Settings: how the agent thinks, written down — Thinking Patterns + Planner + Teams (v0.21.268)

[QR-Madness]

Wave 3.4. 26 keys across three screens. 102 of 208 settings now documented — just under half.

Written around what each pattern costs

That is what the choice actually turns on, and nothing said it before:

• Chain of thought rides in the same call — the only price is the tokens the reasoning occupies. Cheapest pattern here.
• Step-back adds a hidden pre-call before anything streams, so its cost is felt as unresponsiveness rather than slowness.
• Reflection spends output tokens on a draft you never read, and also gates the multi-pass variant.
• Consensus multiplies by k, and the samples are the cost — so pointing them at a cheaper model usually beats lowering k, since the value comes from attempts being independent rather than each being excellent.

Auto's classifier gets the point that matters: keyword heuristics run first and are free, so the common path costs nothing extra. The LLM tiebreak fires only when they're unconfident, only on a message past the length floor, bounded at ~150 output tokens and five seconds.

Delegation help states the three gates in order — the global switch, the per-profile opt-in (which is why nothing delegates on a fresh install), and the per-conversation override — and that depth × parallelism is roughly a turn's worst-case fan-out.

Two more mis-sections

planner.prompt_override and prompt_enhancement.system_prompt are both edited on Feature Prompts. The Task Planner screen says so in prose and doesn't render the control; both inherited a different screen from their config root. Now declared onto the screen that edits them, which moved Feature Prompts out of the "owns no settings" list.

That's five of these across four slices. Checking where a key actually renders is now the first step of a refit rather than a spot-check — noted in the campaign plan.

A wrong empty_means

step_back_model and sc_model were declared follow_role. The code and the UI hint both say empty means the conversation's own model — only the classifier follows a role. Now active_turn_model, so the manifest stops contradicting the screen it describes.

Three more control-less keys surfaced

step_back_timeout_seconds (its model was already shown), planner.max_subtasks (beside the threshold it pairs with), and alloy.non_blocking_delegations.

Verification

• Backend suite 1243 OK; task test:sterile 1243 OK
• Client 499 tests (6 new, covering the three newly-surfaced controls, picker bindings, and that Planner does not render the prompt override); 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: Thinking 13 anchors, Planner 6, Teams 6 — each with a help popover, no nested rows, no horizontal overflow. Manifest confirms the section moves, the corrected empty_means, and 13/13 · 6/6 · 6/6 help coverage.

One process note for the record: the first walk showed zero help popovers, which looked like a real defect. It wasn't — the API caches settings_help.yaml at first use, so it needs a restart after editing that file or the walk lies to you.

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 8:46pm