Summary

toNormalizedUsage() in src/agents/pi-embedded-runner/run.ts (line ~192) intentionally returns only the last API call's input, cacheRead, and cacheWrite tokens instead of the accumulated values from the UsageAccumulator. While output tokens are correctly accumulated, input/cache tokens are taken from lastInput/lastCacheRead/lastCacheWrite. This causes session transcripts and cost tracking to miss ~80% of actual billed tokens in tool-heavy workflows.

This was introduced to fix the context-size display issue (#13698), but it inadvertently broke cost/accounting accuracy.

Impact

On a real deployment (572 sessions, ~7,900 assistant messages, ~7,300 tool calls over March 2026):

Verified independently by Claude Code and OpenAI Codex analyzing the source code — both confirmed no alternative persistence path exists for the full accumulated usage.

Root Cause

In run.ts:

const toNormalizedUsage = (usage: UsageAccumulator) => {
  const lastPromptTokens = usage.lastInput + usage.lastCacheRead + usage.lastCacheWrite;
  return {
    input:      usage.lastInput || undefined,       // ❌ last call only
    output:     usage.output || undefined,           // ✅ accumulated
    cacheRead:  usage.lastCacheRead || undefined,    // ❌ last call only
    cacheWrite: usage.lastCacheWrite || undefined,   // ❌ last call only
    total: lastPromptTokens + usage.output || undefined,
  };
};

The UsageAccumulator correctly accumulates all values via mergeUsageIntoAccumulator() (including input, cacheRead, cacheWrite), but toNormalizedUsage() discards the accumulated prompt-side fields. For a turn with N tool-call round-trips, this loses (N-1)/N of input and cache tokens.

Downstream consumers all read this lossy object:

• Session store (session-store.ts:47) → inputTokens, cacheRead, cacheWrite, estimatedCostUsd
• Transcript JSONL entries → usage block per assistant message
• Gateway transcript fallback (session-utils.fs.ts)
• Diagnostics emitter (agent-runner.ts:585-619)
• Cost calculation (session-cost-usage.ts)

No alternative persistence path captures the full accumulated totals.

Steps to Reproduce

1. Configure OpenClaw with Anthropic provider (prompt caching enabled by default)
2. Start a session with tool-using agent (e.g., exec tool)
3. Send a message that triggers multiple tool calls in one turn:

   Run these 3 commands: echo "hello", echo "world", echo "done"
   

4. Check the transcript JSONL for the assistant message's usage.cacheRead
5. Compare with Anthropic Console usage for the same time window

Expected: usage.cacheRead reflects sum of all API calls in the turn (≈ context_size × number_of_calls)
Actual: usage.cacheRead reflects only the last API call

Suggested Fix

The types already have the separation needed (EmbeddedPiAgentMeta.lastCallUsage exists for context-size display). The fix is:

1. Make toNormalizedUsage() return accumulated values for all fields (restore usage.input, usage.cacheRead, usage.cacheWrite from the accumulator)
2. Continue using lastCallUsage / promptTokens for context-size display (already wired correctly via deriveSessionTotalTokens)
3. Cost/accounting consumers already use agentMeta.usage — they will automatically get correct values

This separates the two conflated concepts:

• Context-size display: needs last call's prompt tokens → lastCallUsage / promptTokens
• Cost tracking: needs accumulated billed tokens → agentMeta.usage

Related Issues

• Bug: sessions.recent returns totalTokens maxed at contextTokens (200000) instead of actual usage #13698 — Original context-size display bug that introduced lastCacheRead (the fix that caused this regression)
• [Bug]: session_status totalTokens always shows contextTokens (100%) after multi-turn tool-use runs #17016 — Same context-size issue, closed as resolved by Bug: sessions.recent returns totalTokens maxed at contextTokens (200000) instead of actual usage #13698 fix
• Usage dashboard undercounts tokens/cost — reset archive transcripts are ignored #40870 — Complementary undercounting issue (reset archive transcripts ignored)

Environment

• OpenClaw 2026.3.23-2
• Anthropic provider with prompt caching
• Heavy tool-use workflows (avg ~1 tool call per assistant message)