Antithesis Workload

Antithesis Workload

Skill version: 2026-08-24 d6b507d

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/product/test_templates/test_composer_reference.md
• SDK reference: https://antithesis.com/docs/reference/sdk.md
• Properties and assertions: https://antithesis.com/docs/concepts/properties_assertions/assertions.md
• Fault injection: https://antithesis.com/docs/product/fault_injection.md

Reference Files

Recommended Workflows

Implement next property

1. Detect implementation status, then recommend one property and explain why (see "Detect implementation status" and "Present and recommend" above)
2. Get user confirmation on which property to implement
3. Read references/component-implementation.md
4. Read references/assertions.md
5. Read references/test-commands.md
6. Read references/interesting-values.md
7. Identify the property's value menu — boundary values plus configured-limit families. See interesting-values.md for the discovery process and worked examples
8. Implement the chosen property: assertions, test commands, and supporting code
9. Write Sometimes assertions as reach claims for the behaviors the workload is designed to drive (see references/assertions.md, "Sometimes Assertions as Workload Reach Claims"). Assert the precondition, not the violation. Use antithesis-launch to run a test, then antithesis-triage to check which assertions fired — unfired ones are evidence the workload isn't reaching its targets and the starting point for iteration

Post-triage iteration

1. Read references/iteration.md
2. Read references/assertions.md if assertions need to change
3. Read references/test-commands.md if command coverage needs to change
4. Update the workload and the relevant Antithesis scratchbook artifacts together

General Guidance

• Keep Antithesis-only code out of production paths. If you must touch shared code, make the change surgical and easy to wall off.
• Prefer simple workload code over highly configurable abstractions.
• Design workloads to be fault-tolerant. Unlike happy-path tests where errors signal bugs, workloads run under fault injection and concurrent activity — transient errors are expected, not exceptional. Make progress toward a goal rather than bail on the first failure.
• Assume antithesis-setup has already made the system runnable in a mostly idle state; this skill owns what the workload does once the system is up.
• Assume antithesis-setup has already installed the relevant SDK and added one minimal bootstrap assertion in the SUT. This skill owns the broader property catalog beyond that initial integration check.
• Write test commands in the project's language, not Bash, so they can reuse the project's clients, helpers, and libraries.

Output

• Assertions for every property in scope, in workload code or carefully chosen SUT locations
• Test commands and supporting workload code under antithesis/test/
• Updates to antithesis/scratchbook/property-catalog.md when the implemented properties change

Self-Review

Before declaring this skill complete, review your work against the criteria below. If your agent supports spawning sub-agents, create a new agent with fresh context to perform this review — give it the path to this skill file and have it read all output artifacts. A fresh-context reviewer catches blind spots that in-context review misses. If your agent does not support sub-agents, perform the review yourself: re-read the success criteria at the top of this file, then systematically check each item below against your actual output.

Review criteria:

• For every property in scope, the implementation covers the code paths, failure scenarios, and instrumentation points described in its evidence file — not just "an assertion exists" but "the assertions cover what the evidence says needs to be covered"
• Each assertion uses the correct SDK assertion type for its property's semantics (Always/AlwaysOrUnreachable for safety, Sometimes(cond) for liveness or meaningful semantic state, Reachable/Unreachable for path and outcome checks)
• Sometimes(true, ...) assertions should be rewritten as Reachable(...).
• Assertion property names are inline constant string literals, unique across the whole project — never built at runtime or passed through variables (static analysis must pre-catalog them; see references/assertions.md "Naming"); no broad property is implemented by reusing one message at multiple unrelated callsites
• Workload-only instrumentation was not used where surgical SUT-side assertions would provide materially better search guidance for rare, dangerous, or timing-sensitive internal states
• Workload code makes progress toward stated goals under transient errors rather than bailing on first failure
• The workload records both attempted and acknowledged operations so later assertions can check bounds (e.g., "counter changed by some value in [acknowledged, attempted]")
• Where retries are used, they're consistent with the SUT's idempotency contract
• Reachable(...) markers are attached to distinct outcomes or branch results, not redundant early path-entry locations on the same straight-line flow
• For bounded inputs in test commands, draws come from property-specific value menus (boundary values for the input type plus configured-limit families from the property's code paths) rather than arbitrary ranges, or the test command documents that the menu axis is not applicable. See interesting-values.md.
• Test commands exist under antithesis/test/ and use valid prefixes (parallel_driver_, singleton_driver_, serial_driver_, first_, eventually_, finally_, anytime_)
• Test commands are written in the project's language, not Bash, and reuse the project's clients and libraries where possible
• No test command is responsible for Antithesis lifecycle signaling; setup_complete is emitted before test commands begin
• Test templates are structured correctly at the path that will map to /opt/antithesis/test/v1/{name}/ in the container
• Helper files or directories are prefixed with helper_ so Antithesis ignores them
• Whatever provenance frontmatter was present in the catalog was displayed when presenting status, so the user could spot a mismatch with the system they're working on
• antithesis/scratchbook/property-catalog.md is updated to reflect the implementation status of every property in scope, with provenance frontmatter reflecting the current codebase state (refresh commit and updated; preserve sut_path and external_references from the existing catalog). The frontmatter format is defined in the antithesis-research skill, references/scratchbook-setup.md.
• Sometimes assertions exist as reach claims for the behaviors the workload is designed to drive, and unfired reach claims from a test run have been used as the iteration signal (see references/assertions.md, "Sometimes Assertions as Workload Reach Claims")
• No reach claim is the negation of an Always or AlwaysOrUnreachable at the same site, and each reach claim still fires when the system is correct (see references/assertions.md, "A Reach Claim Asserts The Precondition, Never The Violation")
• Assertions are in workload code or surgical SUT locations — not scattered across production paths
• Use snouty validate on antithesis/config to ensure that the compose setup can reach setup complete and any configured test-templates work. Make sure to build the latest images before running validate.