Since moving from @zed-industries/codex-acp (0.16.x) to @agentclientprotocol/codex-acp (1.1.x), the model's reasoning shows up as terse summary headlines (e.g. "Planning CI polling loop", "Exploring gh pr checks options") instead of the fuller reasoning prose we used to get. I think this is coming from a hardcoded reasoning-summary setting rather than a model change — same model, same account, and the only thing that changed on our side was the adapter.
Reading the source, sendPrompt in src/CodexAcpClient.ts sets the summary mode on every runTurn:
return await this.codexClient.runTurn({
...
summary: disableSummary ? "none" : "auto",
effort: effort,
...
});disableSummary is derived in src/CodexAcpServer.ts as account.type === "apiKey" || modelLacksReasoning, so for a ChatGPT-auth reasoning model it's always false → summary: "auto" on every turn. There doesn't seem to be any way to request "detailed" (or raw reasoning), and CODEX_CONFIG doesn't reach this value.
For comparison, the old @zed-industries/codex-acp (Rust) never set a summary mode — it inherited Codex's own config defaults (Config::load_with_cli_overrides_and_harness_overrides, so -c model_reasoning_summary=... passthrough worked) and, on the emit side, forwarded both reasoning summary deltas and ReasoningRawContentDelta (the raw reasoning text) into agent_thought. So downstream ACP clients received fuller reasoning content.
Codex itself supports the range — ReasoningSummary = "auto" | "concise" | "detailed" | "none" — so this is purely about the adapter not exposing it.
Ask: make the reasoning-summary mode configurable rather than hardcoding "auto" — e.g. honor a Codex config passthrough (tie-in with #280 / #215), an env var, or an ACP _meta option, and/or allow requesting "detailed". That would let clients recover the fuller reasoning content the previous adapter surfaced.
Environment: @agentclientprotocol/codex-acp@1.1.2, ChatGPT-auth reasoning model, via an ACP harness observer feed.
Related: #280, #215
Since moving from
@zed-industries/codex-acp(0.16.x) to@agentclientprotocol/codex-acp(1.1.x), the model's reasoning shows up as terse summary headlines (e.g. "Planning CI polling loop", "Exploring gh pr checks options") instead of the fuller reasoning prose we used to get. I think this is coming from a hardcoded reasoning-summary setting rather than a model change — same model, same account, and the only thing that changed on our side was the adapter.Reading the source,
sendPromptinsrc/CodexAcpClient.tssets the summary mode on everyrunTurn:disableSummaryis derived insrc/CodexAcpServer.tsasaccount.type === "apiKey" || modelLacksReasoning, so for a ChatGPT-auth reasoning model it's alwaysfalse→summary: "auto"on every turn. There doesn't seem to be any way to request"detailed"(or raw reasoning), andCODEX_CONFIGdoesn't reach this value.For comparison, the old
@zed-industries/codex-acp(Rust) never set a summary mode — it inherited Codex's own config defaults (Config::load_with_cli_overrides_and_harness_overrides, so-c model_reasoning_summary=...passthrough worked) and, on the emit side, forwarded both reasoning summary deltas andReasoningRawContentDelta(the raw reasoning text) intoagent_thought. So downstream ACP clients received fuller reasoning content.Codex itself supports the range —
ReasoningSummary = "auto" | "concise" | "detailed" | "none"— so this is purely about the adapter not exposing it.Ask: make the reasoning-summary mode configurable rather than hardcoding
"auto"— e.g. honor a Codex config passthrough (tie-in with #280 / #215), an env var, or an ACP_metaoption, and/or allow requesting"detailed". That would let clients recover the fuller reasoning content the previous adapter surfaced.Environment:
@agentclientprotocol/codex-acp@1.1.2, ChatGPT-auth reasoning model, via an ACP harness observer feed.Related: #280, #215