You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When /ce:work selects the "parallel subagents" execution strategy (Phase 1 Step 4), it dispatches multiple Agent tool calls for independent implementation units. Currently, the skill does not instruct the use of isolation: \"worktree\" on these Agent calls — subagents all operate on the same working directory, relying on a "non-overlapping files" heuristic.
This creates risks:
Git operations can conflict between concurrent subagents
Test runs can interfere with each other
If the overlap assessment is wrong, file writes race
Evidence: it already works when the model chooses it
In a real session (Opus, 3 plans executed in parallel), the orchestrating agent independently decided to use isolation: \"worktree\" on each Agent call. The result:
3 worktrees created at .claude/worktrees/agent-{id}, each on its own branch
Subagents ran fully isolated with run_in_background: true
Orchestrator reviewed diffs from all 3 worktrees, then merged sequentially
3 merge conflicts arose on overlapping files — resolved cleanly by the orchestrator
The flow worked well, but it was emergent behavior — the model happened to use the parameter. A different model or context could skip it entirely.
Proposal
When the parallel subagents strategy is selected in Phase 1 Step 4, the skill should instruct:
Use isolation: \"worktree\" on each Agent dispatch
Use run_in_background: true for concurrent execution
After all subagents complete, review diffs from each worktree
Merge branches sequentially in dependency order, resolving conflicts
Clean up worktrees (note: .claude/worktrees/ should be in .gitignore to avoid accidental staging)
This makes the parallel strategy reliable rather than model-dependent.
Notes
The serial subagents and inline strategies don't need this — they're sequential by definition
The git-worktree skill's worktree-manager.sh is an alternative to the Agent tool's built-in worktree support, but isolation: \"worktree\" is simpler since the harness manages the lifecycle automatically
Problem
When
/ce:workselects the "parallel subagents" execution strategy (Phase 1 Step 4), it dispatches multiple Agent tool calls for independent implementation units. Currently, the skill does not instruct the use ofisolation: \"worktree\"on these Agent calls — subagents all operate on the same working directory, relying on a "non-overlapping files" heuristic.This creates risks:
Evidence: it already works when the model chooses it
In a real session (Opus, 3 plans executed in parallel), the orchestrating agent independently decided to use
isolation: \"worktree\"on each Agent call. The result:.claude/worktrees/agent-{id}, each on its own branchrun_in_background: trueThe flow worked well, but it was emergent behavior — the model happened to use the parameter. A different model or context could skip it entirely.
Proposal
When the parallel subagents strategy is selected in Phase 1 Step 4, the skill should instruct:
isolation: \"worktree\"on each Agent dispatchrun_in_background: truefor concurrent execution.claude/worktrees/should be in.gitignoreto avoid accidental staging)This makes the parallel strategy reliable rather than model-dependent.
Notes
/lfg-ext) explored a similar pattern for external delegates but hasn't been mergedworktree-manager.shis an alternative to the Agent tool's built-in worktree support, butisolation: \"worktree\"is simpler since the harness manages the lifecycle automatically