Chat2DB Edition
Chat2DB Community
User Problem
The AI Issue Assistant currently appends a public firewall warning for the internal awmg-mcpg Safe Outputs gateway. A complex Markdown response can also be corrupted when its backticks and newlines are placed directly in a shell --body argument, and the required per-run idempotency marker may be omitted. Reporters should receive only the intended response, without workflow infrastructure diagnostics or damaged Markdown.
Concrete Use Case
When a contributor opens or reopens an Issue, the assistant should read the form and publish exactly one concise response. Markdown identifiers and line breaks must survive unchanged, the final chat2db-ai-issue run marker must be present, and no internal firewall block notice should appear in the public comment.
Product Area
Other
Proposed Outcome
Route Safe Outputs through the generated CLI proxy, require complex comment bodies to use the official JSON-on-stdin safeoutputs add_comment . contract, and verify the exact idempotency marker before and after emission. Keep the firewall enabled, preserve read-only agent permissions, and retain the existing custom provider and Secret boundary.
Alternatives Considered
Adding awmgmcpg or awmg-mcpg to network.allowed was rejected because gh-aw validates neither internal single-label host and upstream issue github/gh-aw#45915 identifies that recommendation as misleading. Disabling the firewall was rejected because it would weaken the workflow security boundary. Waiting for a future gh-aw release was rejected because tools.cli-proxy: true is the upstream-tested source-level workaround.
Contribution
I can submit a pull request
Submission Checklist
Chat2DB Edition
Chat2DB Community
User Problem
The AI Issue Assistant currently appends a public firewall warning for the internal
awmg-mcpgSafe Outputs gateway. A complex Markdown response can also be corrupted when its backticks and newlines are placed directly in a shell--bodyargument, and the required per-run idempotency marker may be omitted. Reporters should receive only the intended response, without workflow infrastructure diagnostics or damaged Markdown.Concrete Use Case
When a contributor opens or reopens an Issue, the assistant should read the form and publish exactly one concise response. Markdown identifiers and line breaks must survive unchanged, the final
chat2db-ai-issuerun marker must be present, and no internal firewall block notice should appear in the public comment.Product Area
Other
Proposed Outcome
Route Safe Outputs through the generated CLI proxy, require complex comment bodies to use the official JSON-on-stdin
safeoutputs add_comment .contract, and verify the exact idempotency marker before and after emission. Keep the firewall enabled, preserve read-only agent permissions, and retain the existing custom provider and Secret boundary.Alternatives Considered
Adding
awmgmcpgorawmg-mcpgtonetwork.allowedwas rejected because gh-aw validates neither internal single-label host and upstream issue github/gh-aw#45915 identifies that recommendation as misleading. Disabling the firewall was rejected because it would weaken the workflow security boundary. Waiting for a future gh-aw release was rejected becausetools.cli-proxy: trueis the upstream-tested source-level workaround.Contribution
I can submit a pull request
Submission Checklist