feat: add per-delegate max_turns override by Yagi-Michael · Pull Request #8066 · block/goose

Yagi-Michael · 2026-03-23T05:49:11Z

Summary

Adds a max_turns parameter to the delegate tool, allowing individual delegations to override the max turns limit
Fixes priority order so delegated recipe settings.max_turns and session recipe are both respected
Full priority chain: per-delegate max_turns > delegated recipe settings > session recipe > env var > config > default

Context

The original max_turns implementation landed in subagent_tool.rs, but the summon refactor replaced that file with the unified summon extension (platform_extensions/summon.rs). The per-delegate override was never carried over, so it needs to be wired up in the new location.

Closes #6198

Changes

DelegateParams: new optional max_turns: Option<usize> field
JSON schema: exposes max_turns as integer with minimum: 1
Validation: rejects max_turns: 0
build_task_config: checks params.max_turns, then delegated recipe.settings.max_turns, then falls back to resolve_max_turns
resolve_max_turns: reordered to session recipe > env var (was env var > recipe)

Test plan

Deserialization tests for max_turns presence and absence
Validation tests rejecting zero and accepting positive values
Resolution tests verifying recipe overrides env var
Resolution tests verifying env var fallback when no recipe
Resolution tests verifying default fallback
Env var tests use #[serial] for thread safety

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0dfa76a437

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-03-23T05:54:46Z

crates/goose/src/agents/platform_extensions/summon.rs

+        let max_turns = params
+            .max_turns
+            .unwrap_or_else(|| self.resolve_max_turns(session));


Read delegated recipe max_turns before parent-session fallback

For delegate(source: ...) calls, build_delegate_recipe has already loaded the target recipe and passed it into this function, but the new max-turn logic still ignores that recipe argument and only consults session.recipe. In a normal chat session session.recipe is None, so a source recipe/subrecipe with its own settings.max_turns will silently fall back to the env/default limit instead of honoring the delegated recipe's explicit bound; if the parent session has a different recipe limit, that value wins instead.

Useful? React with 👍 / 👎.

Helpful feedback. Think this is real and not just fluff. Taking a look (using goose to help 🪿 )

Good catch — this was a real bug. resolve_max_turns only checked session.recipe (the parent session's recipe), so a delegated source recipe's settings.max_turns was silently ignored.

Fixed by inserting an .or_else() check on the delegated recipe.settings.max_turns between the per-delegate params.max_turns override and the resolve_max_turns(session) fallback. The priority chain is now:

params.max_turns (per-delegate override)

Delegated recipe's settings.max_turns

Session recipe / env var / config / default

Adds max_turns: Option<usize> to DelegateParams so each delegate() call can cap its own turn budget independently of the recipe-level setting. The original max_turns implementation landed in subagent_tool.rs, but the summon refactor replaced that file with the unified summon extension (platform_extensions/summon.rs). The per-delegate override was never carried over, so it needs to be wired up in the new location. Priority chain (highest to lowest): 1. delegate tool max_turns argument 2. delegated recipe settings.max_turns 3. session recipe settings.max_turns 4. GOOSE_SUBAGENT_MAX_TURNS env var / config 5. Default 25 Rejects max_turns: 0 since it would immediately stop the delegate. Relates to block#6198 Signed-off-by: Michael Yagi <[email protected]>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 03aa967d7a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-03-23T06:56:19Z

crates/goose/src/agents/platform_extensions/summon.rs

+        let max_turns = params
+            .max_turns
+            .or_else(|| {
+                recipe
+                    .settings


Bound delegated max_turns to SessionConfig's u32 range

build_task_config now forwards either params.max_turns or recipe.settings.max_turns as a raw usize, but run_subagent_task later narrows task_config.max_turns to u32 in subagent_handler.rs:175-179. That means delegated values above u32::MAX wrap (for example, 4294967296 becomes 0 and the first turn check in agent.rs stops the subagent immediately), and settings.max_turns: 0 from a delegated recipe now slips through even though validate_delegate_params rejects 0 for tool-call overrides. Since this feature is what starts honoring delegated values, it needs a 1..=u32::MAX guard before with_max_turns.

Useful? React with 👍 / 👎.

Fixed — added .clamp(1, u32::MAX as usize) to the resolved max_turns before passing to with_max_turns. This handles both issues: rejects 0 from delegated recipe settings, and prevents wrapping when narrowed to u32 in subagent_handler.rs:178.

- Collapse .or_else block to single line (cargo fmt) - Use struct initializer instead of field assignment for Session (clippy) - Clamp resolved max_turns to 1..=u32::MAX before passing to TaskConfig, preventing u32 wrapping in SessionConfig and rejecting 0 from recipes Signed-off-by: Michael Yagi <[email protected]>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 500d503356

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-03-23T08:05:56Z

crates/goose/src/agents/platform_extensions/summon.rs

+            .or_else(|| recipe.settings.as_ref().and_then(|s| s.max_turns))
+            .unwrap_or_else(|| self.resolve_max_turns(session))
+            .clamp(1, u32::MAX as usize);


Reject invalid max_turns instead of saturating it

The new .clamp(1, u32::MAX as usize) avoids the wraparound in subagent_handler.rs:175-179, but it still rewrites malformed limits into different budgets. If a delegate call, recipe, or GOOSE_SUBAGENT_MAX_TURNS is set to 0 or above u32::MAX, this now runs the subagent with 1 or 4_294_967_295 turns instead of surfacing the configuration error. Because max_turns is the runaway-safety bound, silently saturating typos here can turn a bad limit into an effectively unbounded delegate.

Useful? React with 👍 / 👎.

Okay codex, one more time :)

Yagi-Michael · 2026-03-23T08:06:03Z

🪿 please let me know how this looks or if would like any changes 🪿

Error on max_turns of 0 or above u32::MAX from recipe settings or env var, which bypass validate_delegate_params. Signed-off-by: Michael Yagi <[email protected]>

Signed-off-by: Douwe Osinga <[email protected]>

DOsinga

Good fix — addresses a real bug (delegated recipe max_turns was silently ignored) and a real edge case (u32 overflow). Nice job engaging with codex feedback and iterating. I cleaned up a few restatement comments and dropped two trivial deserialization tests (they only assert struct layout, not behaviour) in a follow-up commit.

Yagi-Michael · 2026-03-23T15:30:30Z

Woo, thanks! Thank you for the cleanup, will be better at catching those bloating going forward

Yagi-Michael · 2026-03-24T17:11:51Z

Good fix — addresses a real bug (delegated recipe max_turns was silently ignored) and a real edge case (u32 overflow). Nice job engaging with codex feedback and iterating. I cleaned up a few restatement comments and dropped two trivial deserialization tests (they only assert struct layout, not behaviour) in a follow-up commit.

Hi @DOsinga, just curious what the merge cadence is and roughly when can this might get merged in

Signed-off-by: Michael Yagi <[email protected]> Signed-off-by: Douwe Osinga <[email protected]> Co-authored-by: Douwe Osinga <[email protected]> Signed-off-by: Cameron Yick <[email protected]>

chatgpt-codex-connector bot reviewed Mar 23, 2026

View reviewed changes

Yagi-Michael force-pushed the feat/max-turns-per-delegate branch from 0dfa76a to 03aa967 Compare March 23, 2026 06:49

chatgpt-codex-connector bot reviewed Mar 23, 2026

View reviewed changes

Yagi-Michael mentioned this pull request Mar 23, 2026

feat: Add max_turns to recipe and subagent settings #6198

Closed

1 task

Yagi-Michael and others added 2 commits March 23, 2026 01:25

fix: reject invalid max_turns instead of silently clamping

b5a2271

Error on max_turns of 0 or above u32::MAX from recipe settings or env var, which bypass validate_delegate_params. Signed-off-by: Michael Yagi <[email protected]>

chore: remove restatement comments and trivial deserialization tests

19773fe

Signed-off-by: Douwe Osinga <[email protected]>

DOsinga approved these changes Mar 23, 2026

View reviewed changes

DOsinga added this pull request to the merge queue Mar 26, 2026

Merged via the queue into block:main with commit 838d99e Mar 26, 2026
21 checks passed

github-actions bot mentioned this pull request Apr 2, 2026

chore(release): release version 1.30.0 #8261

Open

Conversation

Yagi-Michael commented Mar 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Context

Changes

Test plan

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

Yagi-Michael Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

Yagi-Michael Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

Yagi-Michael Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

Yagi-Michael Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

Yagi-Michael commented Mar 23, 2026

Uh oh!

DOsinga left a comment

Choose a reason for hiding this comment

Uh oh!

Yagi-Michael commented Mar 23, 2026

Uh oh!

Yagi-Michael commented Mar 24, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Yagi-Michael commented Mar 23, 2026 •

edited

Loading