This is your side of the loop. Agents read CLAUDE.md; you read this.
Supersedes plan/docs/workflow.md (review model changed per D-22; lanes
renumbered; stack changed per D-03).
- ✅ Done 2026-09-05: guardrails v0.1.0 is tagged with a working
preparebuild — dialed depends ongithub:ddteeter/agentic-guardrails-scaffolding#v0.1.0. The CLI bin isagentic-guardrails; adoption goes through the repo's adoption.md/init flow. - Run the guardrails live-loop acceptance test
(
docs/live-loop-verification.mdin the guardrails repo — it exists at v0.1.0; the dialed archive merely referenced it) — folded into Phase 0's done criteria, which prove the loop on this repo. - Accounts/keys: Visual Crossing API key, Strava API app (callback URL can
be workers.dev initially), Google OAuth client, extraction-model API key — OpenAI presumptive, per D-32 (107's
extraction rung; the eval may change the vendor), Sentry project (free tier), UptimeRobot (or
similar) pinging
/health, Turnstile site key. Store all secrets as Wrangler secrets, never in files.
- Alerting is exception-based: Sentry emails on new exceptions; the daily
digest cron emails only when an anomaly threshold trips; UptimeRobot pings
/health. A quiet inbox is the healthy state. - Recovery: bad deploy →
wrangler rollback. Bad data → D1 Time Travel point-in-time restore (30-day window). Photos → originals in R2 are the source of truth; derived sizes regenerate. - The rule that keeps this true: any incident that required manual poking gets a follow-up micro-task making that failure self-healing or self-reporting. Ops debt is paid immediately or it compounds.
Single session, you pair. Point Claude Code at docs/tasks/000 and stay
engaged — every decision made here is one agents inherit. Budget an evening
or two. Done criteria include deliberately breaking a branch to watch the
Stop gate + fixer fire on dialed itself.
Every worktree lives in .claude/worktrees/<name>, which is where Claude
Code puts its own (isolation: "worktree" agents, EnterWorktree,
claude --worktree). Make yours there too, as the recipes below do, rather
than beside the checkout: one git worktree list and one directory then
hold all of them, and the checkout's parent folder holds only repos.
- Tools must skip it. Each worktree is a full second copy of the
project inside the checkout, so anything that globs from the root sees
it. It is gitignored, and dependency-cruiser and eslint exclude it by
name (flat config does not read
.gitignore: before the ignore,npm run lintin the primary checkout linted ~2,200 extra files per worktree). A new root-globbing tool needs the same entry. - Cleanup: a worktree stays on disk after its PR merges, and each holds
about 1 GB of
node_modules.npm run worktrees:prunelists every worktree and local branch whose PR has merged or closed; add-- --yesto remove them. It keeps anything with uncommitted changes (a leftoverPR_BODY.mdaside), and any branch without a PR that holds commits no remote has.
git worktree add .claude/worktrees/lane-101 -b lane/101-closet
git worktree add .claude/worktrees/lane-102 -b lane/102-runs
git worktree add .claude/worktrees/lane-103 -b lane/103-weather
git worktree add .claude/worktrees/lane-104 -b lane/104-feed
# each worktree: npm install, then open Claude Code in it
# lanes 105 (onboarding) and 107 (product intelligence) start after 101 merges:
git worktree add .claude/worktrees/lane-105 -b lane/105-onboarding
git worktree add .claude/worktrees/lane-107 -b lane/107-product-intelSession-start prompt (paste into each Claude Code instance):
Read CLAUDE.md, then docs/tasks/. Follow the workflow in CLAUDE.md: write the design doc first using docs/design-doc-template.md and commit it, then proceed to implementation. I review design docs asynchronously — if I comment on yours, reconcile before continuing. Schema changes and new bindings always stop for my approval.
Suggested rollout order if you'd rather stagger than run four at once: 103 (simplest, proves the loop) → 101 + 102 together → 104 (touches the most seams) → 105 + 107 after 101 merges.
Task 090 (validation study) is cut by D-159 and is not dispatched; the call epic learns from v1's verdict history instead.
Four contracts landed after the v1 lanes were written, and none of them was
adopted by the lanes that shipped under them. They are sequential, on
main, one at a time — not a fan-out. Every one rewrites src/ui/ and most
of src/modules/*/components/, and unlike the schema there is no protocol
for component contention: four worktrees here would conflict on nearly every
file.
| # | lane | why it sits here |
|---|---|---|
| 1 | 113 design system port + §AG/§AH | Everything downstream takes its numbers from tokens.js and T1. Writing raw px before this means rewriting it after. Round 10/11's two v1 slices — composition on garment detail, colour collection on F — fold in here (owner, 2026-09-18) because §AH edits GarmentForm and garmentBase, which Part A already rewrites. Part A lands first and completely; then Part B. |
| 2 | 114 motion adoption | Travel distances are spacing steps. |
| 3 | 112 accessibility | The focus outline, 44px targets and 8px gaps are token values — and its reduced-motion requirement is vacuous until 114 gives it something that moves. |
| 4 | 115 desktop | Consumes all three: BREAKPOINT/MEASURE from 113, the tab-switch treatment from 114, focus order from 112. Also the only one that crosses every lane's routes, so it runs alone. |
| 5 | 116 composition reconciliation | The other half of "does the UI match the design". 112–115 make every value enforceable; nothing checks composition — and rounds 6–8 changed it on already-built screens while the repo could not see the drawings. Last, because 113–115 move every screen and 115 adds a second composition at width. |
The split they inherit. CLAUDE.md §Design truth divides the bundle:
contracts are truth for values, artboards are truth for composition. Lanes
113–115 make the values enforceable once, by lint rule and pin test, so no
human ever reviews a font size again. 116 is the composition half, and it
is review rather than enforcement because no rule can check whether a screen
contains the right things in the right order.
Why they are lanes at all. docs/design-deltas.md carried motion
adoption as something lanes would pick up "opportunistically, audited at the
launch gate". Six lanes shipped and none did: of the doctrine's 12 surfaces,
one is implemented. The same would have happened to the token contract —
93 arbitrary values and 17 Mono bypasses accumulated the same way. A
cross-cutting contract with no packet does not get adopted.
Tasks 110 (The Desk) and 111 (dark theme) are separate and unscheduled; neither blocks this sequence.
- Design docs, batched async (~10 min/lane, phone-friendly by construction). Agents do NOT wait for you — read soon after each lands; a comment is an interrupt the agent must reconcile. Check: contract touches say "none" or are justified; test plan names real tests; open questions are answered (agents guess less when you answer fast). Exception — hard stops that do wait: schema changes, new bindings/queues/crons.
- PR review (the hard gate). The gates have already enforced lint/types/dead code/boundaries/tests — spend your attention on: product correctness against docs/product.md, the EXPLAIN output (lane 104), auth scoping tests present, brand/lexicon adherence, and anything the agent flagged in its end-of-task summary.
Between those checkpoints: don't watch. That's what the gates are for.
- Lanes rebase on main at least daily (
git fetch && git rebase origin/main); tell the agent to do it and resolve conflicts as part of its work. - Merge order on simultaneous readiness: 103 → 101 → 102 → 104 → 105 → 107.
- Per-lane route directories (
src/routes/<lane>/) and untouchabledb/schema*.tsstructurally remove the two classic conflict sources.
When a lane's design doc requests a schema change:
- You approve/adjust the proposal.
- Run it as a micro-task on main (you or a short dedicated session):
edit schema →
drizzle-kit generate→ migrate → merge. - Affected lanes rebase. Never let a lane branch carry a migration.
- Gate escalates repeatedly on one lane → read the recurrence state; it usually means the packet under-specified something. Fix the packet, not the agent.
- Agent asks a question mid-lane → answer in chat AND fold the answer into the packet or CLAUDE.md so the next session doesn't re-ask.
- Two lanes need the same new shared utility → it goes in
lib/(orui/) via a micro-task on main, not copy-paste in both lanes.
The launch gate is no longer a list in this file. The production-readiness
audit (docs/reconciliation/2026-09-25-production-readiness-audit.md)
replaced it with two plans, and the owner separated them (decision D-37):
docs/launch/development-plan.md: everything that is code, in five parallel lanes (tasks 125–129; shared rules indocs/tasks/125-129-launch-development.md). All of it lands first.docs/launch/deployment-plan.md: everything that is a dashboard, an account, a legal registration or a config edit, run as its own sweep afterwards. Its §11 holds the rollout gates: the owner alone, then invited friends, then the public (decision D-38).
Keep the launch list in those two files only. A third copy here would drift from them, which is what this section used to do.
Fan-out for the five lanes:
git worktree add .claude/worktrees/lane-125 -b lane/125-ops-platform
git worktree add .claude/worktrees/lane-127 -b lane/127-strava-logging
git worktree add .claude/worktrees/lane-128 -b lane/128-content-safety
# after PR #104 merges:
git worktree add .claude/worktrees/lane-126 -b lane/126-accounts
# after PR #102 merges:
git worktree add .claude/worktrees/lane-129 -b lane/129-feedWhere the old checklist's items went:
| old item | now |
|---|---|
| 1. Run task 106 | Done; the audit found its ban mechanics unfinished, which is 128 · SAF-4 |
| 2. Enable the CSAM scanning tool | Deployment plan §8. Photos stay private behind sign-in (decision D-69, superseding D-46), so the tool would scan nothing; CSAM coverage for public content is decided before the public launch |
| 3. Observability config (tracing off, logs sampled) | Deployment plan §1 |
| 4. Design-deltas round trip before strangers | The development plan's design dependencies, and the R-45 copy pass as a stage-2 gate |
| 5. Motion Doctrine sweep | Closed by task 114 (design-deltas item 12) |
| 6. Publish a privacy policy (R-105) | 126 · ACC-13 (page and links); deployment plan §9 (the text; PR #109 drafted it) |
The call epic is next (see docs/post-mvp.md); its packets get written the
same way (design doc template, worktree lanes) once v1 verdict data exists.
The engine starts from the generic dressing chart; there is no validation
study (D-159).