[Feature]: Add compact per-turn session stats footer

[coygeek]

Before submitting

• I searched existing issues and did not find a duplicate.
• I am describing a concrete problem or use case, not just a vague idea.

Area

apps/web

Problem or use case

T3 Code already exposes some useful run state, such as the current model selection in the composer, elapsed message metadata, and a context-window meter. What is still hard to see at a glance is the compact per-turn/session statistics that make a coding run easy to audit after the fact.

A useful reference is the t3.chat footer style, which shows a small row like:

GPT-5.4 (High) · 59.17 tok/sec · 3276 tokens · Time-to-First: 4.3 sec · 1 tool call

For long coding sessions, this information helps answer questions like:

• Which model/effort produced this response?
• Was the run slow because first-token latency was high, because generation was slow, or because tool work dominated?
• How many tokens did this turn consume?
• How tool-heavy was this turn?
• Is one provider/model configuration clearly performing better than another?

Today, users have to infer this from scattered UI state, context-window popovers, message timing, or logs. That makes performance/provenance comparison much harder than it needs to be.

Proposed solution

Add a compact per-turn stats footer under each assistant response, or in the existing completion-divider/message metadata area, with graceful hiding for unavailable fields.

Suggested fields, when available:

• provider/model/effort snapshot for the turn, e.g. Codex · GPT-5.5 · High
• elapsed duration, e.g. 2m 14s
• output/turn token count and/or total processed tokens
• generation throughput, e.g. 59 tok/sec, computed only when the numerator/denominator are reliable
• time to first assistant output, e.g. TTFT 4.3s
• tool call count, e.g. 7 tool calls

The footer can stay compact by default and put secondary detail in a tooltip/popover. Missing provider fields should simply be omitted, not shown as zero.

Implementation notes from a Repomix pass over main:

• packages/contracts/src/providerRuntime.ts already defines ThreadTokenUsageSnapshot with usedTokens, totalProcessedTokens, input/output/reasoning token fields, toolUses, durationMs, and compaction-related metadata.
• apps/server/src/provider/Layers/CodexAdapter.ts already normalizes Codex thread/tokenUsage/updated notifications into ThreadTokenUsageSnapshot.
• apps/server/src/provider/Layers/ClaudeAdapter.ts already normalizes Claude usage and emits thread.token-usage.updated from result/task progress events.
• apps/server/src/orchestration/Layers/ProviderRuntimeIngestion.ts already projects thread.token-usage.updated into context-window.updated thread activities.
• apps/web/src/lib/contextWindow.ts and apps/web/src/components/chat/ContextWindowMeter.tsx already consume the latest context-window snapshot in the composer footer.
• apps/web/src/components/ChatView.tsx, apps/web/src/session-logic.ts, and apps/web/src/components/chat/MessagesTimeline.tsx already derive/display turn timing from latestTurn.startedAt, latestTurn.completedAt, and assistant message timestamps.

The feature therefore looks feasible as an incremental UI/projection addition rather than a full provider rewrite. The likely missing pieces are turn-local stats instead of only latest-thread context-window state, a reliable first assistant-output timestamp, and a per-turn provider/model/effort snapshot if #2481 does not cover that first.

Why this matters

Coding-agent sessions are often long, mixed-provider, and tool-heavy. A compact stats footer makes the transcript self-auditing: users can compare model behavior, understand slow turns, and reconstruct what happened without opening logs or guessing from the current composer state.

This is especially valuable when experimenting with provider/model choices. A user should be able to look back at a thread and quickly see that one turn was GPT-5.5 High, another was Claude Opus, one had a high time-to-first, and another spent most of its time in tools.

Smallest useful scope

A first pass could be narrow:

• show a compact footer for the latest completed turn / assistant message only
• reuse existing latestTurn.startedAt / completedAt for elapsed time
• count tool activities for the turn from existing tool.* activities
• display available token data from the latest matching context-window.updated activity for that turn
• add a tooltip explaining when a metric is approximate or unavailable

A stronger follow-up could persist/project an explicit per-turn stats object so older turns remain exact after reloads and after later context-window updates.

Alternatives considered

• Expand feat: add usage / quota visibility for Codex sessions and accounts #228: that issue is broader usage/quota visibility for Codex sessions and accounts. This request is narrower: transcript-visible per-turn run stats and performance/provenance metadata.
• Fold into [Feature]: Show the provider and model used for each assistant message #2481: that issue asks for provider/model provenance per assistant message. This request complements it by adding runtime stats such as tokens/sec, time-to-first, token count, and tool count.
• Keep everything in logs/context popovers: useful for debugging, but too hidden for routine transcript review.

Risks or tradeoffs

• Provider parity: not every provider exposes identical token or timing fields. The UI should omit unavailable values instead of fabricating them.
• Metric semantics: tokens/sec can be misleading if it includes tool time, waiting time, or first-token latency. Label the denominator clearly, or only compute it from a well-defined assistant-output window.
• Historical correctness: if the UI reads only the latest context-window snapshot, old turns may show stale/newer usage. A durable per-turn projection is cleaner for full accuracy.
• Visual clutter: the row should be compact and low-contrast, with richer detail behind hover/click.

Examples or references

• t3.chat-style compact footer: Model (effort) · tok/sec · tokens · Time-to-First · tool calls
• Related but broader: feat: add usage / quota visibility for Codex sessions and accounts #228 (usage / quota visibility for Codex sessions and accounts)
• Related provenance work: [Feature]: Show the provider and model used for each assistant message #2481 (Show the provider and model used for each assistant message)

Contribution

• I would be open to helping implement this.

[coygeek]

Sneak Peak:

[nbros]

Why was this abandoned? This seems like a helpful feature (could be optional) to keep track of usage, especially with the current generation of models (Fable 5.1 and GPT 6 Astra) that can act like token furnaces! It would be useful to know which prompts use up a lot of usage.

[joshshiman]

This is such an important feature to me

[coygeek]

I re-opened. For anyone who wants to work on this and link the PR.

[liltaket]

Also want this