## Symptom (real aufgetreten, 2026-08-11)
Ein `review_task(decision="approve")` über MCP lief >5 Min. Claude Codes MCP-Client brach den Call ab mit:
> MCP server "claudedo" tool "review_task" sent no response or progress for 300s; aborting.
Ergebnis: der Merge war **schon gelaufen und comitted** (`0e12be4`, `mergeCommit` am Worktree gesetzt), aber der Task blieb auf `WaitingForReview` hängen —
ClaudeDo-Task: d8199f1f-3df3-447f-8de2-e7aca9ec5064
wait_for_task_change resolved #<number>/bare-number ids up front via TaskIdResolver, which
throws for an unknown number -- breaking the tool's own documented promise that an unknown
id reports status "NotFound" instead of failing the whole call. Resolve per id and fall back
to the original id on a resolution failure so CheckOnceAsync can still report it.
TaskNumberAllocator.AddWithNumberAsync indexed into the app_settings UPDATE...RETURNING result
without checking for an empty result, throwing on a missing singleton row; it also caught any
DbUpdateException as a number collision, burning up to 5 numbers on an unrelated failure (e.g.
FK violation) before the real error surfaced. Now recreates the missing row and only retries on
the actual unique-index collision (SQLite error 19 on tasks.number), rethrowing everything else
immediately.
Audited the other TaskIdResolver.ResolveAsync/ResolveManyAsync call sites (ExternalMcpService,
HandoffMcpTools, ConfigMcpTools, RunHistoryMcpTools, AttachmentMcpTools, LifecycleMcpTools,
BatchMcpTools): none of their tool descriptions promise a found/NotFound flag for the id itself
(BatchMcpTools.BatchGetTasks already handles this correctly via its own per-id try/catch;
PreviewMergeSet promises a per-task "error" field, not a found/NotFound flag; the rest are
single-id tools that already throw on a missing task downstream) -- left throwing behavior as-is.
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.
TaskIdResolver resolves a #123/bare-123 taskId parameter to its GUID
before any lookup, across every External/ MCP tool that takes a task
id, including the batch tools' id arrays (via delegation to the
already-resolving single-entity methods) and update_task's
dependsOnTaskId (empty string still passes through unchanged as the
clear-link sentinel). An unknown number throws a clear error instead
of a silent null. McpToolDocs.TaskNumberHint tells the agent to refer
to tasks as #<number> when reporting to the user, added to the
description of get_task, list_tasks, add_task, update_task_status and
review_task.
Adds Number alongside every task id in External/'s DTOs -- the two
central mappers (ToDto/ToRefDto -> TaskDto/TaskRefDto) plus every
DTO that carries a bare task id and bypasses them (batch results,
queue state, wait-for-change, config, attachments, handoff,
lifecycle, merge-preview-set, worktree list). Input resolution
(#123 as an argument) stays for slice 3.
TaskEntity.Number is a global, monotonically increasing, never-reused
integer (displayed as #123), allocated from AppSettingsEntity.NextTaskNumber
via a single UPDATE...RETURNING statement rather than MAX(number)+1, which
would reissue a deleted task's number. Both insert paths (TaskRepository.
AddAsync and CreateChildAsync) route through the new TaskNumberAllocator,
with a bounded retry on a unique-index collision. One migration adds the
columns, backfills existing rows in creation order, and creates the unique
index afterwards. Data-layer only; MCP/UI wiring is later slices.
Threads a nextPhase parameter (wait/merge/wait_final/merge_final, validated
by the new MergeHelperPhase) from handoff_list_handler through
HubBroadcaster/WorkerHub into InteractiveLaunchSpecService, which now picks
the next session's system prompt (MergeHelperWait/MergeHelperMerge) and
model (HandlerWaitAlias/HandlerMergeAlias) from it instead of hardcoding the
old two-phase Execute prompt -- this also fixes a build break left by the
prior prompt-split task, which removed PromptKind.MergeHelperExecute without
updating its only caller.
Also sets --model/--effort/--permission-mode explicitly for every list-handler
session (Triage included) via PermissionModeResolver instead of inheriting the
CLI's ambient model and hardcoding "auto", and adds Task to the merge-helper
allowlist so the Merge phase can delegate diff reviews to subagents.
SetListConfig previously could only pass VerifyCommand through unchanged;
only the UI hub could set it. Add an optional verifyCommand parameter with
matching clear/merge/create-vs-delete handling, and split the list result
into ListConfigDto so the task-only SetTaskConfig/GetTaskConfig DTOs stay
untouched. Also corrects Worker/CLAUDE.md's claim that tasks can override
verify_command — it's list-only.
Two branches merged into main independently: one added `runVerify` to
PreviewMergeSet ahead of `cancellationToken`, the other added the subset
test that passed the token positionally. Neither conflicted textually, so
the mismatch only surfaced as CS1503 at test-compile time.
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.
MCP_TOOL_TIMEOUT (raised by ClaudeDo's launchers) is a wall-clock cap unrelated
to Claude Code's idle-silence abort (default 300s for HTTP-transport MCP
servers), which no launcher raises. A wait near the 900s recommendation was
silently killed at ~300s in any session, launcher or not.
WaitForTaskChange now reports MCP progress every 30s while polling, which
resets that idle timer. Verified against a real claude -p call (no launcher
env overrides) surviving a 341s wait via 60s pings -- the same silence that
previously aborted at 300s. Tool description and the list-handler prompt no
longer claim ClaudeDo's launchers guarantee a long wait survives.
get_task/batch_get_tasks now return failureReason (max_turns|timeout|error|
cancelled|unknown) plus failureTurnsUsed/failureMaxTurns on a Failed task, so
max_turns (worktree usually fine, continue_task) is distinguishable from a
real error (reset_failed_task) without pulling get_task_log's raw NDJSON.
Classified and stamped onto TaskEntity by TaskRunner.MarkFailed via
TaskStateService.FailAsync; TaskRunEntity also keeps the CLI's raw
terminal_reason/result_subtype/errors for deeper diagnosis. reset_failed_task's
description now warns explicitly that it discards the worktree and points at
continue_task for max_turns. Surfaced on the task card's status-chip tooltip.
batch_get_tasks with includeDescription=true had no size guard: 8 tasks'
full Description/Result serialized to a single 51k-char line, blowing past
the tool-result token budget and forcing a file/PowerShell workaround to
read it back. Roadblock text was also only reachable via a second get_task
call per task, since roadblockCount (in the lean ref) has no text sibling.
- descriptionMaxChars (default 1500, was unlimited) truncates Description
and Result independently, flagged via *Truncated/*FullLength so a caller
never silently works off a cut string.
- fields narrows taskFull to just the named properties, including the new
roadblockText (the tail of Result after TaskRunner's roadblock marker) —
reachable without pulling in the rest of Result/Description.
- taskFull is now BatchTaskDetailDto, a batch-only shape decoupled from
TaskDto so get_task's own contract is untouched.
- The whole response is capped (MaxResponseChars); over that, the call
throws naming which parameter (descriptionMaxChars/fields/taskIds) to
adjust instead of shipping an oversized payload.
10 tasks with 5000-char descriptions (default settings) serialize to
~13.8k chars, comfortably under the 25k-char cap.
Adds a user/MCP-declared task dependency (DependsOnTaskId) distinct from the
planning chain's internal BlockedByTaskId: add_task/update_task can set it,
the queue picker skips a Queued task until the dependency reaches Done, a
Failed/Cancelled dependency leaves the dependent blocked instead of starving
silently, and setting a link rejects self-reference/unknown-id/cycles.
get_task/list_tasks/batch_get_tasks now report blocked/blockedReason, and
wait_for_task_change reports "Blocked" immediately instead of running out its
timeout on a task the picker will never claim.
preview_merge/preview_merge_set gain staleFiles: files a branch touches that
the target branch also changed since the branch's fork point, a more honest
staleness signal than `behind` alone.
Omitted fields on set_task_config/set_list_config used to be indistinguishable
from an explicit clear, so any single-field update silently wiped every other
override (including SessionSkills, which wasn't even a tool parameter). Now
only the fields you pass are changed; clearing an override requires naming it
in the new clearFields array (a sentinel string doesn't work uniformly since
maxTurns is an int). set_list_config's "all four null deletes the config" case
is preserved but only reachable via clearFields. Return values now reflect the
full resulting config.
get_merge_conflicts and resolve_conflict_hunk let an MCP caller resolve a
leaveConflictsInTree merge hunk-by-hunk (ours/theirs/base/literal text)
instead of grep-parsing diff3 markers itself. Reuses the existing
TaskMergeService/ConflictMarkerParser machinery (adds MergeSegment.StartLine
and a non-staging TaskMergeService.WriteConflictFileAsync) rather than a
second parser. A partially resolved file is written without git add so
continue_merge keeps refusing while any hunk still has markers.
## Befund (Batch-Lauf 2026-08-06, Liste "Bandel.Hub")
maxParallelExecutions = 5, 23 Geschwister-Tasks auf demselben Plugin. Die Parallelitaet hat die
Ausfuehrung verkuerzt, aber die Konfliktlast erhoeht: eine CSS-Datei wurde von 10 Tasks
angefasst, drei Razor-Dateien von je 6. Der Engpass des Gesamtdurchlaufs war nicht die
Ausfuehrung, sondern Review und Merge — die sind zwingend seriell. Netto wa
ClaudeDo-Task: 89f989e0-4fcf-43f5-a802-11d63df7a6d1
get_task_diff gained an optional paths param (git pathspec, both stat and
full-diff modes) so a large multi-file task can be narrowed to the files
actually in question instead of shipping the whole diff.
preview_merge_set now also reports Subsets: pairs where one task's changed
files are a proper subset of another's in the same set, the strongest
available post-hoc signal that a task may be redundant with another.
preview_merge could only see file-overlap cleanliness via git merge-tree, never
whether the result compiles -- the two costliest findings of the 2026-08-06 batch
run were both merge-tree-clean but build-broken. When a list has a verify command
configured, a clean preview is now additionally materialized (via a commit-tree +
detached scratch worktree, outside the real repo, always cleaned up) and
built/tested there, without ever touching the real working tree.
preview_merge always attempts a verify run when a command is configured;
preview_merge_set only does when its new runVerify parameter is set (default off),
so a set preview never starts N builds unasked. The post-merge verify gate is
unchanged.
Every tool description now leads with what the tool does AND when to reach for
it, since MCP clients rank tools by that text. Per-parameter prose moved onto
the parameters as [Description], exhaustive result-shape enumerations and
design/history rationale dropped, and the repeated boilerplate clauses
(lean-task-ref, batch cap, refused-while-Running) pulled into McpToolDocs,
which also documents the style for future tools.
Tool-level description text: 20494 -> 13605 chars (-34%); combined with the new
parameter descriptions 18517 (-10%).
Closes gaps that caused wrong calls rather than just verbose ones:
- list_task_attachments returns metadata only, no file content
- run_task_now shares continue_task's single override slot and throws when busy
- list_runs is ordered oldest-first and feeds get_run
- workingDir on create_list/update_list is an existing local git repo path,
unvalidated until the first task run
- get_task_worktree's behind=0 also means the main ref was unreachable
Removes get_task_status_values: a whole tool entry for static reference text.
GetTask's description is now the canonical place for status meanings.
NormalizeTitleWords transliterates umlauts before tokenizing ('für' ->
'fuer'), but TitleStopWords listed 'fur' -- a token the pipeline can
never produce -- so 'für' was never filtered as filler and could push
unrelated titles past the near-duplicate threshold.
TaskRunner.RunAsync created the worktree (PrepareRunDirectoryAsync)
before claiming Running via StartRunningAsync. RunNow dispatches by
task id with no atomic claim of their own, so a Queued task racing
the queue picker's atomic SQL claim could hit WorktreeManager's
branch-collision self-heal, which force-removes and recreates the
winner's live worktree mid-run.
Move the claim before any resource creation and bail out immediately
when it's rejected. OverrideSlotService.RunNow also fast-rejects a
task already Running in the DB (defense in depth). RunCancellationRegistry
now refuses (and logs) a double registration instead of silently
overwriting the first CTS, so a losing dispatch's cleanup can no
longer unregister the winner's cancellation token.
preview_merge/preview_merge_set now report isEmpty (ahead==0, or
HandlerBaseCommit==HandlerHeadCommit for a worktree-less list-handler
task) instead of leaving an empty branch indistinguishable from a small
one. preview_merge also stops throwing for worktree-less handler tasks,
falling back to their fixed commit range. review_task's parent approve
returns emptyChildren, naming the Done children whose review range
contributed nothing before the unit merge lands. TaskRefDto/TaskDto now
expose roadblockCount so a CLAUDEDO_BLOCKED child is identifiable over
MCP, since it still reaches Done per the unified parent model.
WaitingForChildren already counted as "changed" since it's outside Queued/Running,
so waiting on a planning parent returned immediately even though its children were
still running. The new opt-in flag (default false, unchanged behavior) keeps polling
through WaitingForChildren and only reports changed once the parent reaches
WaitingForReview or a terminal status.
list_tasks on a list of ~100 verbosely-described tasks could return
390k+ characters in one call, blowing past the caller's token limit.
Both tools now default to lean TaskRefDto references (no
Description/Result) and take an includeDescription flag to opt back
into the full TaskDto payload — same flag-alongside-nullable-payload
idiom already used by BatchGetTaskResult/TaskConfigResult. get_task is
unchanged.
add_task and batch_add_tasks now report up to 3 open (non-terminal)
tasks in the same list with a strongly overlapping title, so a
parallel agent can notice and mention a likely duplicate instead of
silently creating one. The task is always created regardless. Uses a
cheap normalized-word overlap heuristic (no embeddings/LLM call),
robust to German umlaut/digraph spelling variants. Breaking change:
AddTask now returns AddTaskResult { task, possibleDuplicates } instead
of a bare TaskRefDto; BatchAddTaskResult gained a PossibleDuplicates
field.
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.
Adds a read-only get_effective_run_config(taskId) tool that reports the
model/max-turns/effort/permission-mode/agent-path/system-prompt/skills a
task will actually run with, each tagged with its source (task/list/
preset/global), plus max-turns' raw requested value and clamp status.
Extracted the model/max-turns/agent-path resolution out of
TaskRunner.ResolveConfigAsync into EffectiveRunConfigResolver so the run
path and the new reporting tool share one codepath instead of risking
drift, per docs/explore-notes/worker-task-pipeline.md's max-turns trap.
update_task, update_task_status, add_task, add_subtask, set_my_day,
abort_merge, review_task, and their batch variants now return a lean
TaskRefDto (id/listId/title/status/sortOrder/isMyDay) instead of the
full TaskDto. Those tools were re-sending the caller's own description
text back on every call, wasting a large share of session context on
list-handler-style runs. get_task/list_tasks/batch_get_tasks are
untouched and still return the full DTO.
mika kuns