lsm / neokai / 27848419266

82%

Build Type

push

github

Committed by web-flow

Commit Message

Fix task-agent thinking token counts (#2138)

* fix(daemon): attribute thinking tokens as per-block turn deltas

SDK thinking_tokens events carry a cumulative turn-level estimate, but
the daemon stamped that cumulative value onto every assistant thinking
block. In task-agent/GLM sessions this produced the same token count on
many split thinking blocks.

Now SDKMessageHandler tracks the cumulative estimate and the cumulative
amount already stamped. Each assistant thinking block gets the delta
since the last stamped block; non-positive deltas are omitted so the UI
falls back to character counts. Tracking resets at result messages and
session_state_changed idle to prevent stale cross-turn attribution.

Add unit tests for delta attribution, repeated cumulative estimates,
turn-end reset, and stale/zero/negative delta omission.

* refactor(daemon): reset thinking token tracker when cumulative estimate drops

If the SDK/provider resets the running total for a new thinking block,
the cumulative estimate can drop below the last stamped value. Treat that
as a new block boundary instead of a negative delta, so later thinking
blocks in the same turn still get a token count.

Update tests to cover zero/repeated deltas separately from new-block reset.

* fix(daemon): reset thinking counters on new SDK query init

A turn interrupted or stopped without a result/idle message could leave
lastStampedThinkingTokensEstimate from the previous turn. The next query
would then subtract that stale baseline and undercount or drop its first
thinking block.

Reset the tracker on every SDK system/init message, which fires at the
start of a new query/session and is early enough to clear stale state
before any thinking_tokens events arrive.

Add a test verifying init clears a stale counter from an interrupted turn.

* docs(daemon): clarify thinking token delta heuristic comment

* fix(daemon): reset thinking counters before queued turn replay

The turn-level thinking counter was res... (continued)

Coverage Stats

10810 of 15739 branches covered (68.68%)

Branch coverage included in aggregate %.

22 of 22 new or added lines in 1 file covered. (100.0%)

1 existing line in 1 file now uncovered.

91908 of 109848 relevant lines covered (83.67%)

298.08 hits per line