Replace genai dedup action with agentic duplicate-surfacing workflow

[denelon]

📖 Description

Replaces automatic-issue-deduplication.yml (which used pelikhan/action-genai-issue-dedup@v0 with label_as_duplicate: false, so it produced no visible signal — no label, no comment) with a GitHub Agentic Workflow (gh aw) that surfaces likely duplicates for maintainer review.

On every newly opened issue, a Copilot-engine agent reads the issue and searches existing issues (open and closed) by meaning, not keyword overlap, then — only when confident — applies a new Possible-Duplicate label and posts a single comment listing up to five candidate duplicates with links and one-line rationale.

The workflow detects but never resolves: it never closes the issue, never removes Needs-Triage, and never emits the Duplicate of #NNN moderator trigger. Choosing the canonical issue (earliest / curated preferred) and closing stays a maintainer decision.

Why replace the current action

• The current action runs with label_as_duplicate: false — no label, no comment, so duplicates are still found manually.
• Lexical matching misses semantic duplicates that share little vocabulary. Example: Add command or flag to show manifest verbatim (in CLI) - i.e. winget cat or winget show --verbatim #1458 ("show manifest verbatim") and Verbose show command #159 ("Verbose show command") are the same request but read very differently.
• An agentic workflow reasons over meaning/synonyms, and its actions are constrained by a safe-outputs allowlist (one comment + one label — no close, no assign, no code execution).

Changed files

• Add .github/workflows/duplicate-surfacing.md (agentic spec)
• Add .github/workflows/duplicate-surfacing.lock.yml (compiled via gh aw compile; generated — do not hand-edit)
• Add .github/aw/actions-lock.json (pinned action versions for the framework)
• Update .gitattributes (marks *.lock.yml as generated / merge=ours)
• Remove .github/workflows/automatic-issue-deduplication.yml
• Update .github/actions/spelling/allow.txt (add toolsets and msftbot)

New label required: Possible-Duplicate (advisory), distinct from the authoritative Resolution-Duplicate (applied on human confirmation via the trigger). See the linked issue for the label plan.

Enablement prerequisites (org/repo): agentic workflows enabled + the Copilot engine and its secrets (COPILOT_GITHUB_TOKEN, GH_AW_GITHUB_TOKEN, GH_AW_GITHUB_MCP_SERVER_TOKEN). Same framework microsoft/mxc adopted (#639/#646).

Safety posture: safe-outputs restricts the agent to add-labels: [Possible-Duplicate] (max 1) and add-comment (max 1); MCP GitHub access is repo-scoped; workflow token is issues: read; issue content is treated as untrusted (prompt-injection defenses); trigger is opened only to bound comment volume; egress is firewalled to a pinned domain allowlist.

Authored with GitHub Copilot assistance.

🔗 References

• Resolves Agentic duplicate-surfacing workflow to flag likely duplicate issues for maintainer review #6399
• Prior art: microsoft/mxc agentic issue-triage (Fix typo #639, New Manifest Schema and Structure #646)

🔍 Validation

• gh aw compile duplicate-surfacing → 0 errors, 0 warnings; lock file regenerated deterministically.
• Live trial passed via gh aw trial (simulated issue Add command or flag to show manifest verbatim (in CLI) - i.e. winget cat or winget show --verbatim #1458 against microsoft/winget-cli): the agent applied Possible-Duplicate (HIGH confidence) and surfaced Verbose show command #159 as the duplicate — the exact semantic match lexical/TF-IDF scoring misses (~0.15). No close, no Duplicate of trigger, correct comment format. Trial repo was private and deleted after.

🚀 Rollout & enablement

• Scope (intentionally narrow): this is a contained first step — dedup only, safe-outputs limited to one comment + the Possible-Duplicate label, no code changes, no PRs. Chosen deliberately for a public, first-party repo to build trust before considering broader automation.
• Engine auth — no custom secret required: the Copilot engine runs on the built-in GITHUB_TOKEN together with the copilot-requests: write job permission. A dedicated COPILOT_GITHUB_TOKEN secret is not configured or needed (same setup microsoft/mxc uses, Fix typo #639/New Manifest Schema and Structure #646). Copilot is already available to Actions in the microsoft org.
• Workflow permissions: the repo default GITHUB_TOKEN is read-only; this workflow's safe-outputs jobs request issues: write + pull-requests: write explicitly, which overrides the default — no repo-wide setting change was needed.
• Actions allow-list: the repo restricts actions to a selected allow-list; github/gh-aw-actions/*@* was added so the gh-aw setup action is permitted. GitHub-owned actions/* were already allowed; the gh-aw container images (ghcr.io/github/gh-aw-*) are pulled at runtime and are not gated by the actions allow-list.
• Not adopting Repo Assist here: githubnext's broader Repo Assist workflow (which opens PRs and edits code) is a separate, governed decision — out of scope for this PR.
• Public-repo note: do not set GH_AW_CI_TRIGGER_TOKEN on this repo (abuse risk per gh aw docs).

✅ Checklist

• Signed the Contributor License Agreement
• Linked to an issue
• Updated Release Notes (if applicable)
• Updated documentation (if applicable) — workflow spec is self-documenting; N/A elsewhere
• Updated Copilot instructions (if build, architecture, or conventions changed) — N/A, no convention change

📋 Issue Type

• Bug fix
• Feature
• Task