Testing Matrix
Testing Matrix turns a skill, agent, or prompt-driven workflow into realistic behavioral test prompts with explicit routing, pass criteria, and guided result tracking.
What it tests
- Positive and natural-language activation.
- Nearby requests that should not trigger the target.
- Main-path behavior and promised outputs.
- Missing or ambiguous inputs.
- Dependency failures and safe recovery.
- Authorization, privacy, and destructive-action boundaries.
- Stateful follow-ups, reruns, and regressions.
Tests target observable behavior rather than preferred wording.
Modes
- Matrix only: produce the test plan without running it.
- Guided run: walk through one prompt at a time and record results.
- Results review: evaluate responses or artifacts supplied by the user.
- Regression update: revise coverage after the target changes.
- Repair and retest: diagnose failures; target edits still require separate authorization.
For a general “help me test this” request, the skill defaults to a guided standard run.
Installation
Install the entire codex-testing-matrix directory through your normal Codex skill or plugin workflow. Keep SKILL.md and agents/ together.
Example prompts
Use $codex-testing-matrix to create a standard behavioral test matrix for this skill and guide me through it one prompt at a time.
Create release-level regression coverage for this agent, including dependency failures and permission boundaries.
Review these target-session responses against the attached test matrix and mark each row Pass, Fail, Blocked, N/A, or Not run.
How guided testing works
The matrix tells the user where each test belongs:
- a fresh target session for independent activation and boundary checks;
- the same target session for deliberate stateful sequences;
- an isolated workspace for scripts and file behavior;
- the target application for product-specific behavior;
- the current conversation for planning and result review.
Expected outcomes remain outside the copy-paste test prompt so the system under test is not coached.
Result states
Rows are recorded as Pass, Fail, Blocked, N/A, or Not run. Record fixes and prior failures in the Retest history. Unclear observations stay Blocked with the missing evidence identified. Missing runtime access remains blocked; untested behavior is never converted into a pass.
License
No license has been assigned by this package. Add one before public redistribution.