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): queue the parity campaign, fifteen work orders that finish every remaining item from the launchpad teardown

nirholas6f19eb3476a5
diff --git a/docs/research/competitor-teardown-2026-09.md b/docs/research/competitor-teardown-2026-09.md
index 43ec22b4c..ff93e3083 100644
--- a/docs/research/competitor-teardown-2026-09.md
+++ b/docs/research/competitor-teardown-2026-09.md
@@ -4,6 +4,8 @@ Research date: 2026-09-29. Competitor X is a placeholder name for an AI-agent la
Companion to [launchpad-landscape-2026-09.md](launchpad-landscape-2026-09.md), which covers the wider market.
+**Status (2026-09-30):** Tier 1 shipped on 2026-09-29. Every remaining item is a work order in `prompts/finish/` (015 to 022 runnable, 919 to 925 owner-gated); start from [the campaign context](../../prompts/finish/_context/parity-00-CONTEXT.md).
+
## Bottom line
1. **Competitor X's edge is packaging, not technology.** We already ship most of what it sells: a hosted OAuth MCP server with annotated tools, resources and prompts, agent wallets with allowlists and spend caps, automations and autonomous runs, an agent marketplace with full ownership transfer, a token directory, agent email and cards, x402 selling, 90 locales. What it does better is how it presents these: one connector URL, one CLI command, one fee split explained per $100, and pages that show proof.
@@ -67,7 +69,7 @@ Status comes from a sweep of this repo on 2026-09-29. "Build" is the smallest ch
|---|---|---|---|
| 6 | Public per-agent creator earnings | `api/cron/launcher-claimer.js` claims creator fees into agent wallets. The Fee Bridge is held by the owner. There is no public earnings endpoint. | `GET /api/agents/:id/earnings` (lifetime claimed, pending, recent claims with signatures), shown on the agent profile and `/launches/:mint`. This is the number a builder screenshots. |
| 7 | One analytics page | The data is split across `/agent-economy-volume`, `/economy`, `/marketplace/analytics` and `api/home-stats.js`. | `/analytics`: total volume, agents, launches, creator fees claimed, compute tokens per day, x402 settlements, a growth chart, and a "how it's calculated" footnote for every number. |
-| 8 | $THREE fee flow "per $100" | `/three-token` has live buybacks with Solscan receipts. There is no per-$100 breakdown or published buyback wallet, and our docs disagree about whether bought-back $THREE is burned (fee-bridge says never, launcher-claimer says buyback and burn). | Settle the burn policy (owner call), then add a per-$100 diagram and the public wallet address to `/three-token`, all read from the same config the crons use. |
+| 8 | $THREE fee flow "per $100" | `/three-token` has live buybacks with Solscan receipts. There is no per-$100 breakdown or published buyback wallet, and the policy itself is settled in code: the platform never burns $THREE (`SPLIT_POLICIES` in `api/_lib/token/config.js`); the burn in `launcher-claimer.js` is each agent's own coin. Production reported the treasury and rewards wallets as unconfigured on 2026-09-30. | Add a per-$100 diagram for every split policy and the public wallet addresses to `/three-token`, all read from `GET /api/token/config`. |
| 9 | Success stories | None. | `/stories` built from real records: agents with launches, claimed fees, x402 revenue or marketplace sales. Every metric links to its mint or transaction. Only publish agents whose builder opts in. |
| 10 | `/experiments` | Only trading experiments (`/sniper/experiments`, `docs/trading-experiment.md`). | A general index with a fixed template (question, method, spend, result, what we got wrong). Start from the funnel numbers already in the landscape doc: 788 agents and 3 launches in 30 days is exactly the kind of honest finding this format is for. |
| 11 | Agent leaderboard by SOL earned | `/leaderboard` ranks traders and `api/leaderboard/unified.js` ranks creations. There are no points or seasons. | Add an "agents by SOL earned" tab using item 6's data. Points and seasons can follow once the tab gets used. |
diff --git a/prompts/README.md b/prompts/README.md
index d09b6ba00..dfa484856 100644
--- a/prompts/README.md
+++ b/prompts/README.md
@@ -79,6 +79,7 @@ directory, never this table. For the order to run them in, see [RUN-ORDER.md](RU
| `materialize-` | The physical lane: any forge creation to a high-precision 3D print delivered to a door. Printability engine (analyze/repair/STL/color-3MF), quote + order + Solana USDC checkout incl. the x402 agent lane, the /materialize surface with true-scale AR, fulfillment adapters + operator console, on-chain print certificates with editions and QR provenance, fabrication content gate. Six orders over a shared context file; launches on a manual-fulfillment adapter, no partner dependency. |
| `okx-ai-` | Taking the three.ws 3D Studio listing (agent #2632) to approved status on OKX.AI. Four open orders (real-payment gauntlet, two relisting orders, final audit and launch) plus a runbook. Each batches its OTP and funding needs into one owner message. |
| `openai-pr-` | The OpenAI Apps SDK submission pack. Briefs 01 to 05 shipped; 06 (the tool count drifted to 11 across the kit) and 07 (the go/no-go, never run) remain, with the portal submit owner-only. |
+| `parity-` | Closing the gaps a teardown of a leading agent launchpad found ("Competitor X", [docs/research/competitor-teardown-2026-09.md](../docs/research/competitor-teardown-2026-09.md)): per-agent creator earnings, `/analytics`, an earnings leaderboard, selling a whole agent over x402, `/experiments`, `/stories`, the $THREE fee flow, MCP use cases and safe paid retries (015 to 022, runnable), plus shipping, the npm publish, directory listings, a comparison page, a builder competition, pricing levers and an X mention bot (919 to 925, owner-gated). Shared facts: [finish/_context/parity-00-CONTEXT.md](finish/_context/parity-00-CONTEXT.md). |
| `production-100-` | The run-to-100% campaign: a master index sequencing every open work order toward a mechanically checkable "production ready, roadmap complete" end state, a standing ship-readiness order, closeout orders for residuals no other campaign owned (stranded custodial funds, master-key hygiene, verdict tuning), and the batched owner-action list. Its map was re-measured against code and production on 2026-09-01; start there. |
| `quality-bar-` | The GCP-credit quality campaign. Six open orders (fleet scale, PBR materials, forge UX, design system, mobile, avatar likeness); the reference pipeline, flagship lane, viewers and eval harness shipped. |
| `roadmap-` | Five runnable orders for existing surfaces (generation suite, creation consolidation, parametric avatar editor, developer resources, native home-screen widgets), plus the strategy layer ([the playbook](../docs/internal/fable-playbook.md)) that decides what to run next and [the reuse map](finish/roadmap-REUSE-MAP.md) for license-vetted OSS. |
diff --git a/prompts/RUN-ORDER.md b/prompts/RUN-ORDER.md
index 19d8635a2..645946590 100644
--- a/prompts/RUN-ORDER.md
+++ b/prompts/RUN-ORDER.md
@@ -9,6 +9,7 @@ so you can open one, run it, and open the next without consulting anything. Ever
| Band | What it is | Count |
|---|---|---|
| `001` to `014` | Runnable right now, no gate. Highest value first. Start at `001`. | 14 |
+| `015` to `022` | The `parity-` campaign, added 2026-09-30. Runnable, no gate except where an order names one step. Run 015 before 017 (017 ranks 015's data); 919 ships them. | 8 |
| `100` to `248` | The route-audit swarm, ordered by measured defect class. Parallel-safe. | 149 |
| `300` to `313` | The unbuilt Home campaign, in numeric order (04 gates 05 and 06, 11 gates 20). | 14 |
| `900` and up | Blocked: the last step is an owner action or an outside party. Read before running. | 19 |
diff --git a/prompts/finish/015-parity-01-creator-earnings.md b/prompts/finish/015-parity-01-creator-earnings.md
new file mode 100644
index 000000000..788b9257e
--- /dev/null
+++ b/prompts/finish/015-parity-01-creator-earnings.md
@@ -0,0 +1,78 @@
+# parity 01: public per-agent creator earnings
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now, no gate.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X's strongest proof is one number per agent: creator fees earned, in SOL, on a public page, with the on-chain claims linked. It is the number a builder screenshots. three.ws launches agent coins on pump.fun (90 on mainnet by 2026-09-30) but shows no builder what their coin has earned, except inside the coin page's live pump.fun lookup, which is not stored, not aggregated per agent, and not on the agent's own page.
+
+**Measured 2026-09-30 (read-only queries against production):**
+
+| What | Value |
+|---|---|
+| `pump_agent_mints` on mainnet | 90 |
+| `launcher_runs` | 20,722 `dry_run`, 197 `skipped`, 11 `failed`, 3 `pending`, 0 with a mint: the autonomous launcher has never launched live |
+| `launcher_claims` rows | 0 |
+| `agent_actions` rows of type `pumpfun.collect_creator_fee` | 1 (0.0706 SOL, 2026-07-26) |
+
+Two defects in the claim pipeline, verified in code (re-verify in step 0):
+
+1. `api/cron/launcher-claimer.js` posts `{ agentId, mint, network }` to `/api/pump/collect-creator-fee-agent`, whose zod schema (`collectFeeAgentSchema` in `api/pump/[action].js`) only knows `agent_id` / `avatar_id` and refines on them, so every automatic claim is a 400 recorded as `claim-failed`. Its test mocks `fetch` and misses it.
+2. `api/_lib/launcher-engine.js` treats a launch as failed unless `launch.status === 200`, but `/api/pump/launch-agent` answers a success with 201. The first live autonomous launch would be recorded as `failed` with a null mint, and the claimer (which filters `status in ('confirmed','launched') and mint is not null`) would never claim it. Its tests mock 200.
+
+Also: `launcher_claims` has no migration (it is created on the fly in two files), no index on `agent_id` or `mint`, and nothing reads it.
+
+## Step 0: re-derive the current state
+
+ grep -n "JSON.stringify({ agentId" api/cron/launcher-claimer.js
+ grep -n "launch.status !== 200" api/_lib/launcher-engine.js
+ grep -n "collectFeeAgentSchema" -A9 api/pump/\[action\].js
+ grep -rn "earnings" api/agents api/pump --include=*.js | head
+ sed -n 60,100p api/pump/launch-detail.js
+
+Then re-run the measurement with a throwaway script in `scripts/.tmp-*.mjs` (read `DATABASE_URL` from `.env.local`, SELECT only, delete the script after): counts of `pump_agent_mints` by network, `launcher_runs` by status, `launcher_claims`, and `agent_actions` where `type='pumpfun.collect_creator_fee'`.
+
+## Tasks
+
+1. **Fix the two defects at root**, each with a test that fails before the fix against the real handler contract, not a mocked `fetch` that accepts anything: the claimer sends `agent_id`; the engine accepts any 2xx with a mint (201 included). Give `launcher_claims` a real migration (same columns, plus indexes on `agent_id` and `(mint, network)`) and remove the two on-the-fly `create table` blocks.
+2. **Earnings data model.** A per-coin earnings snapshot table (migration), for example `agent_coin_earnings (mint, network, agent_id, creator, earned_lamports, claimed_lamports, unclaimed_lamports, source, refreshed_at)`, filled by a cron that walks `pump_agent_mints` (mainnet) and reads each coin's fee totals from the same pump.fun fee-sharing source `api/pump/launch-detail.js` already uses (reuse its client and circuit breaker; do not add a second one). Our own recorded claims (`launcher_claims`, `agent_actions` `pumpfun.collect_creator_fee`) join in as the list of claim transactions. Register the cron in `vercel.json` `crons` (it syncs to Cloud Scheduler); update the cron count in CLAUDE.md if `npm run check:claude` requires it.
+3. **Public API.** `GET /api/agents/:id/earnings`: lifetime earned, claimed and unclaimed in SOL and USD (`solPriceUsd()` from `api/_lib/sol-price.js`), per-coin rows, recent claims with signatures and explorer links, `refreshed_at`, and a `method` string that says where the numbers come from. Public (earnings of a public agent are public on-chain anyway), cached (`public, max-age=60, s-maxage=120, stale-while-revalidate=300`), and it answers a clean empty state for an agent with no coins. Add the per-mint figure to `/api/pump/launch-detail` from the snapshot when present so the coin page stops hitting pump.fun on every view.
+4. **Show it.** On the agent page (`/agents/:id`, `pages/avatar-page.html` + `src/avatar-page.js`) and the agent profile (`/agents/:id/profile`, `src/agent-detail.js`): an "Earned" card with the SOL total, USD beside it, the per-coin breakdown and the claims list, each claim linking to Solscan. On `/launches/:mint`, the same figures for that coin. Every state designed: loading skeleton, no coins ("launch a coin" link to `/launch`), coins but nothing earned yet, data stale (show `refreshed_at`), API error.
+5. **Docs.** `docs/api-reference.md` (the endpoint, response shape, method), the MCP resource `three://agents/{agentId}/earnings` if the MCP resources list follows this pattern (check `api/_mcp/resources.js`), `STRUCTURE.md` row, changelog entry.
+
+Do not extend automatic claiming to user-launched coins in this order: claiming signs with the agent's custodial wallet on the owner's behalf, which is a behaviour change for the owner to decide. Name it as a follow-up in the report.
+
+## Definition of done
+
+- [ ] Both defects fixed; each has a test that fails on the old code (show the red run) and passes now.
+- [ ] `launcher_claims` has a migration with the two indexes; `npm run db:status` shows the new migrations and `npm run db:migrate` applied them (read `db:status` first).
+- [ ] The snapshot cron runs once against production data and fills a row for every mainnet coin in `pump_agent_mints` or records why a coin was skipped.
+- [ ] `curl -s localhost:3000/api/agents/<real agent id>/earnings` returns real figures for an agent that launched a coin, and a clean empty state for one that did not.
+- [ ] The agent page, profile and coin page show the figures, browser-verified at 375, 768 and 1440 px, both themes, zero console errors, every state exercised.
+- [ ] `npm test` passes for the touched suites; docs, `STRUCTURE.md` and changelog updated.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| pump.fun's fee API is down or throttled | The circuit breaker in `launch-detail.js` already handles it; the snapshot keeps the last good value and the page shows `refreshed_at`. Never invent a figure. |
+| A coin predates fee sharing and returns nothing | Record it with `source: 'unavailable'` and render "not reported" for that coin. |
+| Devnet coins | Exclude them from earnings; say so in `method`. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/015-parity-01-creator-earnings.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/016-parity-02-analytics.md b/prompts/finish/016-parity-02-analytics.md
new file mode 100644
index 000000000..19e3643e9
--- /dev/null
+++ b/prompts/finish/016-parity-02-analytics.md
@@ -0,0 +1,70 @@
+# parity 02: /analytics, one page of live platform totals and growth
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now, no gate.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X's `/analytics` is the page partners and investors open first: total volume, agents, fees, compute used per day, a growth chart, and a footnote under every number saying how it was calculated. three.ws has more of this data than it does, but it is spread over seven endpoints and four pages, and nobody can see the whole platform in one place.
+
+What exists (from a code read on 2026-09-30; re-verify in step 0):
+
+| Source | Route | What it gives |
+|---|---|---|
+| `api/home-stats.js` | `/api/home-stats` | agents, on-chain agents, widgets, chains, attestations, forge models |
+| `api/platform/stats.js` | `/api/platform/stats` | agents, views, chats, avatars, countries, widgets, chains |
+| `api/pulse.js` `view=stats` | `/api/pulse?view=stats` | 24h tips, launches, trades, payments, marketplace, active wallets, 7-day series |
+| `api/agent-economy/volume.js` | `/api/agent-economy/volume` | agent-to-agent hire volume, daily series, top providers |
+| `api/marketplace/analytics.js` | `/api/marketplace/analytics` | skill sales, buyers, sellers, trial funnel, daily |
+| `api/status.js` | `/api/status` | uptime and latency |
+
+Tables for the rest: agents `agent_identities`; launches `pump_agent_mints` (plus `fixed_supply_launches`); 3D generations `forge_creations` (`status='done'`); LLM tokens `usage_events` where `kind='llm'` (`input_tokens`, `output_tokens`, indexed by `usage_events_llm_time`); x402 settlements `x402_self_facilitator_log` (`action='settle' and ok`), `x402_receipts`, `agent_hires`; marketplace `skill_purchases` (`status='confirmed'`). `/analytics` is free: no route, no page, no `data/pages.json` entry. Charts elsewhere are drawn on a native canvas with a screen-reader table (`drawVolumeChart()` in `src/agent-economy-volume.js`); `lightweight-charts` is installed but is for price charts.
+
+## Step 0: re-derive the current state
+
+ grep -n '"/analytics' vercel.json; ls pages/analytics.html 2>/dev/null
+ for p in home-stats platform/stats "pulse?view=stats" agent-economy/volume marketplace/analytics; do printf "%s " $p; curl -s -o /dev/null -w "%{http_code}\n" "https://three.ws/api/$p"; done
+ sed -n 150,310p src/agent-economy-volume.js
+ grep -n "usage_events_llm_time\|create table.*usage_events" -r api/_lib/migrations | head
+
+Load the `dataviz` skill before choosing chart colors and marks.
+
+## Tasks
+
+1. **One aggregate endpoint.** `GET /api/platform/analytics?window=30d|90d|all`: totals (agents, agents with a wallet, coins launched, 3D models generated, LLM tokens processed, x402 settlements count and USD, marketplace sales count and volume, agent-to-agent hire volume, creator fees earned once order 015's snapshot exists) and a daily series for each over the window. Each figure carries a `method` string. Reuse the queries the existing endpoints already run (import their helpers; do not fork SQL), add the missing ones against the tables above, cache in KV for 5 minutes and at the CDN (`public, s-maxage=300, stale-while-revalidate=600`). A failing sub-query yields `null` for that metric plus an `errors` entry, never a 500 for the page.
+2. **The page.** `/analytics`, wired per the context file's five steps: a headline row of totals, a growth chart per metric (canvas pattern from `src/agent-economy-volume.js`, with the screen-reader table), the window switcher (URL-synced), "updated at" from the response, and a "How each number is calculated" section rendered from the `method` strings. States: skeletons, a metric-level "unavailable" tile, whole-page error with retry.
+3. **Link it.** From the footer's Resources column, `/docs/start-here`, and `llms.txt` (`site.llms` "Economy and $THREE" group in `data/pages.json`). Nav entry at the advanced tier.
+4. **Docs.** `docs/api-reference.md` entry for the endpoint; `STRUCTURE.md` row; changelog entry.
+
+## Definition of done
+
+- [ ] `curl -s localhost:3000/api/platform/analytics?window=30d` returns every metric above (or `null` with a named error) and a daily series, read from production data.
+- [ ] No figure is a literal in the page code (`grep` proof).
+- [ ] `/analytics` renders at 375, 768 and 1440 px in both themes, zero console errors, window switcher and every state exercised.
+- [ ] Linked from the footer, `/docs/start-here` and `llms.txt` (`npm run build:pages` shows it in `public/llms.txt`).
+- [ ] `npm test` passes for touched suites; `npm run audit:links` passes; docs, `STRUCTURE.md`, changelog updated.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| A sub-query is slow on production volume | Pre-aggregate into a daily rollup table filled by a cron, and read the rollup; do not drop the metric. |
+| Two tables disagree on a count (for example launches) | Use `pump_agent_mints` plus `fixed_supply_launches` as the launch source, and say so in `method`. |
+| 015 not shipped | Omit creator fees with `method: "coming with per-agent earnings"`; add them when 015 lands. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/016-parity-02-analytics.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/017-parity-03-earnings-leaderboard.md b/prompts/finish/017-parity-03-earnings-leaderboard.md
new file mode 100644
index 000000000..ab51d910a
--- /dev/null
+++ b/prompts/finish/017-parity-03-earnings-leaderboard.md
@@ -0,0 +1,59 @@
+# parity 03: agents ranked by what they earned, on /leaderboard
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable after order 015 (per-agent earnings) has shipped; step 0 checks.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X's leaderboard ranks agents by what they earned and brings builders back to watch their rank. three.ws's `/leaderboard` (`pages/leaderboard.html` + `src/leaderboard.js`, data from `/api/sniper/leaderboard` via `getLeaderboard` in `api/_lib/trader-stats.js`) ranks trading agents only, by score, realized P&L, win rate or ROI. Nothing ranks agents by creator fees or service income. `/economy`'s "Agents earning" section ranks by rating and buyer count, not money.
+
+Also found on 2026-09-30: two launch counts disagree. `api/leaderboard/unified.js` and `api/cron/leaderboard-rollup.js` count launches from `agent_identities.meta->'token'->>'mint'`, which agent-wallet launches leave empty (as `src/avatar-page.js` itself warns), while `daily-match` and `pulse` use `pump_agent_mints`. The `launches` board undercounts.
+
+## Step 0: re-derive the current state
+
+ ls prompts/finish/015-parity-01-creator-earnings.md 2>/dev/null && echo "015 still open: run it first"
+ grep -rn "agent_coin_earnings\|/earnings" api/agents api/_lib --include=*.js | head
+ sed -n 880,980p api/_lib/trader-stats.js
+ grep -n "meta->'token'->>'mint'" api/leaderboard/unified.js api/cron/leaderboard-rollup.js
+ grep -n "window\|sort" src/leaderboard.js | head -30
+
+If 015 has not shipped, stop here, run 015, and come back: this board has no data without it.
+
+## Tasks
+
+1. **An "Earned" board.** A new tab on `/leaderboard` (the page keeps its trading boards): agents ranked by earnings in the selected window (24h, 7d, 30d, all), where earnings are creator fees (015's snapshot) plus service income (x402 skill sales and hires: `agent_revenue_events`, `agent_hires`), each column shown separately with the total. Row: rank, agent name and portrait linking to `/agents/:id`, total in SOL with USD, the split, and the agent's coin if any. Rank movement versus the previous window when the data supports it.
+2. **API.** `GET /api/leaderboard/earnings?window=&limit=&offset=` with the same caching pattern as the sniper board (`public, max-age=10, s-maxage=20` or longer if the data refreshes slower), paginated, with a `method` string.
+3. **Fix the launch count** in `unified.js` and `leaderboard-rollup.js` to count from `pump_agent_mints` (plus `fixed_supply_launches`), with a test that fails on the old query.
+4. **Share and link.** The board's current window and tab live in the URL. Link the Earned tab from order 015's "Earned" card on agent pages ("#12 this week"). Add an OG image variant if `api/og-leaderboard.js` supports tabs.
+5. **Docs.** API reference entry, `STRUCTURE.md` row update, changelog entry.
+
+## Definition of done
+
+- [ ] `/leaderboard` has an Earned tab showing real agents with real figures for each window, browser-verified at 375, 768 and 1440 px, both themes, zero console errors, empty window state designed.
+- [ ] `GET /api/leaderboard/earnings` returns ranked rows whose totals equal the sum of their per-agent `/api/agents/:id/earnings` figures for the same window (a test or a scripted cross-check on real data).
+- [ ] The launches count fix has a failing-then-passing test.
+- [ ] `npm test` passes for touched suites; docs and changelog updated.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| Few agents have earned anything | Ship the board with what exists and a strong empty-window state ("be the first: launch a coin" linking to `/launch`). Never pad with test agents. |
+| Service income and creator fees are in different currencies | Convert both to SOL and USD at the price used for the window (`api/_lib/sol-price.js`) and state it in `method`. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/017-parity-03-earnings-leaderboard.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/018-parity-04-agent-as-paid-api.md b/prompts/finish/018-parity-04-agent-as-paid-api.md
new file mode 100644
index 000000000..3ac856d7f
--- /dev/null
+++ b/prompts/finish/018-parity-04-agent-as-paid-api.md
@@ -0,0 +1,74 @@
+# parity 04: sell a whole agent as a paid x402 API
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now. The one real paid call at the end spends real USDC, so it is the single gated step.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X's "x402 cloud": an owner flips "service active", sets a USD price per request, and gets a stable public endpoint; a buyer POSTs, gets a 402, pays in SOL or USDC on Solana, and the agent runs only after settlement. three.ws sells individual skills, not the agent, and the skill flow does not work end to end.
+
+**Measured 2026-09-30 (production, read-only):**
+
+| What | Value |
+|---|---|
+| Active rows in `agent_skill_prices` | 376 |
+| Rows in `agent_revenue_events` (every sale ever) | 0 |
+| Rows in `agent_paid_services` | 0 |
+| `agent_payment_intents_status_check` | `CHECK (status IN ('pending','paid','expired','failed'))` |
+
+`consumeIntent` in `api/_lib/x402.js` sets `status='consumed'`, which that constraint rejects, so the legacy per-skill invoke (`api/agents/x402/[action].js`) fails with a database error AFTER the buyer paid. Other defects in that legacy flow (from a code read): `pay-prep` / `pay-confirm` are session-only, so an autonomous agent cannot buy; `verifyPaid` does not check that the intent's payer is the caller; `emit402` and the manifest can name different recipients; and `emit402` links a manifest path (`/api/agents/:id/x402/:skill/manifest`) that has no route.
+
+The standard x402 v2 stack is healthy and is what to build on: `paidEndpoint` in `api/_lib/x402-paid-endpoint.js` (networks default `['solana','base']`, Solana first), `send402` / `verifyPayment` / `settlePayment` in `api/_lib/x402-spec.js`, spent-payment guards in `api/_lib/x402/spent-payments.js`. The closest pattern to copy is `api/x402/skill-call.js`: a per-request `paidEndpoint` with a `payTo` override and an `onSettled` hook for revenue. One agent turn server-side: `api/agents/talk.js` (non-streaming, persona from `meta.brain`, returns `{response, model, usage}`); `sendMessage` in `api/_lib/agents-v1/messages.js` has the full tool loop and billing but no route.
+
+## Step 0: re-derive the current state
+
+ grep -n "consumed" api/_lib/x402.js
+ grep -rn "agent_payment_intents_status_check\|agent_payment_intents" api/_lib/migrations/*.sql | head
+ sed -n 150,270p api/x402/skill-call.js
+ sed -n 90,200p api/agents/talk.js
+ grep -n "resolvePayoutAddress" -A20 api/_lib/payout.js | head -30
+
+Re-run the four production counts above with a SELECT-only throwaway script in `scripts/.tmp-*.mjs` (`DATABASE_URL` from `.env.local`; delete it after).
+
+## Tasks
+
+1. **Stop the legacy flow failing after payment.** A migration that widens `agent_payment_intents_status_check` to include `consumed`, with a test that runs `consumeIntent` against the real constraint definition. Fix the dead manifest link in `emit402`. Make `emit402` and the manifest resolve the recipient the same way. Bind an intent to its payer in `verifyPaid`. Each fix has a test that fails on the old code.
+2. **The agent endpoint.** `POST /api/x402/agents/:agentId` (and a slug alias if agents have slugs): a per-request `paidEndpoint` priced from the agent's service config, `payTo` the agent's payout address (`resolvePayoutAddress`, Solana USDC first), handler runs one turn (extract the core of `api/agents/talk.js` into a shared function both use, rather than copying it), default mode (handler before settle, so a failed turn is never charged), and `onSettled` writes `agent_revenue_events` net of `PLATFORM_FEE_BPS` (`api/_lib/fee.js`). Body `{ "message": "...", "history": [] }` capped in size; response `{ "reply", "model", "usage" }`. Only public agents whose `embed_policy` allows it can be sold.
+3. **Owner switch.** On the agent wallet's Earn tab (`src/agent-wallet-hub/tabs/earn.js` via `src/agent-economy-hub.js`): "Sell this agent as an API" toggle, price per call in USD, a short public description, then the stable endpoint URL with copy buttons, a ready `curl` showing the 402, and earnings from `agent_revenue_events`. Store the config on the agent (for example `meta.api_service = { active, price_usd, description }`), validated server-side.
+4. **Discovery.** List active agent services in `/.well-known/x402.json` (it is generated; find its builder) and in the x402 catalog page, each with price, network and input schema.
+5. **Docs.** `docs/x402.md` section "Sell your agent as an API" with the buyer's flow and a runnable `curl`; `docs/api-reference.md`; changelog entry.
+6. **The one gated step.** Prove a real paid call on mainnet: the owner message names the agent, the price (keep it at the minimum the facilitator accepts), the payer wallet and the recipient, and asks for a yes. Everything else is proven without spending: the unpaid 402 on production shape, a verify against a malformed payment, and the handler path under test.
+
+## Definition of done
+
+- [ ] The constraint migration is applied (`npm run db:status` then `npm run db:migrate`); a test proves `consumeIntent` succeeds against it.
+- [ ] Each legacy defect listed in task 1 has a failing-then-passing test.
+- [ ] `curl -i -X POST localhost:3000/api/x402/agents/<active agent id> -d '{"message":"hi"}'` returns 402 with Solana USDC first in `accepts`, the agent's payout address as `payTo`, and the configured price.
+- [ ] The owner toggle, price, endpoint and earnings render on the Earn tab, browser-verified at 375, 768 and 1440 px, both themes, zero console errors.
+- [ ] Active services appear in `/.well-known/x402.json` (`npm run audit:x402-catalog` passes).
+- [ ] Docs and changelog updated; the owner message for the one real paid call was sent (or the call was made on the owner's yes, with its signature in the report).
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| The agent has no payout address | Fall back exactly as `resolvePayoutAddress` does (the agent's own wallet); refuse to activate a service with no Solana address and say why in the UI. |
+| The facilitator is down or the settle sponsor is dry (`x402_settle` in `/api/healthz`) | Run the `x402-economy-triage` subagent; ship the code; the live paid proof waits in the owner message. |
+| LLM credits or provider errors in the turn | The turn fails before settlement, so the buyer is not charged; surface the error and test that no settlement happened. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/018-parity-04-agent-as-paid-api.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/019-parity-05-experiments.md b/prompts/finish/019-parity-05-experiments.md
new file mode 100644
index 000000000..37d7d03b4
--- /dev/null
+++ b/prompts/finish/019-parity-05-experiments.md
@@ -0,0 +1,61 @@
+# parity 05: /experiments, public write-ups of what we tried and what the numbers said
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now, no gate.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X publishes its internal experiments with the method, the money spent (often under a dollar), the results, and "the parts we got wrong", including one that ended a project. It is the most credible thing on its site because no marketing team would write it. Its own founder-allocation experiment (7,215 launches) found that 0.40% of builders ever earned $500 and 35.6% of tokens never traded once.
+
+three.ws already runs experiments but only for trading: `/sniper/experiments` (`pages/sniper-experiments.html`, `src/sniper-experiments.js`, data from `GET /api/sniper/experiments`, one entry per armed `agent_sniper_strategies` row) and the long-form `blog/autonomous-trading-experiment.html` plus `docs/trading-experiment.md`. There is no general `/experiments` index and no standard write-up format. `/experiments` is free (no route, page or `data/pages.json` entry).
+
+How content is published today: docs pages are `docs/<slug>.md`, served at `/docs/<slug>` by the SPA `docs/index.html` (it fetches `/docs/<slug>.md` and renders it with `marked`), listed in `docs/nav.json` and `data/pages.json`. Blog posts are hand-written `blog/<slug>.html` files indexed in `blog/index.html`; `scripts/inject-blog-seo.mjs` upserts them into `data/pages.json`.
+
+## Step 0: re-derive the current state
+
+ grep -n '"/experiments' vercel.json; ls pages/experiments.html 2>/dev/null
+ curl -s "https://three.ws/api/sniper/experiments?network=mainnet&window=7d" | head -c 600; echo
+ sed -n 1,60p docs/trading-experiment.md
+ sed -n 1,40p docs/research/launchpad-landscape-2026-09.md
+
+## Tasks
+
+1. **The format.** `data/experiments.json`: one entry per experiment with `slug`, `title`, `question`, `published` (date), `spend_usd`, `status` (`running`, `concluded`, `killed`), `doc` (the write-up's docs slug), and optional `live_url` (for example `/sniper/experiments`). Write-ups are `docs/experiments-<slug>.md` following one template, in this order: The question; Method (data window, tables or tools, sample size); Spend; Result (the numbers, with the SELECT or script that produced them in a closing appendix so anyone can re-run it); What we got wrong; What we changed because of it. A small validator (in `scripts/build-page-index.mjs` or its own script run by `build:pages`) fails the build on a missing field or a `doc` that does not exist.
+2. **The index page.** `/experiments`, wired per the context file: cards from `data/experiments.json` with the question, status, spend and date, newest first, filter by status; each opens its write-up at `/docs/experiments-<slug>`. Empty and error states designed.
+3. **Two first write-ups, from real data measured this session:**
+ - **The agent-to-launch funnel.** How many agents were created in the last 30 days, by how many users, and how many launched a coin (`agent_identities`, `pump_agent_mints`); how many token plans were saved; the landscape doc measured 788 agents and 3 launches in the 30 days to 2026-09-28. Re-measure; report what the numbers say and what we changed or will change.
+ - **The trading experiment.** Index the existing one (link `/sniper/experiments` as `live_url`, its doc and blog post), writing only the template sections it lacks.
+ Every number is measured this session by a SELECT-only script whose SQL goes into the appendix. Nothing is estimated or rounded to flatter.
+4. **Link and document.** Footer Resources, `/docs/start-here`, `docs/nav.json` group, `data/pages.json` entries (index and each write-up), `STRUCTURE.md` row, changelog entry per published write-up. A short `docs/experiments.md` (or a section of the index doc) says how to add one.
+
+## Definition of done
+
+- [ ] `/experiments` lists both experiments from `data/experiments.json`, each opening a write-up with all six template sections.
+- [ ] The validator fails the build on a missing field (shown by a red run on a temporary bad entry you then remove).
+- [ ] Every number in the write-ups has its query in the appendix, and re-running the queries reproduces them (evidence in the report).
+- [ ] Browser-verified at 375, 768 and 1440 px, both themes, zero console errors; `npm run audit:docs` and `npm run audit:links` pass.
+- [ ] Page index, nav, `STRUCTURE.md` and changelog updated.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| A finding is unflattering | Publish it. That is the point of the section. State what we changed. |
+| A number cannot be measured (table missing, access denied) | Say "not measured" and why; never substitute an estimate. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/019-parity-05-experiments.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/020-parity-06-success-stories.md b/prompts/finish/020-parity-06-success-stories.md
new file mode 100644
index 000000000..eb8414a45
--- /dev/null
+++ b/prompts/finish/020-parity-06-success-stories.md
@@ -0,0 +1,61 @@
+# parity 06: /stories, opt-in case studies with metrics anyone can verify
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now; it reads order 015's earnings when present and degrades cleanly when not.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X's `/stories` is five case studies, one per kind of builder (a breakout, a creative marketplace, an infrastructure protocol, a trader, a community operator), each with market cap, SOL earned and links to the coin and the builder's X account. It makes different builders see themselves, and every metric can be checked on-chain.
+
+three.ws has the right foundation in `/spotlight` (`pages/spotlight.html` + `src/spotlight.js`, entries at `/spotlight/<id>`, API `api/spotlight/[action].js`, store `api/_lib/spotlight-store.js`, table `agent_showcase` with `title`, `tagline`, `story` up to 4,000 characters, `category`, `demo_url`, `tags`, one live entry per agent, plus `agent_showcase_votes`). Measured 2026-09-30: 15 live entries, 14 `source='curated'` and 1 `community`. What it lacks: no verified metrics on an entry (no coin, fees or revenue), and no moderation path (`status='hidden'` exists in the schema but nothing writes it).
+
+**Consent rule for this order:** `scripts/seed-spotlight.mjs` inserts `curated` entries about public agents without their owners opting in. A story makes claims about someone's results, so only owner-submitted entries (`source='community'`) or entries whose owner explicitly accepted a feature invitation may become stories. `/stories` is free.
+
+## Step 0: re-derive the current state
+
+ sed -n 1,80p api/_lib/migrations/20260901160000_agent_showcase.sql
+ grep -n "case '\|action ===" api/spotlight/\[action\].js | head
+ sed -n 90,140p api/_lib/spotlight-store.js
+ ls prompts/finish/015-parity-01-creator-earnings.md 2>/dev/null && echo "015 open: earnings not available yet"
+
+Count live `agent_showcase` entries by `source` with a SELECT-only throwaway script.
+
+## Tasks
+
+1. **Verified metrics on every spotlight entry**, keyed by `agent_id` and computed server-side (never typed by the builder): coins launched (from `pump_agent_mints`, with mint and Solscan link), creator fees earned (order 015's `/api/agents/:id/earnings` data when present), service income (`agent_revenue_events`, `agent_hires`), chats and actions (already on the entry). Each metric shows its source link. Add them to the entry API response and render a "Verified results" block on `/spotlight/<id>`.
+2. **Story consent.** A `featured_story` consent on the entry, set only by the agent's owner (a checkbox on the submit form, off by default, and an "accept feature" action for an existing entry). Curated entries never become stories without it.
+3. **Moderation.** An admin-only `hide` / `unhide` action (reuse the repo's admin auth pattern; find it in `api/_lib`), logged, with a test.
+4. **`/stories`.** Wired per the context file: consented entries that have at least one verified metric, grouped by category, each card showing the verified numbers, the builder, the coin and a link to the full entry. A strong empty state inviting builders to submit on `/spotlight`, since few entries will qualify at first. Never pad the page with non-consented or curated entries.
+5. **Docs.** Extend the spotlight doc (find it under `docs/`) with verified metrics, consent and moderation; `STRUCTURE.md` row; changelog entry.
+
+## Definition of done
+
+- [ ] Entry API responses carry verified metrics from real tables; a test covers an agent with a coin and one without.
+- [ ] Only owner-consented entries appear on `/stories` (a test proves a curated, non-consented entry is excluded).
+- [ ] Admin hide and unhide work and are tested; a hidden entry disappears from `/spotlight` and `/stories`.
+- [ ] `/stories` and the updated entry page are browser-verified at 375, 768 and 1440 px, both themes, zero console errors, empty state exercised.
+- [ ] Migration applied if one was added (`npm run db:status` first); docs, `STRUCTURE.md` and changelog updated.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| No entry qualifies yet | Ship with the empty state; that is correct. Do not seed, invent or auto-consent entries. |
+| 015 not shipped | Show coins and service income; add creator fees when 015 lands (the block says which metrics are live). |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/020-parity-06-success-stories.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/021-parity-07-three-fee-flow.md b/prompts/finish/021-parity-07-three-fee-flow.md
new file mode 100644
index 000000000..b6795bc15
--- /dev/null
+++ b/prompts/finish/021-parity-07-three-fee-flow.md
@@ -0,0 +1,65 @@
+# parity 07: "where every $100 goes" on /three-token, from the live split policies
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now, no gate for the page. One measured production gap (below) needs an owner decision; it is batched into one message and does not block the page.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X's token page explains its economy in one picture: of every $100 of fees, how much goes to creators, to buybacks, to the treasury, to an ecosystem fund, with the public buyback wallet linked and a live tracker reading it. People trust what they can check. three.ws has a better-specified economy (it is in code: `SPLIT_POLICIES` in `api/_lib/token/config.js`, served publicly by `GET /api/token/config`) and never burns $THREE by design, but `/three-token` does not show where money goes.
+
+**Policy facts (do not contradict them in copy):** the platform never burns $THREE; every split routes to the treasury (which funds buybacks), the holder rewards pool, and on a sale the seller. The "buyback and burn" mentioned in `api/cron/launcher-claimer.js` is about each agent's own coin via pump.fun's buyback vault, not $THREE. Users may burn their own $THREE with the `three_burn` MCP tool; that is theirs, not the platform's.
+
+**Measured 2026-09-30:** `GET https://three.ws/api/token/config` returned `"treasury": null, "treasury_configured": false, "rewards_wallet": null, "rewards_configured": false`. `treasuryWallet()` fails closed with a 503 in production when `THREE_TREASURY_WALLET` is unset, so every $THREE-priced surface that splits to the treasury may be refusing payments. Re-measure in step 0.
+
+## Step 0: re-derive the current state
+
+ curl -s https://three.ws/api/token/config | python3 -m json.tool
+ node scripts/read-service-env.mjs --names | grep -E "THREE_TREASURY_WALLET|THREE_REWARDS_WALLET|THREE_BUYBACK_ENABLED"
+ grep -rn "treasuryWallet()" api --include=*.js | wc -l
+ sed -n 1,60p src/three-token-page.js
+ grep -n "stats\|buyback" api/three-token/\[action\].js | head -20
+ gcloud logging read 'resource.type="cloud_run_revision" resource.labels.service_name="three-ws-api" textPayload:"treasury_unavailable"' --freshness=7d --limit=20 --format='value(timestamp,textPayload)'
+
+Record whether the treasury and rewards wallets are configured, and how many `treasury_unavailable` refusals production logged in the last 7 days (that is the size of the live defect).
+
+## Tasks
+
+1. **The breakdown section on `/three-token`.** Render every split policy from `GET /api/token/config` as "of every $100 paid for <surface>": consumption (paid compute), marketplace sales, scarcity mints, spins, copy-trading performance fees, each as a stacked bar with dollar amounts and role labels (seller or creator, treasury, holder rewards). Name each surface in plain words (map policy keys to human labels in one place in the page code). No split number is typed into the page: all come from the API. State the no-burn policy in one sentence with a link to `docs/three-thesis.md` (published as `/docs/three-thesis` if routed; check).
+2. **Where the money is.** Under the breakdown: the treasury and rewards wallet addresses from the config with copy buttons and Solscan links, their live $THREE and USDC balances (use the existing Solana RPC helpers), and the buyback history the page already shows (reuse, do not duplicate, its data). When a wallet is not configured, say so plainly ("not yet published") instead of hiding the row.
+3. **Make it accessible.** The bars are an inline SVG or CSS with text equivalents; screen readers get the numbers, not just shapes. Loading skeletons and an error state if the config call fails.
+4. **Docs.** `docs/three-thesis.md` gains a short "where every $100 goes" section pointing at the live page and the API (not retyping numbers). Changelog entry.
+5. **The owner message.** Only if step 0 confirms the treasury or rewards wallet is unset: the count of refused payments from the logs, which surfaces are affected, and a request for the two wallet addresses. Setting them is a config-only update (`gcloud run services update three-ws-api --region us-central1 --update-env-vars THREE_TREASURY_WALLET=...,THREE_REWARDS_WALLET=...`), pre-approved as a command once the owner supplies the addresses; never pick an address yourself.
+
+## Definition of done
+
+- [ ] `/three-token` shows the per-$100 breakdown for every policy in `split_policies`, and adding a policy to `SPLIT_POLICIES` makes it appear with no page change (prove by a test or a local run with a temporary extra policy you then remove).
+- [ ] `grep` shows no basis-point or percentage literal from `SPLIT_POLICIES` in the page code.
+- [ ] Treasury and rewards rows render the configured address or an explicit "not yet published" state; balances come from real RPC reads.
+- [ ] Browser-verified at 375, 768 and 1440 px in both themes, zero console errors, loading and error states exercised.
+- [ ] Copy says the platform never burns $THREE and never says otherwise (`grep -i burn` over the diff, each hit justified in the report).
+- [ ] If the wallets were unset, the owner message was sent with the refusal count.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| Wallets unset in production | Build the "not yet published" state; the addresses go in the owner message. The page still ships. |
+| `gcloud` re-auth needed for the log read | Report the refusal count as "not measured" and say why; it does not block the page. |
+| RPC over quota | Use the existing failover chain (`rpc_lanes`); show "balance unavailable" per row, never a stale or invented number. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/021-parity-07-three-fee-flow.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/022-parity-08-mcp-use-cases-and-safe-retries.md b/prompts/finish/022-parity-08-mcp-use-cases-and-safe-retries.md
new file mode 100644
index 000000000..71a0ba9d6
--- /dev/null
+++ b/prompts/finish/022-parity-08-mcp-use-cases-and-safe-retries.md
@@ -0,0 +1,65 @@
+# parity 08: MCP docs as tool-call sequences, and a paid call that never tells a buyer to pay twice
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now, no gate.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Two patterns from Competitor X's developer docs that we lack:
+
+1. **Use cases written as the tool calls an agent makes, in order** ("Launch a token: check readiness, launch, get the page URL"). An agent learns a workflow faster from the call sequence than from a tool list. `docs/mcp.md` has no use-case section (its nearest is "Example Claude workflow"), although `api/_mcp/prompts.js` already encodes 17 guided workflows (`get-started`, `create-agent`, `setup-wallet`, `trade`, `launch-token`, `hire-agent`, `sell-a-skill`, `review-costs`, `setup-automations`, `setup-dca`, `explore-marketplace`, `explore-x402`, `earn-yield`, `perps`, `predictions`, `embed-avatar`, `generate-3d`).
+2. **Errors after payment that say the buyer is safe to retry.** Their paid launch returns `retrySafe: true` on any failure after payment and makes the retry idempotent on the payment proof.
+
+What happens on our x402 v2 stack today (`paidEndpoint` in `api/_lib/x402-paid-endpoint.js`, from a code read on 2026-09-30; re-verify):
+
+- Default mode runs the handler before settling, so a handler failure charges nothing. Good.
+- **Streaming mode settles first**, and the spent-payment claim runs inside `settleAndBuildResponse`. If the handler then throws, the buyer gets a generic 500 that does not say they paid, and retrying with the same payment header gets 409 `payment_replayed` telling them to pay again. They were charged and got nothing.
+- If settle succeeds and the SIWX grant write fails, the answer is 502 `siwx_record_failed`, which also does not say the buyer paid.
+- The payment-identifier idempotency hash (`api/_lib/x402/payment-identifier-server.js`) hashes `body: null` because a comment assumes paid routes are GET-only, yet POST paid routes exist (`api/x402/service.js` and others), so two different POST bodies under one payment identifier can collide.
+- `retry_safe` / `retrySafe` appears nowhere in `api/x402` or `api/_lib/x402*`.
+
+## Step 0: re-derive the current state
+
+ grep -rn "retry_safe\|retrySafe" api/x402 api/_lib/x402* api/_lib/x402 | head
+ grep -n "settleAndBuildResponse\|siwx_record_failed\|payment_replayed\|stream" api/_lib/x402-paid-endpoint.js | head -30
+ grep -n "body: null\|GET-only" api/_lib/x402/payment-identifier-server.js api/_lib/x402-paid-endpoint.js
+ grep -rln "paidEndpoint(" api/x402 | xargs grep -ln "method.*POST\|methods.*POST" | head
+ grep -n "^## " docs/mcp.md
+
+## Tasks
+
+1. **Never ask a paid buyer to pay twice.** In streaming mode, when the handler fails after settlement, record the failure against the spent payment and let the same payment header re-run the handler (once, or a small bounded number of times within a window) instead of answering `payment_replayed`. The error response for any failure after settlement carries `paid: true`, `retry_safe: true`, the settlement transaction and a plain sentence ("You were charged; repeat this exact request with the same payment to get your result"). Apply the same fields to `siwx_record_failed`. Tests: a streaming handler that throws once then succeeds is retried with the same payment and returns the result without a second charge; a completed call replayed still returns 409 as today.
+2. **Hash POST bodies.** Include the request body in the payment-identifier idempotency hash for non-GET routes, fix the misleading comment, and test that two different bodies under one identifier do not share a cached response.
+3. **Document it.** `docs/x402.md` and the API reference's error table (`retry_safe`, `paid`) say what a buyer's client should do.
+4. **Use cases in `docs/mcp.md`.** A "Use cases" section: for each guided prompt in `api/_mcp/prompts.js`, the goal in one line and the ordered tool and resource calls the prompt drives, using the real tool names (derive them from the prompt definitions, not from memory). Mark each step that spends money. Add a check (a test or an extension of an existing MCP audit) that fails when a tool named in the section no longer exists, so the section cannot drift.
+
+## Definition of done
+
+- [ ] The streaming retry test and the replay test pass; a failure after settlement returns `paid: true` and `retry_safe: true` (test).
+- [ ] The POST-body idempotency test passes.
+- [ ] `docs/mcp.md` has a "Use cases" section covering every prompt, and the drift check fails when a named tool is removed (shown by a red run you then revert).
+- [ ] `npm test` passes for the touched suites; `npm run audit:docs` passes; changelog entry added.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| A route relies on the old replay behaviour | Keep 409 for completed calls; only failed-after-settle calls become retryable. Cover the route with a test. |
+| The settle sponsor is dry in production | Irrelevant to this order: everything here is proven under test. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/022-parity-08-mcp-use-cases-and-safe-retries.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/919-parity-09-ship-and-publish.md b/prompts/finish/919-parity-09-ship-and-publish.md
new file mode 100644
index 000000000..15ed37e7e
--- /dev/null
+++ b/prompts/finish/919-parity-09-ship-and-publish.md
@@ -0,0 +1,63 @@
+# parity 09: push, deploy, and publish the three-ws CLI
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Band 900: the last steps are owner actions (push, production deploy, npm publish). Everything before them is yours.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Every hosted MCP server answers an unauthenticated call with a 401 whose body says `Run npx three-ws setup` and links `https://three.ws/docs/cli`, and `/.well-known/mcp.json` advertises the same command. On 2026-09-29 the `three-ws` package had never been published (`npm view three-ws` returned E404), so that command failed for every developer who tried it. The Tier 1 work (`/connect`, `/docs/cli`, `/skill.md`, the curated `llms.txt`, the footer notice, the CLI's `create` and `launch`) is committed on `main` but was not live.
+
+## Step 0: re-derive the current state
+
+ git fetch threews main && git log --oneline threews/main..main | wc -l
+ curl -s https://three.ws/api/version
+ for p in /connect /docs/cli /skill.md /llms.txt; do printf "%s " $p; curl -s -o /dev/null -w "%{http_code} %{content_type}\n" https://three.ws$p; done
+ (cd /tmp && npm view three-ws version 2>&1 | head -1)
+ npm whoami 2>&1 | head -1
+ npx vitest run packages/three-ws-cli/tests
+
+Record each result. Skip any task below whose outcome step 0 shows is already true.
+
+## Tasks
+
+1. **Make the package publish-ready.** In `packages/three-ws-cli`: run `npm pack --dry-run` and confirm the tarball holds `src/`, `README.md` and `LICENSE` and nothing else (no tests, no `node_modules`). Confirm `bin`, `engines`, `publishConfig.access: public` and the dependency ranges in `package.json`. Run the CLI from the packed tarball in a clean temp dir (`npm pack`, then `npx --yes ./three-ws-<version>.tgz version` and `... mcp list --available`) and confirm both work against production. Fix anything that fails at root and commit it.
+2. **Prepare the deploy exactly as the CLAUDE.md Deploy runbook says**, in order: `npm run clean:worktrees`, `npm run prep:worktree` (plan) then `-- --apply`, `npm run build:gcp` inside the worktree, and run the `deploy-preflight` subagent against it. Stop before `npm run deploy:gcp:submit`.
+3. **Send the ONE owner message** (the gate). It asks for three yeses, each with what it does: (a) `git push threews main`, listing the commit count from step 0; (b) the production deploy from the prepared worktree; (c) `npm publish` of `three-ws@<version>` from `packages/three-ws-cli`, which needs an npm account with publish rights (if `npm whoami` failed, ask for `npm login` in the same message). If the owner's current instruction already approves any of these (for example "ship it"), that is the approval: do not ask again.
+4. **After approval**, run the approved steps: the push, then `npm run deploy:gcp:submit`, `npm run deploy:gcp:purge-cdn`, then `npm publish` from the package directory. Remove the deploy worktree when it lands (`git worktree remove --force <path>`).
+5. **Verify live**, every line against production: `curl -s https://three.ws/api/version` shows the pushed commit; `/connect` and `/docs/cli` return 200 and render (Playwright, no console errors); `/skill.md` returns 200 with `text/markdown`; `/llms.txt` starts with the `## Official` block within its first 20 lines; `npm run smoke:prod` passes; and from an empty temp directory `npx --yes three-ws@latest version` prints the published version and `npx --yes three-ws mcp list --available` lists the hosted servers.
+
+## Definition of done
+
+- [ ] `git log --oneline threews/main..main` is empty for the commits that existed at step 0.
+- [ ] `curl -s https://three.ws/api/version` reports a commit at or after the pushed head.
+- [ ] `/connect`, `/docs/cli`, `/skill.md` and `/llms.txt` return 200 on production, `/skill.md` as `text/markdown`.
+- [ ] `npm view three-ws version` prints a version; `npx --yes three-ws@latest mcp list --available` works in an empty directory.
+- [ ] `npm run smoke:prod` exits 0.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| `gcloud` answers `Reauthentication failed` | Retry one real call; if it persists, it goes into the single owner message as `gcloud auth login`, alongside the three approvals. Everything up to the submit is still yours. |
+| `npm whoami` fails | Same: `npm login` joins the single owner message. Finish the publish-readiness task regardless. |
+| Migrations pending at `db:check` | `npm run db:status`, then `npm run db:migrate` (it applies every pending migration; read the status first). |
+| A build step fails in code you did not touch | Root-cause and fix it if it blocks the build; the CLAUDE.md runbook lists the known failure modes. |
+| The npm name is taken at publish time | Stop the publish, report it, and do not rename the package: every 401, `/docs/cli` and `/.well-known/mcp.json` name `three-ws`. |
+| Production smoke fails on a page unrelated to this order | Report it as a named follow-up; it does not block this order unless it is one of the four surfaces above. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/919-parity-09-ship-and-publish.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/920-parity-10-community-skill-registry.md b/prompts/finish/920-parity-10-community-skill-registry.md
new file mode 100644
index 000000000..d2c8615c8
--- /dev/null
+++ b/prompts/finish/920-parity-10-community-skill-registry.md
@@ -0,0 +1,64 @@
+# parity 10: pull community skills from external GitHub registries
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Band 900: the owner chooses which external registries to trust. Build everything else first.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X lists about 50 "community skills" by pulling two public GitHub skill repositories at runtime (one of them a widely used Apache-2.0 library of 44 Solana protocol skills) and caching them for five minutes. Our registry is richer in validation but closed: every skill must be committed to `community-skills/` in this repo (or merged through the mirror, `scripts/sync-community-skills.mjs`). Pulling vetted external registries widens the catalog without us writing or hosting each skill.
+
+The risk is real and is the design constraint: a skill is a set of instructions an agent loads. An external skill is untrusted input. It must never be auto-enabled, must be pinned to a commit, and must be labelled as external everywhere it appears.
+
+## Step 0: re-derive the current state
+
+ sed -n 1,60p api/_lib/community-skills.js
+ sed -n 1,40p api/skills/community.js
+ python3 -c "import json;d=json.load(open('community-skills/registry.json'));print(d['count'], d['schema'], list(d)[:8])"
+ sed -n 1,60p packages/three-ws-cli/src/commands/skills.js
+ grep -rn "COMMUNITY_SKILL_SOURCES" api scripts packages 2>/dev/null
+ node scripts/read-service-env.mjs '^COMMUNITY_SKILL_SOURCES$' --raw
+
+Note the registry schema (`community-skills/registry.json`), how `/skills/community` and `GET /api/skills/community/:slug` read it, how `three-ws skills import` installs one, and whether any external-source code already exists.
+
+## Tasks
+
+1. **Source list from runtime config, not code.** Read the list of external registries from one env var, `COMMUNITY_SKILL_SOURCES`: a JSON array of `{ "repo": "owner/name", "ref": "<commit sha>", "path": "skills", "license": "<SPDX id>" }`. Nothing in committed code, tests or docs names a specific third-party repository (the commit gate); tests use a synthetic repository fixture served by a local HTTP server.
+2. **Fetch, validate, cache.** A server module (beside `api/_lib/community-skills.js`) fetches each source's skill folders at the pinned `ref` from `raw.githubusercontent.com` (tree listing through the public GitHub API, unauthenticated, with `GITHUB_TOKEN` used when set to lift the rate limit), validates every `SKILL.md` with the same frontmatter rules as our registry, drops anything invalid with a logged reason, and caches the merged result (memory plus a short TTL). A source that fails to load never breaks our own registry: our in-image skills always render.
+3. **Label and isolate.** Every external skill carries `origin: "external"`, its source repo, the pinned commit, and its license in the API response. `/skills/community` shows them in their own clearly labelled section with a "not reviewed by three.ws" badge and a link to the exact source commit. They can be imported only by an explicit user action; nothing enables them on any agent automatically; `three-ws skills import` prints the source and commit and asks for confirmation (non-interactive use requires `--yes`).
+4. **Docs.** Extend `community-skills/README.md` (or the doc it links) with how external sources work, the trust model and how an operator adds one (the env var format). Add a `data/changelog.json` entry once live.
+5. **The owner message (the gate).** Propose a concrete source list: for each candidate, the repository, a pinned commit, its license, its skill count and why it is worth including. Candidates must be permissively licensed (MIT, Apache-2.0, BSD). Ask for a yes on the list, then set the approved value with `gcloud run services update three-ws-api --region us-central1 --update-env-vars` (config-only updates are pre-approved once the list is approved; never `--set-env-vars`).
+
+## Definition of done
+
+- [ ] With `COMMUNITY_SKILL_SOURCES` unset, `/skills/community` and its API behave exactly as before (a test proves it).
+- [ ] With the synthetic fixture source set, the API returns the fixture's valid skills with `origin: "external"`, `source`, `ref` and `license`, and omits the invalid one (a test proves each).
+- [ ] A failing source (404, timeout) leaves the in-repo registry intact and is logged (a test proves it).
+- [ ] `three-ws skills import` of an external skill without `--yes` in a non-interactive shell refuses (a test proves it).
+- [ ] `grep -rn` over the diff finds no third-party repository name.
+- [ ] `npm test` passes for the touched suites; `npm run check:rules -- --paths <files>` exits 0.
+- [ ] The owner message with the proposed source list was sent (quote it in the report). Setting the env var is the only step allowed to remain.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| GitHub API rate limit (60/hour unauthenticated) | Use `GITHUB_TOKEN` when present (check `.env`, `.env.local`, the service env); otherwise rely on the cache and fetch by pinned tree once per TTL. |
+| An external `SKILL.md` has non-standard frontmatter | Drop it with a logged reason; never loosen our validator to admit it. |
+| A candidate repository has no license | Exclude it from the proposal and say why. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/920-parity-10-community-skill-registry.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/921-parity-11-directory-listings.md b/prompts/finish/921-parity-11-directory-listings.md
new file mode 100644
index 000000000..d4315a9f7
--- /dev/null
+++ b/prompts/finish/921-parity-11-directory-listings.md
@@ -0,0 +1,57 @@
+# parity 11: list three.ws in the directories agents and builders read
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Band 900: opening a pull request on someone else's repository is posting to an external channel, so the submission itself is owner-gated. Everything up to it is yours.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+The Solana Foundation's `solana-foundation/awesome-solana-ai` README (about 426 stars on 2026-09-29, pushed weekly) is a curated list of AI coding skills, agents and tools for Solana. Competitor X is listed there twice and cites the listing as credibility on its own marketing page. three.ws is not listed. Our entry point for an agent is `https://three.ws/skill.md` (shipped 2026-09-29, order 919 makes it live).
+
+## Step 0: re-derive the current state
+
+ curl -s -o /dev/null -w "%{http_code} %{content_type}\n" https://three.ws/skill.md
+ curl -s https://raw.githubusercontent.com/solana-foundation/awesome-solana-ai/main/README.md | grep -n "^#\|three.ws\|three-ws"
+ curl -s https://raw.githubusercontent.com/solana-foundation/awesome-solana-ai/main/CONTRIBUTING.md
+ which gh; gh auth status 2>&1 | head -3
+
+If `/skill.md` is not 200 on production, order 919 has not landed: finish the preparation below and log that the submission waits on 919.
+
+## Tasks
+
+1. Read the list's CONTRIBUTING rules and match them exactly: section, alphabetical position, line format, description length.
+2. Decide the section from its content: `https://three.ws/skill.md` belongs under **AI Coding Skills > General**; if the list has an **AI Agents** or **Developer Tools** entry that fits the hosted MCP servers (`https://three.ws/connect`), prepare that second line too, and say which is primary.
+3. Write the lines. Draft for the skill: `[three-ws-skill](https://three.ws/skill.md) - AI agent skill for three.ws: generate textured and rigged 3D models and avatars, create agents with their own Solana wallet, launch coins, and pay for or sell services over x402.` Every claim must be true on production the day you submit; re-check each one.
+4. Prepare the pull request in a local clone under the session scratchpad: branch, the one-file README diff, a commit and a PR title and body written for the list's maintainers (what three.ws is in two sentences, why it fits the list, links to `/skill.md`, `/connect` and the repo). Save the diff, title and body to `docs/awesome-solana-ai-submission.md` in this repo and commit it.
+5. Look for other directories of the same kind (Agent Skills registries, MCP server directories we are not yet in; `STRUCTURE.md` and `docs/` list the ones we are). For each one missing us, add its exact submission steps to the same doc.
+6. **The owner message (the gate):** the list of submissions, each with its target, the exact text and the PR body, asking for one yes. On a yes, submit with `gh` from an authenticated account (install `gh` if missing), or hand the owner the one command if the machine has no GitHub auth.
+
+## Definition of done
+
+- [ ] `docs/awesome-solana-ai-submission.md` holds the exact diff, PR title and body, plus any other directory's steps, and is committed.
+- [ ] Every claim in the proposed lines was re-checked against production that day (the report lists each check).
+- [ ] The owner message was sent. The submission itself is the only step allowed to remain, and the report names the PR URL if it was opened.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| `gh` not installed or not authenticated | Prepare everything; the submission command goes in the owner message. |
+| The list's rules forbid self-submission | Say so in the owner message and prepare the text for a third party to submit. |
+| `/skill.md` not live yet | Prepare everything and log that the submission waits on order 919. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/921-parity-11-directory-listings.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/922-parity-12-why-three-ws.md b/prompts/finish/922-parity-12-why-three-ws.md
new file mode 100644
index 000000000..1d7f41735
--- /dev/null
+++ b/prompts/finish/922-parity-12-why-three-ws.md
@@ -0,0 +1,59 @@
+# parity 12: a "Why three.ws" page built on live numbers
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Band 900: a comparison that names other platforms needs the owner's yes before it is committed. Build the whole page without names first; the named comparison is the one gated step.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X's `/why-deploy` page answers "why build here and not elsewhere" with a flywheel diagram, live platform metrics, a builder-infrastructure grid and a comparison table against named rivals. It converts. It is also weak in exactly one place: its figures are typed into the copy, contradict its own analytics page, and cannot be checked. Ours must beat it on that axis: every number live, every claim linked to where a reader can verify it.
+
+Depends on order 016 (`/analytics` and its API). If 016 has not shipped, build that data layer first as part of this order or stop and run 016.
+
+## Step 0: re-derive the current state
+
+ grep -n '"/why\|"/compare' vercel.json | head
+ python3 -c "import json;d=json.load(open('data/pages.json'));print([p['path'] for s in d['sections'] for p in s['pages'] if p['path'] in ('/why','/why-three-ws','/compare','/analytics','/what-is')])"
+ ls prompts/finish/016-parity-02-analytics.md 2>/dev/null && echo "016 still open"
+ sed -n 1,80p docs/research/competitor-teardown-2026-09.md
+
+`/compare` is taken by a coin-comparison tool; do not reuse it. Pick `/why-three-ws` unless step 0 shows it taken.
+
+## Tasks
+
+1. **Page, unnamed version.** `/why-three-ws`, wired per the context file's five steps. Sections, in order: a one-line promise; live proof (agents, launches, creator fees earned, x402 settlements, 3D models generated, each fetched from the 016 API with a "how it's calculated" line); what a builder gets (3D body, custodial Solana wallet, MCP and CLI access, x402 selling, coin launch, marketplace), each linking to the live surface that proves it; the loop (build, get found, earn, reinvest) drawn as an accessible inline SVG; and a comparison against **categories**, not names ("closed agent launchpads", "open-source agent frameworks", "token launchpads without agents"), where every cell is a fact about three.ws that links to its proof.
+2. **States.** Skeletons while the API loads; if a metric fails, that tile says so and the rest of the page renders.
+3. **Docs and index.** `data/pages.json` entry, `STRUCTURE.md` row, nav entry (advanced tier), changelog entry.
+4. **The owner message (the gate).** Offer the named comparison as an optional table: for each platform proposed, the metric, its value, the public source URL and the date read. Every figure must come from a public source you fetched that day (DefiLlama, the platform's own docs); a figure you cannot source is left out. Ask for a yes on the names; on a yes, add the table in a follow-up commit.
+
+## Definition of done
+
+- [ ] `/why-three-ws` renders at 375, 768 and 1440 px in both themes with zero console errors, every metric fetched live (Network tab evidence in the report).
+- [ ] No number on the page is a literal in the HTML or JS (`grep` proof in the report).
+- [ ] Every claim links to a live page or API that proves it; `npm run audit:links` passes.
+- [ ] Nothing committed names another platform unless the owner approved that name.
+- [ ] Page index, `STRUCTURE.md`, nav and changelog updated; `npm run build:pages` clean.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| 016 not shipped | Build its API as part of this order, or run 016 first. |
+| A metric is zero or embarrassing | Show it honestly with context, or leave it off the page; never inflate or backfill. |
+| No public source for a rival's figure | Leave that figure out of the proposal. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/922-parity-12-why-three-ws.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/923-parity-13-builder-competition.md b/prompts/finish/923-parity-13-builder-competition.md
new file mode 100644
index 000000000..564c1ad86
--- /dev/null
+++ b/prompts/finish/923-parity-13-builder-competition.md
@@ -0,0 +1,56 @@
+# parity 13: a tokenized builder competition page
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Band 900: prize money, dates, judges and tracks are owner decisions, and announcing is posting. Build the machinery so the owner's decision is a config change, not a project.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X ran a six-week builder competition on Solana whose entry rule was "build an agent, then launch its token on our platform by the deadline". Its own analytics page reports 195 tokenized entries and about $44.65M of trading volume in that window. The format turns a hackathon into platform launches, and its live leaderboard of entries by fees earned is the retention loop. three.ws has the pieces (agents, launches, earnings once order 015 ships) but no competition surface. `/arena` and tournaments are trading PvP, not this.
+
+## Step 0: re-derive the current state
+
+ grep -n '"/compete\|"/competition\|"/hackathon\|"/build-off' vercel.json | head
+ ls api/_lib/migrations | tail -5
+ ls prompts/finish/015-parity-01-creator-earnings.md 2>/dev/null && echo "015 still open"
+ grep -rln "tournament" api/_lib/migrations | head -3
+
+## Tasks
+
+1. **Config, not code, for the decisions.** One committed file, `data/competitions.json`, holding zero or more competitions: `slug`, `title`, `starts_at`, `ends_at`, `register_by`, `tokenize_by`, `tracks` (name, description), `prizes` (amount, asset, vesting text), `judges`, `rules`. It ships EMPTY (an empty array). The page and API render nothing but an honest "no competition is running" state until the owner adds one.
+2. **Registration.** A migration for `competition_entries` (competition slug, user id, agent id, team name, X handle, created_at, unique per agent per competition), an authenticated `POST /api/competitions/:slug/entries` that checks the window and that the agent belongs to the caller, and `GET` for the public entry list.
+3. **Tokenized entries.** An entry is "tokenized" when its agent launched a coin through three.ws inside the window (from `pump_agent_mints`). Show each entry's coin, its creator fees earned (order 015's data) and 24h volume where available.
+4. **Page.** `/compete/:slug` plus `/compete` (index), wired per the context file: hero with the countdown to the next deadline, tracks, prizes, rules, the register form, and the live leaderboard of entries by fees earned with the method stated under it. Every state designed: before registration opens, open, closed, ended with winners.
+5. **Docs.** `docs/competitions.md` on how an owner runs one (edit `data/competitions.json`, deploy), `STRUCTURE.md` row, changelog entry when the first competition goes live.
+6. **The owner message (the gate).** Ask for the competition's parameters (dates, tracks, prize pool and funding source, judges, eligibility) in one message, with a filled-in example config they can edit. Prize payouts are spends: they go through the spend gate when paid.
+
+## Definition of done
+
+- [ ] With `data/competitions.json` empty, `/compete` renders a clear "no competition running" state and the API returns an empty list (tests prove both).
+- [ ] With a competition fixture, registration enforces the window and agent ownership (tests), and the leaderboard orders tokenized entries by fees earned from real tables.
+- [ ] The migration applies cleanly (`npm run db:status` shows it pending, then applied); `npm test` passes for the touched suites.
+- [ ] The page renders at 375, 768 and 1440 px in both themes with zero console errors.
+- [ ] Docs, page index, `STRUCTURE.md` updated; the owner message was sent with an example config.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| 015 (earnings) not shipped | Rank by volume from `pump_agent_mints`-linked launch data meanwhile, and switch to fees earned when 015 lands; say which in the method line. |
+| No decisions from the owner | Ship with the empty config. That is a complete, correct state. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/923-parity-13-builder-competition.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/924-parity-14-pricing-levers.md b/prompts/finish/924-parity-14-pricing-levers.md
new file mode 100644
index 000000000..a87e05ce1
--- /dev/null
+++ b/prompts/finish/924-parity-14-pricing-levers.md
@@ -0,0 +1,54 @@
+# parity 14: pricing levers (API key tiers with enforced limits, a $THREE deposit bonus)
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Band 900: prices and limits are owner decisions. Build both levers so each is a config change, with defaults that reproduce today's behaviour exactly, then ask.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X sells four API key tiers (free, builder, scale, enterprise) with rising monthly call allowances and falling swap fees, and gives a 3% credit bonus when a user deposits its own coin. Its own docs admit the call limits are recorded but never enforced. Two lessons: tiers give partners a clear paid path, and a tier that is not enforced is worse than none. Today three.ws has one flat limit on the versioned API (`apiV1`, 120 requests per minute per principal, `api/_lib/rate-limit.js`), account tiers for members (`api/_lib/account-tier.js`, pro, team, enterprise), and credit deposits in SOL or $THREE with no deposit bonus (`api/credits/deposit.js`); $THREE holders instead get up to 30% off spend.
+
+## Step 0: re-derive the current state
+
+ grep -n "apiV1" api/_lib/rate-limit.js api/_lib/agents-v1/http.js
+ grep -n "ACCOUNT_TIERS" -A40 api/_lib/account-tier.js | head -60
+ sed -n 1,80p api/credits/deposit.js
+ grep -rn "api_keys" api/_lib/migrations/*.sql | head
+ grep -n "discount\|holder" api/_lib/three-tier.js | head
+
+## Tasks
+
+1. **Key tiers, enforced.** A tier on each API key (default: the tier implied by the owner's account tier, so nothing changes for anyone today). A committed config maps tier to per-minute and per-month limits; `apiV1` reads the key's tier and applies both. Monthly usage is already metered into `usage_events` (kind `api`); the monthly check reads it with a cheap counter, and a 429 past the monthly cap says so (`reason: "monthly_quota"`) and names the reset date. Default config values must equal today's single limit so enabling the code changes no one's behaviour.
+2. **Visibility.** Dashboard → Data API (or the API keys page) shows each key's tier, the month's usage against its cap, and the reset date. `docs/api-reference.md` rate-limit section documents tiers from the config (generated or read, never retyped).
+3. **Deposit bonus lever.** A config value, `three_deposit_bonus_bps`, default `0`. When non-zero, a $THREE deposit credits `amount_usd * (1 + bps/10000)`, the ledger row records the bonus separately, and the credits page states the bonus. SOL deposits are unchanged.
+4. **The owner message (the gate).** Propose concrete tiers (names, per-minute and monthly limits, prices, what each unlocks) and a bonus value, each with the reasoning and today's usage numbers from production (the distribution of monthly API calls per key, read from `usage_events`). Ask for a yes on each lever separately. On a yes, change only the config and commit.
+
+## Definition of done
+
+- [ ] With default config, every existing rate-limit test passes unchanged and a new test proves a key's limits equal today's.
+- [ ] Tests prove: a key over its monthly cap gets 429 with `reason: "monthly_quota"`; a tier change takes effect on the next request; a $THREE deposit with `three_deposit_bonus_bps: 300` credits 3% more and records the bonus separately; with `0` it credits exactly as today.
+- [ ] Usage and tier are visible to the key owner in the dashboard (browser-verified).
+- [ ] Docs updated; the owner message carries the proposal with production usage numbers.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| No monthly counter exists | Build it on the existing `usage_events` metering; do not add a second metering path. |
+| The owner declines tiers | Leave the enforced mechanism in place at today's limits; it is still correct and documented. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/924-parity-14-pricing-levers.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/925-parity-15-mention-to-launch.md b/prompts/finish/925-parity-15-mention-to-launch.md
new file mode 100644
index 000000000..3aaa9c54d
--- /dev/null
+++ b/prompts/finish/925-parity-15-mention-to-launch.md
@@ -0,0 +1,62 @@
+# parity 15: launch a coin by mentioning @trythreews on X
+
+How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Band 900: reading mentions may need a paid X API tier (a new paid API needs approval), and replying on X is posting to an external channel. Build and prove the whole flow in dry-run mode; going live is one env change the owner approves.
+
+## Operating clause (binding)
+
+- Read CLAUDE.md first, then [_context/parity-00-CONTEXT.md](_context/parity-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
+- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report; it never becomes a question that halts work.
+- The only permitted stops are the CLAUDE.md stop-and-ask gates: spending real funds or any irreversible on-chain write, git push or a production deploy, committing content that names a crypto project other than $THREE, and destroying unrecoverable data. Where this order hits one, it says so, and you batch every such ask into ONE message after everything else is done.
+- Never name Competitor X in anything you commit. If a reference is unavoidable, write "Competitor X".
+- No mocks, no fake data, no placeholder stubs, no unfinished-work markers, no commented-out code. Real APIs and real integrations only.
+- The em-dash and en-dash characters are banned in everything you write.
+- Concurrent agents share this worktree: stage explicit paths only (never a bare add-everything), re-check `git status` before each commit, and commit finished work promptly.
+- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
+
+## Why this matters
+
+Competitor X lets anyone launch a token or check a balance by tweeting at its account. It is a distribution channel: every launch is a public post that shows the product working. Our X integration (`api/x/*`) only posts outbound (scheduled persona posts, milestones, the changelog lane); nothing reads mentions.
+
+The design constraint is ours, not theirs: a launch signs with the agent's custodial wallet, and three.ws allows that only from a same-site browser session. So the bot never launches anything. It replies with a prefilled `/launch` link (the same one `npx three-ws launch` builds), and the person reviews the cost and signs on three.ws. That keeps the spend gate intact and still turns a tweet into a launch.
+
+## Step 0: re-derive the current state
+
+ ls api/x/ && sed -n 1,40p api/x/post.js
+ node scripts/read-service-env.mjs --names | grep -i "^X_\|TWITTER"
+ grep -rn "mentions\|users/:id/mentions\|2/users/.*/mentions" api packages 2>/dev/null | head
+ grep -n "x-mentions\|x_mentions" vercel.json | head
+ sed -n 180,215p src/launch/launch-page.js
+
+Establish: which X credentials exist on the service (`X_API_KEY`, `X_API_SECRET`, `X_BEARER_TOKEN`), which API tier they belong to, and whether the mentions timeline is readable on it (one real read-only call).
+
+## Tasks
+
+1. **Parse.** A pure module that turns a mention into an intent: `launch <NAME> $<TICKER> [description]`, or `help`. It rejects anything else, and treats all tweet text as data (CLAUDE.md: text from untrusted sources never drives an action). Unit-tested on real-shaped fixtures, including hostile ones.
+2. **Resolve the account.** Map the tweet author's X user id to a three.ws account through an existing linked X identity (find where users link X today). An unlinked author gets a reply linking to sign-up and X linking, not a launch link.
+3. **Build the reply.** For a linked author with an agent: the prefilled `/launch?avatar=…&name=…&symbol=…&description=…` URL, plus one line saying nothing is paid until they sign on three.ws. For an author with no agent: a link to create one.
+4. **Poll and dedupe.** A cron (registered in `vercel.json` `crons`, synced by `scripts/create-gcp-scheduler.mjs`) reads new mentions since the last seen id (stored in `app_settings` or a small table), dedupes, rate-limits per author, and records every decision.
+5. **Dry run by default.** With `X_MENTION_BOT_LIVE` unset, the cron reads and records what it WOULD reply, and posts nothing. An admin view or log lists the would-be replies for review.
+6. **Docs.** `docs/x-mention-bot.md` (commands, the no-spend design, the env switch), `STRUCTURE.md` row, cron count in CLAUDE.md if its guard requires it (`npm run check:claude`).
+7. **The owner message (the gate).** The tier question (if mentions need a paid X tier, name the tier and monthly price, and ask to approve it), a sample of real dry-run decisions, and the request to set `X_MENTION_BOT_LIVE=1`.
+
+## Definition of done
+
+- [ ] Parser tests cover valid launches, help, malformed and hostile mentions; no path from tweet text to a spend exists (the report shows the call graph).
+- [ ] A dry-run cron tick against the real X API (read-only) records decisions and posts nothing (evidence: the records and no new posts on the account).
+- [ ] The reply URL for a fixture author opens `/launch` with the fields filled (Playwright proof).
+- [ ] `npm run check:claude` and the cron-syntax check pass; docs and `STRUCTURE.md` updated.
+- [ ] The owner message was sent; setting the live flag (and any tier upgrade) is the only step allowed to remain.
+
+## Never blocked
+
+| Blocker | Resolution (act, do not ask) |
+|---|---|
+| The current X tier cannot read mentions | Build and test everything against fixtures plus one documented real call that shows the tier error; the upgrade goes in the owner message. |
+| No linked-X-identity table exists | Find how X OAuth links today (`api/x/*`, user settings); if linking does not exist, the reply for every author is the sign-up link, and account linking becomes a named follow-up. |
+
+## Close out (required)
+
+1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
+2. Commit with explicit paths and a subject that describes the diff (house style: `type(scope): what changed and why a reader cares`).
+3. If every Definition of done line passes, delete this file in that same commit (`git rm prompts/finish/925-parity-15-mention-to-launch.md`) and append a dated entry to [_context/parity-PROGRESS.md](_context/parity-PROGRESS.md) with the commit SHAs. If a line cannot pass inside this session because an owner action or an outside party is the last step, finish everything else, leave this file in place, and log exactly which line remains and who owns it. Never delete this file on a partial.
+4. Final report, in this order: what step 0 measured; what changed (file list with commit SHAs); evidence for each Definition of done line; the single batched owner message, if the order has a gate; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/_context/parity-00-CONTEXT.md b/prompts/finish/_context/parity-00-CONTEXT.md
new file mode 100644
index 000000000..5ceb99fef
--- /dev/null
+++ b/prompts/finish/_context/parity-00-CONTEXT.md
@@ -0,0 +1,65 @@
+# parity: shared context for every `parity-` work order
+
+Not a work order. Never run this file; read it when an order names it.
+
+## What the campaign is
+
+In September 2026 we tore down a leading AI-agent launchpad on Solana (called **Competitor X** everywhere in this repo, by owner directive of 2026-09-29: never name it in committed files) and ranked what three.ws should replicate. The teardown is [docs/research/competitor-teardown-2026-09.md](../../../docs/research/competitor-teardown-2026-09.md). Its bottom line: we already have most of the technology. Competitor X wins on packaging (one connector URL, one CLI command, the fee flow explained per $100) and on proof (earnings, analytics, stories, published experiments). Every `parity-` order closes one of those gaps with real data.
+
+## Already shipped (do not rebuild)
+
+Tier 1 of the teardown landed on 2026-09-29 in five commits: `1ee0a06c3`, `781090002`, `d8080afd2`, `ec8dae368`, `a9380b9b2`.
+
+| Surface | Where |
+|---|---|
+| One-click MCP connect page | `/connect`, [pages/connect.html](../../../pages/connect.html) |
+| CLI docs and package README | `/docs/cli`, [docs/cli.md](../../../docs/cli.md), [packages/three-ws-cli/README.md](../../../packages/three-ws-cli/README.md) |
+| Official domain and $THREE mint notice | footer ([public/footer.html](../../../public/footer.html)), `llms.txt` (`site.official` in [data/pages.json](../../../data/pages.json)) |
+| Curated `llms.txt` | `site.llms` in `data/pages.json`, rendered by [scripts/build-page-index.mjs](../../../scripts/build-page-index.mjs) |
+| Entry-point Agent Skill | `https://three.ws/skill.md`, template [data/skill-md.template.md](../../../data/skill-md.template.md), generated by `scripts/build-skills-pack.mjs` |
+| `npx three-ws create` / `launch` | [packages/three-ws-cli/src/commands/agent.js](../../../packages/three-ws-cli/src/commands/agent.js) |
+| API error table, rate limits, idempotency, integration checklist | [docs/api-reference.md](../../../docs/api-reference.md) |
+
+Not yet live: none of the above was pushed or deployed when this file was written, and the `three-ws` npm package has never been published (order 919 owns both).
+
+## Facts every order relies on
+
+- **$THREE** mint `FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump` on Solana. It is the only coin the platform promotes.
+- **The platform never burns $THREE.** Policy is in code: [api/_lib/token/config.js](../../../api/_lib/token/config.js) (`SPLIT_POLICIES`, no burn leg; the treasury funds buybacks). The "buyback and burn" in `api/cron/launcher-claimer.js` is about each agent's OWN coin through pump.fun's buyback vault, not $THREE. Never write copy that says $THREE is burned by the platform.
+- **Public token config:** `GET /api/token/config` returns the mint, the treasury and rewards wallets, and every split policy in basis points. Read numbers from it; never type them into page copy.
+- **Solana first.** Build and verify on Solana before any EVM leg (CLAUDE.md "Solana first").
+- **Spend gate.** Anything that pays, launches, transfers or settles shows amount, asset, chain and recipient and waits for an explicit yes. Launching signs with the agent's custodial wallet and is allowed only from a same-site browser session (`requireSessionUser` in `api/pump/[action].js`); do not add a bearer-token launch path.
+- **Commit gate.** Committed files never name Competitor X; say "Competitor X" if a reference is unavoidable. Anything naming another crypto project needs the owner's yes before commit.
+- **Numbers on pages come from live APIs.** Competitor X's pages contradict themselves (two revenue-share figures, six tool counts). Every figure we publish is fetched at render time or generated at build time, with a "how it's calculated" line.
+
+## How a new page is wired (all five, every time)
+
+1. The page: `pages/<name>.html` (copy the head block, theme boot script, `nav.js` header and `/i18n.js` tail from [pages/connect.html](../../../pages/connect.html)). Colors and spacing come from the tokens in `public/tokens.css` / `public/style.css`.
+2. The route: a `{ "src": "/<name>/?", "dest": "/<name>.html" }` row in `vercel.json`, near a sibling page.
+3. The build input: a `<name>: resolve(__dirname, 'pages/<name>.html')` line in `vite.config.js`.
+4. The index: an entry in `data/pages.json` with `path`, `title`, `description`, `priority`, `changefreq`, `added`, then `npm run build:pages`. This feeds the sitemap, `llms-full.txt`, `features.json` and the changelog automatically.
+5. The map: a row in `STRUCTURE.md`. Nav entries go in `public/nav-data.js`; the lite (first-visit) menu is capped at 30 items by `tests/onboarding-tier.test.js`, so new items default to `tier: 'advanced'` unless one is removed.
+
+Verify in a real browser: `npm run dev` (port 3000; reuse it if another agent is serving), then drive the page with Playwright at 375, 768 and 1440 px in dark and light themes, with zero console errors and every loading, empty and error state exercised. Throwaway Playwright scripts go in `scripts/` with a `.tmp-` prefix and are deleted after.
+
+## Order index
+
+| Order | What it ships | Gate |
+|---|---|---|
+| [015](../015-parity-01-creator-earnings.md) | Public per-agent creator earnings (API, agent page, coin page) | none |
+| [016](../016-parity-02-analytics.md) | `/analytics`: one page of live platform totals and growth | none |
+| [017](../017-parity-03-earnings-leaderboard.md) | Agents ranked by SOL earned on `/leaderboard` | needs 015 |
+| [018](../018-parity-04-agent-as-paid-api.md) | Sell a whole agent as a paid x402 endpoint | none (buyers pay; test with the spend gate) |
+| [019](../019-parity-05-experiments.md) | `/experiments`: public write-ups of what we tried and what the numbers said | none |
+| [020](../020-parity-06-success-stories.md) | `/stories`: opt-in case studies with verifiable on-chain metrics | none |
+| [021](../021-parity-07-three-fee-flow.md) | "Where every $100 goes" on `/three-token`, from the live split policies | none |
+| [022](../022-parity-08-mcp-use-cases-and-safe-retries.md) | MCP docs as tool-call sequences; `retry_safe` on every post-payment failure | none |
+| [919](../919-parity-09-ship-and-publish.md) | Push, deploy, publish the `three-ws` CLI to npm | owner: push, deploy, npm publish |
+| [920](../920-parity-10-community-skill-registry.md) | Pull community skills from external GitHub registries | owner: the source list names other projects |
+| [921](../921-parity-11-directory-listings.md) | List `skill.md` in the Solana Foundation's awesome-solana-ai | owner: external PR |
+| [922](../922-parity-12-why-three-ws.md) | "Why three.ws" comparison page | owner: names other projects |
+| [923](../923-parity-13-builder-competition.md) | A tokenized builder competition page | owner: prize money |
+| [924](../924-parity-14-pricing-levers.md) | API key tiers with enforced limits; a $THREE deposit bonus | owner: pricing |
+| [925](../925-parity-15-mention-to-launch.md) | Launch a coin by mentioning @trythreews on X | owner: design sign-off, posting |
+
+Progress log: [parity-PROGRESS.md](parity-PROGRESS.md). Append one dated entry per session.
diff --git a/prompts/finish/_context/parity-PROGRESS.md b/prompts/finish/_context/parity-PROGRESS.md
new file mode 100644
index 000000000..17f1aba56
--- /dev/null
+++ b/prompts/finish/_context/parity-PROGRESS.md
@@ -0,0 +1,11 @@
+# parity: progress log
+
+Cross-session handoff for the `parity-` orders ([parity-00-CONTEXT.md](parity-00-CONTEXT.md)). Append one dated entry per session: which order, what shipped (commit SHAs), what was measured, and anything the next session must know. Never edit an earlier entry.
+
+## 2026-09-29: teardown and Tier 1
+
+- Teardown written: `docs/research/competitor-teardown-2026-09.md` (`1cfce0cd1`).
+- Tier 1 shipped locally, not pushed: `/connect` and `/docs/cli` (`1ee0a06c3`), official notice and curated `llms.txt` (`781090002`), `/skill.md` (`d8080afd2`), `three-ws create` / `launch` (`ec8dae368`), API reference error table and checklist (`a9380b9b2`).
+- Found: the `three-ws` npm package was never published, so `npx three-ws setup` (named in every MCP 401 and in `/.well-known/mcp.json`) fails for everyone. Order 919.
+- Corrected in the context file: the teardown called the $THREE burn policy contradictory. It is not. The platform never burns $THREE; the burn in `launcher-claimer.js` is each agent's own coin.
+- Orders 015 to 022 and 919 to 925 written 2026-09-30.