feat(settings): Conversation Context, written up

[QR-Madness]

First section of the Wave 3 cadence — and the one that started this campaign:

"if a user wants to customize something such as modifying the verbatim window budget, they should be advised on how it works, what it does, why you need it."

What changed

All 20 keys of Conversation Context now carry the full five-field treatment. The prose is grounded in what the code actually does rather than restating the field label:

• Verbatim window budget explains that it's measured against the model's real context window, that assembly still hard-caps at window − reserved output so it can never squeeze out the reply, and that a wrongly-declared window is the usual reason a conversation compacts constantly.
• Compaction pre-warm explains why it sits below the verbatim ceiling — that gap is what keeps summarisation out of the middle of your turn.
• Just-in-time backstop explains what it protects against: history dropping before the digest has caught up, i.e. the agent silently forgetting something you said.

The section spans five config roots (context, session.rolling_summary, trajectory_compression, compression, memory), so binding needed a local-name → manifest-key map. sectionBinder generalizes that for the sections still to come, so the next refit is mostly content.

All 17 controls now show their shipped default, mark what's been changed, offer to put it back, and carry help. Constraints and tiers move server-side — harvested from what the JSX already declared — so the UI and the generated docs enforce the same bounds from one source.

Concretely, this closes the Wave 2 gap. Searching "verbatim budget" landed at the top of the section before, because there was no anchor to land on. It now lands on the control itself.

A real defect the new test caught

memory.project_channels claimed ui_section: "context" — inherited from its config root — while no screen has ever rendered it. It's a Projects setting. KeySpec gains a ui_section override so a key can correct or disclaim its root's section, and the manifest stops advertising a home that doesn't exist.

The help-coverage test now holds a list of refit sections rather than naming Recall alone, so each section stays covered once written and the cadence can't quietly regress.

Verification

• 1240 backend tests · 474 client tests · tsc clean · ruff clean
• task docs:check green, zero warnings; generated reference now documents 45 settings (was 25)
• Live on a clean tab, zero console errors: 17 anchors bound, help renders, and "verbatim budget" flashes config:context.verbatim_budget_ratio itself

Cadence status

Refit: Recall (24 keys) · Conversation Context (20 keys).
Remaining: Web Search (~20) · Thinking Patterns (~13) · Consolidation (~45) · Research · Images & Audio · Ambassador · Agent Teams · Task Planner · Model Limits · Prompt Enhancement. Tracked in todo/backlog/genome-advisor.md.

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	Building	Preview	Aug 7, 2026 12:27am