Ruthless cut audit

Status: historical proposal at 2eb2608d (VUH-1475), not a current inventory or an instruction to delete features. Later decisions in ADR 0203, ADR 0207, and ADR 0213 retain both Discord bodies, retire the menu bar and embedded Swarm, and require native agent delivery. The recommendations and counts below describe the original audit.

Date: 2026-09-30, at 2eb2608d.

ADR 0203 sets the test. Every feature must (1) bring life or character to Clankie, or (2) give the owner something the Claude Code and Codex ecosystems don't, and still be needed if Claude and Codex were twice as good tomorrow. This audit applies that test to every app, package, service, captain tool, CLI and TUI command and bundled skill. It reuses the Discord surface review (VUH-1460) and the 2026-09-30 eval baseline.

Headline

• Cut about 25k to 30k lines of about 249k (10–12%) (source plus tests), in seven batches. Apps go from 9 to 7 and packages from 20 to 17. The largest items are the user-session lab body (about 7.4k lines), the vendored herdr-lead dashboard and nine unused process skills (about 6.2k), and a group of features with zero or failing use: the evaluator, goals, headless agent sessions and the Linear webhook (about 4.4k).
• Two of the issue's named candidates are smaller than they looked.
  • The TUI's 25.9k lines are mostly not duplication. About 9.7k lines are the clankie CLI and the service launcher, which everything depends on. About 5k lines are the only place to enter secrets and OAuth logins, or to set up Discord and voice. The chat shell (about 4.5k) does overlap the app, but ADR 0203 keeps the TUI first-class. What is left to cut is about 1.4k lines: a second hosted console, two duplicated command pairs, and slash wrappers nobody has typed.
  • Terminal-typing control is already mostly replaced: the plugin and app-server adapters landed today. What remains to cut is about 0.8k lines, once VUH-1458 and VUH-1459 are verified live.
• The biggest lever is not in the tables. Clankie's own pi operator lane (bash, read, edit and write on a custom loop) is the part of Clankie that most resembles a custom harness, and the Claude Code and Codex seats now do the same job. Whether it goes is decision D1 below. I have not counted it in the totals.

Method and limits

Sources, read for names, counts, types and dates only. No message text, tool arguments or media were opened. Full tables are summarized in the appendix.

• Usage:

  • pi trees and turn-settled.jsonl (1,383 settled turns since 2026-08-29);
  • Claude Code transcripts (since 2026-08-31) and 1,271 Codex rollouts;
  • ~/.zsh_history;
  • the console's prompt history (.data/tui/prompt-history.jsonl, 319 slash commands, no timestamps);
  • Discord receipts, service logs, relay log and memory stores.

  "30d" means 2026-08-31 to 2026-09-30.

• Size: wc -l over tracked .ts/.tsx/.mjs/.js/.rs/.html/.css, source and tests separately. Figures cover the feature's own files. Wiring in app.ts, tools.ts, protocol schemas and settings is extra, typically 10–20% more.

• Native coverage:

  • Claude Code 2.1.285 and codex-cli 0.159.1 docs and --help;
  • the competing products James listed on 2026-09-30:
    • OpenAI Dots: always-on cloud agents with a browser, 4,000+ connectors and an avatar;
    • xAI Grok Bot: shared cloud VM, 220 connectors;
    • Meta Muse: an avatar companion.

  None of the three runs on the owner's machine, leads other labs' agents, or does Discord voice.

• Dead code: knip finds no unused files today (one duplicate export). The cuts below are whole features, not stray dead code.

What this cannot show:

• Agent use is not product use. Most CLI and tool traffic is agents building and testing Clankie. A command with 50 agent calls may have no user.
• Retention is short. Transcripts start 2026-08-31, pi trees are pruned, and events.jsonl is compacted. The prompt history has no dates, so slash-command counts are lifetime counts since 2026-07-10.
• Only this Mac was read. A hosted body or the Windows PC may use things this machine does not.
• HTTP routes are not logged, so app-only API use cannot be counted directly.

How to read the tables

• Criterion: L = brings life or character; U = unique to Clankie (Claude Code, Codex and the three products don't do it); — = neither.
• Use: 30-day count unless marked. "0 ever" means no call in any retained trail.
• Lines: source + tests.
• Covered by: CC = Claude Code, Cx = Codex, Dots = OpenAI Dots.

Reply with IDs, for example "approve all except C13, C16". Each row names the ADRs to mark superseded when it lands.

Cut, fold and keep calls

Batch 1: dead, failing or zero-use (no product loss)

Batch 2: surfaces from the surface review (VUH-1460)

Batch 3: bespoke connectors and helpers the ecosystems cover

Batch 4: thin harness control (after VUH-1458 and VUH-1459 pass live)

Kept here: the owner's direct terminal (herdr-terminal*.ts, runtime-terminals.ts; U, ADR 0144), the census and the fleet cursor, the seat outbox, both seat plugins, agent-transcript, and the remote-fleet transport.

Batch 5: TUI folds

Batch 6: skills (with VUH-1457)

Kept: this-machine (400 reads), trace-clankie, swarm-mcp, desktop-control, work-items, herdr, pokeagents, computer-use-delegation, research-team, c, p, shared-checkout, solution-space, reflect, testing-archive, and the Linear skills. Anything VUH-1457's eval shows to be noise goes in that issue.

Batch 7: package folds and housekeeping

Keep and invest

These pass the test, and most carry real traffic. Full list in the appendix.

Decisions beyond yes or no

• D1: the pi operator lane. Operator conversations run Clankie's own pi loop with shell and file tools. The Claude Code seat (ADR 0152) and, since today, the Codex seat do the same job inside the lab harnesses. Moving operator work to seats would leave pi running only his social and voice lanes, which is where his character lives. Several thousand more lines would then be candidates:

  • routing and request budgeting;
  • the composer catalog;
  • conversation fork and replay;
  • the operator half of captain.ts and conversations.ts;
  • part of turn metrics.

  Blockers:

  • hosted bodies with capped model usage (VUH-1371) have no harness subscription;
  • the app's Clankie thread reads the operator conversation;
  • I could not split the 739 bash calls by lane, so I don't know how much operator work still runs on pi.

  Recommendation: answer it with VUH-1473 (the seat eval) and VUH-1474 (the lead eval) before cutting anything.

• D2: the hosted-* modules (nine files, 1.9k lines plus 2.8k of tests). These run only on managed bodies. Credits, plan quotas and heartbeat look like control plane, which ADR 0183 puts in clankie-ops. ADR 0183 also keeps the single-owner Linux image public. Moving them doesn't shrink the system, only this repo. Move them, or record why they stay?

• D3: media. Cut video (C14) and keep images? Or cut both, since neither has been used since 08-15?

• D4: Rivals (C16). Keep only if Clankie will actually play Rivals this month.

• D5: Activity avatars (from VUH-1460). When C9 lands, drop custom webhook avatars, or move the route to a public origin.

• D6: Vox video. After C10, trim the roughly 3.3k lines of Rust only watch and Go Live use, or leave Vox alone. The surface review recommends leaving it.

Batches and expected reduction

Each batch is one or more commits with pnpm check green, knip clean, and the docs, trace-clankie, the ADR index and docs/architecture.md updated in the same change. Before counts: 157.3k source lines and 91.7k test lines (249k); 9 tracked apps; 20 packages.

After: about 222k–224k lines (10–12% less), 7 apps, 17 packages, two fewer always-on processes (the Activity server and its public tunnel) plus the evaluator's pane, and about 25–30 fewer tool schemas offered to every operator turn. D1 is the only lever that could move this into the tens of thousands again.

What I'm unsure about

• C11's size. I didn't measure how much of discord-presence-core exists only for the second transport. It could be small.
• Zero use on this Mac is not zero use. For hosted and PC use (C7, D2) I can't see the other machines.
• Slash commands have no dates. C8 rests on lifetime counts since 2026-07-10.
• Skills. "No traceable use of Clankie's copy" (C22) can't tell Clankie's copy from James's global copy when a bare name was loaded. The eval baseline is the stronger evidence.
• Unused doesn't prove useless. Goals (C2), video (C14) and Rivals (C16) never got a fair trial. The ADR's test is use plus coverage, and I applied it as written.

Appendix: inventory

Apps and integrations

Packages

Services

Captain tools (lane bank, shared with the seat's MCP)

CLI nouns (agent calls in 30d, Claude + Codex; James's own in brackets)

• Heavily used:
  • console [66], restart [36], pair [12], seat [9]
  • herdr 289, work 205, linear 154, status 107, pair 79, runtime 63, model 60, gateway 68, mcp 54, swarm 52, doctor 49
• Moderate: skills, devices, voice, conversations, send, prompt, accounts, persona, agents, access, awake, login, memory-card, health, discord, seat-sync, memory, connect, autostart, effort, connections, down, browser, seat-hook, remote-access, metrics, telemetry, fleet, play. All keep.
• Near zero: whoami, disconnect, sessions, file, games, workdir, stance, logout, conversation. Keep: each is a thin noun over a kept API. rivals follows C16 and evaluator goes with C1.
• Zero ever:
  • reset and operator-credential: C7;
  • image-model and video-model: C14 (image-model stays if images stay);
  • terminal, keys and deprovision: D2.

TUI slash commands (lifetime since 2026-07-10)

• Used: /conversation 49, /model 33, /auth 26, /discord 18, /connect 18, /effort 16, /provider 14, /new 12, /persona 11, /herdr 9, /reset 9, /remote-access 8, /games 7, /status 6, /voice 6, /chats 5, /help 4, /connections 4. Keep.
• Retired tokens still typed: /approvals 13, /mission 4, /doctrine 2. Nothing to cut. A "that command is gone" hint would help.
• Never typed:
  • wrappers: C8;
  • /goal: C2;
  • /rivals: C16;
  • /routing, /compaction, /setup, /login, /logout, /pair, /reconnect, /terminal: keep (settings, onboarding or hosted).

Bundled skills (35)