External/buzz

Fork 0

mirror of https://github.com/block/buzz.git synced 2026-08-18 06:50:31 +02:00

Files

T

History

00c6a92c19 feat(agents): make thinking effort configurable and portable in exports

Codex thinking effort was unreachable from Buzz, and neither the agent nor the
team JSON export carried effort or any environment configuration. Fix both, and
fix the pre-existing bug that made the live-ACP effort tier dead code.

Root cause of "there is no way to set Codex thinking effort": `reader.rs`
matched session config options by `category == "effort"`, but no real adapter
publishes that category. `codex-acp` 1.1.9 and `claude-agent-acp` 0.58.1 both
emit `category == "thought_level"` (config ids `reasoning_effort` and `effort`
respectively). The `AcpConfigOption` tier therefore never fired for any runtime
and every agent fell through to `RespawnWithEnvVar` — or to `ReadOnly` for
Codex, which had no thinking env var at all. Match `thought_level` and read the
adapter-advertised `config_id` instead of hardcoding `"effort"`, the way
`find_model_config_id` already does for the model field.

Codex effort is settable via environment after all: the `codex-acp` adapter
Buzz spawns ignores CLI args entirely and reads its configuration from the
`CODEX_CONFIG` JSON env var. Probing the real adapter over ACP,
`-c model_reasoning_effort=low` left `reasoning_effort` at `high`, while
`CODEX_CONFIG={"model_reasoning_effort":"low"}` took effect. Declare that
target in the runtime catalog as `thinking_config_json_env_var` +
`thinking_config_json_key` so no per-runtime branch is needed in the UI or the
snapshot code, and `build_codex_config_env` already deep-merges it into the
generated network overlay so the value reaches the child process.

Exports now carry a first-class scalar `thinking_effort` plus an allowlisted
`portable_env`. `CODEX_CONFIG` itself is deliberately not allowlisted: it can
carry `model_providers.*` auth headers, so only the extracted effort scalar is
projected. Import re-applies the allowlist and the normal env validation, so a
hand-edited snapshot cannot inject arbitrary environment, and the first-class
effort field wins when it conflicts with `portable_env`. New fields are
`#[serde(default, skip_serializing_if)]`, so v1 snapshots still decode and
clean records emit no new keys.

Several files were at the desktop file-size ratchet, so the new code is split
into sibling modules rather than growing capped files: the runtime table, the
catalog entry type (Rust and TS), the raw catalog bridge, and three test
fixture/test files.

Co-authored-by: Hardworking Honey <c5c455215c2506cb8ba776518cec804af62d3a0526e32d496a22072e395042b9@buzz.block.builderlab.xyz>
Co-authored-by: Atish Patel <atish@squareup.com>
Signed-off-by: Atish Patel <atish@squareup.com>

2026-08-06 10:12:10 -05:00