Settings: the last 32, the tier sweep, and the gate that keeps it (v0.21.271)

[QR-Madness]

Closes the settings campaign. All 210 writable settings are documented, and the standard is now enforced rather than remembered.

The tail — 32 keys

Model Providers, Model Limits, Model Roles, Prompt Enhancement, Feature Prompts, and the settings no screen renders.

Written to say what each key is for, not to restate its name: the LM Studio timeout is generous because local generation is slow, not because someone was careless; context_limits.models is the escape hatch for the single most damaging catalog error — a model whose real window is read as small, which surfaces as an agent that forgets within a few messages.

The gate is inverted — the point of the whole campaign

Coverage began as an allowlist of refit sections. That held what was written, but a new setting could ship undocumented forever without tripping anything.

SettingsHelpTest now asserts the inverse: every writable key carries the full five-field write-up. The only way past is UNDOCUMENTED_ALLOWLIST, which a separate test asserts is empty — so adding an exception is a visible, deliberate act with its own failing test rather than something that quietly makes the coverage test pass again.

I verified it bites by deleting an authored entry:

AssertionError: ['config:alloy.chain_of_command — no help at all'] != []

A gate that can't fail isn't a gate.

The tier sweep

Depth is the declared tier, rendered from the manifest. The three ad-hoc dialects are gone. The last one labelled "Experimental" over two switches that ship on and are declared essential — which is precisely what writing a judgement into markup costs: the label outlives the judgement.

Two more declarations corrected

• preferences.default_model is set on Model Roles, not nowhere — the eighth mis-section this cadence has found.
• The five llm_settings.* sampling defaults claimed the Model Limits screen while having no control anywhere. They're config-only now, and their help says so. A manifest pointing at a screen that doesn't show the setting is the exact lie this campaign existed to remove.

Binding the gap the walk exposed

Walking every screen showed three field-kit sections the write-up had documented but never wired: Prompt Enhancement, Feature Prompts (through PromptField's new binding), and the two LM Studio limits. Bound. 150 anchors across 13 screens.

Model Providers (12 keys) and Model Roles (4) stay docs-first — bespoke card and role surfaces rather than field-kit sections. Their help reaches the generated reference but not a popover beside the control. That's the one remaining gap between "documented" and "explains itself where you are", and it's in the backlog rather than glossed.

Verification

• Backend suite 1247 OK; task test:sterile 1247 OK
• Client 499 tests; tsc clean; ruff check api/ clean; pyright baseline 0
• task docs:check green, zero warnings; production build OK
• Live: manifest reports 210/210 fully documented, zero gaps. Walked all 21 screens — 150 anchors, no horizontal overflow, zero console errors.

ADR-17 gains three properties: the inverted gate, tier-as-declaration, and the undeclared-root report from v0.21.270.

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 8, 2026 3:02am