adversary

The Lens of the Adversary

Read first: ${CLAUDE_PLUGIN_ROOT}/core/protocol.md and ${CLAUDE_PLUGIN_ROOT}/core/dossier.md.

This is a META lens: it examines the PLAN itself for weakness, not a domain perspective. It runs as the mandatory final step of /lens:socratic (and standalone on any plan/diff). Adversarial critique reliably surfaces what normal planning misses — so a plan is not "done" until it survives this pass.

First: suggest plan mode, then WAIT

If the work is substantial enough to deserve a written plan, SUGGEST to the user that they enter Claude Code's plan mode (or otherwise formalize the plan as a written artifact) — and WAIT for their input. Do NOT toggle plan mode yourself. If they decline, or the work is small, critique the plan as it stands.

The red-team battery

Attack the plan from these angles; every finding must be CONCRETE, not generic:

1. Refutation: name the ONE assumption that, if wrong, breaks the plan. Is it verified or merely assumed? An assumed load-bearing assumption is the top risk.
2. Weakest decision: of the decisions recorded in the dossier/plan, which is least defensible, and why? What alternative was dismissed too fast?
3. What discovery missed: re-run the four excavations (protocol.md) AGAINST the plan — a caveat, rabbit hole, blast-radius entry, or platform-matrix gap the plan does not address. This is where socratic plans most often leak.
4. Reversibility & blast radius: what here is hard to undo, and is the plan treating it with the caution it deserves? Who/what breaks if this ships wrong?
5. "First thing that worked" test: is this the best plan, or the first coherent one? Name one genuinely different approach and say why it lost — or didn't.

On substantial work, fan out: run the angles as independent adversarial passes (or read-only subagents) so no single line of attack dominates, then merge.

Disposition — every surviving finding

Each finding the plan cannot already answer becomes exactly one of:

• revise — change a decision in the dossier (state the new decision);
• reopen — spawn a frontier branch (protocol.md) and hand back to discovery;
• accept — record as an explicit accepted risk, with an owner, in the dossier. Never let a real finding evaporate as "noted".

Output contract

A critique table: angle · finding · disposition (revise/reopen/accept) · the concrete change or risk. Then the verdict: the plan SURVIVES (every finding dispositioned) or needs another discovery round. Append the critique + verdict to the dossier.

Feed the foundry

Gaps (an attack angle this battery lacks, or a recurring plan-weakness pattern) → {"type":"gap","date":"<YYYY-MM-DD>","lens":"adversary","note":"<one line>"} appended to <foundry>/pending-retros.jsonl when a lens config exists; otherwise mention it.