Planning

Planning

When

• Task has more than one action, stage dependencies, or multiple agents/skills.
• Multiple feasible routes exist and need comparison.
• The flow is spinning, evidence is weak, or closeout is being declared too early.

Brain-Storm

• Define the core contradiction first, then list routes.
• Give ≥3 fundamentally different routes (different assumptions, methods, validation styles).
• Every route gets a stable branch_id and wave. Later stages must use those IDs.
• Every route says: why feasible, biggest risk, first validation step.
• If public leaderboard exists: keep self-explore path and public high-score reverse-absorption path as two parallel branches. Both stay alive.
• BFS-like push: same-wave high-value branches do first-round validation in parallel. Do not collapse into one main line early.
• If one branch leads: open recursive widening around it — create ≥2 sibling-variant branches with new branch_id/wave. Keep them alive until compared or ts-critic says marginal gain is low.
• Output stop not allowed if any executable candidate with information gain still exists.

Deep-Reasoning

• Split problem into goals, assumptions, dependencies, validation standards, stop conditions.
• Identify 2–4 main reasoning paths with preconditions.
• Mark the node that should be validated first.
• Mark candidate merge points — require cross-branch comparison evidence before merge.
• Keep branch_id/wave/preconditions/next-validation-node/drop-conditions for every path.
• If one path leads: judge whether same-family variants, lighter versions, or two-stage versions should still be widened. If not clearly no, do not converge.

Plan Construction

• Plan is loop-based, not one-pass. Last step returns to item one unless ts-critic says no more high-value actions.
• Sync plan into OpenCode TODO. Plan without TODO is not landed.
• TODO states: only todo, in-progress, done. No hidden fourth state.
• TODO must NOT contain end / stop / wrap up / delivery complete. Last item = loop-back, review, or next-round entry.
• Rollback rules: done → in-progress (needs more evidence); done → todo (precondition failed, route changed); in-progress → todo (blocked by upstream).
• For multi-branch tasks: TODO must carry branch_id, wave, frontier_status, compare_gate.

Critic-Loop Trigger Signals

Enter this loop when any of these appear:

• One route tuned repeatedly with no new information.
• 2+ consecutive failures without fundamental route change.
• Conclusion from guessing, no evidence.
• done claimed but no validation, no plot check, no result check.
• Search only did one-keyword direct search.
• Tables/logs read but visual analysis not done.
• TODO contains end/stop/wrap up/delivery complete.
• Files listed in summary but not checked for existence.
• Formal document exists but no final ts-critic review.
• Figures in report body but no analysis paragraph after them.
• Experiment loop missing structured results → plots → visual analysis → targeted retest → self-critique → ts-critic review.
• Forecast outputs exist but horizon/schema/plausibility not checked.
• ztxexp not used for benchmark/ablation/multi-seed work.

Pre-Stop Gate

Every time the flow wants to stop:

1. brain-storm — re-list remaining actions, risk-reduction actions, rollback routes.
2. deep-reasoning — judge each candidate: executable? information gain? dependencies?
3. ts-critic — reviews and gives stop signal.

Stop is allowed only when all three agree no high-value action remains.

Output

• Branch map (mermaid when needed)
• Route comparison summary
• Plan steps with serial/parallel notes
• TODO mapping table
• branch_id/wave/frontier status table
• Stop signal