You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Triage 100 recent T3 Code merged PRs for Solenta #1867
Review the fixed set of 100 T3 Code PRs listed in this thread on 2026-11-07 (32398 through 30719 by merge time). Compare their problem statements to current Solenta main and existing issues. Deliver a ranked shortlist plus a disposition for every PR. Research and triage only; no feature implementation.
Plan:
Read upstream PR descriptions and changed-file summaries.
Compare relevant areas with Solenta main and open issues.
Publish a complete disposition table and prioritized recommendations with evidence.
Reviewed 2026-11-07. This is the fixed set of 100 PRs previously listed in thread a66113ea, not a newly refreshed rolling window. Compared upstream PR descriptions and changed-file lists with Solenta origin/main at f46d2384 (v0.29.0). Open Solenta issues were checked for obvious duplicates. The user's older working checkout and unrelated edits were left untouched; source review used a temporary archive of main.
Recommended order
Normalize Git diagnostic language (#32252). A small reliability fix with a direct code match: gitEnv inherits locale and some failure classifiers expect English. Include SSH/WSL rather than fixing only local Git.
Make settings-read failures visible and retryable (#32248). Boot currently converts a rejected settings read to null. Distinguish unavailable settings from defaults; keep a recoverable screen/banner and avoid saving edits over an unknown snapshot. Existing disk quarantine/backup recovery remains relevant but is not the same failure.
Fix dictation word boundaries (#31799). Reproduced in the current pure helper for English and Korean. Address both insertion boundaries and streaming/final transcript consistency.
Support narrow desktop tiling (#28223). Main is capped at a minimum of 1100x700. This is a useful product improvement, with responsive work required in composer, settings, palette and detached threads.
Reject elided file-link candidates (#30762). Reproduced in the path parser. Cheap correction to misleading links for /tmp/.../file, trailing ..., and Unicode ellipsis.
Opt-in cleanup of owned dev servers on settle (#31857). Build on existing onSettle hooks and devserver stop primitives. This is an enhancement, not a claim that Solenta has T3's exact lifecycle bug. Shared/project-wide servers need explicit ownership rules.
Investigate before scheduling a fix
Atomic generated config writes (#32231): direct writes exist, but prove a concurrent external reader and constrain the fix to generated files. Avoid a blind write-everything abstraction.
Browser runtime currency (#30820): inspect the packaged Electron/Chromium version and test an Electron update. T3's standalone Chrome version/hash bump is not portable.
Resume/settle/checkpoint races (#31292, #31815, #30830): borrow crash/interleaving tests; Solenta already has quota and restart recovery, so a title alone does not establish a missing feature.
Foreground Git responsiveness (#32162): profile startup PR refresh and retention on a large history before adding concurrency machinery.
Safari scrollbars and titlebar hit testing (#32398, #31794): Solenta has the relevant CSS seams; require a native/Safari reproduction.
Simulator discovery (#25971): evaluate preserving last-known devices with a clear incomplete/stale state.
Later product choices
The most useful later addition is multilingual local speech (#32035): the current local model is English-only. Seven-day auto-settle (#32325) is a reasonable weekend-friendly default proposal, not a bug fix. Model sorting (#32045) can wait until more than one local model exists. Cloud transcription itself is already shipped through OpenAI/ElevenLabs (#31769 overlap).
Evidence and limits
Two read-only Node probes ran the current helpers, without changing product code:
These are helper reproductions, not full UI/end-to-end tests. Git locale and settings fallback findings are source-confirmed seams; localized-Git and IPC-failure fault injection were not run. Window minimum is source-confirmed; no new narrow layout was implemented. Other Audit entries are test ideas, not confirmed defects. No performance saving from T3 is presented as a Solenta measurement. No application code was changed and no full suite was needed for this research task.
Adopt = worth doing with a direct local gap or product benefit. Audit = useful hypothesis requiring a local reproduction or measurement. Later = optional product/polish work. Covered = the core idea already exists or its distinctive bug path is avoided. Skip = different stack/provider/client or no matching feature. Covered does not mean every edge case has been proven.
Global WebKit scrollbar rules exist in src/index.css. Excluding iframe owners is a plausible Safari fix; reproduce with Solenta embedded content first.
Adopt atomic staging-and-rename for generated config where external CLIs can read concurrently. Direct writes exist in orchServer.js, grok.js, cursor.js and kimi.js; reproduce the shared-reader window and preserve permissions/symlink semantics.
gitEnv in electron/worktrees-git.js inherits locale while services-git.js and worktrees-gc.js match English errors. Normalize Git diagnostic language while preserving other locale settings; cover local, SSH and WSL execution.
Useful Windows fixture discipline, but the three changed tests and Antigravity installer do not exist here. Apply when a corresponding fixture fails; do not add speculative skips.
Solenta also defaults inactivity settlement to 3 days. Consider 7 as a product choice to survive weekends; preserve explicit choices and pinned/open-PR exemptions. Not a correctness fix.
Meaningful speech expansion: Solenta local speech pins an English-only Nemotron model. Offer a tested multilingual model and explicit language support before adding a catalog or benchmarks.
Composer SpeechControls already renders a microphone and inline download confirmation rather than hiding setup. Recheck browser-specific availability separately.
electron/speech.js already supports opt-in OpenAI and ElevenLabs transcription with local default. Grok/OpenRouter are optional expansion, not a missing cloud-transcription feature.
Preserve the configured data root across headless-server install/update operations. T3 service-install code is absent, so reproduce against Solenta headless deployment before claiming the same bug.
src/useCoder.ts boot catches settings.get failures to null. Surface failure and retry; distinguish unread settings from defaults and prevent editing over an unknown snapshot. Store corruption recovery is a separate existing mechanism.
Solenta has Electron drag regions. Test menus/popovers overlapping those regions with native pointer input; add no-drag only where the failure reproduces.
Solenta already persists quota-wait and re-arms it at boot. Borrow the crash-window test: restart after a resume is claimed but before provider start, and prove no lost or duplicate continuation. T3 command-receipt patch does not transfer directly.
Protect foreground Git work from background PR refresh/retention load. Solenta does not share T3's eight-slot pool; profile real startup contention before adding a second scheduler.
Solenta already has canonical URLs, sitemap, robots sitemap reference and SoftwareApplication/Organization sameAs. Remaining site SEO work is tracked by #1837; avoid a duplicate broad SEO task.
Use upload-artifact compression-level: 0 for already-compressed installers when hosted release uploads are added. Current verification evidence upload is not the same workload.
Solenta now has remote workers. Verify the crew view groups local and linked-computer workers by status consistently; do not assume it shares T3's lineage bug.
Simulator discoverRaw throws on listing failure and no last-known list is retained there. Consider cached devices plus a stale/incomplete indicator; test current UI behavior before calling it a shipped UI bug.
Verify todo/status updates remain visually quiet and only executing tools appear live. T3 timeline classifier is different; no confirmed Solenta shimmer defect.
Add explicit opt-in stop-on-settle ownership to managed dev servers/actions. Solenta already has onSettle hooks and devserver stop primitives, but not this per-action lifecycle policy. Preserve output and do not kill a project server still used by another thread.
electron/main.js sets main-window minimum 1100x700 and detached thread minimum 920x700. Allow narrower tiling, then validate composer/settings/palette at narrow and short sizes. Do not lower constants without responsive checks.
Add coarse shutdown phase durations around runAppCleanup and final store flush if quit latency needs investigation. Prefer a few measurements to per-thread tracing.
Borrow a regression where a later babysitting instruction must not become a generated PR description; verify independent worktree bases. Solenta createPr already uses recorded/default base and accepts explicit title/body, so investigate the generation path rather than copying T3 prompts.
useComposerSpeech starts capture before api.start and buffers pending PCM, already following the early-capture principle. Native iPhone model setup is not present.
Pure helper reproduction: applySpeechTranscript concatenates prefix+transcript+suffix, yielding Helloworld and 코드를검토. Add language-aware Unicode boundaries on both sides, preserving deliberate whitespace and selection behavior.
Useful settings design rule: let descriptions use available width and controls size to content. Apply during narrow-window work, with Solenta-specific screenshots.
Test settle requests racing with machine follow-up turns. Solenta currently refuses setSettled while working and clears settled overrides on activity; T3 deferred-settle contract must be evaluated before adding persisted intent.
Sort Usage provider summaries by the metric currently displayed if declaration order is still shown. Keep chart colors/series stable. Needs a focused Solenta UI check.
Solenta store lazy hydration makes shutdown read scope worth profiling. Borrow candidate-first recovery and bounded concurrency only if quit/restart measurements show unnecessary history loads.
Borrow Unicode truncation cases for generated/subagent titles. Verify code-point-safe truncation in Solenta; T3 byte-sanitizer implementation is different.
Verify Linux install docs explain architecture selection, executable permissions and updates for actual Solenta downloads. Do not add T3 AUR instructions.
Test a stopped run's asynchronous checkpoint completing after the next turn starts. Preserve captured run identity/cwd/time; T3 shared checkpoint-scope patch is not a direct port.
Security-relevant principle: keep the embedded browser runtime current. Solenta uses Electron (manifest still ^35.2.0), not T3's separately pinned headless Chrome; audit packaged Electron/Chromium and choose a tested Electron update, not T3's download hash patch.
Pure parser reproduction accepts /tmp/.../report.md; a trailing ellipsis is stripped to the parent path, and Unicode ellipsis splits into separate path candidates. Leave elided paths as plain text before resolving links; retain valid ../ and :line forms.
Useful diff correctness cases: CRLF and final-newline-only changes. Solenta diffView already preserves displayed no-newline markers, but review excerpts join lines; test the actual selection/handoff path before changing it.
Borrow quoted ~/path project-resolution cases wherever Solenta accepts project paths through CLI/MCP/deep links. No exact T3 thread CLI exists here; do not claim its fallback bug.
Solenta acp-server uses its own newline buffer and MCP SDK transport rather than the T3 readline bridge. No production readline MCP framer was found (readline hit is a smoke fixture).
Completed triage of all 100 PRs against Solenta origin/main f46d238 (v0.29.0), with upstream descriptions/file lists and an open-issue duplicate check. Published full assessment in #1867 (comment). Dispositions: 6 Adopt, 21 Audit, 17 Later, 9 Covered, 47 Skip. Read-only helper probes reproduced dictation boundary and elided-path candidate problems. Research task complete; recommendations are not implemented.
Activity
currentbits commentedon Nov 7, 2026
T3 Code: 100-PR triage for Solenta
Reviewed 2026-11-07. This is the fixed set of 100 PRs previously listed in thread a66113ea, not a newly refreshed rolling window. Compared upstream PR descriptions and changed-file lists with Solenta origin/main at
f46d2384(v0.29.0). Open Solenta issues were checked for obvious duplicates. The user's older working checkout and unrelated edits were left untouched; source review used a temporary archive of main.Recommended order
gitEnvinherits locale and some failure classifiers expect English. Include SSH/WSL rather than fixing only local Git./tmp/.../file, trailing..., and Unicode ellipsis.Investigate before scheduling a fix
Later product choices
The most useful later addition is multilingual local speech (#32035): the current local model is English-only. Seven-day auto-settle (#32325) is a reasonable weekend-friendly default proposal, not a bug fix. Model sorting (#32045) can wait until more than one local model exists. Cloud transcription itself is already shipped through OpenAI/ElevenLabs (#31769 overlap).
Evidence and limits
Two read-only Node probes ran the current helpers, without changing product code:
These are helper reproductions, not full UI/end-to-end tests. Git locale and settings fallback findings are source-confirmed seams; localized-Git and IPC-failure fault injection were not run. Window minimum is source-confirmed; no new narrow layout was implemented. Other Audit entries are test ideas, not confirmed defects. No performance saving from T3 is presented as a Solenta measurement. No application code was changed and no full suite was needed for this research task.
Relevant Solenta sources at the reviewed commit:
Complete disposition of all 100 PRs
Adopt: 6, Audit: 21, Later: 17, Covered: 9, Skip: 47.
Adopt = worth doing with a direct local gap or product benefit. Audit = useful hypothesis requiring a local reproduction or measurement. Later = optional product/polish work. Covered = the core idea already exists or its distinctive bug path is avoided. Skip = different stack/provider/client or no matching feature. Covered does not mean every edge case has been proven.
currentbits commentedon Nov 7, 2026
Completed triage of all 100 PRs against Solenta origin/main f46d238 (v0.29.0), with upstream descriptions/file lists and an open-issue duplicate check. Published full assessment in #1867 (comment). Dispositions: 6 Adopt, 21 Audit, 17 Later, 9 Covered, 47 Skip. Read-only helper probes reproduced dictation boundary and elided-path candidate problems. Research task complete; recommendations are not implemented.