just-demand-execution

Run `just-demand . list-active` to confirm the task, select it with `select-task` or `resume` if needed, verify required context files exist, then mark it executing before implementing. Completion flows through `complete-verification`, which records verification, creates a checkpoint commit, and archives the task.

Dispatch only when all six eligibility gates pass (goal stability, boundary independence, context compressibility, result verifiability, capability match, failure recoverability) plus three net-benefit questions. Any uncertainty means the main agent executes, even for multi-file work.

The tester validates results against the task brief, checks visible effects, and applies low-risk local fixes. The advisor provides fresh-context diagnosis and independent recommendations for repeated failures or cross-boundary problems without doing large-scale implementation.

No. These are compatibility-only roles dispatched only when a legacy task already records the role or the user explicitly requests it. New tasks use the main agent plus tester and advisor subagents.

The agent tells the user the subagent path is unavailable and asks whether to retry now or skip one turn, preferably via the question tool. It never silently falls back to a different plan; retrying the same session is usually the right first move.

Execution halts when clarification gates fail: unresolved blocking questions, missing scope or expected behavior fields, or absent task context files like context.md and verify.md. The workflow routes back to clarification rather than guessing user intent.