Research spike: output token limits — pi-mono + Claude CLI patterns

[bonk-moltbot]

What

Research how other agent harnesses limit LLM output to prevent rate limit exhaustion, then decide what limits Anima should enforce.

Why

Anechka hit the 5-hour rate limit by noon today. The root cause: unconstrained output generation — long text responses, verbose tool calls, sub-agents generating large volumes. Existing tickets (#430 think tool char limit, #425 sub-agent tool call limits) cover specific cases, but there is no holistic output budget strategy.

Research targets

1. pi-mono (open source agent harness)

• Repo: https://github.com/badlogic/pi-mono
• How does it limit per-message output tokens?
• Does it enforce limits on sub-agent output?
• Any per-session or per-hour budget tracking?

2. Claude CLI (claude-code)

• What max_tokens does it set per API call?
• Does it vary by context (tool calls vs conversation)?
• Any output budgeting across a session?

Deliverable

A thoughts note with findings + recommendation: should Anima limit output tokens per-message, per-session, or both? For the main agent and for sub-agents separately.

Related

• Think tool character limit not enforced — allows multi-page dumps #430 — Think tool character limit not enforced
• Sub-agent tool call limits with soft landing via tool_choice #425 — Sub-agent tool call limits with soft landing
• Token Economy HUD: rate limit indicator flickers to default values #423 — Token Economy rate limit indicator flickers