Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

docs(prompts): finish removing six work orders the owner retired on 2026-09-09

nirholas25c6683cfaa6
diff --git a/prompts/finish/313-home-19-plans-entitlements.md b/prompts/finish/313-home-19-plans-entitlements.md
deleted file mode 100644
index 6132d1a16..000000000
--- a/prompts/finish/313-home-19-plans-entitlements.md
+++ /dev/null
@@ -1,145 +0,0 @@
-# 19. Plans, entitlements and quotas
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`. Read [00-CONTEXT.md](_context/home-00-CONTEXT.md) first. Orders
-[03](home-03-api-surface.md), [12](home-12-households-rbac.md) and
-[14](309-home-14-reliability-scale.md) must have landed: you cannot price what you have not measured.
-
-## Binding operating clause
-
-Finish 100%. Never end with a question or an unexecuted plan. All CLAUDE.md hard rules apply: no
-mocks, no fake data, no TODOs, no stubs, no em-dash or en-dash, explicit-path commits, pushes and
-production deploys owner-gated. **Setting a price the platform charges is an owner decision**:
-implement the mechanism, propose the numbers, and batch the price approval into one message.
-
----
-
-## Step 0: re-derive the current state
-
-```bash
-grep -n "^export" api/_lib/account-tier.js | head
-node -e "import('./api/_lib/account-tier.js').then(m=>console.log(JSON.stringify(m.ACCOUNT_TIERS,null,1)))" | head -40
-grep -n "^export" api/_lib/three-tier.js | head
-ls api/billing api/credits api/subscriptions 2>/dev/null
-grep -rn "resolveAccountTier\|tierById" api/ --include=*.js -l | head
-```
-
-The platform already has tiers (`api/_lib/account-tier.js`), a $THREE holder ladder
-(`api/_lib/three-tier.js`), credits and subscriptions. **Join them. Do not invent a fourth
-entitlement system for this lane.**
-
-## What actually costs money here
-
-Be honest about the cost model before proposing anything, because it is unusual: the expensive
-part is not compute per action, it is **holding a live socket**.
-
-| Cost | Driver | Measured in |
-|---|---|---|
-| Connection hold | homes connected concurrently, from order 14's per-connection heap | memory, and therefore instance count |
-| SSE streams | dashboards and wall displays left open | same |
-| Agent turns | LLM tokens per voice or chat interaction | the existing brain-router costs |
-| Speech | ASR and TTS per utterance | the existing lanes |
-| Relay (order 10) | sustained outbound sockets from houses | memory plus egress |
-
-A house that is connected and idle costs real money whether or not anyone speaks to it. Any
-entitlement design that meters only actions will be wrong.
-
-## The dimensions to meter
-
-Propose limits per tier along these, with the measured cost behind each:
-
-| Dimension | Why |
-|---|---|
-| Homes connected | the primary cost driver |
-| Household members per home | the enterprise dimension |
-| Concurrent live streams per home | wall displays |
-| Voice minutes per month | the ASR and TTS lane cost |
-| Agent turns per month | LLM cost, probably already metered by an existing counter; find it before adding one |
-| Action-log retention | order 15's decision, and a natural enterprise upsell |
-| Relay connections | if order 10 landed |
-
-## The rules the limits must obey
-
-These are product commitments, not implementation details. Write them into the code as comments
-and into the docs as text:
-
-1. **A limit never blocks a safety action.** Over quota, a user can still lock their door, close
- their garage and arm their alarm. The safe direction is always free. A product that will not
- let someone lock up because they hit a quota is indefensible.
-2. **Downgrading never silently disconnects a house.** It marks connections over the new limit as
- inactive with a clear explanation and lets the user choose which to keep. Nothing is deleted.
-3. **A quota is shown before it is hit**, on the manage surface, with the reset date.
-4. **Enterprise limits are configurable per account**, not hardcoded, because that is what the
- sales conversation is.
-5. **Nothing about the gate is a paid feature.** Confirmations, the audit log's integrity, and the
- role system's safety properties exist on every tier. Selling safety would be wrong and would
- also make the free tier a liability.
-
-Rules 1 and 5 are the ones to fight for if anyone pushes back.
-
-## Proposal, not decision
-
-Produce a table of tiers and numbers, each justified by the measured cost from order 14, and put
-it in your report as a proposal for the owner. Implement the **mechanism** fully so that applying
-a number is a config change. Do not invent a price and ship it.
-
-Also state, honestly, whether the free tier can carry a connected home at all given the measured
-per-connection cost, and if it cannot, what the free experience is instead (a session-scoped
-connection that closes when the tab does, is the obvious candidate).
-
-## Tasks
-
-| # | Task |
-|---|---|
-| 1 | An entitlement resolver for the lane, reading the existing tier system. `api/_lib/home/entitlements.js`. |
-| 2 | Enforcement at the acquisition point (connections), the stream point, and the metered lanes. Never at the safe-action point. |
-| 3 | Usage counters, reusing the platform's existing counter if one exists (`grep -rn "usage_events\|UsageCounter" api/ packages/`). |
-| 4 | The downgrade path: mark inactive, explain, let the user choose. |
-| 5 | The manage-surface quota display with the reset date. |
-| 6 | Per-account overrides for enterprise. |
-| 7 | Tests: the safe-action exemption, the downgrade path, counter accuracy, and per-account overrides. |
-| 8 | The proposal table and the single owner message. |
-
-## Definition of done
-
-- [ ] Over quota, `lock`, `close_cover` and `alarm_arm_away` still execute. Three transcripts. This is the most important line in this order.
-- [ ] Over quota, a new home connection is refused with a designed message naming the limit and the upgrade path, not a 500.
-- [ ] A simulated downgrade marks the excess connections inactive, deletes nothing, and presents the choice. Recorded.
-- [ ] The quota display shows real usage and a real reset date, screenshotted.
-- [ ] A per-account override raises a limit without a code change. Prove it.
-- [ ] Usage counters match reality: run 20 actions and 3 voice turns, then read the counters. Paste both.
-- [ ] No new entitlement system was created: `grep` shows the lane reads `api/_lib/account-tier.js` (and the $THREE ladder where relevant) rather than its own table.
-- [ ] The gate is unaffected by tier. Prove a free-tier account still gets confirmations and a full audit log.
-- [ ] The proposal table names a measured cost per dimension.
-- [ ] `npx vitest run --root .` shows no new failures.
-- [ ] `npm run check:rules -- --paths <your files>` clean.
-
-## Never blocked
-
-| Blocker | Do this |
-|---|---|
-| The price is not decided | You are not deciding it. Build the mechanism, propose numbers with the measured costs, and batch the approval. Ship with generous defaults behind a config value. |
-| The measured per-connection cost makes a free tier look impossible | Say so plainly and propose the session-scoped free experience. An honest constrained free tier beats an unsustainable generous one. |
-| Someone proposes gating confirmations or the audit log behind a paid tier | Refuse, and leave the reason in the code. Safety is not an upsell. |
-| An existing usage counter is hard to reuse | Reusing it is still right. A second counter means two numbers that disagree, and the one on the invoice will be wrong. |
-| $THREE holder tiers complicate the matrix | `api/_lib/three-tier.js` already models the ladder and the discount. Read it and apply it; do not model holding twice. |
-
-## Report format
-
-1. The three safe-action-over-quota transcripts.
-2. The refusal message and the downgrade recording.
-3. The quota display screenshot.
-4. The override proof and the counter accuracy check.
-5. The grep proving no new entitlement system.
-6. The proposal table with measured costs, and the single owner message.
-7. Anything in this file you found to be wrong, and what you changed it to.
-
-## Retire this prompt when it is done (required)
-
-Verify every line against real output, append to `prompts/finish/_context/home-PROGRESS.md`, commit with
-explicit paths, and delete this file in that commit:
-
- git rm prompts/finish/313-home-19-plans-entitlements.md
-
-If the price approval is the only outstanding step, leave the file in place and name the owner
-action. Never delete it on a partial.
diff --git a/prompts/finish/314-home-20-launch-readiness.md b/prompts/finish/314-home-20-launch-readiness.md
deleted file mode 100644
index 98e5d522b..000000000
--- a/prompts/finish/314-home-20-launch-readiness.md
+++ /dev/null
@@ -1,158 +0,0 @@
-# 20. Launch readiness: the go/no-go
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`. Read [00-CONTEXT.md](_context/home-00-CONTEXT.md) first. Run this **last**, after
-every other order in the campaign has been retired, and run it again after any change to the lane.
-
-## Binding operating clause
-
-Finish 100%. Never end with a question or an unexecuted plan. All CLAUDE.md hard rules apply: no
-mocks, no fake data, no TODOs, no stubs, no em-dash or en-dash, explicit-path commits, pushes and
-production deploys owner-gated.
-
-**This order verifies, it does not build.** If a check fails, either fix it here (small) or
-re-open the owning order (large) and say which. A go/no-go that finds nothing because it looked
-at nothing is worse than no gate at all.
-
----
-
-## Step 0: re-derive the current state
-
-```bash
-ls prompts/finish/home-*.md # what is still open. Every open order is a no-go input.
-npx vitest run --root .
-npx playwright test tests/e2e/ 2>&1 | tail -20
-npm run gate 2>&1 | tail -30
-npm run audit:docs && npm run check:claude && npm run check:docs-search
-git log --oneline -20
-```
-
-## The go criteria
-
-Every line is mechanically checkable and every line needs the command output pasted. A line you
-cannot verify is a no-go, not a judgement call.
-
-### Correctness and safety
-
-- [ ] Order 11's eleven security checks all pass, re-run today.
-- [ ] The confirmation-integrity invariant holds: zero rows anywhere in `home_action_log` with `guarded = true`, `confirmed_by is null`, `outcome = 'ok'`. Paste the query and the count.
-- [ ] The order 16 e2e journeys all pass, including the three that assert on a real lock's real state.
-- [ ] No home tool schema anywhere exposes `confirmed`. Re-dump and paste.
-- [ ] The prompt-injection regression passes with the real physical assertion.
-- [ ] SSRF refusals still hold, including the DNS-rebinding pin.
-- [ ] The role matrix passes every cell, and a `guest` still cannot confirm.
-
-### Reliability
-
-- [ ] The order 14 chaos scenarios re-run green, or every deviation is explained.
-- [ ] p95 action latency measured today against the published SLO.
-- [ ] A ten-minute live session leaves heap flat.
-- [ ] Ten consecutive runs of the lane's suites, green.
-
-### Operations
-
-- [ ] `/api/healthz` reports the `home` subsystem with real numbers.
-- [ ] All three alerts are wired and each has been fired at least once in a test, with the transcript.
-- [ ] `docs/home-operations.md` exists and every command in it was run.
-- [ ] The rollback path in `docs/ops/gcp-production.md` was walked, at least in a dry run, and the exact commands are in your report.
-- [ ] Any cron the lane added: `npm run check:cron-drift` clean, `vercel.json` cron count matches CLAUDE.md, `npm run check:claude` passes.
-
-### Product completeness
-
-- [ ] Every state enumerated by orders 05, 06, 07, 08 exists. Walk the list and screenshot each. A missing state is a no-go.
-- [ ] No dead paths: every button acts, every link resolves, every reachable state has a way out.
-- [ ] Zero console errors and zero console warnings from lane code, on every surface.
-- [ ] `npm run audit:web` (authed, per `docs/ops/page-audit.md`) clean for every home route.
-- [ ] axe passes on every home route.
-- [ ] `npm run i18n:lint` clean.
-- [ ] 320px, 768px and 1440px correct on every state.
-
-### Data and privacy
-
-- [ ] The order 15 inventory matches the schema exactly, re-checked today.
-- [ ] Account deletion leaves zero rows across every lane table.
-- [ ] The log-scrub proof re-run: zero hits for base URL, token and entity names.
-- [ ] No persisted entity-state history exists.
-
-### Documentation and release
-
-- [ ] `npm run audit:docs` clean.
-- [ ] Every tutorial command re-run today and still working.
-- [ ] `STRUCTURE.md` and `data/pages.json` cover every surface.
-- [ ] The `data/changelog.json` entry is present, plain-language, and validated by `npm run build:pages`.
-- [ ] `docs/smart-home.md` describes what shipped, with no plan-tense claims left.
-
-### The build
-
-- [ ] `npm run gate` passes.
-- [ ] `npm test` passes with no new failures. Do not pipe it through `tail`; that masks the exit code.
-- [ ] `npm run check:rules -- --base <lane start sha> --head HEAD` clean across the whole lane.
-- [ ] `node scripts/check-secrets.mjs --base <lane start sha> --head HEAD` clean.
-- [ ] The `deploy-preflight` subagent has been run and reports the deploy safe.
-
-## The launch decision
-
-Produce a one-page verdict in your report:
-
-1. **Go or no-go**, stated first, in one line.
-2. Every failing check, what it blocks, and which order owns it.
-3. Every accepted residual, with the risk named and the reason.
-4. The measured numbers: p95 latency, state freshness, connection cost, the version matrix.
-5. The rollback plan in three commands.
-6. The single owner message, containing everything owner-gated in one place: the deploy, the npm
- publishes, the add-on repository publish, and the entitlement prices.
-
-## The deploy itself
-
-Owner-gated. Follow the CLAUDE.md deploy runbook **exactly and in order**, and do not hand-run
-its steps out of order:
-
-```
-npm run clean:worktrees -- --apply
-npm run prep:worktree -- --apply
-npm run build:gcp
-npm run deploy:gcp:submit
-npm run deploy:gcp:purge-cdn
-npm run smoke:prod
-```
-
-Then verify: `curl -s https://three.ws/api/version` returns the shipped SHA and revision.
-
-Prepare all of it, run everything up to the submit, and stop for the owner's yes. Do not deploy.
-
-## Post-launch watch, first 48 hours
-
-Write this into `docs/home-operations.md` before launch, not after:
-
-| Hour | Check |
-|---|---|
-| 0 | `/api/healthz` home block, the first real connection, the first real action |
-| 1 | handshake success rate across all tenants; confirmation expiry rate |
-| 6 | heap trend on the API service; subscriber counts |
-| 24 | p95 latency against the SLO; the action-log integrity query |
-| 48 | the full alert set has not fired spuriously; the error budget consumed so far |
-
-## Never blocked
-
-| Blocker | Do this |
-|---|---|
-| An order is still open | Then it is a no-go input. Say so, name the order, and do not paper over it with a judgement call. |
-| A check cannot be run in this environment | Name it, name what access it needs, and mark it explicitly unverified in the verdict. Never mark it green. |
-| Something fails and the fix looks small | Fix it, re-run the whole section, and note it. |
-| Something fails and the fix looks large | Re-open the owning order, put it in the no-go list, and say what it blocks. |
-| Pressure to ship with a failing safety check | There is no version of this where a door that opens without a confirmation ships. That check is a hard no-go with no override. |
-
-## Report format
-
-The one-page verdict above, then the full command output for every checked line, grouped by
-section. Nothing summarised, nothing claimed without its output.
-
-## Retire this prompt when it is done (required)
-
-This order is **standing**: re-run it before every deploy of the lane. Delete it only when the
-campaign is retired and the lane is in steady operation, in the same commit that removes
-[00-CONTEXT.md](_context/home-00-CONTEXT.md):
-
- git rm prompts/finish/314-home-20-launch-readiness.md
-
-Until then, leave it in place and record each run in `prompts/finish/_context/home-PROGRESS.md`.
diff --git a/prompts/finish/901-production-100-01-ship-readiness.md b/prompts/finish/901-production-100-01-ship-readiness.md
deleted file mode 100644
index 10e5397da..000000000
--- a/prompts/finish/901-production-100-01-ship-readiness.md
+++ /dev/null
@@ -1,115 +0,0 @@
-# P100-01: Ship readiness, the standing deploy order
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`, or say "execute `prompts/finish/901-production-100-01-ship-readiness.md`".
-Read [00-INDEX.md](_context/production-100-00-INDEX.md) and `CLAUDE.md` first.
-
-**This order is standing, not one-shot.** Run it whenever production trails `main` and the
-owner wants (or has approved) a ship. It retires only when the campaign's Definition of 100%
-holds and no further deploys are pending; until then, finishing a run of it means logging in
-[PROGRESS.md](_context/production-100-PROGRESS.md), not deleting the file.
-
-## Binding operating clause
-
-1. Finish 100% of what this machine can do. Never end with a question or an unexecuted plan.
-2. The deploy submission itself is CLAUDE.md stop-and-ask gate 2. Everything before it
- (cleaning, building, verifying) and after it (purge, smoke, health reads) is yours to do
- without asking. If the owner's instruction this session already approves shipping, gate 2
- is satisfied and you run the whole chain.
-3. Hard rules: no mocks, no shortcuts, no `tail` piped onto test commands (it masks exit
- codes), explicit-path commits only, no em-dash characters in anything you write.
-
-## Step 0: re-derive the state (do all of these, trust none from memory)
-
-```bash
-curl -s https://three.ws/api/version # prod commit vs:
-git rev-parse --short main # local main
-curl -s https://three.ws/api/healthz | head -c 2000
-npm run clean:worktrees # report first; add --apply to act
-df -h /workspaces | tail -1 # the 2026-08-04 disk-full trap
-uptime # load average; concurrent agents thrash this box
-npm run db:status # migrations pending behind the deploy?
-```
-
-If production already equals `main` and healthz is clean, this run is a no-op: log that
-measurement in [PROGRESS.md](_context/production-100-PROGRESS.md) and stop.
-
-## Tasks
-
-1. **Reclaim disk.** `npm run clean:worktrees --apply`. Stale deploy worktrees filled the
- disk to 100% on 2026-08-04 and the failure surfaces as a misleading mid-checkout error.
-2. **Get the repo green enough to ship.** `npm run gate` must exit 0. If it is red on
- something in [../fix-queue/](_context/fix-queue-00-INDEX.md), fix it via that order's file (and
- retire that file per the standard) rather than masking it. A red caused by another agent's
- in-flight work: capture the baseline, coordinate or route around it, never ship over it.
-3. **One clean `npm test` on a quiet box.** This has been owed since 2026-08-08: every prior
- attempt died to load average 200+ from concurrent agents, not to code. If the box is
- thrashing (load average well above 16), wait for a quiet window or coordinate; a SIGTERM
- kill at load 250 tells you nothing. Do not pipe through `tail`. Record the summary line.
-4. **Run the preflight agent.** Launch the `deploy-preflight` subagent and act on every
- finding. It checks the load-bearing build order, worktree artifacts, service-account pins,
- and changelog wiring.
-5. **Build in a clean worktree at a pinned SHA**, exactly per the CLAUDE.md deploy runbook:
- `git worktree add --detach /workspaces/.deploy-wt <pinned SHA>`, hardlink `node_modules`,
- `chat/node_modules`, and `character-studio/build` with `cp -al`, copy `.env`, then
- `npm run build:gcp` inside it. Pin the SHA explicitly; `main` has moved 80+ commits during
- a single build before. Never hand-run the chain out of order.
-6. **Present the ship (gate 2), then execute on yes.** Render for the owner: the pinned SHA,
- a one-paragraph summary of what ships (read `git log <prod-sha>..<pinned-sha> --oneline`
- yourself and summarize by theme, do not paste 80 lines), and the single command:
- `gcloud builds submit --config server/cloudbuild.yaml --region us-central1 --project aerial-vehicle-466722-p5 --substitutions=SHORT_SHA=manual$(date +%s)`
- from the worktree. If the owner already approved this session, run it now.
-7. **After the submit lands:** `npm run deploy:gcp:purge-cdn` (synchronous, never `--async`),
- then verify: `curl -s https://three.ws/api/version` shows the pinned SHA,
- `npm run smoke:prod` passes, healthz subsystems are no worse than the step 0 read, and
- any deploy-armed surfaces flipped (whatever this map's other orders were waiting on; as of
- authoring: `/event.json` serving JSON, `/api/fact-check-benchmark` answering
- `source: "database"`, the discovery paging fix live).
-8. **Clean up.** `git worktree remove --force /workspaces/.deploy-wt`. Log the run in
- [PROGRESS.md](_context/production-100-PROGRESS.md): prod SHA before and after, test summary, smoke result,
- anything that flipped.
-
-## Definition of done (per run)
-
-- [ ] `curl -s https://three.ws/api/version` reports the pinned SHA on a new revision.
-- [ ] `npm run smoke:prod` exit 0.
-- [ ] `npm run gate` exit 0 at the shipped SHA.
-- [ ] One clean full `npm test` recorded, with its summary line, on a box under load.
-- [ ] Post-ship healthz no worse than pre-ship, and every regression explained.
-- [ ] Deploy worktree removed; run logged in [PROGRESS.md](_context/production-100-PROGRESS.md).
-
-## Never blocked
-
-| Blocker | Resolution |
-|---|---|
-| Owner has not said yes to the ship | Do everything up to and including step 6's rendering, so the ship is one command, then log and stop. That is a finished run, not a blocked one. |
-| `gcloud` missing or auth dead | `export PATH="$HOME/google-cloud-sdk/bin:$PATH"`; revive auth with `gcloud auth login --no-launch-browser` fed through a fifo. Verify with a real read (`gcloud run services describe three-ws-api --region us-central1 --project aerial-vehicle-466722-p5 --format='value(status.latestReadyRevisionName)'`), not with `auth list`. |
-| Build dies `No space left on device` | Step 1 was skipped. Run it. |
-| Build OOMs or exits 143 | Load, not code. Retry in a quiet window; check `uptime` first. |
-| `check:dist` fails on a missing bundle | The build chain ran out of order or `build:vercel` was hand-run as the frontend build. Re-run `npm run build:gcp` from clean. |
-| `db:check` exits 4 | Pending migrations. Read `npm run db:status`, understand every pending migration (they all apply at once), then `npm run db:migrate` only when they are all intended. |
-| Post-deploy pages look stale | The CDN purge was skipped or run async. Re-run `npm run deploy:gcp:purge-cdn`. |
-
-## Report format
-
-Prod before -> after (SHAs, revision), what shipped (themes), test summary line, smoke result,
-healthz delta, worktrees cleaned, and the one thing most worth shipping next.
-
-## Retire this prompt when it is done (required)
-
-1. Verify every Definition of done line against actual command output in front
- of you. Never claim a line you did not verify.
-2. Record the outcome in this campaign's PROGRESS or INDEX file if it has one.
-3. Commit with explicit paths and a subject that describes the diff (house
- style: type(scope): what changed and why a reader cares), and delete this
- prompt file in that same commit:
-
- git rm prompts/finish/901-production-100-01-ship-readiness.md
-
- A finished order left on disk reads as open work to the next agent, so the
- shrinking directory is the campaign's progress ledger.
-
-If a line genuinely cannot pass inside this session (an external party must
-respond, or an owner-gated action is the final step), finish everything else,
-leave this file in place, and state exactly which line remains and who owns it.
-Never delete this file on a partial.
diff --git a/prompts/finish/903-production-100-04b-fact-check-publish-run.md b/prompts/finish/903-production-100-04b-fact-check-publish-run.md
deleted file mode 100644
index 7d24022e6..000000000
--- a/prompts/finish/903-production-100-04b-fact-check-publish-run.md
+++ /dev/null
@@ -1,165 +0,0 @@
-# P100-04b: The `mixed` fix is written and tested. Nothing has measured it yet.
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`, or say "execute `prompts/finish/903-production-100-04b-fact-check-publish-run.md`".
-Read [00-INDEX.md](_context/production-100-00-INDEX.md) and `CLAUDE.md` first.
-
-This is the remainder of P100-04, whose code half landed on 2026-09-02. That order's
-diagnosis and fix are done and unit-tested; what is left needs credentials this workspace
-does not have and a deploy only the owner can approve.
-
-## Status as of 2026-09-09, read this before Step 0
-
-The title above is out of date and Step 0's exit test is misleading: the fix HAS been
-measured, and `/api/fact-check-benchmark` still shows the 2026-08-10 run **because the new
-one was correctly withheld**, not because nobody ran it. Three passes that day are written up
-in [PROGRESS.md](_context/production-100-PROGRESS.md); the short version:
-
-- The fix is deployed and reachable. `mixed` went from 0% to 10% and produced the verdict six
- times, against zero in the published run.
-- Two other classes regressed, so the run was not published. The cause is not the calculus:
- `groundedSearch` is denied project-wide by a GCP billing hold, every source is Wikipedia,
- and 15 of 38 claims saw zero stance-bearing evidence.
-- **The only thing still blocking this order is OWNER-ACTIONS row 20**, clearing that hold.
- Nothing on this machine routes around a dunning denial, and no amount of re-running changes
- the answer.
-
-So do not open this order expecting to publish a number. Re-check row 20 first. If it is
-still open, the useful work is whatever improves the evidence layer for the run that happens
-after it clears, and PROGRESS records what has already been tried, measured and deliberately
-not shipped.
-
-## Binding operating clause
-
-1. Finish 100%. Never end with a question or an unexecuted plan. Delete this file when the
- definition of done is verified.
-2. **Never publish a number you would not defend.** Respect the degraded-check guard and
- the 10% error-rate ceiling in `api/_lib/fact-check-benchmark.js`. Never weaken either to
- make a run publishable. A run that regresses gets diagnosed, not published.
-3. Hard rules: no mocks, no fixture-fitting, explicit-path commits, no em-dashes.
-
-## What already shipped (do not redo it)
-
-- **`partial` is a fourth stance.** `agents/fact-checker/src/llm-verdict.js` exports
- `STANCES` and its rubric asks for `partial` when a source finds the claim true in one
- respect and wrong, overstated, or only conditionally true in another. Before this, the
- rubric explicitly told the model to collapse such a source onto one side, which made the
- `mixed` verdict unreachable from evidence that agrees with itself.
-- **`computeVerdict` counts partial weight** as stance-bearing evidence that takes neither
- side (`api/x402/fact-check.js`). With zero `partial` sources its output is identical to
- the old function, which is the anti-seesaw guarantee.
-- **The LLM response reader no longer fails silently.** `extractJsonArray` replaces a
- non-greedy regex that stopped at the first `]`; both stages now report
- `stance extraction unreadable` / `query generation unreadable` instead of returning a
- fabricated all-neutral `insufficient`.
-- Tests: `tests/api/fact-check-verdict.test.js` (synthetic stance distributions, clear-cut
- pins) and `tests/fact-check-degradation.test.js` (extraction shapes, degradation
- contract). Docs: `docs/fact-check.md`, `agents/fact-checker/README.md`. Changelog: two
- 2026-09-02 entries.
-
-## Step 0: re-derive the state
-
-```bash
-curl -s https://three.ws/api/version # is the fix live yet?
-curl -s https://three.ws/api/fact-check-benchmark | head -c 1200
-git log --oneline -5 -- api/x402/fact-check.js agents/fact-checker/src/
-npx vitest run tests/api/fact-check-verdict.test.js tests/fact-check-degradation.test.js
-```
-
-If `/api/fact-check-benchmark` already reports a run generated after the fix landed with
-`mixed` above zero, this order is spent: log the outcome in
-[PROGRESS.md](_context/production-100-PROGRESS.md) and delete this file.
-
-## Tasks
-
-1. **Get an LLM lane.** The benchmark's in-process runner needs one, and on 2026-09-02
- this workspace had none: no `GROQ_API_KEY` / `OPENROUTER_API_KEY` / `NVIDIA_API_KEY` in
- `.env` or `.env.local`, and `gcloud` refused every call with "Reauthentication failed.
- cannot prompt during non-interactive execution", which also takes out the Vertex Gemini
- anchor (`GOOGLE_CLOUD_PROJECT` plus ADC). Restore `gcloud` auth first; that alone
- unlocks Vertex, and it unlocks reading the rest off the service env
- (`gcloud run services describe three-ws-api --region us-central1 --project
- aerial-vehicle-466722-p5 --format=yaml`). Search needs no key: Wikipedia and DuckDuckGo
- are keyless rungs, though a `BRAVE_API_KEY` / `TAVILY_API_KEY` materially improves the
- evidence the verdict sees.
-
- **Do not plan on the keyless LLM floor.** It exists and it is real, but it is two rungs,
- not three, and from this workspace's shared egress IP it answered 0 of 8 probes on
- 2026-09-02: OVH 429 and Pollinations 429 on every attempt across 64 seconds, with one
- isolated success after a long idle. LLM7 was the third keyless rung and is not one any
- more (llm7.io retired its anonymous tier; every unauthenticated call is 401
- `invalid_api_key`, the `unused` token its docs used to accept included), so it is gated
- on `LLM7_API_KEY` now. A 40-claim run is 80 LLM turns; the keyless floor will not carry
- it. Get a real key.
-2. **Check the production chain is not itself degraded before spending a run.** On
- 2026-09-02 a live free-lane check returned
- `query generation unavailable: openai 429 billing_not_active`, meaning every free rung
- had already failed and the paid backstop is dead. A benchmark run in that state measures
- provider availability, and the guard will refuse it. Fix the lane first
- (`docs/ops/llm-lanes.md`).
-3. **Run the suite in-process, cache disabled, and capture per-claim detail:**
- ```bash
- FACT_CHECK_DETAIL_FILE=/tmp/fc-detail.json \
- node --env-file=.env scripts/fact-check-benchmark.mjs --in-process
- ```
- That writes `data/_generated/fact-check-benchmark.json` without publishing. Compare its
- `by_class` against the last published run (`mixed` must leave zero, no class may
- regress) and read `/tmp/fc-detail.json` for the per-source stances behind any class that
- moved the wrong way.
-
- **Fidelity trap, and it decides whether step 4 is honest.** The published 2026-08-10 run
- was produced by production, whose search chain leads with Vertex-grounded Google Search.
- A local run without `GOOGLE_CLOUD_PROJECT` and ADC (or a Brave/Tavily/Exa/Serper key)
- silently falls to the keyless rungs, and the evidence is visibly worse: probing the
- `mixed` claims here on 2026-09-02 returned five Wikipedia pages per claim, among them
- "Aunty Donna's Coffee Cafe" for "Coffee is bad for your health" and "Zootopia" for the
- tongue map. That tier is fine for an A/B of the CALCULUS against itself, where both arms
- see the same evidence. It is not the product's accuracy, and publishing it as the
- headline number would understate a chain the public number is supposed to describe.
- Before `--publish`, confirm the run actually used the grounded search rung.
-4. **Publish only an improved, non-degraded run:** re-run with `--publish`. The live
- `/api/fact-check-benchmark` reads the DB row first, so this reaches the public page
- without a deploy, but the CHAIN itself only improves once the code is deployed, so
- sequence it after the ship.
-5. **Verify the page in a real browser:** `/fact-check` renders the run, denominators, and
- the per-class and per-difficulty tables.
-6. **Log the outcome** in [PROGRESS.md](_context/production-100-PROGRESS.md) with the before/after
- per-class table, and add a `data/changelog.json` entry (tag `improvement`) naming the
- new headline number.
-
-## Definition of done
-
-- [ ] A real, non-degraded in-process run exists with a before/after per-class table.
-- [ ] `mixed` is above zero and no class regressed; if one did, the diagnosis is written
- down and the run is NOT published.
-- [ ] An improved run published to the DB and rendered on `/fact-check` in a browser.
-- [ ] Changelog entry written, outcome logged in PROGRESS.md, this file deleted.
-
-## Never blocked
-
-| Blocker | Resolution |
-|---|---|
-| No LLM key anywhere on the machine | Restore `gcloud` auth for the Vertex anchor, then read the service env. If a key genuinely exists nowhere, that is one line for OWNER-ACTIONS naming the single missing variable, not a reason to leave the rest undone. Do not spend the session retrying the keyless rungs: they were 0/8 from this egress IP on 2026-09-02 (task 1). |
-| Tempted to publish a run measured on the keyless search tier | Do not. It measures a chain the public number does not describe (task 3). Report the A/B, leave the published run alone, and say what is missing. |
-| The deploy is owner-gated | Prepare it so the ship is one command (`npm run deploy:gcp:full` after `npm run prep:worktree -- --apply`) and say so. Publishing the run does NOT need the deploy, but running it against the OLD chain measures the old code, so do not publish that as the fix's number. |
-| Accuracy improves on `mixed` but a neighbour regresses | That is the seesaw the calculus was designed against. Read `/tmp/fc-detail.json` for the regressed claims' stances: a `partial` where the source plainly affirms or plainly refutes is a rubric problem, not a calculus one. |
-| The run is refused as degraded | Working as designed. Fix the lane, then re-run. Never lower `MAX_ERROR_RATE`. |
-
-## Retire this prompt when it is done (required)
-
-1. Verify every Definition of done line against actual command output in front
- of you. Never claim a line you did not verify.
-2. Record the outcome in this campaign's PROGRESS or INDEX file if it has one.
-3. Commit with explicit paths and a subject that describes the diff (house
- style: type(scope): what changed and why a reader cares), and delete this
- prompt file in that same commit:
-
- git rm prompts/finish/903-production-100-04b-fact-check-publish-run.md
-
- A finished order left on disk reads as open work to the next agent, so the
- shrinking directory is the campaign's progress ledger.
-
-If a line genuinely cannot pass inside this session (an external party must
-respond, or an owner-gated action is the final step), finish everything else,
-leave this file in place, and state exactly which line remains and who owns it.
-Never delete this file on a partial.
diff --git a/prompts/finish/904-gcp-credits-05-catalog-animation-seeding.md b/prompts/finish/904-gcp-credits-05-catalog-animation-seeding.md
deleted file mode 100644
index 4c3412eb9..000000000
--- a/prompts/finish/904-gcp-credits-05-catalog-animation-seeding.md
+++ /dev/null
@@ -1,149 +0,0 @@
-# GCP-05: Bulk catalog and animation seeding on the self-hosted GPU fleet
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`, or say "execute `prompts/finish/904-gcp-credits-05-catalog-animation-seeding.md`".
-It is complete on its own. Also read `prompts/finish/_context/gcp-credits-README.md` and `CLAUDE.md`.
-
-## Binding operating clause
-
-1. Finish 100%. Never end with a question, a plan you did not execute, or "should I proceed?".
-2. Blockers have pre-answered routes at the bottom. Use them and keep going.
-3. CLAUDE.md hard rules: no mocks, no fake data, no TODO comments, no em-dash or en-dash
- characters. Stage explicit paths only; other agents share this worktree.
-4. GPU spend on the self-hosted fleet is pre-approved (credits). Curation beats volume: a
- catalog of 5,000 junk meshes is worse than today's catalog.
-
-## Mission
-
-Turn credits into durable platform content: a much larger curated avatar catalog and a
-generated text-to-motion clip library, both quality-gated, both browsable at the new scale.
-
-## Step 0: re-derive current state (trust nothing below)
-
-```bash
-gcloud run services list --region us-central1 --project aerial-vehicle-466722-p5 \
- --format="value(metadata.name,status.conditions[0].status)" | grep -E "trellis|text2motion|rig|hunyuan"
-gcloud run services describe three-ws-api --region us-central1 \
- --project aerial-vehicle-466722-p5 --format=yaml | grep -E "GCP_TEXT2MOTION_URL|MODEL_TRELLIS_URL|FORGE_SELFHOST_PRIMARY|SEED_CRON_BATCH"
-sed -n '1,40p' api/cron/forge-seed-cron.js
-node -e "const d=require('./data/motion-prompts.json');console.log(Object.keys(d), d.prompts?.length)"
-
-# clip library: total, and how many are NOT the Mixamo import
-curl -s "https://three.ws/api/animations/library" | python3 -c \
- "import sys,json;d=json.load(sys.stdin);n=[c['name'] for c in d['clips']];print('clips',d.get('total'),'non-mixamo',sum(1 for x in n if not x.startswith('mx-')))"
-
-# catalog size, through the repo's own db helper (psql is not installed here)
-cat > /tmp/catalog-count.mjs <<'EOF'
-import fs from 'node:fs';
-if (!process.env.DATABASE_URL) {
- const m = fs.readFileSync('/workspaces/three.ws/.env', 'utf8').match(/^DATABASE_URL=(.*)$/m);
- if (m) process.env.DATABASE_URL = m[1].replace(/^["']|["']$/g, '');
-}
-const { sql } = await import('/workspaces/three.ws/api/_lib/db.js');
-console.log('avatars', (await sql`select count(*)::int as n from avatars`)[0].n);
-console.log('last 7d', (await sql`select count(*)::int as n from avatars where created_at > now() - interval '7 days'`)[0].n);
-EOF
-node /tmp/catalog-count.mjs
-```
-
-**Measured 2026-08-01, re-measure before trusting it:** 18,622 avatars in the catalog with 104
-added in the previous 7 days, so the seed cron runs but at roughly 15 a day. The clip library
-serves 2,874 clips and **every one is the `mx-` Mixamo import: zero generated clips exist**, so
-section B below is entirely open. `data/motion-prompts.json` (the section B prompt library),
-`api/cron/forge-seed-cron.js`, `api/_lib/seed-prompts.js`, `api/marketplace/animations.js`,
-`api/x402/animation-download.js` and `api/animations/library.js` all exist;
-`scripts/gcp/seed-avatars.mjs` does not.
-
-That shape sets the priorities: the cron is alive but slow and unproven on quality, and the
-generated-motion half has never run. Do not spend the session re-confirming that the cron
-exists.
-
-## Tasks
-
-### A. Avatar catalog seeding (target: several thousand curated avatars)
-
-1. **Point the seed cron at the self-hosted lane** when `FORGE_SELFHOST_PRIMARY=1`, and make
- cadence and batch size env-tunable (`SEED_CRON_BATCH`, default equals current behavior, so
- the change is a no-op when unset).
-2. **Broaden `api/_lib/seed-prompts.js`**: professions, fantasy, sci-fi, styles, with attention
- to what makes a good rigged avatar (humanoid, front-facing, full-body). Keep every prompt
- coin-neutral: no crypto project other than $THREE, ever, in committed content.
-3. **Automated quality gate before publish**: render a thumbnail, run it through the vision
- check (`api/vision.js` / `api/_lib/vision.js`) for humanoid, complete body, not a blob, plus
- mesh sanity (vertex-count bounds, has texture). Rejected assets stay in R2 under a
- `forge/rejected/` prefix with the reason, for tuning. Track and report accept rate.
-4. **Rig the keepers** through the rig worker so catalog entries are animation-ready, and spot
- check a sample in the viewer with idle and walk clips playing. A catalog avatar that T-poses
- is half-built.
-5. **Resumable batch runner** `scripts/gcp/seed-avatars.mjs`: checkpoint file, safe to re-run,
- parallelism tuned to the deployed GPU count, asserts the backend actually used and aborts on
- fallthrough to any paid third-party lane. Run one real batch of 500 end to end, report
- accept rate and cost per accepted asset, then continue only if quality holds.
-
-### B. Animation library seeding (target: several hundred curated clips)
-
-6. **Use `data/motion-prompts.json`** as the prompt library (extend it if categories are thin;
- it is data, never hardcoded).
-7. **Generate, retarget, publish**: clips through `GCP_TEXT2MOTION_URL`, converted to the exact
- canonical-skeleton clip JSON the library already serves (inspect a real clip from
- `GET /api/animations/library` first and match it byte-shape), uploaded under a distinct
- prefix, manifest updated. Quality gate: sane duration, no NaN keyframes, plays on the default
- rig without foot-sliding. Spot-check a sample visually.
-8. **List the generated set** in `api/marketplace/animations.js` under a platform creator
- identity, priced consistently with existing listings, downloadable through the existing x402
- route. Implement a rotating free subset mechanism and record the policy choice you made.
-
-### C. Scale-proofing
-
-9. At 5 to 10 times the current catalog size, verify `/gallery`, `/dashboard/avatars`,
- `/animations` and the marketplace list endpoints paginate rather than fetch-all, that
- thumbnails lazy-load, and that API responses stay bounded. Fix what breaks: pagination,
- indexes on the avatars table, manifest sharding. Test at realistic counts, not ten items.
-
-## Definition of done
-
-- [ ] Seed cron on the self-hosted lane and env-tunable; quality gate live with a measured
- accept rate.
-- [ ] First avatar batch published: rigged, animated in the viewer, browsable in the gallery.
-- [ ] Generated motion clips live in the library and the marketplace, sample visually verified.
-- [ ] Scale test at target counts passes on all four surfaces, with the fixes committed.
-- [ ] Cost per accepted asset and accept rate recorded in `docs/gcp-credits.md`.
-- [ ] `npm test` green; `git diff` reviewed (no esbuild-mangled `api/` files: check `head -1`
- for `__defProp`).
-- [ ] `data/changelog.json` entries for the bigger catalog and the new generated-animation
- collection, in plain language.
-- [ ] `npm run check:rules -- --paths <files you touched>` clean.
-
-## Never blocked (pre-answered)
-
-| Blocker | Do this |
-|---|---|
-| A self-host lane is not Ready | Deploy or scale it (credits approved) using the bind-first pattern in `workers/model-trellis/main.py`. If a lane genuinely cannot come up, seed with the lanes that are up and report which one is missing. |
-| Accept rate is poor | Stop scaling, tune prompts and gates first, and report the numbers honestly. Volume without curation fails this work order. |
-| A batch would fall through to a paid third-party lane | Assert the backend in the batch script and abort. Never let bulk spend leave the credits. |
-| Database migration needed for an index | `npm run db:status` first (it previews), then `npm run db:migrate` applies every pending migration immediately with no dry run. Read the pending list before running it. |
-| Vision QA rejects everything | Inspect ten rejected thumbnails yourself before touching the threshold; a broken render pipeline looks exactly like a strict gate. |
-
-## Report format
-
-Counts published, accept rates, spend, cost per accepted asset, the four scale-test results,
-and any single remaining owner action. No recap of this file.
-
-## Retire this prompt when it is done (required)
-
-1. Verify every Definition of done line against actual command output in front
- of you. Never claim a line you did not verify.
-2. Record the outcome in this campaign's PROGRESS or INDEX file if it has one.
-3. Commit with explicit paths and a subject that describes the diff (house
- style: type(scope): what changed and why a reader cares), and delete this
- prompt file in that same commit:
-
- git rm prompts/finish/904-gcp-credits-05-catalog-animation-seeding.md
-
- A finished order left on disk reads as open work to the next agent, so the
- shrinking directory is the campaign's progress ledger.
-
-If a line genuinely cannot pass inside this session (an external party must
-respond, or an owner-gated action is the final step), finish everything else,
-leave this file in place, and state exactly which line remains and who owns it.
-Never delete this file on a partial.
diff --git a/prompts/finish/916-roadmap-native-widgets.md b/prompts/finish/916-roadmap-native-widgets.md
deleted file mode 100644
index 96b05a566..000000000
--- a/prompts/finish/916-roadmap-native-widgets.md
+++ /dev/null
@@ -1,122 +0,0 @@
-# RM-WIDGETS: Native widgets, the agent on the home screen and the desktop
-
-**Status 2026-09-09: all four tasks are built, and the 2026-09-04 worker fix is
-deployed** (production `/glance-sw.js` is still byte-identical to
-`public/glance-sw.js` at revision 00420). **The 2026-09-09 card fix is not
-deployed**, so a pinned Windows slot renders nothing today: production serves
-commit `880bdcef8` (2026-09-08 18:57Z) and the fix landed in `fe05babf6`
-(2026-09-09 05:41Z). That is now a one-command fact rather than a guess, via
-`npm run check:windows-widget:live` (added 2026-09-09), which binds the deployed
-card with the deployed worker in the board's own engine and currently exits 1.
-Re-verified 2026-09-09 19:44Z against commit `880bdcef8` (revision 00420): the
-11 offline checks and the 15 Apple checks pass, and the live probe's 5 findings
-are all the one undeployed card. That probe is now a registered guard
-(`data/guards.json`, `docs/guards.md`, `/guards`) rather than a command only
-this file mentions.
-**This file stays because two Definition of done lines cannot be closed from this
-machine** (one needs a Windows 11 machine, one needs an Apple Developer account),
-and the campaign's rule is to leave a work order in place until every line of it
-passes. Read "What remains", not the history above it.
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`. Also read `prompts/finish/_context/roadmap-00-README.md` and `CLAUDE.md`.
-
-## What shipped
-
-The card is one server-rendered bitmap plus a small JSON contract, and every
-platform is a rendering job against it. Product documentation:
-[`docs/native-widgets.md`](../../docs/native-widgets.md). Wire format:
-[`specs/GLANCE_CARD.md`](../../specs/GLANCE_CARD.md).
-
-| Task | Where it lives | State |
-| --- | --- | --- |
-| 1. The card endpoint, two encodings | `api/glance/card.js`, `api/glance/mine.js`, `api/glance/token.js`, `api/glance/template.js` | Live in production. Verified 2026-09-03: `mine?format=png` answers `200 image/png` unauthenticated with `x-glance-state: signed-out`, `token` answers `401` with no session, and a real mint on the QA account fetched every size, theme and scale. |
-| 2. Android home screen widget | `solana-mobile/android-overlay/`, applied by `scripts/apply-overlay.mjs` inside `build-apk.sh` | In a signed APK (1.1.0, versionCode 2). All five emulator checks recorded in `solana-mobile/docs/CHECKLIST.md` section 7b. |
-| 3. Windows 11 widget | `widgets` member in `vite.config.js`, `public/glance-sw.js`, `api/glance/template.js` | Live: production `manifest.webmanifest` carries the member and `/glance-sw.js` serves 200. Guarded by `npm run check:windows-widget` (offline, in `npm run gate`) and `npm run check:windows-widget:live` (deployed origin). The card fix of 2026-09-09 is not deployed yet, and the live guard fails on exactly that. |
-| 4. macOS and iOS WidgetKit | `apple/` (shared `GlanceKit/`, extension `GlanceWidget/`, Mac app `macos/`), plus a `GlanceWidgetExtension` target in `ios/native/App/App.xcodeproj` | Written, wired, and guarded by `npm run check:apple-widget`. Not built: no Mac, and no signing identity. |
-
-The one live number is **moves in the last 24 hours**, computed from the agent's
-own action log. The alternatives considered were earnings (empty for most
-accounts, so most widgets would read zero) and reputation (moves too slowly to
-be worth a home screen slot).
-
-## What remains
-
-Neither line is code, and neither can be closed from this machine.
-
-1. **"The Windows widget is live in the manifest and verified in the widgets
- board."** The manifest half is verified in production. The in-board half
- needs a Windows 11 machine with the PWA installed from Edge: open the board
- with `Win + W`, **Add widgets**, pick **Agent glance**, confirm it populates
- with the signed-in account's real card, refreshes, and that its click targets
- land on the right routes. Owner, or anyone with a Windows 11 machine.
-
- **This precondition is now MET, re-measured 2026-09-10.** The guard exists so
- nobody spends the trip on a stale origin, and it has gone green:
- `npm run check:windows-widget:live` now passes all 16 checks against
- `https://three.ws`, including `the deployed Adaptive Card is the one this tree
- generates`, `the deployed worker is byte-identical to public/glance-sw.js` and
- `the deployed card binds the live response and all 3 states in the board's
- engine`. It exited 1 as recently as 2026-09-09 with `the deployed Adaptive Card
- differs from api/_lib/glance-adaptive.js`; a deploy has since carried the card
- fix. **So the board check is worth a human's time now.** Re-run the one command
- before travelling anyway, since a later deploy could regress it. Everything about this line
- that can be checked without a board now is, mechanically: 11 offline checks
- in `npm run gate`, plus 5 more against the deployed origin. What is left for
- the human is genuinely only "a person sees the pixels".
-
- Four defects have been found here so far, none of them visible from Linux and
- each of which would have failed this check on its own. Three came from reading
- the board's contract against the
- [Microsoft widget docs](https://learn.microsoft.com/en-us/microsoft-edge/progressive-web-apps/how-to/widgets)
- on 2026-09-04: `msAcTemplate` is the URL of the Adaptive Card and the worker
- has to fetch it, but the worker was sending the host that URL as a string
- where the card belonged, so a pinned slot had nothing to draw; nothing ever
- registered the periodic sync, so the advertised 15 minute refresh was never
- scheduled; and a slot pinned before the worker activated stayed empty until
- the board next resumed it. All three are fixed in `public/glance-sw.js`,
- pinned by `tests/glance-sw.test.js`, and **live in production since
- revision 00418**.
-
- The fourth was found on 2026-09-09 by expanding the live template in the
- board's own engine (`adaptivecards-templating`, the Microsoft package the
- board runs) instead of reading it with our own idea of a path. The template
- asked for `${stats.0.label}`, which is how JavaScript indexes an array and a
- syntax error in Adaptive Expression Language; the engine throws on the whole
- card rather than skipping the one binding, so the widget rendered **nothing
- at all**. Both the code and the test that covered it shared the same wrong
- assumption, which is why the guard now uses the real engine. Fixed in
- `api/_lib/glance-adaptive.js`; **this one still needs a deploy** (any routine
- deploy of `main` carries it, nothing widget-specific is required).
-
- A fifth class of defect stands behind those four: a fix that is green in the
- tree and absent from the deploy, which is what the widget is living through
- right now. `--live` covers that class, so no future card change can be
- believed shipped on the strength of a local run.
-2. **Signing and shipping the two Apple binaries.** An Apple Developer account
- is the blocker for both halves, not just the App Store one: the Mac app needs
- a Developer ID certificate and the iOS extension needs App Store signing.
- This is owner action 17 in
- [`production-100-OWNER-ACTIONS.md`](_context/production-100-OWNER-ACTIONS.md). Once the
- account exists, `apple/README.md` is the whole build recipe, and the emulator
- or device pass is the same five checks the Android widget went through.
-
-The `?platform=` parameter that points an unlinked Apple widget at the Apple
-hand-off rather than the Android one has since shipped: verified live
-2026-09-04, `mine?platform=ios` and `platform=macos` answer
-`linkUrl=https://three.ws/glance?link=apple` and everything else answers
-`link=android`.
-
-## Retire this prompt when those two lines pass
-
-1. Verify each line against actual output in front of you. Never claim a line
- you did not verify.
-2. Record the outcome in `production-100-00-INDEX.md` and
- `roadmap-00-README.md`.
-3. Commit with explicit paths and a subject that describes the diff, and delete
- this prompt file in that same commit:
-
- git rm prompts/finish/916-roadmap-native-widgets.md
-
- A finished order left on disk reads as open work to the next agent, so the
- shrinking directory is the campaign's progress ledger.
diff --git a/prompts/finish/_context/00-RETIRED-BY-OWNER.md b/prompts/finish/_context/00-RETIRED-BY-OWNER.md
index b6593205f..38087f9dd 100644
--- a/prompts/finish/_context/00-RETIRED-BY-OWNER.md
+++ b/prompts/finish/_context/00-RETIRED-BY-OWNER.md
@@ -67,3 +67,19 @@ Campaign-level orders:
- `916-roadmap-native-widgets.md`
- `917-roadmap-pumpfun-trading.md`
- `918-roadmap-pumpfun-trading-arena.md`
+
+## 2026-09-16: the retirement commit was incomplete, and is now finished
+
+The retirement commit (`1ab0f9754`) lists the 30 orders above but deleted only four of them
+(`309`, `908`, `909`, `910`). The other deletions never reached a commit, so fourteen retired
+orders stayed tracked in `finish/`, read as an open queue, and were edited by later sessions as if
+they were live. On 2026-09-16 the leftovers were removed to match the owner's recorded decision.
+Their final text, including any status notes added after 2026-09-09, is recoverable from the
+parent of the commit that removed each one:
+
+ git log --diff-filter=D --format=%h -- prompts/finish/<name>.md
+ git show <that sha>^:prompts/finish/<name>.md
+
+Nothing was lost by this. The open state those later notes recorded (the Home price decision, the
+translation run, the billing hold) already lives in `home-PROGRESS.md` and
+`production-100-OWNER-ACTIONS.md`.