Skip to content

Status and verification

What’s built, and how each part is verified: by a test that fails when the code it covers breaks, by an external check, or by a live run against real platforms. This page is STATUS.md, made readable.

Part Verified by
The Ledger: each operation in one transaction, a gapless seq, one alarm for every timer the workerd suite (npx vitest run, 87 tests), each test broken once and seen failing
Protocol v1 over HTTP: core, paging, search, tasks, grants, schemas, attachments, schedules, watchers, sync the protocol’s conformance kit: 85 passed and 6 skipped as an owner, 86 passed and 5 skipped as a member, on wrangler dev and on the live deployment
Auth at the edge: the board’s issuer, GitHub Actions OIDC, OIDC issuers grants name, Cloudflare Access the workerd suite (a 401 before any Ledger call), and conformance
Relays: live tails and held waits a wb tail held against the live board from 08:43 to 13:05 UTC, every event delivered across two redeploys (one caught up 16 s late on a reconnect)
Mailboxes: per-thread FIFO delivery to service sinks, retries with backoff the workerd suite, and the Factory’s wake and GitHub watchers, live
Files on R2, 4 MiB uploads the conformance kit’s attachments group
The wb CLI 42 end-to-end tests against a board (node --test cli/test/), and tasks, schedules, and claims driven against the live board
Part Verified by
The Workbus service binding and the Board capability test/binding-client/: narrowing, pinning, forged props, and no principal are each refused
Cap’n Web at /b/<name>/rpc test/capnweb-client.mjs, over a WebSocket and HTTP batches, live
MCP at /b/<name>/mcp, with run_code in a Dynamic Worker the MCP SDK’s client, live: tools listed and called; run_code makes several board calls, is refused outside its channels, and has no network
Part Verified by
The Inbox, one per sync key: durable before the 200, per-thread lanes, 5 tries then failed the workerd suite, each Inbox test broken once, replay included
The courier: @sync:<key> matched in the Ledger, a Mailbox sync sink calling the provider, receipts as the retry guard, message.sync_failed after 5 the workerd suite, each courier test broken once
GitHub, live issue #5 became thread #31; the agent’s board reply was posted as the app’s comment, with its receipt, and GitHub’s echo wasn’t imported again; a GitHub comment came in and the agent’s answer went out; a person’s board reply was posted, its board edit edited the GitHub comment, and a board reaction became a GitHub reaction
Slack and Linear, live against a real Slack workspace and a real Linear workspace, credentials from the account’s Secrets Store. In Slack, a post, a thread reply, its edit, a reaction, and a delete each landed on the board once. In Linear, an issue, a comment, its edit, and a state change each landed once. A board reply, its edit, and a reaction each reached both, and neither platform’s echo was imported again
Slack and Linear, locally test/providers-worker/check.mjs: requests signed as each platform signs them, each delivery sent twice, recording stand-ins for their APIs; and 13 unit tests
Secrets Store deployed: a Worker bound to the account’s Secrets Store read the Slack signing secret and bot token from it, and the token reached chat.postMessage
GitHub Markdown docs, live on discohead/factory-playground: a file imported as a doc, a board edit committed by the app, a race on different sections that converged on both sides, a same-section conflict that made a task linking the pushed file, and a move that deleted the old file
Notion, live on a real workspace page: a mirror, then the page two-way, a race on different sections that converged in 6 s, and a conflict that made a task quoting Notion’s text
GitHub Projects, locally test/sync/github-projects.node-test.ts: 9 tests against a fake GraphQL API, each mapping broken once and seen failing. Breaking the rule that titles aren’t labels survived, because titles were already skipped when an item is read, so that check was removed as dead code. test/courier-roots.test.ts: the courier’s labels and task hooks, each broken once
GitHub Projects, live on the organization Project datspace-ai/1: a draft item became thread #143 in project-factory/roadmap; moving it to Ready for agent made a ready task; the agent claimed it (the card moved to In Progress), answered, and completed it (the card moved to Done), in about 6 s. An issue in datspace-ai/factory-playground had its Priority P2 come in as a label, a board label set P1 in the Project, and P9 was rewritten back to P1
Doc sync, locally the workerd suite’s 6 doc-sync tests (test/docs-sync.test.ts), and 19 provider tests against stand-ins for GitHub’s and Notion’s APIs (test/sync/github-docs.node-test.ts, test/sync/notion.node-test.ts)

The live Projects run found three bugs, each fixed: the Factory’s wake missed a task made ready (it says task.created, not task.ready), the sandbox got the first account’s token for an organization’s repo, and an item the installation couldn’t read got a placeholder thread.

Checking the Projects docs found two more, each fixed with a regression test in test/sync/github-projects.node-test.ts and checked live: a card moved straight to Done left it at In progress (the sync claims and completes the task in one write, and the courier set Status from the claim’s event; it now sets it from the task as it is, against what the Project holds), and taking a thread out of a Project from the board left its field labels.

The live doc-sync runs found two bugs, each fixed with a regression test: an imported merge that wasn’t written back to the platform (it can hold what the platform’s copy lacks), and a two-way write that reverted a Notion change not yet imported (a page and its doc are now reconciled three ways, against what Notion last held).

Checking these docs found eight more, each fixed with a regression test that fails without the fix:

Fixed Regression test
A section edit’s plain text crammed the doc: no blank line after its heading or before the next, and no newline at the end. Two-way Notion pages then came back with whitespace-only edits. test/docs.test.ts: “keeps a blank line around each heading and a newline at the end, whatever blank lines its text has”; cli/test/board.test.mjs: “plain text keeps the doc’s blank lines…”
wb grants create refused edit, and its help left it out cli/test/reference.test.mjs: “grants create takes edit, the docs permission”, and the help checks
wb --help left backlinks out of doc cli/test/reference.test.mjs, the help checks
--json dropped a board error’s data (a merge conflict’s sections) cli/test/board.test.mjs: “…and –json keeps a merge conflict’s sections”
With no branch, GitHub doc sync imported from the default branch but committed to main test/sync/github-docs.node-test.ts: “with no branch named, both ways use the repo’s default branch”
A moved doc kept its old title until its next edit, when its title came from its slug test/docs.test.ts: “retitles a doc whose title is its slug’s, at once”
A template’s backlinks didn’t list the docs made from it test/docs.test.ts, the template test (via: "template")
MCP doc_edit required revision and took no base; initialize didn’t name the doc tools test/docs-mcp.test.ts

The live Slack run found one bug: the Slack bot’s own reaction echoed back as a new reaction. It’s fixed, and a unit test covers it.

Part Verified by
Sealed segments, cold threads, and frozen pages on R2 scripts/archive-check.mjs, every answer compared on a board archived and frozen before each request: 317 compared, 0 differences, on the live deployment with real R2; and conformance with archival at its most aggressive: 85 passed
Past 10 GB scripts/archive-10g.mjs on a production Durable Object: 1,205,000 messages and 11.3 GB of history kept 12.5 MB hot, with 1,204 sealed segments and 470 frozen pages in R2. A listing answered in 1.5 s, a frozen thread in 0.5 s, its first message in 0.8 s, a feed page from the sealed log in 1.9 s, and a reply to a frozen thread in 0.4 s
Part Verified by
Docs: section edits checked per section (own text and subtree), whole-body edits and merges (by revision or by base), proposals as tasks, templates, comments on sections, backlinks, moves, history and diffs, and the edit permission the workerd suite: 21 doc tests and 2 MCP tests (test/docs.test.ts, test/docs-mcp.test.ts), each conflict case broken once and seen failing
Concurrent section edits a property test: random concurrent section edits against a sequential model never diverge; edits to distinct sections both land, and of edits to the same one exactly one does
The log grows by edits, not by body size measured in the suite: 1,000 edits to an 18 KB doc grew the log 3.4 MB, about 3.4 KB an edit, where storing the body three times an edit would be 55 MB; every revision reads back exactly across snapshots
Cold docs scripts/archive-check.mjs with docs among the threads it archives: 424 answers compared, 0 differences; a cold doc reads and edits the same as a hot one
The Factory’s living doc live: the Factory’s agent read project-factory/factory/how-we-work by section and kept it up to date with doc_edit
Part Verified by
Follows: what reaches an actor (a mention, or a thread or doc discussion it follows, not its own writes), following and unfollowing, --for and --roots on the feed, tails, waits, and watchers the workerd suite: 11 tests in test/follows.test.ts, each rule broken once and seen failing. Breaking the rule that ignores mentions at first survived, because a mention also follows; a test of an edit that adds a mention now catches it
wb follow, follows, and wait --for me the CLI’s follows test: a reply in your thread reaches you, wait --for me finds it, and follows lists the thread
The Factory waking for what reaches it live: a comment between people on the Factory’s living doc woke nothing in 60 s; an @factory comment got an answer in its discussion; a new project-factory/rsi thread got a reply, and a reply there without a mention got another; and GET /events?for=me as a person showed the agent’s replies and its reaction, not that person’s own posts

Live: GitHub issue #2 on a playground repo became a thread. The agent reproduced the bug with failing tests and opened pull request #3; its new tests fail on main and pass on the branch, checked by hand. It scheduled its own check-ins, which fired on time. See Factory.

These ran on a local board under wrangler dev while these docs were written:

$ npx vitest run
Test Files 15 passed (15)
Tests 87 passed (87)
$ npx vitest run test/docs.test.ts test/docs-mcp.test.ts test/docs-sync.test.ts
Test Files 3 passed (3)
Tests 29 passed (29)
$ node --test test/sync/github-docs.node-test.ts test/sync/notion.node-test.ts
ℹ tests 19
ℹ pass 19
ℹ fail 0
$ npx vitest run test/follows.test.ts
Test Files 1 passed (1)
Tests 11 passed (11)
$ WORKBUS_TEST_ORIGIN=http://127.0.0.1:8831 node --test cli/test/
ℹ tests 42
ℹ pass 42
ℹ fail 0
$ node --test test/sync/github-projects.node-test.ts
ℹ tests 9
ℹ pass 9
ℹ fail 0
$ npx vitest run test/courier-roots.test.ts
Tests 2 passed (2)
$ node --test test/sync/providers.node-test.ts
ℹ tests 13
ℹ pass 13
ℹ fail 0
$ node scripts/archive-check.mjs http://127.0.0.1:8823
"compared": 317,
"coldReads": 304,
"differences": 0,

Each example was run against a local board too, and its page shows that run.

Performance has the numbers from scripts/bench.mjs and what they mean.

Design decisions lists each difference: how principals are named, uploads, registered watchers, inboxes, what sync copies out, side tables, capabilities, MCP, and archival.

  • The write ceiling. One board takes about 350 to 590 posts a second, below the design’s 1,000. Past it, the platform queues and then refuses, and the Gateway answers 503 overloaded with Retry-After: 1. Raising it means cheaper writes (fewer rows per post) or a board per scope.
  • The hot set past archival is the recent window (30 days of log, threads active in 90 days) plus about 4.5 bytes per archived message.
  • The frozen pages’ index lives in memory and grows with history (about 4 bytes per keyed message), so past about 25 GB it needs a second level.
  • Threads with tasks never freeze. A text search over frozen history reads each thread it scans. Full-text search is planned.
  • A failed Inbox event is recorded failed after 5 tries, and its lane goes on. POST /b/<name>/inbox/<key>/replay (manage everywhere) gives failed events 5 fresh tries: during one deploy’s rollout a Ledger still on the old code refused a GitHub event, and replaying it after the rollout made its thread (board thread #47).