claude-thermos

Stop paying to rebuild your Claude Code cache. When your main agent waits on a subagent for more than 5 minutes, its prompt cache silently expires, and the next turn re-encodes your entire conversation at the write rate instead of reading it back cheap. On long sessions with many subagents that's roughly 20% of your bill. claude-thermos keeps the cache warm so you never pay that tax.

Use

Run Claude Code exactly as you normally would, but through claude-thermos with uvx:

Requires Python 3.11+ and the claude CLI on your PATH.

That's it. Warming runs automatically in the background. To disable it for a run without changing the command, set CLAUDE_WARMER_DISABLE=1.

Tuning (all optional):

Why your cache keeps expiring

Claude Code's prompt cache uses a 5-minute TTL. Every turn, your whole conversation history is served from cache at 0.1x the input price instead of being re-sent at full price, as long as the cache stays alive.

The cache expires if more than 5 minutes pass between requests on the same prefix. The dominant trigger for that gap is not you thinking. It's the main agent blocked on a subagent that runs longer than 5 minutes. A subagent has a different system prompt and tool set, so its requests have a different cache prefix and never refresh the main agent's. While the subagent works, the main agent's cached history ages untouched; past 5 minutes it's gone. When the subagent returns, the main agent resumes with a byte-identical, append-only history, and finds its cache missing, forcing a full re-encode at the 1.25x write rate.

By then the history is large, so the re-encode is expensive: individual collapses re-write 200K to 500K tokens. Measured across roughly 185 local sessions, these rebuilds accounted for about 22% of the total bill, money spent re-encoding content that was already cached moments earlier.

How it works

claude-thermos launches Claude Code behind a small local reverse proxy (it points ANTHROPIC_BASE_URL at a loopback port; all traffic still goes to the real Anthropic API).

Observe. The proxy watches /v1/messages traffic and groups it into sessions and lineages, a lineage being one cache prefix, keyed by model + tool set + system text. The first tool-bearing lineage is the main agent; the rest are subagents.

Detect the danger window. When the main lineage goes idle and a subagent is actively running, the main prefix is at risk of expiring.

Warm. On an interval under the 5-minute TTL, it replays the main agent's last real request as a warm request: identical cacheable prefix, but max_tokens: 1 and no streaming. The single token is thrown away; the point is the prefill, which reads and refreshes the full cached prefix. Warm requests go directly to the API, never through the proxy, so they can't disturb real traffic.

Result. When the subagent finishes, the main agent's cache is still warm. It pays a cheap read instead of a full rewrite.

Each warm costs a cache read (0.1x); each rewrite it prevents would have cost a write (1.25x) on a much larger prefix, so the trade is heavily in your favor.

Event logs & savings

Every session writes to:

events.jsonl records each request/response's token usage plus every warming decision (warm_fired, warm_result, cap_reached, resume_detected, and so on). summary.json is the rollup you'll usually read:

All three cost figures are in base-input-token units (token counts already weighted by their cache multiplier). To turn net_savings into dollars, multiply it by your model's price per input token:

For example, at an input price of $3 / 1M tokens, a net_savings of 1_200_000 is about 1_200_000 × $3 / 1_000_000 = $3.60 saved that session.