LYGO API Token Saver

Minimize pay-to-go xAI/API token spend. Use when user says token saver, save tokens, API budget, pay-to-go burn, or before long Grok/Claude/GPT calls. Prefer local Ollama army; Biophase7 alt xAI key; compact agent behavior.

API key order (Biophase7)

1. Never paste keys in chat or commits.
2. Load vault: python tools/load_biophase7_vault.py → uses XAI_API_KEY_ALT before XAI_API_KEY_MAIN.
3. Frontier harness / probes: default --models stack only; add grok only when user explicitly needs frontier rows.
4. Set LYGO_OPENAI_FRONTIER_MODEL only when OpenAI runs are required.

pxpipe-LYGO (vision context compression)

When prompts/tool dumps are huge and byte-exact hashes are not the focus:

bash

cd lygo-protocol-stack
pip install -r requirements-pxpipe.txt
python tools/run_pxpipe_lygo_proxy.py

See docs/BIOPHASE7_PXPIPE_LYGO.md and skill lygo-pxpipe-lygo. Agent one-liner:

python tools/pxpipe_lygo_for_agent.py --shrink-file <huge.txt> --target grok

Do not compress secrets, seeds, or diff-critical line numbers.

Prefer local silicon (zero API tokens)

bash

# Ollama army cron + queue (127.0.0.1 only)
cd "%LYGO_STACK_ROOT%\.grok\skills\lygo-ollama-army\ollama_command_center\scripts"
python army_cron_once.py

• Drafting, summarizing repo files, grep/explore: spawn_subagent explore or local Ollama — not main chat API for bulk read.
• LFW path: lyra_failsafe() → LYGO_LFW_FALLBACK_MODEL on Ollama when cloud throttled.

Agent behavior (Grok Build)

Do	Don't
Short replies; tables over prose	Re-summarize full session history
grep + read_file with offset/limit	read_file entire 3k-line trees
One spawn_subagent for execute batches	Many sequential full-context turns
todo_write only for 3+ steps	Narrate every tool call
Cache paths in one line ("see path")	Repeat URLs and commit hashes twice
background: true for HF push / long tests	Block chat on 10min uploads
Stop when task done	"Resonance forward" essays

Frontier harness (metered)

bash

python tools/run_falsifiable_vector_test.py --load-vault --models stack
# API spend only if asked:
python tools/run_falsifiable_vector_test.py --load-vault --models grok --limit 3

Full 60× Grok ≈ high token + latency cost — require explicit user consent.

User phrases → mode

• "token saver" / "pay to go" → stack-only tools, Ollama army, alt xAI, minimal chat output.
• "push all" → one subagent; don't re-read diff in main thread.

Self-check

bash

python -c "import os; print('alt' if os.environ.get('XAI_API_KEY_ALT') else 'no-vault')"

Install companion: lygo-ollama-army, lygo-protocol-stack-operator.