v1.6.0 — Plan handoff and Write clarification
Panda Plan
Saved plans end with an execution handoff. Plan asks the user to review the plan
and choose how panda-execute runs it:
- Subagent-driven: a fresh agent implements each task and a fresh reviewer checks it before the next starts, then one whole-change review. Offered only when the host can delegate.
- Native: the current session implements every task, then one review of the whole change.
Plan recommends one mode from the plan's task count, task coupling, and cost of
a mistake, and gives a ready-to-use request such as
“Use panda-execute (native) to implement .panda-suite/plans/<file>.md.” A mode
the user already named is kept; chat-only plans skip the menu. Saving a plan or
choosing a mode does not by itself authorize implementation.
Plans also state per-task interfaces (what each task consumes and produces), size tasks to independently testable deliverables, write steps that each decide one thing, copy project-wide constraints exactly, and list a short review focus of implied but untested inputs.
Panda Execute
A saved plan is the contract. Execute implements every task with the plan's names, signatures, tests, and values; checks off a task only after its checks pass; resumes from the first unchecked task after re-running finished checks; records any deviation under a "Deviations" heading; and continues until every task is verified or a precise blocker remains. With no mode named it runs native. Subagent-driven mode gives each task to a fresh implementer, re-runs the checks itself, and has a fresh reviewer check the task before the next one.
Panda Write
Write names a source of truth for each section and asks before drafting when the
source, scope, or material facts are unclear. Finished documents leave out chat
narration and "limits" notes; gaps go in the delivery message. New documents
default to .panda-suite/docs/YYYY-MM-DD-<topic>.md.
Verification
Fresh-context runs compared previous and revised instructions: the execution mode offer appeared only with the revision; both Execute versions completed the plan fixture, but only the revised one left already-finished work unchanged. See Plan refinement, plan handoff results, and the Write refinement. One run per arm; the sample does not establish reliability across repeated runs.
Skill validation and 30 tooling tests pass.
Compatibility and recovery
No application migrations. To restore the previous instructions, use v1.5.1. Existing plans without checkboxes still run; Execute reports progress in chat.