main
server source: e5909e53c32a9d5b51caf6352d1f17fd73e86db7
client: real web app, Playwright, /usr/bin/google-chrome
provider: @agentclientprotocol/claude-agent-acp, Claude Haiku 4.5 through local proxy
real five-tool turn, no mocks or injected provider frames
last usage_update: {"used":38593,"size":200000,"cost":{"amount":0.08375305,"currency":"USD"}}
idle turn spending: {"inputTokens":50,"outputTokens":670,"cachedReadTokens":227113,"cachedWriteTokens":1373,"totalTokens":229206}
persisted provider thread contextUsage: {"usedTokens":229206}
pr
server source: 698e2e29976948b8833b4ba156c7f1b0ef5ee691
client: real web app, Playwright, /usr/bin/google-chrome
provider: @agentclientprotocol/claude-agent-acp, Claude Haiku 4.5 through local proxy
real five-tool turn, no mocks or injected provider frames
last usage_update: {"used":39655,"size":200000,"cost":{"amount":0.08376335000000001,"currency":"USD"}}
idle turn spending: {"inputTokens":50,"outputTokens":762,"cachedReadTokens":195316,"cachedWriteTokens":39601,"totalTokens":235729}
persisted provider thread contextUsage: {"usedTokens":39655,"maxTokens":200000,"cost":{"amount":0.08376335000000001,"currency":"USD"}}
Nearby behavior through the web composer on the PR:
- Claude ACP -> native Codex gpt-6.1-sol: handoff ready/injected; response "return handoff ready."
- Native Codex follow-up in the same thread: response "follow-up verified."
- Fresh ACP-only thread: two real shell calls completed; idle spending 65888; context 33039/200000 retained with cost; UI tooltip 33k of 200k tokens used.
- Both changed existing test files: 68 tests passed on Blacksmith.
- Existing head CI: 31 successful, 0 failing, 0 pending.
Limits: did not reproduce the exact insufficient-handoff-allowance error, exercise native desktop/mobile, or run a real ACP child session. Mixed-provider tooltip can show an older native Codex turn count; the ACP-only tooltip and persisted provider snapshot were checked separately.