Explore: adopt githubnext/agentics repo-assist as an autonomous AI coworker

[rsnodgrass]

Summary

Evaluate adopting GitHub Next's repo-assist Agentic Workflow as an autonomous AI coworker on this repo, configured to ox conventions.

Filing for future review — not committed to. Captures the design discussion so we don't re-derive the analysis next time someone surfaces this idea.

What repo-assist is

A gh aw-installed GitHub Actions workflow that runs on cron (and via /repo-assist <instructions> comments). Each run picks ~3 tasks from a weighted draw:

• Issue triage / labeling / investigation / fixing
• Perf, testing, and code improvements
• Dependency + CI updates
• PR conflict resolution + author nudges
• Feature work + strategic planning
• Monthly activity summaries

Reads AGENTS.md, stores state in repo "memory", model-driven (Claude).

Install:

gh extension install github/gh-aw
gh aw add-wizard githubnext/agentics/repo-assist

Why this is worth doing

1. Customer-filed GH issues are real inflow. ~30 open issues (Local audio capture / Vox-style integration interest? #611, Ledger finalizer re-runs terminally-failed sessions, causing duplicate-finalize loop and unresolvable rebase conflicts #601, Device does not poll for server-side config updates (name / team changes) #577, bug: Claude Code resume starts a new raw.jsonl instead of continuing the active recording #541, bug: regenerateArtifacts silently skips summary.md when summary.json is missing #534, …), mostly unlabeled. First-pass triage + investigation comments on customer reports is direct human-time savings even before any code changes.
2. Cost model doesn't conflict with house rules. The CLAUDE.md "reviewer attention is scarcest" rule applies to human reviewers. Autonomous agents opening AND closing PRs cost compute, not human minutes — so PR volume isn't the constraint it would be in a human-only loop.
3. Dogfooding upside. SageOx is agentic context infrastructure. Running another autonomous AI coworker inside the canonical SageOx repo, configured to record via sessions/ledger, demonstrates the product working on itself.

Scoped adoption plan (if pursued)

Adopt the full bundle, configured to ox conventions:

1. gh aw add repo-assist on a branch (not main).
2. Tighten AGENTS.md (symlink → CLAUDE.md) with an explicit AI-coworker preamble:
   • Draft PRs only
   • One-line commits, type(scope): summary
   • No "Claude" / "Claude Code" / Co-Authored-By: Claude anywhere
   • SageOx attribution footer when applicable
   • beads-first task tracking (don't open GH issues for internal work)
   • Respect "Required Reviews" paths and canonical-function list in CLAUDE.md
   • LFS rules ( internal/lfs only, no git-lfs shell-out, no filter=lfs in .gitattributes)
3. Restrict on-demand /repo-assist <instructions> trigger to maintainer commenters (or drop on-demand entirely, cron-only) — free-form prompts from arbitrary commenters are the obvious abuse surface.
4. Wire runs into ox session/ledger recording so the work is visible to teammates as any other AI coworker, not a silent bypass channel.
5. Pair with the autonomous-merger pipeline that closes the loop on its PRs (see preconditions).

Preconditions

• Autonomous reviewer/merger path exists (or we commit to building one). Without it, repo-assist PRs pile up as human review tax — which would violate the "human attention is scarcest" rule. The whole adoption case depends on this loop being closed by another agent.
• GH_AW_CI_TRIGGER_TOKEN (optional, auto-rerun CI) — threat-model before enabling.

Out of scope / skip

• Dependency updates — Renovate/Dependabot are more conservative and better understood. Don't let repo-assist drive go.mod / pnpm-lock.yaml bumps.
• Monthly activity summaries — /retro skill already covers this.
• Anything touching CLAUDE.md "Required Reviews" paths without Ryan in the loop (path locations, data access ergonomics, API source of truth).

Verification (trial protocol)

1. Install on a throwaway branch, not main.
2. Watch one cron run end-to-end; audit every PR/comment against ox house style.
3. Confirm autonomous merger catches convention drift (drafts, commit format, attribution).
4. Kill criteria:
   • Convention drift the merging agent doesn't catch
   • Merged-rate < 50% on agent-opened PRs after 2 weeks
   • Any work touching "Required Reviews" paths without human approval

References

• repo-assist docs: https://github.com/githubnext/agentics/blob/main/docs/repo-assist.md
• gh aw extension: https://github.com/github/gh-aw
• Internal design notes (local): ~/.claude/plans/system-instruction-you-are-working-graceful-hartmanis.md (evolved during the session; the final issue text here is the canonical version)