fix(agent-core-v2): count compaction tokens on the full-request basis - #2699
Conversation
- tokensBefore/tokensAfter now include the system prompt and non-deferred tool schemas, matching the measured-anchor basis the context gauge uses between exchanges - the post-compaction ledger rebase carries the same full-request size, so the reported context size no longer dips to a messages-only estimate and jumps back on the next exchange - the PreCompact hook tokenCount uses the same basis
🦋 Changeset detectedLatest commit: 919bd5d The changes in this PR will be included in the next version bump. This PR includes changesets to release 1 package
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
commit: |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 919bd5d1d1
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| compactedCount: originalHistory.length, | ||
| tokensBefore, | ||
| summaryOutputTokens: attempt.usage?.output, | ||
| requestOverheadTokens: this.requestTokens([]), |
There was a problem hiding this comment.
Because compactionRound() passes requestOverheadTokens here before returning to compactionWorker(), the overhead is captured before profile.refreshSystemPrompt() runs. In the supported path where project instructions/profile prompt change before a manual compaction, tokensAfter and the rebase/status event are computed from the old prompt, while the next request uses the refreshed prompt, so the divider/gauge can still jump after the next exchange. Compute the post-compaction size after the refresh or use the refreshed prompt for this overhead estimate.
Useful? React with 👍 / 👎.
| // The window must stay above the harness's fixed request overhead | ||
| // (system prompt + tools, ~14k): the post-compaction size is reported on | ||
| // the full-request basis, so a smaller window could never be satisfied. |
There was a problem hiding this comment.
The scoped agent-core-v2 guide allows comments only in the top-of-file block, but this adds narrative comments inside the test body. Please encode the rationale in the test name/fixture constants or move it to the file header so the test stays within the local convention.
AGENTS.md reference: packages/agent-core-v2/AGENTS.md:L36-L38
Useful? React with 👍 / 👎.
main's MoonshotAI#2699 counts compaction tokens on the full-request basis, so the tower tool schemas (default registry) and the /tower skill catalog entry (system prompt) shift the pinned numbers: +2789 with the default tool set, +173 with the explicit harness tool list. The 20k-window test keeps its shape with a 22k window so the post-compaction floor still fits.
1 participant
Related Issue
No linked issue — the problem is explained below.
Problem
After compaction, the reported token counts read far below the real context size. The compaction result (
tokensBefore/tokensAfter, shown in the compaction divider and used to rebase the context-size gauge) was computed on a messages-only basis — summary plus kept user messages — while the context size reported between exchanges comes from measured LLM usage, i.e. the full request (system prompt + tool schemas + messages + output). In a session with ~39k of fixed request overhead, compacting a ~90k context reported "→ 1.3k" even though the real next request was ~36.5k; the gauge dipped to the small messages-only estimate and then jumped back on the following exchange.What changed
tokensBeforeis now the full-request estimate (system prompt + non-deferred tools + messages) via the existing request-size helper.tokensAfteradds the estimated request overhead (system prompt + non-deferred tool schemas) on top of the measured summary output tokens and the kept user messages, so it stays on the same full-request basis as the measured exchange anchors. The token-counting ledger rebase after compaction carries this full-basis size, so the reported context size no longer dips and jumps.tokenCountuses the same basis.tokensAfterverbatim, so old compaction records restore unchanged.Checklist
gen-changesetsskill, or this PR needs no changeset.gen-docsskill, or this PR needs no doc update.