feat(context-engine): pass incoming prompt to assemble by danhdoan · Pull Request #50848 · openclaw/openclaw

danhdoan · 2026-03-20T04:51:09Z

Add optional prompt parameter to the ContextEngine.assemble() interface so retrieval-oriented context engines can use the user's query for relevance-based assembly.

Summary

Problem: The ContextEngine.assemble() interface has no access to the current user prompt, so retrieval-oriented engines cannot use it for relevance-based context assembly.
Why it matters: Context engines that perform semantic search (e.g. RAG-based memory plugins) need the user's query to retrieve relevant context at assembly time.
What changed: Added optional prompt?: string to the assemble() params in the interface and forwarded params.prompt at the call site in attempt.ts. Added 2 tests.
What did NOT change (scope boundary): LegacyContextEngine is untouched — it doesn't use prompt and remains a pass-through. No behavioral change for existing engines.

Change Type

Scope

Linked Issue/PR

None

User-visible / Behavior Changes

None — additive optional parameter on an internal interface.

Security Impact (required)

New permissions/capabilities? No
Secrets/tokens handling changed? No
New/changed network calls? No
Command/tool execution surface changed? No
Data access scope changed? No

Repro + Verification

Steps

Register a custom context engine plugin that reads params.prompt in assemble()
Send a message to the gateway
Observe prompt is forwarded to the engine's assemble() call

Expected

params.prompt contains the user's message text

Actual

Confirmed via unit tests

Evidence

Failing test/log before + passing after
- 2 new tests: forwards prompt to the underlying engine and omits prompt when not provided — both pass

Compatibility / Migration

Backward compatible? Yes — parameter is optional, all existing engines continue to work unchanged
Config/env changes? No
Migration needed? No

Failure Recovery (if this breaks)

How to disable/revert this change quickly: Revert the commit (1 commit, 3 files)
Known bad symptoms reviewers should watch for: None expected — the param is optional and unused by existing engines

Risks and Mitigations

None — additive optional field with no behavioral change for existing code paths.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f5f8578161

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

greptile-apps · 2026-03-20T04:54:32Z

Greptile Summary

This PR adds an optional prompt?: string field to the ContextEngine.assemble() interface and forwards it at the call site in attempt.ts, enabling retrieval-oriented engines to perform relevance-based context assembly using the user's current query. The change is additive, backward-compatible, and well-scoped.

Key observations:

Behavioral inconsistency at the call site: attempt.ts writes prompt: params.prompt unconditionally, which means when params.prompt is undefined the assembled params object will contain the key prompt with value undefined — not truly absent. The test "omits prompt when not provided" only validates the direct-call path and does not catch this. Engine authors who guard with 'prompt' in params (a natural pattern) will see the property as always-present when called through attempt.ts.
LegacyContextEngine signature drift: The concrete class was left with the old assemble param type (no prompt field), while the test helper class LegacySessionKeyStrictEngine was updated in this same PR. TypeScript accepts this via bivariant method checking, but the implementation diverges from the interface it declares to implement.

Confidence Score: 3/5

Safe to merge with minor fix — the prompt: undefined key presence at the attempt.ts call site is inconsistent with the documented and tested contract.
The interface update and test logic are correct. However, the call site in attempt.ts always includes prompt as a key (possibly undefined) rather than omitting it, which contradicts the "omits prompt when not provided" test contract. An engine using 'prompt' in params as a guard would behave incorrectly. The fix is a one-line conditional spread.
Pay close attention to src/agents/pi-embedded-runner/run/attempt.ts line 2170 and the corresponding test coverage gap.

Comments Outside Diff (1)

src/context-engine/legacy.ts, line 38-43 (link)

LegacyContextEngine.assemble signature not updated to match interface

The ContextEngine interface now includes prompt?: string in the assemble params, but LegacyContextEngine.assemble() still declares the narrower param type without prompt. TypeScript won't error here due to bivariant method checking, but the class no longer accurately reflects the interface it implements.

This is in contrast to LegacySessionKeyStrictEngine (test helper) which was updated in this same PR to include prompt?: string. Consider adding the field here too for parity — it's a one-line no-op addition that keeps the implementation aligned with the interface:

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/context-engine/legacy.ts
Line: 38-43

Comment:
**`LegacyContextEngine.assemble` signature not updated to match interface**

The `ContextEngine` interface now includes `prompt?: string` in the `assemble` params, but `LegacyContextEngine.assemble()` still declares the narrower param type without `prompt`. TypeScript won't error here due to bivariant method checking, but the class no longer accurately reflects the interface it implements.

This is in contrast to `LegacySessionKeyStrictEngine` (test helper) which was updated in this same PR to include `prompt?: string`. Consider adding the field here too for parity — it's a one-line no-op addition that keeps the implementation aligned with the interface:



How can I resolve this? If you propose a fix, please make it concise.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/run/attempt.ts
Line: 2170

Comment:
**`prompt: undefined` key always present, contradicting test intent**

When `params.prompt` is not set, this line produces `prompt: undefined` in the object literal — the key **is** present in the params object, it just has value `undefined`. This directly contradicts the test named `"omits prompt when not provided"` which asserts `expect(calls[0]).not.toHaveProperty("prompt")`.

The test only passes because it calls `engine.assemble()` directly without `prompt`, bypassing this code path entirely. Any retrieval engine that guards with `'prompt' in params` (a reasonable pattern matching the stated contract) will incorrectly receive `true` on every call through `attempt.ts`, even when no user prompt was present.

To honour the documented intent — **truly** omit the key when no prompt is provided — use a conditional spread:

```suggestion
              ...(params.prompt !== undefined ? { prompt: params.prompt } : {}),
```

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: src/context-engine/legacy.ts
Line: 38-43

Comment:
**`LegacyContextEngine.assemble` signature not updated to match interface**

The `ContextEngine` interface now includes `prompt?: string` in the `assemble` params, but `LegacyContextEngine.assemble()` still declares the narrower param type without `prompt`. TypeScript won't error here due to bivariant method checking, but the class no longer accurately reflects the interface it implements.

This is in contrast to `LegacySessionKeyStrictEngine` (test helper) which was updated in this same PR to include `prompt?: string`. Consider adding the field here too for parity — it's a one-line no-op addition that keeps the implementation aligned with the interface:

```suggestion
  async assemble(params: {
    sessionId: string;
    sessionKey?: string;
    messages: AgentMessage[];
    tokenBudget?: number;
    prompt?: string;
  }): Promise<AssembleResult> {
```

How can I resolve this? If you propose a fix, please make it concise.

_{Last reviewed commit: "feat(context-engine)..."}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7eeb826a5c

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0977e912c1

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-03-20T15:24:49Z

+        if (
+          isLegacy &&
+          allowedKeys.some((key) => rejectedKeys.has(key) && hasOwnLegacyCompatKey(params, key))
+        ) {
+          return method(withoutLegacyCompatKeys(params, rejectedKeys));


Keep assemble retries active after sessionKey detection

If a strict legacy engine first rejects sessionKey in bootstrap(), compact(), or another earlier method, isLegacy becomes true with rejectedKeys = {"sessionKey"}. The fast path here then strips only sessionKey and calls assemble() directly, so a legacy engine that also rejects the new prompt field never reaches invokeWithLegacyCompat() and still throws. In the normal run path, runEmbeddedAttempt() calls contextEngine.bootstrap() before assemble() for resumed sessions, so prompt compatibility is still broken on existing session files.

Useful? React with 👍 / 👎.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 86d6fc3e43

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-03-20T15:39:02Z

+    /\bmust not have additional propert(?:y|ies)\b.*['"`]prompt['"`]/i,
+    /\b(?:unexpected|extraneous)\s+(?:property|properties|field|fields|key|keys)\b.*['"`]prompt['"`]/i,
+    /\b(?:unknown|invalid)\s+(?:property|properties|field|fields|key|keys)\b.*['"`]prompt['"`]/i,
+    /['"`]prompt['"`].*\b(?:was|is)\s+not allowed\b/i,


Restrict prompt fallback to true unknown-field validation errors

For prompt-aware engines that reject a supplied prompt for semantic reasons, this new pattern is too broad: messages like 'prompt' is not allowed when search mode is disabled also match here, so invokeWithLegacyCompat() will classify the failure as legacy-schema incompatibility, strip prompt, and silently retry without retrieval instead of surfacing the real error. Because run.ts reuses the resolved engine across retries, one misclassified failure suppresses prompt-based assembly for the rest of that run.

Useful? React with 👍 / 👎.

Add optional `prompt` parameter to the ContextEngine.assemble() interface so retrieval-oriented context engines can use the user's query for relevance-based assembly. Uses conditional spread at the call site to ensure the key is truly absent when no prompt is provided. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

@jalehman