| name | delivery-board |
| description | Operate the Prime-Silo delivery board (delivery/) autonomously — select, claim, execute, verify, and log work contracts without frontier supervision. Use whenever picking up a task from delivery/board/BOARD.md, authoring a gate script, verifying another agent's task, or deciding which model tier should take which contract. |
Delivery-board operator — run the board without supervision
The board at delivery/board/BOARD.md is the single source of work. delivery/README.md is the
law — read it once per session, follow it exactly. This skill adds what README assumes you know:
the mechanics, the tier routing, and the traps that have actually burned agents in this repo.
1. The loop (memorize this shape)
read README → read BOARD → take TOPMOST READY item → claim (board edit + LOG line + commit)
→ read ONLY your contract (delivery/tasks/<ID>.md) → make a worktree if sandbox: worktree
→ TDD: failing test/gate FIRST, watch it fail → implement inside the allowlist → gate green
→ move board line to VERIFY + LOG `ready-for-verify` + commit → STOP (never self-DONE)
- Claim commit message:
chore(delivery): claim <ID> · work commits: feat(<ID>): <what> /
test(<ID>): <scenario name> · verify handoff: chore(delivery): <ID> ready-for-verify.
- LOG format (append-only, never edit):
<ISO-ts> | <ID> | <event> | <agent> | <note>
Events: claimed, ready-for-verify, verified-by, blocked, unblocked, note.
- Agent identity: use
<harness>-<tier> (e.g. claude-sonnet, claude-haiku-verifier).
- Promotion rule: when the last dep of an AUTHORED item is verified DONE, move it to the
BOTTOM of READY with a note
(dep <X> DONE — entered READY <date>). Anyone may promote;
only the human reorders READY.
- Two strikes → blocked. Same failure twice: log
blocked with the exact error and what you
tried. Do not redesign around the contract. Do not brute-force a third attempt.
2. Model-tier routing (who takes what)
| Tier | Takes | Never takes |
|---|
| haiku | Independent VERIFICATION of others' tasks (re-run gate from clean checkout, tick BDD scenarios, move VERIFY→DONE), mechanical single-surface contracts (grep-and-replace passes like C4 execution, LOG/board hygiene, running gates) | Contract authoring, anything with design judgment, multi-file architecture, writing new gates |
| sonnet | Standard contracts: layout/CSS work (C1), SSE/API endpoints (B1), keyframe+binding work (C5), test suites, gate scripts from a spec, docs passes | Contract authoring/splitting, taste-defining UI (first impressions), cross-workstream tradeoffs |
| opus | Design-taste flagships (C3 login/first-run, D1 Studio spec, E-workstream visual), contract authoring/splitting at plan checkpoints, anything touching >2 workstreams, unblocking a task two tiers failed | Bulk mechanical work (wasteful) |
Verification is ALWAYS a different identity than the author — a fresh session/agent with no shared
context, reading only the contract + README, from a clean checkout. Haiku is the default verifier;
escalate to sonnet if the gate involves preview/screenshot judgment.
3. Worktree recipe (sandbox: worktree)
The repo lives under OneDrive; put worktrees OUTSIDE it so sync doesn't thrash:
git -C /c/Users/nsdha/OneDrive/binary16/prime-silo worktree add \
/c/Users/nsdha/.ps-worktrees/<ID> -b task/<ID> main
Work + commit there. When the gate is green, hand off; the merger (or verifier) does
git merge --no-ff task/<ID> on main after verification, then
git worktree remove /c/Users/nsdha/.ps-worktrees/<ID>.
Board/LOG edits are made on MAIN directly (they are coordination, not task payload) —
claim before creating the worktree so the claim is visible to everyone immediately.
If node_modules are needed in the worktree: npm ci --prefix server (only if your gate
actually requires the server; static-analysis gates should stay zero-dep).
4. Gate-script craft (scripts/gates/*.mjs)
Study scripts/gates/c0.mjs and w0.mjs before writing one. House style:
- Zero-dep Node ESM (
node:fs, node:path, node:child_process only). Exit 0 = pass,
non-zero = fail naming the file and the rule violated (a gate that fails silently is a defect).
- Prefer static analysis (grep/parse the source) over spinning up servers. If a preview is
unavoidable (C1 screenshots), make the gate degrade honestly: check what it can statically,
print
MANUAL: <what the verifier must eyeball> for the rest.
- Ratchet pattern for pre-existing violations: record the floor inline (
VIOLATION_FLOOR),
fail only if the count RISES or a new file offends (see c0.mjs).
- A gate must fail-as-designed first: run it against the pre-change tree and watch it exit
non-zero before you implement. That IS the TDD red step for infra tasks.
5. Traps that have actually burned agents here
.env contains BENNY_HMAC_KEY — never staged, never committed. Same for logs/,
brain/, scratch/, site/ (unrelated WIP), and any live run workspace. Stage files BY NAME,
never git add -A.- The tree usually has unrelated uncommitted changes (owner's WIP). Do not touch, stash,
or commit them. Worktrees sidestep this — another reason to use them.
- Allowlist is literal. Needing one file outside it =
blocked, not improvisation. Tests
for your scenarios are implied-allowed (put them in tests/ or next to the gate).
- Budget = changed lines excluding tests/lockfiles. Over budget → log
blocked: needs split.
- Governed CSS is linted by c0.mjs on every commit — no hex outside
colors.css, no
text-align: justify, base font ≥16px, exactly 3 elevation tokens. Read
architecture/DESIGN-SYSTEM.md before touching any CSS. Your task's gate should CALL c0
if you touched governed CSS.
- Windows + Git Bash: shell scripts run under Git Bash (
/c/Users/... paths); Node is the
portable choice for gates. PowerShell 5.1 has no &&.
- Do not call the LAN LM host (192.168.68.125:1234) — a LONGVIEW run may be in flight and
concurrent probes wedge the engine. UI/board tasks never need it.
- Old routes keep working — every UI change ships behind a flag or is additive
(protocol rule:
PRIME_SILO_NEW_SHELL pattern). Breaking the default experience = blocked.
- UI tasks read
.claude/skills/prime-silo-experience/SKILL.md (identity, motion doctrine,
progressive discovery) and, if animating, .claude/skills/animejs-scrollcraft/SKILL.md
(the verified anime.js v4 API — API drift is the #1 historical failure source).
6. Repo map for board work (verified paths)
- Board/law:
delivery/board/BOARD.md, delivery/board/LOG.md, delivery/README.md
- Contracts:
delivery/tasks/<ID>.md · template: delivery/tasks/_TEMPLATE.md
- Plan narrative:
architecture/PLAN-local-power-unified-ui.md (§ per workstream)
- Design law:
architecture/DESIGN-SYSTEM.md + app/L0/_all/mod/_core/framework/css/{colors,layout,type}.css
- Gates:
scripts/gates/ · Login page: server/pages/login.html (+ enter.html, index.html)
- Mascot/onscreen agent:
app/L0/_all/mod/_core/onscreen_agent/panel.html
- Prime-Silo modules:
app/L0/_all/mod/_prime_silo/ · widgets: app/L0/_all/mod/_core/widgets/
- UI-UX findings:
architecture/REQUIREMENTS-ui-ux-refactor.md (C3/C5 source — don't re-audit)
- Launch configs for preview:
../.claude/launch.json (binary16 root; shell = the app on :3000)
7. Definition of done (verifier checklist)
All BDD scenarios have named tests · verify exits 0 from a clean checkout · lint/format clean ·
allowlist respected (diff audit) · budget respected · board+LOG updated at every transition ·
docs/skills updated if the contract's Handoff says so · author ≠ verifier.
