feat(mcp): improve Claude Code ChatGPT browser consults

[pdurlej]

Summary

Adds a focused Claude Code / MCP browser-consult path without adding new MCP tools or changing unrelated API defaults.

• Adds MCP consult preset chatgpt-pro-heavy.
• Adds MCP dryRun:true so Claude/Codex can inspect the resolved Oracle run without opening Chrome or creating a session.
• Adds MCP browserModelStrategy passthrough for parity with the CLI.
• Adds oracle bridge claude-config --local-browser plus explicit --oracle-home-dir and --browser-profile-dir overrides for macOS/local Claude Code setups.
• Documents the "Let Them Fight" workflow: Claude Code asks Oracle, Oracle asks ChatGPT browser mode, and the answer returns through MCP.

The preset targets gpt-5.5-pro directly, matching the current upstream model support from #156.

Public API

• MCP consult:
  • preset?: "chatgpt-pro-heavy"
  • dryRun?: boolean
  • browserModelStrategy?: "select" | "current" | "ignore"
• Bridge CLI:
  • oracle bridge claude-config --local-browser
  • --oracle-home-dir <path>
  • --browser-profile-dir <path>

PR / Issue Map

• Builds on [codex] support GPT-5.5 Pro #156 for official GPT-5.5 / GPT-5.5 Pro model support.
• Complements feat(browser): surface ChatGPT thinking heartbeat #148 by giving Claude/Codex a better long-run browser consult path with heartbeat support when both land.
• Complements feat(browser): coordinate concurrent ChatGPT tabs #150 because multiple MCP/CLI callers may share one browser profile.
• Complements feat(browser): support ChatGPT Deep Research mode #151 by exposing Deep Research/research-mode fields through the same MCP consult surface when that PR lands.
• Does not implement ChatGPT Project Sources management from Feature request: manage ChatGPT Project Sources directly from CLI and MCP #131 / [codex] add ChatGPT project sources support #132.

Validation

Latest branch head: ab735615; no code changes were needed in the final May 4 PR-body cleanup.

• pnpm install --frozen-lockfile -> passed, including prepare/build
• pnpm vitest run tests/mcp/consult.test.ts tests/cli/bridgeClaudeConfig.test.ts -> 2 files passed, 8 tests passed
• pnpm run check -> passed
• pnpm run build -> passed
• pnpm test -> 107 files passed / 18 skipped, 656 tests passed / 41 skipped
• pnpm run test:mcp:unit -> 5 files passed, 7 tests passed
• pnpm run test:mcp:mcporter -> passed
• MCP dry-run smoke: preset:chatgpt-pro-heavy, dryRun:true returned browser mode with gpt-5.5-pro and did not create a session or touch Chrome.
• git diff --check -> passed
• GitHub: mergeable, GitGuardian green

[pdurlej]

Rebuilt this branch cleanly on current upstream/main and pushed as ab735615.

Rebase notes:

• Dropped the old stacked GPT-5.5 browser/model commits that are now covered by [codex] support GPT-5.5 Pro #156.
• Kept this PR focused on MCP consult ergonomics and Claude Code local-browser config.
• Updated preset:"chatgpt-pro-heavy" to target the current gpt-5.5-pro model directly instead of the older gpt-5-pro alias.

Local validation on macOS:

• pnpm install --frozen-lockfile -> passed, including prepare/build
• pnpm vitest run tests/mcp/consult.test.ts tests/cli/bridgeClaudeConfig.test.ts -> 2 files passed, 8 tests passed
• pnpm run check -> passed
• pnpm run build -> passed
• pnpm test -> 107 files passed / 18 skipped, 656 tests passed / 41 skipped
• pnpm run test:mcp:unit -> 5 files passed, 7 tests passed
• pnpm run test:mcp:mcporter -> passed
• MCP dry-run smoke: preset:chatgpt-pro-heavy, dryRun:true returned browser mode with gpt-5.5-pro and did not create a session or touch Chrome.
• git diff --check -> passed

No secrets or account details included.

[pdurlej]

Final refresh from my side:

• No code changes in this last cleanup pass; I refreshed the PR body with the latest scope/test notes.
• Current head: ab735615.
• This remains intentionally narrow: MCP/Claude consult ergonomics and local Claude Code browser config, without pulling in the tab scheduler or Deep Research work.
• Previously validated with focused MCP/bridge tests, pnpm run test:mcp:unit, pnpm run test:mcp:mcporter, pnpm run check, pnpm run build, pnpm test, and dry-run consult coverage.
• GitGuardian is green and GitHub currently reports the PR as mergeable.

[steipete]

Thanks @pdurlej. I shipped the cleaned-up version of this on main as dbef347.

What landed:

• consult preset chatgpt-pro-heavy, hard-coded to gpt-5.5-pro with heavy browser thinking by default.
• MCP dryRun: true support so callers can inspect the resolved Oracle run without opening Chrome or creating a session.
• MCP browserModelStrategy passthrough for parity with the CLI.
• oracle bridge claude-config --local-browser plus explicit local path overrides for same-Mac Claude Code + signed-in ChatGPT browser workflows.
• README, MCP docs, bridge docs, and changelog coverage.

I didn’t merge the PR branch directly because it had become stale against current main and included unrelated browser churn after recent merged work. I replayed the focused commits onto current main and added one fix while landing: bridge config generators now keep stdout machine-readable, so oracle bridge claude-config --local-browser > .mcp.json produces valid JSON instead of including the oracle banner/tip text. There is also a CLI regression test for that.

Validation before push:

• pnpm vitest run tests/mcp/consult.test.ts tests/cli/bridgeClaudeConfig.test.ts -> 9 passed
• pnpm run check -> passed
• pnpm run build -> passed
• pnpm test -> 109 files passed / 18 skipped, 703 tests passed / 42 skipped
• manual stdout parse for node dist/bin/oracle-cli.js bridge claude-config --local-browser -> valid JSON

Closing this PR because the cleaned-up implementation is now on main. Appreciate the focused MCP/Claude Code workflow work.