docs: lead decision 149 and the channel kind brief, after row 129 (Q156 item 4, Mos mos-0726, #1661)

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

docs/plans/2026-09-26_lead-decisions.md

@@ -3358,3 +3358,21 @@ which stay with him. Each item names who decided it and what happened.
     - I6 carries Mos's two checks: when the bus host exits, the core
       container stops too, and the container's Docker logs hold no
       capability token.
+
+149. **The channel kind gets its own row, directly after row 129
+    (2026-10-18, Mos mos-0726, replying to sage-0726).** Q156 item 4
+    (decision 143) planned the `cli` channel module for row 128 and the
+    `webui` module for row 129. My refresh of row 128's brief (0f3eb25ea)
+    dropped the step without recording it, and row 128 landed without it
+    (ade628c8b). `packages/plugins/src/kinds.mjs` has no channel kind.
+    - Row 129 stays UI only on the current human transport, with one
+      HTTP-to-bus seam in `packages/webui/src/chat.mjs` (sage-0724).
+    - Row 164 (#1661, `docs/plans/2026-10-18_channel-kind.md`) adds the
+      channel kind with `cli` and `webui` as its first two modules. It
+      runs after row 129, and no other row goes between them, so the seam
+      is swapped while it's still fresh. Design packet first, read by Mos,
+      because the row works next to the human proof.
+    - This changes the row numbers Jason approved in Q156 item 4, not the
+      order of kinds: channels still come after the providers and before
+      the task tracker. Mos ruled that it doesn't need Jason, and told
+      Jason so in mos-0726.

docs/plans/2026-10-18_channel-kind.md

@@ -0,0 +1,111 @@
+# The channel plug-in kind, 2026-10-18
+
+Lead decision 149 gives Q156 item 4 its own row. Jason approved the
+conversion order in Q156 (lead decision 143): channels come after the
+providers and before the task tracker. The plan named rows 128 and 129 for
+the `cli` and `webui` channel modules. Row 128's refreshed brief
+(0f3eb25ea) dropped that step by mistake, and row 129 stays UI only
+(sage-0724, Mos mos-0726). This row does the conversion right after row
+129, with nothing between them, while row 129's seam is still fresh.
+
+## The channel kind, with cli and webui as its first modules
+
+### Problem
+
+PRD REQ-PLUG-1 lists channels as a plug-in kind, and
+`packages/plugins/src/kinds.mjs` has no `channel` entry. Today the
+channels are hard-coded, as the PRD's Technical considerations list says:
+the human channel is `via: 'cli'` in the broker
+(`packages/bus/src/runtime.mjs:117`, `broker.mjs:526`, `:1867`,
+`landing.mjs:107`), and the WebUI reaches the bus by running the human
+CLI. Row 128 added the CLI's thread verbs (`packages/cli/src/thread.mjs`)
+and row 129 adds the WebUI chat area, whose one HTTP-to-bus seam is
+`packages/webui/src/chat.mjs`. Both call the bus's project and thread ops
+directly. REQ-PLUG-2 says a hard-coded part converts when a row changes
+it. Gitea #1661.
+
+### Owner and reviewer
+
+- Owner: set when the row is briefed.
+- Reviewer: dewey.
+
+### Files owned
+
+Set by the design packet and named in it before any code:
+
+- `packages/plugins/src/kinds.mjs`: one new `channel` entry;
+- `packages/plugins/channel/cli/index.mjs` and
+  `packages/plugins/channel/webui/index.mjs` (new);
+- the call sites in `packages/cli/src/thread.mjs`, `packages/cli/src/cli.mjs`
+  and `packages/webui/src/chat.mjs` that move behind the loader;
+- `packages/plugins/README.md`, the PRD's Technical considerations line
+  for channels, and the tests the packet names.
+
+### Design first
+
+The owner writes a design packet in their `agents/<seat>/work/queue-<row>/`
+directory before any code. Sage reads it and sends it to Mos, because the
+row works next to the human proof. It answers:
+
+1. The kind's exports, interface version and switch. A channel carries
+   messages between a human surface and the bus's project and thread ops.
+   The packet names each export and what the core calls it with.
+2. What switching a channel off means. The organization file's `plugins`
+   key switches modules by name (REQ-PLUG-1). Say what the CLI and the
+   WebUI do with their channel switched off, and how an operator who
+   switched off `cli` gets it back. The packet must not leave a path where
+   a config edit locks the human out of the write path that would undo it.
+3. Where the human proof sits. It stays in core (REQ-PLUG-1, REQ-DEC-3,
+   Q156 item 4). The bus accepts only its own `human-cli.mjs` by literal
+   path (#225), and bus, Console and wrapper share a tree. Name every file
+   the proof reads and show that none of them moves or gains a caller.
+4. The broker's `via` value. Say whether `via: 'cli'` stays as the proof's
+   name for the human transport or becomes a channel name, and what each
+   of the four call sites above reads after the change.
+5. Discord. It isn't connected to the bus, so it stays out. Say what a
+   future Discord module would export so the kind doesn't need a version
+   bump for it.
+
+### What ships
+
+After Sage and Mos accept the design:
+
+- the `channel` kind in `kinds.mjs`, landing with its first two modules,
+  as REQ-PLUG-1 requires of a new kind;
+- `cli` and `webui` as channel modules. The CLI's thread verbs and the
+  WebUI chat area reach the bus only through the loader, never by
+  importing a module;
+- no behavior change. Every existing CLI and WebUI test passes unchanged,
+  except where the packet names a test that imported the moved code;
+- the PRD's Technical considerations line for channels says `cli` and
+  `webui` are modules and Discord isn't connected yet.
+
+Tests:
+
+- the loader loads both modules and refuses each REQ-PLUG-1 case for the
+  channel kind (bad name, missing directory, symlink, missing export,
+  wrong interface version);
+- `mosaic talk --thread` and `mosaic thread read` behave as before row
+  128's tests expect, now through the module;
+- the WebUI chat routes behave as row 129's tests expect, now through the
+  module;
+- the human proof is unchanged: the same `human-cli.mjs` path, the same
+  bytes, and a call that skips the proof still refuses;
+- the answer to design question 2, tested as the packet states it.
+
+Mutants: one per behavior above, with the killing test in the packet.
+
+### Out of scope
+
+- A Discord module, and any change to the Discord connector.
+- Any change to `packages/bus/src/human.mjs` or to how the bus checks the
+  human proof.
+- Plug-ins from outside the repository (Vikunja #286, Jason's call).
+- Any live switch of the `plugins` setting. A module goes live through a
+  re-pin, with Mos's go.
+
+### Gate
+
+Dewey approves on the row's issue, naming the candidate manifest. The
+plugins, cli, bus and webui node suites and every `scripts/test-*.sh` pass
+on Sage's gate rerun. The landing changes nothing live.