Settings: what the web costs, written down — Web Search + Research Mode (v0.21.267)

[QR-Madness]

Wave 3.3 of the settings campaign. 31 keys across two screens, written to the golden standard. 76 of 208 settings now documented, up from 45.

Both screens are about the same thing — how much of the web an agent may read and what that is allowed to cost — so they were written together and cross-reference each other.

The cost story, stated rather than gestured at

• A budget window opens with your turn and closes with it. Call counts and dollar ceilings bind together: whichever runs out first ends that turn's spending.
• Only interactive turns are metered. Work running under delegation or the planner is not inside a window at all. That isn't obvious from anywhere else and it matters if you are reasoning about spend.
• Tavily credit arithmetic appears where it decides something: an advanced search bills 2 credits, so setting it as the default halves how far a call budget stretches in money while the count looks unchanged.
• source_policy gets the asymmetry it turns on. Blocked is a hard floor the model cannot widen past. Preferred is a soft seed that a model's own explicit scope overrides — because silently over-narrowing is how a research turn comes back empty, and the agent reads that as "no such information exists".

A mis-section the refit caught

search.research_per_turn_limit claimed the Web Search screen — inherited from its config root — while it has always rendered on Research Mode, with the rest of that budget. Now declared onto the screen it actually appears on.

Same class of lie memory.project_channels told in Wave 3.1. Second one this cadence has caught, which is roughly the point of writing the help against the code.

Six keys had no control anywhere

Writable over the API, rendered nowhere. Rather than documenting them as unreachable:

• brave_context_max_tokens_per_url — its sibling total budget was already on screen, and a total with no per-page share is half a control.
• The deep-research tuning group — output floor, cache, poll bounds, and the budget weight that the search-budget help explicitly points the reader at. Writing "see the deep-research weight below" and then not showing it would have been worse than either option.

The Research Mode cost projection now reads the live budget weight instead of assuming 3.

Two fixes on the way

The generic round-trip test was shrinking. It skipped keys that declare constraints, so its coverage fell every time a section was refit — down to 43, tripping its own floor of 50. That is backwards: the test exists to catch a declared key that silently fails to persist, and a constrained key is no less prone to that. Constrained keys are now probed within their declared range; coverage is up and the floor rises with it.

Toggle labels wrapped to three lines. A toggle row lays out as [control + label + chrome] [hint] and the hint was winning the width fight — which got worse once labels gained a help affordance beside them. Secondary text wraps now. Verified live across Web Search, Recall and Conversation Context.

Also: advanced in the generated reference no longer claims to be "behind the disclosure in Settings". That was true of the one section this started with and is not a promise every screen keeps.

Verification

• Backend suite green; task test:sterile 1243 OK
• Client 493 tests (6 new, covering the source-policy single binding, the per-page budget, and API-key help); tsc clean
• ruff check api/ clean; pyright baseline 0
• task docs:check green, zero warnings; production build OK
• Live walk, fresh tab, zero console errors: Web Search shows 19 anchors and 21 help popovers (the two API keys carry help despite having no other chrome — secrets have no default to compare or reset to); Research Mode shows 10 anchors including config:search.research_per_turn_limit on the screen it belongs to. Manifest confirms the section move, the constraints and the tiers. 375px pass, no horizontal overflow.

Assisted-by: Opus 5

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agentx-docs-site	Ready	Preview	Aug 7, 2026 4:04pm