feat(models): add live session model switching and /init skill

[lavaman131]

Summary

This PR introduces two major improvements:

1. Live session model switching: Enable runtime model changes for Claude and Copilot agents without requiring a new session
2. Dynamic /init skill: Replace static template files with an intelligent codebase exploration skill that generates tailored project documentation

Key Changes

Model Switching

• Add setActiveSessionModel to CodingAgentClient interface (src/sdk/types.ts:623)
• Claude implementation: Persist model changes on session config for subsequent turns (src/sdk/claude-client.ts:1121)
• Copilot implementation: Rebind active session with updated model and reasoning effort (src/sdk/copilot-client.ts:881)
• Update UnifiedModelOperations to prefer SDK-level model switching over pendingModel fallback
• Normalize Claude model preferences to canonical aliases (opus, sonnet, haiku) by stripping provider prefixes
• Fix model preference persistence to use effective model from modelOps instead of raw config

/init Skill Refactor

• Remove static template files: CLAUDE.md, AGENTS.md, MODULE_DOCUMENTATION.md, and .atomic/settings.json
• Add /init builtin skill that dynamically explores the codebase using sub-agents (src/ui/commands/skill-commands.ts:1106)
• Generated documentation is tailored to actual project structure, tech stack, and conventions
• Clean up additional_files and preserve_files references from agent configs

Other Improvements

• Reformat skill-commands.ts to consistent 4-space indentation
• Add comprehensive test coverage for setActiveSessionModel behavior
• Update model operations tests to verify normalized preference handling

Breaking Changes

None. The changes are backwards compatible.

Test Plan

• Verify model switching works mid-session for both Claude and Copilot agents
• Confirm model preferences persist correctly with canonical aliases
• Run /init skill on a fresh project and verify generated CLAUDE.md/AGENTS.md are populated
• Run existing test suite (bun test) to check for regressions

[claude]

PR Review: fix(models): live session model switching and init skill refactor

Summary

This PR makes two significant changes:

1. Live session model switching - Adds setActiveSessionModel to enable runtime model changes without requiring a new session for Claude and Copilot agents
2. Replace static init templates with /init skill - Removes hardcoded CLAUDE.md, AGENTS.md, and MODULE_DOCUMENTATION.md template files, replacing them with a dynamic /init builtin skill that explores the codebase

---

Code Quality & Best Practices

Strengths

• Good test coverage for new functionality: Tests added for setActiveSessionModel in claude-client.test.ts, model operations test expansion, settings normalization tests, and builtin commands tests
• Clean interface extension: The CodingAgentClient interface properly extends with an optional setActiveSessionModel method
• Proper fallback handling: The UnifiedModelOperations.setModel() prefers SDK model setter when available, with sensible fallback for SDKs that don't support live switching

Issues

1. Inconsistent indentation change in skill-commands.ts (lines 2586-3727): The entire file was reformatted from 2-space to 4-space indentation. This creates unnecessary diff noise and deviates from what appears to be the project's 2-space convention. Consider:

   • Reverting the indentation changes
   • If 4-space is intentional, this should be a separate formatting PR

2. Missing type annotation in src/ui/index.ts:289-295:

   const sdkSetModel = agentType === "opencode" && "setActivePromptModel" in client
     ? async (selectedModel: string) => { ... }
     : agentType && "setActiveSessionModel" in client
       ? async (selectedModel: string, options?: { reasoningEffort?: string }) => { ... }
       : undefined;

   The nested ternary with different function signatures could benefit from explicit typing or a helper function for clarity.

---

Potential Bugs & Issues

1. Race condition potential in CopilotClient.setActiveSessionModel (copilot-client.ts:403-452):

   • The method gets the last active session, unsubscribes from events, calls resumeSession, then re-subscribes
   • If events fire between unsubscribe and re-subscribe, they could be lost
   • Consider keeping the old subscription active until the new one is confirmed

2. No validation for empty model arrays in ClaudeAgentClient.setActiveSessionModel:

   const activeSessions = Array.from(this.sessions.values()).filter((state) => !state.isClosed);
   const activeSession = activeSessions[activeSessions.length - 1];
   if (!activeSession) { return; }

   This silently does nothing if no active session exists. Consider logging a warning or returning an error state.

3. Model preference persistence timing in chat.tsx:3046-3057:

   if (modelOps && 'setPendingReasoningEffort' in modelOps) {
     (modelOps as {...}).setPendingReasoningEffort(reasoningEffort);
   }
   const result = await modelOps?.setModel(selectedModel.id);

   Setting pending reasoning effort before setModel is correct for Copilot (which uses it), but ensure this ordering is intentional for all agent types.

---

Performance Considerations

1. normalizeClaudeModelPreference string operations (settings.ts:37-51):

   const canonical = CLAUDE_CANONICAL_MODELS.find((name) =>
     normalized === name || normalized.includes(name)
   );

   Using includes could match unintended substrings (e.g., "haikuv2" would match "haiku"). Consider using a more precise match:

   normalized === name || normalized.startsWith(`${name}-`) || normalized.endsWith(`-${name}`)

2. Session rebind in Copilot (copilot-client.ts:437):

   • resumeSession appears to rebind the entire session with all tools and config. For frequent model switches, this could be expensive. Consider if there's a lighter-weight API.

---

Security Concerns

1. No input sanitization on model ID in ClaudeAgentClient.setActiveSessionModel:

   const targetModel = stripProviderPrefix(model).trim();
   if (!targetModel) { throw new Error("Model ID cannot be empty."); }

   While basic validation exists, consider validating against a whitelist of known model identifiers to prevent potential injection issues if the model string is used in API calls.

---

Test Coverage

Good:

• Unit tests for setActiveSessionModel behavior
• Tests for model preference normalization including canonical alias stripping
• Test for effective model usage in state updates

Missing:

• Integration test for end-to-end model switching during an active session
• Test for the new /init skill execution
• Test for CopilotClient.setActiveSessionModel (only Claude client is tested)
• Error handling tests for failed resumeSession calls

---

Suggestions

1. Split this PR - The /init skill refactor and model switching are independent features. Separating them would make review and potential rollbacks easier.

2. Add JSDoc comments to setActiveSessionModel interface method explaining:

   • When it's safe to call
   • What happens to conversation history
   • Expected behavior per agent type

3. Consider deprecation path - The removal of CLAUDE.md/AGENTS.md templates is a breaking change for users who relied on atomic init generating these files. Consider:

   • Adding a migration note
   • Keeping a minimal template as fallback if /init skill fails

4. Revert indentation changes in skill-commands.ts or make it a separate formatting commit

---

Overall Assessment

The core model-switching functionality is well-implemented with proper abstraction through the CodingAgentClient interface. The test coverage for the new functionality is reasonable. However, the /init skill and template removal changes deserve separate consideration due to the breaking nature of removing static templates.

Recommendation: Consider splitting this into two PRs:

1. Live session model switching (ready to merge with minor fixes)
2. /init skill refactor (needs more testing and migration planning)