qq928820655/changmodel

Portable DeepSeek Harness plugin for switching models, managing model capabilities, and optionally enabling input-history recall.

Commands

/cmodel [provider|]model[context][/effort]

Examples:

/cmodel gpt-5.6-sol[1m]
/cmodel {gateway}gpt-5.6-sol[1m]
/cmodel openai|{gateway}gpt-5.6-sol[1m]/high

• provider| selects an existing Harness provider route. When omitted, the current session provider is retained.
• {gateway} remains part of the upstream model id. It is not a Harness provider route.
• [context] sets the model context window.
• /effort is applied only when the configured model supports that reasoning level; unsupported values fall back to the default effort.
• Models.dev lookups remove {...} and [...] decorations.

Settings

The changModel settings page provides:

• Provider-scoped model search, add, edit, and delete.
• Context window, maximum output tokens, text/image capability, and custom reasoning levels.
• Models.dev capability lookup with editable bulk application to matching saved models.
• A fixed summary model selection for context compaction when the target Harness supports compaction.
• Optional input-history recall patch controls.
• Subagent role policies with exact, alias, keyword, and priority matching.
• Unmatched-role fallback to a configured subagent default, the parent session model, or deny dispatch.
• Role-level provider, model, reasoning effort, context window, maximum output, and image capability settings.
• Provider-default and model-level API protocol selection for OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages.
• Subagent configuration is persisted under llm-pi-ai.changmodelSubagents; the optional dispatch adapter applies matching rules when the target Harness exposes a pre-dispatch hook.

All model capability changes write to the target Harness's own llm-pi-ai settings. No provider credentials, API keys, sessions, or model configuration are bundled in this plugin.

Optional compatibility features

Input-history recall patches the installed dsh-client-ui-conversation only after semantic anchor validation. The patch manager:

• discovers the target component at runtime;
• creates a versioned local backup;
• refuses to patch when expected Harness structures are missing or ambiguous;
• supports status checks and restoration.

Compaction remains optional. A Harness that does not expose an active compaction service continues to support model switching and settings management.

Model-level protocol overrides are enabled through a managed compatibility patch. The page checks the installed llm-pi-ai structure before applying it, stores a hash-protected backup, refuses incompatible upgrades, and can restore the original file. Enabling or restoring this patch requires a Harness restart. Only protocol adapters provided by the installed Harness are accepted; arbitrary request formats are not inferred.

Installation

Install this package using the DeepSeek Harness plugin mechanism. The package declares a web client bundle and a profile bundle patch.

After installation, restart the Harness. Use the changModel settings page to configure optional features for that specific machine.