tokenmaxx install pi

[DexterStorey]

Towards #11, replacing #14 at a smaller scope: pi only, since that is the harness actually in use.

What

tokenmaxx install pi merges a tokenmaxx-anthropic and a tokenmaxx-openai provider into ~/.pi/agent/models.json, pointed at the local proxy. pi re-reads models.json every time /model opens, so no restart is needed. tokenmaxx uninstall pi removes exactly the two providers we added and nothing else, and doctor shows a pi line whenever pi is on the machine.

An unparseable models.json is never touched: the CLI prints the exact block to add manually instead.

The proxy speaks the ChatGPT backend's dialect

Carried over from #14 unchanged: harnesses send standard Responses API requests — system prompts as messages in the input array, max_output_tokens set — and the ChatGPT codex backend rejects both. On the openai route for OAuth accounts, the proxy now lifts system and developer messages into instructions and drops max_output_tokens. Codex's own requests already have that shape, so the adapter is a no-op for them; API-key accounts hit api.openai.com and are left untouched.

Cut from #14

openclaw and hermes writers, the settings-tab harness panel, and the settings scroll machinery. The CLI plus doctor covers the pi story on its own.

Tested live

• bun run check: 84 tests, typecheck, biome all green.
• uninstall pi → providers gone from models.json, neighbours untouched; install pi → providers back; doctor shows ok pi.
• Real completion through pi 0.84.2 (pi --provider tokenmaxx-openai --model gpt-5.6-sol -p) into the running daemon: answered correctly and metered under the active codex OAuth account in token_events.
• The anthropic route was deliberately not exercised live: it pairs with an API-key account (tokenmaxx login claude --api-key) and none is registered on this machine. The provider entry and route mechanics are identical to the claude routing exercised daily.

🤖 Generated with Claude Code

[github-actions]

try this branch: bun add -g tokenmaxx@0.0.62-alpha.93