Notes on supporting other models

[mcintyre94]

Currently Wisp only uses Claude, but this probably won't be a limitation forever. Sprites support anything + come with Codex/Antigravity preinstalled (IIRC). This issue is just my notes on what we'd need to check before supporting something else:

Basically we need a headless mode with:

• tool use
• an equivalent of --verbose so it prints tool use
• streaming response format (Claude's --output-format stream-json)
• an equivalent of dangerously-skip-permissions that makes this mode useful (I think Sprites sets this by default on other preinstalled ones the same way it does for Claude)
• resume from session, so we can make it function like a multi-turn chat

And ideally also:

• support for our custom ask tool + disallowing their built-in one
• a way to append instructions to their system prompt
• preferably use the same built-in tool names so we don't have to map between them

[mcintyre94]

Reviewed the current codebase against this list — a few things we actually use that aren't mentioned here:

Missing from required:

• -p / non-interactive/print mode — the most fundamental flag and worth calling out explicitly. Any replacement would need a headless mode that doesn't prompt for input.

Missing from nice-to-haves:

• --effort {level} — we expose low/medium/high/max effort selection in the UI. Other models would need an equivalent thinking-budget control, or we'd need to hide the picker for non-Claude models.
• --max-turns {count} — we let users set a turn limit. Other models would need an equivalent flag or we'd handle it client-side.
• --disallowedTools {list} — used in Quick Chat to block dangerous tools (Write, Edit, WebSearch, WebFetch). This is broader than just disabling the built-in ask tool already mentioned.
• MCP support for WispAsk — the custom ask tool is delivered via --mcp-config. "Support for a custom ask tool" already covers the need, but worth noting MCP is the specific mechanism so any replacement would need MCP support, not just a generic tool injection path.

Implementation details worth tracking separately:

• CLAUDE_CODE_OAUTH_TOKEN env var — Claude-specific auth. Other agents use different env var names/token schemes.
• Session history format (~/.claude/sessions JSONL) — we read these files to reconstruct chat history when reopening a session. Separate from --resume; other models would store history elsewhere/differently.
• error_max_turns result subtype — we special-case this in the UI. Other models may surface turn limits differently.