fix: accumulate tokens across multi-step tool calls instead of overwriting - #22032

Fixes #21913

The bug

In multi-step tool-call assistant messages, processor.ts overwrites assistantMessage.tokens on each finish-step event instead of accumulating additive fields. Only the last step's output, reasoning, and cache.write survive. cost was correctly accumulated with +=.

Separately, custom provider models not in models.dev got limit.context: 0, which broke context % display and disabled auto-compaction entirely.

Changes

Token accumulation (processor.ts) — Field-wise accumulation replaces wholesale overwrite:

• input, cache.read: replaced each step (snapshots of current state)
• output, reasoning, cache.write: accumulated with += across steps
• total: derived from components by getUsage() instead of using API's totalTokens (which double-counts cached tokens)

Derived total (index.ts) — getUsage() computes total = input + output + reasoning + cache.read + cache.write. Kept as a top-level field (not inside tokens) so existing consumers of msg.tokens.total continue to work.

Named helpers (message-v2.ts) — MessageV2.promptSize() (input + cache) and MessageV2.totalSize() (all fields) so consumers don't hand-compute the sum inconsistently.

Compaction (overflow.ts) — Uses MessageV2.totalSize() for overflow check. Output/reasoning tokens are included because they become part of the next turn's context.

Custom provider defaults (provider.ts) — limit.context defaults to 128_000 and limit.output to 4_096 via named constants instead of 0.

ACP usage (agent.ts) — Includes cache.write in the used token count.

Test coverage

• 3 modified existing tests: getUsage total assertions (now derived), provider limit defaults (128k/4096), overflow semantics (uses totalSize)
• 12 new tests: promptSize/totalSize helpers, isOverflow pure unit tests, getUsage total derivation, 3-step accumulation simulation mirroring processor.ts logic