Make the lifecycle issue the interface: status dashboard, checkbox decisions, Mermaid plans, one run per command #2228

Problem

Running a proposal through research → triage → plan → activate on octodemo/zava-social-cao-ops-20261014152823#3 showed that the issue is hard to read and to act on:

• No feedback after a command. The only signal was a 👀 reaction that never changed. Commands were re-sent 2–4 times each, so a 4-phase lifecycle used 10 commands. Two research runs failed with nothing posted.
• Preconditions aren't checked. /squad triage with no research artifact produced nothing at all, not even a rejection.
• The lifecycle table isn't trustworthy. It's buried mid-thread and edited silently, the Plan link renders as a placeholder (#aw_eSi0hpoM), Activation never updates, and its JSON is empty ("phases":[]).
• Decisions were resolved without the human. The plan said "Decisions Needed feat: worktree awareness, Scribe auto-commit, and decision consolidation #1–Feature: Issue Assignment & Auto-Triage for Squad Members #4 … are resolved as design choices … no additional decision-gate step is required". D2's acceptance criteria point to "the level of fidelity chosen in Decision feat: GitHub Issues intake, PRD mode, and human team members #2", which nobody chose.
• IDs and references break. "Decision feat: worktree awareness, Scribe auto-commit, and decision consolidation #1" auto-links to issue feat: worktree awareness, Scribe auto-commit, and decision consolidation #1. Items are relabelled at every stage (Rn / # / Item N).
• Dependency graphs are ASCII (1 → 2 → 3) and don't show that decisions gate work items.
• Too many commands and too much noise. There are long copy-paste commands in the body, five "next step" options (including /squad plan program), a repeated footer, and open "Structured data" JSON.
• Hidden deadlines. Evidence expiry (expires_at, 24 h) appears only in the JSON.

Constraints (verified against GitHub's documentation and renderer)

• GitHub strips <button>, <form>, <style> and <script>, so a real click-to-run button isn't possible.
• Reactions have no Actions trigger. Comment boxes can't be prefilled from a link, and neither can workflow_dispatch inputs.
• Mermaid (including classDef styling), callout alerts (> [!IMPORTANT] etc.), task-list checkboxes, <details>, tables and images all render in issues.
• Every checkbox tick fires issues: edited. To conserve runs, Squad must not subscribe to issues: edited. Even a run that immediately skips still counts as a run.

Proposal

1. A status dashboard at the top of the issue body

The bot rewrites only the block between <!-- squad:dashboard:start --> and <!-- squad:dashboard:end -->, and the original proposal stays below it, unchanged. The dashboard shows:

• the proposal and phase progress
• one next action in a > [!IMPORTANT] callout
• evidence expiry in a > [!WARNING] callout
• the work items table and a Mermaid dependency graph
• links to artifacts and to the Cast PR, with its merge state
• collapsed sections for the lifecycle map, run history, provenance/cost/JSON, and other proposals

The bot re-reads the body immediately before writing it.

2. Decisions as checkbox cards; ticking is free and the existing command submits them

• Each decision gets one checkbox per option, with ⭐ marking Squad's recommendation. Use stable IDs (D1, W1, R1) and never #N.
• Ticking a box starts no run, because there's no issues: edited trigger.
• There's no separate submit step. The next phase command (e.g. /squad activate, or /squad plan if decisions come out of triage) reads the current body. It then:
  • checks that the person who sent the command has write access, so the command is their approval
  • echoes the resolved answers back ("D1 = A, D2 = B, …")
  • stops with a clear message if a decision has no box ticked or more than one
• Decisions are never answered automatically. Optionally, use userContentEdits to flag ticks made by someone without write access.
• Body text is still untrusted input. Authority comes from the command and the sender's permission check, never from markers or checkboxes alone.

3. Mermaid instead of ASCII

• Plans render a flowchart with decision nodes, work items coloured by state and owner via classDef, and out-of-scope follow-ups shown dashed.
• A stateDiagram-v2 lifecycle map goes in a collapsed section.

4. One status comment per command, edited as the run progresses

• Post it within seconds, then update it in place: 👀 received → ⏳ running (step x of y, run link) → ✅ done (link to the artifact) or ❌ failed (cause, "nothing changed", retry command).
• A precondition failure gets a one-line ⛔ rejection with the command to run instead.
• If a duplicate command arrives while a run is in progress, say "merged into this run" instead of starting another.
• Optionally, move the reaction along 👀 → 🚀 → ✅/😕.
• Limit concurrent runs to one per issue so duplicates can't race.

5. Less noise

• Shorten commands to /squad research P1.
• Make each artifact end with one next step that links to the dashboard.
• Collapse the footer, the "add this workflow" ad and the structured JSON into <details>.
• Use one set of size labels everywhere (S/M/L).
• Pair risk emoji with text so it isn't conveyed by colour alone.

Optional: fewer runs

• Chain research → triage → plan from a single command and stop at the first point where human input is needed. This saves 2 runs per proposal, at the cost of no chance to redirect between phases.
• Put the decision cards in triage's output, so /squad plan uses the human's answers instead of guessing them.

Run budget

Acceptance

• A dashboard section in the issue body is kept current and is the only place that shows "next action"
• Decisions are checkbox cards with stable IDs; the phase command reads, validates and echoes them; incomplete answers block the phase
• No workflow trigger on issues: edited
• Plans and lifecycle use Mermaid instead of ASCII
• Each command gets one status comment updated in place, covering success, failure and precondition rejection
• Duplicate commands are merged and acknowledged, never run in parallel
• Broken references fixed (#aw_… placeholder, #N used for decisions)