antithesis-workload

What this skill does

# Antithesis Workload

## Purpose and Goal

Implement or improve the Antithesis workload. Success means:

- Properties from the `antithesis-research` skill are mapped to concrete assertions, using both the property catalog and per-property evidence files
- Test commands exist under `antithesis/test/` and exercise the right behaviors
- The first real test templates are created after setup, or existing ones are expanded
- Triage findings turn into workload or property updates instead of staying implicit

Use the `antithesis-research` skill first to build the property catalog. Use the `antithesis-setup` skill to scaffold the infrastructure. Use the `antithesis-triage` skill to review runs, then return here to improve the workload. If the user asks to submit or launch a run, use the `antithesis-launch` skill — do not run `snouty launch` directly.

## Prerequisites

- DO NOT PROCEED if the Antithesis scratchbook (usually at `antithesis/scratchbook/`) doesn't exist. Use the `antithesis-research` skill to create it.
- DO NOT PROCEED if there is no `docker-compose.yaml` for Antithesis present. Use the `antithesis-setup` skill to create it.
- DO NOT PROCEED if `snouty` is not installed. See `https://raw.githubusercontent.com/antithesishq/snouty/refs/heads/main/README.md` for installation options.

# Scoping

Each invocation of the "Implement next property" workflow below focuses on **one property** to keep context manageable and quality high. (Post-triage iteration follows its own scoping based on triage findings.)

If the user asks for multiple properties, recommend doing one at a time — explain that implementation quality degrades as context accumulates, and each property's effort is unpredictable. Ask which one they'd like to start with. If they insist on multiple, proceed — but warn them first.

If the user specifies which property to work on, skip the full catalog scan — but still assess that property's status against its evidence file before proceeding. If it's already fully implemented, tell the user rather than redoing work.

If the user does not specify a property, run the full detection and recommendation flow below.

## Detect implementation status

The detection work below is context-heavy (reading every evidence file, scanning the codebase for assertions). If your agent supports sub-agents, delegate it to a sub-agent that returns a per-property summary (status + brief rationale). This keeps the main implementation agent's context clean.

The detection task: for each property in the catalog, search the existing test and SUT code for Antithesis SDK assertion calls and cross-reference them with the property's evidence file at `antithesis/scratchbook/properties/{slug}.md`. Assess whether the existing assertions cover the code paths, failure scenarios, and instrumentation points the evidence file describes. Classify each property as:

- **Implemented** — assertions cover what the evidence file describes
- **Partially implemented** — some assertions exist but coverage is incomplete
- **Not implemented** — no related assertions found

## Present and recommend

Read whatever provenance frontmatter is present in the catalog and include it when presenting status. Use whatever fields you find — the schema may evolve over time, so don't treat partial or older frontmatter as broken. Describe what's there in plain language. Examples:

- Full schema: "working from the catalog at commit `abc12345abcd` (2026-05-05), SUT path `/foo`, external refs: A, B."
- Partial: "working from the catalog at commit `abc12345abcd` (2026-05-05) — no SUT path or external refs recorded."
- Absent: "working from the catalog (no provenance recorded; predates the convention)."

Workload runs every property cycle, so don't ask the user to re-confirm provenance every run — display it as context and continue. The user will speak up if it no longer matches the system they're working on. The user-facing commit display uses the short hash (first 12 characters) for readability; the frontmatter still stores the full SHA.

**Recommend one property at a time — don't dump the full catalog.** A catalog can easily have dozens or hundreds of properties; an exhaustive status list is overwhelming and rarely what the user needs. This is the default: pick one, explain why, and wait for confirmation.

The exception is when the user explicitly asks to discuss what to work on rather than receive a single recommendation. In that case, give a high-level summary (counts by status, notable clusters, anything unusual) and have the conversation. Surface specific properties only as the conversation calls for them.

Strategy for picking the recommendation:

- **Getting started** — when few or no properties are implemented yet — recommend a **simple** property: one whose test doesn't have a lot of moving parts. The highest-priority properties tend to require the most setup and coordination, which is the wrong place to start. Simple properties build momentum, exercise the SDK and test harness end-to-end, and let you and the user develop a feel for the workflow. Tell the user this is the strategy you're using and why.
- **Once you're cooking** — multiple properties implemented and the harness validated end-to-end — switch to priority-based ordering: partially-implemented properties that need completion, then unimplemented properties that cluster with recently implemented ones (see `antithesis/scratchbook/property-relationships.md`), then other high-priority unimplemented properties. Tell the user when you're making this switch.

Explain **why** you picked the property you picked, and wait for the user to confirm or choose differently before proceeding.

For the chosen property, read both the catalog entry and its evidence file.

## Other scoping questions

Ask the user only for blockers or scoping decisions you cannot infer safely, such as:

- The property catalog location, if it is not the standard `antithesis/scratchbook/property-catalog.md`
- The project language or SDK choice, if the repo does not make it obvious
- Triage findings or known gaps, if iterating on an existing workload

## Definitions and Concepts

- **SUT:** System under test.
- **Test template:** A directory of test commands at `/opt/antithesis/test/v1/{name}/`. Each timeline runs commands from one test template. Files or subdirectories prefixed with `helper_` are ignored by Antithesis, so use that prefix for helper scripts kept alongside commands.
- **Test command:** An executable in a test template with a valid prefix: `parallel_driver_`, `singleton_driver_`, `serial_driver_`, `first_`, `eventually_`, `finally_`, `anytime_`.
- **Timeline:** One linear execution of the SUT and workload. Antithesis runs many timelines in parallel and branches them to search for interesting behaviors.
- **`Always` / `AlwaysOrUnreachable`:** Assertions for safety and correctness properties.
- **`Sometimes(cond)`:** Assertions for liveness or non-trivial semantic states that should occur at least once.
- **`Reachable` / `Unreachable`:** Assertions about whether meaningful outcomes or forbidden paths are exercised.

## Documentation Grounding

Use the `antithesis-documentation` skill to access these pages. Prefer `snouty docs`.

- Test commands reference: `https://antithesis.com/docs/test_templates/test_composer_reference.md`
- SDK reference: `https://antithesis.com/docs/using_antithesis/sdk.md`
- Properties and assertions: `https://antithesis.com/docs/properties_assertions/assertions.md`
- Fault injection: `https://antithesis.com/docs/environment/fault_injection.md`

## Reference Files

| Reference                                | When to read                                      |
| ---------------------------------------- | ------------------------------------------------- |
| `references/component-implementation.md` | Implementing workload-side components or wrappers |
| `references/assertions.md`               | Turning properties into SDK assertions