Oracle

Create a bundle you can hand to an external expert assistant (ChatGPT Pro, Claude, etc.) with real repository context.

This skill produces two artifacts:

• prompt.md — paste this into the expert assistant as your message
• context.zip — upload this to the expert assistant (contains selected repo files + MANIFEST.md)

Both artifacts are written to:

<repo_root>/.agents/oracle/<slug>/

Workflow

1. Pick the expert “role” (what mindset you want)

• debugging → senior engineer debugging with limited context
• code-review → staff engineer reviewing for correctness & maintainability
• architecture → principal engineer reviewing system design
• security → security engineer doing threat modeling
• performance → performance engineer identifying bottlenecks
• data-sql → database engineer reviewing correctness & performance
• ui-ux → expert UI/UX designer reviewing interaction & visuals
• general → general expert second opinion

2. Select a conservative file set Include the smallest set that allows a grounded answer:

• README / docs / ADRs relevant to the task
• The feature folder(s) under discussion
• Direct dependencies (callers/callees), config, types, error handling
• Repro artifacts: failing test output, logs, stack traces (redact secrets)

Do not include secrets: .env, API keys, credentials, private keys, prod configs.

3. Generate the bundle (script) Resolve scripts/oracle.py from this skill directory, and set the target repository with --repo-root (use your current project root):

python3 scripts/oracle.py \
  --repo-root "$PWD" \
  --task "What you want the expert to do" \
  --template debugging \
  --constraint "Key constraint" \
  --verify "Command(s) to validate locally" \
  --entry "path/to/folder::Main feature folder" \
  --entry "path/to/file.ts::Key code path"

The script writes:

• <repo_root>/.agents/oracle/<slug>/prompt.md
• <repo_root>/.agents/oracle/<slug>/context.zip

4. Hand off to the expert assistant

• Upload context.zip
• Paste the contents of prompt.md as the message
• After you get advice: verify locally (tests, logs, benchmarks)

Script flags (quick reference)

• --template {general|debugging|code-review|architecture|security|performance|data-sql|ui-ux}
• --role "Custom role string" (overrides the template default)
• --entry "PATH::REASON" (repeatable; PATH may be a file or directory)
• --entries-from <file> (one PATH::REASON per line)
• --exclude <glob> (repeatable)
• --max-file-bytes <int> (skip very large files; default 2,000,000)
• --estimate-tokens (best-effort token estimate)
• --dry-run (don’t write files; print what would be included)

Output contract for the expert assistant

The generated prompt asks the expert assistant to:

• Read MANIFEST.md first
• Use only evidence supported by the uploaded files
• Cite file paths for concrete claims
• Avoid questions; proceed with explicit assumptions
• Keep output structured and actionable (Answer → Key Points → Next Steps → Risks)