Meta Core AI Agent

META-CORE — Telemetry Layer v1.0

Portable agent-engine core. This repo contains the working, tested part of the architecture: the decoupled telemetry bus that every META-CORE node writes to. The rest of the design (Flow-Core, Control Plane, Chronos V3) is specified in docs/ARCHITECTURE.md and tracked as roadmap.

Core invariant

A core never reports its state to a specific recipient. It publishes a fact via the Transponder. Anyone who needs state reads the event stream.

agent / ADK node ──► transponder.py ──(flock+fsync append)──► event_stream.jsonl
                                                                   │
                         observer.py (read-only) ◄─────────────────┤
                         rotate.sh  (100 MB → archive/*.gz) ◄──────┤
                         log_sync.sh (git batch push, 5 min) ◄─────┘

Layout

Path	Purpose
telemetry/transponder.py	Validates event (schema v1.0, 9-event ENUM), enriches, atomic append. Zero network/git. CLI + emit() API.
telemetry/observer.py	Passive parser: trace trees, handoff depth, cycles, orphan spawns, stale cores. Exit 0/1/2 = OK/WARN/CRITICAL.
telemetry/log_sync.sh	Batch git commit/push only if changes exist; flock; exponential backoff; errors → log_sync.err.
telemetry/rotate.sh	Copy+truncate rotation under the writer lock, gzip into archive/.
schemas/event_bus_payload.v1.json	JSON Schema 2020-12 contract.
deploy/systemd/	Timers for sync (5 min) and observer (1 min).
examples/demo.py	Two-core handoff + spawn simulation.
tests/	stdlib unittest, incl. 8-process concurrent write test.

Quick start

cp .env.example .env
make test     # 14 tests
make demo     # writes demo events, runs observer

export CORE_ID=core-001 EVENT_LOG_PATH=telemetry/event_stream.jsonl
python3 telemetry/transponder.py '{"event":"STEP_COMPLETED","execution_step":"18/42","data":{"status":"success"}}'
python3 telemetry/observer.py --json

From Python:

from telemetry.transponder import emit
emit({"event": "STEP_STARTED", "core_id": "core-001", "trace_id": tid, "span_id": sid, "parent_span_id": parent})

Requirements: Python 3.10+ (stdlib only), bash, git, util-linux (flock). Linux/macOS (uses fcntl).

Event contract (v1.0)

Required: schema, event_id, core_id, event, timestamp. Events: CORE_STARTED CORE_STOPPED HEARTBEAT STEP_STARTED STEP_COMPLETED STEP_FAILED SPAWN_REQUESTED SPAWN_COMPLETED FATAL_ERROR. Execution events (STEP_*, SPAWN_*) always carry trace_id/span_id — generated as a root span if the caller omits them. Optional objects: data, context_diff, guardrail_state, metrics.

Deploy (systemd)

sudo cp -r . /opt/meta-core && sudo useradd -r metacore && sudo chown -R metacore /opt/meta-core
sudo cp deploy/systemd/* /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable --now log-sync.timer observer.timer

Wire observer.service exit code 2 to your kill switch (OnFailure=).

Decisions

• ADR-001 — No git in the Transponder. Push is batched by log_sync.sh (rate limits, latency, lock contention).
• ADR-002 — trace_id conditional: optional for lifecycle/heartbeat, auto-generated for execution events.
• ADR-003 — Rotation is copy+truncate under the writer's flock (rename could lose in-flight appends).

License

TBD