Plan mode upgrade: in-conversation plan tracker, one-click approval, and reasoning-effort tiers

[Luhaozhu]

Summary

This PR upgrades plan mode along three axes and fixes a reasoning-leak bug uncovered along the way.

In-conversation plan tracker (update_plan)

The main agent now carries a lightweight update_plan tool. On complex multi-step tasks it lays out a step checklist before starting, keeps executing in the same turn (no redirect, no approval round-trip), and updates step states as it works. The checklist renders as a slim plan bar pinned above the chat input — with a determinate progress ring, the current step, and an expandable step list — instead of cluttering the message flow. Simple Q&A never triggers it.

One-click approval for manual plan mode

The plan preview card now ships Confirm / Discard buttons (typed confirmation still works as a fallback). The decision is reflected on the card, and execution progress is mirrored into the same plan bar.

Plan-Execute-Verify

Plan prompts now instruct the model to append a final verify-and-fix step whenever the task has checkable deliverables (generated files, code, data processing): open the artifacts, check them against the acceptance criteria, and fix problems in place rather than merely reporting them. Whether and how to verify is the model's call — nothing is hard-coded.

Fix: reasoning chain leaking into the answer

Hybrid reasoning models often omit the opening <think> tag after tool calls. The frontend stripper re-arms on tool results, but update_plan intentionally emits no tool events — so the next iteration's bare reasoning leaked into the visible answer. The stripper now also re-arms on plan_update events, and plan-step outputs strip orphan closing tags as a belt-and-braces measure.

Misc

• Model providers gain a supports_reasoning_effort switch that unlocks High / Max thinking tiers per chat provider (delivered via chat_template_kwargs).
• Community edition hides the admin tag on globally installed plugins (CE has no admin/user hierarchy).

Testing

• End-to-end browser runs: manual plan mode (generate → button approval → execution → plan bar → completion), agent-initiated update_plan checklists across research/spreadsheet/slide tasks, and a negative control confirming simple questions don't trigger planning.
• Reasoning-leak regression verified in high-reasoning mode on both the live stream and the reload/history path.
• npm run build and backend import checks pass.