fix(compaction): avoid cached usage overcount by LZY3538 · Pull Request #99864 · openclaw/openclaw

LZY3538 · 2026-07-04T07:01:49Z

What Problem This Solves

Long cached Anthropic tool loops could report aggregate cache billing buckets as
if they were one live context-window snapshot. That inflated token counts and
could trigger compaction well below the configured reserve threshold.

Why This Change Was Made

Both Anthropic stream adapters now preserve the prompt snapshot from
message_start for context-window accounting while retaining the final
aggregate cache buckets for billing and cost reporting.

The shared context-token helper prefers an explicit prompt snapshot, then
derives the latest call's prompt size from total - output, and finally falls
back to component sums. Accumulated session usage remains component-based.
Metadata persistence and timeout-driven compaction use the same helper so the
decision is consistent across runner paths.

User Impact

Cached Anthropic sessions no longer compact early because repeated cache-read
and cache-write billing totals are mistaken for the current prompt size.
Billing totals remain intact, and sessions above the real context threshold
still compact normally.

Evidence

Exact PR head: b054cb5c5921218b0c5b2b8783efb84cf3728df8.
Local focused proof:
node scripts/run-vitest.mjs src/agents/usage.test.ts src/agents/embedded-agent-runner/run/helpers.test.ts src/llm/providers/anthropic.test.ts src/agents/anthropic-transport-stream.test.ts src/auto-reply/status.test.ts src/agents/embedded-agent-runner/run.empty-error-retry.test.ts src/agents/embedded-agent-runner/run.compaction-loop-guard.test.ts src/agents/embedded-agent-runner/run.overflow-compaction.loop.test.ts src/agents/embedded-agent-runner/run.overflow-compaction.test.ts
- 9 files, 380 tests passed.
Blacksmith Testbox tbx_01kwr9ecvcty3pvvtthtkw60kc:
- verified the exact head SHA;
- repeated the same 380 focused tests successfully;
- pnpm check:changed passed, including production/test typechecks, all core
  oxlint shards, import-cycle checks, and changed-surface guards.
Fresh autoreview:
.agents/skills/autoreview/scripts/autoreview --mode local --prompt-file .local/review-ci-followup.md --stream-engine-output
- no accepted/actionable findings;
- patch correct, confidence 0.98.
Hosted compact jobs still expose four unrelated docs-sync failures. Each was
reproduced unchanged on current origin/main at
1d9fb2773f04ee8828d98077130b06496344ac2a:
src/docs/environment-docs.test.ts,
src/config/talk-defaults.test.ts,
src/agents/minimax-docs.test.ts, and
src/cron/isolated-agent/run-fallback-policy.test.ts.
This PR changes none of those files or docs.

clawsweeper · 2026-07-04T07:03:00Z

Codex review: stale review; fresh review needed.

Summary
The latest durable ClawSweeper review was for head f8f7878534960589affe2b6b736b00d9bcd2ae95, but the PR head is now 5fc8f0ab37c208b7864262269ac55e307d473316. Its old verdict and PR readiness labels are no longer current.

Next step
Run or wait for a fresh ClawSweeper review on the current PR head.

Review history (8 earlier review cycles)

reviewed 2026-07-04T07:15:55.535Z sha 6cc347b :: needs real behavior proof before merge. :: [P1] Preserve cached prompt tokens in the fallback
reviewed 2026-07-04T09:18:23.136Z sha 4bc55b6 :: found issues before merge. :: [P1] Carry a real context snapshot before dropping cache buckets
reviewed 2026-07-04T12:06:46.712Z sha f38d668 :: needs real behavior proof before merge. :: none
reviewed 2026-07-04T12:19:52.464Z sha f38d668 :: needs real behavior proof before merge. :: none
reviewed 2026-07-04T12:27:29.837Z sha f38d668 :: needs maintainer review before merge. :: none
reviewed 2026-07-04T12:38:02.900Z sha f38d668 :: needs maintainer review before merge. :: none
reviewed 2026-07-05T01:34:36.716Z sha 5e6c485 :: needs maintainer review before merge. :: none
reviewed 2026-07-05T07:35:36.799Z sha 3657fa3 :: needs real behavior proof before merge. :: none

clawsweeper · 2026-07-04T07:46:42Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

clawsweeper · 2026-07-04T12:12:03Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

vincentkoc · 2026-07-05T09:24:27Z

Maintainer repair landed as e6ec74c25430f2df94eb6a99fb5e050ebbdd87ad from final PR head 5fc8f0ab37c208b7864262269ac55e307d473316.

The repaired implementation separates cumulative Anthropic billing buckets from the final-iteration context snapshot across provider streams, compaction/overflow decisions, persistence, active-branch transcript recovery, and fork preflight. Partial final usage updates preserve already-accumulated prompt buckets; unavailable snapshots remain fail-closed.

Validation:

node scripts/run-vitest.mjs src/agents/sessions/agent-session.context-usage.test.ts src/agents/usage.test.ts src/gateway/session-utils.fs.test.ts src/gateway/session-transcript-readers.test.ts src/auto-reply/reply/session-fork.runtime.test.ts packages/agent-core/src/harness/compaction/compaction.test.ts packages/ai/src/utils/overflow.test.ts packages/ai/src/providers/anthropic-usage.test.ts packages/ai/src/providers/anthropic.test.ts src/agents/anthropic-transport-stream.test.ts src/agents/embedded-agent-runner/run.timeout-triggered-compaction.test.ts src/agents/embedded-agent-runner.sanitize-session-history.test.ts src/agents/embedded-agent-runner/run/attempt.test.ts src/agents/embedded-agent-runner/run/helpers.test.ts - 838 passed across five shards.
node scripts/generate-docs-map.mjs --check - passed.
git diff --check - passed.
Autoreview - no findings after the partial-final-delta repair; final whole-branch confidence 0.82.
Final exact-head CI - 54/54 jobs passed: https://github.com/openclaw/openclaw/actions/runs/28736168153
Final Workflow Sanity - passed: https://github.com/openclaw/openclaw/actions/runs/28736168157
OPENCLAW_TESTBOX=1 scripts/pr prepare-run 99864 - passed in hosted_exact_head mode.

Known proof gaps: none.

vincentkoc · 2026-07-05T09:33:48Z

Merged via squash.

Prepared head SHA: 5fc8f0ab37c208b7864262269ac55e307d473316
Landed commit: e6ec74c25430f2df94eb6a99fb5e050ebbdd87ad

openclaw-barnacle Bot added the size: S label Jul 4, 2026

LZY3538 marked this pull request as ready for review July 4, 2026 07:07

clawsweeper Bot mentioned this pull request Jul 4, 2026

fix(compaction): exclude turn-aggregated cache fields from context token fallback #99863

Closed

openclaw-barnacle Bot added agents Agent runtime and tooling size: M and removed size: S labels Jul 4, 2026

clawsweeper Bot removed rating: 🧂 unranked krab Not merge-ready due to missing proof or serious correctness/safety concerns. status: 📣 needs proof The PR needs real behavior proof before ClawSweeper can clear the contributor ask. labels Jul 4, 2026

clawsweeper Bot removed rating: 🦞 diamond lobster Very strong PR readiness with only minor maintainer review expected. status: 👀 ready for maintainer look ClawSweeper has no concrete contributor-facing blocker left for this PR. labels Jul 5, 2026

vincentkoc force-pushed the codex/99843-compaction-context-estimate branch 2 times, most recently from db59127 to 3657fa3 Compare July 5, 2026 05:35

vincentkoc force-pushed the codex/99843-compaction-context-estimate branch from 3657fa3 to 405ba6b Compare July 5, 2026 09:00

openclaw-barnacle Bot added docs Improvements or additions to documentation gateway Gateway runtime size: XL and removed size: M labels Jul 5, 2026

vincentkoc force-pushed the codex/99843-compaction-context-estimate branch from 405ba6b to f8f7878 Compare July 5, 2026 09:14

fix(compaction): preserve Anthropic context usage

5fc8f0a

vincentkoc force-pushed the codex/99843-compaction-context-estimate branch from f8f7878 to 5fc8f0a Compare July 5, 2026 09:24

vincentkoc merged commit e6ec74c into openclaw:main Jul 5, 2026
98 checks passed

github-actions Bot mentioned this pull request Jul 5, 2026

📡 Upstream Digest — 2026-07-05 10:03 UTC curtismercier/openclaw-mods#1212

Open

github-actions Bot pushed a commit to Desicool/openclaw that referenced this pull request Jul 6, 2026

fix(compaction): preserve Anthropic context usage (openclaw#99864)

040bd6e

jalehman mentioned this pull request Jul 6, 2026

refactor(sessions): route new session store bypasses through the accessor #101179

Merged

sercada mentioned this pull request Jul 8, 2026

[Bug]: context-overflow-midturn-precheck estimator over-counts ~2.3-2.6x vs billed usage (tool-result chars at 2 chars/token over the untruncated transcript) #101929

Open

This was referenced Jul 8, 2026

fix(agents): reduce tool-result token over-count in mid-turn precheck #101950

Closed

fix(agents): align mid-turn tool-result projection precheck #101952

Closed

ObliviateRickLin mentioned this pull request Jul 15, 2026

fix(sessions): snapshot last-call context, not aggregate cacheRead (#108238) #108272

Closed

This was referenced Jul 15, 2026

fix: prevent cumulative usage from inflating session context #108323

Merged

[Bug]: 2026.7.1 中会话上下文用量把累计 cacheRead 算进 totalTokens，导致误报上下文超限并触发压缩失败 #108238

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

fix(compaction): avoid cached usage overcount#99864

fix(compaction): avoid cached usage overcount#99864
vincentkoc merged 1 commit into
openclaw:mainfrom
LZY3538:codex/99843-compaction-context-estimate

LZY3538 commented Jul 4, 2026 •

edited by vincentkoc

Loading

Uh oh!

clawsweeper Bot commented Jul 4, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jul 4, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jul 4, 2026 •

edited

Loading

Uh oh!

vincentkoc commented Jul 5, 2026 •

edited

Loading

Uh oh!

Uh oh!

vincentkoc commented Jul 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

LZY3538 commented Jul 4, 2026 • edited by vincentkoc Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What Problem This Solves

Why This Change Was Made

User Impact

Evidence

Uh oh!

clawsweeper Bot commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

clawsweeper Bot commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

clawsweeper Bot commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vincentkoc commented Jul 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

vincentkoc commented Jul 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

LZY3538 commented Jul 4, 2026 •

edited by vincentkoc

Loading

clawsweeper Bot commented Jul 4, 2026 •

edited

Loading

clawsweeper Bot commented Jul 4, 2026 •

edited

Loading

clawsweeper Bot commented Jul 4, 2026 •

edited

Loading

vincentkoc commented Jul 5, 2026 •

edited

Loading