| name | create-oracle-review-powered-goal-plan |
| description | Create compact Codex `/goal` plan files that require an implementation loop with Oracle GPT-5.5 Pro review between iterations. Use when the user wants a goal plan that explicitly sequences implementation, strongest-model Oracle review, rework from Oracle findings, repeated review until no material changes remain, and final verification. |
Create Oracle Review Powered Goal Plan
Overview
Create a concise Markdown handoff for Codex CLI /goal <path> that embeds an Oracle-reviewed implementation loop. This skill specializes create-goal-plan: the goal file remains short, but the process requires implementation, local validation, Oracle GPT-5.5 Pro review, rework, and repeated Oracle review until the remaining recommendations are empty or explicitly rejected with rationale.
Workflow
-
Resolve the task context and full plan.
- Prefer the active task
status.md, accepted plan entry, and dedicated plan artifacts under task-progress-artifacts/.
- If no plan is accepted yet, do not imply approval. The goal plan may say that implementation begins only after explicit human acceptance and normal task-status persistence.
- Cite the full plan/status paths in
Context instead of duplicating long planning detail.
-
Write a compact goal plan.
- Use the same shape as
create-goal-plan: first line is one concrete outcome, then exactly Context:, Constraints:, Process:, and Done when:.
- Keep it roughly one screen. The complete architecture, plan, and rationale belong in cited artifacts.
- Include dirty-worktree, privacy, secrets, destructive-action, and external-API boundaries when relevant.
-
Make the Oracle loop explicit in Process.
- Implement in focused phases from the accepted plan.
- Run repo-local tests, linters, validators, smoke tests, or reproductions after each meaningful phase.
- Prepare a tight Oracle prompt with the changed files, acceptance criteria, validation commands already run, known risks, and the exact review question.
- Run the local wrapper with GPT-5.5 Pro browser/manual-login mode:
oracle --engine browser --browser-manual-login --browser-chrome-path "$HOME/pro/botfiles/bin/oracle-chrome-linux" --model "5.5 Pro" --browser-model-strategy select -p "<review prompt>" --file "<focused paths>"
- Wait or reattach until Oracle completes. Do not downgrade for normal latency,
in_progress, prompt-size issues, polling-shell death, or browser upload friction; compact or reattach first. Fallback from GPT-5.5 Pro only after repeated availability or access failures.
- Save Oracle prompt, response, and relevant logs under the task's
task-progress-artifacts/scratchpad/.
- Classify Oracle feedback as must-fix, optional, or rejected-with-rationale. Implement must-fix items, rerun local validation, and repeat Oracle review on the revised diff until no material changes remain.
- Report the created file and invocation.
- Tell the user the exact path.
- Include the exact
/goal <path> command they can paste.
Oracle Review Rules
- Use the local
oracle wrapper, not raw npx, unless debugging the wrapper itself.
- Attach the fewest files that still contain the truth. Never attach
.env, secrets, private keys, raw tokens, or unrelated private data.
- Ask Oracle for findings, risk assessment, concrete recommendations, and test gaps. Do not ask for hidden chain-of-thought.
- Treat Oracle as advisory but high-signal. If rejecting a recommendation, write a short rationale and keep it with the review artifact or final notes.
- Repeat review only after meaningful changes. If the last Oracle response has no material changes or only rejected optional suggestions, the loop can close.
Goal Plan Template
<Implement/fix/build the concrete outcome, using Oracle GPT-5.5 Pro review until no material changes remain.>
Context:
- <full accepted plan/status/artifact paths>
- <files or modules to inspect first>
Constraints:
- <do not change / privacy / secrets / external API / dirty worktree boundaries>
- Use Oracle GPT-5.5 Pro browser/manual-login review; do not downgrade for normal latency.
Process:
- Implement from the accepted plan in scoped phases and run local validation.
- Prepare a focused Oracle review prompt with changed files, criteria, tests run, and known risks; run the explicit GPT-5.5 Pro browser command.
- Save Oracle prompt/response/logs, implement must-fix feedback, rerun validation, and repeat Oracle review until clean or remaining suggestions are rejected with rationale.
Done when:
- <objective verification passes>
- <tests/checks pass>
- Final Oracle GPT-5.5 Pro review has no material changes.
- Final answer summarizes implementation, validation commands, and Oracle iterations.