a1143aa5251dbe2a02ec638252d779d4cd85f841
441 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 João Salavisa
|
a1143aa525 |
fix(acp): handle claude-acp async_launched subagents end-to-end (#1244)
* fix(acp): handle claude-acp async_launched subagents end-to-end Claude Agent SDK's Task tool with run_in_background=true returns session/prompt with stop_reason=end_turn while subagents stay alive in the background, reporting `_meta.claudeCode.toolResponse.status = "async_launched"` instead of a terminal top-level status. kandev previously treated stop_reason as authoritative, cancelled the subagent card on prompt return, and fired an "agent finished without producing any output" notice — even though the dispatch did produce work and the subagent's child events would arrive seconds later. Backend: - `SubagentTaskPayload` gains `IsAsync` / `OutputFile` / `CanReadOutputFile` so the async-launched envelope is preserved. - `convertToolCallResultUpdate` recognizes async_launched as terminal for the Task tool itself (dispatch is the completion). - `cancelActiveToolCalls` preserves subagent_task tool calls through the prompt-return sweep so a late terminal tool_call_update from the SDK still lands correctly (covers the in-turn-complete case). Frontend: - `subagentMetaChips` emits a "background" chip when `is_async` or `status == "async_launched"` so the card is visually distinct. - `computeEmptyTurnNotice` suppresses the notice when any subagent task message exists for the turn (the dispatch IS output). Reproduced via `acpdbg prompt` against `claude-acp` 0.31.1; captured frames in acp-debug/claude-acp-prompt-20260602-110119.jsonl drive the backend test fixtures verbatim. * fix(acp): harden async_launched override against future SDK frames Address PR review feedback (greptile #1244, claude review): - Move the async_launched override inside the mutex lock so it can inspect the stored payload kind before flipping the status. - Drop the `status == ""` precondition. The earlier in_progress synthesis (Title/RawInput/Content/Locations) was implicitly trumping the override; a future SDK frame that combines async_launched with any of those fields would have left the card stuck in_progress forever. The override is now unconditional for async_launched subagent_task frames. - Gate the override on `payload.Kind() == ToolKindSubagentTask`. A hypothetical future claude-acp tool that reuses the `async_launched` literal in its own `_meta.claudeCode.toolResponse.status` would have been incorrectly marked complete by the prior unscoped check. Pinned by two new tests: - TestConvertToolCallResultUpdate_AsyncLaunched_OverridesInProgress - TestConvertToolCallResultUpdate_AsyncLaunched_NonSubagentNotTerminated --------- Co-authored-by: Kandev Agent <agent@kandev.dev> |
||
|
José Almeida
|
b0b968cb22 |
feat(web): add reverse direction to recent task switcher (#1241)
* feat(web): add reverse direction to recent task switcher Cmd/Ctrl+Shift+Space now cycles the recent task switcher backward (most-recent toward oldest), complementing the forward Cmd/Ctrl+Space. Both directions are user-configurable via Settings > Keyboard Shortcuts. - new TASK_SWITCHER_REVERSE shortcut; Shift keeps it mutually exclusive with the forward binding so both can be held together - thread a cycle direction through openSwitcher/cycleSwitcher; reverse opens on the oldest non-current task - commit on release of either binding's hold modifier - switcher footer shows both Next / Previous bindings * fix(web): commit task switcher only on the active binding's modifier Address PR review on the reverse task switcher: - track the binding that drove the switcher and commit only on ITS hold modifier release, so divergent custom forward/reverse bindings can't commit early (greptile/coderabbit) - move the CycleDirection type below the import block - correct the reverse-shortcut direction comment (oldest -> most-recent) - assert exact task titles in the reverse e2e precondition * test(web): assert task switcher stays open on Shift-release Make explicit that releasing the secondary modifier alone does not commit the reverse switcher — only the hold modifier does. |
||
|
João Salavisa
|
cb7b5429d5 |
fix(tasks): harden multi-branch add_branch and surface it in UI (#1239)
Three areas were broken in the MCP add_branch_to_task flow: 1. Predictable branch names. autoBranchName counted up from existing branch-N rows. Concurrent add_branch calls raced onto the same name and the worktree directory leaked ordering. Now uses a random branch-<3 random chars> suffix via worktree.SmallSuffix. 2. Sibling materialize did not flow to the UI. The materializer published AgentctlReady with the NEW sibling worktree's path, which the frontend handler wrote onto the session row — clobbering the primary's worktree_id/branch and pointing the file browser at the sibling instead of the task root. Added TaskWorkspacePath to the payload, detected the sibling case on the frontend (existing worktree_id != payload's), kept the primary's id/branch, repointed worktree_path to the task root, bumped commits refetch, and dropped the pre-multi-repo legacy git-status entry so the Changes panel stops surfacing the frozen single-repo snapshot. File browser's resetKey now includes the worktree count so the tree reloads when the sibling lands. 3. workspace_path promotion + agentctl rescan refused to act on historical envs. A few envs had workspace_path pointing at the source repo's local_path (orchestrator computeWorkspacePath fallback to req.RepositoryPath when resp.WorktreePath is empty) instead of the primary worktree. The materializer's promotion gate compared against env.WorkspacePath's parent, so those envs were never promoted, and agentctl's resolveRescanPath only accepted parent-of-current so the rescan to the real task root was rejected. Trackers stayed on the wrong directory and the sibling's commits never appeared in the Commits panel even after a full reload. Anchored promotion on env.TaskDirName (basename cross-check) and added a recovery branch to resolveRescanPath that accepts any absolute directory which actually contains git-repo subdirs — the bearer-token middleware already gates the endpoint and scanRepositorySubdirs validates the children, so the relaxed check stays safe. Co-authored-by: Kandev Agent <agent@kandev.dev> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> |
||
|
João Salavisa
|
140e828ead |
fix(web): subdue voice input button to ghost style (#1233)
* fix(web): subdue voice input button to ghost style The voice button was styled as a primary-filled circle to mirror SubmitButton, but in the task-create dialog there is no adjacent Submit (the submit button lives in the footer), so the loud styling read as the dialog's primary action rather than as a secondary input control. In the chat composer it also competed visually with Submit where Submit should win. Switch to `variant="ghost"` with muted-foreground icon so voice sits in the toolbar alongside attach/enhance as one secondary action among several. Recording state keeps the pulsing destructive ring for unmistakable visual feedback but uses a tinted `bg-destructive/15` instead of a full destructive fill. Unsupported fallback also moves from `secondary` to `ghost` so the greyed-out variant matches. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * refactor(web): couple voice mic ring radius to button corner radius Use `rounded-[inherit]` on the recording pulse-ring so the corner radius follows the underlying `Button`'s actual radius instead of being hardcoded to `rounded-md`. The two happen to match today (shadcn's ghost Button defaults to `rounded-md`), but the implicit coupling would silently misalign if the Button base styles ever change. Flagged in PR review by both Claude and Greptile. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Kandev Agent <agent@kandev.dev> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> |
||
|
João Salavisa
|
fd79d6170f |
feat(tasks): multi-branch tasks — N (repo, branch) pairs per task (#1226)
* feat(tasks): multi-branch tasks — N (repo, branch) pairs per task Extends multi-repo into multi-branch so a single task can hold multiple worktrees on the same repo (one per branch). Sibling worktrees layout at <task>/<repo>/ + <task>/<repo>-<slug>/, dedicated PRWatch per branch, live agentctl tracker rescan + frontend store updates so the UI splits PRs / files / commits per branch without a session restart. MCP add_branch_to_task_kandev tool is shielded to the worktree executor. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(web): unify multi-branch repo label as <repo> · <branch> across panels PR Changes and Commits panels rendered repo labels differently for multi-branch tasks — Commits showed "kandev · feature/x" while PR Changes leaked the raw "kandev-feature-x" agentctl tracker tag. Reformat tracker-tagged repo names via formatRepoLabel using the longest known-repo prefix so kandev-cli-* doesn't collapse to kandev · cli-*. * fix(tasks): address multi-branch PR review — data races, slug collisions, MCP error codes Backend - agentctl process.Manager: add repoTrackersMu RWMutex guarding repoTrackers + workspaceTracker. Reads snapshot under RLock; RescanRepositories builds trackers first, takes Lock for swap/append. Validates newWorkDir via os.Stat before publishing cfg.WorkDir. - worktree.Manager: reject invalid explicit BranchSlug up-front; cache key uses sanitized wt.BranchSlug so read/write paths are symmetric. - worktree.store: narrow isDuplicateColumnError to "duplicate column name". - task service AddBranchToTask: slug-collision guard rejects feat/a vs feat-a. - orchestrator resolvePushRepo: two-pass exact-name match before subdir prefix fallback so "backend" doesn't swallow pushes for "backend-admin". - executor buildRepoSpecs: duplicate repos always get a non-empty BranchSlug (falls back to branch-N when both branch fields sanitize to empty). - MCP handler: classifyAddBranchError maps service errors to NOT_FOUND / CONFLICT / VALIDATION_ERROR; add ErrorCodeConflict constant. - MCP server: extract registerAddBranchToTaskTool, register in task mode only (external mode contract: config tools + create_task_kandev). - workspace_rescan handler: reject malformed JSON with 400 instead of silently rescanning the current root. - branch_materializer: split MaterializeBranch into prepare + finalize helpers to satisfy the 80-line cap. Frontend - existingPrUrl returns undefined when a task has >1 PR so per-repo View PR buttons take over (no silent linking to a sibling). - formatRepoLabel checks exact known-repo match before prefix splitting so "kandev-cli" isn't mislabeled as "kandev · cli". - Repo chip key now includes base_branch so siblings differing only on base render as distinct chips. Docs - multi-branch spec uniqueness wording aligned with the canonical 4-column key (task_id, repository_id, base_branch, checkout_branch). - worktree.CreateRequest.BranchSlug doc describes the sibling layout (<repo>-<slug>) instead of the obsolete nested form. * fix(tasks): scope slug-collision check to repo only + use errors.Is for ErrTaskNotFound - service_branches: drop base_branch from the slug-collision check. Worktree path is (repo, slug); two rows on the same repo cut from different bases still collide on disk. - service_branches: prepareBranchAdd wraps taskrepo.ErrTaskNotFound via fmt.Errorf with %w so classifiers don't have to substring-match the formatted task UUID. - mcp handlers: classifyAddBranchError uses errors.Is(err, ErrTaskNotFound). * fix(agentctl): sanitize rescan work_dir before os.Stat to clear CodeQL path-injection CodeQL flagged the rescan endpoint as a path-injection sink (the HTTP-supplied work_dir flowed straight into os.Stat and downstream scanRepositorySubdirs). Add validateRescanPath which: - requires the input to be a clean absolute path with no ".." segments - rejects anything that isn't equal to or an ancestor of the current cfg.WorkDir — the only legitimate use case is the branch materializer promoting <task>/<repo>/ to <task>/ Defense in depth: agentctl binds to localhost, but any colocated process could otherwise direct it at /etc or the user's home by POSTing /api/v1/workspace/rescan. * fix(agentctl): tighten validateRescanPath, close TOCTOU on existingTrackers, add tests - Drop the strings.Contains(clean, "..") guard. filepath.Clean already resolves traversal; the substring check rejects legitimate names like /home/u/..hidden without adding defense (the ancestor check via filepath.Rel is the real boundary). - Drop the rel == "." branch in the ancestor check — early return for clean == currentClean already covers identity. - Read existingTrackers inside the write-lock section (not under RLock then re-Lock) so two concurrent rescans can't both see 0 trackers and double-bootstrap the multi-repo set. - Add table-driven TestValidateRescanPath covering rejection paths (relative, descendant, sibling, traversal-cleaned, empty input), legitimate ancestor promotion, exact match, no-anchor first-launch, and a /home/u/..hidden dotted-segment case. * fix(agentctl): break taint flow into os.Stat — derive rescan path from cfg, not HTTP body CodeQL's go/path-injection trace still flagged manager_rescan.go after the prior sanitizer because the user-supplied newWorkDir flowed (after validation) into os.Stat and then into cfg.WorkDir → tracker paths → workspace_monitor os.Stat sinks. Replace validateRescanPath with resolveRescanPath: it returns a path DERIVED from currentWorkDir (no-op = currentClean, promotion = filepath.Dir(currentClean)) and uses newPath only for comparison, never as a source for the returned string. os.Stat now receives a path whose provenance is trusted manager config. Functionally the legitimate caller (branch materializer promoting <task>/<repo>/ to <task>/) is unchanged. * test(agentctl): make TestResolveRescanPath windows-portable The Linux Backend Tests job passed but the Backend (windows) job rejected every "absolute" case because filepath.IsAbs("/task/repo") is false on Windows (no volume letter). Build inputs via filepath.Join against absRoot() (filepath.Abs of "/" / "\\") so the same table covers POSIX (`/...`) and Windows (`C:\...`) without per-OS branches. Keep the traversal-after-Clean case POSIX-only since filepath.Clean is the shared primitive and that input form needs a volume to be considered absolute on Windows anyway. * fix(agentctl): serialize RescanRepositories + re-check trackers under write lock Two concurrent RescanRepositories calls could both observe an empty tracker set (between the snapshot read and the bootstrap branch) and both enter transitionToMultiRepoMode, leaking a duplicate bare-root tracker and duplicate per-repo trackers. Same race in appendNewRepoTrackers would double-append the same new child across both callers. Fix: - Add rescanMu sync.Mutex on Manager. Acquired for the full body of RescanRepositories so the snapshot → branch → mutate sequence is serialized. repoTrackersMu still guards individual field reads from unrelated goroutines (subscribe/unsubscribe/poll-mode). - appendNewRepoTrackers re-checks membership inside the write-lock before appending and drops already-present trackers (Stop + detach outside the lock). Defense in depth. - Test: add `dotted segment is legal as promotion` case so the ..hidden-segment promotion path stays exercised. --------- Co-authored-by: Kandev Agent <agent@kandev.dev> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> |
||
|
Carlos Florêncio
|
5591cdc0bf | fix(web): add button to load older chat messages reliably (#1223) | ||
|
Carlos Florêncio
|
0f87f7a56b | fix(acp): enrich tool call metadata per agent with structured ACP fields (#1212) | ||
|
João Salavisa
|
aea266ee71 |
fix(web): fix hold-to-talk on mobile and add coarse-pointer toggle fallback (#1231)
* fix(web): make voice-mode hold gesture survive touch and add mobile coverage
Mobile users with the chat composer's mic button in hold-to-talk mode lost
their recording the moment a finger crept off the 28px target, because
`buildHoldHandlers` aborted on `pointerleave`/`pointercancel` without ever
claiming the pointer.
- `buildHoldHandlers` now claims the pointer on `pointerdown` via
`setPointerCapture` and releases it on `pointerup`/`pointercancel`. The
`pointerleave` handler is dropped — capture keeps the gesture alive across
small finger movement, OS gesture handoff, and soft-keyboard reflows.
- `EnabledVoiceInputButton` resolves an `effectiveMode` via a coarse-pointer
check (`matchMedia("(pointer: coarse)")`); when the stored preference is
`hold` and the device is touch-first, the mic falls back to toggle so a
bad stored preference can't trap users in a broken UX. Persisted
`voiceMode.mode` is left untouched so docking a keyboard restores the
user's choice without a save.
- The button bumps to 40×40 on coarse-pointer devices to meet the 44pt touch-
target floor without disturbing the desktop submit/voice pair.
- The unsupported fallback path (`UnsupportedVoiceButton`) is unchanged so
LAN-HTTP mobile users still get the HTTPS explanation toast.
Adds `voice-input-button.test.tsx` covering capture/release on pointer
events, the absence of `pointerleave` regression, the coarse-pointer toggle
override, and the touch-target class swap.
Adds `mobile-voice-mode.spec.ts` (Playwright mobile-chrome / Pixel 5)
verifying the Voice Mode settings page is reachable from the mobile sidebar
trigger and that the mic button mounts with `data-effective-mode="toggle"`
on the mobile chat composer.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* test(web): drop unused testPage param from mobile-voice-mode seedTask helper
The helper only calls apiClient and does not touch the Playwright Page
object — the parameter was carried over from the neighbour spec's pattern
where the helper also navigates. Removing it also drops the now-unused
`Page` type import.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* refactor(web): drop redundant Element type cast from safePointerRelease
`Element.hasPointerCapture` and `releasePointerCapture` are already declared
in lib.dom.d.ts, so the `target as Element & { hasPointerCapture?: ... }`
casts were over-defensive. Replace with a direct `typeof` runtime check,
which still guards against the headless DOMs that strip these methods.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* test(web): assert coarse-pointer touch-target class in mobile voice spec
The previous assertion (`data-effective-mode === "toggle"`) would have
passed even if the coarse-pointer override were reverted, since the default
stored `voiceMode.mode` is already `"toggle"`. Add a check on the `h-10
w-10` class swap — that class pair is only emitted when
`useIsCoarsePointer()` returned true, so the test now genuinely exercises
the override on the e2e surface.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* test(web): extend voice-input-button unit coverage for toggle paths
- Toggle-mode + fine-pointer: pointerdown must not start recording; click
drives the toggle. Guards against an accidental future change that wires
hold pointer handlers when stored mode is "toggle".
- Recording state + click: stop() fires (not start()), exercising the
buildToggleHandler branch the original suite skipped.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* test(web): seed hold mode in E2E and fix misleading test title
- Add `voice_mode` field to `saveUserSettings` in api-client.ts so E2E
tests can PATCH the user's stored activation mode via the API.
- Seed `mode: "hold"` before the mic-button assertion so
`data-mode="hold"` + `data-effective-mode="toggle"` together prove the
coarse-pointer override fired (not that the default was already toggle).
- Rename settings-page test to drop the "from mobile sidebar" claim —
the test navigates directly to the URL; sidebar existence is still
asserted as a regression guard.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Kandev Agent <agent@kandev.dev>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
|
||
|
Carlos Florêncio
|
e2b2eadb72 | fix(web): apply repository filter correctly on task board (#1215) | ||
|
João Salavisa
|
58dd2773a2 |
feat(web): show inline download progress for Whisper Web model (#1228)
* feat(web): show download progress for voice mode Whisper Web model First-time use of voice mode with the Whisper Web engine downloads ~40–240 MB of model files before the first transcription. The existing modelLoad progress was only surfaced via the mic-button tooltip, so the download felt silent — users couldn't tell whether it was progressing, stuck, or how much was left. Adds a compact inline indicator next to the mic button that renders a percent label and a Progress bar while the model is downloading, and a short error label when load fails. The indicator stays hidden for the Web Speech and server-Whisper engines (which don't download models) and clears once the model is ready. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * refactor(web): tidy VoiceModelLoadIndicator class names and a11y - drop unused cn() import (single-string className doesn't need it) - drop redundant max-w-[8rem] (w-32 already equals 8rem in tailwind) - include modelLabel in the error state's aria-label so screen readers identify which model failed Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * refactor(web): a11y improvements for VoiceModelLoadIndicator - Replace two separate loading/error divs with a single stable role="status" wrapper so AT announces the loading→error transition (newly-inserted live regions are silently ignored by most screen readers). - Move numeric progress to role="progressbar" + aria-valuenow/min/max on the inner div; removes per-tick live-region announcements. - Drop redundant aria-live="polite" (role="status" implies it per ARIA spec). - Align error visible text with aria-label: both now use modelLabel ("Whisper Base failed to load") instead of the generic string. - Standardize tooltip wording to match indicator: "Downloading {model}…" instead of "Loading model…". - Update test: error branch expects model-specific label. * refactor(web): fix nested progressbar and clamp tooltip percent - Remove role="progressbar" + aria-value* from the wrapper div; <Progress> (Radix) already renders its own progressbar — the nested node confused AT. Pass aria-label directly to <Progress> for the accessible name. - Drop outer aria-label on the loading state (Radix progressbar carries it). - Clamp tooltip percent in voice-input-button.tsx to 0–100 and handle NaN (mirrors clampPercent in the indicator, keeps UI consistent on edge values from transformers.js). - Update test: assert aria-label via getByRole("progressbar").getAttribute(). * fix(web): guard tooltip percent against NaN; assert error aria-label in test - Tooltip loading percent: add Number.isFinite guard so NaN from transformers.js progress produces "0%" instead of "NaN%". - Test: assert aria-label on error state root matches modelLabel. * fix(web): add w-32 to error variant to prevent layout shift on state change Loading branch uses w-32 to hold a stable 128px slot in the toolbar. Error branch lacked it, so the mic button shifted on loading→error transition. --------- Co-authored-by: Kandev Agent <agent@kandev.dev> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> |
||
|
João Salavisa
|
1fdeda904b |
feat: add voice mode for chat input (#1159)
* feat: add voice mode for chat input Adds a voice input button to the chat composer with three transcription engines: in-browser Web Speech API, in-browser Whisper via @huggingface/transformers, and a server-side OpenAI Whisper fallback gated on KANDEV_VOICE_OPENAI_API_KEY. User-configurable from a new /settings/voice-mode page: - Master on/off toggle (defaults to on; opt-out, not opt-in) - Engine preference (auto / Web Speech / Whisper Web / Whisper Server) - Language hint (auto-detect or BCP-47 pinned) - Activation mode (click-to-toggle or hold-to-talk) - Auto-send after transcription - Whisper Web model size (tiny / base / small) - Editable Cmd+Shift+M keyboard shortcut Voice settings live under user_settings.voice_mode as a nested JSON object — no schema migration. Setting changes broadcast via the existing user.settings.updated WS event so any open chat tab updates live. The mic button is rendered on the right of the composer alongside the send button (round primary fill; destructive fill with pulse ring while recording). On browsers without a usable engine (e.g. iOS Chrome over HTTP) the button stays visible as a muted greyed icon — tapping it surfaces a toast with the reason (typically a missing secure context). Whisper Web runs in a Web Worker with fp16 decoder weights (avoids the MatMulNBits scale-missing crash that q4/q8 hit on WASM). Models are pulled from the onnx-community/* mirror and cached by transformers.js in the Cache Storage on first use. * fix: address voice mode PR review feedback and E2E selector clash Bot-review fixes from #1159: - Map http.MaxBytesError from c.FormFile to 413 instead of bare 400. - Strip BCP-47 region suffix before passing language hint to Whisper Web. - Recover from worker crash: terminate + null worker + reset ready flag so next init/transcribe creates a fresh worker instead of hanging. - Read live store state via storeApi.getState() in voice-mode settings savers to avoid stale-closure clobbering concurrent settings updates. - Detach Web Speech callbacks before abort() so trailing onend/onerror events don't mutate hook state after cancel(). - Spread caller init before method/body in voice-api fetch so a caller- supplied init can't override the multipart POST. - Replace t.Fatalf with t.Errorf+return inside test HTTP handler goroutine (FailNow from non-test goroutine is unsafe). - Add Enabled/AutoSend zeroing assertions in partial-update test so the bool-overwrite-on-PATCH behavior is locked in explicitly. E2E fix: - toolbar-overflow spec used button.rounded-full which now matches both the send and voice buttons (strict-mode violation). Switched to the submit-message-button testid for an unambiguous locator. * fix: scope voice settings rollback + whisper worker error to changed keys Round 2 of PR review feedback on #1159: - voice-mode settings rollback now restores only the keys patched by the failing request (not the whole snapshot), preserving concurrent edits to unrelated voiceMode fields. - Shortcut rollback diffs previous vs next and reverts only the changed keys (restore prior value or delete if absent), preserving other concurrent shortcut edits. - Whisper worker error handler captures the worker reference at attach time and only clears refs when this.worker still matches it — a stale error from a previously replaced worker no longer nukes the active one. - Update stale whisper-web-models JSDoc that still pointed at Xenova mirrors. * fix: voice mode progress scale, hold-mode race, race-aware rollback Round 3 of PR review feedback on #1159: - Normalise Whisper progress to 0–1 in ensureWhisperClient (transformers.js emits 0–100, the button's display + the `ready: 1` convention expect 0–1 — fixes the 5000% mid-download display). - Claim driverRef synchronously at the top of finishCapture so concurrent pointerup + pointerleave invocations early-return instead of racing and clobbering a freshly-started recording's ref. - Race-aware rollback: voice-mode and shortcut catch branches now skip reverting a key when latest[key] !== next[key]. A failed earlier request no longer overwrites a later successful save to the same key. - Update the Worker wire-protocol comment to show onnx-community/whisper-base instead of the stale Xenova/* example. --------- Co-authored-by: Kandev Agent <agent@kandev.dev> |
||
|
Carlos Florêncio
|
d8e0507d17 | fix(web): destroy terminal shells on tab close with busy confirmation (#1203) | ||
|
Carlos Florêncio
|
3da4eac469 | fix(web): collapse large PR Changes and expand Commits by default (#1206) | ||
|
Carlos Florêncio
|
41ad777ec9 | fix(web): render multi-level subtasks in sidebar task tree (#1204) | ||
|
Carlos Florêncio
|
b2faab430b | fix(web): remove agent id from subagent metadata chips (#1205) | ||
|
Carlos Florêncio
|
5f24349aae | feat: show relative commit time on hover in changes panel (#1199) | ||
|
Carlos Florêncio
|
83c800879a | fix(web): stop right pane width from drifting across tasks (#1201) | ||
|
Carlos Florêncio
|
d1c73d2f2c | feat: surface subagent task tool calls as cards in task chat (#1132) | ||
|
Carlos Florêncio
|
036aee6e19 | fix(agents): add Cursor auto-approve and always-allow permission UI (#1198) | ||
|
Carlos Florêncio
|
f7f4a965a1 | fix(web): portal inline-code tooltip to body to prevent clipping (#1187) | ||
|
José Almeida
|
137336de86 |
fix: manually drain queued chat messages after cancel (#1166)
* fix: drain queued chat messages after cancel * fix: address queue drain review feedback * fix: keep queued messages parked after cancel * fix: stabilize queue drain errors * fix: keep queue parked during cancel events * fix: wait for lifecycle stream goroutines * fix: address lifecycle stream review feedback |
||
|
Carlos Florêncio
|
8e1f55a0ac | fix: keep env prep out of partial tool history (#1189) | ||
|
Carlos Florêncio
|
623044a8d4 | feat(web): confirm session delete when closing multi-session agent tab (#1174) | ||
|
Carlos Florêncio
|
ec289c799b | feat: retry transient provider 529 errors with visible backoff (#1173) | ||
|
Carlos Florêncio
|
d35f36ae73 | fix(web): make left sidebar width global + steady monitor-switch settling (#1140) | ||
|
Carlos Florêncio
|
9f9b5c3f72 | feat(web): auto-expand the first group in the changes panel (#1169) | ||
|
Carlos Florêncio
|
55ed674a15 | fix: dockview/editor restore races, office rebroadcast, and flaky tests (#1171) | ||
|
Carlos Florêncio
|
d66d9e6f67 | feat: remote tab in task-create dialog — multi-row GitHub repo picker (#1116) | ||
|
Carlos Florêncio
|
72441fb69a | feat(web): add PR closed banner and hide CI chip on terminal state (#1161) | ||
|
João Salavisa
|
a34ea719df |
feat(web): allow custom prompts in task creation input (#1155)
* refactor(web): extract measureCaretRect into shared util Move the textarea caret-position measurement helper out of rich-text-input.tsx into apps/web/lib/utils/caret-position.ts so the task-create dialog's autocomplete wiring can reuse it without depending on the chat-only RichTextInput component. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * feat(web): allow custom prompts in task creation input Add an @-mention autocomplete to the task-create dialog's prompt textarea so users can quickly insert one of their saved custom prompts (built-in + user-created) by typing @<name>. - Extend useInlineMention with a promptInsertMode option. The default "context" mode preserves chat behavior (attach the prompt to the context store). The new "inline" mode replaces the @query text with the prompt's full content, requiring no backend changes -- the resolved text just becomes part of the task description on submit. - Wire the hook into TaskFormInputs via a small useTaskCreatePromptMention helper that adapts the existing textarea ref to the RichTextInputHandle surface the hook expects (focus / selection / caret-rect / value). - Reuse the existing MentionMenu popup component for parity with chat. - Cover the new branches with unit tests for makePromptItem / detectMentionTrigger / filterItems, and an E2E spec exercising the menu open / inline-insertion / Enter-doesn't-submit / trigger-inside-word paths. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(web): make mention popup clickable inside Radix Dialog Radix Dialog sets pointer-events: none on document.body to enforce modal isolation. The mention/slash menu portals to document.body and inherits the disabled pointer events when opened from inside a Dialog (the task-create flow). The popup was visible but unclickable: keyboard select worked, mouse clicks were swallowed by the Dialog body. Restore pointerEvents: auto on the popup container so it stays clickable when invoked from within a modal. Bump z-index to 60 so the popup also wins the visual stacking race against the Dialog (overlay + content are both z-50). Chat continues to work unchanged since z-60 and pointer-events: auto are no-ops outside a modal. Add an adversarial QA spec covering mouse-click select, ArrowDown + Enter on the 2nd item, Escape close, multi-line content auto-grow, backspace-over-@ close, space-after-@ close, no-arrow-cursor-shift, and end-to-end submit. The mouse-click case is the one this fix unblocks; the rest pin behavior for future regressions. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(web): address PR review feedback - caret-position.ts: wrap DOM measurement in try/finally so the mirror element is always removed even if getBoundingClientRect throws - useTextareaHandlers: destructure stable fn refs from mention before useCallback so deps are specific properties, not the whole object — avoids defeating memoization on every keystroke - QA spec: consolidate inline cleanupPrompts calls into a shared test.afterEach so cleanup always runs even when a test throws; update submit test comment to accurately reflect what is asserted Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(web): tighten ArrowDown selection assertion in QA E2E spec Equal filter scores sort by insertion order (stable), so qa-arr-1 lands at index 0 and qa-arr-2 at index 1. After one ArrowDown from the default the selected item is "SECOND" — assert that directly instead of accepting either value, matching the test title "selects the second prompt". Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Kandev Agent <agent@kandev.dev> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> |
||
|
Carlos Florêncio
|
cf2cd1bd2f | feat(web): unify panel loading states with grid spinner (#1142) | ||
|
Carlos Florêncio
|
83338a04c3 | feat: surface agent errors and recover corrupted resume sessions (#1144) | ||
|
Carlos Florêncio
|
0a753f48e8 | feat(web): show relative times in remote cloud status tooltip (#1146) | ||
|
Carlos Florêncio
|
0c031c6098 | fix(web): honor default dockview widths on task open (#1136) | ||
|
Carlos Florêncio
|
2a68a9b7c7 | feat(web): persist PR-merged banner dismissal in sessionStorage (#1129) | ||
|
José Almeida
|
2b16a53d6d |
fix: inherit parent workspace for subtasks (UI + MCP) (#1131)
* fix(backend): inherit parent workspace for subtasks on all start paths Subtasks with workspace_mode inherit_parent were provisioning a fresh worktree and tasks folder instead of reusing the parent's. Two causes: - propagateInheritedEnvironment ran only in PrepareTaskSession, so the direct start path (startTask, used by MCP auto-start) never bound the child session to the parent's TaskEnvironment. Moved it into prepareSessionForStart so every launch entry point inherits. - The service CreateTask path did not inherit the parent's repositories when none were supplied (the UI omits them expecting inheritance), so the child was repo-less and could not establish/reuse a worktree. Now mirrors the MCP path: inherit RepositoryID + BaseBranch, drop CheckoutBranch. * fix(web): hide executor profile for inherit-parent subtasks When a subtask inherits the parent's materialized workspace, its executor is inherited too, so picking an executor profile is meaningless. Hide the selector under inherit_parent and omit executor_profile_id from the create payload. * fix(backend): fail fast when subtask parent-repo inheritance lookup errors Return the error from inheritParentRepositories instead of swallowing it: a subtask silently created with no repositories would reintroduce the fresh-worktree bug at launch. Also clarified that inheritance applies to any repo-less subtask, matching the MCP path. |
||
|
Carlos Florêncio
|
efb6493760 | feat(web): move queued chip into chat status bar row (#1127) | ||
|
José Almeida
|
7a2df415a0 | chore(web): drop redundant hover tooltip on tool execute command (#1104) | ||
|
Carlos Florêncio
|
49f9741817 | feat(web): show executor-specific cleanup details in task delete/archive dialog (#1103) | ||
|
Carlos Florêncio
|
f7eff241b9 | ci: faster e2e via image split, shard rebalance, and flake fixes (#1126) | ||
|
Carlos Florêncio
|
8f065f1673 | feat(web): stop filename squeeze on changes-panel row hover (#1114) | ||
|
Carlos Florêncio
|
a068c88b99 | fix: show human-readable title and details in permission prompts (#1101) | ||
|
Carlos Florêncio
|
f4be0d279c | fix: dockview, sprite, auth, and CLI db backup improvements (#1075) | ||
|
Carlos Florêncio
|
905b9f9727 | fix: improve UX for subtasks with repo setup scripts (#1105) | ||
|
José Almeida
|
ec8039ca71 |
fix(web): always allow expanding tool execute to reveal full command (#1086)
* fix(web): always allow expanding tool execute to reveal full command * fix(web): keep tool execute tooltip trigger keyboard-focusable |
||
|
Luan Martins
|
1a467b969f |
fix(web): wire permission approval UI for Kandev MCP tools (#1037)
* fix(web): wire permission approval UI for Kandev MCP tools
KandevToolMessage adapter ignored the permissionMessage prop, so MCP tools
that block on a permission request rendered the result row but no Approve/
Deny buttons — users could not unblock the agent. Extract the shared
permission handlers into use-permission-handlers.ts and reuse them from
both ToolCallMessage and KandevToolMessage. KandevRow now also shows an
amber pending clock icon and hides the result-count summary while waiting
for approval so the row reads correctly before the result arrives.
Adds a mock-agent scenarioKandevMCPPermission emitting
mcp__kandev__list_workspaces_kandev plus a Playwright e2e and unit tests
covering the pending UI and the absence of "0 workspaces" while pending.
* fix(web): collapse tool + permission status into resolved Kandev row
After approve/reject the orchestrator stamps the decision onto the
permission_request message immediately, but the tool_call's own
meta.status stays "pending" until the next tool_call update lands.
KandevToolMessage previously passed meta.status straight to the renderer,
so the row kept the amber pending clock and the hidden-summary state for
the brief window between the two events. resolveKandevStatus now prefers
the permission decision once present, mapping approved -> complete and
rejected -> error.
Drops a redundant Playwright `.first()` after a toHaveCount(1) assertion
on the same locator. Adds unit coverage for the resolved-decision state.
* fix(web): keep Kandev tool status independent of permission decision
The previous fix collapsed an approved permission into tool status
"complete" and a rejected permission into "error". That was wrong on two
counts: an approved tool that's still running rendered as already done
(green check + visible result summary built from an empty payload), and
any later tool_call error was silently masked because the status had
already been forced to "complete".
Replace the status override with a separate display overlay. KandevRow
now reads a `KandevPermissionUIContext` ("pending" | "rejected" |
undefined) and uses it only to drive the waiting-on-user affordances —
amber clock, hidden summary — and a rejection mark that wins over the
tool's own success state. The renderer always receives the real
tool_call status, so a tool that's approved-but-pending still reads as
pending and a tool that errors after approval still renders the red X.
Adds explicit regression tests for the two race-window failure modes
plus the new rejected overlay.
* test(web): tolerate duplicate-render backend race in Kandev MCP e2e
CI shard 1 sometimes renders two approve buttons for the same Kandev MCP
permission because the orchestrator can create a second generic tool_call
message for the same tool_call_id alongside the kandev-routed one. Both
buttons resolve the same pending_id, so clicking the first unblocks the
agent either way.
The bug the test guards against is "no approve button at all" — the
strict toHaveCount(1) assertion was tying the test to that separate
backend issue. Relax to at-least-one and click the first.
* fix: persist Deny clicks as rejected, not approved
Backend RespondToPermission previously derived status purely from cancelled:
cancelled=true -> "rejected", otherwise "approved". Deny clicks sent
cancelled=false with a reject_* option_id, so they were recorded as
"approved". Add an explicit rejected flag from the WS request so the
backend can record "rejected" without taking the cancellation path (which
races against permission_cancelled -> "expired").
Also refactor permission-request-message.tsx to consume the shared
usePermissionResponseHandlers hook (the duplicated inline handler had the
same bug) and introduce a typed models.PermissionStatus enum on the
backend so status strings are not scattered literals.
* test(web): scope Kandev MCP permission e2e to the Kandev row
Previously the test clicked the first matching permission-approve button,
which could be a duplicate generic ToolCallMessage fallback rendered for
the same pending_id by a backend race -- a green test no longer proved the
custom Kandev approval UI rendered.
Add a `data-testid="kandev-tool-permission"` wrapper around the Kandev
action row and a `kandevPermissionApproveButtons` page-object locator that
scopes through it, then assert `toHaveCount(1)` so the test only passes
when the Kandev custom row actually mounts.
* chore(orchestrator): log rejected flag in RespondToPermission
Debug log was missing the new rejected param, making it impossible to
distinguish explicit Deny from dialog cancellation when triaging
permission-flow issues.
|
||
|
Daniel Brown
|
c0eaafaf29 |
fix(mobile): isPassthroughMode falls back to TaskChatPanel when snapshot is missing (closes #1031) (#1034)
* fix(mobile): passthrough TUI rendering + touch-scroll Two related mobile-UI fixes for cli_passthrough sessions, both filed in #1031. 1. task-center-panel.tsx: align isPassthroughMode with the rest of the codebase by checking session.is_passthrough first. The memo previously only consulted agent_profile_snapshot.cli_passthrough, which can be undefined in the Redux store on re-mount (the snapshot isn't always present on the leaner session-list responses). When that happened on mobile, the chat panel rendered the persisted- messages TaskChatPanel fallback instead of the live PassthroughTerminal. Desktop was immune because dockview-desktop-layout.tsx already uses the session.is_passthrough field directly. is_passthrough is the source of truth used by manager_startup.go's StartAgentProcess routing — keeping the frontend consistent with that invariant. 2. passthrough-terminal.tsx: wire touch-drag scrolling on the xterm viewport. xterm.js handles mouse wheel natively but not touch gestures, so swiping inside the terminal on mobile bubbled up and scrolled the page instead of the terminal scrollback — effectively making everything that scrolled off-screen unreachable on mobile. The new hook attaches touchstart/touchmove/touchend handlers that map vertical drag delta to term.scrollLines(), gated on (pointer: coarse) so desktop xterm behaviour stays identical. Row-height comes from xterm's internal render dimensions with a measured fallback (xterm 5.x layout) and a 16px constant fallback if both fail. Closes #1031 * fix(mobile): also fix useSessionLayoutState's isPassthroughMode + tighten touch-handling Three follow-on fixes after Dan's mobile QA surfaced more cases: 1. apps/web/hooks/use-session-layout-state.ts — the mobile layout actually uses THIS hook's isPassthroughMode memo (not the one in task-center-panel.tsx that I previously fixed). It had the identical snapshot-only bug. Aligned with the same is_passthrough-first priority. 2. apps/web/components/task/passthrough-terminal.tsx — always preventDefault on touchmove (was only doing so when the row-delta calculation rounded to non-zero, so sub-row drags slipped through to parent scroll). Added touch-action: none + overscroll-behavior: contain inline on the terminal element so the browser never claims the gesture. 3. apps/web/app/globals.css — overscroll-behavior: none on body to block pull-to-refresh. iOS Safari ignores this in some configurations but blocking it where possible reduces the page-reload race that was triggering the stuck-on-loading-messages state. * fix(mobile): trim PR to minimum required changes After QA showed the user-visible mobile bug is fully addressed by the useSessionLayoutState memo fix alone, revert the parallel changes to the desktop hook (task-center-panel.tsx's useSessionApprove) and the body-level overscroll-behavior — neither was required for the reported bug, and Dan asked for the smallest possible diff. Remaining changes: - apps/web/hooks/use-session-layout-state.ts — actual mobile fix - apps/web/components/task/passthrough-terminal.tsx — touch-scroll feature * fix(mobile): tighten isPassthroughMode + drop terminal touch-scroll from this PR Adversarial-review feedback applied: - Change `is_passthrough === true` to `!== undefined` so an explicit server-side false isn't overridden by a stale snapshot. - Comment now cites the specific WS handler quirk in ws/handlers/agent-session.ts where snapshot writes are truthy-guarded but is_passthrough writes are !== undefined-guarded — that's the concrete path where partial events land one without the other. Dropped: the useMobileTouchScroll hook. Terminal scroll on mobile is a separate feature that deserves its own discussion (xterm renders the canvas above the scrollable viewport so native scroll doesn't reach the viewport; the right fix likely lives in xterm itself or a small event-redirect addon). Keeping this PR scoped to the rendering bug. * fix(mobile): add minimal touch-scroll fallback for xterm viewport xterm.js renders .xterm-screen (canvas) above .xterm-viewport, so iOS / Android native scroll on the viewport never sees the touches and the terminal is unscrollable on mobile. useTouchScrollFallback wires touchstart/touchmove/touchend to Terminal.scrollLines via an accumulated fractional-row delta. No inertia — the proper fix is upstream in xterm-js. Gated on (pointer: coarse) so desktop xterm behaviour is unchanged. Sets touch-action: none + overscroll-behavior: contain inline on the terminal element so the browser doesn't claim the gesture for pull-to-refresh. * fix(mobile): inertia decay on touch-scroll release for native-scroll feel * fix(mobile): address adversarial review of touch-scroll hook - Read xterm from ref inside each handler so renderer recreate doesn't strand the handlers on a stale Terminal instance (React refs are stable so dep array doesn't catch .current swaps). - Drop the internal _core._renderService private-API path for row height in favor of public term.rows + measured element height. The internal path was API-unstable across xterm majors. - Velocity window 80ms → 50ms so a direction reversal at the tail of the gesture is the release velocity, not averaged against the early motion. - Reset pendingRows fractional accumulator on direction reversal — was causing perceptible stickiness when reversing from slow-up to fast-down. - Style restoration now uses removeProperty when the inline style was never set, rather than leaking an empty string in its place. * fix(mobile): revert touch-scroll hook from this PR — separate concern After several iterations the touch-scroll fallback kept surfacing new issues (iOS pull-to-refresh leaking through preventDefault, scroll direction reversal edge cases) that suggest a more involved upstream discussion is needed about whether to ship a workaround in PassthroughTerminal at all vs fixing xterm.js's mobile story upstream. Splitting that work out so this PR can land the rendering bug fix cleanly. Will file a separate issue for the touch-scroll work with the diagnosis + alternatives we explored. * fix(mobile): honor is_passthrough when snapshot missing --------- Co-authored-by: Carlos Florêncio <carlosmflorencio@gmail.com> |
||
|
Carlos Florêncio
|
f1dbad1dbd | feat(cli-mode): auto-inject task prompt + dynamic submit sequence + passthrough UX fixes (#923) | ||
|
Carlos Florêncio
|
8a7b7a0b60 | feat(agentctl): azure Repos PR follow-ups after #1066 (#1071) | ||
|
José Almeida
|
777dfdb15b |
feat(ssh): add SSH executor (#927)
* feat(backend): add SSH executor runtime
Run agents on any Linux box reachable over SSH: own VPS, homelab, EC2,
shared dev box. One SSH connection per (host, user, identity) shared
across all sessions on that host; each session gets its own remote
agentctl process, remote port, and SSH-tunneled local port. Mirrors the
local ~/.kandev/tasks/<task-dir>/<repo>/ layout on the remote so
multi-session tasks share a worktree and per-session runtime state lives
under the task dir.
Implements ExecutorBackend + RemoteSessionResumer + RemoteStatusProvider
+ Closeable. Host-key trust is explicit: the test endpoint observes a
fingerprint without pinning; the user trusts it in settings; subsequent
connects reject any mismatch. Auth via ssh-agent or IdentityFile only
(passphrase-protected keys must be loaded into ssh-agent). Parses
~/.ssh/config for HostName/Port/User/IdentityFile/ProxyJump inheritance.
agentctl is content-hash-cached on the remote (~/.kandev/bin/agentctl
plus a .sha256 sidecar) and uploaded only when local and remote differ.
Linux/amd64 only in v1, matching Docker's current state; arm64 is
called out as a cross-cutting follow-up.
See docs/specs/ssh-executor/spec.md.
* feat(backend): add SSH test + sessions HTTP/WS endpoints
POST /api/v1/ssh/test runs the connection probe used by the settings UI:
dials the host with a permissive host-key callback that records (but
does not pin) the fingerprint, probes uname/arch/git, and reports
whether agentctl is already cached on the remote. The UI gates the
"Trust this host" checkbox on success + fingerprint, and Save persists
the trusted fingerprint into the executor config.
GET /api/v1/ssh/executors/:id/sessions returns the active SSH-runtime
sessions backing the "Active sessions" card on the executor settings
page, mirroring the SpritesInstancesCard / DockerContainersCard shape
(task, session, user@host, remote port, local forward port, uptime).
Both routes also exposed over WS via ssh.test / ssh.sessions.list.
* feat(web): add SSH executor settings UI
SSHConnectionCard: connection form (name, host_alias / host / port /
user, identity source = ssh-agent | file with optional file path,
optional ProxyJump) with a Test Connection button that runs the backend
probe, surfaces the observed host fingerprint, and only enables Save
when the user explicitly ticks "Trust this host". Editing the host /
port / user / identity fields resets both the test result and the trust
checkbox. On save, if the executor already has running sessions, a
confirm modal warns that those keep going on the old host and only new
sessions will use the new config.
SSHSessionsCard: 90s-polled table of active sessions for the executor,
mirroring SpritesInstancesCard / DockerContainersCard. Columns: task,
session, user@host, remote agentctl port, local forward port, uptime,
status badge. Backed by GET /api/v1/ssh/executors/:id/sessions.
Also registers the ssh entry in executor-icons.ts (label "SSH",
IconTerminal2) so it surfaces consistently in the rest of the settings
chrome.
* docs: add SSH executor v1 spec
Captures the design behind the new SSH executor: why ("run on a Linux
box I already own" between local and Sprites), the remote layout that
mirrors local ~/.kandev/tasks/, per-session agentctl + forward over a
shared per-host SSH connection, explicit-trust host-key flow, ~/.ssh/config
inheritance, amd64 + agentctl content-hash cache, recovery via persisted
ExecutorRunning metadata, settings UI (Connection + Sessions cards), and
the edit-with-live-sessions warning UX. Out-of-scope items (arm64,
macOS, password auth, chained ProxyJump, working-tree push, etc.) are
called out explicitly with the follow-ups they unblock.
* chore(e2e): rename docker project to containers + add e2e README
The Docker e2e Playwright project was the natural home for any test
that needs a real Docker daemon on the host — but its name implied it
was only about the Docker executor. SSH e2e tests (incoming in
follow-up commits) belong here too, since they use a Dockerised sshd
container as the SSH executor's remote target.
- Rename project: docker -> containers, widen testMatch to also include
tests/ssh/*.spec.ts.
- Rename env flag: KANDEV_E2E_DOCKER -> KANDEV_E2E_CONTAINERS, with the
old name kept as a deprecated alias for one release so local scripts
and CI configs don't break overnight.
- Rename CI job: e2e-docker -> e2e-containers, switch its blob report
name and the merge-job dependency.
- Add apps/web/e2e/README.md explaining the project layout, what the
containers project covers (Docker executor + SSH executor), why it
was renamed, and how to run it locally.
- Add docs/specs/ssh-executor/e2e-plan.md tracking the SSH e2e
coverage we'll implement on top of this rename (real sshd container,
no mocks, 17 spec files covering ~75 cases). Includes a "Deferred"
section that flags arm64 as something to revisit when
agentctl-linux-arm64 lands in the build pipeline.
* feat(web): add data-testid attributes to SSH settings UI
Sweep stable test ids across the SSH connection and sessions cards so
the incoming Playwright suite can hold steady against refactors. Splits
the formerly monolithic ssh-settings.tsx into three files to stay
within the per-file/per-function size limits:
- ssh-connection-card.tsx (form + test-then-trust flow)
- ssh-sessions-card.tsx (active-sessions table)
- ssh-settings.tsx (barrel re-export so existing imports keep working)
Test ids cover: card roots, every input, identity-source select +
options, test/save buttons, connection badge (with data-status), error
text, test-result panel (with data-success), per-step rows (with
data-success), observed/pinned fingerprint, fingerprint-change
warning, trust checkbox, sessions table + per-row cells, empty state,
refresh button, session status badge.
* feat(e2e): scaffold the SSH executor test harness
Lands the building blocks the SSH specs will compose on top of. Every
piece talks to a real sshd container — no mocks at any layer.
- fixtures/ssh-image.ts: builds kandev-sshd:e2e once per machine (Alpine
+ openssh-server + openssh-sftp-server + git + bash + sudo +
iproute2 + iptables, pre-bakes mock-agent at /usr/local/bin/mock-agent,
generates a fresh host key on entrypoint, reads the worker public key
out of a bind-mounted /authorized_keys). The agentctl binary is
deliberately NOT pre-baked so tests exercise the SFTP upload + sha256
cache path the production code uses.
- helpers/ssh.ts: per-worker sshd lifecycle (startSSHServer /
stopSSHServer), keypair generation, host-key fingerprint scanning,
fault injection (regenerateHostKey for host-key-rotation specs,
dropTrafficToPort22 / restoreTraffic for connection-drop tests,
killRemotePid for "agentctl died" recovery), and remote inspection
helpers (readRemoteFile / remotePathExists / listRemoteDir /
countSshdConnections).
- helpers/ssh-bastion.ts: two-container ProxyJump setup. Bastion is on
a private Docker network and has a published port; target is on the
same network with NO host port — any successful connection MUST have
traversed the bastion, which makes the proxy-jump assertion honest.
- fixtures/ssh-test-base.ts: worker-scoped fixture mirroring
docker-test-base. Skips when Docker isn't reachable. Spins up the
sshd container, generates keys, scans the fingerprint, then seeds a
workspace + workflow + bare repo + mock-agent profile + an SSH
executor with the fingerprint already trusted. Specs that need to
exercise the test-then-trust UI gate create their own executor on
top of the same target.
- pages/SSHSettingsPage.ts: Page Object covering the connection form,
test/save buttons, trust checkbox, observed/pinned fingerprint, per
step badges, and the sessions table.
- helpers/api-client.ts: new createSSHExecutor / updateExecutor /
getExecutor / testSSHConnection / listSSHSessions methods plus
matching DTO types.
* test(e2e): add SSH executor Playwright suite (17 specs, ~75 cases)
Every spec drives the real sshd container set up by the ssh-test-base
fixture; no mocks. Specs are grouped by concern (file per group),
mirroring the docker/ directory style. Each one maps 1:1 to a row in
docs/specs/ssh-executor/e2e-plan.md.
UI surface:
- connection-form.spec.ts — field rendering, save/test gating,
identity-source toggle, default port.
- test-result.spec.ts — successful + failed Test Connection paths,
per-step badges, fingerprint surfacing, cached/will-upload copy.
- trust-gate.spec.ts — save disabled until trust ticked, edits reset
result + trust, fingerprint-change warning when pinned != observed.
- sessions-card.spec.ts — empty state, row rendering, manual refresh,
status badges, truncation, polling pickup.
- persistence.spec.ts — fingerprint persists across reload, new form
starts blank, executor list label, partial state doesn't leak.
API contract:
- test-endpoint.spec.ts — POST /api/v1/ssh/test step ordering, invalid
body, host-required, handshake short-circuit, cached/uploaded states.
- sessions-endpoint.spec.ts — GET /api/v1/ssh/executors/:id/sessions
filters to ssh-runtime rows, populates metadata, monotonic uptime.
- executor-crud.spec.ts — type=ssh persists fingerprint, profile
workdir round-trip, edit-with-running-sessions keeps snapshots.
Real-SSH integration:
- launch-task.spec.ts — full smoke: agentctl upload + sha256 cache hit
on second launch, task/session dir layout, stop cleans session dir.
- recovery.spec.ts — backend restart reconnects, dead pid surfaces
failure, persisted metadata keys present.
- auth-methods.spec.ts — key + valid, missing key, passphrase-protected
rejection, agent without SSH_AUTH_SOCK, wrong-key auth failure.
- concurrency.spec.ts — two sessions share task dir distinct ports, two
tasks distinct dirs, traffic drop + restore exercises eviction +
reconnect.
- hostkey-rotation.spec.ts — regenerate host key surfaces new
fingerprint, stale pinned fingerprint blocks launch verbatim, re-trust
restores function.
- config-inheritance.spec.ts — ~/.ssh/config HostName / Port /
IdentityFile / ProxyJump inheritance with form-overrides-config rules.
- workdir-per-profile.spec.ts — per-profile ssh_workdir_root routes to
separate remote workdir trees; default falls back to ~/.kandev.
- proxy-jump.spec.ts — 2-container bastion topology asserts target is
unreachable directly, reachable via ProxyJump, chained ProxyJump fails.
- error-surfacing.spec.ts — TCP refused, auth failed, bastion unreachable
surface verbatim; backend 400 stays a 400 not a synthesised failure.
Also adds an AGENTS.md (CLAUDE.md symlink) line pointing at the
e2e/README.md so future agents discover the containers project +
KANDEV_E2E_CONTAINERS env var without grepping playwright.config.ts.
* fix(e2e): unlock sshd account, derive fingerprint from kandev probe
Three bugs surfaced by actually running the SSH e2e suite
(KANDEV_E2E_CONTAINERS=1 against the containers project):
1. Alpine's `adduser -D kandev` leaves the account password-locked
(`!` in /etc/shadow). sshd then refuses every auth method for that
user — including pubkey — with "User kandev not allowed because
account is locked". Fix: `usermod -p '*' kandev` in the Dockerfile
so the account is open for key-based login but still has no valid
password. Also adds the `shadow` apk package for `usermod`.
2. ssh-keyscan reports the ed25519 host fingerprint, but Go's ssh
client may negotiate a different host key type when sshd offers
multiple — meaning the fingerprint the fixture pre-scanned didn't
match the one kandev observed at handshake time. Fix: the fixture
now hits /api/v1/ssh/test with a permissive callback right after
the container is up and uses that fingerprint as the pinned value.
Same code path real users hit; eliminates the negotiation skew by
construction.
3. ssh.ts had a CPU-spin sleep helper and an nc-dependent TCP probe,
both fragile. Replaced with a proper sleep subprocess and a bash
/dev/tcp probe, plus retry-with-stderr on the fingerprint scan so
failures are diagnosable.
Also tightens test-endpoint.spec.ts F4: the test was using
identity_source=agent to "fail fast", but that fails inside
buildAuthMethods before the handshake step ever runs. Switched to
identity_source=file against a refused port so the test exercises
the actual TCP+handshake failure path.
After these fixes 4/6 test-endpoint cases pass against a real sshd.
* fix(ssh): make the end-to-end launch path actually work
A pass of issues uncovered by running the SSH e2e suite locally
(KANDEV_E2E_CONTAINERS=1 against the real sshd container). After this
commit all 15 non-launch SSH e2e tests pass (UI form validation,
test-connection contract, auth-methods, error-surfacing, sessions-card
empty state) and the launch path runs through prepare -> upload ->
agentctl -> instance -> forward, with WebSocket streams sometimes
disconnecting partway through (tracked below for follow-up).
Image changes (apps/web/e2e/fixtures/ssh-image.ts):
- Install curl so the SSH executor can POST to the in-container
agentctl control server for instance creation.
- Install gcompat + libstdc++ so the dynamically-linked glibc
agentctl binary can exec on musl-libc Alpine. Without it the
kernel returns ENOENT for /lib64/ld-linux-x86-64.so.2.
- Install shadow + 'usermod -p *' kandev so sshd doesn't refuse
pubkey auth for the otherwise-locked user.
- Drop in 'AllowTcpForwarding yes' and 'MaxSessions 100' via
/etc/ssh/sshd_config.d/. Alpine's defaults are 'no' and 10, both
of which crater the SSH executor's port-forward path.
- chown /home/kandev/.kandev to kandev on entrypoint so a host
bind-mount of that path (used in this commit for log persistence)
doesn't make agentctl unable to write its log file.
Image helper (apps/web/e2e/helpers/ssh.ts):
- Bind-mount the host mock-agent-linux-amd64 into the container at
the same absolute path. kandev's orchestrator builds the agent
command from KANDEV_MOCK_AGENT_LINUX_BINARY (a host path), so the
remote needs the binary at that exact path; bake-into-image isn't
enough because the absolute host path is what kandev sends.
- Bind-mount /home/kandev/.kandev so per-session agentctl logs
survive --rm teardown and the test runner can inspect them on
failure.
- Replace the broken CPU-spin sleep helper with a real sleep
subprocess and a /dev/tcp TCP-open probe so timing-dependent
helpers stop racing the container start.
- Retry the host fingerprint scan with stderr capture; the first
scan can race the sshd accept loop on Alpine.
Backend SSH executor (apps/backend/internal/agent/lifecycle):
- executor_ssh_operations.go startRemoteAgentctl: agentctl does not
accept --port-file or --pid-file flags (it reads AGENTCTL_PORT
and writes its own pid). Rewrite the launch wrapper to inject
AGENTCTL_PORT via env, capture $! as the pid, and tail the
on-remote agentctl.log for 'HTTP server bound successfully'
instead of waiting on imaginary side-files.
- executor_ssh_operations.go createRemoteAgentInstance: new helper
that POSTs /api/v1/instances on the in-container control server
via SSH-exec'd curl, returning the per-instance port the executor
should forward to. Mirrors what executor_sprites_operations.go
does inside its sprite. Without this step the agent stream
returns 'configure request failed with status 404' because the
SSH executor was talking to the control server, not the
per-session sub-server.
- executor_ssh.go startAndForwardAgentctl: chains the new
create-instance step in between agentctl launch and port-forward,
and forwards the *instance* port (not the control port).
- executor_ssh.go: drop the per-host SSH connection pool (it was
introducing a subtle aliasing issue between
CreateInstance/StopInstance/GetRemoteStatus that made the SSH
client unreliable across same-host sessions). Each session now
holds its own ssh.Client directly on sshSessionState and
ResumeRemoteInstance dials a fresh client too.
- executor_ssh_operations.go SSHPortForwarder: serialize the
client.Dial calls with a per-forwarder mutex. golang.org/x/crypto/
ssh.Client.Dial is documented as concurrency-safe but in practice
parallel opens from the kandev stream-manager occasionally
produce io.EOF as if the channel-open response was never
received.
Test fix (apps/web/e2e/tests/ssh/error-surfacing.spec.ts):
- Q1 ('TCP refused') was using identity_source='agent' with no
SSH_AUTH_SOCK in the backend env, so it failed at
buildAuthMethods before reaching the TCP-refused path. Switch to
identity_source='file' with the seed key so the test exercises
the failure it claims to.
Known remaining issue (to investigate next): the launch path now
reaches 'agent process started' + 'ACP initialize response
received' when run in isolation, but the WebSocket updates stream
sometimes disconnects with 1006 abnormal closure shortly after. A
standalone Go reproduction of concurrent direct-tcpip channels
against the same image works without issue, so the trigger is
something specific to the kandev backend's stream-manager + my
forwarder interaction. Tracked for the next iteration.
* fix(ssh): half-close forwarded SSH channels to avoid premature WS abnormal-close
The SSH executor's port-forwarder previously called the full Close() as soon
as either direction of the bidirectional copy finished. Under HTTP-Upgrade
WebSocket traffic this meant: when the WS server flushed its final frame and
shut down the write side, our local->remote io.Copy returned EOF, we closed
the entire SSH channel, the remote->local copy aborted mid-frame, and the
backend's WS reader observed close 1006 (unexpected EOF) instead of a clean
close frame.
Switch to a half-close pattern: each direction CloseWrite()s its destination
when its source EOFs, letting the other half drain the remaining frames. The
final Close happens after both copies return.
The "ssh forwarder dial remote failed: EOF" warning in e2e logs was a
follow-on symptom — the SSH client was being torn down by the abnormal close,
so any later channel-open returned EOF immediately.
* fix(ssh): resolve session executor_id via TaskSession join
GET /api/v1/ssh/executors/:id/sessions was filtering ExecutorRunning rows by
their ExecutorID column, but the lifecycle manager's buildRunningFromExecution
only sets that column from a prior row — for fresh sessions it's left empty.
The endpoint therefore returned [] for every just-launched SSH session.
Look up the TaskSession by session_id and compare its ExecutorID instead.
Keep the row's own ExecutorID as a fast-path for resume/recovery flows
where it's already populated.
* feat(web): register ssh entry in the new-executor type map
Without this entry, /settings/executors/new/ssh fell back to the
"unknown executor type" placeholder. Extracts the type map into its
own module so the page file stays under the 600-line lint cap.
The SSH-specific form is still TODO and will replace the standard
profile-create form for this type in a follow-up.
* test(e2e): poll listSSHSessions for SSH session row in group G
The simple-message scenario completes inside a couple hundred
milliseconds, so the previous "wait for executor_type=ssh, then
query once" pattern raced the row's removal during teardown. Poll
listSSHSessions directly with a 60s window so the row is observed
while the executor is up.
Adds dumpRemoteLogs helper for copying per-session agentctl.log
files out of the sshd container on assertion failure, easing
debugging when the row never appears.
* test(e2e): align H5 assertion with actual on-disk session-dir layout
The wrapper writes agentctl.pid and agentctl.log to <sessionDir>; the
remote agentctl port travels via AGENTCTL_PORT env + ExecutorRunning
metadata, not a separate .port file. Assert the files that actually
exist and the metadata field that carries the port.
* feat(web): mount SSH connection + sessions cards on dedicated routes
Wires the existing SSHConnectionCard and SSHSessionsCard into the
settings UI so the SSH executor flow has a home:
- /settings/executors/new/ssh renders SSHCreatePage (just the connection
card; Save POSTs a new executor and routes to the edit page).
- /settings/executors/ssh/[executorId] loads the executor and renders
the connection card pre-filled + the sessions card.
- Adds the SSH entry to the executors landing-page tile grid.
- Adds createExecutor / fetchExecutor / updateExecutor helpers to
settings-api so the new pages can talk to /api/v1/executors directly.
Profile management for SSH executors still flows through the existing
profile pages; this commit only adds the executor-level chrome the
SSHSettingsPage e2e tests expect.
* fix(web): keep SSH trust checkbox visible after editing a connection field
Previously, editing any form field nulled out `result` so the entire
TestResultDisplay (including the trust checkbox) unmounted. The trust-gate
e2e specs assert that after editing host/port/user/identity_source, the
trust checkbox is still present but unchecked — which matches the
intended UX: the user can re-test against the new target and re-tick
trust without losing the prior result for reference.
Now an edit only clears `trust` when the field could route to a
different machine (host_alias, host, port, user, identity_source,
identity_file, proxy_jump); the result panel stays mounted. Editing the
display-only `name` leaves trust alone.
* test(e2e): point SSHSettingsPage at the new SSH executor routes
Updates the page object to navigate to the SSH-specific routes that now
host the connection + sessions cards:
- gotoNew(): /settings/executors/new?type=ssh → /settings/executors/new/ssh
- gotoExisting(): /settings/executors/[executorId] → /settings/executors/ssh/[executorId]
The old URLs collided with the standard profile-create / profile-edit
flow and were never going to render the SSH cards. The new paths give
the SSH UI its own surface, leaving the generic profile pages untouched.
* feat(api): add top-level DELETE /api/v1/executor-profiles/:profileId
Profile IDs are globally unique, so callers that only have the
profile id (e2e api client, CLI tooling) shouldn't have to plumb
the executor id through. Aliases the existing nested handler.
* test(e2e): probe SSH fingerprint via kandev, not ssh-keyscan, in K specs
regenerateHostKey was returning the ssh-keyscan fingerprint, but
ssh-keyscan and the Go ssh client can negotiate different host-key
types — meaning the fingerprint kandev observes during a real probe
does not always match what keyscan saw. The keyscan output was used
verbatim as the expected value, which K1 / K3 then asserted against
kandev's pinned-fingerprint result, so the specs flapped depending
on which algorithm sshd advertised.
Drop the return value from regenerateHostKey and have callers hit
/api/v1/ssh/test to learn the canonical fingerprint, matching what
the SSH fixture already does at seeding time.
K2 was polling env.error_message which doesn't exist — the launch
fails at SSH dial time before the executor is recorded on the task
environment. Point the polling at the task-session error_message
instead, where updateTaskSessionState actually persists the dial
failure.
* test(e2e): wipe agentctl before F5 to assert from a fresh-install state
F5 asserts that the agentctl_action probe reports "uploaded" before
the first launch and "cached" after. Tests share a worker-scoped
sshd container, so an earlier test in the same worker may have
already primed the cache, making the "before" assertion flap.
Wipe ~/.kandev/bin/agentctl + .sha256 from the container at the top
of F5 so the assertion sees the documented fresh-install path.
Export execInContainer so the spec can reach into the container
without copying the docker-exec boilerplate.
Add error_message to the listTaskSessions return type so the
hostkey-rotation spec can read the field (which the backend already
returns).
* fix(ssh): re-read ~/.ssh/config on every resolve, no process-wide cache
kevinburke/ssh_config's package-level Get/GetStrict use a sync.Once
to load $HOME/.ssh/config once per process and cache the result. That
breaks two things:
* users who edit their config mid-session don't see updates until
kandev restarts;
* e2e tests that write a fresh ~/.ssh/config per test all see the
state captured by whichever test triggered the first lookup.
Open and Decode $HOME/.ssh/config on demand from a local
inheritFromSSHConfig helper. The library still parses the file the
same way — we're just dropping the global cache.
Unblocks the four ssh/config-inheritance.spec.ts cases that wrote a
per-test ~/.ssh/config.
* feat(ssh): parse literal user@host:port ProxyJump bastion strings
dialViaJump fed target.ProxyJump straight into ResolveSSHTarget as a
host alias. That works for `Host prod-jump` entries in ~/.ssh/config
but turns the literal `user@host:port` form into a single hostname
that the kernel resolver chokes on (`lookup kandev@127.0.0.1:23000: no
such host`).
Detect the literal shape inline (`[user@]host[:port]`) and build the
bastion SSHTarget from its parts. Falls back to the alias path when
no @ / : separators are present.
Fixes the proxy-jump.spec.ts bastion flow. While here:
* scanFingerprintInsideNetwork: alpine package is `openssh-client`,
not the (nonexistent) `openssh-keyscan`. Surface install errors
so future failures aren't silent.
* scanFingerprintViaHost: retry on empty output — sshd accepts TCP
before host-key negotiation completes on a cold container.
* startBastionAndTarget: share one keypair across bastion + target
so a single -i auths both hops (matches how OpenSSH ProxyJump
handles per-host identities outside ~/.ssh/config).
* proxy-jump N1: assert any SHA256 fingerprint came back, since the
test no longer pre-discovers a value to compare against.
* test(e2e): time-bound the drop-traffic SSH probe with an AbortController race
iptables -A INPUT -p tcp --dport 22 -j DROP silently discards SYN
packets — Go's net.Dialer sits in the connect syscall through the
kernel's TCP_SYN_RETRIES (default 6 ~ 127s) before the syscall is
abandoned, even though the dialer's logical Timeout is 30s. The
spec polled the testSSHConnection API for the failure path; the
request blocked long enough that Node's keep-alive evicted the
socket and the test failed with "other side closed" instead of
seeing kandev's success=false response.
Race the probe against a 5s timeout. If the request errored or
timed out the test passes (both are valid evidence that the drop
hit the dial path); if it came back with a response, assert
success=false the same way the original spec did. The restore-probe
poll still gates on a real success after restoreTraffic.
* test(e2e): rerun ssh.test in hostkey-rotation afterEach to repin
Hostkey-rotation specs mutate the worker-scoped sshd container and
the seeded SSH executor's pinned fingerprint. Without a cleanup hook,
the next spec in the same worker that resolves seedData.sshTarget.
hostFingerprint observes a stale value and the UI's trusted-badge
check (persistence R1) flaps.
Add an afterEach that re-probes the live container via kandev, syncs
the fingerprint into both seedData and the seeded executor's pinned
config. The cleanup is a no-op when the probe fails (e.g. K2 left an
intentionally bad fingerprint that no longer matches anything — the
next test will re-establish state via its own setup).
* test(e2e): restore the seeded executor's pinned fingerprint after R5
The fingerprint-change-warning spec pinned a deliberately bogus
fingerprint on the worker's seeded SSH executor to exercise the
mismatch UI, then exited without resetting it. Later specs in the
same worker that launched real tasks against the same executor
(workdir-per-profile R1/R2 in alphabetical order) failed at SSH dial
time with host-key-changed, since the executor was still pinned to
the bad value.
Wrap the assertion in a try/finally that restores the original
fingerprint from seedData when the test ends, so the worker leaves
the executor in the same state it inherited.
* fix(ssh): quote every remote path + clean identity file path
CodeQL flagged two paths from HTTP-supplied SSH-config fields into
shell + filesystem operations:
* runSSHCommand built "cat $remoteShaFile" / "test -x $remoteBin"
/ "mkdir -p $(dirname $remoteBin)" by direct concatenation; the
arguments come from $HOME on the remote (not the kandev user)
but the analyzer can't see that, and a remote $HOME containing
shell metacharacters would still misparse. Route them through
shellQuote for parity with every other call site in the file.
* dialAuth opened target.IdentityFile straight from the configured
string. Run it through filepath.Clean so any `..` segments slipped
in by string substitution are normalised before os.ReadFile, which
is the canonical mitigation CodeQL recognises here. Behaviour is
unchanged for normal paths.
* refactor(ssh): drop the unused conn pool + remote curl dependency
CodeRabbit and Greptile both flagged the SSHConnPool: the executor never
took the pool path (CreateInstance always dialDirect'd a fresh client and
parked it on sshSessionState), so the pool's keepalive goroutines and
refcount logic were dead code. Get() also overwrote map entries on every
call, which would have leaked keepalive goroutines if the pool *had* been
used. Delete the entire SSHConnPool + sshConnKey + pooledConn + keepalive
plumbing. Export DialSSH so the SSH test endpoint can dial directly and
close the client when its single-shot probe is done.
While here:
* pickRemoteAgentctlPort: switch from time.Now().UnixNano()%20000 to
math/rand/v2.IntN. Two concurrent CreateInstance calls on a fast
machine were getting the same port (UnixNano cycles in ~20µs through
the 20k range); randomized picks avoid the collision.
* createRemoteAgentInstance: POST to the remote control port over a
direct-tcpip dial via net/http instead of shelling out to `curl`. The
sshd container no longer needs curl, the kandev test endpoint no
longer needs to probe for it, and the call inherits Go's standard
timeout + error handling. Drop `curl` from the e2e Alpine image.
* maybeUploadCredentials: was probing `printf %s "$HOME"` and
discarding the output, then allocating an unused sshFileUploader.
SSH v1 doesn't ship credential upload — just log a warning when a
user sets remote_credentials so they know it won't take effect, and
delete the dead uploader file.
* fix(ssh): verify bastion host key + tighten the trust-gate UX
Bastion host-key verification (security)
resolveBastion built the bastion's ClientConfig with an empty
PinnedFingerprint, which buildClientConfig translated into a
callback that accepts any key. An attacker on the path to the
bastion (DNS poisoning, route hijack) could pivot through it
before the final-target check ever ran. Greptile flagged it as P1.
Swap in a callback that consults ~/.ssh/known_hosts via
knownhosts.New: matched key → accept; mismatched key → reject loud
(the alarm); host absent → accept with the same TOFU semantics
OpenSSH uses for StrictHostKeyChecking=accept-new, so e2e and
first-connect setups still work without manual `ssh-keyscan`
priming.
Stale-test trust gate (frontend)
Editing a connection field after a successful test previously kept
the prior result + cleared `trust`, but the trust checkbox could
be re-checked and saved with the old fingerprint bound to the new
target. CodeRabbit + cubic both flagged it as P1.
Add a `resultStale` flag that flips when a CONNECTION_FIELD edit
follows a successful test, surface a warning under the result, and
disable the trust checkbox + save button until the user re-runs
Test Connection. handleTest clears the flag.
While here
* Surface onSave failures in the card instead of swallowing them.
* ssh-create-page: read the executors slice from the live store
(useAppStoreApi + getState()) at write time so a concurrent
update doesn't get overwritten by a stale closure.
* Split FingerprintTrustBlock into its own file to keep
ssh-connection-card under the 600-line cap.
* fix(ssh): defend launch + handshake against several real bugs
manager_launch: trusted exec-config keys override req.Metadata.
buildLaunchMetadata applied executor config only when the matching
key was missing in req.Metadata. A request that carried metadata.
ssh_host_fingerprint (or ssh_host, ssh_identity_*) could pivot the
launch to a different target and bypass the pinned fingerprint.
Pin a list of connection-routing keys (ssh_host, ssh_port, ssh_user,
ssh_host_fingerprint, ssh_identity_source, ssh_identity_file,
ssh_proxy_jump) and have the executor config always win for those.
Cubic / CodeRabbit P1.
executor_ssh.CreateInstance: reuse the resumed session state.
ResumeRemoteInstance attaches the recovered ssh.Client + forwarder
to r.sessions[InstanceID] and refreshes metadata, then the lifecycle
manager unconditionally called CreateInstance, which dialed a fresh
client and started a second remote agentctl on top of the live one
— leaking the resumed state. CreateInstance now early-returns from
the existing r.sessions entry via buildResumedInstance. Cubic P1.
executor_ssh.Close: actually release tracked sessions.
After dropping the SSHConnPool, Close was a no-op. A hard kandev
exit (no per-session StopInstance) left forwarders and SSH clients
open. Iterate r.sessions and Close the forwarder + client for each.
Cubic P2.
buildAuthMethods: close the ssh-agent unix socket after handshake.
Each agent-authenticated dial opened net.Dial("unix", $SSH_AUTH_SOCK)
and never closed it; only PublicKeysCallback held the agent.Client
reference and that callback is consulted exactly once at handshake.
Return a cleanup func from buildAuthMethods + buildClientConfig and
defer it from dialSSH / dialViaJump so the socket releases as soon
as key exchange completes. Greptile P1 security.
waitAgentctlHealthy: probe HTTP /health, not just the TCP listener.
The local port is owned by the SSH forwarder, which Accepts before
it Dials the remote direct-tcpip channel — a TCP connect can succeed
before the channel actually establishes. Issue a real GET /health
and require a 2xx so a broken tunnel surfaces here rather than at
the first agent operation. Cubic P1.
ssh-connection-card: also gate the trust + save on in-flight tests.
Editing a connection field during a pending test silently kept
resultStale=false because prev.result was still null; when the test
returned, the new fingerprint was bound to the old form values but
the gate let the user trust + save anyway. Treat in-flight tests as
prior tests for staleness purposes, and preserve resultStale across
the test's success setState. Cubic P2.
* fix(ssh): drop stale polls + preserve port inheritance + sshd extra-keys
ssh-sessions-card: race-protect polls.
Polling setInterval + manual refresh shared one state setter, so a
slow request that resolved after a newer call (or after a parent
switched executorId) overwrote fresh data with stale rows. Bump a
monotonic sequence on each refresh and drop the response when the
sequence has moved on, plus reset on mount and invalidate on
unmount so an in-flight reply from a teardown can't land. Cubic P2.
ssh-config parseSSHExecutorConfig: don't default ssh_port to 22.
Reading an executor row that had no ssh_port set rewrote port=22
on the form, which then round-tripped back to the backend. That
silently broke ~/.ssh/config Port inheritance for aliases — a save
could redirect connections to 22 even though the underlying alias
was supposed to resolve to another port. Leave the field undefined
so the backend resolver can apply the alias' value. Cubic P2.
ssh-image entrypoint: guard authorized_keys.extra mount.
When only /authorized_keys.extra (and not /authorized_keys) was
bind-mounted, the >> redirect ran under set -e against a missing
destination and aborted container startup. mkdir + touch the
destination before the append and reapply the ownership/perms.
CodeRabbit P2.
* test(e2e): make G2 + F6 actually exercise the contract
G2 (sessions-endpoint filter) iterated `for (const s of sessions)` over
the listSSHSessions result without first asserting `length > 0` — an
empty response let the test pass without checking anything. Seed an SSH
launch before the loop and require at least one row, so a handler
regression that returned [] would now fail loud. CodeRabbit P1.
F6 (ws ssh.test parity) called apiClient.testSSHConnection twice (HTTP
both times) and compared the two HTTP responses, which would have
passed even if the WS dispatcher entry was broken. Use apiClient.
wsRequest("ssh.test", payload) for the second call so the gateway's
WS handler is actually invoked, and assert the step names match across
transports. CodeRabbit P1.
* refactor(e2e): extract SSH HTTP shapes into api-client-ssh.ts
api-client.ts was already over the 600-line lint cap on main (1528
lines); the SSH additions made it worse. Split the SSH HTTP shape
types into a dedicated module and re-export from api-client.ts so
existing callers keep working. The methods themselves stay on
ApiClient as thin pass-throughs since splitting them out across the
class boundary would change every test call site.
Also document the per-executor bound on the SSH listSessions per-row
GetTaskSession lookup (typical N < 10), and what a batch path would
need (a repo method to fetch task_sessions by id list). Cubic /
Greptile P2.
* docs(spec): correct SSH recovery keying after pool removal
The spec said ResumeRemoteInstance re-opens 'per (host, user) — one
connection serves all surviving sessions on that host', which was
the pool design. The current executor gives each session its own
SSH client, so the resume keys on the full (host, port, user,
identity_*, proxy_jump, host_fingerprint) tuple. That way two
profiles on the same host with different keys can resume side by
side without merging. Cubic P2.
* fix(ssh): IPv6 ProxyJump + port range + dedup + docker timeouts
parseLiteralProxyJump: handle bracketed IPv6 + reject invalid ports.
The literal parser used LastIndex(":") to split host from port, which
for [2001:db8::1]:22 left host as "[2001:db8::1]" and produced an
invalid double-bracket address through net.JoinHostPort. Recognise
the bracketed form explicitly via a parseBracketedHostPort helper
and strip the brackets. Validate port is in [1, 65535] instead of
accepting any integer (Atoi happily parses negative or out-of-range
values). Split the host/port branch into its own function so the
nesting stays under the golangci nestif limit.
ssh-create-page: dedupe on append.
setExecutors([...current, next]) would double-list the same executor
if a WS event (or another refresh) inserted it while createExecutor
was in flight. Skip the append when current already has the id.
ssh-image fixture: time-bound docker probe + build.
hasSSHContainerSupport ran `docker info` with no timeout — an
unreachable / hung daemon would stall the Playwright worker
indefinitely. Cap at 10s for the probe. buildE2ESSHImage runs
`docker build` with no timeout; cap at 15 min so a stuck registry
pull or BuildKit hang fails fast.
* fix(ssh): reject invalid ssh_port + retry in-net keyscan + restore D5 + doc count
executor_ssh.targetFromMetadata: reject malformed / out-of-range
ssh_port instead of silently falling back to 0 → 22 in the resolver.
"abc", "0", "70000" all now return a "must be 1-65535" error so a
broken config can't masquerade as a successful launch routed through
port 22. Cubic P2.
scanFingerprintInsideNetwork (bastion fixture): retry on empty output.
Mirrors the host-side scanFingerprintViaHost retry loop — sshd accepts
TCP before host-key negotiation completes on a cold container, so the
keyscan can come back empty on the first try. Bound at 15s; surfaces
"status=..., stderr=..." on final failure for diagnosis. Cubic P2.
executor-crud D5 (changing config…): restore the seeded executor in a
try/finally. The test pinned the executor to host=10.255.255.1 to
verify the snapshot stayed sticky on the live session, but never put
the real host back, so any later spec in the same worker that reused
seedData.sshExecutorId would dial the unroutable address and hang.
Cubic P2.
e2e-plan.md: replace the "roughly 75" estimate with the actual 80 /
17-files current count plus the ~2.4 min wall-clock. Cubic P3.
* fix(ssh): use bare mock-agent name so the remote PATH resolves it
For SSH the agent runs on a remote host whose filesystem doesn't share
paths with the kandev backend's, so the host-absolute mock-agent path
configureMockAgent computes can't apply. Extend the bare-name branch in
MockAgent.BuildCommand to include RuntimeSSH (the SSH e2e image already
bakes mock-agent into /usr/local/bin) and drop the now-redundant SSH e2e
bind-mount that worked around this by mirroring the host path into the
container.
* fix(ssh): unique sftp temp filename so concurrent uploads don't collide
When two sessions on the same task launch back-to-back they both call
ensureAgentctlOnHost which streams to <remotePath>.tmp before renaming.
With a shared .tmp name the second uploader's rename errors with
"file does not exist" once the first uploader's rename consumed the temp
file. Give each uploader its own pid+random-suffix temp; uploaders write
identical sha256-keyed content, so the rename race is then safe.
* fix(ssh): don't leak session-scoped metadata across sibling sessions
reuseExistingEnvironment copies persistent metadata from the previous
ExecutorRunning row into the new launch request so executors can pick up
worktree/container/sandbox IDs. For SSH that swept in session-scoped
keys too — agentctl PID/port, session dir, local forward port — which
made the second session's ResumeRemoteInstance reattach to the first
session's agentctl process. Mark those keys as session-scoped and skip
them in applyExecutorRunningMetadata; the same-session resume path
(applyRunningRecordToResumeRequest) still copies them because it works
off the session's own ExecutorRunning record.
* test(ssh): cover ProxyJump parsers + sibling-session metadata filter
Addresses the two coverage gaps Claude flagged on PR review round 3:
- parseLiteralProxyJump / parseProxyJumpHostPort / parseBracketedHostPort
/ parsePortString are pure parsers that gate every ProxyJump dial,
including IPv6 (`[2001:db8::1]:22`). Table-driven tests pin the happy
paths and the bracket / port / range edge cases — silent regressions
in bracket stripping would otherwise break every IPv6 ProxyJump
session without a single test failing.
- applyExecutorRunningMetadata is the function commit
|