tinyml-workflow-agent

Tiny ML Tensorlab Skill

Overview

Build, train, compile, and deploy ML models to embedded MCUs using Tiny ML tensorlab.

---

CRITICAL: Script Location

runner.py lives in this skill's scripts/ directory, not in tinyml-modelzoo.

SCRIPTS_DIR = ~/.claude/plugins/marketplaces/<plugin>/skills/tinyml-workflow-agent/scripts/

Session Setup step 1 shows how to discover this path automatically.

---

CRITICAL RULES — BLOCKING GATES

These two rules apply unconditionally. Read them before any workflow step. They exist because compiler environment state does not persist across shell sessions, and because proceeding to deployment with missing or failed compilation artifacts causes silent, hard-to-debug failures downstream.

Rule 1: Compiler path MUST be exported in the same shell session as the training script

Before running run_tinyml_modelzoo.sh, locate and export the compiler path in the exact same bash call that runs the script. Exporting env vars in a prior step is not sufficient — shell state does not carry across separate tool calls.

Use ccs-project → getCompilers (preferred — CCS already knows what's installed):

Call ccs-project → getCompilers

From the returned list, find the entry matching $TARGET_DEVICE family (e.g., TI C2000 CGT for F28xxx, TI ARM CGT or tiarmclang for MSPM0/AM26x). Extract its installation path. Strip to the root directory containing bin/. Store as CGT_ROOT.

If ccs-project MCP is unavailable, fall back to filesystem search:

# C2000 CGT
find /opt/ti ~/ti $HOME/ti /usr/local/opt/ti 2>/dev/null -name "cl2000" -type f 2>/dev/null | head -1
# ARM CGT
find /opt/ti ~/ti $HOME/ti /usr/local/opt/ti 2>/dev/null -name "armcl" -type f 2>/dev/null | head -1

Strip the binary name → get bin/ → strip bin/ → get CGT_ROOT.

Then export and run in one command:

export C2000_CGT_PATH=<CGT_ROOT>   # or ARM_CGT_PATH, per device — see references/example_running_guide.md
cd "$TINYML_MODELZOO_PATH" && ./run_tinyml_modelzoo.sh "$CCS_PROJECT_PATH/config.yaml"