Agent Skills: Fixing Tests

Fixing Tests

Invariant Principles

1. Tests catch bugs, not checkmarks. Every fix must detect real failures, not just achieve green status.
2. Production bugs are not test issues. Flag and escalate; never silently fix broken behavior.
3. Read before fixing. Read the test file AND the production file under test. Never guess.
4. Verify proves value. Unverified fixes are unfinished fixes.
5. Scope discipline. Fix tests, not features.

Inputs

| Input | Required | Description | |-------|----------|-------------| | test_output | No | Test failure output to analyze (for run_and_fix mode) | | audit_report | No | Green mirage audit findings with patterns and YAML block | | target_tests | No | Specific test files or functions to fix (for general_instructions mode) | | test_command | No | Command to run tests; defaults to project standard |

Input Modes

Detect mode from user input, build work items accordingly.

| Mode | Detection | Action | |------|-----------|--------| | audit_report | Structured findings with patterns 1-10, "GREEN MIRAGE" verdicts, YAML block | Parse YAML, extract findings. Read patterns/assertion-quality-standard.md for assertion quality gate and Full Assertion Principle. | | general_instructions | "Fix tests in X", "test_foo is broken", specific test references | Extract target tests/files | | run_and_fix | "Run tests and fix failures", "get suite green" | Run tests, parse failures |

If unclear: ask user to clarify target.

WorkItem Schema

interface WorkItem {
  id: string;                           // "finding-1", "failure-1", etc.
  priority: "critical" | "important" | "minor" | "unknown";
  test_file: string;
  test_function?: string;
  line_number?: number;
  pattern?: number;                     // 1-10 from green mirage
  pattern_name?: string;
  current_code?: string;                // Problematic test code
  blind_spot?: string;                  // What broken code would pass
  suggested_fix?: string;               // From audit report
  production_file?: string;             // Related production code
  error_type?: "assertion" | "exception" | "timeout" | "skip";
  error_message?: string;
  expected?: string;
  actual?: string;
}

Phase 0: Input Processing

Dispatch subagent (Task tool) with /fix-tests-parse command. Subagent parses input (audit YAML, fallback headers, or general instructions) into WorkItems and determines commit strategy.

Phase 1: Discovery (run_and_fix only)

Skip for audit_report/general_instructions modes.

pytest --tb=short 2>&1 || npm test 2>&1 || cargo test 2>&1

Parse failures into WorkItems with error_type, message, stack trace, expected/actual. If suite passes completely: report "no failures found" and confirm with user before exiting.

Phase 2: Fix Execution

Dispatch subagent (Task tool). Subagent MUST read the referenced files:

First, read these files to understand the quality requirements:
- Read the fix-tests-execute command file for the fix execution protocol
- Read patterns/assertion-quality-standard.md for the complete Assertion Strength Ladder and Full Assertion Principle

Then execute the fix protocol on these work items: [work items]

[Copy in the full Test Writer Template from skills/dispatching-parallel-agents/SKILL.md before dispatching]

THE FULL ASSERTION PRINCIPLE (most important rule):
ALL assertions must assert exact equality against the COMPLETE expected output.
This applies to ALL output -- static, dynamic, or partially dynamic.
assert result == expected_complete_output  -- CORRECT
assert result == f"Today is {datetime.date.today()}"  -- CORRECT (dynamic: construct full expected)
assert "substring" in result               -- BANNED. ALWAYS. NO EXCEPTIONS.
assert dynamic_value in result             -- BANNED. Dynamic content is no excuse for partial check.

When fixing partial assertions on dynamic output: construct the complete expected value
using the same logic as the function, then assert ==. Prefer construct-then-compare
over normalization. Normalization is last resort only for truly unknowable values
(random UUIDs, OS-assigned PIDs, memory addresses).

When fixing partial mock assertions: also check whether ALL mock calls are fully asserted.
Assert EVERY call with ALL args; verify call count. NEVER use mock.ANY -- construct
the expected argument dynamically if it is dynamic.

BANNED PATTERNS (if your fix introduces ANY of these, it is NOT a fix):
- assert "X" in result (bare substring on any output -- static or dynamic)
- assert len(result) > 0 (existence only)
- assert result is not None without value assertion
- assert "X" in result and "Y" in result (multiple partials are still partial)
- assert result == function_under_test(same_input) (tautological)
- mock.ANY in any call assertion
- assert_called() or assert_called_once() without argument verification
- Asserting only some mock calls

Every assertion must be Level 4+ on the Assertion Strength Ladder.
Replacing a Level 1 assertion with a Level 2 assertion is NOT a fix.

2.1 Assertion Quality Gate (ALL modes)

1. Read patterns/assertion-quality-standard.md - the Full Assertion Principle and Assertion Strength Ladder
2. Classify each new/modified assertion on the Assertion Strength Ladder
3. REJECT any assertion at Level 2 (bare substring) or Level 1 (length/existence)
4. REJECT any fix that moves from one BANNED level to another (Pattern 10)
5. Level 3 (structural containment) requires written justification in the code
6. For each new assertion, name the specific production code mutation it catches
7. If you cannot name a mutation, the assertion is too weak; strengthen it

2.2 Production Bug Protocol

PRODUCTION BUG DETECTED

Test: [test_function]
Expected behavior: [what test expects]
Actual behavior: [what code does]

This is not a test issue - production code has a bug.

Options:
A) Fix production bug (then test will pass)
B) Update test to match buggy behavior (not recommended)
C) Skip test, create issue for bug

Your choice: ___

Do NOT silently fix production bugs as "test fixes." </CRITICAL>

Phase 3: Batch Processing

FOR priority IN [critical, important, minor]:
    FOR item IN work_items[priority]:
        Execute Phase 2
        IF stuck after 2 attempts:
            Add to stuck_items[]
            Continue to next item

Stuck Items Report

## Stuck Items

### [item.id]: [test_function]
**Attempted:** [what was tried]
**Blocked by:** [why it didn't work]
**Recommendation:** [manual intervention / more context / etc.]

Phase 3.5: Post-Fix Adversarial Review (MANDATORY)

Dispatch subagent (Task tool):

First, read these files to understand the quality requirements:
- Read patterns/assertion-quality-standard.md (especially The Full Assertion Principle)
- Copy in the full Test Adversary Template from skills/dispatching-parallel-agents/SKILL.md

ROLE: Test Adversary. Your job is to BREAK the new/modified test assertions.

## Context
- Modified test files: [list of files changed during fix phase]
- Git diff of changes: [paste or reference the diff]
- Production files under test: [paths]

## Mandatory Checks

1. IMMEDIATE REJECTION: Flag any assertion that is:
   - assert "X" in result on deterministic output (BANNED)
   - assert len(x) > 0 or assert x is not None (BANNED)
   - A fix that replaced one BANNED pattern with another (Pattern 10)

2. For each new/modified assertion:
   - Classify on Assertion Strength Ladder (must be Level 4+)
   - Determine if function under test is deterministic
   - If deterministic: only Level 5 (exact equality) is acceptable
   - Construct a plausible broken implementation that still passes
   - Verdict: KILLED or SURVIVED

3. Overall verdict:
   - Any SURVIVED or BANNED assertion: FAIL (list required re-fixes)
   - All KILLED + Level 4+: PASS

Return: Per-assertion verdicts and overall PASS/FAIL.

If verdict is FAIL: Re-execute Phase 2 for the failed items with explicit instructions about what went wrong. Do NOT skip re-review.

Phase 4: Final Verification

pytest -v  # or appropriate test command

Summary Report

## Fix Tests Summary

### Input Mode
[audit_report / general_instructions / run_and_fix]

### Metrics
| Metric | Value |
|--------|-------|
| Total items | N |
| Fixed | X |
| Stuck | Y |
| Production bugs | Z |

### Fixes Applied
| Test | File | Issue | Fix | Commit |
|------|------|-------|-----|--------|
| test_foo | test_auth.py | Pattern 2 | Strengthened to full object match | abc123 |

### Test Suite Status
- Before: X passing, Y failing
- After: X passing, Y failing

### Stuck Items (if any)
[List with recommendations]

### Production Bugs Found (if any)
[List with recommended actions]

Re-audit Option (if from audit_report)

Fixes complete. Re-run audit-green-mirage to verify no new mirages?
A) Yes, audit fixed files
B) No, satisfied with fixes

Special Cases

Flaky tests: Identify non-determinism source (time, random, ordering, external state). Mock or control it. Use deterministic waits, not sleep-and-hope.

Implementation-coupled tests: Identify the BEHAVIOR the test should verify. Rewrite to test through the public interface. Remove mocks of the unit under test's own internals; do not remove mocks of external services.

Missing tests entirely: Read production code. Identify key behaviors. Write tests following existing test file patterns in the codebase. Ensure tests would catch real failures.

Slow/bloated tests: Tests taking >5s often hide issues: heavy fixtures, unnecessary I/O, or oversized test data (e.g., 1024x1024 matrix where 4x4 suffices). Separate slow tests with marks (@pytest.mark.slow, @pytest.mark.integration, etc.). Shrink test inputs to the minimum that exercises the behavior. Move real I/O to integration tier. If a fixture takes longer than the test itself, it is too heavy for a unit test.

Over-Engineering

• Creating elaborate test infrastructure for simple fixes
• Adding abstraction layers "for future flexibility"
• Refactoring unrelated code while fixing tests

Under-Testing

• Weakening assertions to make tests pass
• Removing tests instead of fixing them
• Marking tests as skip without fixing

Scope Creep

• Fixing production bugs without flagging them
• Refactoring production code to make tests easier
• Adding features while fixing tests

Blind Fixes

• Applying suggested fixes without reading context
• Copy-pasting fixes without understanding them
• Not verifying fixes actually catch failures </FORBIDDEN>

Self-Check

<RULE>Before completing, ALL boxes must be checked. If ANY unchecked: STOP and fix.</RULE>

• [ ] All work items processed or explicitly marked stuck
• [ ] Each fix verified to pass
• [ ] Each fix verified to catch the failure it should catch
• [ ] Each fix verified to be Level 4+ on the Assertion Strength Ladder (patterns/assertion-quality-standard.md)
• [ ] Each new assertion has a named mutation that would cause it to fail
• [ ] No bare substring checks introduced (assert "X" in result is BANNED on all output)
• [ ] No partial assertions on dynamic output (full expected constructed, not membership checked)
• [ ] All mock calls fully asserted: every call, all args, call count verified
• [ ] No mock.ANY introduced
• [ ] No Pattern 10 violations (partial-to-partial upgrades)
• [ ] Phase 3.5 adversarial review completed with PASS verdict
• [ ] Full test suite ran at end
• [ ] Production bugs flagged, not silently fixed
• [ ] Commits follow agreed strategy
• [ ] Summary report provided

<FINAL_EMPHASIS> Tests exist to catch bugs. Every fix you make must result in tests that actually catch failures, not tests that achieve green checkmarks.

Fix it. Prove it works. Move on. No over-engineering. No under-testing. </FINAL_EMPHASIS>