Local agents and model-free context selection

Local setup is implemented in source and awaits installed-app acceptance with real models. No inference runtime, weights or tokenizer are bundled with Jackalope. The optional setup uses OpenCode and Ollama; it does not require an AI provider subscription. Jackalope's account-access requirement remains separate.

Guided setup

Open Agents → Try a local agent, or use the same entry during onboarding.

1. Inspect total/available system memory, graphics names and free disk space on the expected model drive. Unknown values stay unknown. GPU detection does not establish acceleration or usable VRAM.
2. Review an optional model download and its memory guidance. Sizes are approximate upstream tag sizes; tags and packages can change.
3. Install Ollama and prepare Jackalope's private OpenCode runner, then explicitly start the model download. Windows offers WinGet for Ollama; other platforms use its official setup link. The private runner downloads from Jackalope's pinned manifest and does not require a separate OpenCode installation. Open Ollama and recheck if its service is not running. Download and installer progress, elapsed time, cancellation and failures appear in the selected row. Canceling an installer cannot roll back completed installation steps.
4. Run the local check. Jackalope creates an Ollama alias with a 65,536-token context, asks OpenCode to edit a file in an owned disposable folder, and continues the same session to create a second file. A matching session ID, tool event, exact file contents and unchanged base/alias digests are required. Only then can the user connect the account. Cancellation invalidates the check. An optional title model can be selected from other installed catalog models. Its separate 8,192-token alias must pass a bounded structured-response check; both model selections and digests are rechecked before connecting. This does not download another model or change the coding model's context.

Optional model	Approximate download	Suggested system memory
Qwen 3.5 4B	3.4 GB	16 GB
Qwen 3.5 9B	6.6 GB	24 GB
Qwen3 Coder 30B	19 GB	48 GB

Memory figures are conservative starting points, not measured acceptance thresholds or speed guarantees. Context, GPU memory and other apps affect whether a model fits. Ollama's Windows documentation also requires at least 4 GB for its runtime, separate from weights. The private OpenCode runner uses the version and integrity pins described in agent support.

Model requests use 127.0.0.1:11434. Native download requests bypass proxies and redirects. Download progress comes from Ollama's layer byte counts; totals remain indeterminate until reported. Closing setup cancels its current operation. Ollama retains reusable layers, and tools/models remain installed when setup is skipped. The prepared alias shares weight layers with its base model. To reclaim those layers, remove both models in Ollama. Owned check folders are removed after the check; a locked leftover may remain under the native profiles local-checks directory.

Connecting creates or reuses a named OpenCode account and selects it for future OpenCode work. It preserves other accounts and their model preferences. Onboarding pins the chosen local account to the pending project. Outside onboarding, choose the account in project settings. Finishing local setup refreshes agent detection and selects OpenCode and the new local account in onboarding, including after a reload. If the app's default agent is unavailable, setup selects OpenCode for Ask Jackalope. Existing sessions retain their saved account; model and account restrictions still apply.

By default, OpenCode's main model and small_model point to the same checked local model. Selecting a title model changes only small_model, with a 512-token output limit; it does not route coding work to that model. The setup probe requests a two-minute Ollama keep-alive; subsequent OpenCode requests use Ollama's configured retention behavior. Jackalope does not change global GPU, quantization or retention settings. Keeping two models loaded can increase memory pressure and may be slower. The profile enables only its local provider, disables sharing and project configuration overrides, and removes inherited external configuration paths. Per-task permissions and Jackalope's direct MCP configuration are retained. This is a local inference configuration, not an OS sandbox: agent tools, package installation, enterprise-managed OpenCode configuration and external services have their own boundaries. The file check establishes basic tool/session support, not coding quality, offline installation or suitability for every task.

Sources: Ollama and OpenCode, Windows requirements, Qwen 3.5 tags, Qwen3 Coder, OpenCode configuration.

Context retrieval and routing

• Ask Jackalope receives up to four relevant, complete documentation passages with sources, bounded to about 6 KB. It can answer from sufficient excerpts without a search/read round trip, or request full documents when necessary. Search tools use BM25 too; partial retrieval is never treated as proof of an answer.
• Routing shares repeated account metadata while retaining every eligible model and every original candidate field. It preserves the full task, saved context, and acceptance criteria. Quality and task fit precede headroom among capable models; being small, free or local does not establish fit.
• Local profiles expose only their checked model, and other accounts keep their model choices. Explicit assignments, restrictions, quota reserves and handoff eligibility remain enforced. One eligible candidate, or accounts sharing the same agent and explicit model, need no model-selection call. Unknown CLI defaults and different agent/model choices retain the configured decision method.

Managed local OpenCode helpers can reuse at most two authenticated loopback servers. Requests use separate sessions and the same account, executable path, environment and configuration. Completed responses and usage are read back from the known session, so attached CLI stream timing cannot silently lose the result. Helpers use a separate configuration directory and disable external plugins. A server is discarded on failure/cancellation, runtime shutdown or after two idle minutes. Unsupported startup falls back before submitting a prompt and backs off for ten minutes; submitted requests are never automatically replayed. JACKALOPE_WARM_OPENCODE=off disables reuse for comparisons. Cloud accounts and workers with task-scoped MCP credentials retain independent process lifetimes.

Candidate packing preserves every original field exactly. Prompt size and retrieval coverage do not establish billed-token savings or coding quality.

Verification and remaining acceptance

The optional OpenCode protocol probe runs a real OpenCode executable against an explicitly fake loopback provider:

node scripts/verification/local-opencode.mjs path/to/opencode.exe

The probe checks CLI configuration, tool events and session compatibility against a fake provider; it does not run Ollama or real weights. The ignored native warm_opencode_protocol test uses JACKALOPE_OPENCODE_EXE and a fake loopback provider to check authenticated reuse, separate sessions, cancellation and idle cleanup with the installed CLI. It does not establish real-model speed. Native tests cover account isolation, model restrictions, cancellation ownership, download parsing and lossless prompt packing. Browser fixtures explicitly disable installation, downloads and connection.

Before release, test the packaged app with current Ollama/OpenCode and each offered model on representative Windows hardware; record peak RAM/VRAM, latency, tool reliability, continuation and quality on reviewed tasks. Exercise cancellation, restart, low disk, unavailable services and installer failures. macOS/Linux setup and effective enterprise configuration need separate acceptance. Keep routing quality and real token measurements distinct from retrieval metrics.