status bar: show current context occupancy in addition to summed turn usage

[btipling]

Summary

Keep the context/usage slot's turn totals (they are useful).
Also show current context occupancy: how large the last (or active) model request was — the thing that can hit the window.

Today the slot paints AI SDK result.usage (phase 3 #539 / PR #547, product #327). On a tool loop that value is the sum of every step. A turn that prints 1.3M in · 13.3k out · 1.3M tok · 1.2M cached can be 15 × ~90k-input steps, all inside a 200k/1M window. Totals look like overflow. Occupancy would not.

Keep (do not remove)

Figure	Meaning
Turn in / out / tok / cached	Provider sum across steps — cost / “this turn was huge”
Honesty locks	source: 'provider' only, no client tokenizer, no fake % without a catalog max
Abort/cancel	keep last honest paint; completed turn with nothing usable hides
Host-local	not a cloud meta key

Add

Want	Not
Input tokens on the last step (or max step, if last is a tiny wrap-up call)	Replacing the sum
Optional cache-read on that same step	Using the summed cache as occupancy
Optional later: catalog window → honest % only with a real max	Guessed denominator

Suggested paint (host-folded string, Wasm still generic slot 3), e.g.

90k ctx · 1.3M in · 13.3k out

or a second slot if 96 bytes cannot hold both honestly (truncateStatusValue already ellipsizes).

Success: a 15-step turn on a 200k model still shows the 1.3M in total and ~80–180k ctx.

Constraints

• Still provider-sourced. Prefer result.steps[i].usage (or equivalent). If the SDK only exposes the aggregate, omit occupancy — do not invent a client token count and do not relabel the sum as ctx.
• Completed turn with totals but no per-step usage → keep painting totals; occupancy hidden.
• Cap: USAGE_SUMMARY_MAX_BYTES / STATUS_SLOT_MAX_BYTES (96). Extending the carrier is a NEW field + Caps table in the phase plan (do not silently jam two meters into a 96-byte JSON that then drops).
• Not agent-session (agent session architecture — model-context memory, compaction, and viewport #548). Status-bar / usage-meter honesty. Do not add agent-session or project 3.

Non-goals

• Compaction / changing the 400 / 3.5M fold (inference budget: replace 400-row / 3.5M-char fold with model-window token budget #551 / compaction: model-capability context compaction (not API payload slices) #552)
• Dropping the summed totals
• Client tokenizer as a stand-in

Related

• harness: show live context size from model usage (not client math) #327 (closed) — live context from model usage; shipped totals only
• plan: workspace status bar — phase 3 — context/usage slot (implements #327) #539 / PR feat(#539): context/usage status slot — bounded provider-usage summary (chat+agent, JSON + stream done) folded into the Wasm context slot #547 — carrier + fold
• inference budget: replace 400-row / 3.5M-char fold with model-window token budget #551 — model-window budget (what we send). Different; may consume occupancy later.