Agent Skills: Codex Skill Guide

Codex Skill Guide

Running a Task

Defaults

• Model: gpt-5.2-codex
• Reasoning effort: medium
• Sandbox: workspace-write

Invocation Modes

Mode 1: Template invocation (from commands like /ultra-dev, /ultra-test)

• Use template config directly, NO user interaction
• Templates define model/effort/sandbox/prompt

Mode 2: Regular invocation (user requests codex directly)

1. Display current defaults
2. Use AskUserQuestion:
   • Option A: "Use default config" (Recommended) - gpt-5.2-codex, medium, workspace-write
   • Option B: "Custom config" - then ask model/effort/sandbox separately
3. Execute with chosen config

Configuration Options

Models:

• gpt-5.2-codex (default, optimized for code)
• gpt-5.2 (general purpose)

Reasoning effort:

• low - fast, simple tasks
• medium (default) - balanced
• high - complex analysis
• xhigh - maximum reasoning

Sandbox:

• workspace-write (default) - can run git/ls for context
• read-only - analysis only, no file access
• danger-full-access - requires explicit user permission

Command template

codex exec \
  -m gpt-5.2-codex \
  -c model_reasoning_effort="medium" \
  --sandbox workspace-write \
  --skip-git-repo-check \
  "prompt here"

Execution rules

• Do NOT use 2>/dev/null - stderr contains important error info
• Run the command and show complete output to user
• After completion: "You can resume with 'codex resume'"

Resume Syntax

# Resume with new prompt (correct syntax)
codex exec resume --last "new prompt here"

# Resume reading prompt from stdin
echo "new prompt" | codex exec resume --last -

# Resume with config overrides (flags BEFORE resume)
codex exec -m gpt-5.2-codex resume --last "prompt"

Quick Reference

| Use case | Command | |----------|---------| | Analysis | codex exec -m gpt-5.2-codex --sandbox workspace-write --skip-git-repo-check "prompt" | | With edits | codex exec -m gpt-5.2-codex --sandbox workspace-write --full-auto --skip-git-repo-check "prompt" | | Resume | codex exec resume --last "continue with..." | | Code review | codex exec review (built-in subcommand) |

Following Up

• After every codex command, use AskUserQuestion to confirm next steps or whether to resume.
• When resuming, the session inherits original model/sandbox settings unless overridden.

Error Handling

• If codex exec exits non-zero, show the error and ask user for direction.
• --full-auto requires explicit confirmation in Mode 2 custom config flow.
• --sandbox danger-full-access requires explicit user permission (separate confirmation).
• If output shows warnings, summarize and ask how to proceed.

Review Templates

Use these predefined templates when commands reference codex skill with template: <name>.

research-review

| Config | Value | |--------|-------| | Model | gpt-5.2-codex | | Effort | medium | | Sandbox | read-only |

Prompt:

Review this technical research output against these rules:

[Evidence-First]
- Every claim must have verifiable source (official docs, benchmarks)
- Unverified claims must be marked as "Speculation"
- Priority: 1) Official docs 2) Community practices 3) Inference

[Honesty & Challenge]
- Detect risk underestimation or wishful thinking
- Point out logical gaps explicitly
- No overly optimistic assumptions without evidence

[Architecture Decisions]
- Critical state requirements addressed?
- Migration/rollback plan for breaking changes?
- Persistence/recovery/observability considered?

[Completeness]
- Missing risks or edge cases not considered
- Contradictions between sections

Provide specific issues with file:line references.
Label each finding: Fact | Inference | Speculation
If no critical issues found, respond with "PASS: No blocking issues".

code-review

| Config | Value | |--------|-------| | Model | gpt-5.2-codex | | Effort | high | | Sandbox | read-only |

Prompt:

Review this code diff against these rules:

[Code Quality]
- No TODO/FIXME/placeholder in code
- Modular structure, avoid deep nesting (max 3 levels)
- No hardcoded secrets or credentials

[Security]
- No injection vulnerabilities (SQL, XSS, CSRF, command injection)
- No auth bypass or secrets exposure
- Input validation at system boundaries

[Architecture]
- Critical state (funds/permissions/external API) must be persistable/recoverable
- No in-memory-only storage for critical data
- Breaking API changes require migration plan

[Logic]
- No race conditions or incorrect state handling
- No N+1 queries or memory leaks
- Spec compliance - implementation matches acceptance criteria
- Edge cases handled (boundary values, null, empty, error paths)

[Testing in Code]
- No mocks on core logic (domain/service/state paths must use real deps)
- Test files included should follow Core Logic NO MOCKING rule

Provide specific issues with file:line references and severity (Critical/High/Medium/Low).
If no critical/high issues found, respond with "PASS: No blocking issues".

test-review

| Config | Value | |--------|-------| | Model | gpt-5.2-codex | | Effort | medium | | Sandbox | workspace-write |

Prompt:

Review this test suite against these rules:

[Core Logic Testing - NO MOCKING ALLOWED]
Core Logic = Domain/service/state machine/funds-permission paths
- These paths MUST use real implementations, not mocks
- Repository interfaces: prefer testcontainers with production DB
- Fallback: SQLite/in-memory only when testcontainers unavailable

[External Systems - Test Doubles ALLOWED]
- External APIs, third-party services → testcontainers/sandbox/stub OK
- Must document rationale for each test double

[Coverage]
- Missing edge cases (null, empty, boundary values, error paths)
- Untested critical paths (auth flows, payment, data mutations, deletions)

[Anti-Patterns]
- Flaky tests (time-dependent, order-dependent)
- Tautology assertions (expect(true).toBe(true))
- Empty test bodies
- False confidence - tests that pass but don't verify behavior

[Security Testing]
- Auth/permission tests exist for protected endpoints
- Input validation tests for injection vectors
- Sensitive data handling tests (no plaintext secrets in logs/responses)

Provide specific issues with file:line references.
If no critical issues found, respond with "PASS: No blocking issues".