OpenThoughts-Agent
CLAUDE.md309 starsChanged 4 months agoApache-2.0View on GitHub
Guidance for Claude Code when working in this repository. ot-agent is a distributed LLM training + evaluation system for HPC clusters, with four subsystems — datagen (Harbor/Daytona traces + standard vLLM generation), SFT (LLaMA-Factory), RL (SkyRL/GRPO), eval (terminal-bench / agentic). One unified launcher: python -m hpc.launch --job_type <type>. This file is a thin index. The real, maintained documentation lives under .agents/ — read the relevant piece for the task at hand (skills are also invocable by name via the Skill…
In this repository50 other files it gives its agents.
- analyze-datagen-campaign-summarySkill
- analyze-dataset-token-lengthSkill
- analyze-id-eval-rankingSkill
- analyze-job-history-irisSkill
- analyze-rl-behaviorSkill
- analyze-training-run-irisSkill
- build-tpu-image-irisSkill
- code-create-staged-planSkill
- code-execute-staged-planSkill
- commitSkill
- crud-archive-runSkill
- crud-otagent-supabaseSkill
- crud-purge-below-gate-evalsSkill
- crud-purge-stale-eval-placeholdersSkill
- datagen-create-task-datasetSkill
- datagen-job-cleanupSkill
- datagen-launch-irisSkill
- datagen-launchSkill
- datagen-reduce-dataset-snapshotsSkill
- datagen-standard-launchSkill
- debugSkill
- docs-deslopSkill
- eval-agentic-cleanupSkill
- eval-agentic-launch-irisSkill
- eval-agentic-launchSkill
- eval-standard-cleanupSkill
- eval-standard-launchSkill
- file-issueSkill
- monitor-cron-sweep-irisSkill
- monitor-cron-sweepSkill
- monitor-job-tablesSkill
- monitor-restore-irisSkill
- monitor-restoreSkill
- monitor-restore-unifiedSkill
- rl-agentic-job-cleanupSkill
- rl-agentic-launch-irisSkill
- rl-agentic-launch-jupiterSkill
- rl-job-health-deep-diveSkill
- rl-standard-job-cleanupSkill
- rl-standard-launch-leonardoSkill
- serve-model-vibe-testSkill
- sft-cleanup-hf-onlySkill
- sft-job-cleanupSkill
- sft-launchSkill
- supervisor-initSkill
- utils-cleanup-stale-sandboxesSkill
- utils-fix-permissionsSkill
- utils-reclaim-stale-snapshotsSkill
- write-testsSkill
- writing-styleSkill
The file
We split this CLAUDE.md into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 30
Local clones are ground truth; clusters never diverge: All edits go in local checkouts → push → git pull on the cluster.
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
Guidance for Claude Code when working in this repository.
ot-agent is a distributed LLM training + evaluation system for HPC clusters, with four subsystems —
datagen (Harbor/Daytona traces + standard vLLM generation), SFT (LLaMA-Factory), RL (SkyRL/GRPO),
eval (terminal-bench / agentic). One unified launcher: python -m hpc.launch --job_type <type>.
Source of truth = .agents/
This file is a thin index. The real, maintained documentation lives under .agents/ — read the relevant
piece for the task at hand (skills are also invocable by name via the Skill tool):
.agents/skills/<name>/SKILL.md— operational how-tos, one per task. By prefix:- launch:
rl-agentic-launch-jupiter,rl-agentic-launch-iris(CoreWeave H100),rl-standard-launch-leonardo(non-agentic GRPO),sft-launch(LLaMA-Factory + axolotl · Delphi),datagen-launch(agentic Harbor trace-gen),datagen-standard-launch(Curator +generate.py),eval-agentic-launch,eval-standard-launch(+*-irisvariants). - cleanup:
rl-agentic-job-cleanup(agentic RL + traces),rl-standard-job-cleanup(standard GRPO — model + metrics only, no traces),sft-job-cleanup,datagen-job-cleanup,eval-agentic-cleanup,eval-standard-cleanup. - monitor:
monitor-cron-sweep,monitor-job-tables,rl-job-health-deep-dive(per-RL-job probe → KILL/NO-KILL),monitor-restore(3-hourly sweep loop),monitor-restore-iris-cron. - analysis / data / db:
analyze-rl-behavior,analyze-dataset-token-length,analyze-id-eval-ranking(z-score ranking),datagen-reduce-dataset-snapshots,crud-otagent-supabase,crud-purge-stale-eval-placeholders,crud-purge-below-gate-evals. - code:
code-create-staged-plan→notes/<codebase>/,code-execute-staged-plan→agent_logs/. - build:
build-gpu-rl-image-iris(kaniko in-cluster; Mac can't build it). - role:
supervisor-init— session bootstrap.
- launch:
.agents/projects/<dep>/— facts & gotchas per codebase:ot-agent/(branches + launcher map),marin/,marinskyrl/,harbor/,vllm/,llama-factory/,axolotl/,daytona/,ajudge/; plus technique docs —pyspy/(live-hang py-spy capture method — read this when the operator asks to "go in with py-spy on the live job" / "scope the failure live")..agents/ops/<target>/— machine/cluster particulars (access, paths, env/SIF map, gotchas):jupiter/,leonardo/,torch/,iris/,local/(this Mac),all/(cross-cluster HF/tmux),experiments/(the per-experiment tracker workspace~/Documents/experiments),data/(dataset trackers — e.g.tasktrove.md, the full TaskTrove inventory)..agents/secret.md— untracked, gitignored; holds privileged values (pinggy bank, etc.) pulled out of the committable docs. Referenced by name from skills/ops.
Always (apply before any skill loads)
- Run Python via the otagent env's full interpreter path (symlinks don't work in the sandbox):
/Users/benjaminfeuer/miniconda3/envs/otagent/bin/python. (curatorenv only for Curator datagen.) - Syntax/lint check with the IDE MCP tool
mcp__ide__getDiagnostics, NOTpython -m py_compile/flake8(bash output capture is unreliable here). - Local clones are ground truth; clusters never diverge. All edits go in local checkouts → push →
git pullon the cluster. Branch flow differs by repo:- marin-community forks (
harbor,MarinSkyRL,evalchemy): worktree → PR →main.penfever/workingis RETIRED. A subagent never self-merges a marin PR. OpenThoughts-Agent+vllmfork: stay onpenfever/working(commit → push; may self-merge).- Clusters
git pullthe tracked branch (Python repos are editable installs, live after pull). An unmerged marin-fork fix rides--harbor-ref/--skyrl-ref, never a shared-branch mutation. No untracked/divergent changes on a cluster; no patch-by-rsync; no hand-editing. vLLM (compiled) is built from source on each cluster from the committed fork.
- marin-community forks (
- Standing ML-ops guardrails (full statements in
monitor-restore/ the cleanup skills):enable_db_registration: falsein YAMLs (manual DB register only); ≤6 RUNNING RL jobs per cluster (Daytona); a3 series CONCLUDED; Daytona snapshot caps are HARD (clean stale, never raise); cross-user FK safety pre-check before any Supabase delete/mutate; HF uploads default PUBLIC tolaion/; never kill a RUNNING job without explicit permission.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

