fix(code): report total context after /offload - #5488
Merged
Conversation
github-actions
Bot
added
dcode
open-swe
Bot
changed the title
/offload
Mason Daugherty (mdrxy)
marked this pull request as ready for review
August 13, 2026 19:20
Mason Daugherty (mdrxy)
changed the title
github-actions
Bot
added
feature
![open-swe[bot]](https://avatars.githubusercontent.com/in/1395372?s=60&v=4){kind=small}
Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
The reported branch rebuilt the post-offload figure as `max(0, reported_before - conversation_before) + conversation_after`. When the provider total fell below the local conversation estimate -- a stale `_context_tokens`, or `count_tokens_approximately` overshooting by a single token -- the clamp collapsed the overhead to zero, leaving `tokens_before` on the provider scale (system prompt plus tool schemas included) while `tokens_after` dropped to conversation-only. The whole fixed overhead was then reported as freed context, which for a coding agent with a full tool schema is the dominant term. The percentage could also render as a negative "decrease". Subtract the conversation *delta* from the provider total instead. The two forms are algebraically equal whenever the clamp does not fire, but only this one keeps both figures on the provider's scale. Floor the percentage at zero, since a summary can come out larger than the messages it replaced. Also document what `MessageStore.turn_count` actually counts: rendered user-authored rows, which include local-only flows (`!shell`, most slash commands) that mount a `UserMessage` without invoking the server. That is a broader population than the "conversation turns" the offload report derives from graph state, and the two surfaces used the same word for it. Restore the rationale for slicing the pre-seed message list rather than the post-run one, now that two more counters depend on it, and cover the branches that mutation-tested as unheld: the low-total path, the percentage floor, `turns_offloaded`'s internal-message filter, the archive-failure wording, and the singular labels whose assertions were prefixes of their own plurals.
Mason Daugherty (mdrxy)
force-pushed
the
mdrxy/code/offload-context-report
branch
from
August 18, 2026 01:48
a3b88b5 to
78277ec
Compare
Mason Daugherty (mdrxy)
changed the title
/offload
github-actions
Bot
added
fix
…ntext-report # Conflicts: # libs/code/deepagents_code/app.py # libs/code/tests/unit_tests/test_offload.py
When a generated summary is larger than the messages it replaces, `tokens_after > tokens_before`, but the clamped `max(0, ...)` percentage rendered the result as `0% decrease` alongside the claim "freeing up context window space." Users were told context was freed even though `/offload` increased it. Render the signed outcome instead of clamping it away: - When `tokens_after > tokens_before`, the stats line shows `(increase)` instead of a percentage, and the surrounding text says the summary was larger than the messages it replaced, so context increased — no "freed" claim. - When the operation did reduce tokens, the existing `N% decrease` and "freeing up context window space" text is preserved. The archive-failure path now shares the same `outcome` prefix, so the error message no longer hardcodes "and freed context" either.
Mason Daugherty (mdrxy)
pushed a commit
that referenced
this pull request
Aug 18, 2026
> [!CAUTION] > Merging this PR will automatically publish to **PyPI** and create a **GitHub release**. For the full release process, see [`.github/RELEASING.md`](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md). --- _Release notes preview: keep this section in sync with the package `CHANGELOG.md`. Publish reads the merged CHANGELOG via `release.yml`, not this PR description — keep them aligned anyway so the PR stays an accurate historical record for reviewers and anyone returning later._ --- ## [0.1.57](deepagents-code==0.1.56...deepagents-code==0.1.57) (2026-08-18) ### Features - Added warnings before expensive cold-cache turns and trust user-declared endpoints for cold-cache policies ([#5439](#5439), [#5462](#5462)). - Made the chat input resizable by dragging its top border ([#5524](#5524)). - Added a `multi_select` question type to `ask_user` ([#5097](#5097)). - Added support for ACP approval modes ([#5394](#5394)). - Added `DeepSeek-V4-Pro-0813` to the model picker ([#5512](#5512)). - Show conversation turns alongside message counts ([#5571](#5571)). - Include `TERM_PROGRAM` in the resume hint ([#5548](#5548)). ### Bug Fixes - Report total context after `/offload` ([#5488](#5488)). - Fixed transcript and thread restoration issues, including hydration lag, scrolling resumed threads to the bottom, and hiding empty previous-thread hints ([#5479](#5479), [#5543](#5543), [#5552](#5552)). - Fixed Auto-mode approval handling by binding “yes” to the paired `ask_user` question and avoiding duplicate Auto denial notices ([#5038](#5038), [#5501](#5501)). - Improved reload behavior by keeping the chat input responsive during `/reload`, reporting MCP server changes, and avoiding plugin reload prompt flashes or startup hints ([#5529](#5529), [#5504](#5504), [#5500](#5500), [#5502](#5502)). - Improved dependency update UI by preserving editable fields and hiding dependency details after updates ([#5521](#5521), [#5519](#5519)). - Fixed chat UI polish issues, including detached spacer mount anchors, the unfocused input cursor, and relative timestamp toggle display ([#5516](#5516), [#5258](#5258), [#5503](#5503)). - Refresh the splash version after updates ([#5520](#5520)). _End release notes preview._ --- > [!NOTE] > A **community contributors** list and a **Special thanks** section (crediting the users who filed the issues this release's PRs closed) are appended to the GitHub release notes automatically at publish time (see [Release Pipeline](https://github.com/langchain-ai/deepagents/blob/main/.github/RELEASING.md#release-pipeline), step 3). --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: langchain-oss-automated-triage[bot] <248757908+langchain-oss-automated-triage[bot]@users.noreply.github.com>
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.
/offloadnow reports the change in total context (system prompt, tool schemas, and conversation) instead of passing off a conversation-only estimate as "Context".The previous report computed both sides of the before/after with
count_tokens_approximatelyover conversation messages only, then labeled the resultContext:. For a coding agent the fixed overhead (system prompt plus tool schema) is the dominant term, so the reported totals and the percentage reduction were both misleading — an offload that freed mostly conversation looked like it had freed the whole window.~because the delta is estimated.overhead + conversation_after) matters: the local estimator need only overshoot the provider count by a token for anoverhead = max(0, reported - conversation_before)clamp to collapse the overhead to zero, which would silently report the entire system prompt and tool schema as freed context.Conversation:so the two metrics are not mistaken for one another — percentages are not comparable across the two labels.