Repository navigation
Pebble 1.3.0
Pebble 1.3.0
Requests that a model provider can cache, and fewer round trips. Above 8K tokens of context, every request of a task
now extends the previous one instead of rewriting its start, so a provider's prompt cache can reuse what it has
already read. On the ChatGPT route the model may also ask for several tool calls at once. How much this lowers cost
and time on real tasks is measured in the next benchmark and is not claimed here.
Highlights
- Above 8K, each request of a task now extends the previous one, so prompt caching can reuse it.
- The ChatGPT route can ask for several tool calls in one step; Pebble still performs them one at a time.
- On the Anthropic route the model's earlier reasoning stays valid between steps, since nothing before it changes.
Changed
- Requests above 8K extend each other. Measured on 1.2.0 in a scripted coding task, every request from the
second on kept only the system prompt the same: the task checkpoint was rewritten after each tool result, and the
file audit, the plan and the diagnostics were inserted near the top and rewritten. Now, above 8K:- the task checkpoint is written once, with the local time the task started;
- the file audit and a changed plan are added after the newest tool results, listing only what changed; earlier
ones stay where they were, and only the newest plan is protected when older context is archived; - pinned skills, plugin schemas and connector routes stay where they were first added;
- the system prompt carries the date only, so it is the same for a whole day's tasks.
In the same scripted tasks every request now extends the previous one. Two events still change the start of a
request, on purpose: diagnostics are added and removed as they come and go (one that no longer matches its file is
removed, so the model never acts on it), and ReadSource joins the tools once a web page has been read. At 8K and
below nothing changes.
- Several tool calls per step on the ChatGPT route (above 8K). Pebble 1.0.4's kept benchmark runs spent 99 % of
their time waiting on the model, and the slowest ones read three files in three separate requests. The model may
now ask for independent calls together; Pebble performs them in order, each under its own permission and approval.
For contributors
- Each model request above 8K records how much of the previous request it kept (
model_attempt.prefix), content-free. backend/tests/test_cache_first_35.py;scripts/validate_phase35.py. D64: validate_phase28 records, rather than
asserts, that the first message changes per step; every check on the stale-block binding is unchanged.
Install
On a Mac with Apple silicon, in Terminal:
curl -fsSL https://raw.githubusercontent.com/tohin003/Pebble/main/install.sh | bash
The installer checks the download against SHA256SUMS, keeps any earlier Pebble.app in ~/PebbleBackups, never uses sudo and never changes Gatekeeper. --dry-run shows what it would do; --uninstall moves the app aside and lists what Pebble keeps on your Mac.
To install by hand, download Pebble-macos-arm64.zip and SHA256SUMS into one folder, check the zip with shasum -a 256 -c SHA256SUMS (it should say Pebble-macos-arm64.zip: OK), unpack it, and drag Pebble.app to Applications.
Signing and checksums
- Pebble is signed with the project's own stable signing identity, not an Apple Developer ID, and it is not notarized by Apple.
- SHA256SUMS lists the zip's SHA-256. Checking it shows that the download arrived intact. SHA256SUMS is not signed and comes from the same release, so it does not prove who made the file (integrity, not authenticity).
- A zip downloaded in a browser is quarantined, so macOS may refuse the first launch. Allow Pebble alone: Control-click Pebble.app, choose Open, then Open again (or System Settings, Privacy & Security, Open Anyway). Never turn Gatekeeper off. The one-line installer downloads with curl, which does not quarantine.
SHA-256 of Pebble-macos-arm64.zip: eb96a7b070701864a3be41bec219b239c7740960462bbb9f0b696361371530bd