Repository navigation
Research: which limit and context-window data the model provider and SDK expose #223
Description
Question
What limit and context-window information can an agent instance obtain at runtime: rate-limit headers or usage endpoints from the Meta API and the LiteLLM gateway, model context-window size, and what the pinned Claude Agent SDK surfaces (usage events, rate-limit events, compaction signals)? Which of these are reliable enough to show on a dashboard, and which are only estimates (TokenLedger estimates vs measured)?
Part of map #222. Findings feed the telemetry snapshot contract.
Activity
lbacik commentedon Oct 14, 2026
Research findings are on branch
research/provider-limits-and-context-window, filedocs/research/provider-limits-and-context-window.md(primary sources only: Meta, LiteLLM and Claude Code docs, plusclaude-agent-sdk==0.2.163source and bundled CLI 2.1.286).Summary:
x-ratelimit-{limit,remaining}-{tokens,requests}headers. No usage endpoint. Context window is documented as 1,048,576 tokens for all Muse Spark versions (Contributor variant not listed separately); max output tokens are not documented.GET /v1/modelsexposes no window field.count_tokensis available.x-ratelimit-*(key limits, else provider values),x-litellm-*cost/latency/retry headers andllm_provider-*passthrough;/model/infoand/key/infoare usable out-of-band. Header presence on/v1/messagesand/model/infotoken-limit fields are unverified.RateLimitEventmodels claude.ai subscription windows only, so it is not usable on Meta. Raw backend response headers are not exposed to the SDK.contextWindow/maxTokensare the CLI's assumed values for an unknown model ID (200000 seen in the earlier Meta probe vs 1,048,576 documented); none of the override variables (CLAUDE_CODE_MAX_CONTEXT_TOKENSetc.) is set in the repo. Compaction arrives asSystemMessagecompact_boundarywithcompact_metadata {trigger, pre_tokens, post_tokens?}per the CLI schema.usageviamessage_delta,model_usagetoken counts and compaction events are measured;TokenLedgerestimates, subagent contexts,total_cost_usd, and the SDK context-window figure are estimates or guesses.measured/estimated/configured), treat the context window as configuration, and mark provider rate-limit headroom "unavailable" in v1 unless read out-of-band from LiteLLM.The note ends with five open live checks. Issue left open; map not edited.