feat: complete the feedback loop — session thread, piggyback delivery, watch guidance

[benvinegar]

Started as the session-thread fix; grown (per discussion) into the full feedback-delivery story. Three commits, three layers:

1. Session thread in the viewer (0ebe057)

Comments not attached to a snippet (sideshow comment without --snippet, or POST /api/comments with only a session) were stored and delivered to agents but never rendered in the viewer — the agent believed it replied; the user saw silence. The reverse was impossible too: no UI existed for the user to write a snippet-less comment.

A session thread now sits at the bottom of each session's stream: renders snippet-less comments live over SSE, and adds a "Message the agent…" composer. New snippet cards insert above it. Verified live in the browser; e2e on Chromium + WebKit.

2. Piggyback feedback delivery (7ade106)

Agents previously heard feedback only while blocked on the long-poll — so they either waited needlessly or missed comments until the next checkpoint.

Now every agent write (publish, update, reply) may carry a userFeedback array: comments the user left since the agent's last call. Mechanics:

• Per-session agentSeq cursor (both stores). SqlStore adds the column via a pragma_table_info probe so pre-existing Durable Object boards migrate in place.
• Delivered once: piggyback advances the cursor, and a consumed author=user wait advances it too (no re-delivery in either direction). The viewer's unfiltered reads never touch the cursor.
• The user's own POST /api/comments never carries feedback; only agent-authored writes do.
• MCP (both HTTP and stdio) tool descriptions + instructions teach agents to treat userFeedback as messages from the user; the stdio server passes the field through untouched.

3. Background watch guidance (docs, same commit)

For the stretches where the agent isn't touching sideshow at all: the skill, /setup block, and design guide now teach arming sideshow wait --timeout 600 as a background process — it exits the moment the user comments, which surfaces the output through the harness's background-task notifications; handle and re-arm. (Proven live in a Claude Code session: comment typed in browser → background long-poll exited → agent woke and replied, no polling.) Includes the arm-on-the-right-session caveat learned the hard way.

Together: piggyback covers actively-publishing agents, background watch covers quiet stretches, blocking wait remains for explicit checkpoints.

Tests

• Store contract: agentSeq defaults, monotonic advance, unknown-session no-op; JsonFileStore reload persistence (including pre-agentSeq data files).
• API: piggyback on update/reply, delivered-once, wait-consumes, viewer reads don't consume, MCP result carries the field.
• e2e (Chromium + WebKit): session thread renders agent comments, composer round-trips, cards stay above the thread.
• Full gate green: 46 unit/API tests, 10 e2e, typecheck, lint, format.
• Live-verified against a 0.2.0-era data file (cursor migration) and in a real browser session.

🤖 Generated with Claude Code