openclaw - 💡(How to fix) Fix [Bug]: Codex before_prompt_build receives empty messages after #88262 [1 pull requests]

Bug type

Regression / plugin contract break in the Codex app-server hook surface.

Summary

After PR #88262, the Codex app-server passes messages: [] into before_prompt_build, even when the session transcript was loaded successfully.

That breaks the public hook contract and silently degrades existing history-sensitive plugins. The PR intended to prevent mirrored history from being injected into Codex model input, but it also removed the hook's read-only view of the session history.

These are different surfaces and should not share the same empty array.

Why this is high confidence

• The hook type says before_prompt_build.messages are the "Session messages prepared for this run":
  • hook-before-agent-start.types.ts#L21-L26
• The agent loop docs say the hook runs after session load "with messages":
  • docs/concepts/agent-loop.md#L93-L94
• The merged Codex app-server code loads historyMessages, then replaces the messages passed to before_prompt_build with codexModelInputHistoryMessages, which is hard-coded to []:
  • run-attempt.ts#L767-L777
• The updated test now locks this in by asserting that a hook sees an empty message list even after an existing assistant message was appended:
  • run-attempt.test.ts#L1687-L1703
• In-tree plugins actually use event.messages:
  • active-memory derives recent turns from event.messages: extensions/active-memory/index.ts#L3110-L3119
  • memory-lancedb derives the latest user text from event.messages: extensions/memory-lancedb/index.ts#L1060-L1064

Diagram

flowchart TD
    A["Session JSONL / mirrored transcript"] --> B["readMirroredSessionHistoryMessages(...)"]
    B --> C["historyMessages populated"]
    C --> D["before_prompt_build hook"]
    C --> E["Codex model input"]

    D --> F["Plugin observes recent turns, latest user text, safety state"]
    E --> G["Native Codex thread receives current prompt only on resume"]

    H["Current merged code"] --> I["codexModelInputHistoryMessages = []"]
    I --> D
    I --> E
    I --> J["Hook loses observation history even though model-history dedupe was the real concern"]

Expected behavior

before_prompt_build should continue to receive the session messages that were prepared for the run, or a clearly documented filtered/redacted subset of them.

Separately, the Codex native model input can still use an empty parallel history array when the native Codex thread already owns conversation continuity.

In other words:

flowchart LR
    A["hookObservationMessages"] --> B["before_prompt_build"]
    C["codexModelInputHistoryMessages"] --> D["Codex turn/model input"]
    C --> E["May be [] for native resume"]
    A --> F["Should not be forced to [] just because C is []"]

Actual behavior

before_prompt_build receives [].

That means a plugin cannot tell whether:

• this is the first turn of a session,
• there are previous user constraints,
• the latest user message in the session differs from the current prompt,
• recent turns include context needed for memory recall,
• safety or routing state depends on recent messages.

The plugin sees the same messages: [] for a brand-new session and for a long-running session.

Minimal reproduction shape

1. Add a prior assistant or user message to the OpenClaw session file.
2. Register a before_prompt_build hook that records event.messages.
3. Run a Codex app-server attempt.
4. Inspect the hook input.

With the merged code, event.messages is [].

The current test suite already demonstrates this behavior in run-attempt.test.ts; the test expectation should be reversed or split so that hook-observation history and Codex model-input history are verified independently.

Impact

This is a plugin API compatibility break, not just an internal implementation detail.

Known in-tree impact:

• active-memory will produce weaker or incorrect recall context because recentTurns becomes empty.
• memory-lancedb can fall back to event.prompt even when the session has a more precise latest user message in event.messages.
• Any third-party plugin that uses before_prompt_build.messages for policy, prompt shaping, memory recall, audit, or safety checks now loses the transcript without an API version change.

This can also confuse lossless-context integrations indirectly. Even if the active context-engine path still receives historyMessages during assembly, plugins that run in the same pre-prompt phase no longer see the loaded messages described by the hook contract.

Suggested fix direction

Keep the PR's intended model-input behavior, but split the variables:

• hookObservationMessages: the loaded, filtered session messages passed to before_prompt_build.
• codexModelInputHistoryMessages: the actual history sent as parallel Codex model input, which may remain [].
• diagnosticModelInputMessages: whatever the model diagnostics need to reflect the actual Codex input.

If Codex app-server intentionally wants a different hook contract than other harnesses, the type and docs should be changed explicitly, and the hook should receive a reason field such as messagesUnavailableReason: "codex-native-thread-owns-history". Right now it looks like a silent contract regression.

Evidence checked

• Merged PR: openclaw/openclaw#88262
• Merge commit: 530351e394a19b1dd2943cb08259657a13f90572
• PR head audited locally: 04a4427d0c0b8f2e7bb666bbbcda5c557d033972