Codex review: needs maintainer review before merge. Reviewed July 6, 2026, 4:06 AM ET / 08:06 UTC.

Summary
The branch reorders runReplyAgent so memory flush runs before preflight compaction, updates direct-runtime ordering and exhausted-flush recovery tests, and corrects the memory-flush helper comment.

PR surface: Source +14, Tests +81. Total +95 across 3 files.

Reproducibility: yes. Source inspection of current main and v2026.6.11 shows preflight compaction runs before memory flush, while the docs and linked bug expect the flush before compaction; this read-only review did not run a local long-context Telegram reproduction.

Review metrics: none identified.

Root-cause cluster
Relationship: fixed_by_candidate
Canonical: #84695
Summary: This PR is the candidate fix for the open direct preflight-compaction-before-memory-flush regression; adjacent memoryFlush issues cover different failure modes or follow-up surfaces.

Members:

• canonical: [Bug]: Compaction does not write summary to daily memory — regression of #79022 #84695 - The PR uses closing syntax for this open bug and directly changes the reported runReplyAgent ordering.
• partial_overlap: Bug: in-turn auto-compaction can bypass memoryFlush because memoryFlush only runs pre-turn #62420 - That closed issue concerns in-turn auto-compaction bypassing memoryFlush, which overlaps the memoryFlush durability symptom but not this direct preflight ordering path.
• adjacent_distinct: fix(reply): flush memory for pending in-turn auto-compaction even when tokens drop below threshold #87088 - That open PR targets in-turn catch-up in shouldRunMemoryFlush; this PR reorders the pre-run maintenance phases.
• adjacent_distinct: memoryFlush has no escalation when assistant process dies mid-flush — session permanently over threshold, every subsequent message wedges #85645 - That issue concerns repeated memory-flush failures and wedged replies, not the initial order of preflight compaction versus memory flush.
• adjacent_distinct: fix: replies fail when memory flush is exhausted #100618 - That merged PR fixed memory-flush exhaustion recovery but left current main running preflight compaction before memory flush.
• 2 more in the report.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Merge readiness
Overall: 🐚 platinum hermit
Proof: 🦞 diamond lobster
Patch quality: 🐚 platinum hermit
Result: ready for maintainer review.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

• none.

Risk before merge

• [P1] Merging intentionally changes pre-run session maintenance order, so maintainers should confirm flush-before-compaction is the desired session-state invariant for existing memoryFlush configurations.
• [P1] The PR head was recently force-pushed and some broad CI jobs were still in progress during review; final merge should gate on exact-head CI for dc6888fd6bdab353ef307524c7b2d74fa75bea5f.

Maintainer options:

1. Land after ordering approval (recommended)
   Accept the session-state timing change and let normal exact-head CI/mergeability gates decide the final land step.
2. Revise the recovery policy
   Require a branch update if maintainers want exhausted memory flushes to rotate, compact, or notify under different conditions after the ordering change.
3. Defer to broader memoryFlush design
   Pause this PR if maintainers want to bundle the direct preflight-ordering fix with the adjacent in-turn catch-up or failure-escalation work.

Next step before merge

• No automated repair is needed; the branch is focused and reviewable, and the remaining action is maintainer acceptance of the session-state ordering change plus exact-head merge gates.

Maintainer decision needed

• Question: Should runReplyAgent make memory flush the canonical pre-compaction maintenance step for existing sessions?
• Rationale: The patch is focused and source-supported, but it deliberately changes session-state timing for memoryFlush metadata, compaction count, and degraded recovery behavior in existing conversations.
• Likely owner: steipete — This person has the strongest current history signal on the branch and the intersecting memory-flush exhaustion recovery path.
• Options:
  • Accept flush-before-compaction (recommended): Merge after exact-head CI if maintainers want durable memory to checkpoint the pre-compaction transcript whenever memoryFlush is enabled.
  • Keep preflight-first ordering: Close or redirect this PR if maintainers prefer the current preflight-first order and want the linked memory gap handled through a different mechanism.
  • Narrow before landing: Ask for a revised branch if maintainers want only the recovery tests or a different exhausted-flush policy without changing the canonical order yet.

Security
Cleared: The diff changes reply runtime ordering and colocated tests only; I found no new dependency, workflow, credential, permission, package, or downloaded-code surface.

ClawSweeper PR egg

✨ Hatched: 🌱 uncommon Moonlit Clawlet

Hatch command

Comment @clawsweeper hatch when this PR is hatchable.

Hatchability rules:

• Merged PRs are hatchable.
• Open PRs are hatchable when they are status: 👀 ready for maintainer look, status: 🚀 automerge armed, or labeled clawsweeper:automerge.
• Closed unmerged PRs are hatchable only when one of those hatchable labels is still present in the durable record.

Rarity: 🌱 uncommon.
Trait: sleeps inside passing CI.
Image traits: location CI tidepool; accessory review stamp; palette amber, ink, and glacier blue; mood celebratory; pose balancing on a branch marker; shell smooth pearl shell; lighting calm overcast light; background gentle dashboard dots.
Share on X: post this hatch
Copy: My PR egg hatched a 🌱 uncommon Moonlit Clawlet in ClawSweeper.

Real behavior proof for #84695 / this PR:

I reran the targeted maintenance-order regression locally on the PR branch and verified the memory-flush phase runs before preflight compaction.

Command:

/opt/homebrew/opt/node@22/bin/node scripts/run-vitest.mjs src/auto-reply/reply/agent-runner-direct-runtime-config.test.ts src/auto-reply/reply/agent-runner-memory.test.ts src/auto-reply/reply/reply-state.test.ts

Observed terminal result:

RUN  v4.1.6 /Users/andy/openclaw/openclaw-84695

Test Files  3 passed (3)
Tests  66 passed (66)

Specific behavior proved by the new regression:

expect(runMemoryFlushIfNeededMock.mock.invocationCallOrder[0]).toBeLessThan(
  runPreflightCompactionIfNeededMock.mock.invocationCallOrder[0] ?? Number.POSITIVE_INFINITY,
);

That exercises runReplyAgent and proves the pre-run maintenance ordering is now:

1. runMemoryFlushIfNeeded
2. runPreflightCompactionIfNeeded

This is the behavior the bug needs: the flush gate sees the pre-compaction session state, instead of seeing the lower post-compaction token count after preflight compaction has already run.

Additional validation:

/opt/homebrew/opt/node@22/bin/node /Users/andy/openclaw/node_modules/.bin/tsgo -p test/tsconfig/tsconfig.test.src.json --pretty false --declaration false --singleThreaded --checkers 1
# exit 0

/opt/homebrew/opt/node@22/bin/node /Users/andy/openclaw/node_modules/.bin/oxfmt --check --threads=1 src/auto-reply/reply/agent-runner.ts src/auto-reply/reply/agent-runner-direct-runtime-config.test.ts
# All matched files use the correct format.

@clawsweeper re-review

🦞🧹
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.

Re-review progress:

• State: Complete
• Detail: The targeted re-review finished, the durable review comment was updated, and the synced verdict was routed.
• Run: https://github.com/openclaw/clawsweeper/actions/runs/26204797050
• Updated: 2026-05-21T04:13:47.365Z

Heads up: this PR needs to be updated against current main before the new required Dependency Guard check can pass.

Rebased onto current origin/main and pushed the committed proof at head 26caf5e60ec2.

The proof is now in src/auto-reply/reply/agent-runner.runreplyagent.e2e.test.ts as starts memory flush before preflight compaction with real maintenance helpers; it runs real runReplyAgent and does not mock runMemoryFlushIfNeeded or runPreflightCompactionIfNeeded.

Local verification on this head:

node scripts/run-vitest.mjs src/auto-reply/reply/agent-runner.runreplyagent.e2e.test.ts src/llm/providers/mistral.test.ts src/llm/providers/anthropic.test.ts src/gateway/server-cron.test.ts src/gateway/server-cron-notifications.test.ts
# passed: 3 Vitest shards, 171 tests

pnpm check:test-types
# passed

pnpm lint
# passed

git diff --check
# clean

@clawsweeper re-review

🦞🧹
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.

Added the live Gateway proof ClawSweeper asked for.

Current head: 26caf5e60ec2

This is no longer only e2e/test-harness evidence. I started an isolated foreground Gateway on loopback with diagnostics enabled, sent a real chat.send through openclaw gateway call, and captured the runtime timeline.

Redacted timeline excerpt:

{"type":"span.start","timestamp":"2026-07-01T07:04:45.661Z","name":"gateway.chat_send.dispatch_inbound","phase":"agent-turn","attributes":{"runId":"pr84792-live-gateway-proof","sessionKey":"agent:main:main","provider":"proof-local","model":"proof-model"}}
{"type":"span.start","timestamp":"2026-07-01T07:04:46.038Z","name":"reply.memory_flush","phase":"agent-turn"}
{"type":"span.end","timestamp":"2026-07-01T07:04:46.038Z","name":"reply.memory_flush","phase":"agent-turn","durationMs":0.374}
{"type":"span.start","timestamp":"2026-07-01T07:04:46.038Z","name":"reply.preflight_compaction","phase":"agent-turn"}
{"type":"span.end","timestamp":"2026-07-01T07:04:46.041Z","name":"reply.preflight_compaction","phase":"agent-turn","durationMs":2.642}
{"type":"span.start","timestamp":"2026-07-01T07:04:46.043Z","name":"reply.run_agent_turn","phase":"agent-turn"}

The run then reached provider transport and failed against the intentionally unreachable local endpoint (127.0.0.1:9), after the ordering proof was already captured.

@clawsweeper re-review

PR body now has a literal ## Evidence section with the live Gateway diagnostics from the isolated chat.send run. The latest Real behavior proof workflow is green on head 26caf5e60ec2.

@clawsweeper re-review

Land-ready after maintainer rewrite.

• Memory flush now runs before preflight compaction and passes its returned session forward. Successful preflight recovery keeps the compacted session; exhausted context-overflow failure rotates; cancellation and unrelated preflight failures retain their existing paths.
• Removed the original unrelated Anthropic cast and redundant diagnostics fixture; final diff is 3 files.
• Focused proof: node scripts/run-vitest.mjs src/auto-reply/reply/agent-runner-direct-runtime-config.test.ts src/auto-reply/reply/agent-runner-memory.test.ts src/auto-reply/reply/reply-state.test.ts — 3 files, 98 tests passed on final head.
• Local pnpm build and pnpm check passed. Blacksmith Testbox lease tbx_01kwv3r1kemf11jnashga8a53t ran all 253 Vitest shards; the touched auto-reply shard passed, while unrelated baseline fixture/tooling failures were subsequently separated and the deterministic package fixture was fixed in test(release): match pinned Claude CLI setup #100723.
• Exact-head hosted CI passed all 44 jobs, including broad compact shards and QA smoke: https://github.com/openclaw/openclaw/actions/runs/28776569239
• Fresh local-delta and full-branch autoreviews are clean; wrapper review artifacts validate with zero findings and prepare-run completed for dc6888fd6bdab353ef307524c7b2d74fa75bea5f.

Codex contract check: native pre-turn and mid-turn compaction remain owned by codex-rs/core/src/session/turn.rs:796-819 and codex-rs/core/src/session/turn.rs:345-357, using codex-rs/core/src/session/context_window.rs:24-90; manual app-server compaction remains in codex-rs/app-server/src/request_processors/thread_processor.rs:1771-1782. OpenClaw continues to skip this preflight helper for Codex sessions.

Known proof gap: no live long-context provider conversation was forced; source-level ordering, 98 focused tests, the full remote suite, and exact-head hosted CI cover the changed decision surface. Thanks @TurboTheTurtle.

Merged via squash.

• Prepared head SHA: dc6888fd6bdab353ef307524c7b2d74fa75bea5f
• Landed commit: e24d4426fe8c7a4818e23c17ec9b368d91fb5aaf