feat(coding-agent/modes): upgraded the workflowz keyword into a delegate-first DAG execution mode#2671
feat(coding-agent/modes): upgraded the workflowz keyword into a delegate-first DAG execution mode#2671metaphorics wants to merge 1 commit into
Conversation
…ate-first DAG execution mode - Rewrote the `workflowz` hidden notice into a delegate-first DAG execution contract: decompose the task and fan the work over subagents as an acyclic graph (certainty-before-code, skill survey, plan-agent gate, delegate-as-a-DAG via the eval pipeline/parallel runtime with handle piping, scenario contract, evidence-bound manual QA, TDD, reviewer gate). It references only bundled agents and does NOT change thinking effort. - No new magic-keyword names: the keyword remains `workflowz` only (detection, highlighter, and fast-path unchanged in surface). - Updated the magicKeywords.workflow setting description and the detection + notice-contract regression test.
There was a problem hiding this comment.
Pull request overview
Note
Copilot was unable to run its full agentic suite in this review.
Updates the workflowz keyword’s hidden system notice to a more explicit delegate-first DAG execution contract and aligns tests/docs accordingly.
Changes:
- Rewrote
WORKFLOW_NOTICEcontent to a DAG/discipline contract (plan gate, scenario contract, evidence-bound QA, reviewer gate, etc.) and documented that it does not change thinking effort. - Updated workflow notice tests to assert the new contract text and agent references.
- Refreshed related inline documentation / settings schema description.
Reviewed changes
Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| packages/coding-agent/test/modes/workflow.test.ts | Adjusts tests for keyword detection and the updated workflow notice contract. |
| packages/coding-agent/src/prompts/system/workflow-notice.md | Replaces the hidden notice with a structured delegate-first DAG execution contract. |
| packages/coding-agent/src/modes/workflow.ts | Updates docstrings to match the new notice semantics. |
| packages/coding-agent/src/modes/magic-keywords.ts | Minor comment tweak describing the fast-path check. |
| packages/coding-agent/src/config/settings-schema.ts | Updates the setting description to reflect the new workflow contract. |
| packages/coding-agent/CHANGELOG.md | Notes the workflow notice change under Unreleased. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| it("matches the standalone workflowz keyword delimited by whitespace", () => { | ||
| expect(containsWorkflow("workflowz")).toBe(true); | ||
| expect(containsWorkflow("please workflowz this rollout")).toBe(true); | ||
| expect(containsWorkflow("design the workflowz")).toBe(true); | ||
| expect(containsWorkflow("run these workflowz")).toBe(true); | ||
| }); |
| it("references only bundled agents (no critic / deep_task)", () => { | ||
| expect(WORKFLOW_NOTICE).toContain("agent: plan"); | ||
| expect(WORKFLOW_NOTICE).toContain("agent: reviewer"); | ||
| expect(WORKFLOW_NOTICE).not.toContain("critic"); | ||
| expect(WORKFLOW_NOTICE).not.toContain("deep_task"); | ||
| }); |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 39fc979fc4
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - Keep going until the task is closed — a returned fan-out is a step, not a stopping point. | ||
| </execution> | ||
| ## 9. Fan-out toolkit | ||
| The delegation mechanism is the `eval` runtime (full helper reference lives in the `eval` tool description): `agent()` (one subagent; `schema=` to branch on a validated object; `return_handle` for a DAG node carrying the `agent://` handle), `parallel()` (one wave), `pipeline()` (staged waves, barrier between stages), `completion()` (cheap stateless scoring inside a fan-out), plus `budget`/`log`/`phase`. Compose the harness the task calls for — **adversarial-verify** (N skeptics per finding, each prompted to refute, keep on majority survive), **judge-panel** (N approaches scored by parallel judges, synthesize the winner), **loop-until-dry** (keep spawning finders until K rounds surface nothing new, dedup against everything seen). Scale fan-out to the ask. A returned fan-out is a step, not a stopping point — keep going until every scenario passes. |
There was a problem hiding this comment.
When a workflowz turn follows this new DAG guidance, the notice tells the model to use agent(..., return_handle=...), but the eval helper only accepts prompt, agent_type, model, label, and schema (as shown in src/prompts/tools/eval.md and implemented in src/eval/py/prelude.py). In Python eval this raises TypeError: agent() got an unexpected keyword argument 'return_handle' before any subagent can spawn, so handle-piped DAG workflows break precisely in the mode this notice is meant to enable.
Useful? React with 👍 / 👎.
Summary
Rewrites the
workflowzmagic keyword's hidden notice into a delegate-first DAG execution contract: decompose the task and fan the work over subagents as an acyclic graph (via theevalpipeline/parallelruntime with handle piping), with certainty-before-code, a skill survey, a plan-agent gate, a scenario contract, evidence-bound manual QA, TDD, and a reviewer gate. It references only bundled agents and does NOT change thinking effort.Tests
Detection / highlight / notice-contract regression test;
bun checkgreen.Depends on #2662.
Closes #2670.