fix: support Amp handoff and Codex gpt-5.5

[Srini-B]

Summary

Fixes Amp /handoff and newer Deep Mode compatibility for both local Codex and local Claude routing.

Codex handoff responses can arrive through the internal buffered SSE path without a final message output item, which makes Amp throw Expected message output but none found. Claude smart-mode handoff forces a tool call, and Anthropic rejects or fails that shape when extended thinking is also present. Newer Codex models such as gpt-5.5 also gate access on the advertised Codex CLI version, so the connector now reports the installed Codex CLI identity instead of a stale hard-coded one.

What changed

• Reconstruct buffered Codex handoff output from streamed response events before returning non-streaming Responses JSON to Amp.
• Backfill missing Codex response.completed.response.output message content from output item/content/output text events.
• Compact synthesized Codex message content arrays so skipped content indexes cannot serialize as sparse-array null entries.
• Strip Codex request fields unsupported by the local Codex backend, including stream_options.
• Detect the installed codex --version, use it in the Codex CLI-style User-Agent, and stop sending the stale legacy Version header.
• Strip Anthropic thinking when tool_choice.type is tool or any, while preserving it for compatible modes such as auto.
• Add regression coverage for Codex handoff reconstruction, sparse content-index compaction, Codex version identity, and Anthropic forced-tool request preparation.
• Document the Codex deep-mode and Anthropic smart-mode compatibility behavior.

Validation

• bun run format
• bun run check — 68 pass, 0 fail
• Manually verified Amp /handoff works through:
  • LOCAL_CODEX deep-mode route
  • LOCAL_CLAUDE smart-mode route with create_handoff_context returning a tool_use
• Manually verified Amp Deep Mode works with openai:gpt-5.5 after advertising the installed Codex CLI version.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d2c93a9264

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".

[charles-chenzz]

unsupport stream option still appear, did it require new release or something? sorry for the ping @Srini-B

[Srini-B]

Yes, @nghyane needs to make a release, for now you can run it directly from my fork till its done @charles-chenzz

[charles-chenzz]

Yes, @nghyane needs to make a release, for now you can run it directly from my fork till its done @charles-chenzz

I will check how to run your fork, not familiar with bun ecosystem

[Srini-B]

clone my fork, switch to fix/codex-handoff-buffer-output branch, run bun i and then bun run dev.

[charles-chenzz]

clone my fork, switch to fix/codex-handoff-buffer-output branch, run bun i and then bun run dev.

@Srini-B if I don't have google acc oauth it seems like can't use handoff

[Srini-B]

I have mine linked to it, so I haven't seen this issue, but some of the workings are core to how the AMP CLI works. There is no way getting over it without degrading its usefulness unless we have all of the three accounts linked with. Don't you have a Google account which will use your free quota from antigravity? @charles-chenzz

[charles-chenzz]

I have mine linked to it, so I haven't seen this issue, but some of the workings are core to how the AMP CLI works. There is no way getting over it without degrading its usefulness unless we have all of the three accounts linked with. Don't you have a Google account which will use your free quota from antigravity? @charles-chenzz

I have google account that can link, but I confirmed that antigravity aggressively ban account that oauth to use on other places(my friends got ban for like 3 accounts). I only have one goog account which I don't want to take any risk, so I didn't link at the moments

[Srini-B]

That's a valid point. Depending on usage, the account gets flagged from what I have seen. Since what we are going to use here from the CLI is going to be the Google models and not the claude models, I think the risk is low. Another way to overcome this lack of Gemini models for the usage of handoff commands: you can create a new thread and ask it to refer to your previous thread.

[charles-chenzz]

@Srini-B sorry for the ping, amp officially use gpt5.5 as default deep mode model, does the branch in ur fork repo already support this? if not, any planning to support?

[Srini-B]

@Srini-B sorry for the ping, amp officially use gpt5.5 as default deep mode model, does the branch in ur fork repo already support this? if not, any planning to support?

Yes it does if you're on the latest commit on my fork

[Srini-B]

Things will not be the same if https://ampcode.com/news/neo is enabled for your account. It will use amp credits. you can go back to old way by using amp --take-me-back