feat(openai): add act-don't-ask and tool retry directives for GPT-5 [v3 2/6] by 100yenadmin · Pull Request #66372 · openclaw/openclaw

100yenadmin · 2026-04-14T05:58:56Z

GPT 5.4 Enhancement v3 — PR 2/6

Tracking: #66345 | Issue: #66347
Priority: P0 — HIGH | Gap closure: ~25%

Problem

GPT 5.4 on OpenClaw exhibits two failure modes Hermes avoids:

  User: "Is port 8080 open?"         Tool returns empty result
           │                                   │
  ┌────────┴─────────┐               ┌────────┴─────────┐
  │                  │               │                  │
  ▼                  ▼               ▼                  ▼
Hermes             OpenClaw        Hermes             OpenClaw
  │                  │               │                  │
  │ "act on obvious  │ "prerequisite │ "retry with      │ (no retry
  │  defaults" +     │  lookup"      │  different       │  guidance)
  │  examples        │ (too vague)   │  strategy"       │
  ▼                  ▼               ▼                  ▼
Runs               Asks user:      Retries with       Reports:
`ss -tlnp |        "Which host     broader query      "No results
 grep 8080`        should I check?"                    found."
✅ ACTS             ❌ STALLS        ✅ PERSISTS         ❌ SURRENDERS

Changes

extensions/openai/prompt-overlay.ts — Enhanced OPENAI_GPT5_EXECUTION_BIAS with two new subsections:

Act, Don't Ask: Concrete examples showing when to act on obvious defaults vs when to ask
Tool Persistence: Retry-on-failure directive — diagnose and adjust instead of surrendering

Defense-in-Depth

  ┌──────────────────────────────────────────────┐
  │          Defense-in-Depth Stack               │
  │                                               │
  │  Layer 1: PROMPT (this PR)                    │
  │  ├─ Act-don't-ask → prevents stall           │
  │  ├─ Tool retry → prevents surrender           │
  │  │                                             │
  │  Layer 2: RUNTIME (already exists)             │
  │  ├─ Planning-only detection → catches misses   │
  │  ├─ Ack fast path → "ok do it" acceleration    │
  │  └─ Strict-agentic blocked exit                │
  └──────────────────────────────────────────────┘

Hermes Reference

agent/prompt_builder.py:221-229 — <act_dont_ask> block
agent/prompt_builder.py:199-202 — retry-with-different-strategy directive

Verification

"Is port 8080 open?" → runs ss/netstat, doesn't ask "where?"
"Find all TODO comments" → if first grep misses, retries with different pattern
"What packages are outdated?" → runs npm outdated, doesn't ask which project

GPT 5.4 exhibits two failure modes that Hermes avoids: 1. Clarification stalling: asks "open where?" instead of checking the local machine. Stalls on ambiguity that has an obvious default. 2. First-failure surrender: reports tool failure instead of retrying with a different strategy. Add two new subsections to OPENAI_GPT5_EXECUTION_BIAS: - "Act, Don't Ask" with concrete examples (port check, OS check, time) - "Tool Persistence" with retry-on-failure directive These complement the runtime planning-only retry guard in incomplete-turn.ts by preventing stalls at the prompt level, avoiding wasted turns. Part of GPT 5.4 Enhancement v3 sprint. Closes #66347. Tracking issue: #66345.

greptile-apps · 2026-04-14T06:00:07Z

Greptile Summary

Adds two subsections to OPENAI_GPT5_EXECUTION_BIAS in the OpenAI extension's prompt overlay: ### Act, Don't Ask (act on obvious defaults instead of asking for clarification) and ### Tool Persistence (retry with a different strategy instead of surrendering on empty results). The change is narrowly scoped to prompt text within the extension boundary and follows existing patterns in the file.

Confidence Score: 5/5

Safe to merge — purely additive prompt text change with no logic or schema impact.
Only finding is a P2 style suggestion to restate the existing "real blocker" escape hatch inside the new subsection for internal consistency. No correctness, data-integrity, or reliability concerns.
No files require special attention.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: extensions/openai/prompt-overlay.ts
Line: 81-84

Comment:
**Missing blocker escape hatch in Tool Persistence**

The parent block already says "keep calling tools until the task is done **or you hit a real blocker**", but the new `### Tool Persistence` subsection omits that exit condition. A model reading this subsection for retry guidance could interpret "keep calling tools until (1) complete AND (2) verified" as requiring indefinite retries even when the failure is permanent (e.g., `EACCES`, missing resource, auth error). Worth adding the escape hatch explicitly to stay internally consistent.

```suggestion
### Tool Persistence
If a tool returns empty or partial results, retry with a different query or strategy before giving up.
Keep calling tools until: (1) the task is complete, AND (2) you have verified the result — or you hit a real blocker with no viable alternative.
Do not abandon a viable approach after a single failure — diagnose why it failed and adjust.
```

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

_{Reviews (1): Last reviewed commit: "feat(openai): add act-don't-ask and tool..." | Re-trigger Greptile}

Copilot

Pull request overview

This PR enhances the GPT-5 execution-bias prompt overlay for the OpenAI extension to reduce “clarification stalling” and “first-failure surrender” by adding more explicit agentic directives.

Changes:

Added an “Act, Don’t Ask” subsection with concrete examples to encourage acting on obvious defaults.
Added a “Tool Persistence” subsection instructing retries with alternate strategies when tools return empty/partial results.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2a497f9b3b

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".

- Replace em dashes with periods (per GPT-5 output contract) - Add destructive-action escape hatch to act-don't-ask rule - Remove mandatory verification from tool persistence (conflicts with execution bias principle of acting first)

Copilot

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

Copilot

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

…tence Parity audit: previous iteration drifted from Hermes based on bot reviewer suggestions. Restore exact Hermes wording for both subsections. Act, Don't Ask (Hermes prompt_builder.py lines 221-229): - Restore single quotes in examples (Hermes style) - Restore "don't use user profile" wording - Remove destructive-action escape hatch (not in Hermes) Tool Persistence (Hermes prompt_builder.py lines 198-204): - Restore all four Hermes bullets including "improve correctness" and "do not stop early" - Restore mandatory verification: "(1) task complete AND (2) verified" - Remove "diagnose why it failed and adjust" rewording

Copilot

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

Copilot · 2026-04-14T11:15:09Z

+### Tool Persistence
+- Use tools whenever they improve correctness, completeness, or grounding.
+- Do not stop early when another tool call would materially improve the result.
+- If a tool returns empty or partial results, retry with a different query or strategy before giving up.
+- Keep calling tools until: (1) the task is complete, AND (2) you have verified the result.`;


The PR description’s proposed "Tool Persistence" block includes an additional directive ("Do not abandon a viable approach after a single failure. Diagnose why it failed and adjust."), but that sentence isn’t present in the updated OPENAI_GPT5_EXECUTION_BIAS string. Please either add the missing directive here or update the PR description so it matches what actually ships in this PR.

100yenadmin · 2026-04-16T05:02:07Z

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 05:58

openclaw-barnacle Bot added extensions: openai size: XS labels Apr 14, 2026

Copilot started reviewing on behalf of 100yenadmin April 14, 2026 05:59 View session

greptile-apps Bot reviewed Apr 14, 2026

View reviewed changes

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

Copilot AI reviewed Apr 14, 2026

View reviewed changes

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

chatgpt-codex-connector Bot reviewed Apr 14, 2026

View reviewed changes

Comment thread extensions/openai/prompt-overlay.ts

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

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

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

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

Copilot AI reviewed Apr 14, 2026

View reviewed changes

100yenadmin requested a review from Copilot April 14, 2026 07:47

Copilot started reviewing on behalf of 100yenadmin April 14, 2026 07:47 View session

Copilot AI reviewed Apr 14, 2026

View reviewed changes

100yenadmin requested a review from Copilot April 14, 2026 11:13

Copilot AI reviewed Apr 14, 2026

View reviewed changes

100yenadmin mentioned this pull request Apr 16, 2026

feat(openai,agents): GPT-5.4 prompt discipline + personality bridge + injection scanning #67512

Closed

7 tasks

100yenadmin closed this Apr 16, 2026

100yenadmin mentioned this pull request Apr 22, 2026

[GPT 5.4 v3 — PR 2/6] Strengthen execution bias: act-don't-ask + tool retry directives #66347

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

feat(openai): add act-don't-ask and tool retry directives for GPT-5 [v3 2/6]#66372

feat(openai): add act-don't-ask and tool retry directives for GPT-5 [v3 2/6]#66372
100yenadmin wants to merge 3 commits into
openclaw:mainfrom
electricsheephq:gpt5-v3/act-dont-ask-tool-retry

100yenadmin commented Apr 14, 2026

Uh oh!

greptile-apps Bot commented Apr 14, 2026

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 14, 2026

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

GPT 5.4 Enhancement v3 — PR 2/6

Problem

Changes

Defense-in-Depth

Hermes Reference

Verification

Uh oh!

greptile-apps Bot commented Apr 14, 2026

Greptile Summary

Confidence Score: 5/5

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

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!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

100yenadmin commented Apr 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants