stack/docs/plans/2026-09-12_control-board-mvp.md

Control board MVP — plan

Tracking: Gitea issue #1503 (#1503)

Why

Jason is building blind. He has many agent sessions running across projects and no single place to see what they are doing or who is waiting on him. This plan replaces the registry line as the next thing to build. Registry increment 3 stays parked; it does not resume by inertia. Decision record: MOSAIC-STACK-D-001 (Jason, 2026-09-12).

What "done" looks like for the MVP

One web page lists my running agent sessions across projects, shows each one's current status, and flags which ones are waiting on me.

Acceptance:

• Jason opens one page and sees every live pi agent, grouped by project.
• Each agent shows a state in plain words (working, waiting, error, offline, idle, unknown).
• A "waiting on you" section sits at the top of the page.
• The page refreshes itself. Jason does not have to reload it by hand.
• Every state on the page comes from a status file written by the scanner, not from a guess made in the page itself.

Three steps

Step 1 (today): plan, ticket, scanner

• Write this plan and open the Gitea ticket.
• Build the status scanner script with tests. Running it writes one status file per agent under <dataRoot>/board/.
• This step does not touch any page or launcher. It only produces files.

Step 2: the web page

• Build a page that reads the status files the scanner writes and shows them grouped by project, with the waiting-on-you section on top.
• The page refreshes itself on a short timer. The small local server behind it may re-run the scanner on each refresh so the files stay fresh. No separate daemon is needed for this step.

Step 3: daily use and fixes

• Jason uses the page every day for real work.
• He says what is wrong or missing.
• Fix that one thing, then repeat. No new scope beyond what he actually hits.

Step 3 log

2026-09-12 — Gate A fix: task, active project and workspace per row. Asked by the professor session on Jason's behalf. Each row now shows Task (the session's first user message), Active project (basename of the nearest git checkout above the workspace) and, in the detail row, Workspace (the live #{pane_current_path} of the pane running pi, else the session log's cwd). Anything that cannot be derived shows the word "unknown". Missing signals, for launchers: (1) pi session logs carry no task envelope entry, so Task is the free-text first user message, which for fleet seats is a fleet-comms envelope and for hand-started seats is whatever was typed ("resume", "who are you?"); a launcher that wants a clean task label should emit one as the first line of the first message or write a per-seat status file the scanner can read. (2) Neither roster.yaml nor a seat's profile.json names a project, and every fleet seat runs in ~/.mosaic, which is itself a git checkout, so Active project reads .mosaic for all of them; a project field in the roster would fix that. (3) Workspace is only live while a pi pane exists; for offline seats it is history from the log. Fixture test per field, a tmux-inspect parse test, and a page test (79/79). Sonnet review: APPROVED, no findings.

2026-09-12 — liveness follows the tmux pane, not just the session. Jason found agents marked "waiting" that were actually dead: their tmux session still existed but no longer ran pi (it had exited to a shell or something else). The scanner now checks which program each tmux pane is running (tmux list-panes -s -t '=<session>' -F '#{pane_current_command}') and only counts an agent alive if a pane is running pi. A tmux session with no pi pane is now "offline" instead of "waiting".

2026-09-12 — a "Seen" mark for rows that don't need a reply. Jason noticed most "waiting" rows were completion reports, not real asks, and they kept cluttering "Waiting on you". The page now has a "Seen" button on waiting/error rows; clicking it stores the row's lastActivity in <dataRoot>/board/seen.json (via POST /api/seen) and drops the row out of "Waiting on you" while keeping its state visible. The mark clears itself the moment the agent writes anything new, and "Unsee" puts a row back by hand. seen.json is only ever written by that click, never by a scan.

2026-09-12 — Gate B passed: seat registration on the board. Jason restarted the board on 17153fe1 and relaunched darkwing through agents/darkwing/launch.sh. The row shows the registration (pid alive, not stale), workspace and active project from the record, and the last message updating as the session runs. Task still comes from the first user message because the launch gave no --task; the next real use is scripts/mosaic seat task darkwing "<text>" so the row shows the assigned task instead. Fleet seats stay on the old launchers; only agents/ seats register. Next piece in Jason's order: reply-from-board.

2026-09-12 — Gate C passed: reply-from-board (#1505). Board restarted on 867619dc. Jason sent "pizza?" from the board to filbert (Pi, openai-codex/gpt-6-astra). The message arrived with the trailer, the seat answered in its own session ("Yes. Pepperoni?") and made no send attempt; the row showed the answer on the next scan. Before the trailer the same model followed ms-communications to the letter and tried agent-send.sh -s control-board, which the tool refused (no such pane). That was correct behaviour on both sides, not a model defect; a weaker model that skipped the send was ignoring the skill, not doing better. Sequence next: the ledger script, then the WebUI on the Console design absorbs the board.

Piece 2: reply-from-board

Approved by Jason 2026-09-12 after Gate B ("seat task update works. I want to keep things building"). Same shape as #1503/#1504: one piece, one yes/no gate, small commits with tests on refactor. Darkwing opens the issue and updates CURRENT.md when he picks it up.

What it is

A reply box on the board. Jason types a message on a seat's row and the board delivers it to that seat's tmux pane through the repository's own channel. No new transport, no message store, no WebSocket.

• Page: a one-line text box and a Send button in each row's detail (registered rows only, see below). After Send the row shows one of two receipts: delivered with the time, or failed with the tool's stderr. Nothing else changes on the page; the next scan shows the effect (the seat's newest user message is the text just sent, and a waiting row turns working).
• Server: POST /api/reply with { "agent": "<row id>", "text": "..." }. The handler resolves the row's tmux session, runs tools/tmux/agent-send.sh -s <session> -S "<host>:control-board" -m <text> and returns the exit code, stdout and stderr. The exit code is the receipt; a non-zero code is a failure the page shows verbatim. -S labels the sender so the seat's preamble reads [<host>:control-board -> …]. Fleet sockets pass -L from the registration's tmux.socket when set.
• Which rows: only rows whose registration (#1504) names a tmux session, and whose registration is not stale. Everything else shows "reply needs a registered seat" instead of the box. Seat name alone is not an address; the registration's session is.
• Never: raw tmux send-keys, a queue, retries, a history panel, a broadcast, sending to more than one seat per request. The durable record of what was said is the seat's own session log, which the scanner already reads.

Gate C (yes/no)

Jason answers a seat that is waiting from the board, without opening a terminal, and within one scan the row is working and its Task or last message reflects the reply. One real exchange is the pass.

Tests

• Server: a fake agent-send.sh on PATH captures its argv; assert -s, -S, -L (when the registration has a socket) and the message text.
• Refusals with exit codes surfaced: unknown row, empty text, row without registration, stale registration, tool exit non-zero (stderr returned).
• Page: the box appears only on registered live rows; a failure receipt shows the stderr text.
• Loopback only stays as is; no auth is added for this piece.

Boundaries

• No change to packages/seat or the registration record.
• No change to tools/tmux/agent-send.sh; if it refuses, the refusal is the finding (ms-communications rule), reported on the page and in the issue.
• Fleet seats are out: they are not registered (Jason, 2026-09-12).

How the scanner decides state

The scanner reads the newest pi session log for each agent, plus whether the agent's tmux session is still alive. It does not guess; it only reports what the log and tmux say.

State	Plain-words meaning
working	The agent is in the middle of a turn: thinking or running a tool.
waiting	The agent finished its turn. It is your move now.
error	The agent's last turn ended in an error, was aborted, or was cut off. Go look at it.
offline	There is no live tmux session for this agent right now.
idle	The agent is live but has not had a conversation yet.
unknown	The scanner could not ask tmux (missing or not answering). It does not assume the agent is alive.

Two kinds of agents are scanned today:

• Repo agents: session logs live under .pi/state/<agent>/sessions in a project checkout, on the default tmux socket.
• Fleet agents: session logs live under ~/.mosaic/fleet/agents/<agent>/.pi/agent/sessions, on the tmux socket named mosaic-fleet.

Where files go

• One file per agent: <dataRoot>/board/sessions/<project>/<agent>.json
• One summary file for the whole board: <dataRoot>/board/index.json
• One registration per seat started through mosaic launch: <dataRoot>/seats/<layout>/<seat>/registration.json (#1504). Written by the launcher, read by the board, rewritten on relaunch. Not under board/ because the board never writes it.
• dataRoot comes from ~/.config/mosaic-dev/config.json. If that config is missing or broken, the scanner refuses to run. It does not guess a fallback location.

Board files are derived. They can be deleted and rebuilt at any time by running the scanner again. They are NOT run records and they are not evidence under the repository's write-once rules.

Commands

Scan once and print the table:

node packages/control-board/src/cli.mjs scan --print

Start the page (step 2), then open http://127.0.0.1:7331/ in a browser:

node packages/control-board/src/cli.mjs serve

Tests:

node --test packages/control-board/tests/

Boundaries

• No changes to any launcher script. Corrected 2026-09-12 by Jason's go on the mosaic launch brief (#1504): the four repository launch scripts (agents/{darkwing,dewey,filbert,rocko}/launch.sh) gained a guard that re-enters them through scripts/mosaic launch for registration. The fleet launchers under ~/.mosaic are still untouched; their one-line change is written in the Step 3 log for orch-01 to apply.
• No new files at the repository root.
• No secrets read, stored, or printed.
• No daemon yet. In step 1 the scanner is run by hand. In step 2 the page may trigger it on refresh; nothing runs it on a schedule.
• No changes to packages/mosaic. Still true: seat registration went into a new package, packages/seat, not the hardened registry.
• This plan does not authorize push or merge beyond whatever the existing refactor-branch plan already allows.

2026-09-12 — a place to find what you marked Seen. After using the button, Jason noted the board had no way to show seen rows again. The page now has a collapsed "Seen (N)" section between "Waiting on you" and "By project" that lists every marked row with an Unsee button. It stays open or closed across refreshes.

2026-09-12 — "Hide seen" per project. Jason suggested a checkbox like "Hide offline" so seen rows stop cluttering the fleet table now that the Seen section exists. Each project header has both boxes, on by default, and the note under the table reads "N offline hidden · N seen hidden".

2026-09-12 — Mid-tool-call rule checked and pinned. Requested by the professor session on Jason's behalf. Finding: the scanner already honoured the rule, but only through pi's stopReason (toolUse and tool results were working, stop was waiting; deriveState in scan.mjs). It now also looks at the content: an assistant message with a toolCall block is working whatever its stop reason; an error stop reason still wins; a thinking block does not stop a text turn from being waiting. Three fixture tests (tool call after a question-looking text, tool result last, finished text-only turn) plus three state-table cases. Live check at 14:27Z: orch-01 mid-tool-call → working, rev-code-01 tool result last → working, velma finished text-only turn → waiting; all 13 waiting seats had text-only (or thinking+text) last messages, none had a tool call.

2026-09-12 — Header count "N of N". Jason pointed out that "fleet (38)" sat above a table showing 13 rows. The header now reads "fleet (13 of 38)" while a checkbox hides something and falls back to "fleet (38)" when nothing is hidden. The note under the table stays: the header says how many, the note says why. One static test.

2026-09-12 — Parked until the board passes its gate. Jason's ruling: nothing else starts until he has used the board in place of cycling tmux. Parked, not cancelled, and not to be resumed by inertia:

• Dewey's dashboard mockup selection (task D05, agents/dewey/work/wui/). It gets chosen against a board Jason has used, not before.
• Registry increment 3 (already parked above; restated so it stays parked).
• New seat directories. The roster is frozen at its current seats until the rails pass condition is written (phase C of the 48-hour plan).

2026-09-12 — Rule: a seat mid-tool-call is working, never waiting. Added as an acceptance rule, not a fix log entry, because it is the second way the top section can lie (the first was the dead-session case above). If the newest session-log entry is an assistant message that contains a tool call, or a tool result with no assistant text after it, the state is working even if the last text the agent printed looks like a question. waiting requires the last entry to be an assistant message whose content is text only and whose turn has ended. Verify with one live seat caught mid-tool-call and one that has finished a turn; add a fixture test for each so the rule cannot regress silently.

2026-09-12 — Jason chose Dewey design 3 (Console) as the WebUI starting point. This answers task D05 and lifts the "chosen against a board Jason has used" hold above: the selection is made, on evidence from today's use. It does not start WebUI build work. Sequence stays: seat registration via mosaic launch (in progress), then reply-from-board, then the WebUI on the Console design absorbs the board as its first screen.

2026-09-12 — Seat registration: mosaic launch <seat> (#1504). Gate A passed on Jason's go, relayed by the professor session with a brief. New package packages/seat and wrapper scripts/mosaic:

• scripts/mosaic launch <seat|seat-dir> [--task ..] [-- args] resolves agents/<seat>/launch.sh (repo layout) or a seat directory (fleet layout, <seatDir>/.pi), writes <dataRoot>/seats/<layout>/<seat>/registration.json, then replaces itself with the launch script (process.execve), so the pid, the terminal, the exit code and the tmux foreground command are the script's own. The record: seat, project, task, workspace, tmux, harness, startedAt, pid, plus sessionsDir (the board matches on it, not on the name, because "darkwing" exists in both layouts), seatDir, launchScript, layout, updatedAt, version. Rewritten on every launch.
• scripts/mosaic seat task <seat> <text> changes the task only; refuses when the seat was never launched this way.
• Scanner: registrations are read on every scan; a registered task, project or workspace wins and taskSource, activeProjectSource, workspaceSource read registration; empty or null registered fields leave the derived value; rows without a registration are unchanged. A malformed record is listed in registrationErrors and skipped. Page: a small "registered" tag next to the value and a Registered line in the detail (start, harness, pid, tmux, layout).
• The four repository launch scripts (agents/{darkwing,dewey,filbert, rocko}/launch.sh) gained a five-line guard: unless MOSAIC_LAUNCH_REGISTERED is set or --check is among the arguments, re-enter through scripts/mosaic launch --harness pi|claude-code. So the plain agents/darkwing/launch.sh registers too. rocko registers but is still not on the board: it has no pi session log to scan. Gap, not a bug.
• Fleet launchers under ~/.mosaic are not edited from this repository. Not planned (Jason's ruling, 2026-09-12, relayed by the professor session): fleet seats stay on the old launchers; only agents/ seats use scripts/mosaic. The one-liner below is kept for the record only. One-line change that would do it, in ~/.mosaic/fleet/bin/launch-seat.sh, after SEAT_DIR= is set (the per-seat launch.sh files are shims and need nothing): [[ -n "${MOSAIC_LAUNCH_REGISTERED:-}" || "$CHECK" -eq 1 ]] || exec /mnt/storage/src/mosaic-stack/scripts/mosaic launch --harness pi "$SEAT_DIR" -- "$@" Fleet records carry project and workspace null (roster has no project field; all fleet seats run in ~/.mosaic), so those rows keep their derived values until --project/--workspace are passed or the roster gains a field.
• Sonnet review found one defect before commit: the record path was keyed by seat name alone, so the repo and fleet "darkwing" would have overwritten each other. Fixed by keying on layout and seat (seats/<layout>/<seat>/); seat task refuses a bare name registered in more than one layout unless --layout is given. Second review pass approved. The reviewer also noted process.execve is marked experimental in Node; recorded, accepted (Node 24+ is the floor and the API is the only way to keep the seat's pid).
• Name clash noted: the npm-global mosaic CLI (estate tooling) has no launch or seat command; the repository wrapper is run by path.

Tests: seat package 15 (record write via a fake launch.sh, args and exit passthrough, relaunch rewrite, --task, tmux null, tmuxContext with a fake exec, task update and its refusals, malformed records, samePath through a symlink, recursion guard); control board 79 → 89 (override precedence, empty registration, no registration parity, loadRegistrations errors, matchRegistration by realpath and the same-name-different-layout case, scan on disk, server round trip, page source checks); the two launch-script suites now run each real launch.sh in a fixture that is its own data root and assert the record and the MOSAIC_LAUNCH_REGISTERED value Pi/Claude saw (5/5); registry 69/69; config 24/24. Live check: the real agents/darkwing/launch.sh run through the launcher against a temporary data root, then a scratch board on 7332 reading that root: the darkwing row shows the registered task, "registered" on task and project, and the Registered detail line.

Found on the way: Jason's own launch refused with "local Pi must match package.json (0.85.1)". Commit 557aba0f (#1499) moved the pin without an npm ci in this checkout (node_modules held 0.84.4; the global pi was already 0.85.1, which hid it). Fixed by npm ci; the refusal now names both versions. Jason questioned the pin itself; it is canon (AGENTS.md "Version pin") and stays unless he decides otherwise.

2026-09-12 — Two defects in 69f99323, reported by the professor session and verified here. (1) The darkwing launch test's flock-contention spawn inherited the plain environment, so the launch script re-entered scripts/mosaic with the default config and wrote fixture records for darkwing, dewey and filbert into the real data root (dead pids, workspaces under /tmp), overwriting the record of Jason's live darkwing. Fix: that spawn names the fixture config, and both launch test files now set MOSAIC_CONFIG to a nonexistent path and clear MOSAIC_LAUNCH_REGISTERED for the whole process, so any spawn that forgets fails instead of polluting. The three fixture records were deleted by hand; Jason's next darkwing launch writes the real one. (2) A registration is written before the launch script's own checks, so a refused launch (Jason's Pi pin refusal) left a record with a dead pid that the board showed as registered. The scanner now probes the recorded pid on every pass (pidAlive, signal 0); a gone pid makes the record stale: still on the Registered line, marked stale with registered.alive false, but derived task, project and workspace win. Index gains registrationStale; the CLI summary a stale count. Tests: board 90/90 (was 89), seat 15/15, launch scripts 5/5; the launch tests run against the real data root untouched (checked by timestamp). Also from Jason: fleet seats stay on the old launchers; the orch-01 one-liner above is marked not planned.

Observation from Jason, same day: after relaunching darkwing, the row's "Last message" was still from the earlier session. Cause, not a scanner bug: Pi writes the session log on the first message, and --continue resumes the newest existing log, so until the relaunched seat sends or receives a message there is nothing newer to read. The registration's startedAt and pid are the only trace of the relaunch, which is what the Registered line is for. A later refinement could say "relaunched at X, no messages since" when startedAt is newer than the last activity.

2026-09-12 — Model per row (#1503). Jason: "The running agent (sonnet, opus, gpt-6-astra) should surface in the interface as well." Pi logs name the model twice: a model_change entry (provider, modelId) at launch and on every /model switch, and model/provider on each assistant turn. readSession now keeps the latest of either; the record gains model and provider; the page shows the model in small type under the agent name (provider in the hover) and a Model detail row. A /model switch shows on the next scan even before the seat answers. Blank when the log names no model; rocko has no pi log and stays off the board. Checked on a scratch board against the real data root: all 42 rows carried a model (darkwing and sage gpt-6-astra, filbert glm-5.3, fleet rows a mix of gpt-5.6-* and glm-5.3). Board suite 91/91. Jason's board on 7331 needs a restart to pick this up. Sonnet review found the model hover title escaped an already escaped string (double encoding for names with & or <); fixed before commit, then APPROVED on that basis.

2026-09-12 — Piece 2: reply-from-board (#1505). Built to the brief above. POST /api/reply {agent, text} rescans, finds the row, and runs tools/tmux/agent-send.sh -s <session> -S <host>:control-board -m <text>, plus -L <socket> when the registration names one, with MOSAIC_TMUX_SOCKET stripped from the tool's environment so the registration is the only source of the socket. The answer carries delivered, the exit code, stdout and stderr; a non-zero exit is reported as is. Refusals before the tool runs: bad or empty or long text 400, unknown row 404, no registration or stale registration or no tmux session 409, tool not startable 500. The page puts a one-line box and Send in the detail of rows whose registration is live and has a tmux session; other rows say "reply needs a registered seat". The receipt reads delivered <time> to tmux <session> or failed (exit N): <stderr>; drafts and receipts survive the periodic refresh and the caret is given back to the box; a delivered reply clears the box and rescans at once. Not built, per the brief: raw send-keys, queue, retries, history, broadcast, more than one seat per request; no change to packages/seat, the registration record or agent-send.sh.

Live check on a scratch server (port 7332) with a scratch tmux session cb-reply-test running cat and a temp data root: agent-send.sh accepted dragon-lin:control-board as sender, the text arrived in the pane as [dragon-lin:control-board -> dragon-lin:cb-reply-test] ..., and the page showed delivered. Killing the session also killed the registered pid, so the row went stale and the box gave way to "reply needs a registered seat", as intended. With a live pid pointed at a missing session the receipt read failed (exit 1): ERROR: invalid resolved pane identity and the draft stayed in the box. Tests: a fake agent-send.sh on the server's configured path captures argv and the environment; cases for -s/-S/-m without -L, with -L, exit 4 with stderr, each refusal with the tool never called, a missing tool, and static checks on the page (box only when alive !== false and a tmux session, JSON POST, escaped receipts). Board suite 98/98 (was 91). Gate C is Jason's to run against a real waiting seat; his board on 7331 needs a restart first.

Sonnet review of the piece: one finding, the route comment at the top of serve.mjs had the /api/reply line spliced into the middle of the /api/seen description, so it read as if a reply rescanned and returned the index; the comment was split back into two entries before commit. Noted but not changed: the body limit is bytes and the text limit is characters, so a non-ASCII text near 2000 characters is refused as an oversized body rather than as long text; and a row open in both "Waiting on you" and its project group has two forms with the same id, so the refocus after a refresh may land on the other copy.

2026-09-12 — Gate C, first exchange, and the trailer. Jason answered darkwing from his restarted board with "pizza?". The delivery worked ([dragon-lin:control-board -> dragon-lin:darkwing] pizza? in darkwing's session log) and the row showed the answer as its last message within one scan, so the return path exists: the board reads the transcript. What failed was the seat's side: the communications skill told the seat a reply only exists once agent-send.sh runs, the seat aimed at -s control-board, and the tool refused because control-board is not a tmux session ("invalid resolved pane identity"). That refusal is correct; the board is not made a pane. Fix in the board, relayed as Jason's ruling by the professor session: every message the board sends ends with a fixed line, (control-board: answer in your own session as usual; the board reads your transcript. Do not agent-send to control-board.), exported as REPLY_TRAILER; -S stays <host>:control-board so the preamble stays honest about the origin. The fake agent-send test now checks the message text ends with the trailer; the receipt is unchanged. Committed with the piece.