Claude Code context usage is only emitted at final result, leaving the meter stale during tool loops

[dfaivre-pcs]

Summary

Claude Code context usage remains unavailable or stale throughout a multi-step agent turn, even while assistant responses and tool calls complete. bb already extracts per-request context counts from assistant messages, but only emits the context-window snapshot when it processes the final result. Emit updates during the turn when those counts arrive.

Versions and environment

• bb CLI 0.42.1; local npm-installed bb-app on Linux/WSL.
• Provider: claude-code; installed Claude CLI 2.1.263 (the CLI version does not establish the bundled SDK version).
• Source inspected at 06aeaa994942ae7527dc49d2268c1f801e8542a0.
• Observed in an active personal-workspace thread and an active managed-worktree thread. A local stats plugin reads bb's public thread-event API; the missing data is confirmed directly through that API.
• Existing live threads were inspected; a fresh controlled reproduction on this source commit has not been run.

Steps to reproduce

Suggested minimal reproduction (not separately executed):

1. Start a Claude Code thread with a task requiring several sequential model/tool cycles, for example: Inspect this repository, run its checks, investigate failures, and summarize what you find.

2. While assistant messages and tool calls are completing, query the events on the running bb server:

   curl -sS 'http://127.0.0.1:38886/api/v1/threads/<thread-id>/events?types=thread%2FcontextWindowUsage%2Fupdated,thread%2FtokenUsage%2Fupdated,turn%2Fcompleted&order=desc&limit=20'

3. Compare with turn/started and item/completed events. Check again after the provider's final result.

Expected vs actual

Observed on thr_j3r5kg9m79 while active:

• One turn/started at sequence 9.
• The newest 100 events included 18 item/completed, 17 item/started, and 8 item/agentMessage/delta events.
• The filtered context/token-usage/turn-completed query above returned exactly [].
• A second active Claude thread, thr_ra6tsbrfyy, likewise had no context or token-usage events during its first ongoing turn.

Expected: context-window updates during the turn when the SDK provides a request usage sample, without waiting for the entire agent run to finish. This need not be a token-by-token counter. Missing readings should remain unknown; cumulative tokens must not be substituted for context occupancy.

Evidence

The source suggests the reporting boundary is the cause:

• Assistant-message translation extracts and stores latestRequestContextTokens, but does not emit a context-window delta there.
• Final-result translation emits the contextWindow delta.
• Request context extraction uses input + cache-read + cache-creation tokens.

Suggested direction: emit a context-window delta when a valid top-level assistant request sample arrives, preserving model capacity, estimated status, and turn attribution. Do not count subagent context as the parent context. Add a regression test that observes context events before the final result.

Investigation thread (reporter's local bb instance): http://127.0.0.1:38886/threads/thr_68imcxrdav . All relevant observations are included above so access to that instance is unnecessary.

What was ruled out

• Not merely a rendering or polling problem: the usage events themselves are absent from the public event API.
• Completed assistant messages/tool calls are not being mistaken for a persisted turn/completed: the latter was explicitly queried and absent.
• Searched open and closed issues. Pi context-window usage stays stale during multi-tool turns #2023 describes the same reporting cadence problem for Pi and is closed; this report concerns Claude Code's separate adapter.
• No claim that the provider lacks intermediate usage: the adapter already extracts it from assistant messages.

Suggested priority and effort

Medium: context pressure is invisible during long Claude tool loops, especially a first run or one extended by steering. Waiting for a final result is a workaround; no data loss was established.

AGENT GENERATED

[bb-slop-cop]

🚨 SLOP COP 🚨 · new-issue-autopilot

• Classification: Bug · Priority Medium · Effort Low
• Labels: providers, provider-claude-code, confirmed-repro
• Report: https://get-bb.github.io/reports/issues/3220.html
• Verdict: REPRODUCED · root-cause confidence: high
• Root cause: assistant translation stored valid request-context samples, while only result translation emitted the context-window delta.
• Reproduction: confirmed-repro — the focused translator test failed identically in two clean trusted checkouts before the fix.
• Fix: Publish Claude request occupancy from assistant messages #3221