WAI for coding agents on the toolchain's names: wai mcp (toolsets core and dev, tools named by their operations), the room service's API with catalogue operationIds, wai setup-agent #505

Merged

dcharlot merged 7 commits from wai/dx-agent-mcp-2 into main 2026-10-27 10:14:04 -04:00

Conversation 0 Commits 7 Files changed 24 +8126 -39

2026-10-27

Transaction Science

wai mcp: the tools table's test module after every item (clippy items_after_test_module) ...

Co-Authored-By: Claude Opus 5.5 <[email protected]>

bff1736e36

• View at this point in history

2026-10-27 04:44:52 -04:00

Transaction Science

CI and docs: the agent door under the catalogue's names ...

wai-conformance runs `--lib ops::` beside the mcp and setup-agent tests, and
its comment says what they now hold (the tool list as operations, toolsets,
the dev tools' target, setup-agent's envelope and exit statuses). The room
service's workflow comment says the contract test checks every operationId.
wai/README.md gives the toolsets and their tools, and setup-agent's CLAUDE.md
and --toolset; the room service's README the operationIds; the CHANGELOG
entry is rewritten for the names.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

170c883f9f

• View at this point in history

2026-10-27 04:38:59 -04:00

Transaction Science

joule-moq-room: each operationId is the operation's name in the catalogue ...

openapi.json names its 40 operations as the toolchain's catalogue does,
noun.verb: room.show (GET /room) and room.grant (POST /grant) as the catalogue
gives them, and the rest by its rule (service.show and service.check, the
health; admission.show; mls.tail, mls.import, mls.update; performance.list
and .show, schedule.create, handover.create, window.create, fidelity.create;
time.show, source.fetch, watchchallenge.create, watch.tail, watch.show,
watch.join, watch.update; control.grant, remote.show, remote.verify;
conversation.show, .join, .tail, .import, .update, .leave,
conversationchallenge.create, message.create, acknowledgement.create,
member.delete; callchallenge.create, call.create, call.invite, inbox.show,
answer.create, call.show). The paths stay as the service serves them. The
description lists the service behind wai dev's front (/room-api) as a server.

tests/openapi_contract.rs gives each operation its name beside its handlers and
checks that the description names it so, that each name keeps wai::ops's rule
and names one operation, and that room.show and room.grant are the routes the
catalogue gives them. tests/mcp_dev_room.rs drives `wai mcp --toolset core,dev`
under the new tool names and checks that the tool room_grant and POST /grant
are the one operation room.grant, and that no other tool shares an API
operation's name.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

264181c0cf

• View at this point in history

2026-10-27 04:38:58 -04:00

Transaction Science

wai mcp and wai setup-agent on the toolchain's names ...

The two verbs join the one table, dx::VERBS, beside hello, create, dev and
verify; wai.rs keeps no second command table. A verb ends either as a check
(0 held, 1 did not, 2 usage) or with an exit status of the toolchain's table,
which setup-agent uses.

wai mcp (agent.mcp): each tool is an operation of the catalogue, named by it
with `_` for the dot, and nothing else names it. The tools come in toolsets,
chosen with --toolset (core when none is named):
- core: envelope_inspect, registry_list, registry_show, receipt_verify,
  package_build, conformance_list, conformance_run, docs_search;
- dev: dev_start, dev_stop, room_grant, room_join, room_leave, room_session,
  on the target local alone (a target named must be it).
receipt_verify takes the four kinds the verify tools took one each (delivery,
claim, room_session, audit_bundle), by `kind` or read from the arguments, and
answers with the kind beside the verifier's verdict. tools/list gives each
tool's operation and toolset in its _meta. The development room's program is
found as wai dev finds it: beside wai, then on PATH. envelope_inspect counts
what this build decodes as wai inspect does, the floor included.

wai setup-agent (agent.setup): AGENTS.md always; a CLAUDE.md only where the
project has one, which gains an import of AGENTS.md. --toolset registers the
server with those toolsets. Off a terminal it answers in the toolchain's JSON
envelope (envelope/1: op, ok, error {code, message, hint, retryable}, result,
and on exit 4 the proposed changes and the confirming command), with the
table's exit statuses: 2 usage, 3 permission, 4 confirmation required, 5 not
found, 7 another I/O error.

wai::ops holds the rule operation names keep: noun.verb, lowercase letters,
the verb from a closed list, and the few names the catalogue gives as they are
(room.session among them). The quickstart's wai holds `mcp`, so the server
setup-agent registers is in the archive's binary.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

6d88524a4e

• View at this point in history

2026-10-27 04:38:46 -04:00

Transaction Science

CI and docs: the commands for coding agents ...

wai-conformance runs tests/cli_mcp.rs and tests/cli_setup_agent.rs, with the binary's
unit tests, under the feature `mcp` on each leg. The room service's workflow runs every
target, and now says what tests/openapi_contract.rs and tests/mcp_dev_room.rs hold.
wai/README.md says how a project is set up for a coding agent and what `wai mcp`
offers; CHANGELOG records both commands and the API description.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

d84afcdf1e

• View at this point in history

2026-10-27 04:19:19 -04:00

Transaction Science

joule-moq-room: a development room, the API described in OpenAPI 3.1, held to the code ...

`joule-moq-room dev` (src/dev.rs) runs the relay and the room service in one process
on loopback, with keys made at start and kept only in memory, media held to no
receipts and rate limits open, and says on standard error that it is for development
only. It answers JSON control lines: join (a test principal enters with its grant and
a proof, publishes audio and video, and the answer waits for the clock's join and
publishes), leave, session (the room as live-rooms §9 gives it to a verifier) and
stop. Its log goes to standard error. tokio's io-std feature, for its standard input
and output.

openapi.json describes every route of the service in each of its modes: requests,
answers and statuses. tests/openapi_contract.rs holds it to the code: the routes the
routers match are exactly its operations; each status it lists for an operation is one
the operation's handlers return, and each status they return is listed; and the
service, run beside a relay in each mode (a plain room, end-to-end encryption,
admission by age, co-performance, a watch room, a remote desktop room, a conversation,
calls, and without its relay), answers every operation with a status it lists.

tests/mcp_dev_room.rs builds `wai` with the feature `mcp` in a target directory of its
own and drives `wai mcp` with an MCP client: a development room started, two test
principals granted, joined and publishing, the session verified, a changed record
refused, a leave verified, and a grant from a service off the machine refused.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

60e1007c47

• View at this point in history

2026-10-27 04:19:07 -04:00

Transaction Science

wai mcp and wai setup-agent: WAI's abilities as tools for a coding agent ...

`wai mcp` (the feature `mcp`) is a Model Context Protocol server (2025-06-18) on
standard input and output. Its tools, in one table (src/bin/wai_cli/mcp/tools.rs):
inspect an envelope or a package; list and describe registered capabilities; verify a
delivery receipt, a claim with the context the tool takes, a live-room session and a
stream audit bundle, each answering with the verifier's verdict and reason code;
package a media file under `wai wrap`'s rules, inside the project directory only; list
and run a conformance corpus's independent verifier; search the specification; and,
marked development only, drive a development room on loopback (`joule-moq-room dev`):
start it, mint grants for test principals (any service whose address is not a loopback
address is refused before a connection is made), join and publish as them, read the
session, stop it. Each tool carries safety hints; `--read-only` offers only those that
read or run a verifier.

`wai setup-agent` writes AGENTS.md and CLAUDE.md (which imports it) and registers the
server in the project's MCP configuration files by their documented formats. A file
that exists is changed only after its diff is shown and confirmed (asked on a terminal,
or --confirm; otherwise exit 4 with every diff and the confirming command), --dry-run
writes nothing, nothing outside the project directory is written, and its output is
JSON off a terminal.

Both are reached through one table of verbs in wai.rs, so a name is one line. `wrap`
returns its printed lines and its facts, which the packaging tool reads.

tests/cli_mcp.rs drives the binary with an MCP client: every tool, and the verifiers
agreeing vector by vector with jwp-conformance, claim-conformance,
live-conformance/rooms and live-conformance/audit. tests/cli_setup_agent.rs runs the
command in scratch projects, fresh and with files that exist.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

84a5cb33fa

• View at this point in history

2026-10-27 04:19:07 -04:00