| name | devnet-debug |
| description | Debug hosted/remote Ethereum devnets (ethpandaops networks like glamsterdam-devnet-N) with a Lodestar lens. Builds on the panda `query` + `investigate` skills and the 'debug devnet' runbook for the data/procedure; adds the cross-client 'read the failing client's own logs' method and a ChainSafe infra POV (SSH to Lodestar nodes, ChainSafe Grafana Loki). Hosted devnets, not local Kurtosis. |
Devnet debugging (hosted / remote) — Lodestar lens
This skill layers on panda's own debugging skills — it does not replace them. For the data and the procedure, use panda:
query skill — the panda data/query API. Discover datasources, tables, and worked queries live (panda getting-started, panda schema, panda search examples "<topic>"). Don't hardcode datasource/table names — they're owned by the proxy and change.investigate skill — routes local-vs-remote and loads the canonical runbook. For a hosted devnet, run panda search runbooks "debug devnet" and follow it: it owns the flow (Dora shape, external.otel_logs, ethnode RPC, debug-report output).
devnet-debug adds only what panda has no knowledge of: the Lodestar interpretation, the cross-client correlation method, and a ChainSafe infra secondary POV.
Not this skill: local Kurtosis → kurtosis-devnet; join with a local node → join-devnet; mainnet networking repro → local-mainnet-debug; Lodestar RC metrics → release-metrics; heap leaks → lodestar-heapsnapshots; bulk log triage → log-reader.
Flow
- Preflight panda.
skills/devnet-debug/scripts/ensure-panda-auth.sh — heals the cred-file perms so the containerized panda-server can read them, and reports auth state. Token refresh is handled by panda-server itself; do NOT run scripts/panda/panda-reauth (the browser/GitHub-cookie device flow is disabled per nflaig 2026-06-19). If datasources=null persists, a human runs panda auth login once. Then panda's own readiness check: scripts/debug/check-devnet-routing-readiness.py <network> (exit 2 = datasources not ready; panda datasources --json returning {"datasources": null} means not ready, not "network absent").
- Run the panda procedure.
panda search runbooks "debug devnet" → follow it. Use the query skill / panda search examples for query patterns; discover names live.
- Apply the Lodestar lens (below).
- Drop to ChainSafe infra (below) only when you need Lodestar internals panda doesn't ship.
Triage — symptom → first look
Quick public snapshot (no auth/panda): skills/devnet-debug/scripts/net-health.sh <network> → finality, active forks, node→client topology. Then:
| Symptom | First look |
|---|
| Not finalizing / low participation | dora.get_network_overview (or /api/v1/epoch/latest: finalized, participation) → per-client error signatures in otel-logs |
| A client stuck / on its own fork | Dora /forks (who diverged) → that client's own otel-logs → its live beacon API via the bn- gateway |
| ePBS / Gloas (glamsterdam) | payload-envelope / PTC / builder errors in the stuck client's logs (e.g. payload envelope … not found in forkchoice); confirm GLOAS_FORK_EPOCH (cl/config.yaml) is actually active |
| PeerDAS data columns missing | data-column-sidecar tables (discover via panda schema / search) + custody/sampling errors in CL logs |
| Peers dropping / low count | Lodestar lodestar_peers_by_client_count (:5054 / Prometheus); peer/disconnect lines in otel-logs; pin the offending client |
| Lodestar block production failing | SSH :5052 /eth/v1/debug/fork_choice; ChainSafe Loki debug logs; scripts/debug/devnet-triage.sh <node> |
| Did the network pass its tests? | Assertoor /api/v1/test_runs (failed runs) |
Always read the failing client's own logs — don't infer its bug from Lodestar's side. Once root-caused, hand the mechanism + the exact log lines to that client's team (don't just say "looks like X").
Network metadata & explorers
Every hosted devnet has a landing page https://<network>.ethpandaops.io/ linking all services + machine-readable config — scope a network from there.
config.<network>.ethpandaops.io: /api/v1/nodes/inventory (authoritative node→client map — client, image tag, ENR, peer_id, bn- beacon URI per node), /api/v1/nodes/validator-ranges (validator index → node), /cl/config.yaml (fork schedule, e.g. GLOAS_FORK_EPOCH; far-future 18446744073709551615 = not scheduled), /cl/genesis.ssz, /el/genesis.json. Client versions also in repo ethpandaops/<network>s (…/images.yaml).- Services:
rpc. (EL RPC), beacon. (public CL REST), dora., forkmon., syncoor. (sync tests), assertoor. (network test runner — /api/v1/test_runs & /api/v1/tests JSON for pass/fail), checkpoint-sync., faucet..
- Dora — prefer the panda module:
from ethpandaops import dora → dora.get_network_overview(net) (epoch/slot/finality/participation/validator counts), get_epoch/get_slot/get_validator(s), link_*. Quick raw curl (no auth): /api/v1/epoch/latest (finality + participation), /api/v1/slot/<n>, /api/v1/slots, /api/v1/validators, /api/v1/validator/<idx> (JSON); /forks is the HTML fork view (one row per fork = a split).
Full endpoint catalog + examples: references/network-metadata.md.
Query discipline — prefer panda execute
panda's whole point is the sandbox Python runtime, not raw SQL dumps. A panda clickhouse query returning thousands of rows floods context ("context rot") — the exact thing panda was built to avoid (see the ethpandaops panda post). So:
- Aggregate in the sandbox, return summaries.
panda execute --code '...' with the ethpandaops lib (clickhouse/prometheus/loki/dora/specs) → return df.describe(), group-by counts, or top-N — not raw rows.
- Logs: group by error signature, return the top distinct patterns + counts, then drill into a couple of examples. Don't dump hundreds of lines.
- Raw
panda clickhouse query only for tiny results (a count, a handful of rows).
- Sessions for multi-step:
panda execute --session <id> keeps the sandbox warm; cache an expensive pull to /workspace/*.parquet and reuse it next call.
- Spec constants inline:
from ethpandaops import specs; specs.get_constant("MAX_EFFECTIVE_BALANCE").
Execute-based otel-logs example: references/panda-recipes.md.
Lodestar lens — is it us?
On an interop devnet the high-value move: when a client is stuck/forked, read that client's own logs — don't infer its bug from Lodestar's peer view. The runbook's external.otel_logs carries every client; key by network + host.name (<cl>-<el>-<n>, e.g. prysm-nethermind-2), split CL/EL via log.file.name, match severity on Body. Confirm current table/column names via panda schema / panda search examples before trusting any literal.
When Lodestar is implicated, cross-check head/finality/peers, then go to the secondary POV for internals.
Worked example (Prysm "every node a fork" on glam-devnet-5 → Execution payload envelope … not found in forkchoice) and the exact otel-logs query pattern: references/panda-recipes.md.
Secondary POV — ChainSafe infra (panda has none of this)
Use when you need Lodestar internals panda doesn't carry: live fork-choice dump, debug-level logs, heap/CPU profiles, exact peer-by-client counts.
- SSH (Lodestar nodes only):
ssh devops@lodestar-<el>-<n>.srv.<network>.ethpandaops.io (key ~/.ssh/id_ed25519 = the lodekeeper.keys entry). Beacon REST localhost:5052, metrics localhost:5054/metrics. Other-client nodes reject the key — read theirs via panda otel-logs (logs) or the bn- gateway below (live beacon API). Peer-by-client: lodestar_peers_by_client_count{client="Prysm"}.
- Any-client beacon API (
bn- gateway): reaches every client's beacon API (Prysm/LH/Teku/Nimbus/Grandine/Lodestar), unlike the SSH key. AUTH=$(awk -F': ' '/^<network>:/{print $NF}' ~/.config/ethpandaops/bn-basic-auth) then curl "https://$AUTH@bn-<cl>-<el>-<n>.srv.<network>.ethpandaops.io/<beacon-path>" — HTTP basic auth (user eth). Use for live head/finality/peers/version/syncing of a stuck non-Lodestar node, and Lodestar debug routes (/eth/v1/lodestar/...) on ours. Password is per-devnet and not in this repo — stored 0600 at ~/.config/ethpandaops/bn-basic-auth; get fresh per-devnet creds from Nico / the devnet config. (Bare <node>.<network> without bn- 404s.)
- ChainSafe Grafana Loki: ChainSafe's own Lodestar devnet nodes ship debug-level logs to Loki (datasource 4) under
group="beacon_devnet", network="dev", instances devnet-ax41-0..3. Faster than panda for Lodestar peer/disconnect/sync digs, no OIDC. Token: eval "$(grep '^export GRAFANA' ~/.bashrc)". See grafana-loki skill.
- One-shot per-node snapshot: before a longer run, preflight telemetry with
scripts/debug/devnet-triage.sh <node> --check-only --require-grafana when Loki/Prometheus are needed. Autonomous wrappers should use --check-only --json for machine-readable readiness. Then scripts/debug/devnet-triage.sh <node> pulls that node's recent error logs (Loki) + key metrics (Prometheus) into a markdown report (needs GRAFANA_TOKEN).
Full commands + "which POV when": references/chainsafe-infra.md.
Self-Maintenance
If any commands, file paths, URLs, or configurations in this skill are outdated or no longer work, update this SKILL.md with the correct information after completing your current task. Skills should stay accurate and self-healing — fix what you find broken.
Iteration Log
Track what works and what doesn't after each use:
| Date | Network / issue | What worked | What to improve |
|---|
| 2026-06-16 | glamsterdam-devnet-5 — Prysm "every node a fork" | Cross-client otel-logs surfaced Prysm's own Execution payload envelope … not found in forkchoice lines — named the mechanism, not just "looks like Prysm". config/api/v1/nodes/inventory + the bn- gateway gave topology/versions/live beacon API without SSH. | First pass dumped raw log rows into context (costly) → switched to panda execute aggregation. Confirm live table/datasource names via panda schema / search examples before querying — don't trust the literals here. |
