Consulting An Oracle

Consulting an Oracle

Overview

When in-session debugging has stalled, package the failed investigation into a single self-contained markdown file that a stronger one-shot model can answer cold. The oracle has zero project knowledge and no memory of prior runs — the prompt must stand alone.

Core principle: Just-enough context beats whole-repo dumps. A focused 1500-line prompt outperforms a 30k-line dump every time.

This skill produces a file. It does not call any model. Print the path and suggested invocation; let the user run it. This separation is deliberate: keeping prompt-construction decoupled from model-orchestration keeps the skill model-agnostic (works with any oracle the user picks — Codex, Claude, Gemini, paste-into-chat), keeps the artifact re-runnable later, and avoids coupling the skill to credentials, CLI availability, or cost decisions that belong to the user.

Announce at start: "I'm using the consulting-an-oracle skill to package this investigation."

When to Use

• 3+ failed fix attempts on the same symptom in this session
• User says "ask the oracle", "ask GPT-5", "write a letter to an expert", "draft a prompt for another model", "/tmp/letter.md"
• Investigation is consuming context with no convergence
• Before compaction, when the failed investigation should outlive this session as a one-shot prompt

Don't use when:

• You haven't actually tried to debug yet → use superpowers-ruby:systematic-debugging
• The bug is solved → use superpowers-ruby:compound
• Handing off to another Claude session with full plugin/tooling → use superpowers-ruby:handoff
• The user wants you to keep trying → keep trying

Output

Location: tmp/oracle/<YYYY-MM-DD>-<slug>.md (Rails projects gitignore tmp/ already). Fallback: /tmp/oracle-<slug>.md if tmp/ isn't writable or this isn't a Rails project.

Slug: short kebab-case description of the failing symptom — zeitwerk-constant-loop, n-plus-one-after-cache-add, turbo-stream-double-render. Not the branch name, not the ticket ID.

After writing, print:

Oracle prompt: tmp/oracle/2026-05-06-<slug>.md
Suggested invocation:
  codex --model gpt-5-pro --file tmp/oracle/2026-05-06-<slug>.md
  # or paste into chat.openai.com / claude.ai / gemini.google.com

The Process

Step 1: Detect Project Shape

Run these in parallel and capture output:

cat .ruby-version 2>/dev/null || cat .tool-versions 2>/dev/null
ruby --version
bundle --version 2>/dev/null
grep -E "^\s*(rails|sinatra|hanami|roda|rspec|minitest|sidekiq|good_job|solid_queue|sorbet|rbs|standard|rubocop)" Gemfile 2>/dev/null
test -f Gemfile.lock && grep -E "^\s+(rails|rack|puma|pg|mysql2|sqlite3|redis) \(" Gemfile.lock | head
test -f config/application.rb && grep -E "config\.(autoload|eager_load|cache_classes|active_job|active_record)" config/application.rb
ls config/initializers/ 2>/dev/null
test -f bin/rails && echo "rails app"
test -d app/javascript && ls app/javascript 2>/dev/null
test -f config/importmap.rb && echo "importmap"
test -f Procfile.dev && cat Procfile.dev

Extract: Ruby version, Rails (or other framework) version, DB adapter, test framework, background jobs, asset pipeline, type tooling, linter, Hotwire stack.

Step 2: Pull the Failure

From recent session context (read recent tool results, do not re-run failing commands unless they're cheap):

• The exact failing command (test, request, rake task)
• Verbatim stack trace, top-to-bottom — do not summarize
• Expected vs actual behavior in one sentence each

Step 3: Reconstruct What-Was-Tried

Walk recent session history (your own tool calls and edits in this conversation). Produce 3–7 entries, each:

- Hypothesis: <one line>
  Action:     <what you changed or ran>
  Outcome:    <what happened, including partial successes>

Include partial successes — "this fixed one of the failing tests but the other still fails" is high-signal for the oracle. Don't pad: if there were only 2 attempts, write 2.

Step 4: Pick Attached Files

Start from the failing file and walk one hop:

• The failing file itself
• Classes/modules it references (look at require, constant references, method calls on collaborators)
• The matching test file
• Any initializer in config/initializers/ that touches this area
• db/schema.rb excerpt (only the relevant tables) if ActiveRecord-related
• Relevant routes excerpt if request-handling

Hard cap: ≤8 files, ≤2000 total lines. If you'd exceed it, prefer fewer files with surrounding context over many files with no context.

Step 5: Redact Secrets

This step must run before Step 6, not after. Redacting an already-written file is leak recovery, not prevention — once secrets are on disk, they may be in editor swap files, OS-level backups (Time Machine, Dropbox, iCloud, tmp autosync), shell history, or the user's clipboard before you get a chance to scrub them. Redaction operates on the in-memory prompt body and file contents before the Write tool persists anything.

Scan the prompt body and every attached file for:

• Files: .env*, master.key, credentials.yml.enc, *.pem, *.key → never include, even if asked
• Patterns: (?i)(api[_-]?key|secret|token|password|bearer|authorization)\s*[:=]\s*['"][^'"]+['"]
• Patterns: connection strings with embedded credentials (postgres://user:pass@...)
• Patterns: long base64 / hex strings near words like "key", "token", "secret"

Replace with [REDACTED:<reason>] and keep enough surrounding context for the oracle to understand the structure. Note redactions in the safety footer.

Step 6: Write the File

Use the structure in template.md. Section order matters — role and desired output go first because oracles weight the opening of long prompts most heavily.

Step 7: Report

Print the file path, line count, and suggested invocation. Do not call any model. Do not open the file in an editor.

Ruby/Rails-Specific Suspect List

Always include this section in the oracle prompt — these are the implicit-context items that bite Ruby projects and that an oracle cannot infer:

• Autoloading: Zeitwerk vs Classic? Is the failing constant in app/, lib/, or an engine? Is lib/ in autoload paths? Eager-load mode in this environment?
• Frozen string literals: Magic comment present in the failing file?
• Thread safety: Puma worker/thread count, ActiveRecord connection pool size, any Thread.current usage in the call stack
• Initializer order: If config-related, what loads before this code? Custom initializers that monkey-patch?
• Gem version drift: git log -p Gemfile.lock for recently bumped gems on the failing path
• Monkey patches: Anything in config/initializers/, lib/core_ext/, or app/lib/ that reopens the failing class?
• ActiveSupport gotchas: present?/blank?/try collisions, delegate chains, with_options blocks
• Rails environment: Is the bug environment-specific (dev vs test vs prod)? config/environments/*.rb differences on the relevant flag?
• Background jobs: Is the failing code on a sync path or a job path? Adapter (Sidekiq/Solid Queue/GoodJob)?
• Schema vs model: Recent migrations not yet reflected in db/schema.rb?

Include only the items that might be relevant. Don't pad.

Anti-patterns

• ❌ Dumping the whole repo "to be safe" — destroys signal
• ❌ Summarizing the stack trace — oracles need the verbatim text
• ❌ Including .env "with secrets redacted by hand" — use the redaction step, never trust manual redaction of entire files
• ❌ Writing in second person to the user ("you should check...") — write in the voice of someone briefing a peer
• ❌ Asking the oracle an open-ended question ("what's wrong?") — specify the desired output format
• ❌ Including narrative about the session ("we first tried X, then Joe suggested Y") — the oracle doesn't know Joe; structure as hypothesis/action/outcome

Re-runnability

Oracles are one-shot — the model has no memory of prior runs. Write the prompt so that re-running it later (with the same --file argument) reproduces the same context. That means:

• Absolute file paths, not "the file we were just looking at"
• Git SHA at the top so the oracle knows the snapshot
• No references to "earlier in the session" or "as I mentioned"

Quick Reference

See template.md for the exact section layout.