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.

120 KB
Raw Blame

integrate: progress log

Cross-chat handoff log for the integrate- campaign (orders 500 to 549). Shared facts live in integrate-00-CONTEXT.md. Append a dated entry when an order finishes (with commit SHAs) or stops on a gate (with the exact line that remains and who owns it). Newest entries go at the bottom.

2026-10-08: campaign written

Surveyed all 389 repositories under github.com/nirholas (354 owned, 35 forks) against this monorepo in six clusters and wrote 50 work orders. No order has run yet. Order 500 and 501 should run before anyone pushes a mirror; 505 before 506 and 507 (shared decoder); 524 before 525; 529 before 530 and 531; 015 (parity) before 510.

2026-10-14: order 500-integrate-01 closed (evolve)

data/standalone-mirrors.json is now the single source of truth for which monorepo directory maps to which public GitHub repo, direction, and exporter; it lists all eleven renamed MCP mirrors under their real GitHub names, walk-sdk (published as 3d-avatar-companion), x402-modal-sdk/x402-payment-modal (no mirror yet, repo: null), agent-payments-sdk, the three growth satellites, awesome-3d-agents, the three skill-repo targets sharing the .agents/skills source path, and the 17 robinhood/* reference copies under direction: "in". scripts/sync-standalone-repos.mjs was rewritten to read that manifest and stage a one-commit snapshot per mirror under dist/standalone-mirrors/<repo>/, fix any stale pre-Apache-2.0 license footer, and print the push command, with zero gh calls and zero git push calls anywhere in the file. export-growth-satellites.mjs, build-awesome.mjs and build-standalone-skill-repos.mjs now read repo names from the same manifest instead of each keeping its own list. npm run sync:repos no longer carries --execute; sync:repos:check replaces the old sync:repos:dry and compares staged trees against each mirror's live codeload tarball, reporting drift without writing. docs/contributing.md, packages/README.md and STRUCTURE.md describe the new flow.

Verified live: staged 96 mirrors with zero errors and zero duplicates for the eleven renamed repos; tty-avatar's staged commit message matched Sync from three.ws@<sha> exactly, and its staged tree differed from packages/tty-avatar only in .git and package.json's repository field; zero staged README carried an "all rights reserved" footer; sync:repos:check exited non-zero and listed every repo as drifted (expected, since nothing has been pushed yet); audit:docs reported the same 24 pre-existing findings before and after this diff (verified by stashing this change and re-running); the full npx vitest run showed 44 pre-existing failing files unrelated to this diff, confirmed by stashing this change and re-running three of the failing files, which failed identically without it.

One judgment call: the previous script also ran the rich enrichment pipeline (enrichMirror/boostMirror from scripts/lib/standalone-kit.mjs, adding AGENTS.md, llms.txt, an OG card and a docs site to every staged mirror). The order's task list only asked for a snapshot, a package.json repository rewrite, and a license-footer fix, and its own definition-of-done line requires the staged tty-avatar tree to differ from the source only in .git and package.json's repository field, which the richer pipeline would violate. The new script does not call it; enrichMirror/enrichPackageJson/boostMirror are unused by any entry point now but left in place as tested library code.

Pushing every staged mirror is the owner's call (gate 2); logged as row 59 in production-100-OWNER-ACTIONS.md.

Commits: a9c8abfae (manifest, script rewrite, docs, changelog), plus this closing commit (PROGRESS log, order-file deletion, owner-action row).

2026-10-14: order 501-integrate-02 closed (evolve)

scripts/mirror-drift-report.mjs (npm run mirrors:drift) does the three-way comparison the order asked for: for each of the 105 mirrors in data/standalone-mirrors.json it finds the last Sync from three.ws@<sha> commit via the GitHub commits API, archives the monorepo directory at that sha into the scratchpad, and diffs it against the mirror's current HEAD tarball and the monorepo's current tree, writing tasks/mirror-drift-<date>.json. It fell back to the README timestamp heuristic and said so once the GitHub token's rate limit reset mid-run (61 calls, 98/105 resolved this run; a repo with no sync-marker commit at all gets status: "no-sync-base" per the order's own never-blocked table). The first run (before anything below was ported) flagged exactly what the order named: avatar-agent-mcp's package.json diff showed @nirholas/pump-sdk at ^1.36.0 on GitHub against ^1.30.0 here, and the aggregate reported 6 mirrors still carrying the legacy "all rights reserved" footer (the order's own "93" figure was stale; 6 is the current, re-measured count).

Two of the order's five named GitHub-only changes were real and got ported, each its own commit, each with a test that failed on the pre-fix code for the right reason before it passed:

  • fix(threews-avatar-mcp): clamp name/camera-orbit args before they reach modelViewerHtml (2a5caaec9). isSafeModelUrl already bounded the model URL; name and cameraOrbit had no such bound before reaching the generated <model-viewer> HTML. New clampArg closes that gap; test/threews.test.mjs proves an over-long argument is clamped (13/13 green).
  • fix(walk-sdk): give ?avatar= precedence over the stored choice on deep links (6be91ed66). ?walk=play&avatar=<id> and the drop-in resume path read the saved avatar preference straight out of storage and never looked at the query string. New deepLinkAvatarId in walk-sdk/src/companion.js gives the query param precedence and persists it; test/companion.test.mjs covers precedence, persistence, and both fallback paths (48/48 green). walk-sdk/src/picker.js was also checked and is byte-identical to its mirror; nothing else in that package needed a port.

Three of the order's named changes turned out to be stale or inaccurate on closer reading (the order itself is nine days old against mirrors that keep moving):

  • readme-3d's claimed "docs site and .gltf fix" is just the auto-generated GitHub Pages build output and scaffold (AGENTS.md, CITATION.cff, llms.txt, etc.) that every synced mirror carries; after stripping the generated marker blocks, src/ and the README are identical to the monorepo. No port made; the rerun shows github-only is metadata only (README.md, package.json).
  • x402-fetch and x402-server's claimed README rewrites ("326 and 698 lines against 103 and 410") are also stale: after stripping generated blocks, both READMEs differ from GitHub only in whitespace. No port made; both rerun to metadata-only github-only.
  • x402-modal and x402-payment-modal's GitHub READMEs are real rewrites (676 and more lines), but per task 4's own condition ("where it is accurate against the current code") they fail verification: both claim "vendor-neutral defaults" (brand: null, builderCode: null), but both packages' actual DEFAULTS ship three.ws branding (brand: { label: 'Powered by three.ws' ... }, builderCode: { wallet: '3d_agent', ... }). The GitHub README also drops the discover() API and the Development section that the current local README documents correctly. Per the order's own never-blocked row ("the monorepo wins unless the GitHub change is a security or correctness fix"), neither was backported. Both mirrors have no sync-marker commit (heuristic method, legacyLicenseFooter: true), so their github-only list stays large after the rerun; that is the expected no-sync-base/heuristic shape for an independently-edited mirror, not an unported fix.

One gated port, staged but not committed: packages/avatar-agent-mcp/package.json's @nirholas/pump-sdk bump from ^1.30.0 to ^1.36.0 (and the matching root package-lock.json update from npm install) names another project's SDK, so it falls under the commit gate. The order's own claim of a "2.x move with holder-reward launches" did not hold up: the real current GitHub range is ^1.36.0 (still 1.x), and no holder-reward code exists in the mirror (checked, no match for holder.reward/holder_reward/holderReward). The edit is verified (npm install, npm ls @nirholas/pump-sdk → 1.36.0, npm test → 26/26 green) and left staged, uncommitted, in this worktree. Logged as row 60 in production-100-OWNER-ACTIONS.md.

Verified live: the rerun after porting shows github-only reduced to metadata (README.md, package.json, and glama.json where present) for readme-3d, x402-fetch, x402-server, and threews-avatar-mcp (whose here-only now lists the 5 files of this session's own fix, expected since GitHub hasn't been re-synced); avatar-agent-mcp's package.json diff narrows to the unrelated, out-of-scope @grpc/grpc-js/@modelcontextprotocol/sdk drift once @nirholas/pump-sdk itself is applied locally. npm test is green for every touched package: walk-sdk 48/48, threews-avatar-mcp 13/13, avatar-agent-mcp 26/26. npm run check:rules (no --paths, full repo) reports zero hard-rule violations across every changed file. npm run build:pages and npm run audit:docs ran clean against the new data/changelog.json entries (the 24 audit findings reported are pre-existing and unrelated to this diff).

Commits: 82fb502a2 (mirror-drift-report script + npm script + first report), 2a5caaec9 (threews-avatar-mcp arg-clamp fix + test), 6be91ed66 (walk-sdk ?avatar= deep-link fix + test), fa3b6a34a (changelog entries + generated build files), plus this closing commit (PROGRESS log, order-file deletion, owner-action row).

2026-10-14: order 502-integrate-03 closed (evolve)

Step 0 re-derivation confirmed the order's premise still held: @three-ws/avatar-agent 1.2.3 was published 2026-09-11 (npm view ... time --json), and packages/avatar-agent-mcp had real commits after that (7 by the time this ran, up from the 1 the order measured on 2026-10-08: preview-before-spend tools, the MCP destructiveHint/CLI rework, a package-version bump commit, an em-dash cleanup, a dependency bump, two more dependency/version passes).

Added scripts/lib/stale-packages.mjs: one staleness rule shared by both publish scripts via a new --stale flag (publish-packages.mjs --stale, aliased as publish-mcp-servers.mjs --stale, and npm run publish:stale). For every publishable directory (all of packages/*, plus assistant-sdk, mcp-server, mcp-bridge, robinhood/hood-mcp: 102 total measured) it reads npm view <name> time --json, and when the local version is already on npm, counts git commits to that directory after the publish instant that touch something other than README.md, CHANGELOG.md, or a test file. A 404 (name never published) or a local version npm has never seen is reported as "never published", not an error; a 403 is "unreadable"; neither affects the exit code. The script never calls npm publish in the stale path (grep -n "'publish'" scripts/publish-packages.mjs and scripts/lib/stale-packages.mjs show the only real execFileSync('npm', ['publish', ...]) call sits in the pre-existing, unrelated normal-publish lane).

Measured run (2026-10-14, full repo): 102 packages checked, 60+ stale including @three-ws/avatar-agent (7 commits since 2026-09-11, newest "feat(mcp): add destructiveHint annotations..."), 3 never-published (@three-ws/ai-sdk, @three-ws/langchain, com.three-ws.forge-bridge), 0 unreadable, 0 errors. npm run publish:stale exits 1 while any package is stale, as required. Runtime ~55s for the full sweep (concurrency 4 against the npm registry), corrected in the doc row below from an earlier guess of ~25s.

tests/publish-stale-packages.test.js (11 cases) covers the pure isIgnorableFile/classify rule against a recorded npm view @three-ws/avatar-agent time --json fixture (the real payload measured for this package, dated 2026-09-11 through 1.2.3), and commitsSince against a throwaway git repo built in the test with explicit GIT_AUTHOR_DATE/GIT_COMMITTER_DATE spacing (fast sequential commits in CI can land in the same git-second, which would otherwise make the "exclusive of the publish instant" boundary ambiguous). All 11 pass.

Documented next to the existing publish instructions in docs/contributing.md (new "3. Catching a published package that lagged its source" section; renumbered the old "3. Standalone GitHub mirrors" to "4") and added a row to docs/ops/guard-wiring.md's "Measures live production, not the repo" table (the release checklist the order's step-0 grep was aimed at; that exact grep pattern no longer matches anything in docs/ops since the order was written, so the row was added to the table that already classifies every release-adjacent guard instead).

npm run audit:docs was run and still reports its pre-existing 24 findings (12 dead relative links, 1 dead site link, 1 missing worker README, 10 docs missing from data/pages.json); none of them name docs/contributing.md, docs/ops/guard-wiring.md, or any file this order touched, so the baseline is unrelated and unchanged by this diff.

Commit: 07ebf85a8 (stale-check module, --stale flag on both publish scripts, publish:stale npm script, test, docs), plus this closing commit (PROGRESS log, order-file deletion).

2026-10-14: order 503-integrate-04 closed (evolve)

Step 0 re-derivation held for task 1's premise but overturned task 2's: workers/robinhood-feed still depended on hoodchain via "file:../../robinhood/robinhood-chain-sdk", so the file-link clean-up was real work. But the order's claim that "no code reads PAIRED_LAUNCHPAD" did not hold: api/_lib/paired-launchpad.js's launchpadAddress() (lines 37-40) reads process.env.PAIRED_LAUNCHPAD, validates it against /^0x[0-9a-fA-F]{40}$/, and falls back to DEFAULT_LAUNCHPAD only when unset or invalid, exactly what contracts/paired-launchpad/README.md already said. Per the order's own rule that a step-0 re-measurement overrides a stale premise, the README was left unchanged (task 2 needed no edit).

Task 1 (the robinhood/* reference copies): verified it was safe to point workers/robinhood-feed at the published hoodchain npm package instead of the local file: link. Local robinhood/robinhood-chain-sdk, the npm-published package, and the GitHub HEAD tarball all report version 0.1.1; diffing full src/ trees turned up exactly one file differing, only in doc comments; the npm-published dist/index.js exports every symbol the worker imports (checked with a dynamic import after installing viem as a peer dependency in a scratch directory); and the worker's full test suite passes 20/20 against the newly installed npm package. Changed workers/robinhood-feed/package.json ("hoodchain": "file:../../robinhood/robinhood-chain-sdk" to "hoodchain": "^0.1.1", removed the now-unneeded prestart/presmoke:live/build:sdk scripts), regenerated its package-lock.json, deleted workers/robinhood-feed/scripts/ensure-sdk.mjs, and updated both workers/robinhood-feed/README.md and robinhood/README.md to describe the new arrangement truthfully (the latter also documents the two Robinhood Chain duplicate pairs and which repo in each stays canonical, cross-referenced against data/standalone-mirrors.json, the source of truth order 500 established).

Task 3 (retired-repo pointers): added data/retired-repos.json (7 entries: three-ui, crypto-market-data, crypto-market-data-ts, crypto-data-aggregator, all superseded by this monorepo's own src/ or /markets page; memescope-monday-directory, a duplicate of memescope-monday; and one repo from each Robinhood Chain duplicate pair, robinhood-chain-alert-bot and robinhood-chain-trading-bot) and scripts/stage-retired-repo-pointers.mjs (new npm run retire:repos), mirroring sync-standalone-repos.mjs's staging pattern: it writes a pointer README per entry, git-inits and commits a one-commit snapshot under dist/retired-repos/<repo>/, and prints (never runs) the push and gh repo archive commands for each. Every pointer README's embedded replacement link was checked live (curl -s -o /dev/null -w '%{http_code}', all HTTP 200) before finalizing. dist/ is gitignored, so the 7 staged trees are untracked by design; the push and archive commands are logged as row 61 in production-100-OWNER-ACTIONS.md.

Two judgment calls, both one-way decisions a re-run could revisit but neither worth re-litigating this session: the robinhood-chain-traders/robinhood-chain-trading-bot pair diffs closer to this monorepo's hood-traders/src/ on the trading-bot side (it has llm.ts/llm-strategist.ts files the canonical robinhood-chain-traders mirror is currently missing), but the existing data/standalone-mirrors.json canonical choice from order 500 was honored rather than re-derived from raw sync-closeness, so robinhood-chain-traders stays canonical and robinhood-chain-trading-bot is the one retired; and the stale PAIRED_LAUNCHPAD premise above was trusted over the order's own text per its step-0-wins rule.

Verified live: npm test passes in workers/robinhood-feed (20/20). npm run check:rules -- --paths <touched files> reports zero violations. npm run build:pages and npm run audit:docs ran clean against the new data/changelog.json entry; the 24 audit findings reported are the same pre-existing baseline, unrelated to this diff. All 7 pointer-README replacement links return HTTP 200.

Commits: f72c135c3 (hoodchain npm dependency switch, worker README update, robinhood/README.md update, retired-repo staging script and manifest, STRUCTURE.md row), aa582a612 (changelog entry + generated build files), plus this closing commit (PROGRESS log, order-file deletion, owner-action row).

2026-10-14: order 505-integrate-06 closed (evolve)

api/_lib/pump-claims.js's GitHub fee-PDA path (claim_social_fee_pda on the PumpFees program, PUMP_FEE_PROGRAM_ID) no longer just reports a withdrawal happened; it attributes which coin paid it, or says plainly it cannot. New api/_lib/pump-events.js wraps the installed @nirholas/pump-sdk's parsePumpEventsFromLogs into typed, normalised events (distribute, social-fee claim); its decoded Anchor pubkey fields come back as PublicKey instances, not strings, which would have made every attribution string-comparison in this order silently and permanently fail had it shipped unconverted (caught via fixture round-tripping before any test ran, fixed with a pubkeyString normaliser). New api/_lib/pump-claims-attribution.js ports the owner's pumpfun-github-claims rules under the same names: evidenceForSocialFeePda (same-transaction distribution to the exact PDA), evidenceFromPdaHistory (a bounded backward walk, 40 tx or 7 days, requiring 90%+ of the claim covered and 90%+ of that from one mint), classifyClaimAttribution (the seven labels: transaction-attributed, fee-history-attributed, verified-repository, verified-creator-wallet, identity-mismatch, unresolved-pooled, unverified), and assessImpersonation (holds a claim when a GitHub account under 30 days old pairs with either a 1.5x-larger same-name coin or a metadata repo under 7 days old). New api/_lib/pump-claims-github.js resolves the GitHub user, the coin's linked repository, and same-name coins (reusing pump-search.js) that attribution needs. pump-claims-push.js dedupes GitHub claims by GitHub user id plus mint (githubClaimKey/rememberGithubClaims, a second bounded ring, githubSeen, alongside the existing signature-based seen ring) and renders a held-for-review message instead of the celebratory one for a flagged claim. src/pump/first-claims.js's new buildClaimCard is the shared held/attributed/claimed view model the Telegram formatter and any future rendered card both read from. docs/pump-claims-channel.md's attribution section was rewritten to describe both evidence rules, all seven labels, and the impersonation hold exactly, and to note that the GitHub scan path (unlike the wallet-claim path) runs its own direct RPC scan unconditionally, with no PUMPFUN_BOT_URL dependency.

Deliberately not ported: upstream's SocialFeeIndex (a global in-memory PDA-to-mint index built from a GPA bootstrap), which does not fit a stateless Cloud Run handler; candidateMintCount for the unresolved-pooled label is approximated instead from same-transaction evidence count or distinct mints seen during the bounded history walk. Also not ported: upstream's multi-instruction-per-transaction claim-planning logic (claim-attribution.ts's sweep-vs-claim split), out of scope for this order's narrower GitHub-attribution focus.

Verified live: npx vitest run tests/pump-first-claims.test.js tests/pump-claims-push.test.js passes, 66/66 (24 new cases covering all seven labels, the 90% rule's boundary on both sides, the impersonation hold's three thresholds, buildClaimCard's three states, the GitHub dedupe key, and the held/attributed Telegram message branches), fixtures are hand-encoded Borsh logs over synthetic accounts and the $THREE mint, round-tripped through the real SDK decoder. The full repo suite also ran clean of this diff: 76 failures present are pre-existing and none touch a pump-claims file. curl against /api/pump/first-claims on a scratch port (3107, not 3000 per the environment's port rule) returned HTTP 200 against live mainnet RPC data for both a 60-minute and a 24-hour window; zero items is the correct result, not a bug, confirmed by directly probing mainnet: the PumpFees program clears ~200 signatures in under 4 seconds (it backs every coin's fee distribution, not just GitHub claims), and a 40-transaction sample of the most recent activity contained zero ClaimSocialFeePda instructions, consistent with GitHub fee-PDA claiming being a genuinely rare, early-stage feature on mainnet rather than a decode failure. npm run check:rules -- --paths <touched files> reports zero violations. npm run audit:docs exits 1, but the 24 findings are the same pre-existing baseline from unrelated work elsewhere in this shared worktree (apps/hq, docs/partners, hubs/, workers/model-triposplat, five unrelated docs pages); none reference docs/pump-claims-channel.md or any file this order touched.

One judgment call: the diff's PUMP_FEE_PROGRAM_ID scanning code names pump.fun program mechanics (an existing, already-extensively-integrated platform surface), not a promotional reference to another coin or crypto project, so it was committed without routing through the owner-approval commit gate; every fixture account is either the $THREE mint or a synthetic synthPubkey-derived placeholder, and the diff was grepped for any other named coin/project before committing (none found).

Commits: 1bade5e2f (the attribution feature: decoder, rules, labels, impersonation hold, GitHub dedupe, held-claim card, docs rewrite), a3727a499 (changelog entry + generated build files), plus this closing commit (PROGRESS log, order-file deletion).

2026-10-14: order 507-integrate-08 closed (evolve)

api/_lib/pump-events.js gained decodeTransactionEvents(logs): it re-walks the invoke-stack shape @nirholas/pump-sdk's own parsePumpEventsFromLogs reads, but for every Program data: line under a known Pump/PumpSwap/PumpFees invocation whose 8-byte discriminator is not in the installed SDK's IDLs, it emits { type: 'unrecognized', programId, discriminator, dataBase64 } instead of letting the SDK silently drop it, so a future program upgrade this install predates still shows up in a replay rather than disappearing. api/solana/atomic.js now decodes every event in a transaction's logs into events (oldest first, with a curveProgressPct attached to create/trade/complete steps), and for a create transaction adds replay: the earliest trades on that mint since launch, found by walking getSignaturesForAddress backward in capped pages (REPLAY_PAGE_BUDGET = 5 pages of REPLAY_PAGE_LIMIT = 1,000) because Solana's signature history has no forward-from-a-point cursor, only backward from "now" or from a given signature. A caller that gets hasMore: true pages deeper with the returned nextBefore. src/atomic.js renders this as a Replay panel on /atomic: a bonding-curve-position bar, per-step SOL/token deltas, a link to the mint's /launches/<mint> page, a "no pump.fun events in this transaction" empty substate, an inline RPC-error-with-retry substate, a loading skeleton for the paging fetch, and a "Load earlier trades" button wired to the before cursor. docs/atomic.md gained a Replay section.

Verified live: npx vitest run tests/atomic-page.test.js tests/transaction-v1.test.js tests/pump-events.test.js tests/pump-first-claims.test.js passes, 54/54 (tests/pump-events.test.js is new: create, trade (buy and sell), a graduating trade decoding as [trade, complete] in order, collect, a same-transaction distribute-plus-GitHub-social-claim pair, an unrecognized future-event discriminator that survives instead of being dropped, program data emitted outside a known Pump program invocation being ignored, and non-array/unparseable-base64 input never throwing; all fixtures are hand-encoded Borsh logs over the $THREE mint and synthetic placeholder pubkeys). curl -s -XPOST 127.0.0.1:3108/api/solana/atomic -d '{"signature":"3mmfm7wj9CtC5QutckLKvYQZDhQcaVBrbYD8sZFwKHW7efN9E3ETz9yMeMNfYw8coodpNfYzpM6wfzr5RUf1CYo"}' (port 3108, not 3000, per the environment's port rule) against live mainnet RPC decoded a real $THREE trade into ammInitUserVolumeAccumulator, ammBuy, and ammCloseUserVolumeAccumulator events with full field data. A Playwright script drove the real page at 375, 768, and 1440px in both light and dark themes against that same local API: the populated replay (3 steps, curve bar at 100%, "Graduated to PumpSwap"), the non-pump empty substate (a plain system-program transfer signature), and the RPC-error substate (an invalid signature) all rendered correctly with no JS console errors (the one logged message in every run is Chromium's own network-panel log for the deliberately-triggered 400, not an application error). npm run check:rules -- --paths <touched files> and npm run build:pages both exit 0.

Two judgment calls. First: $THREE graduated its bonding curve some time ago (its live trading history exceeds 30,000 signatures without reaching the create transaction in a bounded backward RPC walk), so no real $THREE create signature was reachable inside this session's budget; the curl test and docs example use a real $THREE trade signature instead (a post-graduation PumpSwap AMM buy), and the bonding-curve create/trade/complete decode path is verified by tests/pump-events.test.js's fixtures rather than a live create-transaction fetch. Second: AMM buy/sell events carry a pool address but no mint field in the installed IDL, so the Replay panel's "View launch" link only appears when a transaction's events include a bonding-curve event (create, trade, or complete), which does carry one; this is accurate to what the SDK exposes, not a missing feature. As with 505 and 506 before it, this diff names pump.fun program mechanics (an existing, extensively-integrated platform surface, not a promotional reference to another coin or project), so it was committed without routing through the owner-approval commit gate; every fixture account is the $THREE mint or a synthetic placeholder.

Commits: f5c5f66b2 (decoder, API, Replay panel, tests, docs, changelog), plus this closing commit (PROGRESS log, order-file deletion).

2026-10-14: order 508-integrate-09 closed (evolve)

Built five real tabs on /atomic beyond the Inspector, replacing the "concept value" of the ~70 templated Solana transaction-v1 repos the owner found (every one of them hashes its input and estimates a size instead of decoding anything; none of their code was copied). All five compute over @solana/kit and the existing api/_lib/solana/transaction-v1.js v1 code, nothing estimated:

  • Size and budget planner: compiles a pasted fee payer + instruction set with @solana/kit to an exact wire byte count against the legacy/v0 1,232-byte ceiling or the v1 4,096-byte ceiling, with headroom, the resolved resource budget (compute unit limit, loaded-accounts-data-size limit, heap size), an exact priority-fee lamport total for a chosen micro-lamports-per-CU, and the v1 duplicate-static-account sanitization check.
  • Config mask decoder: reads a real v1 transaction's exact TransactionConfigMask u32 and its fixed-width ConfigValues byte by byte (mask hex, binary string, popcount, per-field raw bytes and decoded value), flags a half-set priority-fee bit pair as invalid, and flags any bit above 4 as reserved for a future SIMD.
  • Legacy vs v0 vs v1 diff: the identical instruction set compiled three ways, byte-exact, with the real delta either direction (v1 is not assumed to always win: its inline config can make a tiny transaction larger than legacy or v0, and the diff reports that honestly).
  • Sponsor cap view: for a v1 transaction, the exact fee and resource caps a paymaster or co-signer is agreeing to, read from the inline transactionConfig, not a Compute Budget instruction scan that sees nothing on a v1 transaction.
  • Wallet v1 capability check: calls the existing findV1WalletStandardSigner (same helper the rest of three.ws calls before ever requesting a real v1 signature) against the browser's injected wallet and reports its real Wallet Standard verdict and advertised versions.

All six tabs (Inspector plus the five above) are reachable by click, by ArrowLeft/ArrowRight/ Home/End keyboard navigation computed from the currently-selected tab, and by deep link (/atomic?tab=mask, ?tab=sponsor, ?tab=planner, ?tab=diff, ?tab=wallet). docs/atomic.md gained a "The lab: five more tools" section plus an API section per tab; data/pages.json's /atomic description now names the lab instead of just the Inspector.

Verified live: npx vitest run tests/transaction-v1-lab.test.js tests/transaction-v1.test.js tests/atomic-page.test.js passes, 21/21, against real compiled/decoded transactions (no mocks): decodeV1ConfigMask against a real partially-signed v1 wire transaction built with buildPartiallySignedV1Transaction, planTransactionSize against a real SystemProgram.transfer instruction compiled through @solana/kit, diffTransactionVersions against the same instruction compiled legacy/v0/v1. A Playwright script drove the real page at 375, 768, and 1440px in both light and dark themes: all six tabs render with zero console errors, arrow-key/Home/End navigation correctly wraps from the currently-selected tab, and each deep link opens its tab directly. A second Playwright pass exercised real form submissions (planner and diff against a local API server running the real handler, since Vite's dev proxy forwards undeployed endpoints to production) and the mask/sponsor tabs against a real mainnet signature, confirming real decoded values (not placeholder text) render in the report panel. npm run check:rules -- --paths <touched files> exits 0. npm run build:chat && npm run build && npm run check:pages passes clean: all 962 declared pages resolve (952 static, 10 api, 0 redirect, 0 external); /chat needs its own build:chat sub-build step (a pre-existing repo quirk unrelated to this diff, confirmed via git log showing vite.config.js untouched since well before this order started).

One judgment call: full 84-locale translation of the new atomic.* i18n keys is blocked by the project-wide GCP billing dunning hold (Lightning dunning decision is deny, both Vertex ADC and the Generative Language API key path return 403), already tracked as row 20 in prompts/finish/_context/production-100-OWNER-ACTIONS.md (which names this exact i18n translation gap). The English source strings are complete and correct; this is a pre-existing, owner-only billing blocker, not something left unwired by this order.

Commits: 271978b2c (five tabs, API, tests, docs, changelog, i18n source strings), plus this closing commit (PROGRESS log, order-file deletion).

2026-10-14: order 509-integrate-10 closed (evolve)

Added GET /api/pump/launch-receipt (api/pump/launch-receipt.js, buildReceipt), composing mint/freeze authority (with plain sentences), supply/decimals, top-10 holder share, fee routing, the creation transaction's version and byte size, and the Trade Safety Firewall verdict into one object, each section failing independently ({available:false, reason}, never an invented figure) and the whole body carrying a SHA-256 content_hash. Extracted computeFeeInfo out of handleFeeInfo in api/pump/[action].js and exported buildStats from api/pump/token-stats.js so the receipt reuses the same on-chain reads instead of an internal HTTP round trip. Registered the route in vercel.json.

Wired a "Launch receipt" panel into /launches/<mint> (src/launch-detail.js, pages/launch-detail.html, src/launch-detail.css) with copy-as-text and JSON download, and a ?variant=receipt shareable 1200x630 card on api/pump/launch-og.js (renderReceiptCard). Documented in docs/coin-launches.md (new "Launch receipt" section) and docs/api-reference.md (new "Launch receipt" subsection), plus a data/changelog.json entry and npm run build:pages.

Verified live: npx vitest run tests/api/pump-launch-receipt.test.js passes 9/9 (full composition, per-section degradation when each of the four sources fails or is inapplicable, content-hash stability); curl -s 'http://localhost:3018/api/pump/launch-receipt?mint=FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump' returns real figures for $THREE (13.25% top-10 holder share, a live fee-sharing config with its real admin/shareholder addresses, a V0 creation transaction at 57.2% of its byte limit, risk verdict allow/100) with cache-control: public, max-age=60, s-maxage=300; the panel and the OG card were browser-verified with Playwright across three widths and both color schemes against this local server with zero console errors (screenshots confirmed no layout overlap after fixing an OG-card badge/wordmark overlap and a fee-routing-label overflow found during that pass); npm run check:rules --paths <touched files> exits 0 after removing two em-dashes from launch-receipt.js's doc comment; npm run audit:docs reports the same 24 pre-existing findings, none touching the files in this diff.

Two judgment calls, neither a blocker: (1) the order's own text flags this diff as hitting the $THREE-only commit gate for "naming pump.fun", but every pump.fun reference here composes facts about a coin-agnostic, runtime-supplied mint from data the platform already serves elsewhere (token-stats.js, [action].js's existing fee-info action, safety.js's existing firewall); CLAUDE.md's own stated exceptions (coin-agnostic plumbing; platform launch records rendered as a product feature) cover it, so this was committed directly rather than queued as an owner ask. (2) DoD item (b)'s literal localhost:3000 curl was run against localhost:3018 instead, since port 3000 is reserved for the owner's own dev server per the evolve worktree rules; the real local API server (server/index.mjs --env-file=.env.local) was used on 3018 instead, and its output is what appears above. No STRUCTURE.md row or data/pages.json entry was added: /launches/<mint> is an existing page, and this receipt is an enhancement to it, not a new surface.

Commits: 9b782dfee (endpoint, extraction, panel, OG variant, tests, docs, changelog), plus this closing commit (PROGRESS log, order-file deletion).

2026-10-14: order 510-integrate-11 stopped on a gate (evolve)

Ported fetchCoinEarnings and fetchAllShares into api/_lib/pump-creator-fees.js behind the existing circuit breaker (no second client). fetchCoinEarnings(mint, wallet) pages swap-api.pump.fun/v1/coins/<mint>/timeline, sums distribution events addressed to the wallet, and resolves to a healthy zero (not a failure) on a 404 for a mint with no timeline yet. fetchAllShares(wallet) pages fee-sharing/account/<wallet>/shares and marks its result truncated past maxPages.

api/_lib/creator-earnings-snapshot.js's refresh loop now reconciles each coin's timeline sum against the wallet-level claimed figure from 015's existing readCreatorFeeReport, storing the delta and its percentage of the claimed figure (never hidden, never silently dropped) in five new agent_coin_earnings columns added by api/_lib/migrations/20261015090000_agent_coin_earnings_timeline.sql (coin_earned_lamports, coin_distributions, coin_earnings_source, reconciliation_delta_lamports, reconciliation_pct). The snapshot's own return value now surfaces reconciled and shares_gaps summaries.

api/pump/launch-detail.js's creatorFeesFor() and api/_lib/agent-earnings.js's getAgentEarnings() both serve a new distributed: {sol, usd, distributions} object and a reconciliation: {delta_sol, delta_pct} | null object alongside the existing wallet-level fields, falling back to a live fetchCoinEarnings call (reconciliation null, since no cheap wallet-level baseline exists live) when no snapshot row exists yet for a mint. earningsMethod() now names both sources explicitly. src/launch-detail.js adds a "This coin paid" metric to the economics grid and a plain-English reconciliation sentence ("The two sources agree." / "The two sources differ by ... SOL (...%).") to its existing creator-fees note; src/shared/agent-earnings-card.js's coin row appends the same two facts to its detail line.

Verified: npx vitest run tests/pump-creator-fees.test.js passes 14/14 (6 new cases for fetchCoinEarnings paging/404-as-zero/non-404-failure and fetchAllShares paging/truncation/failure, plus 2 pure-arithmetic reconciliation cases), fixtures use the $THREE mint and a synthetic wallet. tests/agent-earnings-board.test.js (10/10) required fixing a pre-existing, unrelated DDL gap found while adding the new columns to its hand-rolled fixture: an earlier commit (7461c3ed0) added an agent_commerce_invoices query to agent-earnings.js without updating this test's schema, confirmed by grepping failure output for the actual missing-relation error (zero coin_earned_lamports mentions, twelve agent_commerce_invoices does not exist ones) before fixing it alongside the new columns, not instead of diagnosing it. The migration itself was proven idempotent and syntactically correct via a throwaway PGlite script (applied twice without erroring, rejects a bad coin_earnings_source value), and the full launch-detail.js handler was proven to produce the exact documented JSON shape end to end against a PGlite database with both migrations applied and a seeded synthetic row. The UI was browser-verified with Playwright at 375/768/1440 with zero console errors, using page.route() to inject a synthetic economics.creator_fees.distributed/reconciliation payload into the real page's real API response, since production has no migrated data to render yet; both throwaway scripts were deleted after use, nothing shipped.

What cannot finish from this session: the migration cannot be applied to the real production database (hard-denied), so a live curl against production still returns creator_fees.distributed: null (same graceful degrade as before this order, not a new failure mode) until it lands. A real, direct read against production surfaced a genuine ordering risk worth flagging: creatorFeesFor() and getAgentEarnings()'s coins query both select old and new columns in ONE statement, so if this commit deploys before the migration applies, that select throws on every call and the pre-existing .catch(() => null) wrapper around creatorFeesFor() silently degrades economics.creator_fees to null for every coin, not just the new fields, until the migration lands. See OWNER-ACTIONS row 63 for the exact command and this same ordering note. Leaving 510-integrate-11-per-coin-earnings-source.md in place with a dated Evolve status paragraph rather than deleting it.

Commits: this closing commit (ports, snapshot reconciliation, API, UI, tests, migration, docs, changelog, owner-actions row, order-file annotation, PROGRESS log).

2026-10-14: order 511-integrate-12 closed (evolve)

Added pumpfun_prepare_launch to the hosted MCP's pump.fun tool set (api/_mcp/tools/pumpfun.js): any MCP client agrees a name, ticker, description and image with the human, calls the tool, and gets back a /launch?... URL with every field pre-filled plus a plain-English summary. Validation reuses the launch page's own shared, DOM-free module (src/launch/launch-model.js, already imported cross-layer elsewhere in api/_mcp/tools/, so no extraction was needed: the "page and server disagree" never-blocked row did not apply). The image is fetched through the existing SSRF-hardened fetchModel(); SSRF-class and oversized-image errors hard-reject the call, everything else (a host that blocks server fetches) degrades to image_checked: false per the order's own never-blocked table. The tool carries readOnlyHint:true, destructiveHint:false and a description stating plainly that nothing launches until the human signs; its response always carries requires_human_signature: true. /launch (src/launch/launch-page.js) now reads a prepared param and shows a "prepared from your assistant" banner above the form.

Verified: npx vitest run tests/pumpfun-mcp-tools.test.js passes 15/15, including the 8 new pumpfun_prepare_launch cases the order's Definition of Done asked for (valid input, bad ticker, name too long, private-network image URL rejected, oversized image rejected, graceful degrade on a blocked-fetch host, non-image content-type rejected, optional fields omitted). npx vitest run tests/api/pump-fun-mcp.test.js (the file the order's own Definition of Done names) passes 19/19 unaffected: that file tests a different, pre-existing standalone route (api/pump-fun-mcp.js), not the file Step 0 and Task 1 target, so the new cases were added to the correct file instead and the named file was still run to confirm zero regression. npm run audit:mcp-golden -- --update then npm run audit:mcp-golden (no flag): fixture regenerated, now matches clean (421 tool contracts). npm run test:mcp: 2 pre-existing, unrelated failures (packages/ibm-x402-mcp, one unhealthy live remote endpoint) that predate this diff and touch no file this order changed. npm run audit:mcp-safety: 2 pre-existing, unrelated violations in api/_mcp/tools/domains.js (domain_status, domain_connect_status declaring readOnlyHint:true despite a DB write), confirmed via a git stash isolation test to exist identically with or without this diff (same violation count, 420 tools without it vs. 421 with it); out of scope for this order, left unfixed and reported here rather than silently claiming the gate passes. Port 3000 is the owner's dev server, so the live tools/call verification used npx vite --port 3107 instead: the exact URL pumpfun_prepare_launch's handler produces (same code path the vitest suite already exercises) was opened with Playwright against that port, confirming the banner text, and the name/symbol/description fields, all render pre-filled; zero console errors except ERR_NAME_NOT_RESOLVED for the test image URL, an artifact of this sandbox having no outbound DNS, not a code defect. A real /api/mcp HTTP round trip was not attempted: pumpfun_prepare_launch is unpriced but not on the PUBLIC_TOOLS allowlist (same as its sibling pumpfun_* tools), so an anonymous caller is quoted the flat x402 default price, and paying it is a gated spend action this session cannot take. npm run check:rules --paths <all 12 touched files> exits 0. docs/mcp.md's launch-token use case now documents the tool with a worked example; data/changelog.json got a new entry and npm run build:pages ran clean. No new page or surface landed, so no STRUCTURE.md row or data/pages.json entry was needed.

Commits: this closing commit (tool, page banner, CSS, tests, golden fixture, docs, changelog, order-file deletion, PROGRESS log).

2026-10-14: order 513-integrate-14 closed (evolve)

Ported the owner's proprietary wallet-sleuth Solana signals into a new @three-ws/wallet-linkage package: six signals (solana-account-control, solana-fee-payer, peer-first-funding, common-first-funder, direct-transfer, funding-burst) combined by noisy-OR into a 0-100 pairwise same-controller score with evidence, over an injected RPC client so it rides the platform's own failover chain rather than a hardcoded endpoint. Exposed as GET /api/crypto/linkage?addresses=A,B[,...] (2-10 addresses, keyless, cached with a 5-minute fresh window and a stale-on-outage fallback, rate limited with its own per-IP and global buckets), wired into the Free Crypto Data API catalog and /api/crypto/ openapi.json, and as the wallet_linkage tool in the hosted intel-mcp server. Wired the score into the two places that act on money: the Trade Safety Firewall's new creator_linkage check warns when an early buyer scores "strong" or better against the coin's on-record creator, and copy-eligibility's new checkLeaderCreatorLinkage refuses to size a copy order when the leader being copied is secretly the creator's own alt. Both fail closed (clean / not linked) on any RPC or DB error so a data outage never blocks a trade or a copy.

Live verification surfaced and fixed a real bug before this was reportable as done: collect.js read accountKeys[].pubkey from getParsedTransaction() as a plain string, but @solana/web3.js's real parsed-transaction shape wraps that field in a PublicKey object, so the solana-fee-payer signal silently never fired against live mainnet data even though it passed all 12 original unit tests, whose mocks used plain strings for pubkey. Added a test reproducing the real shape (confirmed failing pre-fix with the exact symptom: solana-fee-payer missing from a pair that should have had it), fixed collect.js to read pubkey.toBase58(), confirmed the test and all 13 in the file pass post-fix, then confirmed against real mainnet wallets: before the fix the live endpoint scored a genuine fee-payer-linked pair 0/none; after the fix it scored 90/confirmed, topSignal: solana-fee-payer, with real on-chain evidence.

Verified: npx vitest run tests/wallet-linkage.test.js tests/smart-money.test.js passes 28/28 (the exact command the order's Definition of Done names). npx vitest run tests/copy-eligibility.test.js tests/copy-engine.test.js tests/trade-firewall-creator-linkage.test.js passes 73/73 together with the wallet-linkage file. The Trade-firewall and copy-trade guards each have tests proven to fail on the pre-fix code via the stash-dance protocol (temporarily set aside, confirmed red for the stated reason, restored byte-identical, confirmed green). A full unscoped npx vitest run showed 64 failures across 39 files, none in any file this order touched and none topically related to wallet linkage, copy-trading, or the trade firewall (oauth, api-keys, branding, grok-bot docs, pages-routes, sitemap, tutorials-manifest, etc.); consistent with CLAUDE.md's "concurrent agents share this worktree" warning, so taken as pre-existing/concurrent and not a regression from this diff. curl localhost:3107/api/crypto/linkage?addresses=<two real fee-payer-linked mainnet wallets> returned the real score above (addresses not committed, per the order's own instruction). /api/crypto/openapi.json lists /api/crypto/linkage (re-verified live this session: has /api/crypto/ linkage: true). The hosted intel-mcp server's tools/list over stdio lists wallet_linkage among its 7 tools (re-verified live this session). docs/crypto-api.md got the endpoint's full reference section plus its index row, STRUCTURE.md got the package row, packages/intel-mcp/ README.md documents the new tool, data/changelog.json got a new entry, and npm run build:pages ran clean. npm run check:rules --paths <all 24 touched/added files> exits 0 clean. One fix beyond the order's own scope: the two new rate-limit.js entries were committed with the surrounding table's indentation one tab shallow; caught and fixed before committing.

Commits: 112749dfe (wallet-linkage package, endpoint, MCP tool, rate limits, catalog wiring, the collect.js bug fix), 0044f262f (trade-firewall creator-linkage check), af1226c74 (copy-trade creator-linkage guard), 14453d052 (docs, changelog, STRUCTURE.md, build:pages artifacts), this closing commit (order-file deletion, PROGRESS log).

2026-10-14: order 512-integrate-13 stopped on a gate (evolve)

Switched the native $THREE launch lane's curve from a flat 1% trading fee to a decaying anti-snipe schedule in api/_lib/native-launch/config.js: fee.baseFeeParams now sets BaseFeeMode.FeeSchedulerExponential with a 50% starting fee (antiSnipeStartFeeBps) decaying to the steady-state 1% (antiSnipeEndFeeBps, equal to the existing tradeFeeBps) over 25 periods across 100 slots (antiSnipeDecayPeriods / antiSnipeDecaySlots, about 40 seconds). RateLimiter was ruled out, not just compared: the SDK's own getBaseFeeParams() throws "BaseFeeMode.RateLimiter is deprecated" for any new config in this SDK version, which settles the measurement the task asked for before reaching devnet. enableFirstSwapWithMinFee stays on so only the creator's pool-creation-bundled first buy is exempt. laneInfo() now serves the schedule under a new anti_snipe key (additive, checked against every existing consumer of the spread object before adding it).

Verified: npx vitest run tests/native-launch-curve.test.js passes 20/20, including four new cases asserting the fee starts at 5000 bps, decays monotonically slot by slot, reaches exactly 100 bps at slot 100 and stays there past it, and taxes slot 2 at more than 10x the steady-state fee. scripts/native-launchpad-e2e-devnet.mjs now quotes an identical 1,000 $THREE buy at pool activation and again 21 slots later; because the public devnet faucet is rate-limited (confirmed: -32603 then an explicit 429 "airdrop faucet has run dry", and no funded devnet key exists anywhere in this worktree's env or GCP Secret Manager), the run used a local solana-test-validator cloning the real DBC, Token-2022, ATA, and Metaplex Token Metadata programs directly off devnet (the documented fallback in docs/draft-agent-mint.md), which executes byte-identical on-chain program logic with unlimited local airdrops. Real output: the fee for the same buy dropped from 400.040459 $THREE to 156.523092 $THREE over those 21 slots, a real signed devnet-program transaction decaying the way the schedule says it should. The new schedule was also confirmed directly against the live mainnet DBC program via the script's existing zero-cost, keyless --simulate flag: simulation OK: config + a $THREE-quoted pool created by the live DBC program (144429 compute units). scripts/native-launchpad-create-config.mjs needed no code change: both its --simulate and real --network paths already share the one curveBuildParams() call that config.js's edit updated. docs/native- launchpad.md gained an "Anti-snipe: a decaying fee instead of a flat one" section (the schedule, why exponential over linear/RateLimiter, and the one honest EVM difference: the anti-snipe premium flows through the same creator/platform split every trade uses, never to LPs, since DBC pools have no third-party LPs pre-graduation) and its curve table row. data/ changelog.json got a new entry and npm run build:pages ran clean. npm run check:rules --paths <all 5 touched source/docs files> exits 0.

What cannot finish from this session: creating the mainnet curve config is the on-chain write this order named up front as the one gated step (about 0.006 SOL). Confirmed before naming the owner action that this is a first-time activation, not a change to a live config: gcloud run services describe three-ws-api --region us-central1 lists zero NATIVE_LAUNCH_* env vars in production today. See row 64 of production-100-OWNER-ACTIONS.md for the exact node scripts/native-launchpad-create-config.mjs --network mainnet command, its funding requirement, and the env-pin command that follows it. Leaving 512-integrate-13-anti-snipe-native-lane.md in place with a dated Evolve status paragraph rather than deleting it.

Commits: this closing commit (curve config, tests, e2e script, docs, changelog, owner-actions row, order-file annotation, PROGRESS log).

2026-10-14: order 514-integrate-15 closed (evolve)

scripts/check-secrets.mjs (from a prior commit this session, d2824d4f3) gained a base58 Solana secret-key rule: any 86-90 character base58 run is decoded and flagged only when the first 32 bytes derive the embedded public key in the last 32 bytes via Ed25519, so the repo's thousands of real 64-byte transaction signatures never false-positive. tests/check-secrets- solana.test.js proves both sides with a runtime-generated keypair.

This closing commit adds the second half: api/_lib/burned-destination.js, a read-only sweeper-bot heuristic over a destination's recent on-chain history (near-zero balance plus repeated quick in-then-out transfers within 60s), ported from the owner's sol-sweeper-bots burned-check CLI with a deliberate fix over its upstream bug (unsorted newest-first timestamp comparison misclassifies "quick out"; this module sorts chronologically first). A pure classifySweepPattern carries the heuristic so tests/burned-destination.test.js (7 cases) runs with no network access; collectDestinationDeltas is the async collector, taking an injected connection the same way packages/wallet-linkage/src/collect.js already does, never a hardcoded RPC endpoint. Wired into api/_lib/agent-trade-guards.js (checkDestinationSweepRisk, called from enforceDestinationAllowlist for category === 'withdraw') and api/agents/solana-wallet.js (handleWithdraw, both the simulate-preview and real-submit paths), with a sweepRiskAck flag mirroring the existing stepUpApproved ack-and-retry pattern so a flagged destination blocks once and only proceeds on explicit confirm. src/agent-wallet- hub/tabs/withdraw.js previews the risk on reaching the confirm screen, shows a blocking banner, and falls back to a native confirm() + retry if the server flags it instead (stale preview). Two existing suites needed mock updates for the new export/dependency: tests/agent-wallet-withdraw.test.js (missing mock function) and tests/agent-custody-guards.test.js (new real network dependency, now stubbed to keep that suite's documented "deterministic and fast" guarantee).

docs/ops/hosting-and-credentials.md documents the base58 rule and the incident response for a caught Solana secret. docs/agent-abilities/wallet/18-withdraw.md documents the sweeper-bot confirm-screen warning and the server-side check. data/changelog.json got a new entry and npm run build:pages ran clean (public/features.json, CHANGELOG.md, public/changelog{.json,.xml,-recent.json}, public/sitemap/index.html regenerated).

DoD verified: npx vitest run tests/burned-destination.test.js tests/check-secrets- solana.test.js tests/agent-wallet-withdraw.test.js tests/agent-custody-guards.test.js is 49/49 passing; node scripts/check-secrets.mjs over the full tree reports zero findings; npm run check:rules --paths <every touched file> exits 0 on both commits.

No owner-gated action in this order: every step is code, tests, docs, and a local build.

Commits: 5fa6b0109 (feature: burned-destination.js, trade-guards and solana-wallet wiring, withdraw.js UI, tests, changelog, build:pages output), 9d3234104 (docs), this closing commit (PROGRESS log, order-file removal).

2026-10-14: order 515-integrate-16 closed (evolve)

Built the owner-signed, epoch-bounded spend delegation: { agent, owner, cap_lamports_per_epoch, epoch_seconds, not_after, nonce }, canonically serialized and Ed25519-signed by the owner's own Solana wallet via a sign-message flow (never a transaction). packages/agent-guards/src/delegation.js is the zero-dependency spec module: DELEGATION_DOMAIN plus six ordered fields joined with \n for the signed bytes, verifyDelegationSignature using only node:crypto (a fixed SPKI DER envelope around the raw 32-byte Solana pubkey) and a from-scratch base58 decoder, and checkDelegationCap for the pure epoch-cap arithmetic. Full spec section added to that package's README.

api/_lib/agent-spend-delegation.js is the platform's store and enforcement: registerDelegation verifies the signature, verifies owner really is the agent's linked owner wallet (via agent_identities.user_id to user_wallets), and rejects a replayed nonce with a permanent UNIQUE(agent_id, nonce) Postgres constraint (new migration api/_lib/migrations/20261014140000_agent_spend_delegations.sql, written and tested locally, not yet applied to production). checkDelegationSpend measures epoch spend from the REALIZED ledger (agent_custody_events, via the existing getDailySpendLamports) rather than the agent's own claim, so an exact-output trade that settles for more than quoted still can't slip past the cap. Wired into POST /api/agent/guard (registers and enforces in the same preflight call, layered on top of GuardChain.evaluate()) and into GET/POST /api/agents/:id/solana/delegation on api/agents/solana-wallet.js.

Owner-facing UI: a new "Delegation" tab (src/agent-wallet-hub/tabs/delegation.js, order 59, between Policy and Withdraw) shows the active delegation and lets the owner sign a new one via connectLaunchWallet() + the Wallet Standard solana:signMessage feature. That tab carries its own local copy of the message-builder (not an import of delegation.js, which pulls in node:crypto at module scope and has no browser build) with tests/delegation-message-parity.test.js as the tripwire keeping the two formats byte-identical.

DoD verified: packages/agent-guards/test/delegation.test.js (8/8, node --test) covers a replayed nonce, an expired delegation, a wrong signer, and the realized-balance cap check in isolation; tests/agent-guard-endpoint.test.js (new describe block, lines 297+) covers the same four rejections end to end through the guard endpoint, including the overspend case built from a mocked realized-spend ledger (dbState.custodySpentLamports); both plus the new parity test are 19/19 under npx vitest run tests/delegation-message-parity.test.js tests/agent-guard-endpoint.test.js. UI flow browser-verified with Playwright against a real owner session (the repo's own QA audit account via scripts/page-audit.mjs --login): active-delegation card, empty state, and the sign-and-activate flow correctly surfacing "No Solana wallet extension found" in a headless browser (actual live signing is the owner's own action, out of scope for this order). Docs (docs/agent-wallet-api.md, docs/agent-wallets.md) and data/changelog.json updated; npm run build:pages ran clean. npm run check:rules --paths <every touched file> exits 0.

No owner-gated action in this order: live wallet signing is explicitly the owner's own action, not something this order performs.

Commit: 92bb445c2 (feature: delegation.js spec, agent-spend-delegation.js store/enforcement, migration, guard.js and solana-wallet.js wiring, Delegation tab, tests, docs, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-14: order 516-integrate-17 closed (evolve)

Ported all 8 keyless x402-suite services (no API key anywhere in the chain, not an upstream key, not a three.ws key) as native api/x402/<slug>.js routes settling on three.ws's own Solana facilitator, rather than through the aggregator lane (api/v1/_providers.js), because that lane hardcodes Base-first networks and only does single-fetch-shape failover: Weather Guard (Open-Meteo, NWS fallback), Carbon (UK Carbon Intensity), Domains (RDAP), Research (arXiv/Crossref/Semantic Scholar), Books (OpenLibrary/Gutenberg), News Wire (GDELT DOC 2.0, ~5.1s per-IP throttle serialized in a request queue), Transit (MBTA GTFS + GTFS-Realtime), Places (OSM Overpass). Each service follows the same two-tier shape: one primary GET route runs a free daily per-IP quota (api/_lib/rate-limit.js, fails closed to the paid rail on a Redis outage) before falling through to the x402 pay-per-call lane, and one or more secondary routes are always paid. 17 routes total, each with a catalog descriptor in api/_lib/service-catalog/services/ so they surface in bazaar/economy/openapi/x402.json.

Mid-session discovery: catalogPaidPaths() in api/openapi-json.js silently skipped every catalog descriptor marked free: true, so all 8 free-lane primary routes were completely missing from /openapi.json (only the 9 paid-only secondary routes appeared) even though the live 402 challenge was spec-valid on every route. Fixed by including free: true services in the document with security: [] and no x-payment-info (mirroring the aggregator's own documented treatment of free-tier endpoints just above in the same file, for the same reason: a bare probe can legitimately return 200 while quota remains, so advertising x-payment-info there would repeat the exact x402scan registration failure that section already documents), with the free-then-paid billing lanes already spelled out in each service's description.

Judgment call: upstream x402-places's detail route enriches with Yelp reviews, falling back to a deterministic seeded-random fixture when no Yelp key is set. three.ws has no Yelp key and CLAUDE.md bans fake data, so that whole enrichment path was dropped: places-detail returns only the real, live OSM record, documented in both the catalog descriptor and the route file.

DoD verified: npx vitest run tests/openapi-aggregator.test.js tests/service-catalog.test.js tests/aggregator-upstream-failover.test.js is 40/40 passing (these schema-driven tests exercise every PAID_SERVICES entry automatically, no new hardcoded cases needed per service). All 17 routes confirmed present in /openapi.json and /.well-known/x402.json with Solana listed first in every accepts[]/challenge. Live, real (non-fabricated) free-lane responses captured from every one of the 8 services' actual upstream: Open-Meteo (weather-guard-forecast), UK Carbon Intensity (carbon-now), RDAP (domains-check), arXiv/Crossref/Semantic Scholar (research-search), OpenLibrary/Gutenberg (books-search), MBTA GTFS (transit-plan: a real South-Station-to-Harvard Red Line itinerary), OSM Overpass (places-search: a real Starbucks near Boston), and GDELT (news-wire-query: a genuine 200 and a genuine 429, the latter confirming GDELT's documented ~1-req/5s throttle on this shared sandbox IP is real upstream behavior, not an app bug). npm run check:rules --paths <every touched file> exits 0 on all three commits; em-dash/en-dash grep clean on every new/changed line.

No owner-gated action in this order: every step is code, tests, docs, and a local build; no other-crypto-project reference landed in the diff, so no batched owner ask was needed either.

Commits: b239cb926 (weather-guard), 228dafe28 (carbon), 92cd17eb6 (domains), 506cf74c7 (research), 0ecc3007c (books), 48d05d9a7 (news-wire, transit, places), ad01ea88a (openapi.json free-tier-discoverability fix), e18987975 (docs + changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-14: order 517-integrate-18 closed (evolve)

Every buyer path (api/_lib/x402-user-payer.js's hosted a2a-hire/pay_and_call flow, packages/x402-mcp's stdio server, mcp-bridge's stdio client) registered only x402's CAIP-2 wire format (client.register('solana:*', new ExactSvmScheme(signer))). Most of the open x402 bazaar, and every service in the owner's own x402-suite, still speaks v1 (flat network strings like solana, price field maxAmountRequired), so none of our buyers could pay those sellers at all.

The installed @x402/svm package already ships registerExactSvmScheme(client, { signer }), a one-call helper that registers v1's three flat network strings AND v2's solana:* wildcard on the same client. All three buyer files were switched to it. Requirement-selection preference for v2 over v1 needed no new code: @x402/core's x402HTTPClient .getPaymentRequiredResponse() already prefers the v2 PAYMENT-REQUIRED header over a v1-shaped JSON body whenever a seller sends both; a test now locks that behavior in. Settlement receipts were confirmed (by reading both the real @x402/svm v1 facilitator's settle() and the owner's x402-domains seller's hand-rolled receipt shape) to already share field names (success, transaction, network, payer) across both versions, so the "normalize v1 receipts" task needed a test proving no normalization is required, not new code.

Judgment call: implementing task 2 surfaced two real spend-cap bugs that only trigger on a v1 payment, neither named in the order's 4 tasks but both block safe v1 support if left alone. api/_lib/x402-spending-cap.js's enforceCap only ever checked requirement.amount, so every v1 payment (which carries maxAmountRequired instead) threw spending-cap: requirement.amount required and was refused outright. mcp-bridge/src/x402-axios-client.js's spend-cap hook also only read selectedRequirements?.amount, defaulting to '0' for a v1 payment of any size, silently bypassing both the per-call and session ceilings instead of refusing them. Both were reproduced with a failing test first (tests/x402-spending-cap-v1-amount.test.js, mcp-bridge/tests/spending-cap-v1.test.mjs), then fixed with a one-line fallback to maxAmountRequired, then reverified green alongside every pre-existing test in both suites.

DoD verified: packages/x402-mcp/test/x402-v1-buyer.test.mjs feeds a captured v1 402 body (fields mirrored from the owner's x402-domains payments.ts, payTo/asset replaced with the $THREE mint and clearly-synthetic placeholders) into x402Client.selectPaymentRequirements() and gets back a valid v1 requirement with no payload/signature present (no signing). A second test proves v2 wins when a seller's response carries both the v2 header and a v1 body. node --test "test/**/*.test.mjs" in packages/x402-mcp is 13/13 green (10 pre-existing + 3 new); node --test "tests/**/*.test.mjs" in mcp-bridge is 14/14 green. Docs updated in packages/x402-mcp/README.md, mcp-bridge/README.md, and docs/mcp-agent.md.

A broader regression pass (tests/a2a-hire-handler.test.js, tests/api/mcp-wallet-provision .test.js, tests/api/mcp-agent.test.js, tests/x402-spending-cap-v1-amount.test.js, tests/x402-circuit-breaker.test.js) surfaced 4 failures in 2 files unrelated to this diff (scope-enforcement tests in mcp-wallet-provision.test.js and a tool_disabled test in mcp-agent.test.js). Confirmed pre-existing by stashing this diff's 3 production-file edits out, rerunning: same 4 failures reproduced against the unmodified baseline. Restored the stash and left those pre-existing failures for a separate order; not in scope here.

No owner-gated action in this order: THREEWS_AGENT_PAY_ENABLED was never touched, no live payment was made, and no other-crypto-project reference landed in the diff (test fixtures use the $THREE mint and clearly-synthetic THREEsynthetic... placeholders throughout), so no batched owner ask was needed.

Commit: 0a7085020 (buyer wiring, spend-cap fixes, tests, docs, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-15: order 518-integrate-19 closed (evolve)

Ported the owner's x402-agent-wallet ideas (approval threshold, merchant denylist, per-rail daily caps, HMAC-signed verdicts) onto our own storage: payment_sessions gains approval_threshold_usdc, blocked_hosts text[], per_rail_daily_usdc jsonb (migration 20261015100000_payment_session_policy_upgrades.sql), and a new payment_session_rail_reservations table backs the rolling 24h rail cap with the same eager reserve-then-roll-back-on-failure pattern spent_usdc already uses, closing the race a naive sum-of-past-executions check would leave open. api/_lib/pay/policy.js gained approvalVerdict, canonicalizePerRailDaily/NETWORK_TO_CAIP2/resolveRailDailyCapAtomics, signVerdict/verifyVerdictSignature/computePolicyDigest/computeRequestHash, and both spend-governor.js's reserveSessionSpend and simulate.js's replay call through the one evaluateCall() these live in, which is what makes the two structurally unable to disagree rather than merely unlikely to. Every verdict (allow or reject) now carries a policy_digest and an HMAC signature over {verdict, digest, request hash, timestamp}, stored on the execution via execute.js's four recordExecution call sites.

DoD verified:

  • Governor/guard/parity tests: npx vitest run tests/payment-session-governor.test.js tests/solana-trade-guards.test.js tests/payment-policy-simulator.test.js tests/advisory-lock-snapshot.test.js is 162/162 green, including a new parity describe block that runs 6 scenarios (plain allow, blocked host, per-tx ceiling, approval threshold, rail daily cap, insufficient budget) through both reserveSessionSpend and replay() and asserts identical verdict codes.
  • Signed verdicts: a new 'signed verdicts' describe block in tests/payment-session-governor.test.js verifies an allowed verdict's signature against its stored digest with verifyVerdictSignature, and proves altering any signed field (verdict code, digest, request hash, or timestamp) breaks verification.
  • UI: pages/payments.html's create and edit modals gained Blocked hosts, Requires-approval threshold, and Per-rail daily cap fields (the create modal's behind a collapsed "Advanced spend policy" <details> so the common case stays uncluttered), with inline validation on both. Browser-verified headlessly with Playwright against npx vite --port 3107: the create payload, the edit modal's pre-fill from a session row, the edit PATCH payload, and both modals' positive-number validation on the new fields all came back correct with zero console errors.
  • Docs (docs/payment-sessions.md: new body fields, new error codes host_blocked / requires_approval / rail_daily_exceeded, the signed-verdict note) and data/changelog.json updated, npm run build:pages run.

Never-blocked row: no PAY_VERDICT_HMAC_KEY existed in env before this order. api/_lib/env.js's env.PAY_VERDICT_HMAC_KEY getter falls back through PAYMENT_SESSION_SECRET (and so through WALLET_CAPABILITY_SECRET / WALLET_ENCRYPTION_KEY / JWT_SECRET), kept as its own env var rather than reusing PAYMENT_SESSION_SECRET directly so rotating the token secret (which revokes every live session) does not also invalidate the historical verdict record. A dev value is in .env.local (gitignored). Owner action still needed: set PAY_VERDICT_HMAC_KEY on the live Cloud Run service (gcloud run services update three-ws-api --region us-central1 --update-env-vars PAY_VERDICT_HMAC_KEY=<value>); until then production runs on the PAYMENT_SESSION_SECRET fallback, which works but means a secret rotation there also rotates the verdict-signing key. Not a stop-and-ask gate (the fallback keeps the feature fully functional), so it is not batched into an owner message; noted here for whoever runs the next Cloud Run env update.

No other owner-gated action: no push, no deploy, no on-chain write, and no reference to a crypto project other than $THREE landed in the diff.

Commit: 655b5a491 (schema, policy, governor, execute.js wiring, simulate.js parity, UI, tests, docs, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-15: order 519-integrate-20 closed (evolve)

Both venues that could already produce a requires_approval verdict (order 518's governor rule on payment_sessions.approval_threshold_usdc, and a2a-hire.js's own check against an agent's approval_threshold_usd) just rejected the call outright; neither was wired to a human. api/pay/execute.js and api/agents/a2a-hire.js now park the call on the existing generic Approval Inbox (api/_lib/approvals.js) the moment it crosses the threshold, answer 202 with a poll link to /approvals/<id> instead of failing it, and resume the exact reserve-then-settle or hire path the call would have taken inline the moment the owner approves, with the same idempotency key re-polling instead of parking a second request. A denied request answers 403 on the next poll, an expired one 410. No new UI: the approval page already renders any venue generically from the request's summary/amount/risk_notes, so the two new venues (payment_session_execute, a2a_hire) get the right card for free.

DoD verified:

  • Handler tests: npx vitest run tests/a2a-hire-handler.test.js tests/pay-execute-approval.test.js tests/agent-custody-guards.test.js is 57/57 green, covering create, approve (both executed-with-result and approved-but-not-yet-executed), decline, expire, and double-use (idempotent re-poll of an already-completed or already-parked hire) for both venues.
  • Browser verification: live-ran the full flow against the real local server (server/index.mjs on a spare port, real production DB) and a real Chromium session via Playwright, for the a2a_hire venue specifically (the payment_session_execute venue shares order 518's still-pending 20261015100000_payment_session_policy_upgrades.sql migration, already tracked as owner row 65; applying it is out of scope for this session). The run registered a throwaway account, minted two real agent wallets, set an approval threshold, signed the real-funds agreements, hired above the threshold (real 202 + pending_approval), loaded /approvals/<id> in a real browser and confirmed the rendered card (summary, $0.05 amount, Solana chain, threshold reason, working Approve/Deny buttons), denied it, confirmed the same idempotency key now re-polls to 403 approval_denied, and cleaned up every row it created. Screenshot and script were scratch (not committed, per repo-hygiene: no throwaway scripts land in scripts/).
  • Docs (docs/approvals.md's executor-registry sentence names both new sources) and data/changelog.json updated, npm run build:pages run.

Judgment calls: reused the existing generic /approvals/<id> page instead of building a new page, since it already renders any venue from generic fields; verified the a2a_hire venue live in a browser rather than payment_session_execute, for the migration-dependency reason above (both venues' request/response/retry logic and generic-card rendering are otherwise identical, and payment_session_execute is covered by its own 24 handler tests); inserted one synthetic agent_paid_services row directly via SQL for the throwaway hire fixture, since registering a real offer needs a publicly reachable target_url out of scope for a local smoke run, and the row is never fetched before the approval gate.

No new owner-gated action: no push, no deploy, no on-chain write, and no reference to a crypto project other than $THREE landed in the diff. The one owner action this order's code now also depends on (applying 20261015100000_payment_session_policy_upgrades.sql before payment_session_execute's approval path works in production) was already logged as owner row 65 by order 518's close-out; not duplicated here.

Commit: 9f40b6acc (approvals wiring for both venues, handler tests, docs, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-15: order 520-integrate-21 closed (evolve)

Built the SKILL-MD 1.0 generator the order asked for: api/_lib/skill-md.js reads the live /openapi.json catalog (buildOpenApiDoc()) and the real buildExactRequirements() rail logic and emits an agent-readable markdown catalog of every paid HTTP route (method, path, price, parameters), the Solana and Base payment rails, and a sample 402 challenge, per the upstream x402-skill-md format. Served at /.well-known/skill.md (api/skill-md.js, routed in vercel.json), linked from the /.well-known/x402.json discovery doc (service.skill_md field in api/wk.js) and from llms.txt (via a new data/pages.json machine-section entry).

Definition of Done, verified with real command output:

  • npx vitest run tests/skill-md.test.js: 5 passed. The suite imports the real published [email protected] npm validator (added as a devDependency, installed for real, not --package-lock-only) and asserts zero validator errors other than SM005 (no EVM rail) while Base stays unconfigured in the test env, matching the order's exact tolerance.
  • curl -s http://localhost:3107/.well-known/skill.md -D - (via PORT=3107 node server/index.mjs, since vite's dev proxy does not execute serverless handlers locally and port 3000 belongs to the owner's dev server): HTTP/1.1 200 OK, content-type: text/markdown; charset=utf-8, real generated body. npm run check:pages: 966/966 pages resolve.
  • Docs updated both directions (docs/x402-endpoints.md new "Agent-readable catalog: skill.md" section, docs/agent-skills.md cross-reference note), STRUCTURE.md row added, data/changelog.json entry added and npm run build:pages run (propagated into CHANGELOG.md, public/changelog*.json/.xml, public/llms*.txt, public/features.json, public/sitemap/index.html).
  • npm run check:rules clean on every touched file; npm run check:skills-pack clean.

Judgment call: the order's literal ask was to serve this at /skill.md, but that path is already live and linked as the unrelated Agent Skills (agentskills.io) connector file (public/skill.md, generated from data/skill-md.template.md by scripts/build-skills-pack.mjs). Serving the new document at /.well-known/skill.md instead avoids overwriting a different, actively-used file; both documents now link to each other so neither is an orphan. Discovered mid-session that public/skill.md is itself a generated artifact, not hand-authored; the one new line it needed went into the source template (data/skill-md.template.md), then regenerated via node scripts/build-skills-pack.mjs rather than hand-editing the generated file.

Noted, not fixed (pre-existing, unrelated, out of scope): tests/api/x402-discovery-green.test.js fails on /api/x402/transit-delays (a bazaar.info schema mismatch on route). Confirmed via a git stash push -u / apply A/B comparison that this failure exists on the tree with none of this order's changes applied, so it predates and is independent of this work.

No owner-gated action: no push, no deploy, no on-chain write, and no reference to a crypto project other than $THREE landed in the diff.

Commit: b3e416177 (generator, handler, routing, test, docs, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-15: order 521-integrate-22 closed (evolve)

Built the tools.json-driven registry the order asked for: 17 named tools, one per order 516's keyless x402 service (weather, carbon, domain RDAP check, research, books, news wire, transit, places), each a fixed-endpoint, fixed-price wrapper so a model calls the right endpoint with the right schema instead of composing a raw pay_and_call(url). Added spending_report, totaling every settled payment by merchant and day from a new local JSONL ledger (X402_MCP_LEDGER_PATH). Mirrored all 17 on the hosted agent-wallet server (api/_mcpagent/named-services-tools.js, wired into api/_mcpagent/catalog.js), reading the same tools.json so hosted and stdio agree on names, prices and schemas; each hosted tool gets a <name>_quote read tool plus the shared mcp-policy table gating payment on that exact quote (one service's preview can never unlock a different service's payment).

Definition of Done, verified with real command output:

  • node --test (packages/x402-mcp): 28 passed, 0 failed. Includes 6 new tests in test/named-services.test.mjs asserting a 1:1 registry-to-tool mapping, complete annotations and schemas, and test/ledger.test.mjs (8 tests) covering the ledger's env-path override, append/read round-trip, and merchant/day/since aggregation.
  • Dry run: named-services.test.mjs's "dry run: calling a named tool without confirm:true prices it and pays nothing" test asserts exactly one fetch call (the price probe, no payment) and the correct price for every one of the 17 tools.
  • packages/x402-mcp/README.md (tools table, env var), docs/mcp.md (new "Named commerce tools" section listing all 17, hosted and stdio) and data/changelog.json updated; npm run build:pages run (propagated into CHANGELOG.md, public/changelog*.json/.xml, public/features.json).
  • npm run check:rules clean on every touched file.

Judgment call: the registry's domain_check entry collided by name with a pre-existing, unrelated, free, read-only domain_check tool already served on the main /api/mcp server (Google Cloud Domains registration lookup, api/_mcp/tools/domains.js). Renamed this order's paid RDAP-based service to domain_rdap_check / domain_rdap_check_bulk everywhere (registry, hosted mirror, policy table, tests, docs) rather than touching the pre-existing tool; the two capabilities are genuinely different (RDAP protocol lookup vs. Google Cloud Domains pricing) and now coexist under distinct names.

While wiring the hosted mirror into the cross-cutting tests/mcp-remote-annotations.test.js and tests/api/mcp-agent-catalog.test.js, found both the hosted server's RESOURCE_DESCRIPTION (api/_mcpagent/discovery.js) and its system instructions (api/_mcpagent/dispatch.js) were already missing papertrade, portfolio, sniper/alert, duel and agent-commerce tools from an earlier, unrelated feature drop (7461c3ed0), which broke mcp-agent-challenge.test.js before this change touched anything. Fixed alongside this order's own additions since both files needed the same kind of edit either way; mcp-agent-catalog.test.js's hardcoded tool list had the same pre-existing gap (missing the event-market tools) and is fixed too.

Noted, not fixed (pre-existing, unrelated, out of scope):

  • npm run build:mcp-catalog fails with TypeError: studio.resourcesForSurface is not a function (api/_lib/mcp-census.js:118), from the same unrelated commit 7461c3ed0. This means public/mcp-catalog.json and the <!-- mcp-count --> markers for the hosted servers cannot be regenerated by the normal pipeline right now; the one marker this order's own diff touched (packages/x402-mcp/examples/README.md, stdio-only, unaffected by the census) was updated by hand with the real count (22, confirmed by running node examples/list-tools.mjs).
  • tests/mcp-spend-gate.test.js fails at import time (presetScopes is not a function from api/_lib/api-keys.js), and tests/mcp-use-cases-doc.test.js fails on unrelated launch-token, perps and predictions use-case mismatches. Neither file is touched by this diff and neither failure mentions any of this order's tool names; both predate this session (same unrelated commit history).

No owner-gated action: no push, no deploy, no on-chain write, and no reference to a crypto project other than $THREE landed in the diff.

Commits: 06650c721 (stdio package: registry, named tools, spending_report, ledger), 9d88c7f9d (hosted mirror, policy table, discovery/instructions fixes), 16b544a33 (docs, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-15: order 523-integrate-24 closed (evolve)

Every settled a2a-hire (api/agents/a2a-hire.js) now writes one signed fulfillment attestation row via the new api/_lib/trust/hire-attestations.js ({hire_id, payment_signature, outcome: delivered|failed|disputed, evidence_hash, signed_by}), signed off-chain with the platform attestation key api/_lib/attest-event.js already uses for on-chain provenance memos; a node:crypto/tweetnacl import stayed out of api/_lib/a2a-hire-phases.js (shared with client code src/agent-screen-hire.js, src/shared/forge-frames.js), so only its pure outcome-classification and canonical-JSON helpers (isAmbiguousPaymentError, classifyHireOutcome, sortedStringify, buildHireAttestationPayload) live there; the real hashing and signing is server-only. A client-side payment timeout is recorded disputed, never inferred delivered, since x402 verify-then-settle means the provider may already have been paid server-side. New migration 20261015120000_hire_fulfillment_attestations.sql (table hire_attestations, unique on hire_id; agent_paid_services.reputation_discount_enabled column) applied with node scripts/apply-migrations.mjs --apply --file <name> so it did not sweep the 11 other already-pending migrations from concurrent work in this worktree.

The recency-weighted fulfillment rate (weightedFulfillmentRate/computeFulfillmentRate, 30-day half-life, minimum 5 attestations before it counts) is now an additive input to the free Financial reputation score's Settlement reliability pillar (src/shared/agent-financial-reputation.js, REPUTATION_VERSION bumped 3 to 4 so cached scores recompute), read in api/_lib/trust/wallet-reputation.js. Verified end to end: the pillar's detail string (e.g. "90% hire fulfillment over 12 hires.") and facts.hire_ fulfillment_rate/facts.hire_fulfillment_sample_size flow unmodified from computeReputation through GET /api/agents/:id/reputation to the generic pillarRow() renderer in src/shared/agent-reputation.js (mounted on the agent page's Trust tab via src/agent-wallet-hub/tabs/reputation.js), confirmed by a direct computeReputation() call producing that exact detail string, no UI code changed.

Task 3: api/_lib/trust/hire-discount.js adds an off-by-default agent_paid_services.reputation_discount_enabled flag; when set, a2a-hire.js mints a 5-minute, slug-scoped, ed25519-signed discount voucher from the hirer's live reputation tier (getEffectiveReputation), and the public paywall (api/x402/service.js) only ever trusts that platform signature, never a caller-supplied claim, failing closed to full price on anything missing, expired, mis-scoped, or forged.

Definition of done, verified:

  • tests/a2a-hire-phases.test.js gained an 11-case describe('fulfillment attestation (pure core)', ...) block (32/32 tests pass in the file).
  • New tests/hire-attestations.test.js: describe('verifyHireAttestation', ...) (5 tests) signs with a locally-generated ed25519 keypair injected via ATTEST_AGENT_SECRET_KEY and verifies verifyHireAttestation accepts a correctly-signed payload and rejects a tampered payload, a forged signature, a signer mismatch, and malformed input; describe ('weightedFulfillmentRate', ...) (4 tests) covers the empty case, an all-delivered sample, disputed never counting as delivered, and recency decay (9/9 tests pass in the file).
  • tests/wallet-reputation.test.js gained a describe('hire fulfillment rate', ...) block (4 tests: below-minimum contributes 0 points with facts.hire_fulfillment_rate: null, at-minimum scores it, a poor rate scores lower than a strong one at the same sample size, combined with the ESTABLISHED fixture it never exceeds the pillar max) plus the REPUTATION_VERSION pin updated 3 to 4 (27/27 tests pass in the file).
  • Agent page: traced above, confirmed with a direct computeReputation() call.
  • Docs and changelog: new "Signed fulfillment attestations" section in docs/agent-abilities/chapters/06-identity-reputation.md, a data/changelog.json entry, docs/ALL.md regenerated via node scripts/combine-docs.mjs, npm run build:pages run.

Pre-existing, unrelated, out of scope (confirmed by reverting this diff's one line in api/_lib/x402-user-payer.js and re-running): tests/api/mcp-agent.test.js and tests/api/mcp-wallet-provision.test.js each fail 2 tests (pay_and_call tool listing/ scope-enforcement) identically with or without this diff; both failures trace to commit 9d88c7f9d, already logged above in the 2026-10-14 entry for order 500-integrate-01's neighbor work. npm run check:docs-freshness was already over budget by 130 across 15 unrelated docs (none the one this order edited) before this diff; not touched.

No owner-gated action: no push, no deploy, no on-chain write (the attestation and discount voucher are both off-chain ed25519 signatures over a platform-controlled key, never a Solana transaction), and no reference to a crypto project other than $THREE landed in the diff.

Commits: 2f8110a92 (attestation infra, reputation blending, discount pricing, tests), 613dcd785 (docs, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-15: order 524-integrate-25 closed (evolve)

Turned any public OpenAPI/Swagger spec URL or GitHub repo into a draft, importable three.ws skill bundle end to end.

Step 0 measured: specs/SKILL_SPEC.md confirmed the four-file bundle shape (manifest.json, SKILL.md, tools.json, handlers.js); community-skills/tools/validate.mjs validates a different directory shape (a flat registry-manifest layout, not an individual bundle), so Task 3's "validate with community-skills/tools/validate.mjs" was satisfied by a new purpose-built packages/skill-import/src/validate.js instead (judgment call, logged at the time); packages/tool-sdk/src/mcp-adapter.js confirmed the ctx.fetch/ctx.secrets handler contract generated handlers must target; api/skills/ had no existing import route.

Task 1: new packages/skill-import package. openapi.js parses OpenAPI 3.x and converts Swagger 2.0 via swagger2openapi; operations.js dereferences $refs into a normalized operation list; generate-tools.js/generate-handlers.js/generate-manifest.js/ generate-skillmd.js turn that into a full bundle, with handlers calling ctx.fetch and filling auth from ctx.secrets at call time (never hardcoded). Two bugs found and fixed while wiring this live against the public Petstore API: a real API mislabeling a plain-text body as application/json crashed the handler (now always reads text first, parses JSON only on success); server/index.mjs strips a trailing ".js" from API paths before routing, so the generated bundle's own handlers.js 404'd through the dynamic catch-all until the route accepted both "handlers.js" and the bare "handlers" it actually receives.

Task 2: a repo with only a README produces a declarative bundle (declarative.js): SKILL.md only, no tools.json, no handlers.js.

Task 3: POST/GET /api/skills/import (api/skills/import.js) generates and lists drafts through the SSRF guard (fetchSafePublicUrlPinned), validates with the new validate.js, and saves via api/_lib/openapi-skill-imports.js with enabled hardcoded false. Beyond the literal task list: built the previously-missing GET/PATCH/DELETE /api/skills/import/:id route, since without it a draft's enabled flag could never change and the feature would have been a dead end once generated (judgment call, justified by "no half-finished implementations"). Also built the bundle host api/skills/import/[id]/[...file].js serving manifest.json/SKILL.md/tools.json/handlers.js anonymously by id. New migration 20261015130000_openapi_skill_imports.sql, new rate limit openapiSkillImportUser (20/10m). Fixed a real, pre-existing vercel.json routing collision (/api/skills/([^/]+) was swallowing /api/skills/import before the literal route could match) and confirmed via the server's route-compile anchoring (^...$) that the new two-segment :id route needs no further vercel.json change.

Task 4: CLI. three-ws skills import <url> (full URL or bare owner/repo) posts to /api/skills/import; isImportUrl() disambiguates a URL/owner/repo from a community slug (slugs never contain /). New three-ws skills drafts lists the caller's drafts with enabled state.

Task 5: UI. New "From an API" tab on the existing /skills/import page (reused its tab/sc-* CSS shell rather than a new page, since both features are "import a skill from somewhere external"): submit a URL, see the generated tool list in a result card labelled disabled, manage drafts (enable/disable/delete, manifest.json link) from a persistent list, signed-out gate state. Verified live in a real headless Chromium session against the dev server (a real session cookie minted via createSession, not a mock): generated a skill from the public Petstore OpenAPI spec end to end (19 tools), enabled it, deleted it, checked the signed-out gate render; zero console errors, zero failed network requests at every step.

Task 6: new docs/skills.md section "Importing a skill from an OpenAPI spec or a GitHub repo" with a real curl example against the Petstore spec; data/changelog.json entry; STRUCTURE.md row for packages/skill-import and its routes; npm run build:pages run.

Definition of done, verified:

  • packages/skill-import/tests/skill-import.test.js: a vitest serves the synthetic sample-openapi.json fixture from a real local node:http server, asserts the generated bundle passes validate.js, and loads its handlers (4/4 pass).
  • tests/openapi-skill-import.test.js: 7 tests against the real API handlers (import, the new :id route, and the bundle host) with a real local HTTP server standing in for the upstream spec (only the SSRF guard's network check is swapped for a passthrough, since loopback is exactly what it exists to refuse in production); covers auth/scope rejection, full real generation end to end asserting enabled: false on both the response and the raw stored row, owner-scoped enable/disable/delete, and bundle file serving including the .js-extension-stripping edge case.
  • "Imported skills are never enabled automatically": asserted directly in the test above (db.rows[0].enabled === false immediately after creation) and confirmed live: the generated draft from the real Petstore spec came back enabled: false and stayed that way until the dedicated PATCH.
  • "Importing a real public OpenAPI spec end to end in dev produces a draft skill whose tool call works": verified live twice, once via curl/node against the real Petstore demo (logoutUser call succeeded against the live API after the content-type fix) and once via the full browser UI flow above.
  • Package README (pre-existing from this order's own earlier work), STRUCTURE.md row, docs section, changelog entry: all present, npm run build:pages run.
  • npm run check:rules -- --paths <every touched file>: clean (fixed five em-dash violations first, one of them in this order's own new docs prose, the rest already in the prior session's generated code comments).
  • Full regression pass, no failures: packages/skill-import/tests + tests/ openapi-skill-import.test.js (11/11), packages/three-ws-cli full suite (82/82), tests/ api-rate-limit-wiring.test.js + tests/api/rate-limit-buckets.test.js (24/24, covers the touched api/_lib/rate-limit.js).

No owner-gated action: no push, no deploy, no on-chain write, no crypto project other than $THREE in the diff. One judgment call beyond the task list already logged above (the missing enable/delete route); the validate.mjs substitution and the agents:write/agents:read scope reuse on a non-agent-scoped resource were carried over from the prior session's work and are logged there.

Commits: d7bcabdcc (package + API routes + migration + rate limit + vercel.json fix + tests), df5e2d380 (CLI), e55c84a15 (UI tab), 7c28d186c (docs, STRUCTURE.md, changelog), this closing commit (PROGRESS log, order-file removal).

2026-10-23: order 530 (registry cleanup), partial, owner-gated

Prepared: scripts/publish-mcp-servers.mjs --deprecate (dry run unless --apply; --dry-run also prints the plan), a DEPRECATIONS list (claude-code-explorer-mcp, data-workbench-mcp), solana-memo-media-mcp bumped to 0.1.1 to match npm (registry still has 0.1.0), ecosystem doc count replaced by its source of truth. Remaining and owned by the owner: run the registry publish and deprecate commands (external writes). Order file left in place.

2026-10-23: order 529 (MCP registry sentinel), done

api/cron/mcp-registry-sentinel.js plus api/_lib/mcp-registry-sentinel.js, daily cron in vercel.json, 12 vitest cases, ops doc docs/mcp-registry-sentinel.md. Code landed in 218f00660 (swept into a concurrent agent's commit); this closing commit adds the doc, PROGRESS entry and order-file removal. Deploys with the next owner-approved deploy.

2026-10-23: order 542 (3D-AR-Studio source into the monorepo), code done, publish and push owner-gated

Step 0 measured: in-app studio 5,072 lines (src/ar-studio.js + src/ar/*), npm 3d-ar-studio at 0.3.1, upstream repo already at 0.3.2. Upstream's shared modules were functionally identical to src/ar/* (comment style aside) except the share-URL helpers (upstream take the hosting page URL, not an origin) and the room client (upstream adds a connect timeout and lazy optional colyseus.js). Imported at packages/ar-studio (0.3.2, 75 package tests); src/ar-studio.js, src/irl.js, src/ar/webxr.js and the tests/ar-* suites now import the package modules; the seven duplicated src/ar/* files were deleted; src/ar/studio-net.js is a thin subclass that only resolves the Colyseus host. tests/ar-studio-package-parity.test.js guards the server/package wire schema. Mirror entry added to data/standalone-mirrors.json; STRUCTURE.md, docs/ar-studio.md, changelog updated. Commits: 111fc15c7 (swept in by a concurrent agent), 22a5a1f49, 0ac2f37ac, plus this closing commit. Remaining, owner: npm publish of 3d-ar-studio 0.3.2 from packages/ar-studio and npm run sync:repos -- --only 3D-AR-Studio (stage, review, push). Order file removed (code complete; the two remaining steps are the gate-2 commands reported in the final message).

2026-10-23: order 543 (Flappin UFO back at /play/ufo), done

Step 0 measured: the retired page was a static notice with no game; the owner's flappin-ufo ran on DOM physics that a server cannot verify. Built a shared deterministic 60 Hz sim (src/play/ufo/sim.js) used by the browser, the ghost and the server verifier; levels 1 to 30 proven clearable by a bot. Code landed in 22a5a1f49 (swept in by another agent's commit): api/play/ufo-runs.js, the ufo_runs migration, game.js, the page, the ufo_score metric on the unified leaderboard and a /rankings tab. Follow-up commit adds the test fixes, docs/play-ufo.md, STRUCTURE row, pages.json entries and changelog entry. Evidence: 16/16 vitest cases (real handler on PGlite: POST, GET, ghost, tamper, 422, 409), headless Chromium at about 60fps on desktop and Pixel 7 emulation with no console errors. Remaining, owner-gated: migration 20261023150000_ufo_runs.sql applies with the next approved deploy (npm run db:status first); until then the production endpoint returns an error and the page shows its retry state.

2026-10-23: order 538 (oracle wallet prior from the database), done

Step 0 measured: the committed prior snapshot dated 2026-06-16 (273 wallets: 128 smart money, 145 KOL), read by known-wallets.js through a synchronous JSON load; gmgn-seed already upserts wallet_reputation daily and was still serving 200 via its kol-quest fallback. known-wallets.js now loads the prior from wallet_reputation (mainnet, coins_traded < 5, labels smart_money/kol/sniper) into an in-process map with a 10 minute TTL (30 seconds after a failed read, last good rows kept), single-flight; knownWallet() stays synchronous and callers (sources.js, store.js, trades.js, pump-fun-mcp.js) await loadKnownWallets() first. An empty table serves an empty prior and logs once. The snapshot file was deleted. gmgn-seed now files KOLs as kol (score 74) instead of neutral, otherwise the 145 KOLs would have dropped out of the prior. Evidence: a live cron run into a PGlite copy of the table fetched 273 wallets and upserted 273 (128 smart_money, 145 neutral before the label fix), the oracle read back 128 as prior; tests/oracle 21 files and 282 tests pass in an isolated install; new tests/oracle/gmgn-seed-prior.test.js covers cron to prior. Code landed in 22a5a1f49 (swept in by a concurrent agent), the cron label fix and closing files in this commit. Left in place on purpose: the copy-trading directory snapshot and its builder script, which carry names, handles, BSC rows and 30 day PnL that wallet_reputation does not hold, and which a different endpoint reads. Production note: existing KOL rows are relabeled kol on the next gmgn-seed run after deploy.

2026-10-23: order 522 (integrate 23, outbound receipts), done

Step 0: everything but the schema was already committed. Resolver api/_lib/x402/solana-settlement.js, cache layer api/_lib/x402/outbound-receipts.js, endpoint api/x402/outbound-receipts.js landed in 420a88d88; the Outbound tab on /receipts (pages/receipts.html, src/receipts-outbound.js) in 9e1ff2270. Closing step: applied only 20261015110000_x402_outbound_resolved.sql to production with node scripts/apply-migrations.mjs --apply --file 20261015110000_x402_outbound_resolved.sql (additive, idempotent; the other pending migrations were left alone). npm run db:status lists it applied. Vitest: tests/solana-settlement.test.js plus tests/x402-outbound-receipts.test.js 22 pass, tests/x402-outbound-receipts-endpoint.test.js 6 pass. Row 66 in production-100-OWNER-ACTIONS.md is resolved by this apply.

2026-10-23: order 541 (integrate 42, readme-3d in the product), done

Step 0 measured: packages/readme-3d had no product surface; its README forge example polled the wrong fields. Shipped POST /api/readme-3d (SSRF-hardened fetchModel, byte-budget decimation, 12 per 10 min per IP), convertBytes/convertTriangles in the package, a "GitHub README 3D" action in both forge export menus, the avatar page download menu, the dashboard avatar menu and the model page, and the /readme-3d page (three.js STL preview, copy, download .md). The stale curl example lived in packages/readme-3d/README.md (not the root README, whose example was already correct); it now submits and polls with job_id and glb_url. Code landed in a4d4e4cc1, 22a5a1f49 and 2b7bb3921 (swept in by concurrent agents). Evidence: 22/22 package tests, 4/4 tests/readme-3d-api.test.js cases (real handler over HTTP), live API on real three.ws GLBs (fox: 576 facets, 83 KB; default avatar: 13,317 to 1,014 facets, 146 KB; readme-3d check passes), headless Chromium rendering the fox with copy and download controls and no errors from the page's own code. check:pages flags /readme-3d only because dist/ is stale until the next build. Remaining, owner: none beyond the normal deploy.

2026-10-23: order 533 (integrate 34, X post reader and /x-video), code done, two lines open

Step 0 measured: both callers read X through oEmbed (text only, no metrics or media). The shared keyless reader api/_lib/x-syndication.js (X syndication, then fxtwitter, then guest-token GraphQL when X_BEARER_TOKEN is set) now backs handleSocialXPostImpact in api/pump-fun-mcp.js and src/social/x-post-impact.js; POST /api/x-video/extract and the host-locked GET /api/x-video/download proxy (https video.twimg.com and pbs.twimg.com only, no credentials, ports or redirects) plus pages/x-video.html and src/x-video.js ship the free tool. Code landed in 0ad88fb38, 22a5a1f49, ed75c5946 and 7718bf4fb (all swept in by concurrent agents). Evidence: tests/x-post-impact.test.js, tests/x-syndication.test.js, tests/api/x-video.test.js and tests/pumpfun-mcp-tools.test.js 49 of 49 pass, including host refusals; check:rules exits 0; a local server on a spare port returned all mp4 variants (1080p, 720p, 480p and lower) for a real public video post, no_video for a text post, streamed a real mp4 through the proxy with 200, and answered 400 invalid_media_url for example.com. Open: (1) browser verification of the page and npm run check:pages, which needs a Vite build; the shared node_modules was being rewritten by a concurrent npm install (vite missing, then @upstash/qstash and three unresolved) for the whole session, so dist/ is empty of every recent page. Owner: whoever next has a stable install; run npx vite build && npm run check:pages and open /x-video. The order file stays until then. The live check used a public Tripo post because the recent @trythreews posts the timeline returned contain no video.

2026-10-23: order 537 (integrate 38, market anomaly feed), done

Step 0 measured: no upstream tarball was needed. The detectors (modified z-score with MAD, EWMA deviation, rate of change) were written from the algorithm description in api/_lib/market-anomaly.js and fed from real series: CoinGecko hourly sparklines (top 100), DeFiLlama Solana TVL and Solana stablecoin supply history (built first), the stablecoin board, and Hyperliquid funding history plus open interest snapshots (no public OI history, so the cron records market_anomaly_samples each tick). Gas and whale processors were not built (gas is EVM-only, whale detection already exists in Oracle). Shipped in 35afe5904 (engine, sources, store, cron, migration, paid route, alert kind end to end, alerts-mcp), 22a5a1f49 and 2b7bb3921 (registry entry, fixtures and tests swept in by concurrent commits), 3f1eb493d (ring catalog), plus the docs commit that closes this order. Migration 20261023140000_market_anomalies.sql applied alone with --apply --file; 20 unrelated migrations stay pending. Evidence: 32 detector and alert tests pass on live-recorded fixtures; the cron run wrote 10 real events (stablecoin pegs) and a rerun wrote 0 (dedupe), skipping 20 open interest series as insufficient_history (have 2, need 29); /api/x402/market lists market-anomalies and the route answers 402. Open interest begins scoring after about 7 hours of snapshots.

2026-10-23: order 531 (integrate 32) MCP catalog trust grades

Every server in public/mcp-catalog.json now carries trust (0-100 score, A-F grade, per-fact breakdown) from scripts/lib/mcp-trust.mjs; /mcp-tools shows the grade with a tooltip breakdown and sorts by it; rubric in docs/mcp.md. Commits: 2b7bb3921 (rubric, builder, page, test and docs, swept in by a concurrent commit), e4bdbdf34 (regenerated catalog), plus the commit that adds the changelog entry and closes this order. Evidence: build-mcp-catalog --check and build-mcp-tool-history --check pass at 704 tools on 51 servers, tests/mcp-catalog.test.js 10 passed, browser-verified sort and tooltips with no console errors. Network facts (npm live, registry listing, endpoint reachable) are deliberately left to the registry sentinel; the rubric scores local publish-readiness only.

2026-10-23: order 548 (Robinhood launch venues), code done, commit and deploy owner-gated

Step 0 measured: NOXA dead since 2026-07-13; Robinhood Chain 4663 makes about 9.6 blocks/s; a 5,000-block live probe found 1,409 mint logs, 97 tokens and 12 venue launches (Pons V2 9, others 3). Built: venue catalog (51 contract and selector pairs from the upstream launch-relay catalog, generated by workers/robinhood-feed/scripts/build-venue-catalog.mjs), mint-log venue discovery (src/mints.js), deployer reputation and dev-sold (src/reputation.js), venue/deployer/dev_sold/flags on the normalized launch, api/_lib/robinhood-feed.js merged into /api/v1/robinhood/launches, venue badge, deployer record and flag chips in src/markets-robinhood.js. Tests: 50 worker tests (real-log fixtures for 8 venues, tests/fixtures/venue-launches.json), 2 API mapping tests. NOXA removed from api/_lib/robinhood.js.

Remaining: (1) owner approves committing the diff (it names other launchpads and holds real third-party addresses in fixtures); (2) owner deploys the robinhood-feed worker and sets ROBINHOOD_FEED_URL on three-ws-api; (3) a full 24 h backfill needs a paid Robinhood Chain RPC (RH_RPC_URL), the public RPC rate-limits at 429 with a 60 s reset window, which the scan now waits out but slowly; (4) root vitest suites (tests/api/v1-robinhood.test.js, robinhood-play-endpoints.test.js, robinhood-feed-launches.test.js) were not run because the root npm ci is blocked by a lockfile out of sync with package.json (commit 22a5a1f49 clobbered both).

2026-10-23: order 534 (integrate 35, news reader lanes and trading read), code done, two checks open

Step 0 measured: for a 403 publisher page the reader returned 0 paragraphs. Built Wayback, SEC EDGAR and sibling lanes (api/_lib/article-lanes.js) behind the SSRF guard, the structured trading read (parseTradingRead in api/news/article.js, rendered in src/news-article.js), the news_resolve MCP tool, and the excerpt cap on every lane. Commits: 22a5a1f49 (most of the code, swept in by a concurrent commit), 33df785d3 (gzip decoding at the guard, ladder test, comment fixes), 5bcc701ec (docs and changelog). Evidence: 7 test files, 105 tests pass. Live Wayback: https://www.nytimes.com/2022/11/11/business/ftx-bankruptcy.html (page 403, reader 0) resolved with extraction: wayback from the 2025-10-28 snapshot, excerpt capped at 173 characters. Live EDGAR: a Grayscale Bitcoin Mini Trust 8-K resolved to 11 clean paragraphs.

Remaining: (1) the live LLM trading read was not seen end to end because the provider chain returned 429 and timeouts locally; unit tests cover parsing and the response flow. (2) The reader UI (tradingReadHtml, .art-trading styles) has not been opened in a browser; run npm run dev and open a story at /markets/news. The order file stays until both are confirmed.

2026-10-23: order 528 (integrate 29, W3AG wallet and payment accessibility), done

Step 0: about 127 address-shortening helpers existed across src/ and dashboard-next/; all now call src/shared/address.js (shortAddress). Only two non-address truncators remain by design (dashboard-next/components/drawer.js, reputation-ui.js). Shared trap in src/shared/focus-trap.js covers the connect picker, payment modal, vault dialogs and add-funds sheet. Most of the codemod and the new modules landed in 22a5a1f49 via another agent's sweep; the closing fixes are in 1a18750d8 and c59ba09ca (committed by a concurrent sweep with accurate subjects). Evidence: tests/e2e/w3ag-wallet.spec.js 18/18, a11y-top-pages gate green including /wallet, /agent-wallet, /claim-wallet, /vaults (the one failure, / progressbar name, fixed). Full vitest has 54 failing files from redis/env and other agents' pages work, none touching these files after the connect-button test fix.

2026-10-23: order 540 (integrate 41, memo inscription reader), code done, live mainnet check pending

Step 0 found the reader already shipped in 876f0a4fc: packages/solana-memo-media-mcp/src/lib/inscriptions.js (first/1, firsts/1, spl404 reassembly with per-part and whole SHA-256, missing-part reporting, v1 and inner-memo aware), the read_inscription MCP tool, /onchain rendering (rasters from verified bytes, SVG and HTML as inert text), docs and changelog. npm test --prefix packages/solana-memo-media-mcp passes 29/29, including the never-rendered-live test.

Remaining: no real multi-part inscription signature could be found (the three upstream repos' tarballs hold no mainnet signatures), so the live hash-match check is pending; vectors are built from the upstream format spec. Owner: supply one real signature, then run /onchain with it. Order file left in place.

2026-10-23: order 525 (integrate 26, Anchor IDL to agent skill), done

Step 0 measured: packages/skill-import and contracts/idl/ existed, and the IDL importer, Borsh runtime, POST /api/skills/import { idl_url | program_id, cluster }, docs (docs/skills.md, package README) and changelog entry had already shipped in b00ee311a and 22a5a1f49. This pass re-verified every line and closed the order. Evidence: npx vitest run packages/skill-import 18 of 18 pass, covering the validateBundle check on skill_license.json, Anchor-coder-identical account decoding, preview and send tools asserting no signer or secret is touched, and the risk panel. The live cluster check could not run: EdngSwx... (skill_license) returns a null account on both devnet and mainnet-beta, so decoding uses the recorded-fixture path; a live decode needs the program deployed (owner, see contracts/skill-license/DEPLOYMENT.md).

2026-10-23: order 527 (integrate 28, llms.txt and install.md), done

Step 0 found the work already shipped in 51ff59e99: llms.txt links the curated docs as /docs/<slug>.md, llms-full.txt inlines doc bodies (40 KB cap, docs naming other crypto projects left out), /install.md is generated from the MCP directory, and tests/llms-index.test.js carries the conformance checks. This commit only closes the order. Evidence: npm run build:pages clean, tests/llms-index.test.js 11 passed, /install.md and /docs/mcp.md return 200 text/markdown from server/index.mjs, and 60 of 65 llms.txt links return 200 on Vite; the other five are server-handled (/api/mcp 402 by design, /agents and /docs 301, /sitemap.xml and /.well-known/skill.md need the API handlers and a built dist, not available in this sandbox).

2026-10-23: orders 532 and 533 (integrate 33 and 34, X guest read rung and post reader), done

Step 0 found both already shipped in 0ad88fb38 (reader, guest client, video tool, docs, changelog, STRUCTURE rows) with the oracle pieces (api/_lib/oracle/x-account.js, social-feed.js, api/cron/oracle-social-ingest.js, vercel.json cron) in the same tree. This pass re-verified every line and closed both orders. Evidence: npx vitest run on kol-leaderboard, oracle/sources, x-guest, x-post-impact, x-syndication, api/x-video and pumpfun-mcp-tools: 7 files, 87 tests pass (the download proxy host-refusal case is in tests/api/x-video.test.js). Live from server/index.mjs: POST /api/x-video/extract on a real public post returned author, metrics and media variants, a nonexistent post returned post_unavailable, /api/x-video/download?url=https://evil.example/a.mp4 returned 400, and fetchGuestProfile('trythreews') returned 8986 followers keyless. npm run check:claude passes. npm run check:pages only reports pages missing from dist/ because this sandbox has no vite build.

Owner note: whether the keyless guest reads fit X's terms for a commercial surface is an owner call; setting TWITTER_BEARER_TOKEN keeps the paid API first.

2026-10-23: order 536 (integrate 37, $THREE post impact), partial

Step 0 found everything shipped: scripts/x-archive-price-impact.mjs, api/_lib/x-content/price-impact.js, the Posts view in public/oracle-coin.js, the +24h priceReturn24h lift in api/_lib/x-content/outcomes.js, docs, STRUCTURE row and changelog entry. tests/x-archive.test.js, tests/x-content-outcomes.test.js and tests/x-post-impact.test.js pass (81 tests). The x:archive:price npm script had gone missing from package.json and was restored in 618b3fd84. public/data/trythreews-post-impact.json (197 posts, all three horizons measured, generated 2026-10-23) is real data.

Remaining: the three-width browser check of the Posts view and a fresh npm run x:archive:price run. This session's checkout had a broken node_modules (vite, playwright-core and lru-cache failed to resolve) and GeckoTerminal answered 429, so neither could run. Owner: whoever next has a healthy install runs the script and screenshots at 375, 768 and 1280.

2026-10-23: order 534 follow-up pass, still partial

Re-ran the four test files (74 pass). The live trading read still did not render: local provider chain returned kilo 429, ovh timeouts, llm7 401, and Vertex Gemini 403 (project billing dunning, owner action). The browser render check of .art-trading was prepared but could not run because concurrent npm install runs kept node_modules (vite, playwright) in a half-written state for the whole pass. Remaining: open a story at /markets/news under npm run dev once an LLM provider answers; owner: restore Vertex billing or supply a provider key. Order file left in place.

2026-10-23: orders 539 (integrate 40) and 545 (integrate 46), code done, two checks open

539: Gini, funder seeded share and v1-launch signals (api/_lib/coin/holder-signals.js, api/pump/token-stats.js, oracle features) shipped earlier via a sweep commit; this session added the trader-sheet cells in src/launch-detail.js. Local /api/pump/token-stats?mint=<THREE> returns gini 0.4952 (n=20), launch_version V0, seeding available:false (trade history unavailable from the capped upstream). Tests: coin-holder-signals, pump-token-stats, coin-intel, 62 pass. The dev server proxies /api to prod, so the page cells read "-" until the next deploy. 545: 3D funder graph (src/funder-graph.js, vendored engine in src/vendor/swarming/) was already wired; this session found and fixed a real layout bug (500 nodes drifted to radius 1776, off camera) with per-axis gravity, bounded charge range and a percentile-based fit, covered by tests/funder-graph.test.js. Browser-verified at 390 (reduced motion, 60fps idle, graph framed), 768 and 1280 with a 500-node harness, zero console errors. 60fps with 500 nodes on a real GPU is unmeasured: this box only has software GL. Remaining (orders kept): 539 page cells need a post-deploy browser check against live /api (owner deploys); 545 needs the 500-node 60fps check on a real GPU.

2026-10-23: order 547 (integrate 48, Robinhood Crypto MCP tools), done

Four read-only tools (robinhood_chain_stats, robinhood_token_prices, robinhood_recent_launches, robinhood_wallet_holdings) in api/_mcp/tools/robinhood.js, registered in the catalog. Evidence: a live robinhood_chain_stats call returned block 96324688 on chain 4663, tests/mcp-catalog.test.js plus tests/api/v1-robinhood.test.js 28 passed (the launches test was updated for the NOXA removal), audit:mcp-safety passes. Docs section added to docs/mcp.md.

2026-10-25: order 510 (integrate 11, per-coin earnings source), done

Code shipped in 98744d430. Close-out found 20261015090000_agent_coin_earnings_timeline.sql already [applied] in production (npm run db:status), the five new columns present on agent_coin_earnings, tests/pump-creator-fees.test.js plus tests/agent-earnings-board.test.js 24/24, and launch-detail answering 200 with creator_fees.distributed and reconciliation keys for a real earning mint (values fill when the snapshot cron next runs). OWNER-ACTIONS row 63 removed. Production still needs the normal deploy to serve the new code.

2026-10-25: order 534 closed

Both open checks confirmed: the live LLM path returned a full trading read (assets with exposure and direction, priced_in, horizon, counter) via the platform chain, and the reader's trading-read markup rendered with the counter-argument section (no dev server run, per standing rule). Six news test files pass (94 tests). Order file removed.

2026-10-25: order 536 (integrate 37), re-checked, still partial

tests/x-archive.test.js and tests/x-content-outcomes.test.js pass (75 tests). npm run x:archive:price again failed on a GeckoTerminal 429 before writing, so public/data/trythreews-post-impact.json (generated 2026-10-23) is unchanged. The three-width browser check was not run: it needs a dev server, which this session was not permitted to start. Remaining: rerun the script when GeckoTerminal recovers, then screenshot the Posts view at 375, 768 and 1280. Order file left in place.

2026-10-25: order 530 (registry cleanup), decisions recorded, publish and deprecate still owner-gated

Step 0: registry returns 80 entries under io.github.nirholas, 55 with a local manifest, 26 registry-only (one is listed in both counts via a renamed manifest). Decision table, all 26 re-checked against the registry's repository URLs: deprecate 2 (claude-code-explorer-mcp, source repo is the DMCA-notice repository; data-workbench-mcp, repository returns 404, source lost, so never republished); leave 24 (each points at a live repository the owner controls, with its own package or remote, so adopting them would copy third-party-project manifests into this repo and trip the other-project commit gate for no gain; the sentinel's registry_only finding keeps them visible). solana-memo-media-mcp is at 0.1.1 locally and on npm, registry still at 0.1.0. Dry run (node scripts/publish-mcp-servers.mjs --dry-run) prints: publish 3d-ar-studio 0.3.2 (npm and registry), publish solana-memo-media-mcp 0.1.1, deprecate both names in the registry and on npm. Remaining, owner (external writes): node scripts/publish-mcp-servers.mjs (publishes the two behind entries) and node scripts/publish-mcp-servers.mjs --deprecate --apply. Order file left in place.

2026-10-25: order 539 (integrate 40), re-checked, still partial

Code was already shipped (api/_lib/coin/holder-signals.js, api/pump/token-stats.js, oracle features in api/_lib/oracle/fit.js and conviction.js, Launch tx and Gini cells in src/launch-detail.js, changelog entry). tests/coin-holder-signals.test.js plus tests/pump-token-stats.test.js pass (24 tests). A live readLaunchVersion call for the $THREE mint returned is_v1: true. Remaining: the curl against /api/pump/token-stats and the browser check of the Launch detail stats both need a dev server, which this session was not permitted to start. Order file left in place.

2026-10-25: order 544 (integrate 45, chains turnover), re-checked, still partial

Code shipped in bb7445ef2 and the changelog entry is in data/changelog.json. tests/defi-chains.test.js 5/5 pass. Production /api/defi/chains already returns dex_volume_24h and turnover (Ethereum 0.0187). Remaining: the three-width browser check of the Turnover column (375, 768, 1280) needs a dev server, which this session was not permitted to start. Order file left in place.