Local llm bridge

Identity: The Local Gemma Sub-Agent Dispatcher

Dispatches bounded tasks directly to the optimized local Gemma 4 12B server at http://localhost:8089/v1/chat/completions. No routing proxy involved. Uses the run_agent.py task router with cli=llama.

[!IMPORTANT] Requires llama-server running on port 8089. Check: curl http://localhost:8089/health Start: ./run_server.sh in the local-llm-bench workspace. Thinking is disabled server-side (--reasoning off) — no special flags needed.

---

Why This Is Fast

The routing proxy (Mode A) carries ~29K tokens of Claude Code system prompt — at ~30 tok/s prefill that costs 60+ seconds per context boundary crossing.

This skill (Mode B) sends only the task prompt — typically 50–500 tokens. At 7+ tok/s generation on M1 Metal with a small context:

Output length	Typical response time
50 tokens	~7s
100 tokens	~14s
200 tokens	~28s

Default max_tokens=120 keeps responses terse. Override via code if needed.

---

Orchestration Pattern: run_agent.py

python ./scripts/run_agent.py \
  <PERSONA_FILE> <INPUT_FILE> <OUTPUT_FILE> "<INSTRUCTION>" \
  --cli llama --max-tokens 120

Example — code review

python ./scripts/run_agent.py \
  agents/refactor-expert.md \
  target.py \
  review.md \
  "List the top 3 issues in this code. Be terse." \
  --cli llama

Example — summarize a diff (longer output)

python ./scripts/run_agent.py \
  /dev/null \
  changes.diff \
  summary.md \
  "Summarize this diff in 2 sentences. Focus on risk." \
  --cli llama --max-tokens 200

Example — instruction only (no input file)

python ./scripts/run_agent.py \
  /dev/null /dev/null \
  answer.md \
  "What is the capital of France? One word." \
  --cli llama --max-tokens 10

---

Prompt Budget Guidelines

Keep prompts lean — this is the primary performance lever:

• Persona: 100–300 tokens (enough to set role and tone)
• Source file: keep under 2,000 tokens where possible; trim to the relevant section
• Instruction: 1–3 sentences; specific and bounded
• Expected output: terse — list form, not prose paragraphs

Avoid: pasting full file trees, long conversation histories, or open-ended "analyze everything" instructions.

---

Hardware Details (M1 Mac, 16GB)

Parameter	Value
Server	llama-server :8089
Model	Gemma 4 12B UD-Q4_K_XL
GPU offload	-ngl 99 (full Metal)
Flash Attention	-fa on
Batch sizes	-b 2048 -ub 512
KV cache quant	-ctk q8_0 -ctv q8_0
Thinking	disabled (--reasoning off)
Context	32768 tokens (1 slot)

---

Persona Registry (agents/)

Persona	Use For
security-auditor.md	Vulnerability review, risk assessment
refactor-expert.md	Code cleanup, readability, DRY
architect-review.md	Design review, modularity check

---

Co-located Scripts (scripts/)

All scripts are symlinked from the canonical plugins/cli-agents/scripts/ so the skill is self-contained when installed in isolation.

Script	Purpose
run_agent.py	Task router — cli=llama dispatches here
kv_cache_orchestrator.py	KV slot save/restore for repeated persona calls
run_server.py	Start llama-server with authoritative parameters
test_run_agent.py	37 tests: command builders, isolated-flag security contract, llama payload

routing_proxy.py is NOT included — it is the Mode A API compatibility shim and is not part of this skill's execution path.

---

Smoke Test

curl http://localhost:8089/health
python ./scripts/run_agent.py /dev/null /dev/null /tmp/test.md "Say hello in one word." --cli llama
cat /tmp/test.md

---

Health Check

curl http://localhost:8089/health          # must return {"status":"ok"}
# If down: python ./scripts/run_server.py