🤖 feat: default workflow agents to exec#3611
Conversation
Default workflow child agents now use exec unless a workflow step explicitly selects another agent. Deep Research pins its discovery stages to explore and uses exec with fable/high for verification and synthesis. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `4.90`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=14.90 -->
|
@codex review |
|
Codex Review: Didn't find any major issues. 🚀 Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Remove hardcoded verifier and synthesis model/thinking overrides so Deep Research exec steps inherit the configured exec sub-agent defaults. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `9.38`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=19.38 -->
|
@codex review Updated Deep Research to remove the hardcoded fable/high model and thinking overrides so exec verification/synthesis steps inherit the configured exec sub-agent defaults. |
|
Codex Review: Didn't find any major issues. What shall we delve into next? Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Dedupe the identical inline argument object declared by WorkflowTaskServiceLike.create and createMany into a named WorkflowTaskCreateArgs interface so the agent-task creation shape has a single source of truth. PR #3611 just added onRefusal to both copies, demonstrating the drift risk. Behavior-preserving: pure type-level refactor, no runtime change.
1 participant
Summary
Default workflow-owned child agents to
execand let workflow authors opt individual steps intoexplore, model aliases, and thinking levels. Deep Research now keeps read-only discovery onexplorewhile running verification/synthesis through exec agents that inherit the configured exec sub-agent defaults.Background
Workflow agents previously defaulted to
explore, which made workflows safe for read-only orchestration but too restrictive for stages that need execution-oriented behavior. Deep Research in particular benefits from read-only explore agents for source discovery and exec agents for adversarial verification.Implementation
DEFAULT_WORKFLOW_AGENT_ID = "exec"used by CLI, ORPC, and AI service workflow adapters.agent()options somodelandthinkingare normalized into task-levelmodelStringandthinkingLeveloverrides before launch/replay hashing when workflow authors explicitly set them.explorefor scoping/search/fetch andexecfor verification and synthesis without hardcoded model/thinking overrides.Validation
bun test src/node/services/workflows/WorkflowTaskServiceAdapter.test.ts src/node/services/workflows/WorkflowRunner.test.tsbun scripts/gen_builtin_skills.ts checkmake fmt-checkmake typecheckMUX_ESLINT_CONCURRENCY=1 make lintMUX_ESLINT_CONCURRENCY=1 make static-checkRisks
Moderate workflow behavior change: workflows that relied on omitted agent options now run
execby default. Read-only stages should explicitly setagentId: "explore"; Deep Research has been updated accordingly.Generated with
mux• Model:openai:gpt-5.5• Thinking:xhigh• Cost:$19.38