Antithesis Review Inputs

Skill version: 2026-09-23 a8fe900

Purpose and Goal

Examine a workload and report how well its input generation explores state space. The system under test is a black box — the review works entirely from the workload code and any documentation the user provides.

This skill is read-only. It produces a report of findings. It does not modify the workload, open PRs, or suggest specific code changes.

The framing is "untapped exploration" — findings are ceilings on what the workload can reach, not defects. A workload exhibiting any anti-pattern may still be productive. The question is whether it could reach more interesting behavior with structural changes to how it uses randomness.

When to Use

When a user isn't finding bugs and wants to understand whether the workload's input generation is limiting what Antithesis can explore. Also useful as a periodic health check on workloads — both human-written and agent-generated.

Core Principle

Randomize the policy per run, not just the choices under a fixed policy. Each run should explore a different region of behavior by biasing its configuration differently. The skill reviews how well the workload does this.

Workflow

1. Locate the workload

Find the test commands and workload code. Common locations:

• antithesis/test/ — test command scripts
• Workload source files referenced by test commands
• Client libraries or helpers used by the workload

If the workload location isn't obvious, ask the user.

2. Read the anti-patterns reference

Read references/anti-patterns.md before examining the workload. It contains the full catalog of input generation anti-patterns organized by category, with explanations of why each limits exploration and what better looks like.

3. Examine the workload

Read the workload code and look for structural patterns that match the anti-patterns. Focus on what is visible in the code:

Action selection — How are operations chosen?

• Is there a fixed set that's always fully active, or per-run subsetting?
• Are action weights fixed, uniform, or varied per-run?
• Is ordering imposed by the harness or emergent from randomness?
• Are there symmetric pairs always active together?

Data generation — How are inputs produced?

• Are values hardcoded, drawn from fixed pools, or generated?
• Do ranges vary per-run, or are bounds hardcoded?
• Does input quality vary (valid, invalid, mixed), or is it always valid?
• Do payload sizes span orders of magnitude, or stay in a narrow band?

Structure — How is the workload organized?

• Is the concurrency level fixed or varied per-run?
• Does operation pacing vary, or is it always steady/fast?
• Does the workload always start from empty state?

4. Build the findings

For each anti-pattern you identify in the workload:

• Name the pattern
• Point to where in the code you see it (file and line)
• Explain what state space it leaves unexplored
• Describe what the workload could reach if this changed

The explanation should be concrete and specific to this workload, not a generic restatement of the anti-pattern. "Your three actions (put, get, delete) are always all active with equal weight — runs where only deletes fire would test behavior with zero keys" is better than "all actions are always active."

5. Tier the findings

Not all findings deserve equal weight. Tier them:

• High signal — The workload has a clear structural limitation that prevents it from reaching entire regions of state space. Example: all actions always active with no subsetting, all values hardcoded.
• Worth considering — The workload could explore more space with changes, but the limitation is more contextual or less severe. Example: no null/absent value variation, fixed concurrency level.
• Observation — Something the reviewer noticed that might matter depending on the system. Example: encoding variation, seed data identical across runs.

When you can't assess a finding's severity without domain knowledge (e.g., "this range is hardcoded, but I don't know what the system accepts"), say so. Present the structural fact and let the user supply the domain judgment.

6. Present the report

Structure the report as:

1. What the workload does — A brief summary of what the reviewer sees: what actions the workload takes, how it generates data, how it's structured. This is the reviewer's model of the workload, presented so the user can correct misunderstandings before reading findings.

2. Findings — Organized by tier (high signal first), grouped by category when multiple findings cluster. Each finding names the pattern, points to the code, and explains the unexplored state space in terms specific to this workload.

3. What's working well — Patterns the workload already does right. If the workload has per-run action subsetting, or varies payload sizes, say so. Findings-only reports feel like a critique; acknowledging what works frames the report as an assessment.

Tone

The report must read as "here's what you could explore that you're not reaching" — never as "your workload is bad." A workload exhibiting multiple anti-patterns may still be finding bugs; the goal is to help it find more.

Don't frame findings as "you need to do this." Frame them as "this is state space you're not reaching, and here's what reaching it would look like." The user decides what's worth changing.

Scope Boundaries

This skill reviews input generation — how the workload drives the system into diverse states. It does not review:

• Assertion quality — Whether the workload would notice a bug if it reached one. That's a separate concern.
• System correctness — Whether the system under test is behaving correctly. The SUT is a black box.
• Workload correctness — Whether the workload code has bugs. The reviewer reads the workload's structure, not its correctness.
• Infrastructure — Whether the Antithesis setup, Docker configuration, or test templates are correct.

If the reviewer notices something outside its scope that seems important (e.g., the workload has no assertions at all), it can mention it briefly as an observation, but should not review it in depth.

Reference Files