Workshop Signals

A live GitHub Copilot CLI canvas extension that shows the pulse of every desk in your Workshop. It reads the agent signals your desks emit and renders them as a compact, always-current dashboard in a side panel — so you can direct the work instead of polling each desk by hand.

It replaces the old Blazor WorkshopRoom dashboard (src/WorkshopRoom/) with a native canvas that runs inside GHCP, with no separate web app to launch.

What it is

Each desk in the Workshop leaves signals behind — small stones on the trail — by writing JSON files into its .signals/ folder. Cairn scans those folders, picks the most recent signal per desk, and renders:

• Score bars for the desk's self-assessment (intent, confidence, accuracy, completeness).
• Patterns the desk reported: what worked ✓, what was hard △, and skill gaps ✗.
• Escalations — desks that raised their hand, with what they're blocked on and their recommendation, pinned to the top and pulsing red.

How to open it

The dashboard is registered as the 🪨 Cairn canvas (signals-dashboard). Ask Copilot to open it and pass your workshop root as workshopDir:

Open the 🪨 Cairn canvas with workshopDir set to the folder that contains my desks/ directory.

workshopDir must be the absolute path to the workshop root — the folder that holds desks/ (and optionally classroom/). If omitted, it falls back to the current working directory.

Features

• Signal scanning — walks desks/*/.signals/ and classroom/*/.signals/, reading the newest *.json per desk (mirrors SignalReader.cs).
• Score bars — color-coded intent / confidence / accuracy / completeness, scored out of 5.
• Escalation alerts — escalation signals sort to the very top, render with a pulsing red border, and surface the blocker + recommendation.
• Active desks first — sorted escalations → recent signals → desks with no signal yet, then by recency.
• Stash / restore — pause a workstream by stashing its desk. Stashed desks drop off the active view and auto-expire after a 48-hour TTL; restore any time before then. Stash state lives in .desk-stash.json at the workshop root.
• Auto-refresh — the panel refreshes every 5 seconds using a background fetch (no full page reload), so scores and escalations stay current smoothly.
• Summary bar — desk count, how many are reporting vs. awaiting, an escalation badge, and average scores across the room.
• Cost-aware desk launch — open starts a repo-profile desk that keeps the verified Workshop root available while suppressing ambient plugin MCPs. connected preserves every configured MCP for work that needs external systems. Agency remains the preferred wrapper when installed.
• Local Delegation — orthogonal off/on control. When available and on, the frontier desk may use the installed local-agent-delegation skill for bounded, independently gated read/evidence work. Fail-closed: no skill or no qualified route receipt means the toggle cannot take effect, and no local-savings credit is awarded.

Agent actions

The canvas also exposes actions Copilot can invoke directly:

• refresh — force a rescan and return current signal data as JSON.
• stash — stash a desk by deskName.
• restore — restore a stashed desk by deskName.
• open_desk — open a desk with optional profile: "repo" | "connected" and optional localDelegation: "off" | "on", model, reasoningEffort, and context.

Desk launch profiles

repo is the default. At launch, Cairn asks Copilot for the enabled plugin-scoped MCP inventory and disables those ambient servers for the topic desk. User-, workspace-, organization-, and built-in resources are left alone. If discovery fails, Copilot plugin MCP suppression fails open; Agency repo mode still omits Agency's own default MCPs.

When Agency is installed, Cairn keeps the existing agency copilot launch and adds Agency's --no-default-mcps in repo mode. Outside Agency, the same profile is applied directly to Copilot CLI.

Both profiles pass --add-dir <workshop-root> so a desk can intentionally read another desk's journal or artifact without receiving access outside the room.

Set WORKSHOP_DESK_PROFILE=connected to retain the historical default for the main open button. The separate connected button is always available when repo mode is the default.

Independent sessions and launch preferences

Every open starts a new session with its own UUID. Cairn removes inherited Agency/Copilot session identity and log-directory variables from the child environment and clears them again inside the terminal command. This also keeps an already-running terminal from reconnecting a new desk to its parent's session. Existing sessions and their histories are not changed or resumed. Terminal cleanup matches variable names case-insensitively, including when a running terminal supplies mixed-case state. Linux invokes /bin/sh by that name so BusyBox shell symlinks keep their applet identity.

When opening the canvas, optionally pass these fields alongside workshopDir. Both dashboard launch buttons use that panel's selection. The open_desk action accepts the same fields as overrides for one launch:

Field	Accepted values
model	A Copilot CLI model ID, or auto
reasoningEffort	none, minimal, low, medium, high, xhigh, max
context	default, long_context

Selected values become explicit --model, --reasoning-effort, and --context arguments for either Agency or plain Copilot, independently of the tool profile and Local Delegation. Omitted values remain under Copilot CLI's normal defaults; Cairn does not change shared settings or impose a model on other users. Invalid selections fail explicitly instead of silently using defaults. The open_desk action returns validation failures as { error: "..." }, like other invalid action inputs. Reopening a panel without preference fields clears its previous selection; per-launch overrides do not change the panel selection.

The launch response reports the requested launchPreferences. Model availability and context capacity depend on the CLI and model: a requested long_context argument alone is not proof of the effective context window.

Local Delegation

Local Delegation is not a third desk profile and does not replace the frontier model. It is a separate permission bit:

repo / connected        = which MCPs and tools the frontier desk can see
Local Delegation off/on = whether the frontier desk may invoke a bounded local worker

Availability is fail-closed. Cairn enables the lane only when:

1. the local-agent-delegation skill is installed, and
2. a qualified route is declared via WORKSHOP_LOCAL_DELEGATION_ROUTE_ID or a receipt at ~/.copilot/local-agent-runs/qualified-route.json (status: "qualified", safe route_id).

Operator preference is stored user-locally under ~/.copilot/workshop-local-delegation/ (keyed by the canonical workshop path), never in the cloned workshop — a repo cannot ship preference: on. When the preference is on but availability fails, opens still launch as frontier desks and surface the reason — they never silently fall back with savings credit.

When effective, Cairn sets WORKSHOP_LOCAL_DELEGATION=enabled on the launched process, shows an open toast/badge (Local Delegation effective · route …), and may append one short ASCII line to -i (Local Delegation env is enabled.) when that combined prompt stays quote-free and under the length guard. Full policy still lives in the env flag plus the installed local-agent-delegation skill — never a long multi-sentence -i appendix (Windows Terminal reparse). The runtime, launcher, and gates remain owned by Sealed Delegation.

Signal shape

Cairn reads the agent-signals protocol used across the Workshop:

{
  "signal_type": "execution",
  "agent_name": "desk-name",
  "self_assessment": { "intent": 5, "confidence": 4, "accuracy": 4, "completeness": 3 },
  "patterns": { "what_worked": "...", "what_was_hard": "...", "skill_gap": "..." },
  "escalation": { "reason": "...", "blocked_on": "...", "recommendation": "..." }
}

escalation is only present on signal_type: "escalation" signals.

Replaces the Blazor WorkshopRoom

This canvas supersedes the standalone Blazor dashboard in src/WorkshopRoom/. The data is the truth and the UI is just a view — Cairn renders the same signal data natively inside GHCP, so there's no separate server to run.