wshobson--agents
Seth Hobson
baa5bd7997
* feat(dgx-spark-ops): scaffold plugin and register in marketplace Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(dgx-spark-ops): add spark-environment-setup skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(dgx-spark-ops): tighten spark-environment-setup per review Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(dgx-spark-ops): add spark-training-gotchas skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(dgx-spark-ops): align preflight.sh output contract with docs Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(dgx-spark-ops): add spark-memory-thermal-ops skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(dgx-spark-ops): trim spark-memory-thermal-ops and source 68GB anchor Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(dgx-spark-ops): add dgx-spark-ops-engineer agent Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(dgx-spark-ops): defer agent facts to skills Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(dgx-spark-ops): add /spark-preflight command Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): scaffold plugin and register in marketplace Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add finetuning-method-selection skill with dated model catalog Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add lora-qlora-recipes skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add preference-optimization skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add grpo-rlvr-training skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(llm-finetuning): add isolation warning to execution reward Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add vision-sft skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add dataset-curation skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(llm-finetuning): correct loss-masking example in dataset-curation Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add eval-harness-first skill (Phase 0 gate) Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(llm-finetuning): bring eval-harness-first into line band Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(llm-finetuning): reclaim byte headroom in eval-harness-first Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add trace-to-training-data skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add checkpoint-promotion skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add quantized-export skill Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add llm-finetuning-architect agent Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add llm-finetuning-training-engineer agent Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add llm-finetuning-eval-engineer agent Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add /finetune phase-gated lifecycle command Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(llm-finetuning): thread checkpoint path into /finetune Phase 6 Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * feat(llm-finetuning): add /promote-checkpoint command Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(llm-finetuning): thread checkpoint path via phase5.output in /promote-checkpoint Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix(llm-finetuning): robust checkpoint discovery and goldens fingerprint in re-gate Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * docs: register llm-finetuning and dgx-spark-ops (94 plugins, 203 agents, 175 skills, 109 commands) Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * docs(agents): add fine-tuning and spark-ops agent entries (203 agents) Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * chore: generate per-harness artifacts for llm-finetuning and dgx-spark-ops Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix: final-review cleanups for fine-tuning plugins Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix: address PR #624 review feedback (G1 NGC detection, Phase 3 fallback dispatch, registry metadata) Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix: enforce sandbox boundary in execution grader and reward examples Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix: address CodeRabbit review findings on PR #624 Verified and fixed 37 of 43 outstanding CodeRabbit findings across the llm-finetuning and dgx-spark-ops plugins (skipping 6 confirmed false positives/already-fixed, with reasons in the disposition report). Highlights: cross-file contracts (goldens fingerprint persistence, paired-arena stage numbering, canonical golden-ID field, RERUN resolution before promotion) now match between finetune.md, promote-checkpoint.md, and checkpoint-promotion's templates. TRL API usages (SFTConfig.max_length, trl.experimental ORPO/CPO imports) verified against live TRL docs rather than blindly renamed. Several runnable examples hardened against real failure modes: malformed judge output, empty arena results, non-distinct DPO pairs, unbounded rejection-sampling fan-out, non-deterministic smoke-test comparisons, silent FP8-to-bf16 fallback, and orphaned background thermal-sampler processes. Container detection (G9) and LoRA adapter-size math fixed in both dgx-spark-ops and llm-finetuning where the same bugs were independently present in each plugin. Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix: apply dogfood friction-log remediations (F1-F32) from DGX Spark run Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf * fix: address CodeRabbit round-2 findings (digest pins, suite-size math, monkeypatch scoping) Claude-Session: https://claude.ai/code/session_01RsN3Vz5fZRTdVMkNtmxhSf