# Generate Regression Cases

> Convert incident reports and audit failures into candidate regression cases that prevent re-occurrence — each case captures a concrete observable symptom, the target EOU, the failure class (F-code), and an expected-behavior predicate. Cases are written as candidates; activation requires human owner approval. <example> Context: A diagnosis was filed for a path-drift incident; owner wants regression coverage so the failure cannot return silently. user: "$generate-regression-cases foundry/incidents/inc-0042.yml" assistant: "I'll extract the observable symptom, classify it under the F-code taxonomy, and write a regression case under foundry/self-evolution/regression/cases/. activation_status defaults to candidate." </example> <example> Context: User wants regression cases for a batch of audit findings. user: "$generate-regression-cases foundry/audits/eou-audits/*.audit.yml" assistant: "I'll iterate each finding, generate at most one regression case per observable symptom, and refuse to generate cases for findings

- Skill: `xiaolai/generate-regression-cases` (Agent Skill)
- Install (CLI): `npx skillmds@latest add xiaolai/generate-regression-cases`
- Raw SKILL.md: https://api.skillmd.com/api/skills/xiaolai/generate-regression-cases/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: xiaolai (https://skillmd.com/u/xiaolai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/xiaolai/generate-regression-cases

---


# Generate Regression Cases

Convert `$source` (one or more incident reports or audit failures) into candidate regression cases.

## Inputs

- `$source` (required) — path to one or more incident reports or audit failure files. Accepts a direct file path or a glob (e.g., `foundry/audits/incidents/*.yml`). Every matched file must contain at least one observable failure symptom.

## Required reading

1. `foundry/failure-taxonomy.yml` — classify each failure by F-code
2. `foundry/constitution.yml` — invariants that constrain what counts as a valid regression case
3. `schemas/incident.schema.yml` — validate structured YAML incident files against this schema
4. `schemas/regression-case.schema.yml` — the schema for the output regression case files

## Stop conditions

Halt and report before generating cases if:
- `$source` resolves to no readable files.
- No observable failure symptom can be extracted from any source file — labels and conclusions without at least one observable symptom (error message, wrong output, missing artifact, schema mismatch) are not accepted as source input.
- `schemas/regression-case.schema.yml` does not exist — cannot validate output structure.

## Procedure

1. Load `$source`. If it is a glob, read all matching files.
2. For each incident or audit failure, extract the observable failure symptom.
3. Classify the failure using `foundry/failure-taxonomy.yml` (F1–F12).
4. Derive the regression case: the minimal input that reproducibly triggers the failure.
5. Construct the fixture path: `foundry/self-evolution/regression/fixtures/{case_id}/`.
6. Check `foundry/self-evolution/regression/cases/` for existing cases that cover this failure — skip if a duplicate exists; record the duplicate case ID instead.
7. For each new case, write the regression case file with all required fields.
8. Set `activation_status: candidate` on all generated cases — never `active`.

## Required fields per case (per `schemas/regression-case.schema.yml`)

```yaml
id:                   # regression-{eou_id}-{failure_class}-{YYYYMMDD}
source_incident_id:   # incident or audit ID this case was derived from
failure_class:        # F1-F12 code
eou_id:               # EOU that exhibited the failure
fixture_path:         # foundry/self-evolution/regression/fixtures/{case_id}/
failure_shape_assertion:  # observable output that indicates the failure is occurring (non-empty)
activation_status:    # candidate — never active without human approval
```

## Output

Write each case to `foundry/self-evolution/regression/cases/{case_id}.regression.yml`.
Create the fixture directory at `foundry/self-evolution/regression/fixtures/{case_id}/`.
Record the run in `foundry/runs/{eou_id}/{run_id}.yml`.

## Constraints

- Do not set `activation_status: active`. Cases are candidates until the human owner promotes them.
- Do not generate cases without a concrete source incident or audit failure.
- Each case must include at least one `failure_shape_assertion` — what observable output signals the failure is occurring. Cases asserting only pass are invalid.
- If a duplicate case exists for this failure, record the duplicate ID and skip generation.

## Scope Note

**Upstream:** receives incidents (`foundry/incidents/`) or audit findings (`foundry/audits/eou-audits/`).

**Downstream:** produces regression cases at `foundry/self-evolution/regression/cases/` at activation_status candidate. Cases enter the regression suite only after human owner approval — this skill never sets activation_status to active.

**Related:** `$eou-diagnose` (sibling — F-code classification of the same incident); `$ecp-propose` (sibling — proposes change; regression cases pin the failure so the change can be verified).

**Pipeline:** `incident | audit failure → generate-regression-cases → human approval → activation_status: active`