feat(google): add Gemini execution guidance via prompt overlay [v3 6/6] by 100yenadmin · Pull Request #66379 · openclaw/openclaw

100yenadmin · 2026-04-14T06:04:11Z

GPT 5.4 Enhancement v3 — PR 6/6

Tracking: #66345 | Issue: #66352
Priority: P2 — LOW | Type: Provider Support

Problem

Hermes Agent includes Gemini/Gemma-specific operational guidance that addresses known failure modes. OpenClaw's Google extension has no prompt overlay — Gemini models get only the generic execution bias. This was filed as part of the same v3 for GPT 5.4 given the 0/10 parity scoring gap for Gemini.

Changes

New: extensions/google/prompt-overlay.ts

GOOGLE_GEMINI_EXECUTION_GUIDANCE — 7 operational directives:
- Absolute paths for all file operations
- Verify-first: check file contents before modifying
- Dependency checks: verify package availability
- Conciseness: brief explanations, focus on actions
- Parallel tool calls: batch independent operations
- Non-interactive commands: use -y/--yes flags
- Keep going: work autonomously until done
resolveGoogleSystemPromptContribution() → returns guidance via sectionOverrides.tool_enforcement

Modified: extensions/google/provider-registration.ts

Wire resolveSystemPromptContribution into Google provider

Modified: src/agents/system-prompt-contribution.ts + src/agents/system-prompt.ts

Add tool_enforcement section ID (shared with PR 3)

Architecture

  ┌────────────────────────────────────────────────────┐
  │  Provider Prompt Overlay Pattern                    │
  │                                                     │
  │  extensions/openai/prompt-overlay.ts  (existing)    │
  │  ├── GPT-5 execution bias                           │
  │  ├── GPT-5 tool enforcement                         │
  │  ├── GPT-5 output contract                          │
  │  └── Friendly personality overlay                   │
  │                                                     │
  │  extensions/google/prompt-overlay.ts  (this PR)     │
  │  └── Gemini operational directives                  │
  │      (absolute paths, verify first, parallel calls) │
  │                                                     │
  │  Both use: sectionOverrides.tool_enforcement        │
  └────────────────────────────────────────────────────┘

Hermes Reference

agent/prompt_builder.py lines 256-276 — GOOGLE_MODEL_OPERATIONAL_GUIDANCE.

Verification

Gemini models receive operational directives in system prompt
Non-Google models don't receive Gemini directives
File-manipulation tasks use absolute paths

Add Gemini-specific operational directives that address known failure modes: relative paths, missing dependency checks, interactive CLI prompts hanging, and sequential-when-parallel tool calls. Changes: - New: extensions/google/prompt-overlay.ts with GOOGLE_GEMINI_EXECUTION_GUIDANCE constant and resolver - Wire resolveSystemPromptContribution into Google provider registration - Add tool_enforcement section ID (shared with PR 3) to support the sectionOverrides mechanism Ported from Hermes Agent's GOOGLE_MODEL_OPERATIONAL_GUIDANCE (prompt_builder.py lines 256-276). Part of GPT 5.4 Enhancement v3 sprint. Closes #66352. Tracking issue: #66345.

greptile-apps · 2026-04-14T06:06:18Z

Greptile Summary

This PR adds a Gemini-specific prompt overlay for the Google extension, wiring operational directives (absolute paths, verify-first, parallel tool calls, etc.) into the tool_enforcement section of the agent system prompt. The core changes to system-prompt-contribution.ts and system-prompt.ts are clean and minimal.

The new extensions/google/prompt-overlay.ts imports ProviderSystemPromptContribution directly from ../../src/agents/system-prompt-contribution.js, which violates the extension boundary rule — extension production code must only cross into core via openclaw/plugin-sdk/*. This will be caught by the check-additional CI gate. The fix is to either drop the explicit return-type annotation (letting it be inferred from the hook's declared signature, as the OpenAI extension does) or expose the type through the plugin SDK in the same change.

Confidence Score: 4/5

Not safe to merge until the src/** import in the extension is replaced with the SDK-safe pattern.
One P1 finding blocks merge: the direct ../../src/agents/system-prompt-contribution.js import from extension production code violates the documented and CI-enforced extension boundary, and will fail check-additional. Everything else is sound — the prompt content, the core section plumbing, and the overlay wiring are correct.
extensions/google/prompt-overlay.ts — fix the src/** import before landing.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: extensions/google/prompt-overlay.ts
Line: 2

Comment:
**Extension boundary violation: direct import from `src/**`**

`ProviderSystemPromptContribution` is imported directly from `../../src/agents/system-prompt-contribution.js`, which crosses into core `src/**` from extension production code. CLAUDE.md is explicit: "extensions must cross into core only through `openclaw/plugin-sdk/*`." The `check-additional` CI gate enforces this boundary.

The OpenAI extension (`extensions/openai/prompt-overlay.ts`) avoids this entirely — it returns the shape without annotating the return type, letting it be inferred from the hook's declared signature in `src/plugins/types.ts`. Either drop the explicit type annotation here and rely on inference, or expose `ProviderSystemPromptContribution` through a `openclaw/plugin-sdk/*` subpath in the same change.

```suggestion
import type { ProviderSystemPromptContribution } from "openclaw/plugin-sdk/provider-entry";
```
*(Only valid if/when the type is exported from that subpath; otherwise drop the import and remove the explicit return type annotation below.)*

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

---

This is a comment left during a code review.
Path: extensions/google/provider-registration.ts
Line: 62-66

Comment:
**Hardcoded provider ID makes `shouldApplyGooglePromptOverlay` a no-op guard**

`modelProviderId` is always passed as the literal `"google"`, so the guard inside `resolveGoogleSystemPromptContribution` will unconditionally pass for every invocation of this hook. The `shouldApplyGooglePromptOverlay` logic (and the other two entries in `GOOGLE_PROVIDER_IDS`) becomes dead code from this call site. If the intent is that all aliases (`google-vertex`, `google-antigravity`) also receive the overlay, pass `ctx.provider` instead so the guard is live and the function's semantics are preserved.

```suggestion
    resolveSystemPromptContribution: (ctx) =>
      resolveGoogleSystemPromptContribution({
        modelProviderId: ctx.provider,
        modelId: ctx.modelId,
      }),
```

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

_{Reviews (1): Last reviewed commit: "feat(google): add Gemini execution guida..." | Re-trigger Greptile}

Copilot

Pull request overview

Adds a Google (Gemini) provider prompt overlay to inject model-specific operational/tooling guidance into the agent system prompt via a new tool_enforcement overridable section.

Changes:

Introduces extensions/google/prompt-overlay.ts defining Gemini operational directives and a resolver returning sectionOverrides.tool_enforcement.
Wires resolveSystemPromptContribution into the Google provider registration.
Extends system prompt contribution plumbing by adding the tool_enforcement section id and inserting it into the assembled system prompt.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

File	Description
`src/agents/system-prompt.ts`	Inserts an overridable `tool_enforcement` section into the built system prompt.
`src/agents/system-prompt-contribution.ts`	Adds `tool_enforcement` to `ProviderSystemPromptSectionId`.
`extensions/google/provider-registration.ts`	Wires Google provider `resolveSystemPromptContribution` to the new overlay resolver.
`extensions/google/prompt-overlay.ts`	New Google overlay defining Gemini directives and returning them via `sectionOverrides.tool_enforcement`.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 199932cb75

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

- Remove direct src/ import (extension boundary violation). Use inline return type instead of importing ProviderSystemPromptContribution. - Wire resolveSystemPromptContribution to Gemini CLI provider (google-gemini-cli) in addition to the base google provider. - Add google-gemini-cli to GOOGLE_PROVIDER_IDS set. - Remove "Absolute paths" directive (conflicts with sandbox guidance that recommends relative paths in sandboxed environments). - Replace em dashes with commas/periods per GPT-5 output contract.

Copilot

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.

Parity audit: previous iteration rewrote the Gemini directives based on bot suggestions (em dash ban, generic wording), diverging from Hermes's GOOGLE_MODEL_OPERATIONAL_GUIDANCE (prompt_builder.py:258-276). Restore Hermes text verbatim: - "# Google model operational directives" heading (was ## Gemini ...) - Em dashes restored in Conciseness and Keep going bullets - "Don't stop with a plan — execute it" exact Hermes phrasing Kept deviations, each documented in a comment above the constant: 1. "Absolute paths" bullet remains omitted — OpenClaw's sandbox container workdir differs from host workspace, and the core system prompt explicitly tells the model to prefer relative paths so exec and file tools work consistently (src/agents/system-prompt.ts:600). Injecting Hermes's absolute-path rule would directly contradict this and break exec. High-confidence, documented exception. 2. Tool ID substitution: read_file/search_files -> read or exec (mechanical rename; no grep/find/search_files tools exist in the OpenClaw catalog).

100yenadmin · 2026-04-16T05:02:17Z

Superseded by consolidated final-sprint PRs. This PR's changes are incorporated into:

feat(openai,agents): GPT-5.4 prompt discipline + personality bridge + injection scanning #67512 (OpenAI prompt stack + injection scanning)
feat(google): add Gemini execution guidance via prompt overlay [v3 final 2/3] #67513 (Google Gemini guidance)
feat(agents): task system parity — cancelled status, merge mode, activeForm, plan hydration [Phase 3.A] #67514 (Task-system parity bundle)

The v3 PRs were split across 6 separate PRs which made review harder and touched the same files redundantly. The final-sprint consolidation combines them into 3 clean, independently-mergeable PRs rebased on current main.

Copilot AI review requested due to automatic review settings April 14, 2026 06:04

openclaw-barnacle Bot added the agents Agent runtime and tooling label Apr 14, 2026

100yenadmin mentioned this pull request Apr 14, 2026

GPT 5.4 Enhancement v3: Hermes Parity Sprint — Prompt-Level Tool Enforcement & Execution Discipline #66345

Closed

4 tasks

openclaw-barnacle Bot added the size: XS label Apr 14, 2026

Copilot started reviewing on behalf of 100yenadmin April 14, 2026 06:04 View session

greptile-apps Bot reviewed Apr 14, 2026

View reviewed changes

Comment thread extensions/google/prompt-overlay.ts Outdated

Comment thread extensions/google/provider-registration.ts

Copilot AI reviewed Apr 14, 2026

View reviewed changes

Comment thread extensions/google/prompt-overlay.ts

Comment thread extensions/google/prompt-overlay.ts Outdated

Comment thread extensions/google/prompt-overlay.ts

Comment thread extensions/google/provider-registration.ts

Comment thread extensions/google/provider-registration.ts

chatgpt-codex-connector Bot reviewed Apr 14, 2026

View reviewed changes

Comment thread extensions/google/prompt-overlay.ts Outdated

Comment thread extensions/google/provider-registration.ts

openclaw-barnacle Bot added size: S and removed size: XS labels Apr 14, 2026

100yenadmin requested a review from Copilot April 14, 2026 07:46

Copilot started reviewing on behalf of 100yenadmin April 14, 2026 07:46 View session

Copilot AI reviewed Apr 14, 2026

View reviewed changes

100yenadmin mentioned this pull request Apr 14, 2026

feat(agents): add tool_enforcement as first-class provider prompt section [v3 3/6] #66373

Closed

100yenadmin closed this Apr 16, 2026

This was referenced Apr 16, 2026

feat(google): add Gemini execution guidance via prompt overlay [v3 final 2/3] #67518

Closed

[GPT 5.4 v3 — PR 6/6] Add Google Gemini execution guidance via prompt overlay #66352

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

feat(google): add Gemini execution guidance via prompt overlay [v3 6/6]#66379

feat(google): add Gemini execution guidance via prompt overlay [v3 6/6]#66379
100yenadmin wants to merge 3 commits into
openclaw:mainfrom
electricsheephq:gpt5-v3/google-gemini-execution-guidance

100yenadmin commented Apr 14, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot commented Apr 14, 2026

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

100yenadmin commented Apr 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

100yenadmin commented Apr 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

GPT 5.4 Enhancement v3 — PR 6/6

Problem

Changes

Architecture

Hermes Reference

Verification

Uh oh!

greptile-apps Bot commented Apr 14, 2026

Greptile Summary

Confidence Score: 4/5

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

100yenadmin commented Apr 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

100yenadmin commented Apr 14, 2026 •

edited

Loading