Add GPT-5 specific system prompt template

[enyst]

This PR adds a new system_prompt_gpt_5.j2 template tailored for GPT-5 usage within the OpenHands Software Agent SDK.

Key points:

• Reuses the existing system_prompt.j2 base to keep shared behavior and policies centralized.
• Adapts relevant concepts from OpenAI's Codex CLI prompt, focusing on planning, responsiveness, validation, and presentation of results.
• Integrates the existing task_tracker tool as the primary mechanism for long-horizon planning and progress tracking (replacing the update_plan concept from the original prompt).

This template is intended as a starting point for GPT-5-specific behavior and can be iterated on as we gain experience running GPT-5 in this environment.

Co-authored-by: openhands [email protected]

@enyst can click here to continue refining the PR

---

Agent Server images for this PR

• GHCR package: https://github.com/OpenHands/agent-sdk/pkgs/container/agent-server

Variants & Base Images

Variant	Architectures	Base Image	Docs / Tags
java	amd64, arm64	eclipse-temurin:17-jdk	Link
python	amd64, arm64	nikolaik/python-nodejs:python3.12-nodejs22	Link
golang	amd64, arm64	golang:1.21-bookworm	Link

Pull (multi-arch manifest)

# Each variant is a multi-arch manifest supporting both amd64 and arm64
docker pull ghcr.io/openhands/agent-server:6e1e2c2-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-6e1e2c2-python \
  ghcr.io/openhands/agent-server:6e1e2c2-python

All tags pushed for this build

ghcr.io/openhands/agent-server:6e1e2c2-golang-amd64
ghcr.io/openhands/agent-server:6e1e2c2-golang_tag_1.21-bookworm-amd64
ghcr.io/openhands/agent-server:6e1e2c2-golang-arm64
ghcr.io/openhands/agent-server:6e1e2c2-golang_tag_1.21-bookworm-arm64
ghcr.io/openhands/agent-server:6e1e2c2-java-amd64
ghcr.io/openhands/agent-server:6e1e2c2-eclipse-temurin_tag_17-jdk-amd64
ghcr.io/openhands/agent-server:6e1e2c2-java-arm64
ghcr.io/openhands/agent-server:6e1e2c2-eclipse-temurin_tag_17-jdk-arm64
ghcr.io/openhands/agent-server:6e1e2c2-python-amd64
ghcr.io/openhands/agent-server:6e1e2c2-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-amd64
ghcr.io/openhands/agent-server:6e1e2c2-python-arm64
ghcr.io/openhands/agent-server:6e1e2c2-nikolaik_s_python-nodejs_tag_python3.12-nodejs22-arm64
ghcr.io/openhands/agent-server:6e1e2c2-golang
ghcr.io/openhands/agent-server:6e1e2c2-java
ghcr.io/openhands/agent-server:6e1e2c2-python

About Multi-Architecture Support

• Each variant tag (e.g., 6e1e2c2-python) is a multi-arch manifest supporting both amd64 and arm64
• Docker automatically pulls the correct architecture for your platform
• Individual architecture tags (e.g., 6e1e2c2-python-amd64) are also available if needed

[enyst]

What if we delete this one 😅

I don't know if it's necessary to forbid it? GPT-5 does some of these, but it's rare and, personally, I found it welcome 🤷

[enyst]

This makes GPT-5 talk

[enyst]

These 2 paragraphs are a bit disputable. They're drafted by OpenAI for codex-cli, to shape the behavior of the LLM towards the user. Personally I like the style of GPT-5 right now, on final answers; and, this guidance is a tad annoying when displaying answers on GitHub, because they will be unformatted. 🤔

[enyst]

Suggested change

	The user is working on the same computer as you, and has access to your work. As such there's no need to show the full contents of large files you have already written unless the user explicitly asks for them. Similarly, if you've created or modified files using `apply_patch`, there's no need to tell users to "save the file" or "copy the code into a file"—just reference the file path.
	The user has access to your work. As such there's no need to show the full contents of large files you have already written unless the user explicitly asks for them. Similarly, if you've created or modified files using `apply_patch`, there's no need to tell users to "save the file" or "copy the code into a file"—just reference the file path.

😅 Ditto, drafted specifically for use in cli

[enyst]

This needs dogfooding (well most of this does)

• GPT-5 seems to me good with this (though sometimes it skips tests)
• GPT-5.1, however, is much more inclined to imagine it's in a restricted environment, and/or that it cannot use the network, in general gives up early. I feel like maybe declaring tests as optional will lead to it not running tests too often.

[enyst]

Suggested change

	When referencing files in your response, follow these rules:
	- Use inline code to make file paths clickable.
	- Each reference should have a stand-alone path, even if it's the same file.
	- Accepted: absolute, workspace-relative, `a/` or `b/` diff prefixes, or bare filename/suffix.
	- Line/column (1-based, optional): `:line[:column]` or `#Lline[Ccolumn]` (column defaults to 1).
	- Do not use URIs like `file://`, `vscode://`, or `https://`.
	- Do not provide ranges of lines.
	- Examples: `src/app.ts`, `src/app.ts:42`, `b/server/index.js#L10`, `C:\repo\project\main.rs:12:5`.

[enyst]

It's very curious how much text goes into the codex-cli system prompt referring to styling and formatting. Maybe that's not a bad idea... Clearly OpenAI doesn't use the same system prompt for codex web and codex-cli, at least in part.

[enyst]

😅 IIRC our default prompt has some restraining wording too. And in practice, with it, it seems to me that GPT-5 has mostly this behavior described here: does not usually fix unrelated bugs, does check "around" the issue, and notifies if it sees something suspicious. I can think of exceptions when it fixes stuff.

[enyst]

This wording is twice in this prompt 🤔