assign-to-user / unassign-from-user safe outputs are ignored

[eran-medan]

Steps to reproduce

1. Implement safe outputs that re-assign a user (assign to user A, un-assign to existing asignee) based on the formal docs
2. Run a trial in a case where assignment is needed (e.g. have a rule - if alice is assigned assign bob and unassign alice, else do the opposite etc)
3. Workflow run successfully and reports success on these actions, no errors.
4. The action is not performed

Version:

gh aw version v0.45.0

[pelikhan]

do you have a run URL we can look at?

[pelikhan]

/scout research the source of these safe outputs

[github-actions]

🔍 Scout Research Report

Triggered by @pelikhan

Executive Summary

The assign-to-user and unassign-from-user safe outputs are not actually being ignored - they are fully implemented and functional in gh-aw v0.45.0. These features were added in February 2026 (PRs #15219 and #16132). The issue reported is likely due to a misconfiguration or misunderstanding of how the safe outputs system works rather than a bug in the implementation.

View Detailed Research Findings

Research Findings

1. Implementation Timeline & Source

Both safe output types were recently implemented with full feature parity:

unassign-from-user Safe Output - Added February 12, 2026

• Source PR: Add unassign-from-user safe output handler #15219 (Add unassign-from-user safe output handler)
• Merged: February 12, 2026 at 19:48 UTC
• Purpose: Enable workflows to remove user assignments from issues/PRs (symmetry with assign-to-user)

assign-to-user Schema Updates - Added February 16, 2026

• Source PR: Add unassign_from_user and assign_to_user to compile-time safe outputs schema #16132 (Add unassign_from_user and assign_to_user to compile-time safe outputs schema)
• Merged: February 16, 2026 at 16:24 UTC
• Purpose: Fixed missing compile-time schema definitions (tools existed at runtime but were missing from compiler validation)

2. Architecture Overview

The safe outputs system has a two-phase architecture:

Phase 1: Agent Execution (MCP Server)

• Location: actions/setup/js/safe_outputs_mcp_server.cjs
• Purpose: Provides tools to the AI agent during execution
• Tools Defined: actions/setup/js/safe_outputs_tools.json (lines 445-474)
• Handler Registration: Tools are registered via safe_outputs_tools_loader.cjs
• Output: Writes NDJSON entries to GH_AW_SAFE_OUTPUTS file

Tool Definitions (from safe_outputs_tools.json):

{
  "name": "assign_to_user",
  "description": "Assign one or more GitHub users to an issue..."
}
{
  "name": "unassign_from_user", 
  "description": "Remove one or more assignees from an issue..."
}

Phase 2: Safe Output Processing (Handler Manager)

• Location: actions/setup/js/safe_output_handler_manager.cjs OR safe_output_unified_handler_manager.cjs
• Handler Files:
  • actions/setup/js/assign_to_user.cjs (line 57-58 in HANDLER_MAP)
  • actions/setup/js/unassign_from_user.cjs (line 58 in HANDLER_MAP)
• Execution: Reads NDJSON output and executes GitHub API calls
• GitHub API:
  • assign_to_user → github.rest.issues.addAssignees() (line 121-126 in assign_to_user.cjs)
  • unassign_from_user → github.rest.issues.removeAssignees() (unassign_from_user.cjs)

3. Handler Implementation Details

assign_to_user.cjs (actions/setup/js/assign_to_user.cjs):

• Factory pattern with main(config) export
• Configuration support: allowed, max, target, target-repo, allowed-repos
• Cross-repository support via resolveAndValidateRepo() from repo_helpers.cjs
• Staged mode support (GH_AW_SAFE_OUTPUTS_STAGED=true for dry-run)
• Comprehensive validation and error handling

unassign_from_user.cjs (actions/setup/js/unassign_from_user.cjs):

• Identical architecture to assign_to_user.cjs
• Uses github.rest.issues.removeAssignees() API
• Same configuration options and validation logic

4. Compiler Configuration

Go Compiler (pkg/workflow/):

• assign_to_user.go - Parser and config struct (AssignToUserConfig)
• unassign_from_user.go - Parser and config struct (UnassignFromUserConfig)
• Both embed BaseSafeOutputConfig and SafeOutputTargetConfig
• Schema validation in pkg/parser/schemas/main_workflow_schema.json
• Handler registration in safe_output_handler_manager.cjs HANDLER_MAP

Workflow Compilation:

• Compiler generates GH_AW_SAFE_OUTPUTS_HANDLER_CONFIG environment variable
• Example from daily-assign-issue-to-user.lock.yml:

  {
    "add_comment": {"max": 1, "target": "*"},
    "assign_to_user": {"max": 1, "target": "*"},
    "missing_data": {},
    "missing_tool": {}
  }

• Handler manager is called via: safe_output_handler_manager.cjs:294

5. Configuration Format

safe-outputs:
  assign-to-user:
    max: 3                          # Max operations per run (default: 1)
    allowed: [user1, user2]         # Optional allowlist for usernames
    target: "*"                     # Any issue, "triggering", or specific number
    target-repo: "owner/repo"       # Optional default target repository
    allowed-repos: ["owner/other"]  # Cross-repo operations allowlist

  unassign-from-user:
    max: 3                          # Max operations per run (default: 1)
    allowed: [user1, user2]         # Optional allowlist
    target: "*"                     # Same targeting options
    target-repo: "owner/repo"       # Optional default target
    allowed-repos: ["owner/other"]  # Cross-repo allowlist

6. Testing & Validation

Test Coverage:

• assign_to_user.test.cjs - 11 comprehensive tests
• unassign_from_user.test.cjs - 11 comprehensive tests
• Integration tests: pkg/cli/workflows/test-assign-to-user.md, test-unassign-from-user.md
• Smoke test: .github/workflows/smoke-codex.md includes both handlers

Staged Mode Support (PR #15689 - February 14, 2026):

• Both handlers support GH_AW_SAFE_OUTPUTS_STAGED=true for dry-run preview
• Returns {success: true, staged: true, previewInfo: {...}} without making API calls

7. Known Working Workflows

Currently using these safe outputs:

• smoke-codex.lock.yml - Uses unassign_from_user to remove githubactionagent from PRs
• daily-assign-issue-to-user.lock.yml - Uses assign_to_user for automated issue assignment

Potential Root Causes for Reported Issue

Based on my research, if these safe outputs are "being ignored," the likely causes are:

1. Missing Configuration: The workflow frontmatter doesn't declare assign-to-user or unassign-from-user in safe-outputs: section

   • Fix: Add the safe output type to your workflow's YAML frontmatter

2. Version Mismatch: Using gh-aw version < v0.45.0 (these features are from Feb 2026)

   • Fix: Upgrade to latest gh-aw version

3. Configuration Validation Failure: The allowed list excludes the target users

   • Fix: Check allowed: configuration or remove it to allow all users

4. Target Mismatch: The target: setting doesn't match the issue being processed

   • Fix: Use target: "*" for any issue or verify targeting configuration

5. Max Limit Reached: The handler already processed max items in this run

   • Fix: Increase max: value or check for multiple tool calls

6. Repository Permissions: Cross-repository operations without proper target-repo or allowed-repos config

   • Fix: Add cross-repository configuration if needed

7. Agent Not Calling Tools: The agent writes markdown text instead of using MCP tools

   • Fix: Verify agent prompt instructs to use safe output tools

8. Workflow Not Recompiled: Using outdated .lock.yml file from before these features existed

   • Fix: Run gh aw compile to regenerate workflow files

Recommendations

1. Verify Workflow Configuration:

   # Check if safe outputs are declared in your workflow
   grep -A10 "safe-outputs:" .github/workflows/your-workflow.md
   
   # Check compiled workflow for handler config
   grep "GH_AW_SAFE_OUTPUTS_HANDLER_CONFIG" .github/workflows/your-workflow.lock.yml

2. Check Agent Output:

   # Look for NDJSON entries in agent output file
   cat /tmp/gh-aw/agent/output.ndjson
   
   # Should contain entries like:
   # {"type":"assign_to_user","issue_number":123,"assignee":"username"}
   # {"type":"unassign_from_user","issue_number":123,"assignee":"username"}

3. Verify Handler Execution:

   • Check GitHub Actions workflow logs for "Process Safe Outputs" step
   • Look for handler processing messages: "Assigning N users to issue #X"
   • Check for error messages or warnings about skipped operations

4. Test with Minimal Configuration:

   safe-outputs:
     assign-to-user:
       max: 3

5. Request Run URL: Ask the user to provide a specific workflow run URL to diagnose the actual failure

Key Sources

• PR Add unassign-from-user safe output handler #15219: Add unassign-from-user safe output handler

  • Add unassign-from-user safe output handler #15219

• PR Add unassign_from_user and assign_to_user to compile-time safe outputs schema #16132: Add unassign_from_user and assign_to_user to compile-time schema

  • Add unassign_from_user and assign_to_user to compile-time safe outputs schema #16132

• PR [docs] Update documentation for unassign-from-user safe output #15310: Documentation updates for unassign-from-user

  • [docs] Update documentation for unassign-from-user safe output #15310

• Documentation: Safe Outputs Reference

  • File: docs/src/content/docs/reference/safe-outputs.md
  • Lines 55-56: Assign to User and Unassign from User sections

• Source Files:

  • actions/setup/js/assign_to_user.cjs - Handler implementation
  • actions/setup/js/unassign_from_user.cjs - Handler implementation
  • pkg/workflow/assign_to_user.go - Compiler config parser
  • pkg/workflow/unassign_from_user.go - Compiler config parser

Suggested Next Steps

1. Ask for workflow run URL - Cannot diagnose without seeing actual execution logs
2. Verify workflow YAML - Check frontmatter declares these safe output types
3. Check agent output - Verify agent is calling the MCP tools (not just writing markdown)
4. Verify gh-aw version - Ensure using v0.45.0 or later
5. Test with smoke-codex workflow - Known working example to compare against

---

CONCLUSION: The safe outputs assign-to-user and unassign-from-user are fully implemented and functional. The reported issue is likely a configuration problem rather than a code defect. A workflow run URL is needed to diagnose the specific failure.

🔭 Intelligence gathered by Scout for issue #16457

[pelikhan]

Add field assign-to-user.unassign-first: true which unassigns the user before re-assigning.