Surface review: the Discord Activity and the user-session lab body

Status: retirement proposal declined by James on 2026-09-30 in ADR 0203. Both Discord surfaces remain supported. The audit below is historical.

Date: 2026-09-30

ADR 0203 makes the app, Discord text and voice, and the TUI first-class, and asks two remaining surfaces to justify themselves: the Discord Activity (apps/discord-activity) and the personal-lab user body (apps/discord-user-session). The test is the ADR's own: would it still be needed if Claude and Codex were twice as good tomorrow, and is it used weekly?

Recommendation

Both retirements are reversible from git. If James wants a hedge, the cheaper one is to retire the Activity now and keep the lab body one more month behind its existing off switch; the case for that is weaker than it looks (see What would change this).

Method and limits

Sources, read for counts, types and timestamps only. No message text, media or transcript was opened.

• ~/.local/state/clankie/discord-live-receipts.jsonl (29,185 receipts, 2026-07-25 to now), discord-bridge.log, activity.log, tunnel.log.
• ~/.clankie/events.jsonl (176 presence session snapshots, 46 embodiment.* session claims across the file).
• Tool-call names (never arguments) in the durable Pi trees under ~/.clankie/captain/{rooms,voice,turns}; 108 files modified in the last 31 days.
• ~/.config/clankie/settings.json for the active body and lab-body switches.
• git log, wc -l and grep over the checkout at dca0902a.

"Last month" below means 2026-08-30 to 2026-09-30.

What this cannot show:

• Viewers. The Activity server logs no viewer connect or disconnect, and its stdout log carries no timestamps. It only reports dropped frames with a viewer count. So there is no per-day viewer number, and the Discord Developer Portal analytics were not checked.
• Other machines. Only this Mac was read. A lab body running on the Windows PC or a hosted box would leave its receipts there.
• Usage before 2026-07-25. The retained trails begin that day.

Discord Activity

Who used it in the last month

Effectively nobody, and the audience is capped by Discord regardless.

• The Activity is launched by the captain's discord_watch_start tool on the bot ("I posted the live play launch in voice"). Every call in any retained Pi tree is on 2026-08-18 or 2026-08-19 (six calls). None since. No activity_start appears in events.jsonl.
• All 176 presence snapshots carry activityInstances: [].
• The Activity only shows a game, and the game barely ran. gba-play journals end on 2026-08-30. events.jsonl shows two play sessions claimed that day (18:14 and 18:23 UTC), then one claim on 2026-09-29 that was refused as world_unreachable. No Activity launch accompanies any of them.
• activity.log holds five backpressure lines, all "1 viewers", and no dates. They are consistent with one person (James) watching, at some point since 2026-07-25.
• Discord restricts an unverified Activity to the app team's developers and testers in servers with under 25 members (README). The lab guild is the only place it can ever run. Going public means Discord verification.

What only it can do

Show live GBA frames with synchronized cartridge sound to the people in a voice channel, from the official bot, with no user account involved. Nothing else in the repo does that.

I did not verify whether a PokeAgents world exposes its own watch view. The architecture doc lists the Activity as an "optional game watch surface" (architecture.md). The app is the first-class place to watch Clankie, and I did not check whether it renders game frames.

It also has one duty that is not the game: it serves persona avatar PNGs at /avatars/agent-<persona>-<sha>.png through the public tunnel, because a Discord webhook avatar_url must be a public HTTPS URL. 504 avatar files exist. The consumer is channel projection posts; there were seven such receipts, the last on 2026-09-01. personas.presentation builds the URL from discord.activityTunnelHostname and returns no avatar when that is unset (apps/clankie/src/captain/personas.ts), so removing the host degrades to default-avatar webhook posts rather than failing.

Code and maintenance cost

Roughly 1.6k lines of its own, plus wiring. ADRs 0047 and 0114 own it; 0147 gave it the avatar route.

Git churn is low: 25 commits since 2026-07-25 but two since 2026-08-30 (a docs drift sweep and the avatar route), +90/-13 lines. The cost is not code churn. It is standing infrastructure:

• The launcher supervises two extra processes all day, the Activity server and a named cloudflared tunnel. Both are running now.
• tunnel.log is 45k lines, 19k of them errors (QUIC timeouts and reconnects, plus 1.4k failed requests with "origin unreachable" on 2026-08-16 and 17). A public, internet-facing listener exists for a surface with no users.
• Setup is heavy for a DIY install: a Discord application, an Activity URL mapping, a Cloudflare zone on an active account, a named tunnel, a DNS route and a hand-written ~/.cloudflared/config.yml. This serves the self-hosted audience poorly, and the hosted audience can never use it while it is unverified.

Recommendation: retire

It fails the weekly-use test and cannot reach a public audience without Discord verification. Carrying it out is small:

1. Decide the avatar duty. Either drop custom webhook avatars (posts fall back to the default, no code change beyond removing the route), or move the one route (about 30 lines and a directory read) into the clankie service and expose it on an origin that already exists. I did not find a public origin in this repo to host it; api.clankie.bot lives in the private clankie-ops.
2. Delete apps/discord-activity, packages/rendered-surface-client, the producer credential, the activity and tunnel launcher services, the activity* settings keys, and the activity_start path in the bot runtime. Keep rendered-surface.ts only if play still uses the frame contract.
3. Update docs/discord-media.md, the play README, docs/architecture.md, trace-clankie (service ids) and mark ADR 0047 and 0114 superseded.
4. Stop the named tunnel and remove the Cloudflare route.

Loses: a way to show Clankie's game inside Discord. The game itself, play voice and the app are unaffected.

User-session lab body

Who used it in the last month

Nobody, on this machine, and no trace of it having run since the trails begin.

• settings.json has activeBody: "bot" and userSessionEnabled: false, with empty guild, channel and voice allowlists.
• discord-user-session-receipts.jsonl does not exist anywhere under ~/.local or ~/.clankie. The bot's receipt log from the same day does.
• There is no discord-user-session.log or discord-user-session-service.json in the state directory. Every other service has both.
• All 176 presence snapshots are discord:bot:... sessions with transportKind: bot. None is a user session.
• discord-live-receipts.jsonl (bot) has zero stream, watch or publish types, as expected.
• Bot voice is what actually ran: 66 joins in the receipt log since 2026-07-25, eight of them on 2026-09-29.

Commits show the body was built hard between 2026-07-25 and 2026-08-19 (watch and Go Live through Vox). I found no receipt showing either ran live here. That is absence in retained trails, not proof it never ran.

What only it can do

Two things, both because Discord withholds them from bots:

• Screen-share watch: join a share, decode a JPEG per second, four-frame window, the captain's observe_share.
• Go Live publish: H264 video to a channel.

It also duplicates the bot for text and voice (shared through packages/discord-presence-core), and adds nothing there. It cannot register slash commands. Automating a normal account violates Discord's terms and can get the account terminated; the body carries a recorded owner opt-in for exactly that reason. That risk is awkward next to the hosted product, where ADR 0203 lists Discord text and voice as first-class without it.

Code and maintenance cost

Roughly 6.3k lines in the app and its tests plus about 3.3k lines of Rust that only watch and publish use (of Vox's 15.3k). Also: settings keys, the TUI /discord lab flow and opt-in, launcher service, credential-broker acceptance, and ADRs 0024, 0048, 0098, 0100 and 0128.

Git churn is higher than the Activity: 33 commits since 2026-07-25, seven since 2026-08-30 (+270/-16). The recent ones are shared voice behavior (voice arrival choice, leave decision, transcript wording) that had to be threaded through both bodies. That is the standing tax: every change to how Clankie behaves in a voice room has a second body to wire and test, for a body that is off.

Recommendation: retire

Its unique abilities had zero observed use, and it is the largest piece of code guarding a capability James is not exercising. Removing it also removes the two-body wiring from every voice and text change, and the account-termination risk from the product story.

Carrying it out:

1. Tag the last commit that has it (for example lab-body-final) so it can be restored.
2. Delete apps/discord-user-session, stream-watch-observation.ts, the observe_share tool, the userSession* settings keys, the launcher service and alias (user-session, lab), the TUI lab-body flow, the discord_user_session credential provider and the shared-core branches that only exist for a second transport.
3. Decide Vox separately: leave the video Rust in place, since Vox is its own AGPL package with its own provenance record and costs nothing at runtime, or trim about 3.3k lines that nothing else calls. I recommend leaving it for now and revisiting once the Discord side is gone.
4. Update docs/discord-media.md, credentials.md, architecture.md, trace-clankie (the user-session receipt row and gotchas), the Discord READMEs and mark ADRs 0024, 0048 and 0098 superseded.

Loses: screen-share watch and Go Live. If James wants either back, it returns from the tag as a deliberate feature with an owner, rather than as a dormant body.

What would change this

• Keep the Activity if James expects to run Clankie's game for other people in Discord soon, and is willing to pursue Discord verification. Then the missing work is verification and an audience, not code.
• Keep the lab body if screen-share watch is a capability he wants to grow. The switch is one setting, but I would then want a named use, since nothing in the last month exercised it.
• My evidence is thinnest for the Activity's viewer count (no logging) and for other machines. If either the PC or a hosted box runs a lab body, that changes the second recommendation, and it is a quick check for someone with access.

Decisions for James

1. Retire the Activity? If yes, avatars: drop them, or move the route to a public origin (which one)?
2. Retire the lab body? If yes, leave Vox's video code or trim it?
3. Confirm no lab body or Activity runs on another machine before deletion.