[feat] Ambient context-usage meter sourced from the model catalog

[ashrafchowdury]

Stacked PR. Base branch is fe-feat/session-context-info (Arda's original
indicator)

Context

This extends Arda's context-usage indicator. Two problems with that first version:

1. The context-window denominator came from MODEL_CONTEXT_WINDOWS, a hardcoded ~25-entry
   map in the frontend, matched to the selected model by longest-substring guessing. Every
   new model needed a manual edit, and the substring match was fragile.
2. The indicator showed two raw token measures side by side with developer labels
   (Ctx 407.0k / 1.0M (39%) · Σ 407.0k (39%)). That side-by-side was an internal A/B to
   pick one measure, and it shipped to end users who do not know what Ctx, Σ, or a
   "token" is.

Changes

Context window now comes from the model catalog, not a frontend map.
The SDK already ships each model's context_window on the harness-capabilities document
the frontend fetches (from model_catalog.py), so the denominator was already in the
browser. Deleted the hardcoded map and its substring resolver. A new selector resolves the
window by exact model id.

Before (hand-edited per model, substring-matched):

const MODEL_CONTEXT_WINDOWS = {
"gpt-4o": 128_000,
"claude-3-5-sonnet": 200_000,
// ~25 more...
}

After (exact-id lookup into the catalog that is already on the capabilities doc):

contextWindowForModel(capabilities, harness, modelId)
// → capabilities[harness].model_catalog.find(e => e.id === modelId)?.context_window

No new endpoint. Every model in the picker is covered automatically, and the match is
exact instead of a guess.

The indicator is now an ambient meter, not a dual token readout.
Dropped the Σ running-sum measure from both the data layer and the UI. It double-counted
resent history and never dropped, so it was the misleading one. Kept occupancy: the latest
turn's total tokens, which is how full the window is right now.

Before:
Ctx 407.0k / 1.0M (39%) · Σ 407.0k (39%)

After: a slim fill bar plus one plain line.
▓▓▓▓░░░░░░ Context 39% used

The line escalates with fill: neutral gray with room, amber at 75%, a faded red with a
warning icon and "Context almost full · 95%" at 90%. Exact counts move to the tooltip in
compact form (407k / 1M tokens), with a one-line explanation of what the context window
is and that older messages get summarized as it fills.

Prop and type changes for reviewers:

• ContextBudgetIndicator takes maxTokens: number | null instead of model: string.
• ContextBudget dropped runningSumTokens, runningSumPct, and perTurnTokens.
• useAgentModelKeyStatus now also returns harness (read from agent.harness.kind).
• New public export contextWindowForModel from @agenta/entities/workflow.

Tests / notes

• @agenta/entities builds clean; OSS tsc --noEmit shows no new errors in the touched
  files; ESLint and prettier pass.
• The indicator was the only reader of the dropped runningSum/perTurn fields, so no
  other consumer changed.
• Follow-up, not in this PR: some catalog entries carry context_window: null (e.g. the
  Claude default alias). Those models show a raw token count with no bar or percent.
  Filling that in is a data change in model_catalog.py, not frontend.

What to QA

• Open an agent session and send a few turns. The composer footer shows a gray bar and
  "Context N% used", where N tracks the latest turn.
• Hover it. The tooltip shows compact tokens (407k / 1M tokens) and the plain explanation.
• Switch the agent between a large-window model (a Pi 1M model) and a 200k model. The
  percent and bar rescale to the new denominator, confirming the catalog lookup rather than
  a hardcoded number.
• Regression: pick a model the old map never listed. It still shows a percent (resolved
  from the catalog) instead of a bare token count.
• Tiers: to see amber and red without a huge session, temporarily force occupancy near 80%
  and 95% in ContextBudgetIndicator (thresholds are >=0.75 amber, >=0.9 red).

Previews

When context fills up

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Jul 21, 2026 11:18am

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: ee9a0d2b-9bce-4ed1-b215-27cc9951dfaf

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/extend-arda's-session-context-work

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}