidate findings.
• On-demand subagent — invoked from a Claude Code session via Agent({subagent_type: "liz-patterns:drift-watcher"}) or liz-patterns:agentic-upskilling to interpret a single project’s recent activity.

Lives at services/self-observer/ (Render cron self-observer).

Why it exists

Telemetry produces signals. Signals are not patterns. Without a dedicated interpretation layer, every consumer would have to re-invent interpretation locally, and every interpretation would die with the session that ran it.

The Observer is also why intent is observer-derived rather than emitted as a telemetry field. Intent is one specific instance of what the Observer produces.

How it interacts with the platform

flowchart TB
    T[Telemetry pipeline<br/>OTLP/HTTP + hooks.jsonl]
    M[Memory MCP<br/>memory-mcp]
    O[Observer<br/>self-observer cron + subagent]
    A[Architecture Registry<br/>durable structural facts]
    C[Candidate Registry<br/>pre-promotion patterns]
    OBSY[Observatory<br/>lens-equipped surface]
    T -->|signals| O
    M -->|prior memos + corrections| O
    O -->|candidates| C
    O -->|synthesis memos| M
    O -->|structural drift findings| A
    C --> OBSY
    A --> OBSY

The Observer reads from telemetry + memory and writes to the Architecture + Candidate registries plus back into memory (as synthesis memos). It does not call the Observatory directly — the Observatory reads the registries.

Setup

The Observer is platform-level. You stand it up as part of your deployment; your projects consume its output.

Consuming the existing deployment (default): nothing to install. The Observer is already running as the self-observer Render cron. Patterns it produces show up in the Architecture + Candidate registries and surface in the Observatory.

Self-hosting the Observer:

1. Copy services/self-observer/ into your Tapestry deployment.
2. Deploy as a Render cron — schedule from render.yaml (search for self-observer).
3. Required env vars (see services/self-observer/config.py):
   • MEMORY_BASE_URL — points to your Memory MCP
   • CANDIDATE_REGISTRY_URL — points to your Candidate Registry
   • ARCHITECTURE_REGISTRY_URL — points to your Architecture Registry
   • OTEL_EXPORTER_OTLP_ENDPOINT + OTEL_EXPORTER_OTLP_HEADERS — for emitting the observer’s own runtime telemetry
   • GITHUB_TOKEN — for scanning your repos
4. The cron’s signal-rule definitions live in services/self-observer/signal_rules.py; tune them to your project set.

See Platform dependencies for the full Render + Grafana setup.

Verify

• The cron ran recently: Render dashboard → self-observer → Logs → look for a synthesis memo emission within the last 6h.
• The Observer is producing candidates: query the Candidate Registry — curl https://your-registry-host.example.com/candidates?limit=5 should return recent entries with derivation_method set to observer.
• Memory contains synthesis memos: call memory_recall with the platform’s standard observer tag (e.g., ["self-observer-synthesis"]); recent memos should return.
• The Observatory surfaces Observer findings: open the Observatory console; the Observer lens should show non-zero findings if the cron has run.

Troubleshoot

Symptom	Likely cause	Where to look
Observatory cards look empty	Cron hasn’t run since the last reset, or no signals to interpret	Render logs for self-observer; if cron is fine, check that telemetry is flowing (see Telemetry)
Candidates not appearing in registry	Cron is running but registry POST is failing	Check the self-observer cron’s stdout for HTTP errors against CANDIDATE_REGISTRY_URL
Synthesis memo missing	Memory MCP unreachable	See Memory troubleshoot table
On-demand subagent never returns	liz-patterns or tapestry-patterns plugin not installed	/plugin list in Claude Code; reinstall if missing
Cron firing but no findings	signal_rules.py thresholds too high for the project’s signal volume	Lower the thresholds; re-run; check whether findings appear

Related

• The observer (explanation) — the conceptual page
• Memory, Telemetry, Registry, Observatory — the components the Observer interacts with