Files
stack/agents/dewey/work/queue-76/build.md
T
jason.woltjeandClaude Opus 5.5 86821889ed dewey: row 76 round 2 packet, rebase on 729da72b and decisions 99-100 (#1570)
Rebased on row 59 (a8512007). Darkwing's N1-N5 and the Q1 proxy line
(lead decision 99), Filbert's N1-N4 (lead decision 100): a 408 is unknown
on the send box and the board Reply, the receipt shows the code once, focus
and caret kept, a role no longer held kept, Inbox current on a message
trail, README notes. 35 of 35 mutants killed; gate green on 729da72b.
Manifest sha256 04ecf1b57c594a35d34d53277b2df915a571f4fa9f7cc29146dbedfa612f77d4.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-11 21:38:09 -05:00

520 lines
30 KiB
Markdown

# Row 76 (#1570): Console sends a human message to a role, round 2
Owner: Dewey. Reviewers: Darkwing and Filbert. Brief:
`docs/plans/2026-10-11_dogfood-cutover.md` § "Console sends a message to a
role", blob `f5d56554` (rev 552). Jason's Q95 answer is lead decision 95;
the reply view question is lead decision 97, option (a). Base `729da72b`
(row 59 landed in `a8512007`). None of the candidate's paths changed
between `729da72b` and `refactor` HEAD (`729da72b`) when the packet was
built.
Round 2 is the rebase on row 59, Darkwing's round 1 notes N1 to N5 with
a Q1 line in the README (lead decision 99), and Filbert's round 1 notes
(lead decision 100). They are listed separately below: "Round 2: rebase
on row 59" for the merged hunks, then "Round 2: Darkwing's notes (lead
decision 99)" and "Round 2: Filbert's notes (lead decision 100)" for the
content changes.
The rest of this file is round 1's, updated where those changes touch it,
with the mutants, gate and demonstration rerun on the round 2 tree.
`build.patch` is `git diff --binary 729da72b -- packages/webui`, the two
new test files included. Apply it with `git apply --index` to a detached
worktree of the base; `candidate-manifest.sha256` then checks 9/9 OK.
## What ships
`serve --public-host NAME --auth-header HEADER --allow-user ID --send`.
`--send` is off by default; without it the Console behaves as after row 74
(`preview.test.mjs` and `serve.test.mjs` unchanged and green).
**Start (`cli.mjs`, `serve.mjs`).** `--send` is the one flag without a
value. Without all three preview flags the CLI exits 2 with `refused:
--send needs the preview mode (...)`, and `startServer` throws the same
for a direct caller, or for a `send` that isn't a function. With `--send`
and no system config the start refuses (`--send needs the bus configured
(...)`, the reason redacted as before) instead of serving a box that can't
send; a read-only Console still serves without the config. The `Preview:`
line says reads and human messages to roles, one line per send below.
`busReader` is split into `busCaller` (the human transport call, the
business resolved per request as before) and `views` over it, so the
reads and the send use the same transport.
**The route.** With `--send`, `POST /api/bus/message` is the one write
preview mode lets past its read-only check; any other method on that path
is 405, and every other write is still 403 `read-only`. Order, each on
every request:
1. The shared checks from row 74 run first: repeated Host 400, HEADER with
a Host other than NAME 403 `proxy Host refused`, a NAME request without
HEADER = ID once 401, a sent Origin other than `https://NAME` 403. These
write no log line.
2. `busSend` checks HEADER = ID once again itself (401), so the route
never runs on a loopback Host: a loopback request without HEADER is
401, with it 403 `proxy Host refused`. There is no loopback send.
3. Origin sent once and exactly `https://NAME` (a missing Origin passes
the shared check and stops here), else 403 `cross-origin request
refused`.
4. `Sec-Fetch-Site` sent once and exactly `same-origin`, else 403
`cross-site request refused`.
5. `Content-Type: application/json`, at most 16384 bytes, a JSON object
with `to` and `body`, optionally `in_reply_to`, no other key; else 400
`invalid-request`.
6. `message.send` with exactly those keys. The broker judges `to` and the
body. Its limit, 32768 characters, is past the request limit, so 16 KiB
binds first.
Every answer after step 2 carries `sent`: `false` for a refusal, `"unknown"`
when the message may exist. Success is 200 `{id, at}`.
**Outcomes (`bus.mjs` `sendOutcome`).** The broker writes in one
transaction, so any code it returns means nothing was written:
`{error: code, sent: false}` at the read routes' statuses (`busStatus`),
with `unknown-role`, `message-not-found` and `request-too-large` added as
400. `outcome-unknown`, `invalid-response` and `response-too-large` from
the transport, any error that isn't a `BusReadError`, and an answer
without a valid message id leave it open: 503 `outcome-unknown` with
`sent: "unknown"`. Nothing retries.
**Log.** One stdout line per request that gets past step 2: the time,
`send`, `user=` (JSON-quoted), `to=` (JSON-quoted, cut to 64 characters, or
`none`), then `id=<message id>` or `error=<code>`. Never the body. A test
shows a `to` with a newline can't forge a second line.
**`/api/config`** gains `send`: true only with `--send` on a NAME request,
so the operator's loopback page never shows a box that would get 401.
**The page (`public/bus.js`).** With `send` true the Inbox has "Message a
role": a picker of held roles from the agents read, a text box and Send.
- The draft (role and text) lives in page memory, so the ten-second
refresh redraws the form without losing it; focus and the caret are
kept. A reload loses it. A role no longer held stays selected, marked
"(not held now)".
- Send is disabled while a send is pending; a second press sends nothing.
An empty role or text is refused in the page without a request.
- One POST per Send, `redirect: 'manual'`. Sent: an id came back; the
receipt shows it and links `#/trail/message/<id>` (a new trail kind: the
message and its deliveries, back link to the Inbox, not-found for an id
the trail doesn't hold). The text is cleared unless it was edited during
the send.
- Not sent: a 4xx other than 408, or the server's `sent: false` (a 503
`no-bus-host` too), or a redirect. Only the sign-in proxy redirects, when its session
has expired, so the request never reached the Console; the receipt asks
for a reload. The draft is kept.
- Anything else, a lost answer and a 408 included: outcome unknown. The receipt says
the message may have been sent, that nothing was sent again, and that
there is no trail to link without an id; the draft is kept.
- The box says a reply comes back as a message the Console doesn't show
yet, and the next `mosaic talk` prints it after sending its own message
(decision 97).
**Reply in the board (`public/app.js`).** A refused Reply said "Delivery
may be unknown; check the seat before sending again" for a 403. A 4xx other
than 408 now says "Not sent."; no answer, an unreadable one, a 5xx or a
408 keeps the old text. Row 74 left this for this row.
**README.** A "Message a role" section: the command, what `--send` grants
(anyone the proxy lets through as ID, and any local process that reaches
the port and sets HEADER with `Host: NAME`; decisions 62 and 95), the
request and each check, the broker's refusals and their statuses, the
unknown outcome, the log line, the page's three receipts, and the reply
path with `mosaic talk`'s limits (decision 97). The routes list gains the
message trail; the "nothing writes" lines name the exception. Folded in
from row 74 per the row note: Filbert N1 ("its first signed-in request"),
Darkwing's tunnel sentence (anything reaching a tunnel's local end can
send a loopback Host without HEADER, so bind it to loopback), and a
pointer to `scripts/webui-service.sh`, which takes no `--send`.
## Decisions and tradeoffs
1. **No loopback send.** The brief asks for the header on every request,
loopback included. A loopback request can't carry it (row 74 refuses
HEADER with a loopback Host), so the operator's browser on `127.0.0.1`
reads but can't send. `/api/config` hides the box there.
2. **The route checks the header itself** even though the shared check
already did for NAME requests. It costs one line and keeps the route
closed if the shared check changes.
3. **A missing Origin is refused on the send** (the shared check allows a
missing one for reads). The brief asks for the request's own origin,
and browsers send Origin on a POST.
4. **`Sec-Fetch-Site` is compared exactly**, so `Same-Origin` and a
repeated header are refused. Browsers send the lowercase value once.
5. **Which errors mean "not sent".** The brief maps broker codes like the
reads'. I treat every coded broker answer as definite, since the broker
writes in one transaction, and only the transport's three lost-answer
codes, an uncoded error or a missing id as unknown. The page adds 4xx
and the redirect as definite. A proxy 4xx means the request wasn't
accepted for forwarding; a proxy 5xx (502, 504) can come after the
Console got it, so it stays unknown, and so does a 408 (round 2,
Filbert's N1).
6. **The draft is in page memory, not storage.** The brief asks for it to
survive the periodic refresh. Keeping it out of `localStorage` keeps a
message body off disk on a shared browser profile.
7. **`--send` without the bus refuses to start.** The brief doesn't say.
Serving a box that can only answer `not-configured` seemed worse than
a refusal at start.
## Where the brief and the code differ (for Sage)
- "Replies show in the inbox as today": they don't; the inbox shows
decisions. Decision 97 settled it: the README and the box give `mosaic
talk`'s limits, and the reply view is row 81 (#1575).
- "Points at the trail" on an unknown outcome: there's no id to link, so
the receipt points at the role's side and the bus trail from a terminal.
- The log's `error=` value is the code as answered, unquoted; for the 403s
it has spaces (`error=cross-origin request refused`). It is always last
on the line.
## Round 2: rebase on row 59
Round 1's candidate was on `3d298920` (candidate manifest `ca870314`,
request comment 27594). Row 59 landed first in `a8512007`, so I applied
round 1's `build.patch` to `729da72b` with `git apply --3way --index`.
Every file applied cleanly, with no conflict.
Compared by added and removed lines, file by file, round 2's patch is
identical to round 1's in all nine files. Only the context around the
hunks moved. Two manifest digests change, `packages/webui/README.md` and
`packages/webui/src/public/bus.js`, because row 59 changed those files
too. The other seven digests are the same as round 1's.
Row 59's hunks in this row's files, and how each merged:
- `packages/webui/README.md`: row 59 adds the "Agent state (row 59,
#1548)" section, lines 270-280 in the candidate, between "Inbox, tasks,
agents and trails" and "Shell". My nearest hunks are in "Inbox, tasks,
agents and trails" above it (the route list gains
`#/trail/message/<id>`, and "Nothing writes" names the send box) and in
"Data and boundaries" below it (without `--send` no route writes).
Neither touches row 59's section. It merged cleanly.
- `packages/webui/src/public/bus.js`: row 59 adds `stateMark` and
`attention` and changes `agentsView`, lines 367-390 in the candidate
(State and Attention columns). My hunks sit above it (the header
comment, and the send form's `draft` and `composerHtml`, lines 246-287)
and below it (the message trail, the router, focus restore, and the
submit handler). Neither touches the agents view. It merged cleanly.
`stateMark`, `attention`, `plural`, `blockChip`, `first` and `short`
are each defined once.
Row 59's other files (`packages/bus/README.md`, `broker.mjs`,
`agents.test.mjs`, the webui `bus.css`, `bus-browser.test.mjs` and
`bus.test.mjs`) aren't files this row touches. The mutants, the suites and the
demonstration below ran on the round 2 tree.
## Round 2: Darkwing's notes (lead decision 99)
Content changes made after the rebase, against Darkwing's round 1 review
(comment 27612, `agents/darkwing/work/queue-76-review/review-r1.md`). They
change three files: `packages/webui/README.md`, `src/public/bus.js` and
`tests/send-browser.test.mjs`. Each was a small change; none is left for a
follow-up row.
- **N1, the board Reply's "Not sent." (`app.js:378`) held by a browser
test.** `send-browser.test.mjs`'s stub now answers `/api/reply` from a
queue when one is set. The test opens an inspection, types a reply, and
answers a 403 and then a 502. The 403 must show `failed: refused here.
Not sent.` and the 502 `failed: board unreachable. Delivery may be
unknown; check the seat before sending again.`, the reply text kept both
times, and the board's capture list stays empty. `app.js` itself is
unchanged from round 1. The mutants `reply-4xx-unknown` and
`reply-5xx-refused` check that the test catches both directions.
- **N2, the caret across a refresh.** The refresh case now selects
characters 3 to 7 in the text box, presses Refresh, and asserts focus
and the selection `{start: 3, end: 7}`, after the pattern at
`browser.test.mjs:55`. Mutants `page-no-caret` (not restored) and
`page-caret-not-kept` (not saved).
- **N3, README "Message a role".** "It is recorded exactly like one sent
from the CLI" is now "It goes the same way as one sent from the CLI,
with class INFO where `mosaic talk` uses REQUEST". The class stays INFO.
The box's note in `bus.js` said the same thing ("recorded by the bus
exactly like one sent from the CLI"); it now says "the same way as one
sent from the CLI, with class INFO". No test pins that sentence.
- **N4, README reply path.** "talk prints each one and records it
delivered" is now: the broker records each one delivered when
`message.receive` returns it, so the next receive doesn't return it
again; talk prints each one and marks it read (`message.read`,
`cli.mjs:208`).
- **N5, the code shown twice in a refusal receipt.** The broker's
refusals arrive as `not sent: the bus refused it (unknown-role)`, and
the receipt added ` (unknown-role)` again, after its own "Not sent to
coder:". The receipt now drops the server's leading `not sent: ` and
adds the code only when the message doesn't already carry `(code)`. It
reads `Not sent to coder: the bus refused it (unknown-role). The draft is
kept.` A refusal whose message lacks the code (a 4xx from a check before
the bus, or a proxy's) still gets it. Dropping the prefix goes past N5's
letter; without it the receipt said "Not sent to coder: not sent: ...".
The browser test now asserts the two receipts as exact strings (the 400
`unknown-role` and the 503 `no-bus-host`). Mutants `page-code-twice` and
`page-keeps-not-sent-prefix`.
- **Q1, the proxy contract.** A new bullet in "What the proxy must do":
the proxy passes the Console's status and body through and doesn't
rewrite upstream errors (nginx's `proxy_intercept_errors` with an
`error_page`, for example), because the page reads a 4xx as "not sent"
and a rewritten 503 `outcome-unknown` would then report a message that
may exist as not sent.
- **Q2** stays as built (Darkwing: yes, refuse at start).
## Round 2: Filbert's notes (lead decision 100)
Filbert's round 1 review (comment 27637, approve). These change
`packages/webui/README.md`, `src/public/bus.js`, `src/public/app.js`,
`tests/send.test.mjs` and `tests/send-browser.test.mjs`. N1 is the one
behaviour change; N3 is tests only. Each was a small change; none is left
for a follow-up row. Filbert's N2 is Darkwing's N1 and their N4 is
Darkwing's N5, both above.
- **N1, a 408 is unknown, not "not sent".** A proxy can answer 408 after
it has forwarded the request (Envoy's `stream_idle_timeout`, when no
upstream headers arrived in time). The send box's check
(`bus.js`, the `refused` branch) and the board Reply's (`app.js:378`) now
both leave 408 out of the 4xx they read as not sent; a 408 shows the
unknown receipt. A 408 that carries the Console's own `sent: false` would
still read as not sent, but the Console never answers 408. Tests: the
send box's unknown loop gains a 408, and the board Reply case a third
answer, 408, which must say delivery may be unknown with the text kept.
README: the "Outcome unknown" bullet says a 408 counts as unknown, here
and on the Reply, and the Q1 proxy bullet says "a 4xx other than 408".
Mutants `page-408-refused` and `reply-408-refused`.
- **N3, Filbert's surviving mutants, each now held by a test:**
- `p-no-field-focus` and `p-no-caret`: the refresh case leaves focus in
the text box (the button is clicked from script, which doesn't move
focus), then asserts focus and the selection (Darkwing's N2 above).
Mine: `page-no-field-focus`, `page-no-caret`, `page-caret-not-kept`.
- `p-not-held-dropped`: with `pm` selected, the test ends `pm`'s run in
the broker (`endLaunch`, which releases its claim) and refreshes. The
picker must still hold `pm`, selected, as `pm (not held now)`. Mine:
`page-not-held-dropped`.
- `p-inbox-not-current`: on `#/trail/message/<id>` the Inbox is the
current section (`aria-current="page"`). Mine:
`page-inbox-not-current`.
- `s-reply-truthy`: the passthrough test sends `in_reply_to: ""` and
asserts the key reaches `send` with the empty value. Mine:
`send-reply-truthy`.
- `s-plain-error-code`: the plain `Error` in the outcomes table
(`send.test.mjs`, the `ENOENT` case) now carries `code: 'ENOENT'`, as a
socket error does, and must still be 503 `outcome-unknown`. Mine:
`outcome-plain-error-code`.
- `s-log-cut-65`: the forged-line test asserts the logged `to` is the
first 64 characters exactly, and a 65-character role logs without its
last one. Mine: `send-log-cut-65`.
- `s-array-ok` is equivalent, as Filbert says; no test.
## Rebase on row 60 (round 1)
I built this row on `bbd22f1b`. Row 60 landed first (`1771bf98`), so I
applied my patch to `3d298920` with `git apply --3way`. Row 60's hunks in
this row's files, and how each merged:
- `packages/webui/src/serve.mjs`: row 60's `BUS_VERBS` line (adds
`credentials`). It merged cleanly; my changes there are elsewhere in the
file.
- `packages/webui/README.md`: row 60's credentials read and its line in
the exposure list. Merged cleanly, in different paragraphs from my
"Message a role" section and the "nothing writes" lines.
- `packages/webui/src/bus.mjs`: the header comment, the one conflict. Row
60 named its `credentials` read and I named the one write. The merged
comment says both:
```
// Only the four verbs of the Q1 module (lead decision 56), row 57's
// `decisions` and row 60's `credentials` reads are reachable as reads: no
// read writes, not even seen (Q4). The one write is row 76's human
// `message.send` (lead decision 95), built by the CLI only with --send. The
// bus does the proof; a server started inside an agent run gets
// `human-required` from the broker.
```
Row 60's `tests/bus.test.mjs` change isn't in a file this row touches. The
mutants, the suites and the demonstration below ran on round 2's tree (the
rebase on `729da72b` with every round 2 change).
## Files
| File | Change |
|---|---|
| `packages/webui/src/serve.mjs` | `busSend`, the route under `send`, `body()` limit parameter, `send`/`log` options, `/api/config` `send` |
| `packages/webui/src/bus.mjs` | `busCaller` split out of `busReader`; `sendOutcome`; three codes added to the status map |
| `packages/webui/src/cli.mjs` | `--send`, refused without the preview flags or the bus; the `Preview:` line |
| `packages/webui/src/public/bus.js` | the "Message a role" box, the message trail kind, focus and caret kept across a refresh; round 2: the refusal receipt shows the code once, the box's note, a 408 stays unknown |
| `packages/webui/src/public/app.js` | the Reply 4xx text; round 2: a 408 stays unknown |
| `packages/webui/README.md` | the "Message a role" section and the folded-in row 74 notes; round 2: Darkwing's N3 and N4, the Q1 proxy bullet, the 408 line |
| `packages/webui/tests/send.test.mjs` | new, 13 tests; round 2: `in_reply_to: ""`, a coded plain error, the 64-character log cut |
| `packages/webui/tests/send-browser.test.mjs` | new, 1 browser test; round 2: the board Reply 403, 502 and 408, focus and caret, a role no longer held, exact refusal receipts, a 408 send, Inbox current on a message trail |
| `packages/webui/tests/shell.test.mjs` | the pinned write boundary: bus.js has exactly one write, this one |
## Tests
`send.test.mjs` (13): `startServer` refuses `send` without the preview
mode or as a non-function; the CLI refuses `--send` without the flags and
without a configured bus (exit 2, nothing on stdout, the config path not
named); the CLI with `--send` and a configured bus but no bus host running
prints the `Preview:` line, answers a send 503 `no-bus-host` with `sent:
false` through the real human transport, and logs one line without the
body; without `--send` the route is 403 `read-only` and
config says no send; with `--send` every other write (POST, PUT, DELETE,
PATCH on the board, bus and page routes) is still 403 `read-only` and
nothing reaches the board or bus; config says send only on NAME; 401 for no
header, another user, a suffix, two values, loopback without the header
and `localhost` with no Origin, and 403 `proxy Host refused` for loopback
with it, none sent or logged; each request check (no Origin, no or wrong
or repeated `Sec-Fetch-Site`, no or wrong Content-Type, a form, 16385
bytes, extra keys, `business`, missing `to` or `body`, not JSON, null, an
array, a string) refused with `sent: false`, logged without the body, and
16384 bytes accepted; wrong Origins refused by the shared check; GET and
PUT 405; exactly `to`, `body` and `in_reply_to` passed on, an empty
`in_reply_to` included; a crafted `to` can't forge a log line, and `to` is
cut at 64 characters; each bus refusal with its status and `sent:
false`, each open outcome 503 `unknown` (a plain error with a code too), a missing id unknown, and one
call each (no retry); the human transport request is exactly the business,
`message.send` and the keys; against a real broker the message is written
once as a human send (`from_role` human, `from_run` human-cli, class INFO)
and `unknown-role`, `invalid-request` and `message-not-found` map to 400.
`send-browser.test.mjs` drives the page in Chrome with an in-process
broker for the roles and the trail and a stub for the POST (a send passes
only with the proxy's https Origin, which a loopback page can't have): no
box without `send`; the held roles; empty role and text refused without a
request; the draft, focus and caret across a refresh; a role no longer
held kept selected as "(not held now)"; a 400 and a 503
`sent: false` refused with the draft kept, each receipt asserted exactly
with its code once; a redirect refused as `sign-in`; an unknown
outcome, a network failure, a 200 without an id and a 408 each unknown, with no
trail link and no resend; one POST per Send, pinned as
`{method, headers, body, redirect: 'manual'}`; Send disabled while
pending and a second press ignored; text typed during a send kept; a plain
success clears the text and keeps the role; the receipt's trail link opens
the message's trail, with the Inbox the current section; an unknown
message id is not found; the board's Reply answered 403 says "Not sent."
and 502 and 408 say delivery may be unknown, the reply text kept each time
and nothing captured.
## Mutants
`mutants.py` makes 35 single-site edits, one at a time, in a separate
worktree of round 2's tree (never the build tree). Each mutant runs against
the whole webui suite, and the file is restored from the copy read before
the edit. The baseline passed, the worktree was unchanged afterwards, and
**35 of 35 were killed** (2026-10-12T01:54:19Z to 02:29:20Z). `mutants.txt`
lists each mutant with the tests that killed it.
Round 1's 21:
- Route (`serve.mjs`): the route's own header check removed; Origin
unchecked; `Sec-Fetch-Site` unchecked, or `same-site` let through; the
limit raised to 64 KiB; extra keys allowed; `in_reply_to` dropped; the
id check removed; the log's `to` unquoted and uncut; `/api/config`
`send` true on loopback; the route open without `--send`; every write
opened by `--send` (`send-opens-writes`, killed by the "every other
write is still 403" test).
- Outcomes (`bus.mjs`): only `outcome-unknown` left open; `unknown-role`
dropped from the status map.
- Page (`bus.js`): no pending guard; the text cleared though edited during
the send; a 5xx `sent: false` treated as unknown; an unknown outcome
resent; a redirect treated as unknown; redirects followed; the message
trail shown for any id.
Round 2's 14, one or more for each reviewer note that changed behaviour or
asked for a test:
- Board Reply (`app.js`): a 4xx read as unknown (`reply-4xx-unknown`); a
5xx read as not sent (`reply-5xx-refused`); a 408 read as not sent
(`reply-408-refused`). Darkwing's N1, Filbert's N2 and N1.
- Send box (`bus.js`): the caret not restored (`page-no-caret`), restored
to the end (`page-caret-not-kept`), focus not restored
(`page-no-field-focus`); the code shown twice (`page-code-twice`); the
"not sent: " prefix kept (`page-keeps-not-sent-prefix`); a 408 read as
not sent (`page-408-refused`); a role no longer held dropped from the
picker (`page-not-held-dropped`); the Inbox not current on a message
trail (`page-inbox-not-current`). Darkwing's N2 and N5, Filbert's N1 and
N3.
- Route and outcomes: an empty `in_reply_to` dropped (`send-reply-truthy`);
a plain error with a `code` read as definite
(`outcome-plain-error-code`); the log's `to` cut at 65
(`send-log-cut-65`). Filbert's N3.
Round 1's run on the `bbd22f1b` build and its 21-of-21 run on the row 60
rebase are not repeated here; this run covers all 35 on the final tree.
## Demonstration
`demo.mjs` runs the candidate's `startServer` with `--send`'s options
against a real in-process broker (pm, cto and coder; coder claimed),
driven over `node:http` on loopback with the headers the proxy and the
browser send (`Host: NAME`, `x-authentik-uid: jason`, Origin,
`Sec-Fetch-Site`). Three servers: `send`, `lost` (the broker writes, then
the call throws `outcome-unknown`, as when the transport loses the
answer) and `readonly` (no `--send`). `demo.txt` is its output on the gate
worktree of round 2's tree:
```text
== a send, and a reply to the coder
to coder 200 {"id":"9db1d4eb-2abd-4c13-a388-2ed3d9503c07","at":"2026-10-12T02:37:11.862Z"}
to coder, in_reply_to the coder's message 200 {"id":"c4b38d2a-34df-4dea-b31b-897f7de01777","at":"2026-10-12T02:37:11.877Z"}
== refused before the bus
no sign-in header (401, no log line) 401 {"error":"sign-in required"}
loopback Host with the header 403 {"error":"proxy Host refused","message":"the proxy must forward Host unchanged"}
another Origin (shared check, no log line) 403 {"error":"cross-origin request refused"}
no Origin 403 {"error":"cross-origin request refused","message":"a send needs the Console's own Origin","sent":false}
Sec-Fetch-Site: same-site 403 {"error":"cross-site request refused","message":"a send needs Sec-Fetch-Site: same-origin","sent":false}
Content-Type: text/plain 400 {"error":"invalid-request","message":"Content-Type must be application/json","sent":false}
16385 bytes 400 {"error":"invalid-request","message":"body larger than 16384 bytes","sent":false}
an extra key 400 {"error":"invalid-request","message":"a send takes to, body and an optional in_reply_to, nothing else","sent":false}
GET 405 {"error":"method not allowed"}
== refused by the broker
a role the business lacks 400 {"error":"unknown-role","message":"not sent: the bus refused it (unknown-role)","sent":false}
an empty body 400 {"error":"invalid-request","message":"not sent: the bus refused it (invalid-request)","sent":false}
in_reply_to a missing message 400 {"error":"message-not-found","message":"not sent: the bus refused it (message-not-found)","sent":false}
== the answer lost after the broker wrote
unknown, not retried 503 {"error":"outcome-unknown","message":"the message may have been sent; nothing was retried. Check the trail before sending again","sent":"unknown"}
== without --send
POST is read-only 403 {"error":"read-only","message":"the Console preview is read-only"}
== what the broker holds
9db1d4eb from=human/human-cli to=coder class=INFO in_reply_to=- body="Please look at row 76 before the cutover."
c4b38d2a from=human/human-cli to=coder class=INFO in_reply_to=8c236a3d-c415-4252-818c-015cdd81d185 body="refactor"
pm receives 1 message(s) from the lost-answer send: ["Lost answer test"]
== the Console log, one line per send past sign-in
2026-10-12T02:37:11.862Z send user="jason" to="coder" id=9db1d4eb-2abd-4c13-a388-2ed3d9503c07
2026-10-12T02:37:11.877Z send user="jason" to="coder" id=c4b38d2a-34df-4dea-b31b-897f7de01777
2026-10-12T02:37:11.881Z send user="jason" to=none error=cross-origin request refused
2026-10-12T02:37:11.882Z send user="jason" to=none error=cross-site request refused
2026-10-12T02:37:11.883Z send user="jason" to=none error=invalid-request
2026-10-12T02:37:11.885Z send user="jason" to=none error=invalid-request
2026-10-12T02:37:11.886Z send user="jason" to=none error=invalid-request
2026-10-12T02:37:11.892Z send user="jason" to="auditor" error=unknown-role
2026-10-12T02:37:11.899Z send user="jason" to="coder" error=invalid-request
2026-10-12T02:37:11.905Z send user="jason" to="coder" error=message-not-found
2026-10-12T02:37:11.913Z send user="jason" to="pm" error=outcome-unknown
a body in the log: false
```
## Gate (my run on the candidate)
Detached worktree of `729da72b`, `build.patch` applied with `git apply
--index`, the manifest (sha256 `04ecf1b5…`) 9/9 OK before and after.
Suites one after another, `TMPDIR` in my scratch, Docker on,
2026-10-12T02:29:20Z to 02:37:11Z:
| Suite | Result |
|---|---|
| `packages/webui` | 76 pass, 0 fail, 0 skipped, the browser tests included (14 new) |
| `packages/bus` | 94 pass, 0 fail (88 in round 1; row 59 added 6) |
| `packages/cli` | 114 pass, 0 fail |
| `packages/control-board` | 124 pass, 0 fail |
| business, conversation, discord, harness, ledger, mosaic, queue, runs, seat, tasks | all 0 fail (60, 182, 178, 59, 78, 69, 150, 41, 27, 71 pass) |
| build-tokens `--check` | exit 0 |
| test-auth, test-conductor, test-config, test-discord, test-extension-package, test-foundation | exit 0 (15, 17, 24, 66, 18, 44 passed) |
| test-queue | 27 passed, 0 failed; its `verify`/`render --check` case skips outside the canonical checkout |
| test-release | 14 passed, 0 failed (Docker, nothing skipped) |
| test-task | 98 passed, 0 failed (Docker, nothing skipped) |
`git status` afterwards shows only the patch's 9 staged files and the
`node_modules` link. `core.hooksPath` stayed unset.
## Questions for reviewers
Round 1's two questions are answered. Q1 (decision 5, a proxy 4xx shown
as "not sent"): Darkwing said yes, with a proxy-contract line, now added;
Filbert said yes for nginx, Caddy and Traefik except 408, now unknown.
Q2 (decision 7, refusing `--send` at start without the bus config): both
said yes; kept as built. Round 2 has no new question.