📋 Stack position: This is [Plan Mode 9/9] (OPTIONAL) — the GPT-5 prompt foundation, an independent track that can land before, alongside, or after the main 8-PR plan-mode stack.

• Main stack (sequential): [Plan Mode 1/9] Plan-state foundation is currently open; [Plan Mode 2/9] opens after 1/9 merges; etc.
• This PR (9/9): independent of the main stack — does NOT need to merge in order. Can be reviewed and merged whenever the GPT-5 prompt-overlay area is ready for changes.
• Why optional: this PR's content is GPT-5.4 prompt-layer churn (system prompt reorder, context-file injection scan, agent scope refinement). Maintainers may want to defer it until they've evaluated whether the recent main-branch GPT-5 prompt evolution (e.g., agents: split GPT-5 prompt and retry behavior #65597 split GPT-5 prompt and retry behavior, OpenAI: reduce repeated heartbeat alerts #65148 strengthen heartbeat overlay guidance, openai: add heartbeat guidance to GPT-5 overlay #65069 add heartbeat guidance to GPT-5 overlay) and this PR's overlay are coherent — that's a separate review concern from the rest of the plan-mode work.

Numbering history: this PR was originally [Plan Mode 1/9] (the foundation). Per maintainer-readability feedback, the numbering was reordered so [Plan Mode 1/9] is now the plan-state foundation (which actually starts the plan-mode feature) and the GPT-5 prompt foundation moved to 9/9 as an optional independent track. The original 9-part fork stack (#10-#18 on 100yenadmin/openclaw-1, all closed) used the old numbering.

Summary

This PR carves the GPT-5 and prompt-layer foundation out of the original plan-mode umbrella PR (#68939, closed; replaced by the 8-PR sequential decomposition).

It exists to land the model-facing groundwork: prompt overlay behavior, agent scope, context-file injection scanning, and system-prompt ordering. Keeping this work separate lets reviewers validate prompt semantics without also reviewing plan persistence, approval gates, UI, or channel behavior.

What This PR Includes

• OpenAI prompt overlay (extensions/openai/prompt-overlay.ts) — GPT-5 specific overlay configuration
• Agent scope (src/agents/agent-scope.ts) — reorganized scope semantics for the new prompt foundation
• Context-file injection scanning (src/agents/context-file-injection-scan.ts) — security-oriented scan for prompt-injection patterns in context files
• System-prompt ordering (src/agents/system-prompt.ts, src/agents/pi-embedded-runner/system-prompt.ts, src/agents/system-prompt-contribution.ts) — reordered to put GPT-5-specific guidance first
• GPT-5 boot reorder test (src/agents/system-prompt-gpt5-boot-reorder.test.ts)

Files In Scope

Primary review targets (11 files):

• extensions/openai/prompt-overlay.ts — overlay logic
• extensions/openai/index.test.ts — overlay tests
• src/agents/agent-scope.ts, src/agents/agent-scope.test.ts
• src/agents/context-file-injection-scan.ts, src/agents/context-file-injection-scan.test.ts
• src/agents/system-prompt.ts
• src/agents/system-prompt-contribution.ts
• src/agents/system-prompt-gpt5-boot-reorder.test.ts
• src/agents/pi-embedded-runner/system-prompt.ts
• src/auto-reply/reply/commands-system-prompt.ts

Reviewer Guide

1. Start with extensions/openai/prompt-overlay.ts — the core overlay change
2. Then src/agents/system-prompt.ts + src/agents/pi-embedded-runner/system-prompt.ts — reorder logic
3. Finally src/agents/context-file-injection-scan.ts — security additions

Validation:

• Run pnpm vitest run extensions/openai/index.test.ts src/agents/system-prompt-gpt5-boot-reorder.test.ts src/agents/context-file-injection-scan.test.ts
• Manual: spin up gateway with a GPT-5 model + verify the system prompt order matches expected sequence

What This PR Does NOT Include

• Plan persistence — see [Plan Mode 1/9] Plan-state foundation (open now)
• Approval gates / mutation gate / nudges — see [Plan Mode 2/9] Core backend MVP (next)
• Plan UI — see [Plan Mode 5/9] Web UI + i18n (later)
• Channel integration — see [Plan Mode 6/9] Text channels + Telegram (later)

Issue references

• Refs feat(openai,agents): GPT-5.4 prompt discipline + personality bridge + injection scanning #67512 (GPT-5.4 prompt discipline + personality bridge + injection scanning) — this PR carries that issue's prompt-foundation portion

Architecture references

• docs/plans/PLAN-MODE-ARCHITECTURE.md — see "GPT-5 Prompt Layer" section (added by this PR or earlier)

Test Status

• Targeted unit tests for prompt-overlay + injection-scan + boot-reorder — all passing on the original branch base
• ⚠️ NOTE: this PR is currently rebased on 1bd92975c2 (main from ~534 commits ago). It will need a focused rebase against current upstream/main before merge — the prompt-overlay area has had significant churn upstream (agents: split GPT-5 prompt and retry behavior #65597 GPT-5 prompt split, OpenAI: reduce repeated heartbeat alerts #65148 / openai: add heartbeat guidance to GPT-5 overlay #65069 heartbeat overlay strengthening, etc.) and the merge requires GPT-5.4 domain judgment about how this PR's overlay coexists with the recent main evolution. This is the "deep dive" planned after the 8-part stack lands.

Carry-forward commitments

• This PR's foundation is required by NOTHING in the main 8-PR stack (verified: each of Parts 1-8 in the new numbering does not touch extensions/openai/prompt-overlay.ts or the system-prompt.ts files this PR modifies). Hence it can land in any order relative to the main stack.

Status

DRAFT — pending the focused rebase + GPT-5 prompt evolution audit.