fix: credential detection and provider auth status (#120)

## Summary

Fix credential detection bugs and add credential availability status to the API. Consolidate Claude fallback models and add `sonnet` alias.

Builds on #109 (OAuth token support).

Related issues:
- Fixes #117 (Claude, Codex not showing up in gigacode)
- Related to #113 (Default agent should be Claude Code)

## Changes

### Credential detection fixes
- **`agent-credentials/src/lib.rs`**: Fix `?` operator bug in `extract_claude_credentials` - now continues to next config path if one is missing instead of returning early

### API credential status
- **`sandbox-agent/src/router.rs`**: Add `credentialsAvailable` field to `AgentInfo` struct
- **`/v1/agents`** endpoint now reports whether each agent has valid credentials

### OpenCode provider improvements
- **`sandbox-agent/src/opencode_compat.rs`**: Build `connected` array based on actual credential availability, not just model presence
- Check provider-specific credentials for OpenCode groups (e.g., `opencode:anthropic` only connected if Anthropic creds available)
- Add logging when credential extraction fails in model cache building

### Fallback model consolidation
- Renamed `claude_oauth_fallback_models()` → `claude_fallback_models()` (used for all fallback cases, not just OAuth)
- Added `sonnet` to fallback models (confirmed working via headless CLI test)
- Added `codex_fallback_models()` for Codex when credentials missing
- Added comment explaining aliases work for both API and OAuth users

### Documentation
- **`docs/credentials.mdx`**: New reference doc covering credential sources, extraction behavior, and error handling
- Documents that extraction failures are silent (not errors)
- Documents that agents spawn without credential pre-validation

### Inspector UI
- **`AgentsTab.tsx`**: Added credential status pill showing "Authenticated" or "No Credentials"

## Error Handling Philosophy

- **Extraction failures are silent**: Missing/malformed config files don't error, just continue to next source
- **Agents spawn without credential validation**: No pre-flight auth check; agent's native error surfaces if credentials are missing
- **Fallback models for UI**: When credentials missing, show alias-based models so users can still configure sessions

## Validation

- Tested Claude Code model aliases via headless CLI:
  - `claude --model default --print "say hi"` ✓
  - `claude --model sonnet --print "say hi"` ✓
  - `claude --model haiku --print "say hi"` ✓
- Build passes
- TypeScript types regenerated with `credentialsAvailable` field

research/agents/claude.md

@@ -279,6 +279,44 @@ x-api-key: <ANTHROPIC_API_KEY>
 anthropic-version: 2023-06-01
 ```
 
+## Command Execution & Process Management
+
+### Agent Tool Execution
+
+The agent executes commands via the `Bash` tool. This is synchronous - the agent blocks until the command exits. Tool schema:
+
+```json
+{
+  "command": "string",
+  "timeout": "number",
+  "workingDirectory": "string"
+}
+```
+
+There is no background process support. If the agent needs a long-running process (e.g., dev server), it uses shell backgrounding (`&`) within a single `Bash` tool call.
+
+### User-Initiated Command Execution (`!` prefix)
+
+Claude Code's TUI supports `!command` syntax where the user types `!npm test` to run a command directly. The output is injected into the conversation as a user message so the agent can see it on the next turn.
+
+**This is a client-side TUI feature only.** It is not exposed in the API schema or streaming protocol. The CLI runs the command locally and stuffs the output into the next user message. There is no protocol-level concept of "user ran a command" vs "agent ran a command."
+
+### No External Command Injection API
+
+External clients (SDKs, frontends) cannot programmatically inject command results into Claude's conversation context. The only way to provide command output to the agent is:
+- Include it in the user prompt text
+- Use the `!` prefix in the interactive TUI
+
+### Comparison
+
+| Capability | Supported? | Notes |
+|-----------|-----------|-------|
+| Agent runs commands | Yes (`Bash` tool) | Synchronous, blocks agent turn |
+| User runs commands → agent sees output | Yes (`!cmd` in TUI) | Client-side only, not in protocol |
+| External API for command injection | No | |
+| Background process management | No | Shell `&` only |
+| PTY / interactive terminal | No | |
+
 ## Notes
 
 - Claude CLI manages its own OAuth refresh internally