Agent Skills: Verification Before Completion

Verification Before Completion

Overview

Sniper validates CODE QUALITY. Verification validates FUNCTIONAL RESOLUTION. Both are needed before closing any task.

Sniper catches linter errors, SOLID violations, and code style issues. Verification ensures the original request is actually fulfilled -- the right behavior, the right output, the right fix. A task can pass sniper with zero errors and still be functionally wrong.

| Aspect | Sniper | Verification | |--------|--------|-------------| | Focus | Code quality | Functional correctness | | Checks | Linting, SOLID, style | Acceptance criteria, regressions, side effects | | Runs | After any code change | Before marking task as complete | | Result | Clean code | Solved problem |

Agent Workflow

When invoked, follow the 6-step verification process below before marking any task as completed.

6-Step Verification Process

Step 1: Re-read the original request Go back to the original issue, task description, or user message. Read it word by word. Do not rely on memory or assumptions.

Step 2: List ALL acceptance criteria Extract every explicit and implicit requirement from the original request. Number them. If the request is vague, list what a reasonable user would expect.

Step 3: Verify each criterion with evidence For each criterion, provide concrete evidence of resolution:

• Test output showing the expected behavior
• Log output confirming the fix
• Screenshot of the UI change
• Code diff showing the implementation

Step 4: Check for regressions Run the full test suite. Compare results before and after. No new failures, no new warnings.

Step 5: Check for side effects Review every modified file. Confirm no accidental changes to unrelated code. Verify dependencies and configuration are unchanged unless required.

Step 6: Confirm functional resolution -- challenge, then write the artifact Before writing "Original problem is FUNCTIONALLY resolved," ALWAYS route the claim through the challenger agent (or challenge skill), fresh-context: claim = "functionally resolved" + evidence from Steps 3-5, NEVER the investigation reasoning. This is systematic -- every Verify gate, no exception, exactly like sniper runs at every eXamine. Only write the "FUNCTIONALLY resolved" verdict after a CONFIRMED result (or an UNCERTAIN explicitly accepted by the owner). A REFUTED verdict must be resolved (fix and re-verify) before the claim reaches the owner -- soft-gate, not a hard veto.

Write .claude/apex/docs/verify-{task-slug}.md (template: references/verify-template.md): every verification step checked, one evidence item per criterion (command output, log excerpt, screenshot path, or diff), plus the challenge verdict from this step. A context-only "it works" declaration does not survive a session boundary; the written artifact is the guardrail gates (sniper, later elicitation passes) actually check. Then state explicitly in your response: "Original problem is FUNCTIONALLY resolved" with a summary of evidence, or list what remains unresolved.

{task-slug}: derive per apex-methodology/references/init-tracking.md (git branch slug or active TaskCreate id) -- same pattern used by the elicitation skill's artifact.

Reference Guide

| Resource | Path | Content | |----------|------|---------| | Checklist | references/checklist.md | Full verification checklist with all categories | | Common Misses | references/common-misses.md | Frequently forgotten verification items | | Artifact Template | references/verify-template.md | verify-{task-slug}.md template + task-slug derivation |

Integration with APEX

Verification runs between eLicit and eXamine in the APEX workflow:

Analyze -> Plan -> Execute -> eLicit -> [VERIFICATION] -> eXamine (sniper)

This ensures functional correctness is confirmed before code quality validation. A task is only complete when both verification and sniper pass.

Critical Rules

| Rule | Reason | |------|--------| | Never skip re-reading the original request | Prevents solving the wrong problem | | Evidence required for every criterion | "It works" is not evidence | | Full test suite, not just new tests | Catches regressions | | Review ALL modified files | Catches accidental side effects | | Both verification AND sniper must pass | Quality without correctness is useless | | Step 6 writes verify-{task-slug}.md to disk | In-context self-review without persisted state regresses across sessions |