Help coding agents work on an NVFlare-based Auto-FL harness that follows an autoresearch-style loop. Use when the user wants to create, edit, debug, or extend program.md, task folders such as tasks/cifar10/ and tasks/vlm_med/, task-local job.py, client.py, model.py, shared custom_aggregators.py, mutation policies, results.tsv logging, or coding-agent prompts for a bounded federated-learning research loop. This skill is specifically for NVFlare harness work where the Client API loop, DIFF upload contract, and NUM_STEPS_CURRENT_ROUND metadata must stay intact unless the user explicitly asks for a protocol upgrade.
Instrucciones de origen · Vista previa de solo lectura
name
autofl-nvflare
description
Help coding agents work on an NVFlare-based Auto-FL harness that follows an autoresearch-style loop. Use when the user wants to create, edit, debug, or extend program.md, task folders such as tasks/cifar10/ and tasks/vlm_med/, task-local job.py, client.py, model.py, shared custom_aggregators.py, mutation policies, results.tsv logging, or coding-agent prompts for a bounded federated-learning research loop. This skill is specifically for NVFlare harness work where the Client API loop, DIFF upload contract, and NUM_STEPS_CURRENT_ROUND metadata must stay intact unless the user explicitly asks for a protocol upgrade.
autofl-nvflare
Use this skill to keep edits to the Auto-FL NVFlare starter coherent, safe, and aligned with an autoresearch-style operating model.
Entry point
When the target repo includes program.md, read it first and treat it as the general control plane. Then read the active task profile; use tasks/cifar10/profile.md when the human does not specify another profile.
Use the active task's mutation_schema.yaml for bounded mutation details only when program.md or the active task profile points you there, or when choosing a mutation axis. Use AGENTS.md / CLAUDE.md only as thin local guardrails.
Core rules
Preserve these invariants unless the user explicitly asks for a protocol change:
the same selected model_arch on server and clients for a run
the active max_model_params cap for architecture campaigns
Preferred mutation order
Client-local changes in the active task's client.py
optimizer family
scheduler settings
local epochs, fixed local training steps, batch size, workers
weight decay
gradient clipping
label smoothing
FedProx local loss
extra scalar metrics
Aggregation changes in shared tasks/shared/custom_aggregators.py
weighted aggregation refinements
FedAvg/FedOpt-style DIFF aggregation that stays inside the existing FLModel contract
explicit SCAFFOLD control-variate metadata when the user has opted into that protocol mode
clipping / robust aggregation
median or trimmed-mean style logic
Recipe changes in the active task's job.py
rounds
clients
cross_site_eval
launch_external_process
client_memory_gc_rounds
Registered architecture changes in the active task's model.py
named model_arch variants
parameter-count checks through max_model_params
no new dependencies
Do not change model architecture outside registered model_arch variants or the active max_model_params budget. Do not add server-coupled protocol fields outside an explicitly requested protocol mode.
FedProx is compatible as a client-local loss term. FedOpt is compatible only when it is implemented inside the custom aggregator over already-received DIFFs. SCAFFOLD is available only as an explicit opt-in mode that uses FLModel.meta for scaffold_c_diff and scaffold_c_global.
Required workflow
After making edits:
use the interpreter and dependency rules from the active task profile. For the default CIFAR-10 profile, set and use PYTHON=.venv/bin/python by default, unless the human explicitly provides a different PYTHON value; treat the selected value as authoritative, verify it with test -x "$PYTHON" and "$PYTHON" -c "import sys; assert sys.version_info[:2] == (3, 12), sys.version; print(sys.executable)", and do not search for alternate interpreters with glob or discovery commands such as ls /usr/bin/python*, ls /workspace/.venv*/bin/python*, or which python
do not create virtual environments or install dependencies unless the user explicitly asks; if the active profile's interpreter is missing or invalid and no override was provided, tell the user to rerun that profile's preflight instead of guessing
when initializing a campaign, use a descriptive branch tag with the pattern <node>-<campaign-topic>-YYYYMMDD, such as h100-fedavgm-20260430 or h100-archsearch-20260430; run bash scripts/init_run.sh <tag> before validation, baseline, or candidates; verify git branch --show-current starts with autoresearch/; never run experiments on main, upstream/main, the starter branch, or a shared feature branch; never use date-only branch names
run the active task profile's static checks and syntax validation, with TASK_DIR set to the active task
run the client contract validator against the active task's client.py, not a stale root-level path
run the active task profile's smoke command if the prepared environment has nvflare; for non-CIFAR tasks, pass the task-specific SMOKE_ARGS or use scripts/run_iteration.sh with the active task budget
follow the active task profile's local hardware and candidate-width rules. For the default CIFAR-10/H100 profile, launch up to PARALLEL_CANDIDATES=4 same-budget candidates concurrently on one local H100 when memory allows, and reduce the width if candidates hit CUDA OOM or host contention
use the active task profile's default candidate budget unless told otherwise. For the default CIFAR-10/H100 profile, that budget is 8 clients, 20 communication rounds, 4 local epochs, local_train_steps=0, training batch size 64, eval batch size 1024, alpha 0.5, seed 0, model_arch=moderate_cnn, max_model_params=5000000, weighted aggregation, deterministic client training, final global evaluation on site-1, and a 1200-second timeout; local epochs or may be swept under that runtime cap, but do not vary both in the same narrow sweep
References
Read these when relevant:
references/mutation-schema.md for the allowed mutation surface
references/runbook.md for the recommended iteration loop and reporting format
references/provenance.md for acknowledgement and attribution guidance
local_train_steps
use unique RUN_LOG and job --name values for each candidate; if the active profile requires one local GPU, pin each run with CUDA_VISIBLE_DEVICES=0 instead of spreading candidates across devices
record the outcome in results.tsv; run_iteration.sh initializes the header before launching logged runs, and successful runs are appended as candidate, which means unreviewed, not kept
after every completed batch, update reviewed results.tsv statuses before launching the next batch: promote the selected survivor to keep, mark reviewed non-survivors as discard, leave crashes as crash, and leave only unresolved active rows as candidate; prefer scripts/finalize_batch_status.py --last "${PARALLEL_CANDIDATES:-4}"
commit that ledger on the active autoresearch/ branch after baseline and completed runs/checkpoints, and commit surviving code changes as soon as they are kept rather than carrying them uncommitted into the next batch
if a candidate implements a paper-derived method, include a compact source ref in the results.tsv description field and fuller citation details in templates/mutation_report.md
rank the completed batch against the ledger before deciding whether to keep, narrow, or revert; rank primarily by score, use runtime as a coarse secondary signal, and prefer the faster/simpler candidate when scores are within noise
after setup and baseline, continue launching same-budget candidate batches until manually interrupted; do not ask whether to keep going
after every finalized batch, run scripts/plateau_watchdog.py results.tsv; if it prints recommendation=literature, stop local jitter sweeps and run the Camyla-inspired literature loop from program.md: time it with scripts/log_literature_review.py --start / --finish, generate diverse queries, triage primary papers, extract challenge cards, score contract-safe proposals in templates/literature_loop.md, record the literature event row in results.tsv, and launch the top compatible candidate batch next; if it prints recommendation=continue, do not log another literature row for a routine missed batch, and keep iterating locally unless repeated crashes share one root cause or no non-duplicate safe axis remains
report the mutation hypothesis, changed files, commands run, observed outcome, literature basis, run analysis, and next mutation
