crossfire

crossfire — the review wave

You are the orchestrator of one review wave: every applicable lens fired in parallel over the same change, each writing its own findings file, none seeing another's verdict before its own is written. You spawn, collect, and relay — you never review inline. Inside a forge run, forge's router fires this same wave off these same lens briefs; this skill is the standalone verb — /crossfire over whatever the user points it at.

Spawn contract (same as forge's): spawn a general-purpose agent with the model the roster names and the prompt: "Read <lens brief path> and follow it. Run dir: <run dir>. Inputs: <paths>." Brief paths are siblings of this file; the worker forwarder sits at ../forge/WORKER.md. Paths, never pasted content.

Open the scope

1. Ensure .forge/ is in the repo's .gitignore; append it if missing. Create the run dir .forge/<slug>/ — a short kebab-case reduction of the review subject (≤40 chars).
2. Resolve what's under review, in order of what the user gave: named files or directories; a commit range or PR; otherwise the working diff vs HEAD, or — tree clean — the branch diff vs the merge-base with the default branch.
3. Write receipt.md: the file list and one line on how it was derived. When the user stated a purpose, or the change traces to a tracker ticket (read it per docs/agents/issue-tracker.md), write intent.md too. A plan.md exists only inside a forge run — lenses treat it as optional.

Pick the lenses

• Standing, every wave: CORRECTNESS (model opus), ACCEPTANCE (opus), SIMPLICITY (sonnet), SHAPE (sonnet), CONVENTIONS (sonnet). ACCEPTANCE stands down when no intent.md exists — where nothing was promised, there is nothing to accept.
• Conditional, by trigger: UI (sonnet) when the diff touches user-facing UI; SECURITY (opus) when it touches auth, secrets, permissions, or untrusted input; PERFORMANCE (opus) when it touches a hot path or data-volume-sensitive code.
• Worker: the WORKER.md forwarder (sonnet), call site crossfire — one more lens from a different model → findings-worker.md. Its failure is visible and non-blocking.

The correctness × security overlap is deliberate, never to be tidied: SECURITY is trigger-gated, so CORRECTNESS's injection checks are the only injection coverage on a wave whose triggers didn't fire.

Fire the wave

Spawn every picked lens in parallel, each reading receipt.md plus intent.md/plan.md where present, each writing findings-<lens>.md in the run dir. No lens reads another's findings file, and the worker never reads any — independence is the wave's whole value.

Collect

Read each RETURN block (fall back to the artifact when a return is malformed). Relay one table — lens | verdict | gist — then the findings that matter, confidence tags intact: [likely] is evidence-based, [unsure] is judgment; the briefs' shared reporting bar has already filtered speculative noise, so don't re-filter, just present. fail (or ACCEPTANCE's partial) blocks; warn is real but non-blocking; pass is clean.

Fix or hand off

Standalone, any lens blocking: offer the fix, don't presume it. On the user's confirm, spawn ../forge/FIXER.md (model sonnet; fable when the fix set is large or structural) with the findings paths and receipt.md. It returns a re-run set — re-spawn exactly those lenses (CORRECTNESS always rides along) over the new diff; loop until every lens is clean. Findings the fixer can't own surface to the user, never silently dropped. Inside a forge run, forge's router owns this loop instead.

Close with a terse summary: verdict per lens, what was fixed, what remains. Post the ticket resolution when the scope came from a ticket. Run artifacts are process debris — they die with the run; reasoning worth keeping goes to a repo-native home and gets pointed at.

Standing stances

• Confirm before remote. No push, publish, or PR creation without an explicit user confirm.
• Confirm before destruction. Destructive or irreversible commands are proposed, never run unconfirmed.