[warp-agent-staging]

What this feature does

Add an independent UI verification pass to an existing factory. The foreman delegates a recorded PR commit to a verifier on the Warp Agent harness and posts observed results and shareable evidence links as a PR comment. Approval and merging remain with people.

Summary

Replaces the Computer Use scaffold with a task-oriented verification recipe covering runtime preparation, maintainer-requested intake, capture policy, evidence handoff, and validation.

Review status: Revisions committed and pushed at c4d4b65ec667051e1d92f30068abedc77ecd499d. The parent and this child now include main at e8d547c4. Validation below covers the revised article; live factory checks remain unverified as listed below.

Changes

src/content/docs/factories/use-cases/computer-use-verification.mdx

• Preserves the approved title, Verifying changes with Computer Use in your factory, and uses existing-factory framing and natural explanatory prose.
• Uses concise prerequisites, a configuration summary, seven focused setup steps, and five verification tests.
• Routes label events to the existing foreman, which supplies PR context to a configured VERIFY agent and publishes a comment instead of promising automatic attachment to an existing PR description.
• Shows the Warp Agent model shorthand and Warp-hosted execution. Omits the redundant optional Computer Use model pin from the example.
• Explains that a custom verifier has no built-in code-host skills, needs checkout/build access, and requires an app-specific startup and verification skill.
• Binds the result to the checked-out commit SHA, reports stale results when the PR changes, and distinguishes observed behavior from inference.
• Adds untrusted-content boundaries, maintainer review before executing PR code, a stop-and-ask path for forks, and no implementation, approval, or merging.
• Explains current-run artifact ownership, permission-checked links, public embeds, disabled sharing, and blocked or unverified results when evidence is unavailable.
• Separates trigger exclusions from prompt checks and explains that pushing a new commit alone does not retrigger a label automation.

Content design plan

Audience and JTBD: A team shipping user-facing changes that needs rendered-flow evidence for an existing pull request beyond code-level tests.

Problem: Computer Use, custom agents, and artifact sharing have canonical references, but factory users need a concrete handoff from a PR request to a commit-specific verification report.

Goals: Prepare a verifier on a supported runtime; request verification through a scoped label event; exercise visible criteria at the recorded commit; return evidence under team policy; preserve human review and merge controls.

Purpose and value: Connect factory delegation to Computer Use and shareable media without implying that child-run completion edits a PR or proves every criterion.

Content type: Procedural recipe with a configuration summary and example workflow.

Skill and template: draft_procedural / .agents/templates/procedural.md, with repository review and documentation-quality guidance.

High-impact scenarios: Covers same-repository GitHub PRs, a separate verifier on the Warp Agent harness, app startup, criterion-specific captures, and PR-comment evidence. Excludes production credentials, automatic merging, a full untrusted-fork execution setup, and an automatic verification stage for every implementation.

Editorial review

The reader task is an independent verification pass, not creation of a dedicated factory. Harness availability, runner configuration, agent creation, and artifact policy remain owned by their canonical references. Repetitive best-practice bullets were moved beside the relevant setup or validation actions. The original promise that a capture simply lands in the existing PR description was replaced with an explicit child-to-foreman link handoff and PR comment. Useful startup, security, and success criteria remain in the procedure.

Source files consulted

• warpdotdev/warp-server at 27c9fb8049c27037c15833ab711f7b7a131bf6ac: logic/factoryfile/agenttype.go, logic/factoryfile/resolution/resolution.go, logic/ai/ambient_agents/sources/agent_config.go, logic/ai/ambient_agents/workers/common/task_utils.go, and logic/ai/multi_agent/utils/output/tool_call/shared/run_agents.go. Checked the VERIFY type, model/harness inheritance, Warp-hosted selection, Computer Use run defaults, and configured-child dispatch.
• At the same server revision: logic/ai/multi_agent/runtime/get_media_artifact_links.go, logic/ai/multi_agent/runtime/get_artifacts_for_pull_request_description.go, and logic/ai/multi_agent/artifact_attachments/mode.go. Checked current-run scoping, owning-team policy, private-link defaults, opt-in public embedding, and disabled sharing.
• logic/ai/ambient_agents/automations/provider_github.go at that revision: newly applied-label matching and repository/target-branch filters.
• Canonical Computer Use, browser use, screenshots/videos, custom-agent creation, agent configuration, runner, secrets, and factory-definition documentation in the main-synced checkout.
• The related media-sharing product spec was consulted for problem framing only. Its historical rollout note was not used as current availability evidence.

Validation and rollout evidence

• The exact example tree passes the production app.warp.dev factory-files parser with no diagnostics or deferred resolutions. State-dependent apply checks were not run. No factory or live automation was created.
• Production schemas retrieved October 27, 2026: agent.schema.json SHA256 91c8330a767b50d7881b2f7fb2f84e3daa7d2528dfd83c15a416442442ce3e77; automation.schema.json SHA256 61c15fe6e32dd9e3f918739499b2f2720488609e510cd7d455a7016fd0427fce. Parser/schema acceptance is not end-to-end execution evidence.
• Targeted rollout inspection at the server revision above found all nine checked core factory, Computer Use, media-sharing, and supported third-party foreman harness gates enabled in checked-in production configuration. This does not establish deployment or a live factory run. Broad co-occurrence matches were excluded after checking the actual paths. No private flag names or raw rollout output are published.
• Typecheck: 0 errors. Heading-variable tests: 5/5. Build, heading-TOC validation, homepage JSON-LD, compression contract, and whitespace checks pass.
• Main-sync regressions also pass: tracking redirect 6/6, tracking route validation, Kapa WebMCP 2/2, and support escalation 4/4.
• Rendered title and seven setup steps/five test steps checked. All 19 internal links and anchors resolve.
• Style lint reports one glossary warning in this article for the shipped Embed option, plus five inherited provider-name warnings in sibling recipes. Existing build/deprecation warnings remain.
• Removed the unsupported October 7 stable-release assertion. No recipe-specific stable-release date was established.

Unverified claims

• No live factory verification was run. Confirm app setup, commit checkout, browser access, screenshots/recordings, and named-agent delegation with a controlled PR.
• Actual repository credentials, private dependency access, runner compatibility, workspace configuration, and entitlements were not exercised by parser validation.
• Reviewer access to private evidence links, enabled public embedding, disabled sharing, and the complete child-to-foreman PR-comment handoff were checked against source but not exercised end to end in a factory.
• Draft, closed, fork, stale-head, failed-build, and unreachable-criterion behavior are prompt requirements to test, not guarantees enforced by the label filter or a completed live test.

Documentation risk

Risk: engineering-review-required
Rationale: Documents factory agent configuration, execution of PR code, Computer Use compatibility, artifact privacy, and a multi-agent evidence workflow.
Docs override: none

PR stack

• #895 — Navigation and structure
• #896 — Issue triage
• #897 — Code review
• #898 — Issue implementation
• #899 — Production incident resolution
• #900 — Computer Use verification
• #901 — End-to-end SDLC automation

Dependency: This PR targets factory/factory-use-cases-structure. PR 899 remains a separate, unmerged child and is not included in this revision.

Human review focus

• Confirm the explicit PR-comment evidence workflow and the verifier's app-specific skill requirements.
• Review the credentialed-code execution and artifact-sharing boundaries.
• Exercise a controlled live factory run before relying on the recipe.

Agent review conversation

Co-Authored-By: Oz oz-agent@warp.dev
Co-Authored-By: Warp agent@warp.dev

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
docs	Ready	Preview	Oct 28, 2026 2:43am UTC

[warp-agent-staging]

This PR was generated with Warp.

Comment @warp-staging-factory on this PR to send it follow-up work.