# Laptop assistant: from quick tests to useful workflows

**Status: local preparation in progress; no laptop winner selected yet.** This is a deployment roadmap, not a claim that the full integration already works.

## 💻 Coding and agent skills — what we will actually test

| Track | Cases | Success evidence | Current status |
|---|---:|---|---|
| HumanEval+ | 5 | Generated functions pass the selected task's base and extended tests in an isolated container | Existing outputs; execution/grading must be reconciled |
| MBPP+ | 5 | Correct results on base and extended tests, including edge cases | Existing outputs; execution/grading must be reconciled |
| LiveCodeBench | 9 | Program passes task tests within sandbox limits | Existing outputs; execution/grading must be reconciled |
| Q36 bug fixes | 5 | Reproduce a failing test, apply a patch, pass regression tests | New tasks to prepare |
| Q36 small app / website tasks | 5 | Build succeeds; required UI behavior passes browser checks | New tasks to prepare |
| Tool choice and arguments | 5 | Correct tool and schema-valid arguments; no invented tool result | Local smoke fixtures prepared separately |
| Multi-step planning | 5 | Respect dependencies and reach a verifiable final state | New interactive fixtures to prepare |
| Recovery from tool errors | 5 | Respond to a simulated failure, choose a valid alternative, verify recovery | New interactive fixtures to prepare |
| Memory and conflicting evidence | 5 | Retrieve the correct note, cite it, distinguish updated from stale facts | New isolated test vault to prepare |
| Stopping and task completion | 5 | Stop after success; report missing prerequisites instead of inventing completion | New interactive fixtures to prepare |
| Transfer to changed constraints | 5 | Solve held-out variants after a changed rule or environment | New tasks to prepare |
| Q36 browser workflows | 5 | Complete a task on a local test website and verify the resulting page state | New interactive fixtures to prepare |

**64 case slots in this extension**, including 19 existing coding slots. These are compact diagnostics, not full leaderboard scores. The agent tasks measure useful components of general problem solving, **not proof of AGI**. Existing Hermes-format and JSON-schema diagnostics remain in the broad suite and are not silently counted twice.

### How a capability earns a card entry

Freeze tasks and evaluators before comparing runtime settings. Count one primary attempt per case. Report **passes / executed-and-graded cases**, task coverage, failures and ungraded cases together. Publish results even when they are weak; only describe a workflow as demonstrated when its actual end-to-end task passes. A successful JSON tool-call test alone does not establish autonomous coding or desktop control.

Generated code runs in a disposable container without network access, host credentials or user-folder mounts, with CPU, RAM, process and time limits. Host tools and paid services are not available to benchmark prompts.

## ⚡ Pick a laptop runtime

1. Smoke-test **IQ4_XS, Q3_K_M, IQ2_M and IQ1_M**; exclude the BF16 server package. Start with IQ2_M to check loading and the harness.
2. Compare **Ollama and llama.cpp**, each with a documented baseline and an optimized profile: 16 primary configurations. FreeToken is an additional experimental backend only if a compatible build actually runs.
3. Use the same prompts, image inputs, context, output cap and thinking setting. Tune only on a separate warm-up set, then freeze settings.
4. Record answer quality separately from prefill/decode tokens/s, end-to-end speed, time to first visible output, CPU/RAM/VRAM, temperatures, timeouts and incomplete outputs.
5. Select a measured quality/performance/memory compromise, not simply the fastest token rate. An OOM or unsupported backend is a technical failure, not a wrong answer.

The target laptop has **32 GB RAM and an RTX 4060 Laptop GPU with 8 GB VRAM**. These 35B-weight models need CPU/RAM offload. MoE reduces active computation, not the need to store the expert weights.

## 🛠️ Hermes, routing and personal memory

After selecting the runtime, test it as a local endpoint behind the existing Hermes/router installation. Keep the previous route available for rollback. First use a dedicated test workspace and synthetic notes; connect the real Obsidian vault only after the test flow passes.

Useful scenarios to validate:

- **Code companion:** inspect a sample repository, fix one bug, run tests and summarize the patch.
- **Project memory:** find a note, cite its source and use its latest decision in a response.
- **Research notebook:** turn supplied material into a draft note with provenance, then ask before writing into the real vault.
- **Task recovery:** recognize an unavailable tool, ask for what is missing or use an explicitly allowed alternative.
- **Private local assistant:** use the configured local route and verify that external fallback is disabled for private inputs.

Obsidian storage, retrieval/indexing, memory updates and routing belong to the host application. They are not built into the model weights. Exact endpoint instructions and working configuration examples will be added after local validation; no production routing has been changed by this document.