Diagnose a stated code, JAX, Marin, Iris, Zephyr, or TPU fault or startup/performance regression; do not activate for ordinary implementation or optimization without a symptom.
The English summary is being prepared. Showing the original description.
See how to installReview the instructions the agent will receive before you install.
Keep working notes in the active task. Do not add repository debug-log files.
At the start of a fresh investigation, use consult-echo when repository policy
requires prior-work search. During continued diagnosis or mitigation of the
same incident, reuse results already read instead of invoking consult-echo
again unless the scope or freshness requirement changes or the user asks for a
new search. After diagnosing a live infrastructure incident, use
write-ops-log to publish its standalone Echo record and link it from the
associated PR or issue. A code bug or local debugging session is not an
incident unless it caused a service, production run, or shared operational
system to fail or degrade.
Read lib/iris/AGENTS.md or lib/zephyr/AGENTS.md for context, then follow
the matching OPS.md section:
| Symptom | Read |
|---|---|
| Stuck job, scheduling failure, resource leak, controller stalled | lib/iris/OPS.md → SQL Queries, Process Inspection & Profiling, Known Bugs, Troubleshooting |
| Iris task misbehaving, container inspection, profiling a running task | lib/iris/OPS.md → Task Operations, Process Inspection & Profiling |
| Zephyr pipeline slow / stragglers / data skew / worker failures | lib/zephyr/OPS.md → Diagnostic Patterns, Observability |
TPU bad node (No accelerator found, FAILED_PRECONDITION, Device or resource busy) | lib/iris/OPS.md → TPU Bad-Node Recovery |
Read the guardrails beside the commands. Never modify the controller database,
prefer iris process profile over SSH, and never run a full
iris cluster restart without approval. After a TPU recovery or Zephyr fix,
return to the active Iris job-monitoring or babysit-zephyr loop.
For code bugs, reproduce the failure, identify the smallest falsifiable hypothesis, change one cause at a time, and test the behavior that failed. Let exceptions propagate unless added context changes the diagnosis.
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Run an explicitly requested Zephyr control/treatment benchmark on the same pre-normalized sample and compare Finelog stage metrics.
The English summary is being prepared. Showing the original description.
Register a named Hugging Face dataset for Marin by inspecting its schema and adding the appropriate experiments/datasets module.
The English summary is being prepared. Showing the original description.
Implement a fully specified request for a GCP IAM resource grant or Cloud Run IAP viewer in marin-iac.
The English summary is being prepared. Showing the original description.
Add or change a named Pallas/Mosaic kernel, including its reference implementation, correctness tests, wrapper, or requested tuning.
The English summary is being prepared. Showing the original description.
Archive explicitly named legacy experiment files by publishing a dated tag and canonical issue breadcrumbs.
The English summary is being prepared. Showing the original description.
Launch or continuously monitor a specified Zephyr pipeline on Iris only when asked to babysit it; restart only after explicit approval.
The English summary is being prepared. Showing the original description.