An API builder's kept conversation now records whether it ran in Deep mode or Clean room (`transcript.json` version 2), and is only resumed in that same mode; a version-1 transcript counts as Deep. A run without a new request whose transcript came from the other mode studies again instead of carrying on, and a chosen extractor is no longer started in Clean room.
The Android binary XML reader stays inside the document's declared size, reads a node's fields after the header length it declares, and requires each attribute record and typed value to fit within its own element. Nodes that are too short or attribute tables with an impossible layout mark the document truncated with a warning, instead of borrowing bytes from a sibling or from past the document's end.
Keyed archives now follow object references inside arrays, inline dictionaries and objects that are only an alias for another, both in the object list and when projecting an object. A `$top` that isn't a dictionary is no longer taken for an archive, and the text a projection copies from repeated references shares one 64 MiB budget.
The SQLite reader found a table's column list at the first `(` in its CREATE TABLE, so a table named `"export (1)"` or `[Sheet1 (2)]` lost its columns, rowid alias and WITHOUT ROWID order. The list now opens at the first parenthesis outside quotes.
A relocation's symbol name is read only from a linked symbol table and string table of the right types, and only when the symbol index falls inside that table, so relocations keep their record but no longer borrow names from neighbouring sections.
Solana program section headers are read from records of the declared entry size, which must be at least 64 bytes, and a truncated section header table is reported instead of fields being read from the next record or past the end.
A contract's CBOR metadata trailer is split off only when it decodes completely. Text that isn't UTF-8, or arrays and maps longer than the reader keeps, now invalidate the trailer, so a plausible length suffix alone no longer removes instructions from the analyzed bytecode.
A GNU build ID is read only when its description, after the padded owner name, fits inside the note region, so it can no longer borrow bytes from whatever follows the region.
When a section or segment holds more file bytes than it maps into memory, the padding past its virtual size no longer gets a virtual address, and reads at an address stop at the mapped end instead of running into the padding.
Pointer reads follow the image's width: 16-bit binaries read two-byte pointers instead of four, and an unknown width reads nothing instead of guessing.
A GNU base-256 number with its sign bit set is negative, which an unsigned tar header field such as a member's size can't be; it is now rejected instead of being read as a huge size.
Resource compilers write an empty version string with no value at all, and the PE reader dropped those. Strings such as `OleSelfRegister`, whose presence is what matters, are now kept with an empty value.
A doubled quote inside a quoted column name now stands for one quote instead of ending the name, and single-quoted names, which SQLite accepts, lose their quotes. Keys that name such columns find them, so rowid aliases and WITHOUT ROWID order hold.
The safetensors reader checks tensor sizes for the 4-, 6- and 8-bit float types too, refuses a header over 100 MB, warns about metadata values that aren't strings instead of showing them as JSON, shortens long architecture and name fields, and reports tensor data ranges that overlap, leave gaps or leave bytes unclaimed at the end.
The binary plist reader accepts the `bplist0x` versions and reports others as unsupported. An unknown simple marker, a date marker of the wrong width or a dictionary key that isn't a string is now an error, instead of being shown as null or as an invented key.
A WITHOUT ROWID row stored before ALTER TABLE ADD COLUMN has fewer fields than the table has columns. Those rows now show the added columns' DEFAULT values, as rowid tables already did, instead of NULL.
A comma or quote inside a comment in a CREATE TABLE no longer splits or swallows column definitions, a comment before or right after a column name isn't read as part of it, and a comment inside a definition stays out of its type, so INTEGER PRIMARY KEY still marks the rowid alias.
docs/opencode.md now says that runs need OpenCode 2.x, that the install script and npm's `opencode-ai` still give 1.x, and that 2.x comes from npm's `@opencode/cli`, the desktop app's bundled CLI or the GitHub releases, with how to point the backend at it. It also warns not to start OpenCode during a build, since its shared service takes over the build's session. STATUS.md records the Linux live runs, and the roadmap gains the two problems they found.
JFFS2 extents and nodes, ROMFS payloads and header pointers, cramfs names, SquashFS metadata tables and file blocks, and tar PAX records now stay inside the sizes their images record. Out-of-bounds or truncated parts are reported as warnings and mark the recovery incomplete instead of being read past or silently dropped. PAX records need their terminator and a bounded name, and split octal fields in tar headers are rejected.
On Windows, a .cmd wrapper whose only command is one quoted, absolute .exe path followed by %* is started as that .exe, so its arguments skip cmd.exe (no % expansion, no 8191-character cap, line breaks allowed). Any other wrapper still runs as it is. The test that starts the .exe is Windows-only and isn't run by the Linux gate.
Fixes#1237
An API backend's timeout_secs now bounds the wait for its answer to start as well as the whole answer, so a request queued behind another tool's on a shared Ollama isn't cut off at five minutes; unset keeps five minutes to start and an hour in all, and a wait that runs out says to set it.
Fixes#1239
In Settings → AI, a card for an AI found on this computer lays its buttons out first and wraps the name and reason in the width they leave, so a long reason no longer runs under the Add button.
OpenCode model discovery, which also runs before every OpenCode turn, now uses --standalone unless the person chose a --server. Plain opencode api started OpenCode's shared service, which resumes every unfinished session on the computer, in any project, and carries on with its commands.
Part of #1115.
A VIRTUAL generated column stores no field in a row, so stored values no longer shift one column left: such a column is shown as not read, with a note, in rowid and WITHOUT ROWID tables alike, and a rowid alias after it is filled again.
The older, cut-off copy of the standalone-server sentence in opencode.md's configuration section goes; the newer copy, which also says where the file goes over SSH, stays.
Five STATUS rows that merges had kept twice, with different updates in each copy, are now one row each, the most complete copy (OpenCode models and thinking now says once that SSH runs get the tools), and opencode.md loses its older doubled paragraph. A pitfall says how to merge a conflicting table row.
A --resume of the project's Deep session, or a --continue, saved in the Claude Code backend's extra arguments no longer bypasses the mode check: a Clean room build or question is refused before the agent starts, with how to fix it. Other agents, other sessions and Deep runs are unaffected.
A binary plist whose array or set names one stored object many times, as Python's plistlib writes equal values, is no longer refused; only its references must fit in the file. The node cap is checked before memory is reserved, and copied data and text share a byte budget, so repeats can't expand without bound.
When the shared investigation chat switches from Deep to Clean room, it starts a new agent session instead of resuming the Deep one, which may carry what Deep mode read; a context with no recorded mode counts as Deep.
Play, screenshots and self-tests start a Python project's one package with a __main__.py as python3 -m <name>, and a Maven or Gradle project's one runnable jar (its manifest names a Main-Class, read without loading the jar) with java -jar. Maven now builds with package, so the jar exists after our own build.
The four README pictures are re-rendered from the current app, showing what has landed since October, and the snapshot example gives every scene a made-up projects folder so no picture shows a real account's path.
A TEXT value cut at the 4 KiB cap ends at a character boundary in UTF-8 and UTF-16, so its shown prefix gains no replacement character it did not have; the value is still marked as cut.
craft.json keeps the mode a builder's saved session ran in (version 4, with a migration that counts an older saved session as Deep). A build or question resumes a session only in the mode it was saved in, so Clean room never continues one that may have read inside the original, and a Deep session is kept for a later Deep run.
Estonian (et) joins the app's languages with a complete catalogue, in Settings and the coverage check, and requests in Estonian keep their project kind: a game (mäng) or a website (veebileht, koduleht, veebisait), with the head of the main clause deciding when both are named.
A reply in Qwen's <tool_call> markup that a server handed back as text is read as the calls it makes, with arguments typed from the tool's schema; a call to a tool the model wasn't offered gets a plain answer naming its tools, instead of ending the turn.
A builder in the API loop, which can't open pictures, is no longer sent to look at screenshots: every brief and check that pointed it at a picture tells it what to use instead (capture.json's layouts, the UI tree, or asking the person). Agent apps keep the picture wording.
Source-map data URLs ignore their fragment, ZIP directories are told by their attributes, rejected ASAR names count toward the limit, a HAR with a UTF-8 BOM parses, .NET bundle members stay before the manifest, Play starts a Python project's one installed console script, and Godot and Electron get an absolute project path.
Supersedes #1140, #1141, #1142, #1143, #1144, #1145 and #1146.
Co-authored-by: Afloat <ShuffleMonkey@163.com>
Each backend in a fallback chain gets its own refit after its server refuses a request for its window, and a refused request is asked again only once the conversation was made shorter.
A remote Claude Code or Codex build checks that craft-anything-cli is on the host before it starts, as OpenCode's already did, and a test pins that a remote tool server is only sent flags every release knows.
The check before an Aider run reads the settings Aider itself reads: .env and .aider.conf.yml in the folder it runs in and at the top of its git repository, an --env-file, then the home folder, so a project that names its model or key is no longer refused.
A claude wrapper saved by name (claude.cmd) is kept when it starts Claude Code, and a link that leads to the Claude desktop app is passed over like the app itself.
Git's own author variables are no longer taken for secrets while the hook runs, so a claim that names its author passes; other AUTH-named variables are still checked, and a Unix home folder is listed once.
A stopped, failed or timed-out OpenCode turn deletes the session it started, so OpenCode's own service can't resume it later. A turn records its session while it runs, so one left by a killed app is deleted at the next turn or by doctor, on this computer or over SSH.
Part of #1115. Includes #1176.
A long adb or Frida output line is cut at the last character boundary within the cap, so a multi-byte character at the cut no longer keeps the whole line.
* Claim the build's note about a replaced program (#721)
* Say in a build's feed when it runs another program than the saved one (#721)
* Leave a bare program name found as the same agent out of the note
* Say which program to choose when a build runs another one
Aider gets its task with --message and flags that stop it asking, updating, editing .gitignore or opening a page; a chat outside a build gets --dry-run. A run is refused before launch when no model or key is set up or an OpenRouter model has no OpenRouter key (#72), and Clean room is refused because Aider's reads can't be checked.
Supersedes #1061.
Co-authored-by: Hasnain <161151300+hasnain607@users.noreply.github.com>
An API backend's streamed answer is bounded by the AI's timeout_secs (one hour when unset), so a slow local model can finish a long answer, and a stream that runs out of time says to set it.