Harden stale workflow and add validation - #65

Summary

• harden stale issue eligibility, exemption handling, bot detection, missing-data stops, partial-write recovery, and prompt-injection guidance
• use read-only GitHub CLI proxy access for complete timeline and actor-attributed reaction data
• bound detailed candidate scans while keeping current GitHub state authoritative over cache memory
• add safer Dosu migration and canary guidance, contributor checks, and a repository audit
• add CI for source synchronization, strict validation, compilation, and generated-file drift
• regenerate the workflow and runtime pins with gh-aw v0.83.4

Why

The previous workflow asked the agent to make decisions using timeline and reaction data that its configured tools did not guarantee. It could also close an already-stale issue after an exempt label was added, treat cached state too strongly, scan an unbounded backlog, and turn one legacy operation cap into several independent write caps during migration.

This change makes incomplete data stop writes, makes exemptions win at every decision point, treats cache state as advisory, adds bounded processing, and documents a controlled migration path.

Security review

• ANTHROPIC_API_KEY remains the restricted secret required by the selected Claude engine. No secret value is stored in the repository.
• COPILOT_GITHUB_TOKEN is newly declared by the generated v0.83.4 runtime for its optional OAuth-token check. It is not the selected engine credential.
• No action repositories were added or removed. Existing compiler-owned actions were refreshed, including cache v6.1.0, checkout v7.0.1, setup-node v7.0.0, upload-artifact v7.0.1, and gh-aw setup v0.83.4.
• The generated manifest now records digest-pinned firewall, API proxy, CLI proxy, MCP gateway, Node, and GitHub MCP containers.
• No workflow source redirect was added or changed.
• The generated safe-output and conclusion jobs still request pull-requests: write because the shared label handlers support issues and pull requests. The authored policy remains issues-only, and this compiler limitation is documented for explicit review.

Validation

• gh-aw v0.83.4
• repository workflow strict validation, 0 errors and 0 warnings
• distribution workflow strict validation, 0 errors and the expected ANTHROPIC_API_KEY security-review warning
• strict no-emit compile with schema validation and Actionlint, 0 errors and 0 warnings
• repeated strict compilation produced identical generated-file hashes
• distribution and repository workflow sources are byte-identical
• Actionlint passed for the new validation workflow
• git diff --check passed
• the hosted Validate workflow passed on draft PR Harden stale workflow and add validation #65

The earlier audit also installed and strictly validated the workflow in a temporary clone of dosu-ai/test and produced a trial preview without modifying GitHub. That check used gh-aw v0.83.1 and did not execute the model.

Draft status

No LLM-backed behavioral canary has run. dosu-ai/test lacked the required engine secret and had a large existing backlog, so a live run there would have been unsafe.

Before this is ready for launch, run the behavioral matrix in an isolated fixture repository with a manual trigger, a candidate scan limit no higher than 10, and one write per safe-output type. The audit also tracks the generated pull-request permission, security policy, tagged release, and deterministic pre-write checks as open work.