[AI-assisted] fix(agents): invalidate context engine cache by brokemac79 · Pull Request #78163 · openclaw/openclaw

brokemac79 · 2026-05-06T00:55:46Z

Summary

Clear the context-engine loop hook's cached assembled view when the source message history shrinks.
Clear that cache when context-engine assembly fails, so fallback uses the current source messages.
Add focused regressions for both stale-cache paths.

Real behavior proof

Behavior or issue addressed: installContextEngineLoopHook could return a previously cached assembled context view after the live source history shrank or after a later assemble() failure, allowing stale pre-reset context to be reused.

Real environment tested: Local Windows checkout of this PR branch at commit 2663018ee2, using Node/pnpm from the repository toolchain.

Exact steps or command run after this patch: Inspected the patched hook and ran git diff --check, corepack pnpm exec oxfmt --check --threads=1 src/agents/pi-embedded-runner/tool-result-context-guard.ts src/agents/pi-embedded-runner/tool-result-context-guard.test.ts CHANGELOG.md, corepack pnpm test src/agents/pi-embedded-runner/tool-result-context-guard.test.ts, and corepack pnpm check:changed.

Evidence after fix: Focused regression tests now prove the cache invalidates in both reported failure modes:

clears the cached assembled view when the source history shrinks
clears the cached assembled view when assemble fails

The focused test shard passed:

Test Files  1 passed (1)
Tests       34 passed (34)

Observed result after fix: When source history shrinks, the hook clears the cached assembled view and returns the fresh source messages. When assemble() throws after a previous successful cached view, the hook clears that cache and subsequent unchanged iterations no longer reuse stale assembled messages.

What was not tested: I did not run a live Gateway session reset with a real context-engine plugin. This proof uses the core hook's focused regression tests and source-level behavior.

Validation

git diff --check
corepack pnpm exec oxfmt --check --threads=1 src/agents/pi-embedded-runner/tool-result-context-guard.ts src/agents/pi-embedded-runner/tool-result-context-guard.test.ts CHANGELOG.md
corepack pnpm test src/agents/pi-embedded-runner/tool-result-context-guard.test.ts
corepack pnpm check:changed

clawsweeper · 2026-05-06T00:58:45Z

Codex review: needs real behavior proof before merge.

Summary
The PR invalidates installContextEngineLoopHook's cached assembled context view on source-history shrink or assembly failure, adds focused regression tests, and updates the changelog.

Reproducibility: yes. source-level: current main keeps lastAssembledView, returns it on unchanged calls, and leaves it intact after assembly failure. I did not run a live Gateway/lossless-claw reset path in this read-only review.

Real behavior proof
Needs real behavior proof before merge: The PR supplies focused test/check output and says no live Gateway session reset with a real context-engine plugin was run, so it still needs after-fix real behavior proof before merge. After adding proof, update the PR body; ClawSweeper should re-review automatically. If it does not, ask a maintainer to comment @clawsweeper re-review.

Next step before merge
Contributor live runtime proof and maintainer check-gate judgment remain the blockers; there is no narrow code repair for automation to make on this PR branch.

Security
Cleared: The diff only changes agent cache logic, focused tests, and the changelog; no concrete security or supply-chain concern was found in the patch.

Review details

Best possible solution:

Land the narrow cache invalidation after live Gateway/context-engine proof shows reset and assemble-failure paths no longer reuse stale pre-reset messages, and after required checks are green or explicitly deemed unrelated.

Do we have a high-confidence way to reproduce the issue?

Yes, source-level: current main keeps lastAssembledView, returns it on unchanged calls, and leaves it intact after assembly failure. I did not run a live Gateway/lossless-claw reset path in this read-only review.

Is this the best way to solve the issue?

Yes, the code direction is the narrow maintainable fix: track the source length that produced the cached assembled view and clear the cache on shrink or engine failure. The remaining blocker is proof and gate readiness, not a definite implementation defect.

What I checked:

Current main cache reuse path: Current main stores lastAssembledView and returns it on no-new-message iterations without checking whether the source history has shrunk. (src/agents/pi-embedded-runner/tool-result-context-guard.ts:252, 97b07eaeaf38)
Current main assemble-failure path: The catch path falls back to source messages but does not clear a previously cached assembled view, so a later unchanged call can reuse stale assembled messages. (src/agents/pi-embedded-runner/tool-result-context-guard.ts:323, 97b07eaeaf38)
Runtime installation path: runEmbeddedAttempt installs installContextEngineLoopHook when the active context engine owns compaction, so this is core runtime behavior for context-engine plugins. (src/agents/pi-embedded-runner/run/attempt.ts:1738, 97b07eaeaf38)
Context-engine contract: The docs describe assemble as the hook called before each model run, returning the ordered messages used for the prompt. Public docs: docs/concepts/context-engine.md. (docs/concepts/context-engine.md:70, 97b07eaeaf38)
PR patch direction: The PR adds lastAssembledFromSourceLength, clears the assembled cache when source length shrinks, records source length on cache writes, and clears the cache on assemble failure. (src/agents/pi-embedded-runner/tool-result-context-guard.ts:252, 66d59ba08fca)
Regression coverage: The PR adds tests for source-history shrink and assemble-failure cache invalidation, while current main already has cached-view reuse coverage. (src/agents/pi-embedded-runner/tool-result-context-guard.test.ts:760, 66d59ba08fca)

Likely related people:

Bikkies: Commit fecd4fcc5568 introduced the per-iteration context-engine ingest/assemble hook and the focused hook tests where this cache lives. (role: introduced behavior; confidence: high; commits: fecd4fcc5568; files: src/agents/pi-embedded-runner/tool-result-context-guard.ts, src/agents/pi-embedded-runner/tool-result-context-guard.test.ts, src/agents/pi-embedded-runner/run/attempt.ts)
jalehman: The introducing commit lists jalehman as co-author/reviewer, later context-engine prompt-cache work touched the same lifecycle boundary, and this PR is currently assigned to jalehman with a changelog attribution commit. (role: adjacent context-engine owner/reviewer; confidence: high; commits: fecd4fcc5568, a327b6750dbe, 66d59ba08fca; files: src/agents/pi-embedded-runner/tool-result-context-guard.ts, src/agents/pi-embedded-runner/run/attempt.ts, CHANGELOG.md)

Remaining risk / open question:

Contributor-supplied after-fix proof is still test/source-level only; no live Gateway reset or real context-engine plugin run is shown on the PR branch.
Latest head CI was not fully green at inspection time, with aggregate security failures and multiple queued broad checks needing maintainer interpretation or rerun.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 97b07eaeaf38.

brokemac79 · 2026-05-06T06:29:05Z

Rebased this branch onto current openclaw/main and resolved the CHANGELOG.md conflict by keeping both upstream's plugin/Teams entries and this PR's Agents/context engine entry.

New head: ec9e1054e1

Focused validation after the rebase:

corepack pnpm test src/agents/pi-embedded-runner/tool-result-context-guard.test.ts
corepack pnpm exec oxfmt --check CHANGELOG.md src/agents/pi-embedded-runner/tool-result-context-guard.test.ts src/agents/pi-embedded-runner/tool-result-context-guard.ts
corepack pnpm check:changed
corepack pnpm check:changelog-attributions
git diff --check origin/main...HEAD

brokemac79 · 2026-05-06T07:38:44Z

Rebased this branch onto current openclaw/main after upstream moved again, resolving the CHANGELOG.md conflict by keeping both upstream's Gateway/sessions entry and this PR's Agents/context engine entry.

New head: c575501fdd

Focused validation after the rebase:

corepack pnpm test src/agents/pi-embedded-runner/tool-result-context-guard.test.ts
corepack pnpm exec oxfmt --check CHANGELOG.md src/agents/pi-embedded-runner/tool-result-context-guard.test.ts src/agents/pi-embedded-runner/tool-result-context-guard.ts
corepack pnpm check:changed
corepack pnpm check:changelog-attributions
git diff --check origin/main...HEAD

brokemac79 · 2026-05-06T07:42:42Z

Fresh CI on c575501fdd has Real behavior proof, OpenGrep, changed-path scanning, and the focused/security checks passing. The remaining red checks are outside this PR's diff:

check-lint: src/gateway/server.cron.test.ts:194 — unnecessary spread operator
check-test-types: src/gateway/server.cron.test.ts:444 — Property 'run' does not exist on type '{ start: () => Promise<void>; stop: () => void; }'

This PR only changes:

CHANGELOG.md
src/agents/pi-embedded-runner/tool-result-context-guard.ts
src/agents/pi-embedded-runner/tool-result-context-guard.test.ts

I am not folding unrelated gateway cron test fixes into this context-engine PR; leaving this for upstream/main or maintainer rerun.

jalehman · 2026-05-06T21:33:56Z

Merged via squash.

Prepared head SHA: 6ed3add79759107e1e77053cde93014d475400b3
Merge commit: 609a5d70a5501b230f4f8cff6aabe8ac537f6e8a

Thanks @brokemac79!

@jalehman

…78163) Merged via squash. Prepared head SHA: 6ed3add Co-authored-by: brokemac79 <255583030+brokemac79@users.noreply.github.com> Co-authored-by: jalehman <550978+jalehman@users.noreply.github.com> Reviewed-by: @jalehman

@jalehman

Merged via squash. Prepared head SHA: 6ed3add Co-authored-by: brokemac79 <255583030+brokemac79@users.noreply.github.com> Co-authored-by: jalehman <550978+jalehman@users.noreply.github.com> Reviewed-by: @jalehman (cherry picked from commit 609a5d7)

@jalehman

…78163) Merged via squash. Prepared head SHA: 6ed3add Co-authored-by: brokemac79 <255583030+brokemac79@users.noreply.github.com> Co-authored-by: jalehman <550978+jalehman@users.noreply.github.com> Reviewed-by: @jalehman

@jalehman

…78163) Merged via squash. Prepared head SHA: 6ed3add Co-authored-by: brokemac79 <255583030+brokemac79@users.noreply.github.com> Co-authored-by: jalehman <550978+jalehman@users.noreply.github.com> Reviewed-by: @jalehman

openclaw-barnacle Bot added agents Agent runtime and tooling size: S proof: supplied External PR includes structured after-fix real behavior proof. labels May 6, 2026

brokemac79 force-pushed the fix-context-engine-cache-invalidation branch from 2663018 to ec9e105 Compare May 6, 2026 06:28

brokemac79 force-pushed the fix-context-engine-cache-invalidation branch from ec9e105 to c575501 Compare May 6, 2026 07:38

jalehman self-assigned this May 6, 2026

jalehman force-pushed the fix-context-engine-cache-invalidation branch 4 times, most recently from f1a4693 to 546fced Compare May 6, 2026 21:31

brokemac79 and others added 2 commits May 6, 2026 14:32

fix(agents): invalidate context engine cache

8c4fc90

docs: add changelog attribution

6ed3add

jalehman force-pushed the fix-context-engine-cache-invalidation branch from 546fced to 6ed3add Compare May 6, 2026 21:33

jalehman merged commit 609a5d7 into openclaw:main May 6, 2026
19 checks passed

steipete mentioned this pull request May 6, 2026

installContextEngineLoopHook caches lastAssembledView indefinitely without invalidation #77968

Closed

github-actions Bot mentioned this pull request May 6, 2026

📡 Upstream Digest — 2026-05-06 22:41 UTC curtismercier/openclaw-mods#787

Open

github-actions Bot mentioned this pull request May 9, 2026

Update ghcr.io/openclaw/openclaw Docker tag to v2026.5.7 - autoclosed claytono/infra#1977

Closed

1 task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[AI-assisted] fix(agents): invalidate context engine cache#78163

[AI-assisted] fix(agents): invalidate context engine cache#78163
jalehman merged 2 commits intoopenclaw:mainfrom
brokemac79:fix-context-engine-cache-invalidation

brokemac79 commented May 6, 2026

Uh oh!

clawsweeper Bot commented May 6, 2026 •

edited

Loading

Uh oh!

brokemac79 commented May 6, 2026

Uh oh!

brokemac79 commented May 6, 2026

Uh oh!

brokemac79 commented May 6, 2026

Uh oh!

Uh oh!

jalehman commented May 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

brokemac79 commented May 6, 2026

Summary

Real behavior proof

Validation

Uh oh!

clawsweeper Bot commented May 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

brokemac79 commented May 6, 2026

Uh oh!

brokemac79 commented May 6, 2026

Uh oh!

brokemac79 commented May 6, 2026

Uh oh!

Uh oh!

jalehman commented May 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

clawsweeper Bot commented May 6, 2026 •

edited

Loading