# Kaiju Coder 7 Local Test Instructions

Use these commands from the repo root. The public release name is Kaiju Coder 7. Internally, this build is backed by the v1.8 adapter under `runs/qwen36-27b-lora-v1.8-business-owner/adapter`. The release-candidate raw model path is the merged full model on Gojira B at `/home/richardecholsai5/kaiju-coder/models/Kaiju-Coder-Qwen3.6-27B-v1.8-merged`. The deterministic harness commands work locally now; the fastest current runtime is vLLM bitsandbytes on Gojira B over Tailscale with the local OpenCode fast proxy.

## Run The Local Release-Candidate Gate

```bash
python3 scripts/run_kaiju_business_owner_rc_smoke.py
```

This validates reviewed data, checks v1.7 targets, builds the oversampled business-owner SFT file, smokes the local OpenAI-compatible harness API, runs the hard router suite, and runs static artifact checks.

For release status, read `release/COMPLETION_AUDIT.md` and `release/HUGGINGFACE_RELEASE_DRAFT.md`.

## Merge The v1.8 Adapter

Use this if the merged full model must be rebuilt:

```bash
KAIJU_LORA_ADAPTER=/workspace/kaiju-coder/runs/qwen36-27b-lora-v1.8-business-owner/adapter \
KAIJU_MERGED_MODEL_DIR=/workspace/kaiju-coder/models/Kaiju-Coder-Qwen3.6-27B-v1.8-merged \
  ./scripts/run-gojira-b-qwen36-lora-merge.sh
```

## Start Kaiju Coder 7 Serving

Use this for the fastest current model-side candidate:

```bash
KAIJU_VLLM_CONTEXT=16384 \
KAIJU_VLLM_QUANTIZATION=bitsandbytes \
KAIJU_VLLM_LOAD_FORMAT=bitsandbytes \
KAIJU_VLLM_GPU_UTIL=0.90 \
  ./scripts/start-qwen36-merged-vllm.sh
```

Confirm readiness:

```bash
curl http://100.109.109.14:18084/v1/models
```

Then keep the Mac-side fast proxy pointed at that vLLM endpoint:

```bash
KAIJU_OPENAI_BASE_URL=http://100.109.109.14:18084/v1 \
python3 scripts/kaiju_opencode_fast_proxy.py --host 127.0.0.1 --port 18181
```

The high-context `32768` target has benchmark evidence in
`release/SERVING_BENCHMARKS.md`, but the current speed/default path is 16k
runtime-quantized vLLM plus the local fast proxy.

## Prepare Merged-Model Hugging Face Metadata

Use this before any full merged-model upload review. It syncs release metadata
into the Gojira-B model folder but does not upload or read Hugging Face tokens.
If the remote merged folder is root-owned, the helper automatically uses
passwordless sudo for rsync without changing model ownership:

```bash
bash scripts/prepare_hf_merged_model_metadata.sh
KAIJU_MERGED_METADATA_APPLY=1 bash scripts/prepare_hf_merged_model_metadata.sh
bash scripts/upload_hf_merged_model_from_gojira_b.sh
```

## Install And Smoke OpenCode

```bash
python3 scripts/install_kaiju_opencode_profile.py
opencode run -m kaiju/kaiju-coder-7 --agent kaiju-coder-7 \
  --dir /tmp/kaiju-opencode-loopguard-smoke \
  --dangerously-skip-permissions \
  'Create loopguard.txt with exactly: Kaiju Coder 7 loop guard installed'
```

The installer writes the `kaiju` provider, the lean `kaiju-coder-7` agent, and
the scoped no-autocontinue plugin at
`~/.config/opencode/kaiju-no-autocontinue.mjs`.

## Run The Deterministic Harness Smoke

```bash
python3 scripts/run_kaiju_api_harness_smoke.py
```

## Run A Direct Model Eval

```bash
python3 evals/run_openai_compat_smoke.py \
  --base-url http://100.109.109.14:18084/v1 \
  --model kaiju-coder-7 \
  --tasks evals/tasks/smoke.jsonl \
  --max-tasks 1 \
  --timeout 300 \
  --max-tokens 768 \
  --temperature 0 \
  --disable-thinking \
  --system-prompt-file prompts/kaiju-coder-api-system.md
```

For the selected final business-owner checkpoint, run the focused v1.8
business-owner pack and then score it. Raw merged model generation is slow, so
use the harness for practical paid website delivery until broader raw website
evals pass at acceptable latency:

```bash
python3 evals/run_openai_compat_smoke.py \
  --base-url http://100.109.109.14:18084/v1 \
  --model kaiju-coder-7 \
  --tasks evals/tasks/business-owner-v18-comparison.jsonl \
  --timeout 900 \
  --max-tokens 2500 \
  --temperature 0 \
  --disable-thinking \
  --stream \
  --system-prompt-file prompts/kaiju-coder-api-system.md

python3 evals/score_quality_gate.py runs/evals/<merged-v18-run>/results.jsonl
```

Current merged evidence:

- Probe: `1,155` visible chars in `60.17s`.
- Proposal rerun: `1/1` paid-ready, `4.0/4.0`, `4,014` chars in `212.72s`.
- Jah credits backend: `4.0/4.0`, `9,718` chars in `566.36s`.

## Dynamic LoRA Serving Caveat

Do not use dynamic SGLang LoRA serving as release evidence for v1.8. The adapter-name-only path can be base-equivalent, and the corrected selector `qwen36-27b:kaiju_v18_business_owner` crashes this SGLang build with a fused-module LoRA buffer shape mismatch. Use the merged full-model path above.

## Run The Business-Owner Harness

```bash
python3 evals/run_router_harness_eval.py --tasks evals/tasks/router-hard-harness.jsonl
python3 evals/run_router_static_checks.py runs/evals/<router-run>/results.jsonl
```

## Manual Prompt To Try First

```text
Build me the full Kiyomi 7.7.7 AI company operating pack for a local business owner. I need the launch kit, website, content engine, connector checklist, intake CRM, money report, automations, operator handbook, lead generator, sales closer, ROI dashboard, and Workshop golden run. Make it owner-ready with no developer setup required.
```

Expected shape:

- A project folder with multiple files, not advice only.
- Complete HTML where HTML is requested.
- Lead/sales CSVs.
- Connector verification gates.
- ROI audit gate.
- Workshop golden-run gate.
- Clear owner commands such as `/kiyomi` and `/kiyomi-do`.