All issues

[code-review] the webrtc peer treats ICE Disconnected the same as Failed and tears down the DataChannel immediately: pion usually self-heals a transient disconnect by regathering candidates, but this stack has no ICE restart path at all, so every wifi hiccup falls back to relay and later rebuilds the whole tunnel from scratch - a new PeerConnection plus full signaling round that costs the mobile side roughly 50MB per rebuild per the code comments - meaning protocol-level recovery is abandoned in favor of an expensive outer-loop rebuild on each jitter; also the TURN env pair silently drops the relay server when only one of username/credential is set, so symmetric-NAT peers just fail with no log explaining why

#1854

· topcheer opened

on Sep 7, 2026

[code-review] the Jujutsu backend decides cleanliness by substring-matching the prose line The working copy has no changes, but jj only started printing that wording in 0.26 (upstream changed it from is clean in Feb 2025): on jj 0.25 and older a perfectly clean repo never matches, IsClean returns false forever, and both agentruntime consumers treat every working tree as dirty - the same file already switched Status to the machine-readable diff --summary form in an earlier fix, IsClean just never followed

#1853

· topcheer opened

on Sep 7, 2026

[code-review] the matrix M_UNKNOWN_POS self-heal only clears the token on disk: the mautrix sync loop caches nextBatch in a local variable read once before the loop starts and never re-reads the store, and the custom OnFailedSync returns nil so the loop continues retrying with the exact same rejected token every two seconds - the adapter shows connected while receiving zero events, disk churns twice as often as the default syncer, and the wedge only clears on process restart, which is precisely the symptom the fix claimed to eliminate; the store file is also keyed by adapter name alone so two workspaces running the same-named adapter overwrite each others token and lose offline messages on restart

#1851

· topcheer opened

on Sep 7, 2026

[code-review] WrapTransport only injects the smart proxy chain when the caller passes a nil base transport: with a non-nil base it clones as-is, and since callers like the A2A mTLS client build their base without a Proxy field, the clone silently keeps nil and the client goes direct - the code comment right there claims proxy support is preserved, so in corporate proxy environments the mTLS peer connection simply cannot reach the network; Truncate also returns the input unchanged when the max is negative, and panel callers pass width arithmetic (w-10) that goes negative on narrow terminals, dumping untruncated text into fixed-width panels

#1849

· topcheer opened

on Sep 7, 2026

[code-review] the sse transport acceptance fix never normalized the transport value: the Start switch accepts sse alongside http but nothing anywhere assigns c.transport = http, so every downstream branch that compares == http misses - the first Initialize falls into the stdio default, writes to a stdin pipe that was never established, fails with stdin closed, and the bridge swallows the error exactly as before the fix; the GET notification stream never starts and oauth is never wired, while the migration layer still drops sse on the assumption the client does not support it - a pending test asserts only Start() returning nil, so CI stays green while end-to-end sse is broken

#1848

· topcheer opened

on Sep 7, 2026

[code-review] the Windows symlink fallback copy path drops the Close error, so a delayed write failure on a network redirector (SMB) at CloseHandle time is silently swallowed: io.Copy returns nil, the chmod succeeds, and the caller believes a link was created when it actually holds a truncated copy with no link semantics - classic defer Close antipattern, fixed by explicitly closing and propagating the error

#1845

· topcheer opened

on Sep 7, 2026

[code-review] the Windows shell fallback chain ends with a PATH lookup of bash.exe, which on any machine with WSL installed resolves to the System32 WSL entry point: it accepts -c but runs the command inside the Linux subsystem where drive-letter paths, the working directory, git, and npm all break or misbehave - and the spec is mislabeled git-bash; reachable on slim images lacking both PowerShell and Git Bash, fix is to reject the System32 path

#1842

· topcheer opened

on Sep 7, 2026

[business-logic] the race-verify hint is structurally starved: it registers as a default-severity checker at the tail of the registry under a single-warning budget, so any critical hit or any earlier-registered default checker in the same write crowds it out to a debug log - its real visibility is far below what the detector documents; its delta check is count-based, so the most dangerous race-introducing refactor (removing a mutex and adding a goroutine - net score zero) never triggers despite the header promising modified code is covered; meanwhile the verify-command word list counts make clean, gofmt, and bare make invocations as verification, letting a pure listing command erase a failed-test urgency flag, while missing bazel, gradle, mvn, dotnet, yarn, pnpm, and vitest test runners so a fully-verified agent gets scolded for not building

#1841

· topcheer opened

on Sep 7, 2026

[business-logic] the guidance budget marks a hint tag as seen before asking whether it may be delivered, so when the shared byte pool rejects the hint the tag is already occupied and every later attempt this turn - including critical-tagged ones, since the dedup check runs before the critical bypass - is silently dropped without the hint ever being delivered (the delivered-not-returned family fixed for quotas remains in the dedup slot); the byte pool itself is first-come-first-served across parallel tool results, so advisory hints in early results can crowd out a hardcoded-secret or git-destructive warning arriving in a later result; dedup losses are invisible to the suppression summary, and iteration-level injections bypass conflict arbitration entirely

#1840

· topcheer opened

on Sep 7, 2026

[code-review] the liveness probe relies on signal zero, which succeeds for a zombie - a process that exited but whose parent never reaped it - so IsProcessAlive returns true and the crash-recovery journal keeps treating the dead instance as alive, never cleaning its anchor; empirically confirmed with a real zombie, though the misjudgment direction is fail-safe (leaked anchor rather than wrongful deletion) and the window requires a hung parent

#1839

· topcheer opened

on Sep 7, 2026

[business-logic] required-parameter validation treats an explicitly empty string as missing, but for fields where empty is the semantics - edit_file new_text for deletion, write_file content for empty files, batch_replace replacement for removal - that classification rejects legal calls pre-dispatch: the schema lists these as required while the handlers deliberately omit them from their own checks, the strict gate runs first so the handler intent never executes, and the model is told to provide all required parameters - a fix that is impossible for a deletion, trapping it in the futile retry loop the presence-only validator was built to prevent (its own comment documents the decision the empty-string check silently overrode); that same ordering also makes the richer validator with schema-description errors unreachable dead code

#1838

· topcheer opened

on Sep 7, 2026

[code-review] the home directory helper reads HOME on every platform while the Go standard library deliberately ignores it on Windows, so a Git Bash or MSYS-style HOME like /c/Users/x flows through unchanged and joins into a drive-root-relative path - config and credentials land in the wrong place and split from the many call sites that use os.UserHomeDir directly in the same process, producing two different homes; separately the config-dir join has no empty-home guard the util layer does have, so with HOME unset in a container or service the join yields /.ggcode at the filesystem root and the loader tries to read and create it - silently succeeding when running as root

#1837

· topcheer opened

on Sep 7, 2026