tao-run-deft-aoi

Skill: tao-run-deft-aoi

When to Use This Skill

Use this skill when the user wants an agent to run the full DEFT AOI improvement loop for an NVIDIA TAO VisualChangeNet / ChangeNet PCB inspection model: baseline evaluation, RCA, synthetic defect generation, data mining, retraining, and deployment gating until a KPI target is met.

• "Run the DEFT loop"
• "Fine-tune until FAR below 0.1% at recall=100%"
• "Improve my AOI ChangeNet model using RCA and synthetic defects"
• "Iterate training until false accept rate meets the target"

Do not use this skill for a single standalone TAO training run, one-off inference, generic anomaly generation, or RCA-only analysis. Use the relevant agent directly when the user asks for only that step.

Base Model

The loop operates on NVIDIA TAO Visual ChangeNet classify with the NVIDIA C-RADIOv2-B backbone, fine-tuned end-to-end. The architecture is defined in specs/baseline_spec.yaml — that file is the source of truth. All pretrained weights come from HuggingFace (HF_TOKEN required); NGC_KEY only gates container pulls. ChangeNet backbone resolution + the staged-file/HF-URL fallback for model.backbone.pretrained_backbone_path are owned by references/visual-changenet.md. SigLIP for k-NN mining is owned by references/tao-mine-aoi-images.md. AnomalyGen-side checkpoints (Cosmos-Predict2, T5, NVDINOV2, C-RADIO-V3, DINOv2-large, SAM2, Qwen3-VL — ~22 GB for 2B-only, ~140 GB with 14B + T5-11b) live under <workspace>/augmentation/anomalygen/base_checkpoints/; the paidf-anomalygen container auto-downloads them on first use. The PCB reference dataset under <workspace>/augmentation/anomalygen/datasets/<project>/ is also auto-fetchable. See references/paidf-anomalygen.md.

Train AutoML Policy

DEFT AOI owns the iterative data-improvement loop, retraining cadence, and KPI checkpoint selection. For this workflow only, bypass model-level AutoML even when the underlying Visual ChangeNet model metadata has automl_enabled: true.

automl_policy: off is a workflow argument to the Visual ChangeNet skill invocation (the value the parent passes when calling tao-skill-bank:tao-train-visual-changenet via the Skill tool), not a TAO spec field. Two cases:

• Direct docker run visual_changenet train -e <spec> (the path this workflow actually uses inline): no action needed. The TAO entrypoint is plain training by default; AutoML lives behind a different code path that the SDK orchestrates. Effectively, every direct docker run is already automl_policy: off.
• SDK-orchestrated dispatch (Brev/SLURM/k8s with the SDK building the command): pass automl_policy: off to VisualChangeNetSDK.train(...) or the equivalent runner argument. The SDK uses it to pick the plain-train command instead of the AutoML wrapper.

Never add automl_policy or a workflow key to the spec YAML. TAO's Hydra ExperimentConfig schema does not recognize these keys and the train job fails at config-merge time with Error merging '<spec>.yaml' with schema: Key 'workflow' not in 'ExperimentConfig'. This is a workflow-level override only; do not change model metadata, and do not apply this policy to other workflows.

Launch Intake

After the user confirms they want to run this workflow, ask which supported platform they intend to run on. Generate the platform choices with:

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/list_tao_platforms.py \
  --skill-bank ${TAO_SKILL_BANK_PATH:-~/tao-skills-external} --format text

After platform selection, run:

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/list_tao_platforms.py \
  --skill-bank ${TAO_SKILL_BANK_PATH:-~/tao-skills-external} \
  --platform <platform> --format text

Ask only for credentials relevant to that platform, plus model-specific credentials required by the selected workflow.

Agent Behavior

There is exactly one user gate: pre-flight confirmation. Print the Pre-Flight Summary (see references/preflight.md → Pre-Flight Summary), then STOP and wait for the user to type "go", "yes", "looks good", or similar explicit approval. Do not launch any side-effecting step (docker run, training, SDG, mutations under ${RESULTS_DIR}/) before that approval — reading specs, listing files, docker image inspect, and populating the summary table are fine. "Autonomous" describes behavior after this gate, not before it. Do not skip the gate even if the user's original prompt sounded urgent ("just run it", "go ahead") — the summary itself is the artifact they need to see before approving.

After the gate, the skill is fully autonomous. Run the entire loop without asking for confirmation. Do not pause between steps. Do not ask "want me to continue?" — just continue. Only stop if a step fails with an unrecoverable error or a hard-stop gate fires. Print a one-line status update at each step milestone so the user can follow progress.

Auto-mode required. The post-gate loop fires constant side-effecting calls (docker run, ${RESULTS_DIR}/ writes); without auto-accept / bypass-permissions mode it stalls on the first prompt. Remind the user at the Pre-Flight Summary to enable auto-mode (shift+tab) before approving.

Blocker recovery. Fix recoverable blockers yourself — missing image (pull), unstaged C-RADIO backbone (stage .pth per references/visual-changenet.md), missing pydeps (venv), absent AnomalyGen assets (paidf auto-fetches) — then resume the Pre-Flight step you were on (<blocker> cleared → resuming step N) and continue to the Summary. Halt only for what you can't fix (missing workspace/specs/CSVs/credentials, empty pool, leakage). A fix is not the user gate.

Revised plan. If any run parameter changes after the original summary was shown (user imposes a time limit, overrides epochs, changes max_iterations, etc.), always re-run Pre-Flight and show an updated summary before proceeding.

Workflow

Execute the loop in this order (full detail in references/pipeline-and-state.md → Pipeline + Stage Execution):

1. Pre-Flight. Run every check in references/preflight.md. Resolve workspace, specs, CSVs, checkpoints, container images. Hard stop only on missing input you can't resolve yourself (see ## Agent Behavior → Blocker recovery).
2. Baseline. If deft_state.json already has iterations.baseline.stage_completed == "train" and a best_ckpt_path pointing at an existing file (the upstream automl-deft-pipeline pre-seeds these from its Phase 1 AutoML winner — see its Phase 1 → Phase 2 handoff), skip the train sub-step and resume at inference -> evaluate against the pre-seeded checkpoint. Otherwise run train -> inference -> evaluate by invoking the tao-skill-bank:tao-train-visual-changenet skill. Either way, then rca by invoking tao-skill-bank:tao-analyze-gaps-visual-changenet. Read references/visual-changenet.md and references/tao-analyze-gaps-visual-changenet.md first for DEFT-loop-specific args (mounts, output dirs, deft_state.json updates).
3. Iterate. For each iteration up to max_iterations, execute Pipeline steps 1-7. Between every step, re-read results/loop_log.jsonl tail + results/deft_state.json from disk — disk is canonical.
4. Stop when the KPI target is met, max_iterations is reached, or a hard-stop gate fires (silent-drop, AMP allocation mismatch, train/val leakage). Never auto-retry hard stops.
5. Render results/DEFT_Loop_Report.html after each completed iteration (and once more at loop end) by spawning the reporter subagent (agents/reporter.md). Per-stage renders are not done — every stage already appends one line to loop_log.jsonl, which is enough for a tail-watching user; the HTML render carries an iteration's worth of state and one render per iteration keeps the per-loop token cost roughly linear in iteration count, not in stage count. Do not render inline.

All pipeline stages run inline in the parent context — the parent invokes the underlying tao-skill-bank:* skills directly via the Skill tool, layering DEFT-loop conventions on top via the matching references/*.md file. The only delegated work is HTML report rendering, handled by the reporter subagent in a fresh context so an end-of-loop render is never silently dropped when the parent's context is saturated. See references/scripts-and-agents.md → Agents for the reporter spawn contract.

Using Bundled Scripts

Run bundled scripts from scripts/ via run_script() when the harness provides it (a Claude Code plugin runtime helper, not a function defined in this repo); otherwise fall back to direct python. Resolve every path argument to an absolute host path first. Never write loop_log.jsonl via echo or inline jq — the seq invariant requires reading the live tail through next_seq(). See references/scripts-and-agents.md for the full Available Scripts table, the agents/reporter.md spawn contract, the Stage Reference Modules stage→skill mapping, the path-rule invariant, and the workflow-level AutoML-policy pitfall. For per-script invocation examples, see references/SCRIPT_USAGE.md.

Stage Reference Modules

Each pipeline stage maps to one underlying skill in the bank; the matching references/*.md file layers DEFT-loop conventions (mounts, output dirs, deft_state.json updates, log_stage.py summary string) on top of the skill's generic instructions. Read the reference file first, then invoke the skill via the Skill tool. If a reference file is missing, stop and ask the user to reinstall the plugin. The full stage→reference→skill→ownership table lives in references/scripts-and-agents.md → Stage Reference Modules. The stages: train/evaluate (references/visual-changenet.md), anomalygen (references/paidf-anomalygen.md), rca (references/tao-analyze-gaps-visual-changenet.md), routing (references/tao-route-visual-changenet-samples.md), and data_mining (references/tao-mine-aoi-images.md).

Path rule (invariant). Use absolute host paths under ${RESULTS_DIR}/iter${ITER}/ for every stage's output, mount <workspace> into the container at the same path, pre-create dirs world-writable, and reject any config containing output: /results/... or any path outside <workspace>.

Data, Pre-Flight, Pipeline, and State references

Required input — max_iterations. No default; ask the user if not supplied and do not proceed past Pre-Flight without it. If the user gives a time limit instead, convert it to an estimated max_iterations using the per-iteration runtime figure in references/preflight.md and surface the estimate for confirmation. All other run parameters have defaults — never ask about a parameter with a default. The full defaults list and the Pre-Flight Summary the user approves at the single gate are in references/preflight.md.

Gating

Run the full Pre-Flight (references/preflight.md), print the Pre-Flight Summary, then STOP at the one user gate. After approval, run the baseline (with the pre-seed/skip-train logic) and the 7-step iteration Pipeline, all detailed in references/pipeline-and-state.md.

Hard-stop and never auto-retry on: any stage status=error; train/validation leakage (the mid-iteration check on mining_filter/mining_pool.csv right after mining, and the post-assembly check on the combined CSV); a missing or zero-row mining pool; a failed CSV existence check; silent-drop; and AMP allocation mismatch. The loop stops when the KPI target is met, max_iterations is reached, or an unrecoverable gate fires. Each terminal path runs the loop-end sequence: append the final loop_stop entry via scripts/log_stage.py, backfill token usage with scripts/align_token_usage.py, spawn the reporter agent one final time (trigger="loop-end"), then run scripts/prepare_inference_spec.py — skipped only when no valid checkpoint exists. Per-stage state cadence (one loop_log.jsonl entry per stage, seq=last+1 from disk, disk is canonical, HTML render once per iteration and at loop end) is specified in references/pipeline-and-state.md.