agentleFS
GoogleKnown publishergoogle/heir/.agents/skills/reproduce_failure/SKILL.md

reproduce-failure

Skill948 starsChanged 7 months agoApache-2.0View on GitHub

Describes a decision tree of steps and skills to utilize when, starting from a test failure or the failure of `bazel run` command using heir-opt, you would like to produce a reproducing input IR that fails on a particular compiler pass.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run it to ensure it reproduces the user's error.

    line 56
  2. Very specific:

    Use the dump_intermediate_ir skill to augment the bazel run command with appropriate flags, so that it writes relevant files that contain …

    line 60
  3. Very specific:

    If you know of a particular GitHub issue number, use the number, like issue_1480.mlir.

    line 130
  4. Very specific:

    If you have a starting test file or target, name it similar to that file, e.g., dot_product_8f_regression.mlir

    line 132
  5. Very specific:

    If it makes sense, incorporate the name of the pass, e.g., dot_product_secret_to_ckks.mlir

    line 134
  6. Very specific:

    If no natural name makes sense, use RENAME_ME_regression_test.mlir

    line 136
  7. Distinctive:

    Ensure that at each step you can still reproduce the failure.

    line 140

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Reproduce a test failure

Overview

This skill guides the agent in identifying an IR and compiler pass that reproduces a failure from a lit or e2e test target, or from a bazel run command that runs heir-opt.

Usage

To reproduce a failure, follow these steps.

Identify the starting point of the process

The user may input one of multiple "starting points" for a failure.

  1. A direct bazel run command.
  2. A lit test target or bazel test command pointing to such a target, such as //tests/Transforms/convert_to_ciphertext_semantics:assign_layout.mlir.test,
  3. A file path to a lit test, such as third_party/heir/tests/Transforms/convert_to_ciphertext_semantics/assign_layout.mlir
  4. An end-to-end test target, such as //tests/Examples/lattigo/ckks/dot_product_8f:dotproduct8f_test,
  5. A filepath to an end-to-end test file, such as third_party/heir/tests/Examples/lattigo/ckks/dot_product_8f/dot_product_8f_test.go

Reiterate to the user that you understand which of these options are the starting point, and if the user provides a starting point that doesn't fit into one of these options, warn them and then ask for confirmation when improvising how to convert their failure to a bazel run command in the next step.

Identify the right bazel run command

For option 1 (A direct bazel run command), nothing is needed, continue to the next step.

For option 2, use the lit_to_bazel skill.

For option 3, use the e2e_to_bazel skill.

At the end of this step, you should have a bazel run command that reproduces the user's error, but that command may run many compiler passes in a pipeline. The remaining steps will reduce the bazel run command to a more minimal reproducer.

Run it to ensure it reproduces the user's error.

Identify the specific IR, pass, and pass options

Use the dump_intermediate_ir skill to augment the bazel run command with appropriate flags, so that it writes relevant files that contain the IR and pass options.

Then inspect the right file (usually the last file dumped before an error occurs) to identify the IR and pass options. The dumped IR should include near the top of the file, a comment like

// -----// IR Dump Before CanonicalizerPass: canonicalize{cse-between-iterations=false max-iterations=5 max-num-rewrites=-1 region-simplify=normal test-convergence=false top-down=true} //----- //

The part after : is the relevant pass name and options that will be needed in the next step. The rest of the file content is the IR that will be needed in the next step.

Construct the bazel run command for the individual pass

Given the information from previous steps, construct a bazel run command of the form:

bazel run //tools:heir-opt -- \
--pass-pipeline="<PASS_AND_OPTIONS>" \
<IR_FILE>

where <PASS_AND_OPTIONS> corresponds to the "pass name and options" from the previous step, and <IR_FILE> corresponds to the dumped IR for that pass.

At this step, run the new bazel run command and make sure that it reproduces the error.

Create a reproducing lit test file

Copy the IR to a new lit regression test file with the bazel run command modified to a // RUN: line.

The new test should live in third_party/heir/tests/Regression/ and have at the top of the file a comment describing the origin of the reproducer along with the // RUN: line.

For example, if the reduced bazel run command from the previous step was

bazel run //tools:heir-opt -- \
--pass-pipeline="canonicalze{cse-between-iterations=true}" \
/tmp/mlir/foo.mlir

and the original command was bazel run //tools:heir-opt -- --big-pipeline /path/to/file.mlir

Then the corresponding lit file should start with

// This file is a minimal reproducer of a failure originally produced with
//
//    heir-opt --big-pipeline /path/to/file.mlir
//
// RUN: heir-opt --pass-pipeline="canonicalze{cse-between-iterations=true}" %s

Below that header, place the reproducing IR from the previous step. Then confirm the test exercises the test failure once more by running bazel test //tests/Regression:<name_of_file.mlir>.test.

The <name_of_file.mlir> should be chosen appropriately according to the following heuristic:

  • If you know of a particular GitHub issue number, use the number, like issue_1480.mlir.
  • If you have a starting test file or target, name it similar to that file, e.g., dot_product_8f_regression.mlir
  • If it makes sense, incorporate the name of the pass, e.g., dot_product_secret_to_ckks.mlir
  • If no natural name makes sense, use RENAME_ME_regression_test.mlir

Gotchas

  • Ensure that at each step you can still reproduce the failure. If at any step, a newly reduced command fails to reproduce the failure, stop and report this to the user, asking for assistance.
Copy this checklist and track progress:

- [ ] Step 1: Identify the starting point of the process
- [ ] Step 2: Identify the right `bazel run` command
- [ ] Step 3: Identify the specific IR, pass, and pass options
- [ ] Step 4: Construct the `bazel run` command for the individual pass
- [ ] Step 5: Create a reproducing `lit` test file

Each step corresponds to a section mentioned above.
<!-- mdformat global-off -->

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.