4.8 KiB
4.8 KiB
Stackchain AI Lab — Agent Engineering Report
Audience: agent engineers
Date: 2026-07-08
Status: Operations active, prototypes shipping, release candidates pending human test only
1. Current State — What We Have Built
| Layer | Artifact | Status | Evidence |
|---|---|---|---|
| Infra | Gitea 1.22.0 on VPS, Caddy-free routing | Live | https://forge.alexanderwhitestone.com/git/ |
| Org | stackchain org, 4 repos, 5 users |
Live | Repos: hackathon, dashboard, agent-todo-tracker, project-template |
| Auth | Least-privilege tokens per agent/user | Live | timmy org-admin, turdbot bot, admin CLI |
| Pipeline | CI + RC workflows in .gitea/workflows |
Live | Push to main auto-tags + drafts RC |
| Loop | Autonomous issue→PR→review→merge/no-spam | Locked | Stale 72h, close +24h, velocity #Velocity |
| Agent | agent-todo-tracker CLI shipped |
Done | init/create/list/update/close, SQLite storage |
| Kanban | Hermes board + Gitea label sync | Wired | scripts/gitea_sync.py |
| Creative | Hou3 dispatcher profile (Tencent Hy3) | Active | Assigned to #16/#17/#18, composing deliverables |
2. Operational Constraints — Read This First
- No Caddy changes for Gitea work. Period.
- No human devops. Humans only: review prototype, test RC, approve ship.
- Cached config:
/etc/gitea/app.iniROOT_URL=https://forge.alexanderwhitestone.com/git/ - Asset path:
/var/lib/gitea/public/assets/— verified exists - Pages status: Repo Pages API returns 404; enablement gated on human config handoff
- Dupe cleanup:
john/JohnGaltremoved;john52galtretained - Admin roster:
rockachopa,timmy,grepples,john52galt
3. Market Fit Matrix — What We Can Build Next
| Idea | Market Signal | Fit | Build Order |
|---|---|---|---|
| Adaptive UI dashboard | Developers want intelligent, context-aware interfaces | Strong | P0 — already scoped in #8, needs fleshed panels |
| Agent Task Harness | Multi-agent orchestration is table stakes for agentic startups | Strong | P1 — deterministic routing, retry, circuit break |
| Agent Memory Store | Every agent needs persistent memory across sessions | Strong | P1 — shared KV store, embeddings optional |
| PR Review Agent | Automated code review reduces human review load | Strong | P1 — inline comments, label-based auto-merge |
| Repo Scaffolder Agent | Natural-language project generation lowers onboarding friction | Medium | P2 — templates + agent-driven file gen |
| Hou3 creative pipeline | Marketing/docs automation for launch velocity | Medium | P2 — release notes, overlay variants, README polish |
| Gitea Pages site | Public landing page for the lab | Medium | Blocked on Pages enablement |
| Kanban-to-issue sync | Single pane of glass for ops | Medium | Wired but untested at scale |
4. Recommended Sequencing — Take It to the Next Level
Phase A — Close the Loop (0–48h)
- Ship dashboard panels. Turn the scaffold in
stackchain-dashboardinto a clickable, adaptive UI. Map panels to concrete Gitea views: review queue, issue triage, repo health. - Wire Hou3 output to PRs/issues. Move from “drafting” to “posting review comments” on assigned issues. Close #16, #17, #18 with actual deliverables.
- Run the first full autonomous cycle. Pick one P1 issue, branch it, implement, open PR, self-review, merge. Prove the loop without spam.
Phase B — Harden (48h–2w)
- Add tests. The dashboard and
agent-todo-trackerneed real pytest suites. CI already runs them — fill the gap. - Token rotation cadence. Automate via cron or a Hermes job. Don’t rely on manual rotation.
- Resource audit. 8GB RAM VPS. Model: Gitea + Caddy + N agents + dashboard. Profile before adding more workers.
Phase C — Scale (2w+)
- Multi-repo velocity. Queue issues across
stackchain/*, not justhackathon. - External preview URLs. Each RC gets a unique staging target. Humans click, test, approve.
- Hou3 specialization. Split creative dispatch into its own queue with different acceptance criteria than code agents.
5. Risk Register
| Risk | Likelihood | Mitigation |
|---|---|---|
| Gitea Pages never enables via API | Medium | Surface as explicit human handoff in #14; do not block other work |
| 8GB RAM exhaustion | Medium | Add memory watchdog; serialize worker starts |
| Duplicate accounts / auth drift | Low | Single admin path through Timmy; audit weekly |
| CI workflow YAML drift | Medium | One canonical workflow source; no mirroring |
| Hou3 repeats itself / spams comments | Medium | Seen-state tracking in dispatch.state; max-comment guard |
6. One-Liner Direction
Ship adaptive UI, prove the loop, harden tests, keep humans on prototype/RC gate only.