martin-hausleitner/martins-awesome-skills/oracle-ai-research-e2e

Oracle AI Research E2E

Use this skill to prove that Peter Steinberger's Oracle 0.13 layer and the local ai-research-browser CLI work together, without turning Oracle into a bypass for account, plan, model, quota, or screenshot guards.

The public-safe local harness is:

python3 skills/software-development/oracle-ai-research-e2e/scripts/oracle_ai_research_e2e_check.py

What To Prove

Check all of these before claiming the integration works:

• @steipete/oracle@0.13.0 appears in generated commands.
• oracle-plan emits --browser-attach-running, --remote-chrome, --browser-research, status, reattach, and session --render.
• workflow-run --oracle-mode assist|runner writes Oracle evidence into the same workflow payload.
• runner mode blocks when local login/account/plan/feature/screenshot guards fail.
• ChatGPT Pro/Extended Pro/GPT-5.5 Pro are blocked before UI typing.
• Slash feature fallback does not type into the composer before the pre-submit guard.
• Public GitHub artifacts are sanitized; no real cookies, tokens, user paths, prompts, or private screenshots are committed.

Safe Default Check

From the repo root:

python3 skills/software-development/oracle-ai-research-e2e/scripts/oracle_ai_research_e2e_check.py \
  --json

This runs only local, public-safe checks:

• syntax compile of ai_research_browser.py
• Oracle-focused unit tests
• ChatGPT model-safety and slash-guard regression tests
• oracle-plan command generation
• workflow-run --oracle-mode runner blocked-path proof
• oracle-e2e-smoke opt-in blocking proof
• public-safety audit

On Linux CI hosts without Brave/Comet/Chrome profiles, the browser-dependent workflow-run runner proof is reported as skipped: browser-not-discovered-on-this-host; on the local macOS Hermes machine it should run and prove blocked-by-local-guards.

Real Provider E2E

Run real E2E only when the browser CDP session and paid-quota use are intentional:

AI_RESEARCH_BROWSER_E2E=1 \
python3 skills/software-development/oracle-ai-research-e2e/scripts/oracle_ai_research_e2e_check.py \
  --live \
  --browser brave \
  --profile work \
  --provider chatgpt \
  --mode thinking \
  --cdp-port 9223 \
  --json

Use one flow first. Do not start a broad matrix until the single target has status=verified or an equally strong captured output plus screenshot evidence.

Local Hermes Install

To install this skill into Hermes:

mkdir -p ~/.hermes/skills/software-development
cp -R skills/software-development/oracle-ai-research-e2e ~/.hermes/skills/software-development/

After install, restart or refresh the agent if the skill list is cached.

From a Hermes skill copy, pass the repo root when the current directory is not the public skills repo:

python3 ~/.hermes/skills/software-development/oracle-ai-research-e2e/scripts/oracle_ai_research_e2e_check.py \
  --repo-root /path/to/martins-awesome-skills \
  --quick \
  --json

Common Mistakes

• Do not treat oracle-plan alone as a live E2E result. It proves command integration only.
• Do not run real E2E without AI_RESEARCH_BROWSER_E2E=1.
• Do not use Oracle runner when local guards are blocked; the CLI should mark blocked-by-local-guards.
• Do not use ChatGPT Pro/Extended Pro/GPT-5.5 Pro for routine tests.
• Do not commit real provider screenshots or full prompts unless they are explicitly sanitized.