[shensmart]

Why

cacheRatio (cache reads over the whole prompt) has been in shared/pricing.ts since v0.6.0, but it only reached the UI inside the CACHE cell's bottom line when the turn also had a known price. For unpriced models — including anything not in the built-in table and without an override — that line fell back to –, so the ratio disappeared even though the token counts were right there. The session and overview summaries never showed it at all, which is where a cache-hit number is most useful (it is a session-level property, not a per-turn one).

What

Render the existing ratio in three places, without touching the data layer:

• Session and overview summary rows — a new cache hit stat next to the raw cache token count, aggregated over turns.
• Turn history CACHE cell — shown independent of pricing. When a price is known the cell keeps its current shape ($0.01 · 69%); when it is not, it now shows just 69% instead of –.
• In-progress turn line — appended to the existing token line: in 125.9k · cache 2.58M (95%) · out 20.8k · 33 calls. This replaces a near-duplicate (95% cached) suffix on the same line.

Because input is already provider-normalized to the fresh-token form by semantics.ts, the ratio is comparable between exclusive-input (Claude) and inclusive-input (Codex) adapters without further changes. Derived at render time, so existing ledger records are reinterpreted correctly — no migration, no new persisted field.

Verification

• tsc --noEmit clean; node --test shared/*.test.ts server/*.test.ts → 65/65 pass (the existing cacheRatio tests already cover the arithmetic, including the cache-write denominator).
• Rendered against real data through the 0.11.2 desktop web client: overview shows cache hit 74%, the panel's per-turn table shows 69% / 83%, and the in-progress line shows (95%). Console clean.
• paseo plugin ls → running, no error.

Notes

• This is UI-only; pricing, cost residual handling, and the ledger file format are untouched.
• Happy to split the summary stat out if you would rather keep the turn table and summary changes separate.

---

Update: average output throughput, and a pill that fits

Two follow-ups on top of the cache-hit work above.

Can a plugin see token speed? Not the model's rate. Paseo's usage API carries only token counts, cost and context-window fill — no per-token or first-token timing — so the only signal is the span between a turn's start and end. Summing that per summary gives an average output throughput (output tokens ÷ wall clock). Wall clock covers thinking, tool calls, and waiting on the provider, so it is a floor on generation speed, not the token rate; the UI labels it avg out and the README says so explicitly rather than implying a measurement the protocol cannot make.

Summary.durationMs is added as a non-negative field defaulting to 0, so a client reading summaries from a server that predates it still parses. It is summed next to the token counts in both summaryFromRows and the overview accumulator, and rows without a timing contribute nothing rather than dropping the turn.

The pill now leads with tokens, hit rate, and throughput. Context-window fill is dropped from the pill: the chat surface already shows it next to the composer, so it was duplicating the host at the expense of a number that cannot be seen there.

	before	after
unpriced model	6.50M tok · ctx 12%	26.77M·95%·74 t/s
priced model	≈$0.43 · ctx 47%	≈$0.43·90%·74 t/s

The host caps the pill's button at 160px, and the icon, gaps and padding take 50px of it, so the label has about 110px before the host ellipsizes. Measured advance widths per glyph (12px system-ui, tabular numerals) put % at 11.1px against 1 at 5.6px, so the label is assembled against a measured width table instead of a character count, and drops the throughput (then the hit rate) rather than letting the host clip it. The count itself is never dropped. pillLabel is a pure function with tests covering the ordering and the budget.

Formatting helpers move to shared/format.ts so the pill, panel, and overview render from one implementation.

Verification

• tsc --noEmit clean; node --test shared/*.test.ts server/*.test.ts → 75/75 pass.
• Rendered against live data through the 0.11.2 desktop web client at 1440px and 980px: pill reads 26.77M·95%·74 t/s at 157px, unclipped; panel summary shows cache hit 95% and avg out 74 tok/s; per-turn CACHE column shows 69% / 83% / 96% / 97%. Console clean.
• Noted for review: the throughput figure is a wall-clock average, so it will read low on turns dominated by tool calls or long thinking. If you would rather it not ship at all than ship under that caveat, say so and I will drop that part and keep the pill on tokens + hit rate.

[shensmart]

Superseded by a fresh PR: the branch was rebuilt so the pill design (compact 33.99M(96%) 70t/s label, context fill removed) and the per-turn throughput landed as one reviewed change instead of follow-up commits. No upstream activity on this one. Closing to keep the queue clean.