A handler session only learned the chain was capped when handoff_list_handler
threw at the end of its work, stranding open tasks. The round count already
lived in HandoffRoundTracker the whole time -- it just never reached the
brief. InteractiveLaunchSpecService now reads HandoffRoundTracker.GetState
(read-only, doesn't advance the round) and renders round N of MaxHandoffs
plus the remaining count into the handoff brief, replacing the old one-line
finalNote. The rejection error also names the next action (start a new
handler run) instead of just failing.
`src/ClaudeDo.Worker/External/ExternalMcpService.cs` ist **1897 Zeilen** (nachgezählt 2026-09-02; die Datei ist seit dem Anlegen dieses Tasks von 1696 weiter gewachsen): ~30 DTO-Records oben plus ~25 Tool-Methoden — während **13** Geschwisterdateien `*McpTools.cs` im selben Ordner das Topic-Split-Muster schon tragen (`ListMcpTools`, `ConfigMcpTools`, `BatchMcpTools`, `NotesMcpTools`, `QueueStateMc
ClaudeDo-Task: 1862e160-ffde-4a50-aa03-bb7251f56088
From `docs/open.md` (re-verified against the code 2026-09-02, still open): nothing reconciles `git worktree list` against the `worktrees` table. `ExternalMcpService.ListWorktrees` and the Worktrees overview read DB rows only; `WorktreeManager` calls git solely inside its self-heal path (`ListWorktreePathsForBranchAsync`). Observed once: 22 git-registered worktrees + 26 `claudedo/*` branches in `C:
ClaudeDo-Task: 765fed13-db66-44e4-9285-c4b2d415b443
Agents can write into the user's daily notes, but only within limits:
max 10 agent notes a day, max 500 characters, add only — no edit, no
delete. A new from_agent column separates their lines from the user's
so the cap counts the right ones and the UI can mark them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
IBaseDirtyChecker, IInteractiveLaunchSpecService and the LaunchSpec wrapper
had one implementation and one caller each; ProcessRunnerAdapter existed only
to give a static class an interface, and InstallArtifactLocator used
inheritance for two constructor arguments. The external MCP container now
shares its singletons through a Share<T> helper instead of 17 near-identical
registrations.
Set Stateless = true on the internal (/mcp on the SignalR port) and external
MCP HTTP transports. Two effects:
- No Mcp-Session-Id, so a worker restart can no longer 404 a session that
outlives it -- ConPTY tiles in the UI process and externally registered
claude sessions keep working across a restart.
- A 2026-07-28 client is no longer refused back to the initialize handshake.
A stateful server rejects that revision on purpose (see the SDK's
StreamableHttpHandler.s_sessionSupportingProtocolVersions), so the new
per-request protocol path was unreachable regardless of SDK version.
Nothing here used the stateful-only features (sampling, elicitation, resource
subscriptions, unsolicited notifications). In-tool progress notifications ride
the POST's own response stream and are unaffected -- which matters, since ~20
tools use them to hold off the client's 300s idle abort. All three consumers
(TaskRunner, PlanningSessionManager, the installer's RegisterMcpStep) already
register type: "http", so dropping the legacy SSE endpoint breaks nothing.
Verified against the running worker across two restarts: tools/list returns
57 tools over a bare POST with no initialize and no session id, and a full
2026-07-28 tools/call round-trip returns real data.
Nine interfaces had exactly one implementation and no test double — they existed
only to be named twice in a DI registration: IFindingsStore, IFindingsStoreLocator,
IPrimeScheduleSignal, IRefineRunner, IWeekReportService, IMergeCoordinator,
IMissionControlPane, IOnlineLoginService, ITaskListFilter. Consumers now depend on
the concrete type; the DTO records that shared those files moved next to their
implementation. IInteractiveLaunchSpecService stays — it carries 54 lines of
contract documentation, which is not ceremony.
IMergeCoordinator in particular had a redundant null object: MergeCoordinator with
a null Handler already no-ops, and every test used the real class with Handler set.
Filtering/ collapses from 8 files to 1. ITaskListFilter and TaskListFilterBase were
a double abstraction over four predicates, with MatchesAsContext => false declared
in both. SmartFlagFilter also compiled its expression twice (its own _flag plus the
inherited Matches cache) — it now uses the cache.
StaticTokenAuthProvider was in src but production uses ZitadelAuthProvider; it is
a test double, so it moves to the test project. Its own test goes away with it.
Klassifiziert einen echten Usage-Limit-Abbruch als eigene FailureReason
"usage_limit" (TaskRunner.ClassifyFailureReason: nur bei terminal_reason
"api_error" plus einem Limit-Muster im gerenderten Fehlertext, nicht an
Status==Failed allein). Neuer Toggle AutoContinueOnUsageLimit (app_settings,
Default aus) unter Settings → General → "Usage limit stop":
- UsageLimitAutoContinueCoordinator feuert pro Task genau einmal ContinueTask
über OverrideSlotService, sobald das 5h-Fenster (UsageState.Snapshot.FiveHour
.ResetsAt) tatsächlich zurückgesetzt ist; ein persistenter Marker
(TaskEntity.UsageLimitAutoContinuedAt) verhindert einen zweiten Anlauf bei
einem erneuten Limit-Treffer.
- QueueService schedult zusätzlich einen exakten Wake-Timer auf den
Reset-Zeitpunkt, statt nur auf den 30s-Backstop zu warten.
- Fail-open durchgängig: kein Snapshot/keine Reset-Zeit → kein Timer, kein
Continue, kein Throw. Toggle aus ändert das heutige Verhalten nicht.
Migration AddUsageLimitAutoContinue fügt beide Spalten hinzu; die von
`dotnet ef migrations add` mitgescaffoldete leere UpdateData auf app_settings
(columns/values: []) erzeugte ungültiges SQL ("near WHERE") und wurde entfernt
— TaskNumberMigrationTests deckte das über den vollen Migrate()-Pfad auf.
Teil A: update_task bekommt einen optionalen isManual-Parameter (null =
unveraendert); TaskDto/TaskRefDto spiegeln IsManual zurueck.
Teil B: neuer submit_task_for_review MCP-Tool (LifecycleMcpTools) laesst
einen Handler-Task selbst auf WaitingForReview gehen -- fuer einen
worktree-losen Task wird HandlerHeadCommit vom aktuellen HEAD gestempelt,
sonst werden offene Aenderungen committet. Die Submit-Logik ist aus
WorkerHub.SubmitTaskForReview in InteractiveReviewSubmissionService
extrahiert; der Hub ruft sie jetzt nur noch auf. Der Merge-Prompt-Default
weist die Endzweige (merge ohne Rerun, merge_final) an, den eigenen
Handler-Task nach der Summary einzureichen.
The six Lifecycle/*Recovery hosted services now broadcast one
OperationProgress("startup-recovery", <phase>, current, total) message each
after they finish, instead of leaving the UI on a bare "connecting" text
during worker startup. IslandsShellViewModel subscribes and swaps in
"Recovering... (i/n)" (existing ops.worker.startupRecovery key, no locale
changes) while Worker.IsReconnecting is true, and clears it once actually
connected so a later transient reconnect doesn't replay stale text.
OperationProgress broadcasts to Clients.All with no replay-on-connect, so a
UI that hasn't finished its SignalR handshake yet can miss some or all of
these messages and simply keep showing "connecting" as before -- accepted
rather than adding a cached-state + reconnect-replay path (mirroring
RefreshExternalMergeConflictsAsync) for what is a fast, best-effort,
local-only startup sweep with no UI-visible failure mode beyond that.
HandoffMcpTools.HandoffListHandler only validated the phase name and
broadcast it -- nothing bounded how many times a ConPTY session could
hand off wait<->merge, so a model that skimmed past the prompt's
"final round" line could loop indefinitely. Add HandoffRoundTracker,
an in-memory per-handler-task counter (a list-handler run always
creates a fresh handler task, so no reset logic is needed): past 4
handoffs (two full wait/merge cycles) a non-final nextPhase is
coerced to its "_final" counterpart, and any handoff after a
merge_final round for that task is rejected outright.
Queuing (update_task_status, batch_update_task_status) and run_task_now
now surface a non-blocking baseDirty warning (separate modified/untracked
counts) when the list's working dir has uncommitted changes at enqueue
time, since a new worktree forks from the commit tip and silently misses
them. BaseDirtyChecker caches per working dir for a few seconds so a
batch queue over many tasks in one list only shells out to git once. The
UI surfaces the same warning via the footer error strip on queue actions.
Path.GetDirectoryName returned workingDir itself instead of its parent
whenever list.WorkingDir ended in a directory separator, placing
.claudedo-worktrees inside the target repo's own working tree. Add
WorktreeRootResolver to normalize the trailing separator before deriving
the sibling root and to guard that the resulting worktree path never
lands inside workingDir, used by both WorktreeManager and
PlanningSessionManager. Add a startup sweep (LegacyWorktreeFolderRecovery)
that warns when a list's working dir already contains a leftover
.claudedo-worktrees folder from before the fix.
TaskStateService.CancelAsync allowed cancelling a WaitingForReview task
even while PlanningMergeOrchestrator was mid-drain on it: ApproveReview
awaits the whole multi-subtask merge synchronously, so a concurrent
CancelReview (UI or MCP) could flip the parent to Cancelled while the
orchestrator kept merging children onto the target branch, then
FinalizeParentDoneAsync would find the parent no longer WaitingForReview
and give up - leaving the merged diffs stranded with no rollback.
CancelAsync now rejects with a clear reason when HasActiveMerge(taskId)
is true. TaskStateService can't take PlanningMergeOrchestrator as a
direct constructor dependency (circular back to ITaskStateService), so
it takes a lazily-resolved Func<IActiveMergeState> instead, mirroring
the existing Func<ITaskStateService> cycle-break already used for
PlanningChainCoordinator.
UI polish: DetailsIslandViewModel.IsMergeDraining gates
CancelReviewCommand's CanExecute (same shape as
WorktreesOverviewModalViewModel.IsMerging), and the command's catch now
raises ErrorReported instead of swallowing the rejection silently.
MaxTimeoutSeconds was 170s against runs that take tens of minutes, forcing
a dozen full-context wait rounds per long-running batch. Raise it to 900s
and raise MCP_TOOL_TIMEOUT in lockstep (ClaudeProcess + every
InteractiveLaunchSpecService launch spec) to 930000ms so the client
connection actually stays open that long instead of aborting first.
Add get_queue_state (QueueStateMcpTools): configured vs. effective
parallel-slot count (via QueueService.GetSlotCountsAsync, extracted from
the former GetEffectiveMaxParallelAsync), active slots with taskId +
startedAt including the run_task_now override slot, and queued tasks in
pick order -- so a caller can observe queue occupancy instead of inferring
it from maxParallelExecutions.
EF Core logs every SQL statement at Information, which BroadcastLogSink
buffers into LogRingBuffer regardless of level, flooding the Log
Visualizer overlay and worker log file with SQL noise that buries
hand-curated business events. Override both to Warning so real
EF Core / ASP.NET Core problems still surface.
The merge-helper ("Let Claude handle it") system prompt now calls a new
handoff_list_handler MCP tool once every surviving task is enhanced,
instead of continuing into Phases 3-5 in the same session -- avoiding
paying for Phases 0-2's dedupe/rewrite context on every polling round of
the run/review/merge phases.
The tool broadcasts HandoffRequested; Mission Control opens a second
ConPTY tile for the SAME handler task id (no new task, HandlerBaseCommit
untouched) running a fresh handoff brief that starts at Phase 3. The
original tile stays open. Adds PromptKind.MergeHelperHandoff,
InteractiveLaunchSpecService.BuildForMergeHelperHandoffAsync, and the
GetMergeHelperHandoffLaunchSpec hub method.
The usage monitor polled the undocumented OAuth usage endpoint every 60s and
earned 429s. It now polls every 5 min while any task is Running and every
15 min while idle (usage_poll_interval_active_seconds / _idle_seconds, both
clamped to >= 60; the old single usage_poll_interval_seconds key is gone).
A 429 comes back as UsageRateLimitedException carrying Retry-After and adds
exponential backoff on top, capped at 30 min and never shorter than the normal
cadence; the strike count resets on the first success. The schedule arithmetic
is the pure static UsagePollSchedule.NextDelay.
Since the idle cadence is slow on purpose, WorkerHub.RefreshUsage drives
UsageMonitorService.RefreshNowAsync behind a Refresh now button in the Usage
Monitor modal: an out-of-band poll that pushes the loop's next-due time out so
no double poll follows, with a 10s cooldown so click-spam can't earn a 429.
Staleness now measures against the slower (idle) interval so an idle worker
isn't flagged stale just for not polling.
EnsureExists blindly seeded ~/.todo-app/prompts/*.md with the bundled
default and never revisited it, so any file created by opening the
Files settings tab shadowed every later default change permanently
(SuggestImprovement/AskUser sections never reached real runs since
2026-06-04). PromptFiles now hashes what a file was seeded/saved with
in prompts/.defaults.json: Classify() tells missing/current-default/
known-past-default/edited apart, ReconcileStaleDefaults() drops files
that only ever matched a superseded default, and real edits are left
alone and surfaced in the Files tab with a diff + reset-to-default
action. QuarantineOrphans() moves stale-named leftovers (agent.md,
planning.md) into prompts/_orphans instead of silently deleting them.
Wired as a Worker startup sweep (PromptFileRecovery) alongside the
existing OrphanRecovery/AttachmentOrphanRecovery services.
Adds GetUsageSnapshot/GetModelUsage/GetTaskUsage to WorkerHub (backed by a
shared UsageSnapshotBuilder), a UsageUpdated broadcast fired after every
UsageMonitorService poll cycle, and records the resolved model on each
task_runs row so per-model/per-task usage can be reported from history.
Per-list optional VerifyCommand (list_config.verify_command) runs via
VerifyCommandRunner in the list's working dir right after a successful
merge/continue-merge, before the task is allowed to reach Done. A
non-zero exit or timeout leaves the merge in place but keeps the task
out of Done and reports StatusVerifyFailed with an output excerpt
through MergeResultDto/review_task; no command configured behaves
exactly as before. Merges against the same repo are now serialized
per working dir so a running verify can't be interrupted by a second
merge landing mid-build. Adds the field to the List Settings modal
(en/de localized) and covers success/failure/timeout in
TaskMergeServiceTests + VerifyCommandRunnerTests.
> **Stand 2026-08-05 (List-Handler):** Der Roadblock aus dem letzten Lauf ist erledigt. Beide Voraussetzungen sind jetzt auf `main` gemerged: die `app_settings`-Schwellen `UsageGateFiveHourPct`/`UsageGateSevenDayPct` (Merge-Commit `b1efcdc`) und `UsageState`/`IUsageClient`/`UsageMonitorService` unter `src/ClaudeDo.Worker/Usage/` (Merge-Commit `b126a21`). Dein Worktree ist frisch von diesem `main`
ClaudeDo-Task: 06a7cc32-6ab7-4758-98f4-bee77149b2bf
Replaces the list handler's Start-Sleep + blind get_task poll (Phase 3
of the merge-helper prompt) with a blocking MCP tool that returns as
soon as a task leaves Queued/Running, or times out. Implemented as an
async DB poll (short-lived DbContext, 500ms delay, no held connection)
rather than hooking HubBroadcaster, keeping the existing broadcast
callers untouched. timeoutSeconds is clamped server-side to 170s, under
the list handler's 200s MCP_TOOL_TIMEOUT.
Aggregates ~/.claude/projects/**/*.jsonl assistant messages by date/model/
scope (ClaudeDo worktree-or-sandbox cwd vs Other), deduped by requestId
(falls back to message.id), with a per-file length+mtime cache so repeat
calls skip unchanged files.
Adds Usage/ with ClaudeOAuthUsageClient (reads the access token Claude
Code keeps fresh at ~/.claude/.credentials.json, calls the Anthropic
OAuth usage endpoint, defensively parses buckets/limits), UsageState
(threadsafe last-good-snapshot holder that never regresses on
failure), and UsageMonitorService (BackgroundService polling on the
new usage_poll_interval_seconds config, min 15s, one poll at startup,
warns at most once per distinct error).
- CancelAsync now signals the running Claude process of the cancelled task and
its cascaded children via the new RunCancellationRegistry (queue + override
slots register their CTS there) instead of only flipping DB state.
- external MCP review_task 'approve' now mirrors the hub's ApproveReview:
unit merge for parents, ApproveAndMergeAsync for childless tasks, optional
targetBranch; ReviewTaskResult carries mergeStatus/conflicts.
Nullable filter/patch params across the External/ MCP tool classes (ListTasks,
UpdateTask, AddSubtask, ReviewTask, SetMyDay, SetListConfig/SetTaskConfig,
CreateList/UpdateList) lacked C# default values, so the generated tool schema
marked them required — MCP clients omitting them (the common case) failed.
Gave every such parameter a default value.
Also registered a call-tool filter (ExternalMcpExceptionFilter) on the external
MCP host that translates InvalidOperationException/ArgumentException into
McpException, since the SDK's own catch-all discards ex.Message for any other
exception type and returns a generic "An error occurred invoking 'X'." string.
Added a reflection-based schema test sweeping every [McpServerToolType] class
to guard against reintroducing a required-but-nullable parameter.
The embedded ConPTY terminal replaced the in-app streaming interactive session, so
delete the dead stack: StreamingClaudeSession, InteractiveSessionService,
ProcessClaudeStreamTransport, IClaudeStreamTransport, ILiveSession, LiveSessionRegistry,
IdleSessionReaper (+ WorkerConfig.InteractiveIdleTimeoutMinutes), the WorkerHub
interactive methods + HubBroadcaster events, IWorkerClient interactive members, the
TaskMonitorViewModel composer + SessionTerminalView composer markup, and the old
'Run interactively' entry. AskUser/PendingQuestionRegistry, the autonomous path,
planning, ResumeTaskInTerminal, and all ConPTY code are kept. Localization pruned.
Adds InteractiveLaunchSpecService + GetInteractiveLaunchSpec hub method that
prepares a task worktree (session-skills seeding, run env) and returns a
LaunchSpec {cwd,exe,args,env} for a UI-hosted ConPTY terminal. Reuses
ISessionSkillSeeder, TaskRunner.UnionSkillNames, and WindowsTerminalLauncher
resume-arg/resolve logic. Guards mirror ResumeTaskInTerminal; a never-run task
yields a fresh-start spec instead of an error.
Interactive/streaming sessions are persistent claude.exe processes that
wait on stdin and never exit on their own. The only teardown was an
explicit StopInteractiveSession from the UI — there is no client-disconnect
or shutdown sweep — so an abandoned chat (UI closed, navigated away,
crashed) kept its claude.exe (+ conhost) alive for the worker's whole
lifetime. Under a long-running autostart worker these accumulate to dozens
of orphaned child processes.
LiveSessionRegistry now tracks per-session activity (Touch on every output
line and user action) and exposes ReapIdleAsync, which stops sessions idle
past a timeout while skipping any with a turn in flight. IdleSessionReaper
(BackgroundService) sweeps every 5 min; idle timeout defaults to 30 min,
configurable via interactive_idle_timeout_minutes (0 disables).
InteractiveSessionService resolves a task's list working dir + seeded prompt,
spawns a StreamingClaudeSession (claude stream-json in the list dir, model+auto
as before), registers it in LiveSessionRegistry, streams output over TaskMessage,
and broadcasts InteractiveSessionStarted/Ended (an exit watcher fires Ended). The
hub's OpenInteractiveTerminalAsync now starts this in-app session; SendInteractiveMessage
and StopInteractiveSession route to it. The external Windows-Terminal interactive
launch (LaunchInteractiveAsync / InteractiveLaunchContext / OpenInteractiveAsync) is
removed; planning sessions keep their terminal launch.
A running task can call mcp__claudedo_run__AskUser(question) to block (up to 3
min) on a human answer. PendingQuestionRegistry holds the pending question +
TaskCompletionSource; the tool broadcasts TaskQuestionAsked, awaits the answer
(WorkerHub.AnswerTaskQuestion resolves it), and returns it as the tool result —
or a 'proceed on your judgment' fallback on timeout. The run stays Running
throughout (no status/schema change). ClaudeProcess raises MCP_TOOL_TIMEOUT so
the 60s HTTP-MCP cap doesn't kill the wait; the run MCP is now wired for every
task, not just standalone ones. System prompt updated to reconcile 'unattended'.
Add seven best-effort batch variants of the single-entity external MCP
tools: batch_get_tasks, batch_add_tasks, batch_update_task_status,
batch_cancel_tasks, batch_delete_tasks, batch_set_my_day, and
batch_cleanup_task_worktrees. Each loops the existing ExternalMcpService
methods sequentially (scoped DbContext is not thread-safe), returns a
per-item result array so a failing item never aborts the rest, and
rejects empty or over-100-item batches. Merge/review stay single-task.
wt.exe treats ';' as a command/tab delimiter in every argument, with no escape
that survives quoting (microsoft/terminal#13264), so a task description
containing ';' spawned extra terminals on "Run interactively" and planning start.
Route the launch as wt -> powershell -> claude and pass the free-text prompt via
$env:CLAUDEDO_LAUNCH_PROMPT so it never reaches the wt command line; PowerShell
binds the variable as a single argument (embedded quotes escaped for PS 5.1).
Also clarify the launcher, which serves interactive runs too (not just planning):
IPlanningTerminalLauncher -> ITerminalLauncher, WindowsTerminalPlanningLauncher ->
WindowsTerminalLauncher, LaunchStart/Resume -> LaunchPlanning{Start,Resume}Async.
Mika Kuns