0.1.3 — Two fixes for "Prompt is too long"

Two fixes for Prompt is too long, which chunking is supposed to make impossible.

Requests no longer inherit your global Claude Code context

Runs failed reporting ~1,149k tokens against a conversation of only ~301k. The missing ~848k were MCP tool definitions and skills pulled from the user's global Claude Code configuration into what should be an isolated prompt-in, text-out request.

The Agent SDK gates each context source separately, and settingSources: [] disables only settings files and CLAUDE.md. MCP servers from ~/.claude.json, project .mcp.json and plugins load regardless, and the CLI performs skill discovery even when the skills option is omitted — omitted is documented as not meaning off.

Every switch is now set explicitly in one tested buildQueryOptions(): tools, skills, mcpServers, strictMcpConfig, plugins and settingSources. Verified against an environment carrying ~850k tokens of MCP config: a probe request now totals 171 input tokens.

The chunk budget is clamped to the summary model's context window

maxChunkTokens (600k, sized for Sonnet's 1M window) was converted straight into a character budget regardless of the model actually configured. On any 200k-context model — Haiku, older pinned ids, anything unrecognised — a single chunk could exceed the whole window.

maxChunkTokens is now a cap rather than a promise: it is clamped to the summary model's window minus a 32k reserve for the system prompt, SDK scaffolding and the response. Known 1M-window models keep the full budget; everything else is floored at a conservative 200k, so the worst case for an unrecognised large model is more chunks than strictly necessary.

Also

• npm publishing on version tags via CI.
• Claude SDK, CLI and OpenTUI dependency bumps.

Full changelog: 0.1.2...0.1.3