CRITICAL WORKFLOW REQUIREMENT

You MUST NOT add comments that describe the change they just made (e.g., “removed”, “legacy”, “cleanup”, “hotfix”, “flag removed”, “temporary workaround”).
Only add comments for genuinely non‑obvious, persistent logic or external invariants. Keep such comments short (max 2 lines).
When migrating or refactoring code, do not leave legacy code. Remove all deprecated or unused code.
Put change reasoning in your plan/final message — not in code.

Work Summary

Keep set_work_summary up to date during the turn.
You may call any tools before set_work_summary when you need more context.
Call set_work_summary once you understand the work clearly enough to label it well, and update it again when the focus materially changes.
Prefer calling it with both fields: set_work_summary({ theme: "...", now: "..." }).
theme answers: what is this pane about overall? Keep it stable across many turns.
now answers: what are you about to do next? Update it whenever the next concrete step changes.
Keep both labels concrete and under 48 characters.
Since the summary line has dedicated space, prefer richer phrases that help a forgetful human re-orient instantly.
Good theme examples: Tmux status summary workflow, Agent tracker integration, Flutter auth onboarding.
Good now examples: Patch summary enforcement, Read restore path handling, Wait for user reply.
Bad labels: Working, Coding, Debugging, Researching, Task, Fixing stuff.
Bad now phrasing: Debugging summary enforcement, Reading restore path handling, Waiting on user reply.
If you are blocked or waiting, keep the theme and change now, for example Wait for user reply or Wait for tests.
If the labels are missing or stale, stop and update them first.
Repeating the same theme across turns is acceptable when the overall mission has not changed.

Trigger bursts only when needed; otherwise continue normal execution.
Choose burst size by complexity:
- low: 2 subagents
- medium: 3 subagents
- high/risky: 4-5 subagents
Use one burst round by default.
Run a second round only if confidence is still low.
Assign non-overlapping scopes to reduce duplicate findings.

Each discovery subagent returns compact, evidence-based output:

Limit each subagent to maximum 5 findings.