* docs(formats): classify single-app formats and add ASAR/AXML/ARSC boundary tests (storytold/craft-libs#778)
* fix(formats): accept Truncated or Malformed for short ARSC header
* test(formats): use < 8 byte header in arsc short header test
* Cut over MCP line and JSON-RPC envelope framing to craft-libs (#778)
* mcp: clone request id before worker thread spawn
* chore: update Cargo.lock for craft-jsonrpc
* mcp: add deterministic stdio transcript parity and framing bounds tests
* chore: repin craft-jsonrpc and craft-mcp-server to craft-libs main
* Plan honest usage counts and a cost-limit design in the roadmap (R5.3)
* Say a limit that holds has to treat the turn's price as an estimate (R5.3)
---------
Co-authored-by: Simon Koch-Sultan <[email protected]>
Co-authored-by: Tanish Manku <[email protected]>
* Claim R1.5: mcp-http-cleartext-guard
* Validate MCP HTTP URLs and refuse cleartext or loopback redirects (R1.5)
- Parse `url_allowed` inputs with `ureq::http::Uri`, reject userinfo (`@`), validate `https://` hosts/IPv6 brackets and `u16` ports, and delegate `http://` loopback checks to `craft_anything_platform::standin_server::host_is_loopback` so malformed or spoofed authorities (`http://[::1].evil.example.com`, `http://evil.example.com\@127.0.0.1`, `https:///...`) cannot pass.
- Trim `ServerUrl::url` in `Http::new` to match `url_allowed`.
- Enforce `.https_only(true)` on `https://` MCP requests (reporting `RequireHttpsOnly` without echoing the redirect target URI) and `.max_redirects(0)` on loopback `http://` requests so plugin headers and payloads never follow redirects off loopback or across cleartext HTTP.
* Reject mid-host brackets and signed ports, and drain POST bodies in tests (R1.5)
- Reject mid-host square brackets (`!host.contains(['[', ']'])`) and leading `+` signs on ports across `https://` and loopback `http://` authorities in `url_allowed`.
- Drain incoming HTTP `POST` headers and `Content-Length` bodies (and clear inherited socket non-blocking mode) before sending `302` replies in the redirect regression fixtures.
* Follow loopback MCP HTTP redirects that stay on this computer (R1.5)
* Keep MCP redirect hops in one deadline, cancellable, and without configured headers off the configured origin (R1.5)
* Lengthen the per-hop delay in the MCP redirect deadline test and update STATUS (R1.5)
* Keep MCP redirect hops within the caller's deadline and the configured origin (R1.5)
Pass the caller's deadline into post without a one-second minimum, compare origins by scheme, host and effective port, refuse a session started by another origin, keep the path for query-only redirects, give the cancellation notice 2 s, and report a late answer body as a timeout.
* fix(engines): keep malformed MCP redirect targets out of errors and send cancel notices only for sent calls (R1.5)
* Discard the cached MCP client after a send-side protocol error (R1.5)
---------
Co-authored-by: Tanish Manku <[email protected]>
* Check both comparison backends before building either (R2.9)
* Keep the API refusal test independent of provider keys (R2.9)
---------
Co-authored-by: Tanish Manku <[email protected]>
* Check which list lookups run at the summary cap (R5.7)
* Record the list lookup boundary regression (R5.7)
* Use the workspace minimum Rust API for the lookup test (R5.7)
---------
Co-authored-by: Tanish Manku <[email protected]>
* Claim R2.2 Codex asset tool approval (#1464)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* Let a Codex builder call the paid asset tools within the project's budget (#1464)
codex exec can't ask for approval, so it refused assets_image_generate and
assets_model_generate before our server saw the call. Approve them on Codex's
side too, like rebuild_start: our own limits still apply (the project's
generated-assets switch, its budget, and the question before going over it).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: Claude Opus 5.5 <[email protected]>
Co-authored-by: Tanish Manku <[email protected]>
* Find a build only Cargo places by reading files, starting no process
The file lookup (the Project screen's Play and "Pin this build", a saved Live
pin) now also reads CARGO_BUILD_TARGET_DIR, build.target-dir in Cargo's home
config, and the closest workspace's target/. Every folder outside the app still
gives only a program whose build record names the app's sources.
* Say which of Cargo's rules the file lookup approximates; read a Cargo.toml only when it is a file under 1 MiB
A relative CARGO_HOME means no home config is read (not ~/.cargo).
* Bound the read of Cargo's config too; a decoy for CARGO_BUILD_TARGET_DIR; STATUS says what a Live pin asks and which rules aren't followed
* Test that Play looks in the folder Cargo names; say exactly which settings files are capped and what asks Cargo
---------
Co-authored-by: Tanish Manku <[email protected]>
* Claim: keep development builds from filling the disk with debug information (R5.2)
Co-Authored-By: Claude Fable 5.2 <[email protected]>
* Build development code with line tables only and dependencies without debug information (R5.2)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* Say how to bring back dependencies' debug information too
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: Claude Fable 5.2 <[email protected]>
Co-authored-by: Tanish Manku <[email protected]>
* Claim the live Go project checks (R2.7)
* Exercise rebuild checks with a real Go project (R2.7)
* Record live Go project check evidence (R2.7)
---------
Co-authored-by: Tanish Manku <[email protected]>
Every job that changes an AI hands back a snapshot of the AI list and its options, and the app showed whichever finished last. A snapshot read before a save but finished after it put the old values back on screen until the next refresh, which is also why claude_codes_account_folder_is_set_in_settings failed about one CI run in ten. Snapshots are now numbered as they're read, and one older than the snapshot already shown is dropped.
Co-authored-by: Tanish Manku <[email protected]>
* Claim formatting in batches that fit Windows' command line (R5.2)
* Format in batches that fit Windows' command line: cargo xtask fmt (R5.2)
* Say what cargo xtask fmt formats, and run the rustfmt RUSTFMT names
---------
Co-authored-by: Tanish Manku <[email protected]>
* Claim unreadable folder entry names (R3.11)
* Report unreadable entry names without merging file identities (R3.11)
* Format unreadable entry regression iterator
* Keep invalid filename fixtures off Apple targets
APFS rejects the raw filename bytes used by the two disk-backed regressions.
Keep the iterator-only limit coverage enabled on Unix and record the fixture
constraint in the contributor pitfalls.
Signed-off-by: Afloat <[email protected]>
---------
Signed-off-by: Afloat <[email protected]>
* Claim R2.9: start a new Claude Code conversation when the saved one is too long (#2042)
* Start a new Claude Code conversation when the saved one has grown too long (R2.9)
A conversation past what the model accepts made Claude Code refuse every later turn with "Prompt is too long", and Try again resumed the same conversation, so the project was stuck (#2042). That refusal is now read like a saved conversation that isn't on this computer: the saved session is dropped and the next message starts a new one from what the project already has.
---------
Co-authored-by: Tanish Manku <[email protected]>
egui-wgpu requests textures of 8192 px at most whatever the GPU supports, and the window's surface can't be larger, so on a display wider than that the app aborted before its window showed (#1639). The device request now keeps egui-wgpu's settings and raises the texture limit to the GPU's own, never lower than before.
Co-authored-by: Tanish Manku <[email protected]>
* Correct what Arch's Ghidra package installs, and what a Ghidra file setting accepts
Arch's extra repository ships ghidra (12.2), with /usr/bin/ghidra and
/usr/bin/ghidra-analyzeHeadless as links into /opt/ghidra; the earlier
text said it had no package and described a start script that was
written for the test. That real layout is found with each setting.
A Ghidra file setting is kept when it is a short text file naming the
launcher, which is not checked to run it; the wording said it was.
* Say that any file in a Ghidra folder gives its launcher, and that only Arch's own package was tried
A readable sqlite_schema row could still hold a partial value (an overflow chain that ends
early, loops or leaves the file, or text cut at a schema budget), so a table's column list could
be short while the schema was reported complete. Such records now set schema_truncated and add
a counted integrity note; the object's readable name and metadata are kept.
A squashfs 2.x or 3.x superblock was taken as a filesystem on its magic, version and bytes used
alone. The firmware scan now also requires its block size to match its block_log, from 4 KiB to
1 MiB, as version 4's already must. A superblock cut before those fields is no longer reported.
A note added while an API-backend builder worked waited for its whole turn to end, which can be
hundreds of steps. After each tool step the loop now claims the waiting notes and adds them
after that step's results, in the wording the coding-agent hook uses. Claimed notes are marked
read, so the build shows them in the feed and doesn't send them again after the turn.
An API backend with requests_per_minute set held each request back without a word, so a paced
build looked stuck. A wait of two seconds or more now puts a line in the feed saying how long
it waits and why (#668).
The schema listing dropped its b-tree walk's incomplete flag and notes, so a schema with an
unreadable child page or row was listed as the whole schema, with complete coverage. The
overview now marks such a schema truncated and carries the walk's first notes, and
db_sqlite_show adds a schema_incomplete limitation, partial coverage, and an unknown-table
error that says only the listed part was searched.
The walk listed a folder's files before opening any of its subfolders, so when the file limit stopped it, a file inside an earlier folder was dropped for later ones and the inventory's sort couldn't bring it back. It now opens a subfolder where its name sorts in its parent, and results are sorted folder by folder (`path_order`), so a limit cuts the tree where a listing of it would end. Only the listings of the folders on the current path are held, within the same entry limit. When the entry limit runs out early, fewer entries are reported than before: the honest prefix.
The scan read every squashfs region's block size at offset 12, where only version 4 keeps it;
in 2.x and 3.x that is the uid table's start, and the block size is at offset 51. Older images
now report the block size they declare, and their description no longer names a compression
read from a field those versions don't have.
The build steps filled {editor}, {player} and {ext} with chained replace calls, so an editor
installed under a folder named like a later placeholder reached the builder with a different
path. The one-pass fill moves to platform::text::fill, which both the build steps and i18n::trf
use.