MARSHAL v0.0.4

MARSHAL v0.0.4 — One shared channel

Agents in a project now work from one ordered channel instead of from a
briefing that went stale the moment it was written. Each agent drops what it
does into it as it works, and each reads its own view of it — filtered by what
you decided before the work started.

The channel

Every agent contributes as it works. Claude Code and Codex write
append-only history that MARSHAL reads as it grows. OpenCode and Antigravity
keep SQLite, opened read-only and never written to, which SQLite serves to
readers while a writer holds the file. No agent waits until it exits.

An agent joining late joins the conversation. It used to start from empty,
on the reasoning that its briefing had already summarised what came before. A
summary is not the work. A reader now resumes from its own cursor, so an agent
opening while another is five steps into a task sees those five steps.

An agent that was closed catches up. Entries accumulate whether or not the
recipient is running. Each carries the time it happened, and a boundary marks
where a session begins, so nothing older reads as though it just arrived. A
full view drops its oldest entries rather than sealing itself, because the
recent work is the part a returning agent needs.

.marshal/stream/events.jsonl   the channel — ordered, append-only
.marshal/stream/cursors.json   how far each agent has looked
.marshal/inbox/<agent>.md      that agent's view of it
.marshal/live-peers            who joins, and who each agent sees

Choosing who sees whom

Two decisions, both made before the work starts, with /memory peers or in
.marshal/live-peers:

participants: claude, codex, opencode, agy
agy: all                  # every other agent
claude: all
codex: opencode, agy      # not claude
opencode: none            # contributes, reads nothing

Joining and seeing are separate. An agent can contribute while reading almost
nothing, and that is an arrangement rather than a gap: models differ in what
they can use. A capable one does better seeing everything the others did; a
smaller one does worse, because context it cannot follow is context it can be
confused by. The two directions between any pair may disagree — a reviewer can
read the implementer without the implementer reading the reviewer.

An agent is never shown its own work, and that is not a setting. It already
knows what it did.

Capture reports itself

Memory capture was always live; nothing said so, because the only report came
after the agent exited. The native CLI owns the terminal for a whole session,
so MARSHAL cannot draw a counter — it writes one instead, to
.marshal/<agent>/live-status.json: records imported, entries shown, last sync
and any capture error.

Workspace

• Completion works on arguments. The menu used to close after a command and
  a space, exactly where you had most reason to expect it, and Tab then took the
  first candidate silently. It now opens for a command's arguments too, and Tab
  steps through the candidates the way a shell does, leaving the menu up.
• Enter settles, a second Enter runs. Settling on /memory is where you
  reach for Tab again to complete peers; running on the same keystroke ran
  something half-written.
• Antigravity is Agy cli in the menu, after the command you actually type.
• Gemini leaves the provider menu. Its adapter, its doctor probe and
  Import Gemini JSONL are unchanged — this narrows what the TUI offers, not
  what MARSHAL can run.

Known limits

Delivery is a pull. A running CLI owns its own input; MARSHAL cannot
interrupt it. The view is a file that is current whenever the agent looks, and
the briefing says so rather than implying the agent is kept in sync.

Verification

Tested against the installed CLIs on 2026-09-22 — Claude Code 2.1.278, Codex
0.155.1, OpenCode 1.18.16, agy 1.2.7 — with real sessions rather than
doubles:

• Each agent was run and asked to emit a marker. The watcher read all four real
  stores and put 22 entries into the channel, with all four markers present.
• Each agent was then asked to read its own view. All four correctly named the
  other three and quoted their markers. None reported its own.
• With codex: none set, a real Codex session was run inside the real TUI.
  The channel went from 22 entries to 64 with that session's marker in it, while
  codex's own view stayed at its header — none of the other three agents'
  markers reached it. The data was there and it was withheld.

The arrangement space is tested exhaustively rather than by example: 65,536
configurations written, read back and checked to still mean the same thing for
all sixteen author/reader pairs, and every filter a reader can have rendered to
a real file and read back.

Also in this release

Internal, with no change in behaviour: the CI gate no longer runs the whole
suite twice and is split by cost; the store suite migrates once into a template
instead of 311 times, taking it from 400 seconds to 184 under the race
detector; and the PTY waits scale with the race detector, which had been
failing on a loaded runner and reading as a broken TUI rather than a slow one.