feat(browser): recommend conversation mode in dry-runs

[pdurlej]

Summary

• add an advisory browser conversation-mode recommendation for dry-runs: one-shot, multi-turn, deep-research, or project
• surface the recommendation in CLI dry-run output, browser preview JSON, and MCP consult dry-run structured content as optional recommendedConversationMode
• document that recommendations are guidance only and never switch projects, invent follow-ups, upload Project Sources, or change archive behavior

Safety / Scope

• no browser DOM, model picker, composer, or Project Sources implementation changes
• no dependency on feat(browser): manage ChatGPT Project Sources #173; Project Sources are mentioned only as persistent ChatGPT Project context in docs
• existing dry-run/preview fields keep their meaning; MCP field is optional

Tests

• pnpm install --frozen-lockfile
• pnpm vitest run tests/cli/conversationMode.test.ts tests/cli/dryRun.test.ts tests/cli/dryRun.coverage.test.ts tests/mcp/consult.test.ts -> 4 files passed, 24 tests passed
• pnpm run check
• pnpm run build
• pnpm test -> 116 files passed, 18 skipped; 796 tests passed, 41 skipped
• git diff --check

Notes

This is intended to follow the multi-turn guardrails work, but it remains independently reviewable. If #180 lands first, I can rebase this PR to pick up the wording and resolve any small docs/dry-run conflicts.

[pdurlej]

Rebased onto current main after #178, #179, and #180 landed. Conflict resolution preserved the new browser-control/multi-turn guardrail output and added the advisory conversation-mode recommendation after the control plan, without changing runtime behavior.

Local verification after rebase:

pnpm vitest run tests/cli/conversationMode.test.ts tests/cli/dryRun.test.ts tests/cli/dryRun.coverage.test.ts tests/mcp/consult.test.ts
# 4 passed, 27 tests passed

pnpm run check
pnpm run build
pnpm test
# 118 passed, 18 skipped; 823 tests passed, 41 skipped

git diff --check

GitHub CI is green on Ubuntu, macOS, and Windows.

[steipete]

Thanks for rebasing this and for keeping the behavior advisory-only. I took another pass now that #173 landed. I am going to close this one rather than continue the stack.

Reasoning:

• The useful guardrail from this stack already landed in docs(browser): add multi-turn consult guardrails #180: browser follow-ups are explicit only, and Oracle does not invent multi-turn prompts.
• feat(browser): manage ChatGPT Project Sources #173 now gives Project Sources a real command surface (project-sources list|add) instead of a dry-run recommendation.
• This PR adds a prompt-content heuristic (project for prompts that look like ongoing architecture/product streams). That is plausible, but it is also easy for a CLI to become noisy or over-directive when it guesses workflow mode from prose.
• For agent workflows, I would rather keep dry-runs factual: what Oracle will do, which browser/project URL it will use, which follow-ups/files/sources were explicitly provided, and what will or will not mutate state.

The design may still be useful later if we want an explicit oracle plan/advisor mode or if MCP clients ask for workflow recommendations separately from dry-run execution. I do not think it belongs in the core dry-run path right now.

Appreciate the careful scope: no browser DOM changes, optional MCP field, and clear docs. This is being closed as product-direction deferral, not because the implementation was careless.