app-server: JSON-RPC `-32600` is overloaded across "no rollout found" and "failed to load configuration" — please make the two discriminable

Summary

Over the codex app-server JSON-RPC transport, error code -32600 is returned for at least two completely unrelated failure conditions, and the only way to tell them apart is to substring-match the human-readable message. An external client that wants to react correctly (auto-recover vs. surface to the user) has to parse English strings, which is brittle.

The two meanings we hit:

1. Stale resume — recoverable. thread/resume (or a read on) a thread whose rollout no longer exists returns -32600 with a message like no rollout found / thread not found / not loaded. The correct client behavior is to fall back to thread/start (a fresh thread).
2. Malformed project config — user must fix. A bad project-local .codex/config.toml (e.g. an invalid type / unknown variant) surfaces as -32600 "failed to load configuration: …config.toml:LINE:COL: …". The correct client behavior is the opposite: do not retry/fallback; show the file path + parse error to the user.

Because both arrive as -32600, a client that auto-falls-back on -32600 will silently paper over a broken config file, and a client that surfaces -32600 to the user will nag them about recoverable stale-resume cases.

Environment

• Codex CLI 0.139.0 (Homebrew cask), macOS (Darwin arm64).
• Transport: codex app-server --listen stdio://, newline-delimited JSON-RPC 2.0.
• Third-party app-server client (not the TUI).

Observed error payloads

Case 1 (stale resume):

{"id":N,"error":{"code":-32600,"message":"thread/read failed: thread not loaded: <uuid>"}}

(also seen as no rollout found for thread id <uuid>)

Case 2 (bad config):

{"id":N,"error":{"code":-32600,"message":"failed to load configuration: /path/.codex/config.toml:228:1: invalid type: ..."}}

Request

Make the two discriminable without string parsing, e.g. any of:

• Distinct codes (a dedicated code for config-load failure, and/or for thread-not-found), or
• A structured error.data discriminator, e.g. { "data": { "kind": "no_rollout" | "config_load" | ... } }, or
• For the resume half specifically, the thread/resumeFailed { thread_id, reason } notification already proposed in #22064.

A machine-readable discriminator on error.data would be the lowest-risk fix and would cover both halves at once.

Related

• #22064 — codex exec resume <missing-uuid>: noisy "thread not found"; proposes thread/read→thread/start fallback and a thread/resumeFailed {thread_id, reason} notification (covers Case 1; cites -32600 literally).
• #16872 — app-server WebSocket: turns complete but rollout never materializes → sibling thread/resume fails "no rollout found" (Case 1, pure app-server).
• #19476 — failed to load configuration: config.toml:…: invalid type (Case 2 symptom; does not note the -32600 framing or the collision).
• PR #6938 (merged) — introduced richer v2 app-server error codes/events; the infrastructure exists but isn't applied to these two cases.

Note: the resume-half symptom (#22064/#16872) and the config-parse symptom (#19476) are each reported separately, but the overloading of -32600 across both — and the resulting inability to disambiguate from the payload — does not appear to be filed anywhere. That's the specific ask here.