Repository navigation
[repo-assist] Add DeepWiki MCP tool + guidance and fix 4 triage bugs (supersedes #50962, #50963) [skip ci] - #50964
Conversation
Combines and supersedes #50962 (DeepWiki MCP) and #50963 (triage-logic fixes) into a single coherent repo-assist.md source and one clean `gh aw compile` of repo-assist.lock.yml, so the workflow can be manually tested end-to-end against one lock file rather than two independently regenerated ones that would clobber each other. DeepWiki (from #50962): - Add read-only DeepWiki MCP server (mcp.deepwiki.com/mcp), scoped to read_wiki_structure / read_wiki_contents / ask_question. - Add "Using DeepWiki (cross-repo grounding)" body section: orientation only for tt-metal/tt-umd/tt-kmd/tt-isa-documentation, verify-before- asserting discipline, and cache pointers-not-facts (dated, re-derivable) in repo-memory rather than frozen DeepWiki answers. Triage fixes (from #50963): - Task 7: require verified zero merged PRs AND zero prior issues/comments (via GitHub search) before welcoming a "new" contributor. - Guidelines: strip benign "Firewall blocked ... awmgmcpg" boilerplate from public comments (network.allowed is a dead end at this compiler version; documented in a frontmatter NOTE). - Task 1/2: scope fresh comment+label triage to issues active in ~90 days and stop defaulting [HELP]/question issues to the `bug` label. - Task 7 + Anti-spam: combine triage + welcome into one comment when both apply to the same issue in a single run. lock.yml regenerated once with gh-aw v0.82.14 (matching base main pin); 0 errors / 0 warnings, deterministic recompile. Do not hand-edit the lock file. Co-Authored-By: Claude <noreply@anthropic.com> Co-Authored-By: BrAIn <brain@tenstorrent.com>
There was a problem hiding this comment.
Pull request overview
Adds DeepWiki cross-repository grounding and refines Repo Assist triage behavior.
Changes:
- Adds a read-only DeepWiki MCP server and verification guidance.
- Tightens contributor detection, issue triage, and comment deduplication.
- Regenerates the compiled workflow lock.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
.github/workflows/repo-assist.md |
Configures DeepWiki and updates triage instructions. |
.github/workflows/repo-assist.lock.yml |
Regenerates the executable workflow configuration. |
Comments suppressed due to low confidence (1)
.github/workflows/repo-assist.md:354
- [🟡] Bug: Prompt text cannot suppress the generated firewall footer
In gh-aw v0.82.14, actions/setup/js/messages_footer.cjs:654-673 reads blocked domains and appends the warning after the agent-produced content. This instruction only controls what the agent submits, so the awmgmcpg warning will still be attached to every public safe output.
Why it matters: the firewall-boilerplate bug advertised as fixed by this PR will reproduce unchanged on the next run. The fix must suppress this internal hostname in gh-aw's footer generation (or upgrade to a release that does) and then regenerate the lock file.
- **Never forward firewall boilerplate into comments**: do not copy or reproduce any `⚠️ Firewall blocked …` warning block (e.g. the benign `awmgmcpg` MCP-gateway notice) into issue/PR comments or descriptions. `awmgmcpg` is gh-aw's own internal MCP Gateway sidecar hostname, not a real missing dependency, and it cannot be silenced via `network.allowed` at the current compiler version (see the NOTE in the frontmatter). Treat any such block as internal-only noise and strip it from anything you post publicly.
Addresses Copilot review comment: the 90-day freshness restriction only
lived in Task 2, so Task 1 ("Triage and Label Open Issues") could still
apply first-time labels to long-stale issues — the exact behavior this
PR intends to stop.
- Task 1: scope first-time labelling to issues active in ~90 days,
mirroring Task 2. First-time labels are not applied cold to stale
issues; correcting/removing an already-wrong label stays allowed.
- Guidelines/"Systematic": clarify that oldest-first cursor order still
respects the freshness gate (does not license first-time
comments/labels on stale issues).
- Regenerated repo-assist.lock.yml via `gh aw compile` (gh-aw v0.82.14).
frontmatter_hash unchanged, body_hash updated, output deterministic.
Co-Authored-By: BrAIn <brain@tenstorrent.com>
|
🤖 Repo Assist — review comments addressed Went through the review feedback on this PR. 1 unresolved review comment (from the Copilot reviewer); no human inline comments or issue comments were open.
What changed (
Lock regeneration: Nothing was pushed back on — the single comment was a correct catch. The thread has been resolved. |
|
/codeowners bypass |
tenstorrent-github-bot
left a comment
There was a problem hiding this comment.
✅ CodeOwners bypass approval granted by @blozano-tt (metalium-developers-infra team)
|
✅ CodeOwners Bypass Approval Granted This PR has been approved by @blozano-tt (metalium-developers-infra team) using the bypass mechanism. |
Summary
This PR combines and supersedes two open repo-assist PRs so there is a single coherent
repo-assist.mdsource and one cleangh aw compileto test against:Why combine them
.github/workflows/repo-assist.lock.ymlis a deterministic function of the whole.github/workflows/repo-assist.mdsource (both itsfrontmatter_hashandbody_hashare recomputed on every compile). Each of the two PRs changed both hashes, so merging them independently would have one PR's lock regeneration clobber the other's. Per @wilder's request, they are merged here into one source file and compiled once so the workflow can be manually tested end-to-end against a single lock.What's included
DeepWiki MCP tool + guidance (from #50962)
https://mcp.deepwiki.com/mcp) scoped toread_wiki_structure,read_wiki_contents,ask_question. No new write permissions — purely an additional read-only tool.tt-metal,tt-umd,tt-kmd,tt-isa-documentation), with a strict verify-before-asserting discipline — anything load-bearing must be confirmed against current code viabash/gh, never asserted from DeepWiki alone.Triage-logic fixes (from #50963)
⚠️ Firewall blocked … awmgmcpgblock from public comments.awmgmcpgis gh-aw's own internal MCP-gateway sidecar; it cannot be silenced vianetwork.allowedat this compiler version (confirmed dead end, documented in a frontmatter NOTE), so it's handled at the instruction level.[HELP]/question issues no longer default to thebuglabel.Compile / testing notes
repo-assist.lock.ymlwas regenerated once withgh aw compileon the combined source using gh-aw v0.82.14 — the exactcompiler_versionpinned in the basemainlock'sgh-aw-metadata. Result: 0 errors, 0 warnings, and a second compile is byte-identical (deterministic). The lock diff reflects only real content changes (DeepWiki added to the MCP gateway config +GH_AW_ALLOWED_DOMAINS, updatedgh-aw-metadatahashes). The.lock.ymlis generated — please don't hand-edit; regenerate viagh aw compile repo-assist./repo-assist <a cross-repo question>(e.g. something about tt-umd device init) on a test issue. One run then exercises both change-sets: the DeepWiki tool for cross-repo grounding, and the fixed contributor-detection / labeling / comment-dedup behavior.Once this is verified and merged, #50962 and #50963 are being closed in favor of it.
🤖 Generated with Claude Code
CI Status
Auto-generated on every push. Badges update live. Click a badge to filter runs by this branch.