Code graph and memory for your coding agents, in one Rust program. For Claude Code, Codex, Cursor and any MCP client.
Meta Global AI Developer Hackathon 2026
Graphicle is entered in the startup track of Meta's Global AI Developer Hackathon, 26 October to 4 November 2026. The event's rules ask entries built on an existing product to say clearly what is new. This page does that, and it is updated as work lands.
Before the event
- Graphicle 0.1.8, released on 26 October 2026 before the kickoff, is the baseline.
- Written on 26 October before the kickoff, and not released before the event:
- a first version of
graphicle pack, which fills a token budget with the parts of a repository a question needs; - first versions of
graphicle ask --packandgraphicle plan; - a Meta Model API profile for Muse Spark, and Muse Code in
graphicle setup; - the harness and question sets of the whole-repository benchmark, and a first tuning round on its two development repositories (kubernetes and home-assistant).
- a first version of
Built during the event
Everything below was started after the kickoff, at 26 October 2026, 16:00 UTC. Each item names the release it shipped in.
The commit ids are those of Graphicle's development repository, with the times they were committed (UTC); this public repository receives each release as one commit. From 0.1.10 on, items give times without ids. The commands that call Meta's models were tested against stand-ins for the Meta Model API (the tests' local servers and demo/mock_meta.py), not against the live API; the 0.1.9 notes say so feature by feature. From 30 October 2026, the flow and the benchmarks' runs on Muse Spark and Muse Wonder also ran against the live Meta Model API, and the open-weights runs against Muse Spark Open on Fireworks' serverless API (docs/flow.md gives the flow's runs).
- Cache-aware packs, the spend log and
graphicle fix(0.1.9).graphicle pack --layout cacheputs first a repository part that is the same bytes for every question about a commit, so Meta's prompt cache can serve it. With--layout cache,ask --pack,planandfixsend a prompt cache key to the Meta Model API; they ask it for exact token counts near the input limit, log each call's tokens and cost (graphicle usage), and refuse a call past a profile's cost caps.graphicle fixmakes a change for an issue on a new branch in a new worktree and runs your tests on it. Commits3e0d7e8(26 October, 17:08),854c0ccto925f8c8(5 commits, 26 October, 17:42 to 18:29), and two rounds of review fixes:1e566f8to7ad616b(2 commits, 26 October, 19:17) andad89536to4233438(2 commits, 27 October, 01:20).graphicle fixpacks in the standard layout by default, so the whole budget goes to the files the issue needs:4578a5a(27 October, 09:13). - The whole-repository benchmark (its kit and page, 0.1.9). The grid of model runs, fixed before any model call; the runner, with the prompt cache and spending limits; the cost of a question with the prompt cache, computed at list prices and not billed; and the held-out baseline of the pack written before the event. Commits
c5ca2cato0883169(11 commits, 26 October, 16:58 to 27 October, 01:00). Three development repositories for TypeScript, Java and C++ (angular, keycloak and godot):a1f83cf(27 October, 02:39). The benchmark's model runs were made later, on 30 October. - Pack ranking for TypeScript, Java and C++ (0.1.9). A second tuning round on the five development repositories, then one measurement on the three held-out repositories, whose numbers are in the 0.1.9 notes:
cbbf016(27 October, 05:08). An issue whose title names a version is ranked as written instead of being refused as an empty question:acf5f4e(27 October, 07:20). Then the review's fix to how a pack reads an issue's environment sections: only a label below the title opens one, an issue form's question counts, and a<name> version: <value>line is left out; the results of all 200 development questions are unchanged, and the held-out questions were not measured again:b8023f1to2955bba(8 commits, 27 October, 08:56 to 11:34). - Pixel to code and the hosted demo (0.1.9).
graphicle pixelfinds the code that draws an element of a screenshot: SAM 3.1 finds the element, a vision model (Muse Spark by default) reads it, and Graphicle looks up its strings in the code.demo/is a small web app for trying Graphicle with Meta's models; it is not deployed. Commits2dff779to8c352e8(3 commits, 27 October, 01:51 to 02:18) and2ec27e2(27 October, 02:41). graphicle reviewand its GitHub Action (0.1.9). A review of a change from a cache-aware pack, each comment checked against the changed lines and its citations before it is shown, and a composite GitHub Action that posts one review on a pull request. Commits92e89fcto04dbf74(3 commits, 27 October, 02:21 to 02:26).- Graphicle inside Muse Code (0.1.9). A native Muse Code plugin (the MCP server, a skill and hooks), session memory through Muse Code's session logs, a two-session memory demo, a harness that measures Muse Code with and without the plugin, and docs for Claude Code and Codex on Meta's models. Commits
586c339to99a1a1d(6 commits, 27 October, 06:22 to 07:00). The review's fixes,647941atoe580002(6 commits, 27 October, 08:56 to 09:18): the plugin is a preview that you add withgraphicle connect muse --hooks, since it has not yet run in Muse Code itself; its server is marked optional; the docs say to rungraphicle disconnect musebefore uninstalling; and the probe also records what Muse Code does when a plugin's server cannot start. - See it, find it, fix it, check it (0.1.9).
graphicle seedrives a web app with computer use on the Meta Model API until a bug shows;graphicle fix --focusputs the placesgraphicle pixelfinds first in the pack;graphicle pixelfollows the keys of English JSON locale files; anddemo/flow/flow.pyruns the whole path, from the bug on the screen to a tested fix, with a receipt. Commitsd3f3ec3tofd37c39(6 commits, 27 October, 06:26 to 06:57), and the review's fixes,dc48880to7e50e4f(9 commits, 27 October, 07:55 to 08:59): the Meta key kept from the browser driver, the tests and the app's build, a browser that blocks service workers and closes WebSockets to other origins, one budget for a whole run, the model's reasoning replayed from step to step, screenshots counted by their size, and the docs saying the flow has run only against the stand-in. - Setup where it cannot write a file (0.1.9). As in Claude Code's sandbox, which does not let it write
.git/hooksor.mcp.json: a file setup cannot write is skipped with its reason, and the summary gives the lines to add by hand. Reported by Thomas Thomsen in #15. Commitf29a86b(27 October, 08:55). Then the rest of the issue: a hooks file setup cannot write (Claude Code's.claude/settings.local.json, Codex's~/.codex/hooks.json) is skipped with its reason while the agent stays connected, and a read-only.gitkeeps its exclude file while the summary gives the lines to add:e1c0788to66643b6(4 commits, 27 October, 11:52 to 11:59). - Also during the event (0.1.9). The fix for four zstd advisories (
2f0ced3, 27 October, 04:58); the README's section "Agent memory, compared" and its follow-ups (7e1dd42,2b17d73,7c65015,7ac0c0bandb4ca138, 27 October, 05:32 to 05:58), which describe memory comparisons run before the event; a test's formatting (f20cb41, 27 October, 06:54); and this page (f7a67fd, 26 October, 16:42). - The 0.1.9 release itself: the version numbers, the 0.1.9 notes and this list, on top of the work above.
- Whole top files in standard packs (0.1.10). A standard pack of 32,768 tokens or less holds its first ranked file whole when the file fits in half the budget, and one of 131,072 tokens or more its first two when each fits in a quarter; the settings were tuned on the development repositories, then measured once on the held-out ones and on SWE-bench Pro (where a diagnosis of 0.1.9 chose the change's directions, though not its settings), and the 0.1.10 notes give the numbers. Written 27 October, 19:36 to 20:52 UTC.
- A folder's HTML files kept as they are (0.1.10). Packs, search and edits read a page's own lines instead of the text extracted from it. The code review's fixes: a page past the text-file limit or not in UTF-8 is converted as before, the pages an earlier version converted are synced again,
graphicle reviewescapes markup in what it posts to GitHub, an XML pack keeps a page's closing tags inside CDATA, pinned passages keep their room, and HTML files no longer count as notes. Written 27 October, 17:44 to 19:36; the review's fixes 28 October, 03:40. - SWE-bench Pro and line-level recall (0.1.10). A kit that measures without a model how often a context holds every file and every line a fix changes, on 633 tasks of SWE-bench Pro's public set, with the scripts of its model runs, whose results are not in this release, and new arms and scoring for the whole-repository runner. The same count at line level for the whole-repository questions. The public tables, the kits' settings and the entries of the 0.1.10 notes. Written 27 October, 06:04 to 21:02, and 28 October, 07:48 to 08:04.
- Graph commands before the first extract (0.1.10). A
graphicle graphcommand that reads the graph file says how to build it when it does not exist yet, as Thomas Thomsen reported in #15. Written 28 October, 07:00. - The code review's fixes to the 0.1.10 notes, the kits, the benchmark pages and a sync error message (0.1.10). Written 28 October, 08:56 to 09:02.
- A key over plain HTTP to hosts you list (0.1.10). A model profile with an API key may use plain HTTP to a server on your own network whose exact host is listed in
GRAPHICLE_PROVIDER_HTTP_KEY_HOSTS, which is off by default and sends the key in clear text, as asked in discussion #16. Written 30 October, 06:07 to 06:19 UTC. - The 0.1.10 release itself: the version numbers, the 0.1.10 notes and this list, on top of the work above.
graphicle fixreads its failures (0.1.11). A round's test output keeps the lines that report failures when it is cut, and the next round reads the files that the failure and the model name, when the pack does not hold them whole. Written 30 October, 16:04 to 16:50 UTC.- The flow, run live (0.1.11). The check keeps its budget when the fix runs on another profile; computer use defaults to Muse Spark, which ran it live on 30 October (Muse Wonder answered the computer tool with HTTP 400 on the key tested); the docs give the live runs and the Muse Code plugin's first live session. Written 30 October, 17:20 UTC, and 31 October, 10:39 UTC.
- Benchmark kits for the live runs (0.1.11). A
fireworksprofile for the whole-repository runner (Muse Spark Open on Fireworks' serverless API), and the Muse Code kit reading the events of Muse Code 1.4.14. Written 30 October, 10:21 and 16:41 UTC. - libz-sys 1.1.31 (0.1.11), for RUSTSEC-2026-0418, without the exception 0.1.10 carried for it, and the third-party licence pages that list it. Written 31 October, 10:49 and 11:25 UTC.
- Three
askfailures, reported in #17 (0.1.11). An answer that cites more than the 12 sources its format allows keeps the first 12 instead of failing;--workflow reflectskips a source it cannot have instead of failing; andask --pack,plan,fixandreviewuse a model key saved in the app, asaskanddoctordo. Written 31 October, 11:24 to 11:57 UTC. - The Muse Code plugin's session start (0.1.11) stays within Muse Code's limit on a hook's output (16,384 bytes; the plugin now prints at most 14,000), with a smaller session map for Muse Code, and the docs give the plugin's first live session as recorded. Written 31 October, 11:39 to 12:16 UTC.
- The review's fixes to the fix rounds (0.1.11). Failure lines anywhere in a test output of up to 128 KiB are kept, the added files take only the room the request has left, a round that went without its files says so and does not hand them on, and a round reads only the project's own regular files, never a symbolic link. Written 31 October, 12:08 to 12:42 UTC.
- The 0.1.11 release itself: the version numbers, the 0.1.11 notes and this list, on top of the work above.