feat(agent): cumulative prompt/completion token-split accessors#1
Open
indykish wants to merge 1 commit into
Open
feat(agent): cumulative prompt/completion token-split accessors#1indykish wants to merge 1 commit into
indykish wants to merge 1 commit into
Conversation
indykish
added a commit
to agentsfleet/agentsfleet
that referenced
this pull request
Jun 12, 2026
…h commit The fork patch is now reviewed (agentsfleet/nullclaw#1) and released as tag v2026.5.29-zmb.1 at the same commit (127b5ac4). Swapped the pin ref from the branch name to the release tag — same commit, so the content hash is byte-identical (verified via zig fetch), a more semantic + immutable reference. Spec §1 + Depends-on updated to cite the release. Both build graphs compile against the tag pin. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
…e accessors The agent already normalizes prompt_tokens/completion_tokens per response but folds only total_tokens. Accumulate both splits at the turn and summary-compaction sites and expose promptTokensUsed()/completionTokensUsed() beside tokensUsed(), so embedders can meter prompt-side and completion-side spend separately without re-deriving usage from history. clearSessionState() zeroes both new counters alongside total_tokens (@hasField-guarded, matching the existing pattern), so /new and /reset don't leave by-side metering carrying stale per-session figures while tokensUsed() reads 0. The "/new clears history" test asserts it. Carried on this fork atop v2026.5.29 until upstream exposes split accessors. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
indykish
force-pushed
the
patch/split-token-accessors-v2026.5.29
branch
from
127b5ac to
dd85a21
Compare
Comment on lines
2443
to
+2445
| self.total_tokens += normalized_usage.total_tokens; | ||
| self.prompt_tokens_total += normalized_usage.prompt_tokens; | ||
| self.completion_tokens_total += normalized_usage.completion_tokens; |
There was a problem hiding this comment.
When a provider sets total_tokens > 0 but leaves prompt_tokens = 0 and completion_tokens = 0 (neither normalization block fires because total_tokens is already non-zero), the accumulators for prompt_tokens_total and completion_tokens_total receive 0 for that turn while total_tokens advances correctly. Over multiple such turns, promptTokensUsed() + completionTokensUsed() falls short of tokensUsed() with no signal to the caller. Since the stated purpose of these accessors is by-side billing, an embedder that relies on them for cost metering will silently under-charge prompt-side and completion-side spend for any provider that only surfaces a total — a hard-to-detect billing gap.
Prompt To Fix With AI
This is a comment left during a code review.
Path: src/agent/root.zig
Line: 2443-2445
Comment:
**Split counters silently zero-out when provider omits per-side data**
When a provider sets `total_tokens > 0` but leaves `prompt_tokens = 0` and `completion_tokens = 0` (neither normalization block fires because `total_tokens` is already non-zero), the accumulators for `prompt_tokens_total` and `completion_tokens_total` receive 0 for that turn while `total_tokens` advances correctly. Over multiple such turns, `promptTokensUsed() + completionTokensUsed()` falls short of `tokensUsed()` with no signal to the caller. Since the stated purpose of these accessors is by-side billing, an embedder that relies on them for cost metering will silently under-charge prompt-side and completion-side spend for any provider that only surfaces a total — a hard-to-detect billing gap.
How can I resolve this? If you propose a fix, please make it concise.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Carries one commit atop upstream tag v2026.5.29: the
Agentalready normalizesprompt_tokens/completion_tokensper response but folds onlytotal_tokens. This accumulates both splits at the turn and summary-compaction sites and exposespromptTokensUsed()/completionTokensUsed()beside the existingtokensUsed(), so an embedder can meter prompt-side and completion-side spend separately without re-deriving usage from history.Why a fork
usezombiebills platform-model token spend by side and needs these accessors. We hold no push/release rights on upstreamnullclaw/nullclaw, so the patch rides this fork (usezombie/nullclaw) until upstream exposes equivalent accessors — at which point the consumer pin returns to the upstream tag and this fork is dropped.Changes
prompt_tokens_total/completion_tokens_totalcumulative fields onAgent(default 0).promptTokensUsed()/completionTokensUsed()accessors.Agent tokens trackingtest extended to cover the splits.Verification
zig build test --summary all— the pristine tag and this patched branch fail only the same 5 pre-existing redis-environment integration tests (those diallocalhost:6379, which on a dev box is an unrelated container; clean in upstream CI). No new failures. Released asv2026.5.29-zmb.1.Greptile Summary
Adds cumulative
prompt_tokens_total/completion_tokens_totalfields toAgent, folds them at both the per-turn response and summary-compaction sites, and exposespromptTokensUsed()/completionTokensUsed()alongside the existingtokensUsed()so embedders can meter by token side. TheclearSessionStatereset and/new-command test are also updated.root.zig): twou64fields default to 0, accumulated at the same two fold-sites astotal_tokens, exposed via typed accessors.commands.zig):clearSessionStategains@hasField-guarded zeroing of both new fields, mirroring the existingtotal_tokenspattern; tested by an extended/newassertion inroot.zig.Confidence Score: 4/5
The change is safe to merge for non-billing consumers; the billing-accuracy gap for providers that report total but not per-side tokens should be addressed before production metering is switched on.
The accumulation logic is correct for all providers that supply split data, and the session-reset path is handled cleanly. The one real concern is that neither normalization block back-fills split tokens when a provider gives total_tokens without per-side data — those turns leave prompt_tokens_total and completion_tokens_total at zero while total_tokens advances, so promptTokensUsed() + completionTokensUsed() can silently diverge from tokensUsed(). For the stated purpose of by-side billing metering, that gap is a current defect on the hot path.
src/agent/root.zig — the accumulation sites at lines 2443–2445 and 2864–2865 need a guard or documented caveat for the case where a provider omits per-side token counts.
Important Files Changed
Flowchart
%%{init: {'theme': 'neutral'}}%% flowchart TD A[LLM Response received] --> B[Normalize TokenUsage] B --> C{total==0 and splits present?} C -- Yes --> D[total = prompt + completion] C -- No --> E{total==0 and all zero and text present?} D --> F[Fold into Agent counters] E -- Yes --> G[completion = estimated tokens\ntotal = estimated tokens] E -- No --> F G --> F F --> H[total_tokens accumulates] F --> I[prompt_tokens_total accumulates] F --> J[completion_tokens_total accumulates] H --> K[tokensUsed] I --> L[promptTokensUsed] J --> M[completionTokensUsed] C -- No --> W{total greater than 0 but splits zero?} W -- Yes --> X[splits stay 0 - silent gap] X --> FPrompt To Fix All With AI
Reviews (2): Last reviewed commit: "feat(agent): accumulate prompt/completio..." | Re-trigger Greptile