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.
i18n::trf filled {name} placeholders with one replace per argument, so a value that looked like
a later placeholder was filled in turn. It now fills them in one pass over the translated
template: inserted values stay literal, and unknown names and stray braces are kept as written.
The saved-builder replacement notice uses it instead of its own split.
Review follow-up: folder_picked now refuses a folder whose name isn't text before dispatching either tool, with a message, and sends the exact UTF-8 path. Game files keeps one map of in-flight jobs with their kind and whether the screen still waits for them, instead of a looks list and a listing id. A worker failure is always logged; a current folder scan or listing clears only its own spinner and says so, a failed triage holds neither, and a superseded job says nothing. Export save errors say so without stopping the archive listing, even when the archive path matches a reopened listing.
The same-folder regression now starts both looks through folder_picked and feeds results through CraftApp dispatch; new regressions cover the non-text refusal, a failed triage keeping the scan running, a superseded failure staying quiet, and an export error not clearing loading.
Picking game A, B and A again while the first A look ran could still apply the older A result, because the screen compared the shown folder's path. Worse, a folder whose name has bytes that aren't text was sent as a lossy display string, so every completion failed the same path check and the screen loaded forever. act now returns its job id; GameFiles records the folder looks and the archive listing still expected, and tool_done applies only those. A repeated pick ignores the older answer, and a non-text folder's current answer finishes with its error.
Regressions: a stale same-folder look is ignored and the newer one fills the list; a non-text folder's current error clears looking and is shown. The three existing wrong-folder/archive guards still pass.
Picking game A, B, then A again while the first A look runs can still apply the older A result, because the screen compares the shown folder's path. Record the pending change before fixing it.
A folder walk says which folders it couldn't list and why, but five tools that walk a folder (two for Apple bundles, the app source import and comparison, the managed source check and the JavaScript source restore) read only "cut short". A folder that couldn't be opened was reported as a limit, with neither its name nor the reason, and some results still said they were complete.
- Each of those tools now names the folders and their reasons, and marks its result partial when a walk was cut or a folder couldn't be listed.
- When nothing is found and a folder couldn't be listed, the error says so ("no .cs files under …; 1 folder(s) couldn't be listed…") with what to do, since what was looked for may be in it.
- Two inputs with the same folder name are told apart in the notes ("dist (input 2)").
Codex now shows the model it uses ("Codex default" until one is chosen) with Change…, like Claude Code (#2109). The choices are its default (no model is passed, so its own settings decide) and the models `codex debug models` lists, in its order, leaving out the ones it hides; over SSH the remote Codex is asked. `ai model Codex` lists and sets them the same way. Until now the model could only be set by adding `cli_model` to config.toml by hand.
Follow-up to #2124. The pseudo-C could still show an argument's load in the call line after an instruction that overwrote the bytes it read, for writes the decompiler measured wrongly: 64-bit ARM pair and exclusive stores (both registers count, and non-temporal pairs are stores too), x86 bit writes through a register index, `movdir64b`, and a store that overwrites part of a Windows stack argument. Those are measured now. A register is treated as the stack frame only while the function really keeps its stack there, so a stack pointer loaded from elsewhere is no longer assumed clear of global data. 64-bit ARM loads also show the size they read (`*(uint8_t*)…`), and a pair store shows as two typed writes.
Measured on two real programs before and after (every function, default limits): an x86-64 `grep` folds 3,021 register arguments where it folded 3,023, the two lost being loads from a global with a store through a data pointer in between; an AArch64 `ls` folds 1,813 either way.
Follow-up to #2113. A window loading `file://localhost/ui/index.html` was looked up as `localhost/ui/index.html`, which could match the wrong page; `localhost` and an empty host both mean this computer now. And a page reference ending in a `.` or `..` part (`./boot.js/.`) names a folder, so it is no longer matched to the file before it.
Follow-up to #2116. A result with thousands of lists at its top level was still gone through in full, though its first 64 already decide what is listed. The top level is read first and the work stops at the 64th list; nested lists are only looked for when there is room left. What is listed doesn't change.
Follow-up to #2119, which stopped a load being shown after a store that may have changed the memory it read, but still let it through in several cases: on 64-bit ARM, where the size of an access wasn't known and a 16-byte pair store looked clear of a load 8 bytes on; when a stack address had an index, or the two addresses used different segments; after an instruction the decompiler doesn't model, such as a byte store on 32-bit ARM or an x86 string operation; and for a value loaded from a fixed address. A Win64 argument passed on the stack is also no longer shown with the value stored before a later write to its slot. Where it can't tell, the decompiler now keeps the load as its own statement.
Follow-up to #2108. A column added later to a SQLite table fills older rows with its default. Two kinds of default were read as empty: a blob written with a plus sign in front (`DEFAULT+X'41'`), which #2108's splitting broke, and a bare word containing accented or other non-ASCII letters or a `$` (`DEFAULT été`, `DEFAULT a$b`). Both are read as SQLite reads them. A test table that SQLite itself would refuse to create is replaced by one it accepts.
Follow-up to #2107. Listing a result's pageable lists looked each list name up from the top of the result again, so a large result with many names took one full walk per name (about 150 ms for 5,000 names in the author's measurement, under 1 ms now). The result is walked once, recording what each name reaches first, in the same order as before. What is listed doesn't change.