wabi-finish-loop

name	wabi-finish-loop
description	Run the Wabi website-finish implementation loop — Hermes captain + OpenCode DeepSeek-v4 workers, card cadence, free-model stall recovery, and Tim deploy batches. Use when Ronin asks to finish wabi.chat via the loop, burn SuperGrok/DeepSeek battery on DoD waves, or continue multi-card website polish without model-hopping.
version	1.0.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["wabi","finish-loop","opencode","deepseek","deploy"],"related_skills":["wabi-deploy-debug","opencode","coding-agent-orchestration"]}}

Wabi website-finish loop

Class-level workflow for burning down docs/plans/*website-finish* (or equivalent DoD) on live wabi.chat / Tim.

Roles (locked)

Role	Who	Model
Captain	This Hermes session	SuperGrok / stable captain — diagnose, verify, deploy, finish stalled cards
Worker primary	OpenCode CLI	opencode/deepseek-v4-flash-free (Ronin confirmed more reliable than Laguna, 2026-08-02)
Worker backup	OpenCode CLI	opencode/deepseek-v4-flash-free (user: "backup should be deepseek")

Never dispatch OpenCode Grok (opencode/grok-4.5, grok-build-0.1) for this loop. Zen Grok needs Zen billing; Ronin's SuperGrok is Hermes xAI OAuth and does not unlock OpenCode Zen Grok. User corrected 2026-07-23: use Laguna not Grok; backup DeepSeek free.

Smoke before first card (try primary, then backup):

# CAUTION: opencode free-tier can silently rate-limit — the model prints only the
# build banner and exits 0 with NO actual model response. A banner-only output is
# NOT a successful smoke test; the smoke prompt must return the literal string.
opencode run 'Respond with exactly: OPENCODE_SMOKE_OK' -m opencode/deepseek-v4-flash-free
# if fail / stall rate-limit: try -m opencode/laguna-s-2.1-free
# must print OPENCODE_SMOKE_OK (banner alone is not success)

NOTE: On Bazzite, the shell MOTD banner can flood worker stdout and cause exit 143/SIGTERM before real work starts. Always use pty=true in the Hermes terminal() call for the wrapper. Workers spend ~30s in MOTD before reaching the real build banner. Deepseek smoke test may timeout on first attempt (60-90s) but succeeds on retry.

Full-loop cadence (user preference)

When Ronin says full loop, stop checking in per chunk, or do all of the loop:

1. Do not pause for multi-paragraph status after every card.
2. Chain: write card → dispatch DeepSeek → wait/plateau → verify → next card (or parallel non-overlapping surfaces).
3. Surface him only for blockers, real-browser checks, or batch deploy pulses.
4. Lead with plan once, then execute autonomously.

Card dispatch

1. Prompts under /tmp/wabi-loop/card-*.md with allowed/forbidden paths, no commits, report path under docs/plans/.
2. Python subprocess so the prompt is a positional arg (not -f alone):

   subprocess.Popen(['opencode','run', prompt, '-m','opencode/deepseek-v4-flash-free', '-f','docs/plans/...'], ...)
   

   Use pty=true in the Hermes terminal() call to avoid the Bazzite MOTD banner blocking the wrapped command (workers spend ~30s in MOTD before real work starts). Deepseek smoke test may timeout on first attempt (60–90s) but succeeds on retry. If deepseek unavailable, fall back to opencode/laguna-s-2.1-free (smoke-test it too).
3. One writer per file set. Captain may work non-overlapping modules while a worker runs.
4. Baseline git status --short before each dispatch.

Free-model stall recovery

Laguna/free workers often plateau: log size frozen ≥5 min, PID still warm, handlers half-written, no report.

1. Kill by exact PID (pgrep -af 'opencode run').
2. Harvest partial diffs; finish wire-ups in-session (captain).
3. Write the card report yourself if missing.
4. Common worker bugs to fix on harvest:
   • Invalid Svelte $: items: Type[] = ... → $: items = (...) as Type[]
   • ContextMenu uses open + on:close, not onClose prop
   • Handlers written but never bound to on:contextmenu / never rendered

Do not wait out a full free-model idle timeout hoping it resumes.

Model preference (2026-08-02): Ronin confirmed deepseek-v4-flash-free is "by far more reliable" than laguna-s-2.1-free (which he calls "Ling" and finds "maddening"). When dispatch is viable, use opencode/deepseek-v4-flash-free as primary worker; laguna only as fallback if deepseek unavailable. Smoke-test before dispatch.

Model-down detection (2026-08-02): When the entire opencode.ai free tier silently stops invoking the model — smoke test prints > build · deepseek-v4-flash-free with NO literal response, exits 0, and opencode.log shows Unexpected server error or No payment method — DO NOT retry dispatch. The free tier revokes ~90 min after initial availability. Switch to direct Hermes implementation. See references/model-down-detection.md (shared with opencode-dispatch skill).

Design-polish multi-screen (OpenCode + Hermes)

When Ronin asks to comb the whole app (audit/prompt-design-polish.md / dispatch-design-polish.py):

1. One screen per worker (1 chat … 6 empty + separate login/hub). Parallel only on disjoint paths + separate logs.
2. Worker = mechanical tokens only; Hermes taste + bug review. Token compliance ≠ visual redesign — login can look "untouched" if LaunchPanel was zero-edited; say so.
3. Split design vs functional commits; bun run check stays at known baseline.
4. Never stash-and-forget another session's crash fixes for a "clean" design STATIC_BUILD — ships design without MessageContent/MessageList fixes → live SUBSCRIBE_FAIL / each_key_duplicate. Commit crash fixes before/with the design bake (wabi-postdeploy-runtime design-deploy stash cascade).

Deploy batch (after FE/server cards)

Binary embeds SPA (working tree is the SPA source):

cd frontend && rm -rf build .svelte-kit && STATIC_BUILD=1 bun run build
# must see index.html + _app/ — not handler.js
cd .. && cargo build --release -p wabi-server
scp target/release/wabi-server root@100.96.11.45:/home/tim/Desktop/Wabi/target/release/wabi-server.new
# Tim: SHA match → stop → rm BOTH locks → mv .new → up -d
# Prove :3001 + :8088 health + public CSS hash

Never terser minify. Headless cannot verify UI — Ronin real browser + hard refresh.

Runtime footguns: wabi-deploy-debug / wabi-postdeploy-runtime. wabi-deploy is often user-owned — hermes curator adopt wabi-deploy before agent can patch it.

Channel-type surfaces (recurring UX/API)

When finishing wiki/forum/gallery (not plain chat):

1. No double chrome — Chat already has the channel header. Do not stack SurfaceHeader with the same title + giant primary CTA. Compact toolbar row only (New page / New thread).
2. Unique icons — not all #. Sidebar + header: forum / wiki / gallery distinct glyphs.
3. Create-channel — Forum + Wiki must be enabled options (never greyed "coming soon" if backend kinds exist).
4. Settings per kind — ChannelSettingsModal must branch: wiki/forum/gallery get their own title + options; hide chat retention / spoiler / purge on non-chat kinds.
5. API paths — Wiki is /api/wiki/{channelId}/pages (NOT /api/channels/.../wiki/pages). Forum is /api/forum/{channelId}/threads. Albums /api/albums.
6. Create-then-read race — adapter run() must barrier().wait_for(commit_seq) after commit or create_album / create_forum_thread return 500 "created but not found in projection".

Detail: references/channel-type-surfaces.md.

References

• references/loop-playbook.md — condensed checklist + model lock
• references/channel-type-surfaces.md — wiki/forum/gallery chrome, paths, projection barrier
• references/silent-rate-limit-detection.md — detecting banner-only silent rate-limit exits from opencode free-tier workers