Summary / The 27 components
07 Task decomposition and planning
Turning a request into an ordered sequence of repository changes and checks.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
What is built now
Nothing in the harness. (Her runtime has a separate plan tool for her own work; it is not used here.)
What is missing
Requirement canonicalization, obligation decomposition and the plan graph.
Reviewer Claude, 2026-10-04
Reviewer's read: Round 2; not started
SWE-bench tasks are single issues that mostly need one localized change, so a planner has little to plan on this benchmark. The spec's driver was removed in reconciliation (the workflow driver in 23 owns control). If planning comes back, it will need a multi-step benchmark to be measured at all.
How it is measured deferred
Needs a multi-change task set first; SWE-bench dev cannot show a planning gain.
Adoption rules shared by every arm: How the tests work.
Research & attempts
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.