Adversarial Reactor

Adversarial Reactor Gate

A saga/dataflow-orchestration review gate for Elixir code built on Reactor — pass/fail: a single blind reviewer subagent judges the work against this gate's rules with an adversarial mandate, and the work passes only when every rule is PASS or N/A. This skill renders verdicts; it never fixes the work.

The rules target one failure mode: code that assumes guarantees Reactor never made — that rollback is automatic, that cleanup in compensate fires on downstream failure, that :ok absorbs an error, that steps run in the order they were written, that :retry is bounded by default, that {:halt, reason} aborts, that map steps parallelize, that process-local state follows a step into its task process. These bugs compile cleanly, pass happy-path tests, and surface as orphaned payments, infinite retry loops, and races. Each rule carries an Evidence of violation paragraph so a reviewer can decide PASS/FAIL/N/A from artifact evidence alone. The sibling gates judge different layers: adversarial-elixir paradigm fit, adversarial-beam runtime semantics — run them alongside for full coverage of OTP-shaped work.

When to Apply

• An Elixir feature that defines or modifies Reactor workflows (use Reactor modules, Reactor.Step implementations, Reactor.run call sites) is about to merge and needs an objective PASS/FAIL on its orchestration semantics.
• An agent (Claude, Codex) authored a Reactor saga and you want an independent check that it did not park cleanup in the wrong callback, leave side effects irreversible, or assume lexical step order.
• A workflow is being promoted from happy-path prototype to production (payments, provisioning, imports) and its failure/rollback/retry paths need to be surfaced as verdicts.
• An incident postmortem fixed one orphaned-side-effect or retry-storm bug and you want the same class hunted across the reactors in the affected area.

Do not apply to targets with no Reactor usage (the reviewer prompt's precondition aborts with "GATE NOT APPLICABLE"), or when the user wants explanations and refactors rather than a verdict. General OTP/BEAM concerns (supervision, PubSub, ETS) belong to adversarial-beam; Elixir paradigm fit belongs to adversarial-elixir.

Review Protocol

Follow these steps exactly — the gate's value is that every review runs the same way.

1. Identify the target. Pin down exactly what is under review (a diff, a set of files, a PR) and note the ref/paths so the review runs against an unambiguous, fixed target. Record the stack facts (from mix.exs, mix.lock, test_helper.exs) — Reactor version, Ecto/sandbox presence, mocking library, telemetry consumers. Include the repo root in the target description — several rules must search beyond the diff for mounted step modules' callbacks, Reactor.run call sites, and middleware modules (the reviewer prompt lists them).
2. Load the rules. Read references/_sections.md and every rule file in references/ (all saga-*.md, retry-*.md, dep-*.md, step-*.md, comp-*.md, conc-*.md, obs-*.md files).
3. Compose the reviewer prompt. Fill references/reviewer-prompt.md with the rules, the target, and the stack facts. The composed prompt must be fully self-contained — a reviewer sees no conversation history, so nothing may refer to context outside the prompt.
4. Dispatch one blind reviewer. Launch a single Task subagent whose entire input is the composed prompt — no conversation context, no commentary alongside it.
5. Render fail-closed. The reviewer's structured output is the verdict — there is no merge step. Overall verdict is PASS only when every rule is PASS or N/A; any single FAIL fails the gate. Never average, weigh severity, or waive a rule — a "minor" FAIL is a FAIL. If the reviewer returns "GATE NOT APPLICABLE" (no Reactor usage in the target), stop and report that instead of a verdict.
6. Render the verdict. Fill assets/templates/verdict.md. On FAIL, aggregate the reviewer's "missing for PASS" suggestions into the fix list, each with its location, ordered by category importance. Every rule whose result is FAIL must appear in the fix list with a change concrete enough to apply as written — if the reviewer's suggestion only restates the violation, derive the fix from the rule's correct example before rendering.

If the same rule flips verdicts across re-reviews of an unchanged target, or a human reads the evidence and overrides the verdict, that is a decidability bug in the rule — record it in gotchas.md and sharpen the rule; do not override the gate.

Verdict Format

The reviewer returns, per rule: PASS | FAIL | N/A, evidence (file:line or a quote — required for PASS as well as FAIL), and for every FAIL, the fix that flips the rule to PASS once applied — the named change plus its location, never a restatement of the violation. The final report follows assets/templates/verdict.md.

Rule Categories

Gotchas

Read gotchas.md before dispatching the reviewer — it pre-records scope guards (shapes-not-brands, what counts as a side effect, the diff-vs-repo search obligations) so the reviewer does not judge outside the rules.

Related Skills

• adversarial-beam — the runtime-semantics sibling gate: supervision, backpressure, event delivery, shared state, distribution. A Reactor workflow lives inside an OTP app; run both on substantial work.
• adversarial-elixir — the paradigm-fit sibling gate: OO/enterprise habits ported onto the BEAM. Same protocol, complementary rules.
• staff-level-elixir — the advisory sibling: greenfield "which tool, which convention" guidance and judgment calls the gates deliberately exclude. Use it to write or fix code; use the gates to verdict it.

Reference Files