Audited all 12 prompt kinds against the code they drive. Every real defect sat on
the boundary between prompt text and the --allowedTools the launcher passes.
- Planning: "Use nothing else" after a six-tool list forbade the brainstorming
Skill the same prompt demands two paragraphs earlier. WindowsTerminalLauncher
allowlists mcp__claudedo__*,Read,Grep,Glob,WebFetch,WebSearch,Skill -- name them,
and tell the planner to ground subtasks in the repo with Read/Grep/Glob.
- System: SuggestImprovement is only allowlisted when ParentTaskId is null and
PlanningPhase is None, and TaskRunMcpService throws for any child, but this
prompt reaches every run. Planning children were told to use a tool they lack.
- MergeHelperExecute: derived "effective max-turns" from task/list/preset by hand,
which misses TaskRunner's MaxTurnsCeiling clamp. Call get_effective_run_config
instead -- built for exactly this and reports the clamp.
- MergeHelperExecute: quoted the override-slot error as the raw lowercase throw
rather than the string ExternalMcpService actually surfaces.
- Refine: listed Read/Grep/Glob unconditionally though RefinePrompt.BuildArgs only
appends them when a repo is available.
Two findings deliberately left open, both needing a code decision rather than a
prompt edit: the System prompt's worktree claim is false for a list without a
WorkingDir (task runs in a plain sandbox dir), and 'fable' is missing from both the
prompt's cost ordering and ModelRegistry.ByCostAscending.
The execute prompt told the handler to wait with timeoutSeconds up to 170 -- a
leftover from the retired MCP_TOOL_TIMEOUT=200000ms era. The real server-side
clamp is TaskWaitMcpTools.MaxTimeoutSeconds = 900 and every launcher sets
930000ms, so the handler was making ~5x the wait_for_task_change calls it needed
and burning turns on re-waiting.
It also never passed treatWaitingForChildrenAsBusy, and only waited on ids that
were Queued or Running. A task with children reports "changed" the moment it
reaches WaitingForChildren, so such a task both dropped out of the wait set and
signalled completion early -- the handler could reach review/merge while
children were still running.
RenderTemplate replaced tokens one key at a time over a StringBuilder, so a
token appearing inside an already-substituted value got substituted again on a
later pass. The prompt briefs only escaped this because their callers happen to
pass "tasks" last -- reordering the dictionary or adding a fourth token would
have started rewriting user-authored task descriptions, which after the enhance
phase carry file paths and config snippets.
Single-pass regex over the template instead; unknown tokens still pass through.
Both merge-helper ConPTY sessions loaded PromptKind.MergeHelper, so the
post-handoff session received the phase 0-2 dedupe/enhance instructions and was
told to ignore them by its brief alone.
Split into MergeHelperTriage (phases 0-2 + handoff) and MergeHelperExecute
(phases 3-5), so each session carries only its own phases. Consolidated the
generic ask-the-user rule to one place per prompt, scoped Phase 5's summary to
what the execute session actually knows, and moved the dedupe/enhance bilanz to
the triage handoff. Regression guards assert neither prompt carries the other's
phase headings and that the shared-checkout git rule stays in execute.
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.
Phase 2 can now propose splitting a bundled/blocked task via add_task/add_subtask
instead of forcing it into one description. Phase 3 reports each task's effective
max-turns and asks before queuing a substantial task with a low value. Phase 4
checks Phase 2's file lists for cross-task collisions and may reorder merges with
a stated reason instead of always following brief order.
Phase 0 forced a batch_get_tasks full-fetch across every task just to see
descriptions, which blew past the client's token limit on larger lists.
brief.md lives on disk and has no such limit, so descriptions now render
there directly (fenced with an extended backtick run, indented under the
list bullet, so embedded headings/lists/code fences can't break the task
list's own structure). Phase 0 now treats the brief as the primary source
and only falls back to batch_get_tasks for fields it doesn't carry.
Phase 1 previously asked the user to confirm the absence of duplicates
even when no candidate pair was found. The handler now decides that
itself and moves straight to Phase 2; per-pair questions remain when
at least one candidate exists.
mika kuns