V2 cross-provider sub-agent loses NEW_TASK body in encrypted_content

[webmastertorch]

Summary

With multiAgentMode: "v2", a Codex parent using the native gpt-5.6-sol model can successfully spawn a child with the hidden model override set to xai/grok-4.5, but the routed child does not receive the task body.

The child turn_context confirms that the model override was applied. However, its NEW_TASK message contains an empty plaintext payload followed by a Fernet-looking encrypted_content block. The routed child then reports that no concrete task was provided, or infers an unrelated task from the surrounding system context.

Environment

• OpenCodex: 2.7.7
• Codex Desktop / CLI: 0.144.0-alpha.4
• OS: Windows
• Proxy mode: loopback, openai_base_url = "http://127.0.0.1:10100/v1"
• Parent model: gpt-5.6-sol
• Child model: xai/grok-4.5
• xAI adapter: openai-chat
• multiAgentMode: v2
• fork_turns: "none"

Reproduction

1. Configure OpenCodex V2 with both native OpenAI and routed xAI models visible.
2. Start a new Codex Desktop session on gpt-5.6-sol.
3. Call spawn_agent with a self-contained task, fork_turns: "none", model: "xai/grok-4.5", and a valid reasoning effort.
4. Inspect the child session and its response.

Example child input shape recorded by Codex:

Message Type: NEW_TASK
Task name: /root/backend_scan
Sender: /root
Payload:
<encrypted_content containing a gAAAA... token>

The child model override is applied correctly, but the child answers along the lines of:

I don't see a concrete task yet—only environment and session setup.

The same result occurred across multiple independently spawned routed children.

Expected behavior

The cross-provider child receives the self-contained task text and executes it.

Actual behavior

The routed child receives the task only as backend ciphertext and cannot read it. No plaintext task reaches the model.

Relevant implementation observations

• src/responses/parser.ts explicitly treats encrypted_content as opaque to routed models and substitutes [encrypted content omitted] for tool-result content.
• src/server/responses.ts has sanitizeEncryptedContentInPlace, but it preserves Fernet-looking backend ciphertext.
• The routed-parent compatibility rewrite around responses.ts lines 456-463 runs only for native-bound requests (!requestedModelId.includes("/")).

This appears to leave the native-parent -> routed-child V2 path without a readable task body. The native backend minted the ciphertext, while the routed provider cannot decrypt it.

Questions

1. Is native-parent -> routed-child model override expected to work on the V2 collaboration surface?
2. If yes, where should OpenCodex preserve or recover the plaintext task before forwarding the child request?
3. If this cannot be fixed at the proxy layer because the task is already backend-encrypted, should V2 guidance avoid fork_turns: "none" for cross-provider children or document a required partial-fork workaround?
4. Is V1 currently the recommended mode for reliable heterogeneous-provider sub-agents?

I can provide sanitized session fragments or test a proposed patch if useful.

[lidge-jun]

already issued by bug in upstream
will be patched soon upstream and here both
im keep pushing it!

[Wibias]

Shouldnt this be fixed by #94 @webmastertorch or am I confusing something here? Just woke up though, lol.

[Wibias]

@webmastertorch Can u test v2.7.8 please

[lidge-jun]

@Wibias spawning errors in every routers fuck oai lol

[Wibias]

yea just encountered something similar by using fable before and changing into gpt-sol as fable became rate-limited.. On it.. @lidge-jun lol

Edit:

Existing history produced under Claude Fable 5
↓
Switch thread to GPT-5.6 Sol
↓
Sol has a smaller context window
950,000 → 353,400
↓
Codex decides the oversized history must be compacted first
↓
Codex creates a compaction context using the previous model
anthropic/claude-fable-5
↓
Fable quota is exhausted
↓
Anthropic returns 429
↓
Pre-sampling compaction fails
↓
The actual Sol request is never sent

edit-2: welp, thats actually an codex problem itself, yikes.

[Wibias]

Spinned up an issue on codex github.. Lets see.. openai/codex#32453

[lidge-jun]

@Wibias use v1 at gui everything is stable

[Wibias]

@lidge-jun Yea, i guess for now.

[lidge-jun]

Status update after a fresh root-cause pass on current dev:

• Still reproducible. A native v2 parent spawning a routed child delivers NEW_TASK with empty plaintext + a pure Fernet encrypted_content block (rewritten: 0, parsedContent: "Message Type: NEW_TASK\nPayload:\n"). The model override applies; the task body is lost.
• Root cause is client-side, not the proxy. The Codex client encrypts the spawn message for the native backend; the proxy cannot decrypt Fernet ciphertext for a routed child. The earlier plaintext-parked-in-encrypted-slot case (closed PR fix: preserve routed sub-agent task payloads by sanitizing before parse #94's patch, landed on dev) is a different, already-handled shape — regression-pinned in tests/multi-agent-compat.test.ts.
• Docs are now truthful (dev commit b6281b7): README ×3 and the sub-agent-surface guide ×3 no longer claim cross-model v2 spawns unconditionally stick; they name this limitation and point at v1 as the reliable heterogeneous-provider surface.
• Next step: an upstream openai/codex issue asking that SpawnAgentArgs.message be retained as plaintext (or duplicated beside the ciphertext) when the child's resolved model isn't served by the encrypting backend. Draft is ready and will be filed separately.

Note for readers of the earlier comments: the compaction/429 reports mixed in here are a separate issue (upstream #32453-family), not this encryption path. Keeping this open to track the upstream fix.