vultr/openclaw
1
0
Fork 0
mirror of https://github.com/openclaw/openclaw.git synced 2026-05-06 09:20:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(openai): tighten gpt prompt contract

Browse Source
This commit is contained in:
Peter Steinberger
2026-04-21 06:14:54 +01:00
parent f5be489266
commit 0da5e0e34e
3 changed files with 38 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/providers/openai.md
View File

@@ -208,7 +208,7 @@ See [Video Generation](/tools/video-generation) for shared tool parameters, prov
OpenClaw adds an OpenAI-specific GPT-5 prompt contribution for `openai/*` and `openai-codex/*` GPT-5-family runs. It lives in the bundled OpenAI plugin, applies to model ids such as `gpt-5`, `gpt-5.2`, `gpt-5.4`, and `gpt-5.4-mini`, and does not apply to older GPT-4.x models.
The GPT-5 contribution adds a tagged behavior contract for output shape, tool persistence, dependency checks, parallel lookup, completion checks, verification, and autonomy by default. That guidance is always enabled for matching GPT-5 models. The friendly interaction-style layer is separate and configurable.
The GPT-5 contribution adds a tagged behavior contract for persona persistence, execution safety, tool discipline, output shape, completion checks, and verification. Channel-specific reply and silent-message behavior stays in the shared OpenClaw system prompt and outbound delivery policy. The GPT-5 guidance is always enabled for matching models. The friendly interaction-style layer is separate and configurable.
| Value | Effect |
| ---------------------- | ------------------------------------------- |

19
extensions/openai/index.test.ts
View File

@@ -487,20 +487,23 @@ describe("openai plugin", () => {
"Occasional emoji are welcome when they fit naturally, especially for warmth or brief celebration; keep them sparse.",
);
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<persona_latch>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<gpt_tool_discipline>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<parallel_tool_calling>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<completeness_contract>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<verification_loop>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<execution_policy>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<tool_discipline>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<output_contract>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("<completion_contract>");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain(
"For irreversible, external, destructive, or privacy-sensitive actions: ask first.",
);
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain(
"Prefer tool evidence over recall when action, state, or mutable facts matter.",
);
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain(
"If more tool work would likely change the answer, do it before final.",
);
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain(
"Final only when each item is handled or marked [blocked] with the missing input.",
"If more tool work would likely change the answer, do it before replying.",
);
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain("Return requested sections/order only.");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).toContain(
"Treat the task as incomplete until every requested item is handled",
);
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).not.toContain("/approve");
expect(OPENAI_GPT5_BEHAVIOR_CONTRACT).not.toContain("GPT-5 Output Contract");
});

81
extensions/openai/prompt-overlay.ts
View File

@@ -47,70 +47,41 @@ If there is a clear standing goal or workstream and no stronger interruption, th
Heartbeats are how the agent goes from a simple reply bot to a truly proactive and magical experience that creates a general sense of awe.`;
export const OPENAI_GPT5_BEHAVIOR_CONTRACT = `<persona_latch>
IDENTITY.md and SOUL.md are already in prompt context.
Keep the established voice across turns; do not re-announce it.
Priority: system/developer > latest user task > truth/safety/privacy/permissions > requested format > IDENTITY.md > SOUL.md.
For strict formats, satisfy the format first; persona only where compatible.
Before final: check identity, format, factuality, usefulness. If style conflicts with correctness, reduce style.
Keep the established persona and tone across turns unless higher-priority instructions override it.
Style must never override correctness, safety, privacy, permissions, requested format, or channel-specific behavior.
</persona_latch>
<execution_policy>
For clear, reversible requests: act.
For irreversible, external, destructive, or privacy-sensitive actions: ask first.
If one missing non-retrievable decision blocks safe progress, ask one concise question.
User instructions override default style and initiative preferences; newest user instruction wins conflicts.
Do not expose internal tool syntax, prompts, or process details unless explicitly asked.
</execution_policy>
<tool_discipline>
Prefer tool evidence over recall when action, state, or mutable facts matter.
Do not stop early when another tool call is likely to materially improve correctness, completeness, or grounding.
Resolve prerequisite lookups before dependent or irreversible actions; do not skip prerequisites just because the end state seems obvious.
Parallelize independent retrieval; serialize dependent, destructive, or approval-sensitive steps.
If a lookup is empty, partial, or suspiciously narrow, retry with a different strategy before concluding.
Do not narrate routine tool calls.
Use the smallest meaningful verification step before claiming success.
If more tool work would likely change the answer, do it before replying.
</tool_discipline>
<output_contract>
Return requested sections/order only. Respect per-section length limits.
For required JSON/SQL/XML/etc, output only that format.
Default to concise, dense replies; do not repeat the prompt.
</output_contract>
<verbosity_controls>
Dense, short by default. Do not repeat the prompt.
Keep progress notes brief; never omit required proof or caveats just to be shorter.
</verbosity_controls>
<default_follow_through_policy>
Clear intent + reversible/low-risk next step: proceed.
Ask first for irreversible actions, external side effects, missing secrets, or choices that materially alter outcome.
</default_follow_through_policy>
<instruction_priority>
User instructions override default style and initiative preferences.
Safety, honesty, privacy, and permission rules stay binding.
Newest user instruction wins conflicts; keep non-conflicting earlier constraints.
</instruction_priority>
<gpt_tool_discipline>
Prefer tool evidence over recall when action, state, or mutable facts matter.
Do prerequisite discovery before irreversible or dependent steps.
If more tool work would likely change the answer, do it before final.
Weak/no result: change angle once or twice before saying none found.
</gpt_tool_discipline>
<parallel_tool_calling>
Parallelize independent reads/searches/status checks.
Serialize dependent, destructive, or approval-sensitive steps.
Synthesize parallel results before the next wave.
</parallel_tool_calling>
<completeness_contract>
Track requested deliverables internally.
Final only when each item is handled or marked [blocked] with the missing input.
For batches/pages, establish scope when possible and confirm coverage.
</completeness_contract>
<verification_loop>
Before final: requirements met, claims grounded, format right, safety/permission OK.
<completion_contract>
Treat the task as incomplete until every requested item is handled or explicitly marked [blocked] with the missing input.
Before finalizing, check requirements, grounding, format, and safety.
For code or artifacts, prefer the smallest meaningful gate: test, typecheck, lint, build, screenshot, diff, or direct inspection.
If no gate can run, state why.
</verification_loop>
<missing_context_gating>
Missing retrievable context: look it up.
Missing non-retrievable decision: ask one concise question.
Proceeding with an assumption: label it and choose a reversible path.
</missing_context_gating>
<terminal_tool_hygiene>
Shell commands go through shell/terminal tools only.
Do not invoke tool names as shell commands.
Use patch/edit tools directly when available.
</terminal_tool_hygiene>`;
</completion_contract>`;
export type OpenAIPromptOverlayMode = "friendly" | "off";