collie/cache-rules.toml.example at main · idIing/collie · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
# Your own prompt-cache lifetimes. This file goes next to your `.env`; docs/configure.md, under
# "Your own slash commands", names that directory for each kind of install. On a Herdr plugin:
#   cp cache-rules.toml.example "$(herdr plugin config-dir herdr.collie)/cache-rules.toml"
#
# This file configures ONE thing: how long Collie believes a harness's prompt cache stays warm. Each
# row moves exactly one shipped rule, by id. Everything else about the chip is unchanged.
#
# WHEN YOU NEED IT. Collie ships a TTL per harness and provider, each one read off a vendor's own
# page on a recorded date. A vendor can move that number in a blog post, and a gateway in front of
# your agent can change it without telling anybody. Until Collie ships a new rule, this is the lever.
#
# `collie doctor` tells you when a shipped claim is going stale: its `cache-claims` line warns on any
# rule nobody has re-checked in 180 days, and its `cache-rules` line names any row below that this
# file got wrong. Its `cache-env` line fires when you have set `ENABLE_PROMPT_CACHING_1H` or
# `FORCE_PROMPT_CACHING_5M` in your shell — the bridge runs as its own systemd unit and cannot read
# your agent's environment, so a variable you set for Claude Code has to be mirrored here to be seen.
#
# FOUR FIELDS ARE REQUIRED AND ONE IS NOT. `id`, `ttl_seconds`, `source_url` and `retrieved` must all
# be there, or the row is dropped with a line in the log and in `collie doctor`. You may move a
# number; you may not remove the page and date it came from — that is the whole contract the rule
# catalog rests on. A `retrieved` date later than today is dropped the same way: nobody read a page on
# a day that has not come. `note` is optional and is shown in the pane sheet.
#
# A MEASURED NUMBER STILL WINS. Claude Code writes which cache window it used into its own
# transcript, so a session it says is on the one-hour TTL stays on the one-hour TTL whatever this
# file says. An override sits above the shipped rule and below a live measurement.
#
# Rule ids, and what each one is:
#   claude.subscription   Claude Code on a Claude subscription (Pro/Max)
#   claude.api            Claude Code on an API key or third-party provider
#   codex.subscription    Codex CLI signed in with a ChatGPT plan
#   codex.api             Codex CLI on an OpenAI API key
#   pi.anthropic          pi talking to Anthropic        opencode.anthropic
#   pi.openai             pi talking to OpenAI           opencode.openai
#   pi.google             pi talking to Google           opencode.google
#   pi.unknown            pi on an upstream that documents no TTL, and the opencode twin
#
# Edits are live: the bridge re-reads this file behind an mtime check, so no restart is needed.

# A gateway in front of Claude Code that always asks for the one-hour window.
# [[rule]]
# id = "claude.api"
# ttl_seconds = 3600
# source_url = "https://platform.claude.com/docs/en/build-with-claude/prompt-caching"
# retrieved = "2026-09-12"
# note = "our gateway sends ttl 1h on every request"

# Codex on a model whose 30-minute window you confirmed on OpenAI's own page.
# [[rule]]
# id = "codex.api"
# ttl_seconds = 1800
# source_url = "https://developers.openai.com/api/docs/guides/prompt-caching"
# retrieved = "2026-09-12"