Private workbench for Upscale AI — learning, drafts, and personal scripts. Confidential — Upscale AI, Inc. Do not distribute.
Not a team/product repo: promote code out to org repos when something should be shared.
| File | Role |
|---|---|
| CONSTITUTION.md | Initial v0 — runtime model, layers, boot order, invariants; amend by your judgment only |
| GLUON.md | Gluon identity, challenge, fail-closed (GLUON-REQUIREMENTS-SPEC.md imported) |
| README.md | Operating discipline + fresh-agent litmus test (this file) |
| TASKS.md | Open work — source of truth; labels DT###; optional due |
| TASKS-LOG.md | Task activity log (append-only, loose format) |
| META.md | Px priorities, triage, hatch vs run — do not re-derive unless asked |
Offboarded: Cursor canvas and chat history are not authoritative. Git files are.
Until org remote + PR/review workflow exists, go slowly. The human owns merge-worthy decisions.
- Closing or reopening hatch
- Declaring a phase “done” (onboarding, offboard, access)
- Renaming/moving the repo root
- Changing source-of-truth files (
TASKS.md,META.md, this README) - Bulk deletes or large refactors
- Pushing to any remote with confidential content
- Double-check with the human before a big step — state what will change and wait for explicit OK.
- One concern per commit — big steps get their own commit message; do not bundle with drive-by edits.
- Small steps — prefer several small commits over one “cleanup” commit.
- No silent policy changes — if behavior changes, update
README.mdorMETA.mdin the same commit. - Tasks follow work — update TASKS.md when status changes (same session or immediate follow-up commit).
- Git commit/push — human explicit only (commit, push, or pasted command). Agent may
git add/ show diff when asked; do not rungit commitorgit pushwithout clear approval. Enforce via DT119 when built (not rushed).
- PR + human review required for big steps
- Until then: local
mainonly; no push of confidential material without explicit approval (TASKS.mdp1-remote,p1-privacy)
North star (full bar): kill agent, erase chat, clone repo, new model/host — Gluon recovers fully from git. See CONSTITUTION.md North star — full Gluon recovery. Not there yet; that’s ok — this section is the test we run today.
Only you run this test — in a new machine or new agent chat, no prior context.
git clone(or copy)~/diwakar-workandcdinto it.- Open CONSTITUTION.md → confirm stateless boot, single-writer, three layers.
- Open GLUON.md → confirm challenge template and fail-closed rules.
- Open README.md (this file) → confirm deliberate-pace rules are visible.
- Open TASKS.md → list open tasks; confirm Priority and Title columns; find P0 row.
- Open META.md → confirm Px order; confirm hatch status matches your expectation.
- Ask the agent: “What is open in P0? Where are rules? Was hatch closed?”
- Pass: answers cite only these files, no canvas/chat; agent asks before proposing a big step.
- Pass:
TASKS.mdopen table sortedP0 → P1 → TRIAGE → P2 → P3; within each banddoing→next→open. - Pass: without challenge, agent refuses
git commit/ file writes (see GLUON.md). - Fail: agent invents tasks, merges big changes without asking, or re-derives Px/WFQ unprompted → fix docs, not the agent.
- Fail: any
doingornextrow appears below anopenrow in the same band → sort order broken.
| Name | Role |
|---|---|
| Lepton | This MacBook Pro — local dev, clone, commit (~/diwakar-work) |
| Proton | Future cloud workstation (GCP/AWS workhorse) — not provisioned yet |
| Gluon | Agent that lives in this repo; context = committed files only |
| What | Where |
|---|---|
| Working copy (Gluon) | ~/diwakar-work on Lepton — this is the repo |
origin |
git@github.com:upscale-ai-network/agent-template.git — empty org home for Gluon/template |
| Ignore | ~/work/agent-template — colleague’s dummy clone only; not a second checkout to use |
Push: git push -u origin main once you have write on agent-template (TASKS.md DT103). Until then, work local; no dtundlam/diwakar-work remote configured.
| Path | Purpose |
|---|---|
CONSTITUTION.md |
Gluon runtime contract |
TASKS.md / TASKS-LOG.md |
Work queue + log |
META.md |
Px / triage |
dt100/ |
DT100 done — A3 md + pptx (dt100/README.md) |
dt122/ |
DT122 doing — ccc / B6 md + pptx (dt122/README.md) |
scratch/ |
Chat scratch / mini tasks (not task-scoped) |
archive/ |
Portable snapshots (e.g. zsh dotfiles) |
src/py/ |
Sample programs (pytest wiring smoke) |
src/gluon_cli/ |
uv run build-decks / check-decks entry points |
scripts/ |
PPTX pipeline Python — see scripts/PPTX-PIPELINE.md |
tests/ |
uv sync --group dev && uv run pytest tests/ -q |
assets/ |
Shared pipeline slide, whiteboard photos, templates |
Removed (2026-06-04): Apple Silicon MLX brief, demos, PDFs/HTML — history in git; future copy from archive repo if needed.
git clone https://github.com/upscale-ai-network/agent-template.git diwakar-work
cd diwakar-workPrimary (Mac): write Gluon · after uv sync:
uv run build-decks # A3 + B6
uv run build-decks-a3 # A3 only
uv run check-decks # litmus (md, PNGs, pptx)
uv sync --group dev && uv run pytest tests/ -q
uv run pytest tests/ -m workflow -q # md→pptx regression (needs npx)See tests/WORKFLOW-REGRESSION.md. Entry points live in src/gluon_cli/ ([project.scripts] in pyproject.toml).
Zombie standby (Linux vm): read-only · failover not declared — readiness probe only.
Lambda (shell at repo clone; you own PWD):
git fetch origin main && git reset --hard origin/main
uv sync --group dev
uv run check-decksWrapper: ./scripts/zombie-hatch-audit.sh · No node/npx, no pytest, no regen. See CHECKPOINT.md.
First-time hatch (optional zsh dotfiles): ./scripts/bootstrap-gluon-zombie.sh --full
Heavier regen+pytest (optional): ~/diwakar-work/scripts/zombie-pull-build.sh — may dirty .pptx; git restore dt100/*.pptx dt122/*.pptx before leave.
.
├── CONSTITUTION.md
├── TASKS.md
├── TASKS-LOG.md
├── META.md
├── README.md
├── dt100/ # DT100 done — A3
├── dt122/ # DT122 — ccc / B6
├── scratch/ # non-task scratch notes
└── assets/
├── logical-pipeline-boss-slide.png
└── pics/