I'm on it! Simon Rosenberg can continue tracking my progress here.

Summary Task: Investigate whether

effort

is configurable for ACP agents (Claude Code and Codex harnesses) in

software-agent-sdk

, and file an issue if so. Investigation result

effort

is not exposed as a first-class parameter on

ACPAgentSettings

for any ACP harness: • Codex (

@zed-industries/codex-acp

): partially configurable, but only by encoding it into

acp_model

as

<model>/<effort>

(e.g.

gpt-5.5/high

). The curated

_CODEX_MODELS

list enumerates all 8 model×effort combos as separate model ids. • Claude Code (

@agentclientprotocol/claude-agent-acp

): not exposed at all.

LLM.extended_thinking_budget

exists for direct-LLM agents but is not wired to the ACP subprocess. • Gemini CLI (bonus): same as Claude Code — no effort field. Action taken Filed feature-request issue OpenHands/software-agent-sdk#3699 proposing a first-class

acp_effort: Literal["low","medium","high","xhigh"] | None

field on

ACPAgentSettings

, with per-provider translation (Codex → slash suffix on the model id; Claude Code →

MAX_THINKING_TOKENS

env var; Gemini/custom → no-op). Labels applied:

enhancement

,

acp

,

sdk

. The issue includes an AI-disclosure note that it was opened by OpenHands on the user's behalf. Completeness checklist • Investigated Codex harness effort handling — found slash-encoded model ids in

_CODEX_MODELS

. • Investigated Claude Code harness effort handling — confirmed no field exists and

extended_thinking_budget

isn't propagated. • Since effort is (partially) configurable for Codex, filed an issue as instructed — proposing a uniform first-class API. • No code changes were made, so nothing to commit or push. Conciseness No code was modified; the only artifact produced is the GitHub issue. Nothing extraneous to revert.

Summary I investigated whether reasoning effort is configurable across the two agent paths in

software-agent-sdk

and filed two feature‑request issues. Investigation results 1. ACP agents (Claude Code & Codex harnesses) — not a first‑class parameter: • Codex (

@zed-industries/codex-acp

): effort is only selectable by jamming it into the

acp_model

string as

<model>/<effort>

(e.g.

gpt-5.5/high

). The curated

_CODEX_MODELS

list in

openhands-sdk/openhands/sdk/settings/acp_providers.py

enumerates the 8 model×effort combos as if they were distinct models. • Claude Code (

@agentclientprotocol/claude-agent-acp

): no effort/thinking‑budget field at all. The non‑ACP

LLM.extended_thinking_budget

does not propagate to the ACP subprocess; the only workaround is smuggling

MAX_THINKING_TOKENS

via the secrets channel. 2. Regular OpenHands agents on GPT / Claude LLMs — configurable, but inconsistently: • OpenAI GPT‑5 / o‑series →

LLM.reasoning_effort: Literal["low","medium","high","xhigh","none"]

(default

"high"

) • Anthropic Claude (Sonnet 4.5/4.6, Haiku 4.5, Fable 5) →

LLM.extended_thinking_budget: int

token budget (default

200_000

) • Two different field names + units for the same concept; defaults are silently asymmetric across providers. Checklist of user requests • Investigate effort configurability for ACP harnesses (Claude Code, Codex) → done, findings above. • If configurable, file an SDK issue → filed OpenHands/software-agent-sdk#3699 proposing a first‑class

ACPAgentSettings.acp_effort

field, with per‑provider translation (Codex →

<model>/<effort>

composition; Claude Code →

MAX_THINKING_TOKENS

env mapping). Labels:

enhancement

,

acp

,

sdk

. AI disclosure included. • Investigate effort configurability for regular agents on GPT / Claude → done, findings above. • If configurable, file another SDK issue → filed OpenHands/software-agent-sdk#3701 proposing a single provider‑agnostic

LLM.effort

knob that the SDK translates to

reasoning_effort

(OpenAI) or

extended_thinking_budget

(Anthropic via a

low/medium/high/xhigh

→ token‑budget mapping), with a validation rule against double‑setting. Cross‑linked to #3699 so the ACP and direct‑LLM surfaces ship aligned. Labels:

enhancement

,

llm

,

sdk

. AI disclosure included. Code changes None. Both tasks were investigation + issue filing only — no commits, no PRs, no local edits to revert. The only artifacts written were two ephemeral issue‑body files under

/tmp/

, which are not in version control. Let me know if you'd like me to start implementation PRs for either #3699 or #3701.

so ideally it'd just magically show up in the chat box

or maybe when you connect an anthropic account you get a bunch of different effort profiles auto-created?