ci: validate batch preset skill toggles - #1
Draft
vipinchan wants to merge 19 commits into
Draft
Conversation
…rge incident) A device that had only ever pushed hit a permanent 'object merge aborted (zero changes)' the first time it actually merged. Three compounding causes, found on a real library: - A committed atomic-write leftover (x.json.tmp.<uuid>) under the metadata namespace: the strict validator rejects any non-json file there. Temp/OS junk inside the managed metadata subdirs is now dropped from the merged tree at plan stage, and every commit path deletes such leftovers from the working tree first (a push-only machine never runs the reconcile cleanup). - Unclaimed skill dirs from old app versions in the shared history: rule 4 now grandfathers dirs that were already unclaimed in either merge input — the invariant defends against merges introducing orphans, not against pre-existing ones syncing through. - The failure card only showed the outermost error context: git-classified command errors now carry the full anyhow chain, so the actual validation rule appears in the UI.
"Back Up Now" with "local changes: 0 · remote updates: 1" read as a push that might overwrite the remote — the opposite of what it does. The pending state now names its three situations: remote-only updates get their own title, description and a "Sync Now" button, and both mixed-state descriptions say explicitly that syncing merges per skill and neither side overwrites the other.
…dex review) Two findings from a codex review of the incident fixes, both accepted: - The old-client single-parent tip validation tolerated unclaimed dirs but not committed .tmp. metadata junk, so a legacy tip carrying one still hard-blocked before the plan-stage drop could run. Input-tip validation (validate_input_tip) now skips metadata-namespace junk; the merged tree stays strict — the planner guarantees junk-free output there, so junk would mean an engine bug. - The legacy-dirt integration test committed its tmp file through the app path, which now cleans temp files first — the 'junk already in history' topology was silently uncovered. The test commits via raw git (as the incident device's history actually looks), asserts the junk is really in the remote history, and was mutation-verified: disabling the plan-stage drop makes it fail. The old-client test now also carries committed junk to pin the tip-validation tolerance.
… watcher self-write mute (xingkongliang#248) (xingkongliang#285) * fix(startup): stop the ~8s stranded-target backfill from blocking launch (xingkongliang#248) `backfill_stranded_agent_targets` ran a full multi-agent filesystem scan + per-skill hashing synchronously inside Tauri `setup()`, freezing the window. Measured ~7.9s on every launch: its `has_candidate` fast-path never went false because un-repairable stranded skills (diverged locals we intentionally skip) keep a `source_ref` with no target row forever, so the expensive scan re-ran indefinitely. - Move the backfill into a background `spawn_blocking` after the window shows, mirroring the existing credential-migration pattern. `setup()` body: ~8000ms -> ~15-25ms. - Gate the scan on a signature of the stranded candidate set. An unchanged set skips the scan entirely; a newly-stranded skill re-arms it. Recorded post-repair so it converges in one launch. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * test(git_backup): persist git identity in pre-protocol restore test The test built its repo with transient `-c user.email/user.name` flags that only apply to its own git closure. The production commit path (commit_all_unlocked via plain run_git) then found no identity on CI runners lacking a global git config, panicking with "Author identity unknown". Persist the identity into repo-local config after init so the production path has an author, matching the convention in the other commit_all_unlocked tests. --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Render a hand-drawn (star-history.com style) star-growth SVG from the owner-authenticated gh stargazer timeline and commit it as a static asset. GitHub restricted the stargazer-history API to repo owners in 2026-06, so the live star-history.com embed no longer works for this repo; generating locally with the owner token is now the only reliable path. Auto-refreshed during release:prepare (best-effort, never blocks). Handwritten look uses the bundled Patrick Hand font (OFL), embedded into the SVG so it renders on GitHub without external fonts.
- Backfill gate: fold available-adapter set, DB content hash, and local/ central mtime fingerprints into the stranded-candidate signature so the scan re-arms when repairability changes (tool newly installed/enabled, diverged local restored, center edited) — not only when the set changes. - Backfill race: re-hash both sides of a skill immediately before the sync write; the scan snapshot can be seconds old and the backfill now runs concurrently with the user, so a fresh local edit must never be clobbered by the central copy. - Watcher mute: defer (not drop) relevant events that arrive inside the 1.2s self-write mute window and flush once it closes, so an external edit during the window can no longer leave the UI stale.
- Path-scoped watcher mute: record the dirs the app actually writes and swallow only THEIR echoes; foreign paths seen inside the window defer. Blanket-deferring everything re-introduced one redundant refresh per self-write batch — the exact waste xingkongliang#248 removed. - Backfill gate fingerprint: replace max-mtime+entry-count (churned by .DS_Store-style junk, blind to same-mtime restores) with the same content hash the repair itself compares, plus an existence marker. - Backfill pre-write check: explicit existence checks and fail-closed hash comparison (hash_directory treats missing dirs as empty and its errors were compared as equal None). - Tests: three-state mute verdict table, path-scoped classification, diverged-then-restored re-arm scenario (replaces the mtime-bump test, which the content fingerprint correctly no longer trips on).
Wrap the README star chart in a link to the tool that generates it (github.com/xingkongliang/star-history-svg) and correct the in-image watermark from "star-history.com" (which did not produce this chart) to "star-history-svg".
…ling back to container (xingkongliang#280) Fixes xingkongliang#278. find_skill_dir() now bail!s when a specific skill_id is requested but matches nothing, instead of falling through to the skills/ container or root — which had installed the entire repo as one skill. Fallback preserved only for skill_id == None (enumeration flows). Verified: 287 Rust tests pass; codex read-only review confirmed no caller regressions.
Follow-up to xingkongliang#280 (codex review): lock in the legitimate case where a requested skill_id matches the root SKILL.md's frontmatter `name`, so the xingkongliang#278 bail! never regresses single-skill repos whose id resolves via the recursive name match rather than a directory basename.
Per codex review of 60564f7: the positive-only assertion could pass for the wrong reason — a root SKILL.md also satisfies the skill_id==None root fallback, so the test wouldn't fail if the old xingkongliang#278 fall-through were reintroduced. Add a sibling assertion that a non-matching id on the same root-skill repo errors, proving the root is returned via name_match, not fallback.
…ingkongliang#241) Replaces the two native <select> dropdowns for auto-update (check interval, auto-apply) with the segmented-control button pattern already used elsewhere on the Settings page (theme, sync mode, tray, etc.). Wrapper and active/inactive classes match the existing controls verbatim; adds aria-pressed. Values and change handlers unchanged — behavior identical. Verified: npm run build (tsc -b + vite build) passes.
…ly files (xingkongliang#252) On Windows, relocating the central library to another drive could leave read-only Git pack files at the destination after an incomplete move. The next launch tried to fs::copy over them, hit "access denied (os error 5)", and the error bubbled to the .expect in run() — which executes before the window, logger, and panic hook exist. The process died silently: no UI, no log, no panic file. The app simply wouldn't open. Root cause is overwriting an existing read-only file. Fix: - migrate_repo_if_needed only ever moves into an empty/absent target. A non-empty target (real data or failed-attempt debris, indistinguishable) is never blind-merged — the app falls back to the intact source instead, so it can't overwrite newer data with older, and a fresh target makes the copy structurally incapable of overwriting a read-only pack. - The function is now infallible (returns MigrationOutcome, never Err), so a migration failure can no longer panic through run()'s pre-window .expect. On failure it runs the session against the source via a runtime base-dir override and shows an amber `migration_incomplete` banner (en/zh/zh-TW). - Detailed errors are deferred via record_startup_error and flushed in setup() once the logger exists (the pre-logger log::error! was a no-op). - Migration is skipped when a CLI base override (--skills-root/--path) is active, so the shared pending marker can't hijack an explicit root. Hardened after a grok review pass: - Compare source/target canonically (fs::canonicalize), so a cosmetic path difference (case-insensitive volume, 8.3, symlink) isn't taken for a real move and looped forever on `migration_incomplete`. - Reworded the fallback banner so it never tells the user to empty the path shown under "Central Repo Path" — under the override that chip shows the live library (get_central_repo_path returns base_dir()), so the old wording risked pointing a user at their own data. +6 Rust tests. Full suite 393 passing; npm run build clean.
… paths - backup: refresh presets after clone/reclone/sync/github-connect/restore and first-run restore so the sidebar preset list isn't empty until an app restart — metadata reindex already restored them to the DB (xingkongliang#302) - workspace: overview cards count each agent's real on-disk skills (not just the managed library) so an agent whose skills were installed outside Skills Manager no longer shows 0, matching the per-agent detail badge (xingkongliang#287) - icons: invert monochrome-black agent icons (codex, roo_code) under the dark theme so they aren't invisible on dark backgrounds; augment excluded because it already self-adapts via prefers-color-scheme (xingkongliang#279, xingkongliang#304) - windows: embed an app manifest declaring longPathAware so installing skills with deeply nested paths no longer fails with 'path too long' (xingkongliang#298, xingkongliang#299) codex(gpt-5.5)-reviewed; follow-ups applied: conflict-resolution and background auto-backup completion also refresh presets/skills; overview counts depend on the managedSkills ref and rebuild fresh so failed scans fall back to managed.
- skill_store: 新增 batch_add_skills_to_scenario / batch_remove_skills_from_scenario， 单事务内完成批量 INSERT/DELETE - presets: 新增 batch_toggle_preset_skills Tauri 命令，metadata 写 1 次 + tray 刷 1 次 - lib.rs: 注册 batch_toggle_preset_skills - tauri.ts: 新增 batchTogglePresetSkills API - MySkills.tsx: handleBatchTogglePreset 从 N 次独立 IPC 改为 1 次批量调用 - N 选 30 时: 30 次 IPC + metadata 写入 + tray 重建 → 1 次
vipinchan
force-pushed
the
perf/batch-preset-skill-toggle
branch
from
July 26, 2026 16:58
1a719cd to
4a24f89
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Temporary draft PR used to run the fork's full CI matrix for the branch submitted upstream in xingkongliang#336. This PR will be closed after CI validation.