Repository navigation
Triage the 300 T3 Code PRs into useful Solenta features and fixes #1868
Description
Expand the completed 100-PR assessment (#1867) to the fixed latest-300 merged snapshot from 2026-11-07. Include Claude usage and limits, compare promising items to Solenta main, and deliver a deduplicated complete useful-feature list with priority, upstream links, and confidence. Research only; no implementation.
Plan:
- Read the remaining 200 PR descriptions and group overlapping changes.
- Check candidate features against current Solenta code and existing issues.
- Publish the full useful-feature list, distinguishing missing capabilities, existing features needing verification, and irrelevant patches.
Activity
currentbits commentedon Nov 7, 2026
Useful T3 Code features for Solenta: 300-PR triage
Reviewed 2026-11-07 against Solenta origin/main
f46d2384(v0.29.0). Scope: the fixed 300 merged PRs previously fetched in this thread, newest 2026-11-07 08:45:10 UTC through oldest 2026-11-04 23:54:06 UTC. All 300 titles and description summaries were screened; promising candidates were compared with Solenta source. This expands #1867, grouping related patches by user benefit rather than treating every PR as a separate feature.34 candidate features/improvements, plus 12 validation groups. High/Medium are relative recommendations, not a committed roadmap. Optional/Later are product choices or dependencies. Validation groups are not confirmed bugs. This is research only; no implementation or feature issues were started.
Important corrections and existing coverage
electron/providerUsageManaged.jsreads Anthropic/api/oauth/usage;electron/runner-provider-claude.jsandproviderUsage.jsingest liverate_limit_eventwindows. CLIProxy would be an additional account source.src/components/UsageView.tsxsorts provider rows by the currently displayed metric. The previous 100-PR report left this as a later verification item; it is now ruled out as new work.src/uiPrefs.tshas Summary/Normal/Verbose. Only stronger hide-successful/final-only behavior is a potential addition./clearsettles the old thread and opens a new draft. Neither is the proposed fresh session inside the same visible thread.Candidate features and improvements
Claude and usage
1. Claude compaction-aware context meter (High) — Read the active Claude settings/launch override so the meter shows the actual auto-compaction threshold, not just the model ceiling. No autoCompactWindow/autoCompactThreshold handling was found in Solenta main. Sources: #29220.
2. Ultracode as a separate toggle from effort (High) — Solenta currently includes ultracode in CLAUDE_EFFORTS. Support Low/Medium/High plus Ultracode independently, and expose it only for supported CLI/model combinations. Sources: #27034, #30781.
3. CLIProxy account usage alongside native accounts (Optional) — Useful if the user uses CLIProxy: show its Claude session/weekly quotas and normalize management.html URLs. Solenta already has native Anthropic OAuth usage and live Claude rate-limit events; this would add a source, not introduce Claude usage tracking. Sources: #25579.
Sessions and automations
4. Fresh provider context within the same Solenta thread (High) — Keep the visible history and worktree but start a new provider conversation, with a clear session divider and no accidental old-context handoff. Solenta Restart agent session deliberately preserves the sessionId; /clear currently settles the thread and opens a new draft. Preserve that established /clear behavior unless explicitly changing the product contract. Sources: #30323, #30325.
5. Automation destination and session modes (High) — Choose new thread per run, continue a fixed thread, or reuse one visible thread with fresh provider context each run. Current runAutomation creates a new thread for each run. Fresh-context mode depends on the preceding capability. Sources: #30326.
6. Opt-in stop-on-settle for managed dev servers/actions (High) — Build on onSettle hooks and existing process-group stop primitives. Preserve output; do not stop a project-wide server still used by another active thread. Sources: #31857.
7. Weekend-friendly inactivity settlement (Optional) — Consider a seven-day default instead of three. Preserve explicit values, disabled settings, pins, open PRs and existing settlement exemptions. This is a product choice, not a confirmed defect. Sources: #32325.
Review and GitLab
8. Add any PR comment to agent context (High) — Let the user ask whether a reviewer is right or what a comment means, without automatically requesting a fix. Include source PR/comment links and append to the existing draft. Sources: #30579.
9. PR composer completions for users and references (Medium) — Complete @user, #issue/PR and GitLab !MR in PR/MR comments and descriptions. This is distinct from existing project/composer mentions. Sources: #30747, #30857.
10. Editable checklist items in PR descriptions (Medium) — Save checkbox changes through the existing description update API, preserving formatting and current server contents rather than overwriting unrelated edits. Sources: #30704.
11. GitLab request-changes review verdict (Medium) — Add the reviewer state and submission action, respecting host permissions. Solenta has basic GitLab MR support; this is richer review parity. Sources: #30549.
12. Complete GitLab discussion/finding coverage (Medium) — Include file-level, image and unresolved overview discussions; mark discussions outdated when their position no longer matches the MR diff refs. Keep them actionable as findings. Sources: #30425, #30397, #30423.
13. GitLab labels, reviewer search and accurate review filters (Medium) — Search eligible reviewers beyond page one, retain requested reviewers, edit labels, and make review/check filters traverse enough results. Solenta already lists individual GitLab pipeline jobs; do not reimplement that existing feature. Sources: #30483, #22489, #30433.
14. GitLab fork controls and post-merge cleanup (Medium) — Offer maintainer edits for eligible fork MRs, permission-aware source-branch deletion, and the same remove-agent-credits preference when merging. Each needs the source project/author permission checks. Sources: #30518, #30453, #30441.
15. GitLab reference navigation and watcher metadata (Medium) — Link issues/MRs/SHAs/users, comment timestamps and author profiles. Detect edited bot comments for watchers and use real bot metadata for filtering rather than name guesses. Sources: #30524, #30411, #30418, #30403, #30383.
16. Private and SSO GitHub PR media (Medium) — Carry the actual PR repository into attachment resolution, including fork checkouts. Resolve signed upload URLs in the correct repository/SSO context without forwarding bearer credentials to asset hosts. Sources: #21631, #21632.
17. Safe rich PR media and CRLF rendering (Medium) — Keep before/after video tables and code fences intact. Test CRLF descriptions and media-like URLs inside code; adopt only the rendering behavior that matches Solenta Markdown and sanitization. Sources: #25988, #30352.
18. Show the comparison base next to change totals (Medium) — Make a large or unexpected diff understandable with an explicit vs label. Verify the currently displayed Solenta view before adding a duplicate label; existing recorded-base handling should remain. Sources: #30085.
Speech
19. Multilingual local dictation (High) — Current local speech pins an English-only Nemotron model. Add a tested multilingual option and language/model compatibility checks; do not silently recommend an English-only download for another language. Sources: #32035.
20. Reuse installed speech models from selected folders (Medium) — Avoid duplicate large downloads. Discover compatible files read-only and verify their size/hash before using them; do not modify another app's model store. Sources: #30224.
21. Choose a linked computer for transcription (Medium) — Explicit per-device opt-in to a named authenticated host, with cancellation on revocation and no automatic destination switch. Useful for clients without a local runtime. Depends on an intentional product decision to transmit audio. Sources: #30625, #30623, #31059.
22. Dictate into review comments and browser annotations (Medium) — Reuse the composer speech service and consent rather than granting microphone access to an embedded webpage. Keep drafts full width and preserve the target editor/session identity. Sources: #30624, #32142.
23. Unicode-aware dictation insertion (High) — Confirmed helper gap: prefix+transcript+suffix yields Helloworld, 코드를검토 and Hello worldnext. Handle both boundaries, language-dependent spacing, selection replacement, and provisional/final transcript consistency. Sources: #31799.
24. A compact, testable speech-model chooser (Later) — Once multiple local models exist, show installed/downloading plus a short recommended set; offer language/downloaded filters and size/speed/accuracy sorting. Keep microphone/model tests beside their settings and use one transcript output. Sources: #30410, #31526, #32045.
Everyday UX
25. Desktop windows that fit narrow tiles (High) — Main minimum is 1100x700; detached-thread minimum is 920x700. Lower limits only with composer/settings/palette/approval layout validation at narrow and short sizes. Sources: #28223, #31826.
26. Host/owner/repo clone organization (Optional) — Offer ghq-style destinations to avoid api/docs/fork name collisions. Preserve chosen folders and make the option explicit rather than moving existing projects. Sources: #24904.
27. Correct transcript/terminal path links (High) — Confirmed elided-path candidate gap; keep shortened .../… paths plain. Also test Next/Svelte/TanStack/pnpm paths containing brackets, parentheses, $, + and @ against Solenta's actual link providers, preserving line/column suffixes. Sources: #30762, #30603.
28. Hide successful tools or show final answers only (Optional) — Extend existing Summary/Normal/Verbose transcript density if desired. Those modes already cover a large part of the benefit; a final-answer-only mode and hide-successful behavior are the additional choices, not a need to rebuild transcript folding. Sources: #30000.
Reliability and credentials
29. Visible settings-load failure and Retry (High) — Boot currently catches settings.get rejection to null. Distinguish unavailable settings from defaults and avoid saving edits against an unknown snapshot. Disk quarantine/backup recovery is a separate existing safeguard. Sources: #32248.
30. Locale-independent Git failure classification (High) — gitEnv inherits locale while some callers match English errors. Normalize Git diagnostic language without changing hooks' character locale; include SSH and WSL paths. Sources: #32252.
31. Atomic generated provider/MCP configuration (Medium) — Direct generated-file writes exist. Reproduce a concurrent external reader, then stage and rename the affected files while retaining permissions and intentional symlinks. A useful reliability technique, not a proven torn-read incident in Solenta. Sources: #32231.
32. Secret-manager references in credentials and request cards (Medium) — Extend existing encrypted secrets and thread secret_request cards to a provider such as 1Password. Resolve on use, make references inherently sensitive, show dependencies before provider removal, and provide explicit retry without repeatedly querying a locked vault on unrelated saves. This is reference integration, not a missing secret-card feature. Sources: #31670, #31652, #31651, #31403, #31226, #31654.
33. Better provider/tool failure explanations (Medium) — Show the actionable provider error, validation refusal or preview-network cause, bounded and redacted, rather than raw JSON or a generic retry-looking failure. Verify each actual Solenta path before changing it. Sources: #30331, #30295, #30564, #30318.
Documentation
34. Clear install, data-root and privacy guidance (Later) — Audit real Solenta Linux architecture/launch/update steps, custom worktree/data roots, retry commands and separate diagnostic-data controls. Use Solenta's behavior and paths, not T3's relay/mobile/AUR wording. Sources: #30819, #31000, #30942, #30689, #30480.
Validate before filing a defect or scheduling implementation
V1. Claude quota recovery as one coherent lifecycle. Solenta already has native quota readings, live updates, quota timers and boot re-arm. Test a single recovery notice, no duplicate continuation, restart-before-provider-start, and no paid child restart after a parent is put away. Attribute machine continuations clearly. No missing usage endpoint was established. Sources: #29292, #31292, #30363, #29838.
V2. Model-picker and actual-launch consistency. Solenta has its own signed catalog/CLI alignment path. Test that refreshed capability/effort selections reach the launched Claude process; do not import T3's manifest format. Sources: #31717.
V3. Rewound Codex history import. Solenta reads rollout files; test paginated history_base chains and new rollout filenames so imports retain history before a rewind. Sources: #30652.
V4. Checkpoint restoration and run identity. Test a turn creating .gitignore, Stop followed by another turn before capture completes, and CRLF/no-final-newline excerpts. Solenta restore currently uses reset --hard and differs from T3; do not copy repeated git clean or delete ignored files blindly. Sources: #29795, #30830, #30717.
V5. Branch/PR identity after external Git changes. Test an agent switching branches, merges updating divergence, local upstream remote=. and publish failure after remote creation. Reflect actual branch and partial success without linking/settling the wrong PR. Sources: #31599, #31342, #30563, #29688.
V6. PR responsiveness and isolation. Use already-held data and mutation replies where sound; prevent a slow host/background refresh from blocking unrelated PRs or foreground Git. Profile Solenta first; T3 timings and lock-striping code are not Solenta measurements. Sources: #27798, #25761, #30291, #30293, #30294, #32162.
V7. Preserve user drafts and characters. Repeated Add to agent should merge/dedupe selections without replacing typed text; test emoji-safe labels and path-input caret stability. Verify current paths instead of claiming all these defects exist. Sources: #24990, #30390, #30345, #30391, #30369, #30367, #31642.
V8. Worker results, pending questions and remote status. Keep pending cards answerable independently of bounded history, preserve sibling completions through interruption, open linked-computer worker prompts in the correct thread, and show remote-disconnected work honestly. Preserve explicit Stop and existing Solenta question protections. Sources: #28204, #26435, #26492, #29285, #31269, #29000, #29622, #31815.
V9. Slow-link and burst streaming resilience. Stress large configuration/history and repeated full tool snapshots under slow/reconnecting clients. Coalesce superseded updates before charging buffers and retain control state without dropping the visible history. Solenta's WebSocket protocol differs. Sources: #16317, #23758, #28487, #31266.
V10. Browser and process resource lifecycle. Measure inactive preview processes, tab/resource ownership, packaged Electron/Chromium version, shutdown phases and child exit completion, and per-platform process-monitor cost. Solenta uses Electron previews and its own teardown, not T3's headless browser pool. Never claim the upstream memory/CPU savings transfer. Sources: #30306, #24186, #30820, #22572, #31850, #31807, #30381, #19288.
V11. Simulator and native interaction resilience. Test retaining last-known devices with a stale indicator, popovers over drag regions, Safari iframe-owner scrollbars, and sidebar drag autoscroll. These have matching seams but no complete Solenta UI reproduction yet. Sources: #25971, #31794, #32398, #31924, #27691, #30346.
V12. Fail closed on an unreadable selected credential. Trace existing encrypted locked-key handling through forge/provider fallback. A locked selected credential should not silently authenticate as a different account. SecretPrompt already uses type=password; do not re-add that fix. Apply reference cleanup/retry semantics if secret providers are introduced. Sources: #31646, #31651, #31403, #31226, #31653.
Exclusions and dependencies
Native iOS/Android font, keyboard, Liquid Glass, Nitro/lexer, notification and view-prop patches do not transfer to Solenta's Electron/React client. Effect tracing/schema/lint refactors, T3 SQLite projection/WAL patches, Pi/fx-specific fixes, and T3's pnpm/Blacksmith/Android release optimizations have no matching implementation. Basic hygiene ideas can inform future work but are not independent product features.
Speech catalog sorting depends on a real multi-model catalog; remote speech depends on an explicit audio destination/consent design; fresh-session automation depends on a genuine provider-context boundary. GitLab additions are optional investments in broader forge parity, with per-action permissions and compatibility checks.
No newly added direct Claude OAuth usage-endpoint change was found in this 300-PR set. Do not treat the absence of an upstream PR as proof that Solenta's current endpoint always works. There was no live-account request or paid provider turn in this research.
Source evidence in Solenta
Verification limits
Source review and prior read-only helper probes establish the current dictation boundary/path-candidate findings and the source-level settings/window/Git-locale seams. New V-items are local test proposals, not reproduced failures. No app files were changed; the older working checkout and unrelated edits were untouched. No full test suite was run for this research task. Upstream performance numbers are not Solenta measurements.
currentbits commentedon Nov 7, 2026
Completed the fixed 300-PR triage against origin/main f46d238. Published 34 deduplicated candidate features/improvements plus 12 validation groups with upstream links, source evidence, priorities and existing coverage: #1868 (comment). Native Claude usage, metric sorting, existing transcript density, secret cards and other shipped capabilities are explicitly excluded as new features. No application code changed.