Note
This project is an experiment. It explores what System 1 models can do as message routers. A System 1 model gives a fast, intuitive answer without step-by-step reasoning. TypeSafe Jev and Laya are System 1 models. Here, one of them only classifies each message and selects a preset. It does not read your code or plan the work, so it can select the wrong preset. Do not use this plugin in production. Examine each routing notice, and select a preset manually when the choice is important.
Auto Mode for Paseo is a Paseo plugin. It reads each new message and sends it to the best preset for the task. Each preset has its own provider, model, reasoning setting, and instructions. A preset can use Codex, Claude, OpenCode, or another provider installed in Paseo. TypeSafe Jev or a local Laya model classifies the message. The OpenAI Decisions API is a third, experimental classifier.
The conversation picker contains Auto and preset names. Model names stay in the plugin settings. Paseo owns workspaces and native provider authentication.
- Paseo 0.8.0 or later with plugins enabled.
- Node.js 24 and npm.
- An installed and authenticated Paseo provider for each enabled preset.
- For Jev: a TypeSafe API key.
- For the OpenAI Decisions API: an OpenAI API key with access to that API.
- For Laya: no separate Python installation is needed on Windows.
Clone this repository. Then run these commands in the project directory:
npm ci
npm run check
paseo plugin install .On Windows, use npm.cmd if PowerShell blocks npm.ps1.
If paseo is absent from PATH on Windows, use the project's launcher:
npm.cmd run plugin:install
npm.cmd run plugin:status
npm.cmd run plugin:reloadThe launcher checks PASEO_CLI, then PATH, then the Paseo Desktop installation
under LOCALAPPDATA. Use npm.cmd run paseo -- --help for other commands.
Set PASEO_CLI to the CLI file if Paseo is installed in a different location.
If the agent sandbox blocks the installation, use a supported read grant or
run the affected command through the host's approval flow. The same applies
when gh cannot access the user's existing GitHub login from the sandbox.
No credentials are copied into the project.
Open Settings > Plugins > Auto Mode for Paseo. Select a classifier. Jev remains the default for existing settings. Valid changes save automatically after a short pause in typing. Invalid fields show an inline error. Other valid changes can still save.
Select Jev (TypeSafe API). Enter the TypeSafe API key.
Click Save beside the key to apply it. Click Cancel to discard the key edit.
Other setting changes never submit an unconfirmed key.
You can also set TYPESAFE_API_KEY in the Paseo daemon environment.
The key and Jev model fields appear only when Jev is selected.
On Windows, run the local installer from the plugin directory:
npm.cmd run laya:install -- -Device cudaThe installer downloads Python 3.12 into .laya-python, creates .venv-laya and .laya-cache,
installs laya==0.3.5, installs the CUDA build of PyTorch when selected, preloads the chosen model, and saves the
Laya settings for the Paseo daemon. Both directories stay inside the plugin and
are ignored by Git. Use -Device auto to let PyTorch choose the device. Use
-Device cpu when no compatible NVIDIA GPU is available.
The first install downloads the Python runtime, PyTorch, Laya dependencies, and the selected Hugging Face model. It can take several minutes. The plugin starts the cached model on later turns.
On macOS or Linux, create a Python environment in the project directory:
python3.12 -m venv .venv-laya
.venv-laya/bin/python -m pip install laya==0.3.5Select Laya (local, experimental) in plugin settings when the installer is not used. Set Python executable and Model cache to the absolute paths for that environment. Do not include shell commands or arguments. Select Multilingual for Portuguese. Select English for English-only work. Use Typed decisions only when evaluating that specialized checkpoint. Choose CPU, CUDA, or Automatic as the device. An explicit device mismatch stops the turn. Save the settings. Send a short message to load the model.
The first load downloads model files from Hugging Face. The model stays loaded between classifications until the plugin unloads, the Laya configuration changes, or classification fails. Startup has a 120-second limit. Inference has a 30-second limit. Preload the model before the first Paseo turn when manual installation takes longer than the startup limit.
Laya does not require a TypeSafe key. A saved key stays stored when switching classifiers. The plugin does not pass it to Laya. A Laya failure stops the turn. It never sends the request to Jev as a fallback.
Laya quality for this routing task has not been benchmarked. Its confidence score is not a guarantee of correct intent. Evaluate representative requests in your language before using it for unattended work. Both classifiers can misjudge scope or task depth. Overlapping scopes can be ambiguous. Jev is the recommended classifier for Auto preset selection. In local smoke tests with the default presets and two custom presets, Jev selected the expected preset in 16 of 18 requests and Laya in 14 of 18. Jev tagged 12 of 12 sample scopes correctly and Laya 10 of 12. Routing tests validate selection rules. They are not an accuracy benchmark. See the Laya limitations.
The Decisions API is an OpenAI endpoint that answers fixed questions with typed answers. It is a public beta. OpenAI can change its request format.
Select OpenAI Decisions API (beta, experimental). Enter the OpenAI API key.
Click Save beside the key to apply it.
You can also set OPENAI_API_KEY in the Paseo daemon environment.
A key in the settings file has priority over OPENAI_API_KEY.
Keep Model as gpt-6-luna unless OpenAI publishes another Decisions model.
The plugin sends the same questions and the same bounded data as it sends to Jev. Preset scope questions use the API's predicate type. Other questions use its choice type. The plugin does not send images to the Decisions API. Each request has a 20-second limit.
The OpenAI key and the TypeSafe key are separate values. The plugin never uses one for the other service. A Decisions API failure stops the turn. It never sends the request to Jev or Laya as a fallback.
Decisions API quality for this routing task has not been benchmarked. In one local smoke test with the default presets, the Decisions API and Jev each selected the expected preset in 15 of 15 requests. Laya selected it in 14 of 15. The Decisions API returned the expected intent in 15 of 18 requests, Jev in 18 of 18, and Laya in 10 of 18. Each Decisions API intent error returned discussion where implementation or review was expected. Such an error adds no-edit instructions to a turn that requested a change. The Decisions API and Jev tagged 12 of 12 sample scopes correctly. Laya tagged 11 of 12. The median classification time was about 300 ms for the Decisions API and for Jev. This test used a different request set from the Laya test above. It ran once. Evaluate representative requests before using the Decisions API for unattended work.
Open Settings > Plugins > Auto Mode for Paseo. Click Refresh to load the providers and models available on this daemon. Choose a provider for each preset. Select a model from Available models. Choose a supported reasoning setting. Select Provider default to use the model's default. Edit the instructions and description. Valid changes save automatically. A model release does not require a plugin update.
| Preset | Automatic role |
|---|---|
| Tech Lead | Code delivery, debugging, and implementation validation |
| Staff | Technical strategy, architecture, and consequential tradeoffs |
| Critic | Reviews, critiques, and validation of existing work |
| Reporter | Facts, progress, changes, and open questions |
| Writer | Prose, documentation, explanations, and user-facing text |
Each preset has its own settings block. You can rename, disable, or remove any preset, including the defaults. Complete a new preset's required fields to save it. An incomplete new preset stays local while you edit it. The initial presets are Tech Lead, Staff, Critic, Reporter, and Writer. They are editable starting points. With the initial scopes, "Are these changes good?" fits Critic and "What changed?" fits Reporter. No preset name or ID has special routing behavior.
Edit Scope to define when Auto should use a preset. Include its responsibility, examples, and limits. The full scope stays visible in a multiline field. Auto sends the complete scope of each enabled preset to Jev or Laya. The limits are 240 characters for each scope and 32 presets. Names and IDs do not influence selection. Instructions, model, provider, and reasoning settings do not influence role selection. Edit Instructions to tell the selected model how to work.
Used for assigns each preset to fixed task types: Review, Implement, Design, Report, Write, or Other. When you stop editing a Scope, the settings screen asks the configured classifier to detect the type. Select types to set them manually. Use Detect from scope to return to detection. A preset with an edited scope and no current detection is untagged and competes for every task type.
Set Task depth to the most demanding work the configured setup can handle: Light, Standard, Deep, or Expert. Jev/Laya estimates the required depth from the request, recent context, and aggregate workspace change counts. A short question about a large change can require a deep review. Size alone does not imply difficulty. This field is separate from the provider's Reasoning effort setting. Review task depth when changing a preset's model. The plugin does not infer model capability.
For each message, Jev/Laya also chooses one task type. Auto compares scopes only among the presets used for that type and untagged presets. A higher scope score cannot send a review to a writing preset. When no preset is used for the type, all presets compete and the turn summary reports that. Other never reports it. Auto selects the highest scope score among those presets that support the required depth. When none supports it, Auto uses the deepest available setup and reports that limit. Low scores do not block short or generic messages such as "testing" or "hello". Equal scores prefer the lowest sufficient depth, then stable preset ID order. If no enabled preset has a scope, Auto asks you to configure one or choose manually. Leave a scope empty to make that preset available only for manual selection. Disabled or deleted presets are excluded. Manual selection takes priority. Default and custom presets use the same selection rules, including after renaming. Previous routing thresholds no longer apply.
Use Restore to add deleted defaults back. This action keeps your existing presets and their settings. Saving settings does not restore deleted presets.
The selected preset supplies its provider, model, effort, work mode, and instructions. The classifier still controls intent and automatic Plan and Fast decisions. It cannot change the preset's instructions or grant Full access. Preset scopes are classification data. They do not authorize edits or enable Plan. Each started turn shows one compact notice with the selected preset, model, intent, required task depth, mode, and reasoning setting. Any capability fallback appears in that notice. The new task-depth field starts from the built-in value for existing default presets. Other existing presets start at Standard. All values remain editable.
Workspace context contains only counts of changed files, added and removed lines, binary files, and untracked files. It covers uncommitted changes against HEAD. It does not inspect file contents or measure committed branch changes. Filenames, paths, and raw Git output are excluded. Unavailable or empty counts do not imply an easy task. Laya receives these counts only in a separate task-depth pass. Its scope scores come from a pass without them, because the counts reduced scope accuracy.
The model catalog comes from Paseo. Native execution uses Paseo's installed providers and existing credentials. Settings offer only the models, reasoning levels, and work modes published by that provider. Changing the provider clears dependent choices. Changing the model clears the reasoning level. Refresh loads new releases. Reasoning effort is hidden when the selected model has no reasoning options.
| Provider | Automatic approvals and Plan | Explicit Full access |
|---|---|---|
| Codex | Intent-based sandbox and Codex Auto-review | Supported; Plan remains read-only |
| Claude | Native Auto mode by default; native Plan when Plan is enabled | Native bypass mode; Plan takes priority |
| OpenCode | Published work and planning modes | Native permission options; Plan takes priority |
| Other installed providers | Published modes, or the provider's defaults when modes are unavailable | Published bypass mode when supported; otherwise normal approvals |
The preset's Work mode can override the automatic work-mode choice. It lists ordinary work modes. Plan and bypass are controlled by the conversation. The field is hidden when the provider publishes no work modes. Discussion and review do not enable Plan by themselves. With Plan off, native providers use the preset's work mode. Under automatic approvals, discussion and review receive no-edit instructions. These modes retain each provider's own permission semantics. They do not imply a shared operating-system sandbox. When Plan is requested but unsupported, the plugin uses normal approvals and instructions to analyze without edits. The turn notice explains this limitation. Native provider options are documented in Paseo provider options. Unknown models stop the turn and direct you to the preset settings. An unavailable reasoning setting uses the model default. Fast uses the model's published Fast feature when supported. Otherwise, execution continues at normal speed. The turn notice explains these fallbacks.
Most providers run each turn in a new native Paseo agent. The agent is archived after completion or cancellation. Its full timeline remains in Paseo. No second session store is added. Each new agent receives up to 24 conversational messages, at most 8,000 characters each. This handoff excludes tool output, reasoning, and previous image data. A resumed chat replays this bounded conversational history. Only the smaller six-message context goes to the classifier.
Select Auto Mode for Paseo as the provider. Select Auto or a preset in the conversation. A manual preset applies to the next successful turn by default. Select Keep selected preset to use it for later turns.
The mode control has three options:
- Auto (plan or work) lets the classifier choose whether this turn needs Plan.
- Work on request answers, reviews, or implements with automatic Plan disabled.
- Plan only requests analysis without edits and uses the provider's planning mode when available.
The mode icons identify automatic choice, work, and planning. Modes do not select a preset or grant Full access. Use the separate controls for those choices.
The Fast control has three options:
- Auto starts with Fast off. The classifier can enable Fast for an urgent task.
- On requests faster processing when the model supports it.
- Off uses normal speed.
Fast availability and quota use depend on the provider and model.
The default permission setting is Automatic approvals. Each provider applies it as the support table shows. The classifier cannot enable Full access. Only an explicit user selection can enable Full access. Plan takes priority when Full access is selected. The plugin does not restore Full access when a session reopens.
A setting change during a turn applies to the next turn. Steering support depends on the provider. If a provider does not support steering, wait for the run to finish or interrupt it before you send another message.
The selected classifier receives the following data. Jev sends it to TypeSafe. The OpenAI Decisions API classifier sends it to OpenAI. Laya processes it in a local Python process:
- The new text message.
- Up to six recent user messages, assistant answers, or plans.
- Up to 1,000 characters from each recent item.
The plugin does not send tool output or private reasoning to either classifier. The plugin stores the short routing context in Paseo session data.
The plugin stores a TypeSafe API key in
~/.paseo/auto-mode-for-paseo.local.json as plain text. A key in this file has
priority over TYPESAFE_API_KEY. The settings API does not return the key to
the client. An empty key field keeps the saved key.
The same rules apply to the OpenAI API key and OPENAI_API_KEY.
Laya requests have a 16,000-character message limit and a 64 KiB JSON limit. At most eight classifications can wait or run at once. They run in order. The worker rejects any supplied state or question that would be truncated by the model. Default model limits are 512 tokens for English and 1,024 for the other checkpoints, including questions. Long messages or recent context can therefore stop a turn. Shorten the message or start a chat with less context. The plugin does not send images, tool results, or credentials to Laya.
Provider authentication stays in Paseo and the installed provider CLI. The plugin does not request vendor API keys.
With automatic approvals, discussion and review request analysis without edits. Plan is a separate decision. Native execution uses the preset's work mode when Plan is off. Each provider enforces its native permissions as described above. No-edit instructions do not add a sandbox. Unknown intent values and invalid scores stop the turn.
The provider supports text, images, streamed responses, Plan questions, approvals, interrupts, steering, and history replay. It does not support composer commands.
Paseo renders the image preview in its native composer. Auto Mode for Paseo sends the image to the selected provider with the message. It sends only the message text to the selected classifier for routing.
Attach up to four PNG, JPEG, WebP, or GIF images in one message. Each image must be 5 MiB or smaller.
Run all checks:
npm run checkThe tests use simulated TypeSafe, OpenAI, Laya, and Paseo provider services. They do not need
credentials, model downloads, or a running Paseo daemon.
Python bridge tests use a fake Laya module and need Python 3.10 or later.
Set LAYA_TEST_PYTHON if the executable is not named python.
These tests skip locally when Python is absent. CI requires them.
Read CONTRIBUTING.md before you submit a change. Read ARCHITECTURE.md for the module boundaries and routing rules. Report security problems as described in SECURITY.md.
Reload the plugin after you update its source:
paseo plugin reload auto-mode-for-paseoThis project uses the MIT License.
