Multi-step tool calls overwrite token counts instead of accumulating (output, reasoning, cache.write lost) #21913

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 token counts survive.

Root cause: processor.ts line ~362:

ctx.assistantMessage.tokens = usage.tokens  // overwrite!

While ctx.assistantMessage.cost += usage.cost correctly accumulates, tokens is replaced wholesale.

Impact

For an assistant message with N tool-call steps:

Also fixes:

• Context % display and compaction used total (which double-counted cached tokens) instead of deriving from components
• Custom provider models without limit.context defaulted to 0, breaking context % display and disabling auto-compaction entirely

Changes

1. Token accumulation in processor.ts

Replace the overwrite with field-wise accumulation:

const prev = ctx.assistantMessage.tokens
ctx.assistantMessage.tokens = {
  total: usage.total,
  input: usage.tokens.input,
  output: (prev?.output ?? 0) + usage.tokens.output,
  reasoning: (prev?.reasoning ?? 0) + usage.tokens.reasoning,
  cache: {
    read: usage.tokens.cache.read,
    write: (prev?.cache?.write ?? 0) + usage.tokens.cache.write,
  },
}

2. Derive total from components in getUsage()

total is now computed as input + output + reasoning + cache.read + cache.write instead of using the API's totalTokens (which double-counts cached tokens since AI SDK v6 includes them in inputTokens).

3. Add MessageV2.promptSize() and MessageV2.totalSize() helpers

• totalSize = input + output + reasoning + cache.read + cache.write — full conversation size after this turn (used for context % display and compaction threshold)
• promptSize = input + cache.read + cache.write — current prompt footprint (input tokens sent to the LLM)

overflow.ts uses totalSize for compaction because output/reasoning tokens become part of the context on the next turn.

4. Fix limit.context default from 0 to 128,000

Custom provider models not in models.dev no longer get context: 0 which broke context % display and disabled auto-compaction. Also limit.output defaults to 4,096 instead of 0.

5. ACP usage fix

acp/agent.ts now includes cache.write in the used token count for usage reporting (input + cache.read + cache.write).

Related issues

• github copilot: 2x token (counts total=total+cached) compacts at 64k, instead of 128k #3314 — 2x token display (total included cached, then cached was added again → now fixed by deriving total from components)
• OpenAI reasoning tokens double-counted in cost calculation #17566 / PR fix(opencode): don't double-count OpenAI reasoning tokens in cost calculation #17567 — OpenAI reasoning tokens double-counted in cost (reasoning cost fix, separate from this)
• Cache write tokens not accounted for when using OpenRouter, causing cost underestimation #18440 — Cache write tokens not read from OpenRouter
• Cost tracking ($ Spent) does not work for custom provider models #17223 — Cost tracking $0 for custom providers
• Token usage always shows 0 when using @ai-sdk/openai with custom baseURL (Responses API) #20885 — Token usage 0 with @ai-sdk/openai Responses API