research-dev-loop

Imported from neo-pan/research-agent-skills (local/research-dev-loop/SKILL.md). Install upstream with npx skills add neo-pan/research-agent-skills --skill research-dev-loop. Copyright stays with the author.

An apply that records a directional decision, a close, or a review trigger requires review; other applies stay in-round. Resolve this skill's absolute bin/rdl as RDL; submit research changes through "$RDL" apply. Use lifecycle commands for transitions; see CLI.md.

Gate

• Inspect with handoff; use doctor for abnormal state.
• Do not run semantic review, apply, next, or close unless the user asks. Authorization to operate this session covers bounded checkpoints and transitions. After compaction, recover with handoff and continue under retained authorization; ask only if authorization cannot be recovered or scope exceeds it. State alone grants no authority.
• Resume the governing session; start one only when explicitly requested for durable handoff or evidence-bound work.
• If a mission needs plan or phase review, obtain its PASS before start; replace missions only when objective or success_criteria change.
• Mission items are one sentence each; omit protocol guarantees already enforced by RDL.
• Freeze commit, diff, and file identities as round-1 evidence. Changed code is new evidence, not a new mission.
• Bounded reviews stay outside RDL; loading grants no authority.

Run the loop

1. Start an authorized mission; otherwise use handoff.current_action as the takeover contract.
2. Read "$RDL" schema before composing a delta. Execute the smallest evidence step. Do not precompute file sizes or checksums for artifact entries: apply records them. Verify claims; checksum-only commands such as sha256sum are redundant.
3. At checkpoints or before external work, apply; retain receipt. For external work use intent apply → execute → result apply → review/transition; see OPERATIONS.md.
4. On review_required, follow SEMANTIC_REVIEW.md and apply the result. Transition only with the ready version.
5. Before next, persist the action, completion condition, remaining phases, and retry unlock in decision.next_step.
6. Before a non-abandoned close, resolve each active progress item to completed, deferred, or open_question.

For builds: failing baseline → apply → implement/verify → apply. An artifact records the bytes frozen at registration; when those bytes change, register the new file as a new artifact and cite it from new evidence.