VUH-2796: folding Clankie's user-facing lanes

Work: VUH-2796. Guidance: Anthropic, When to fold subagents into your main agent loop (2026-10-22).

1. Reply latency per lane (2026-10-24)

Read-only analysis of this Mac's durable trails. No live room was touched. Window: 2026-10-06 (ADR 0229 handoffs live) to 2026-10-24, unless noted.

End to end for Discord text is the reply message's snowflake time minus the incoming message's (discord.text.reply receipts). Hops come from turn-settled.jsonl (Pi), the lead's Claude transcript (turn_duration, channel created_at, reply tool calls) and voice discord.voice.response fields.

Pi handoff hops (n=74): ingest and admission p50 3.8 s, p95 39 s; model turn(s) p50 26 s, p95 280 s; return p50 0.8 s, p95 1.5 s. Each handoff starts from an empty Pi context (contextTokensStart 0) and ends at p50 40k tokens after p50 one tool call, so most of each turn's context is the base prompt and tool definitions, loaded again for every message.

Owner main channel hops (escalations, n=68 with reply): Discord to channel event p50 3.2 s; lead-busy queue p50 6.4 s, p95 166 s, max 82 min; think to reply p50 12 s, p95 57 s; deliver p50 0 s, p95 5.5 s. Since 10-10 the lead ran 2,863 worker-report turns, 1,069 watch turns and about 1,170 other service and lead events, against 334 owner-typed turns and 66 owner Discord messages: about 93% of its turns are fleet traffic. One observed case (VUH-2796's own request) held the reply behind a parallel hire_agent call that timed out after 63 s. MCP calls without a read-only hint run one at a time.

Excluded: six Pi replies settled on 2026-10-06 and 10-11/12 but reached Discord only at 10-12 04:37, after a bridge outage. They measure the outage, not a hop.

Caveats: models and load differ across periods, so the before/after for the main channel is observational. The voice sample is small. Console turns are measured from transcript insertion, which excludes TUI-side time.

Reproduce (scripts in latency/):

clankie conversations list > /tmp/convs.json
python3 -I latency/pi_turns.py ~/.clankie/captain/turn-settled.jsonl 2026-10-06
python3 -I latency/lead_turns.py LEAD_TRANSCRIPT.jsonl 2026-10-10 /tmp/lead.json
python3 -I latency/main_channel_hops.py LEAD_TRANSCRIPT.jsonl 2026-10-06 /tmp/esc.json
python3 -I latency/discord_text_hops.py 2026-10-06 /tmp/convs.json /tmp/esc.json

2. Manual lanes eval (2026-10-24)

lanes-eval.ts and its cases are described in the evals guide. The suite has ten golden prompts across four lanes (social room, owner DM, voice, console), two repetitions by default and 88 turns per campaign. Each repetition gets a private scratch service, and nothing reaches Discord or the live service. The runner refuses to run under CI, and no package script, check or test imports it.

Smoke (--smoke, scripted local model, zero spend): 88 of 88 turns settled across 4 lanes and 2 repetitions, with an empty isolation list. Each turn holds only its own reply, its own tools (memory once, only for remember in the owner DM and console) and its own usage. The synchronous-200 replay and refused-401 regressions passed. --fail-after-credential exited 1 and left no copied credential. The scripted model only names the case it heard, so its score says nothing about Clankie. The tools offered per turn confirm each lane's authority: social room 23, voice 18, owner DM 92, console 70. The scratch stack has no connected services or browser, so live counts are higher.

Baseline: not run. A real campaign spends API quota, so it waits for James to choose the credential and model (asked on VUH-2796).

3. First fold candidate (proposed, not built)

ADR 0291 (proposed) would fold Pi text-room handoffs back into each room's durable loop, keeping background forks only for long work. The parent is the room conversation, not the lead. The measurements above argue against folding rooms into the lead's loop: about 93% of its turns are fleet traffic, and owner DMs already wait up to 166 s at p95 behind them.

Observational before and after for ADR 0229 itself, Discord text outside the main channel, with outage-held replies over an hour excluded:

Durable room turns started at a median of 48k tokens of room history. Fresh threads start at zero and end at a median of 40k. Models differed between the periods (mostly gpt-6-astra before, gpt-6.1-sol after), so this suggests the trade-off but does not settle it.

4. Lead seat tool surface (2026-10-24)

Measured from the live native Claude lead seat: its latest deferred_tools_delta attachment and every tool_use from 2026-10-10 to 10-24.

The seat offers 337 deferred tools plus Claude Code's core built-ins. Claude Code already keeps MCP definitions out of the prompt until ToolSearch loads them, so only the names sit in context. The cost is the lookups and the noise, not prompt size.

Of the lead's 124 tools, 76 were never called: all 23 Minecraft, 7 music, 6 PokeAgent, 4 email, the image, video and diagram tools, and most browser and tidy tools. Twenty tools carry 97% of the lead's Clankie calls: message_seat, linear_save_issue, record_drift_correction, linear_save_comment, readopt_seat, worker_reports, hire_agent, close_worker_pane, acknowledge_worker_reports, schedule_wake, linear_wake, linear_get_issue, mcp_tool_call, linear_list_issues, herdr_watch, linear_list_comments, reply, memory, waive_review and settle_hire_receipt.

Recommendations, in order (none built):

1. Stop inheriting unrelated MCP servers into the lead seat. That removes 182 tools nobody called, including a banking server and Gmail send, from an autonomous seat. Documented levers: disableClaudeAiConnectors (or ENABLE_CLAUDEAI_MCP_SERVERS=false) for claude.ai connectors, and --strict-mcp-config --mcp-config or an allowedMcpServers list for user servers. The docs do not say whether strict mode keeps plugin servers, so test it with /mcp first. This changes James's own tools in that seat, so it needs his decision.
2. Mark the twenty hot tools always loaded with _meta["anthropic/alwaysLoad"]: true in the lane tools/list. That removes most of the 341 ToolSearch round trips and keeps about 20 Clankie tools plus Claude's core in context, inside the article's 30–50 range.
3. Group the rest by intent, off the lead's default list. Play (Minecraft, PokeAgent), media (music, voice, image, video, diagrams, YouTube, skin), email and browser belong to rooms and Pi lanes. The lead reaches them through mcp_tool_search and skills if it ever needs them. That takes the lead's Clankie list from 124 to about 70.
4. Annotate read-only tools with readOnlyHint. Claude Code documents that read-only calls run in parallel. The replay-safe list in ADR 0284 is a starting classification. Separately, the 63 s reply wait in step 1 came from a hire_agent startup timeout in the same batch. The docs do not say whether non-read-only MCP calls are serialized, so the timeout itself is a follow-up: hire_agent should return when startup is pending, not block.

Sources: Claude Code MCP, managed MCP, settings reference, CLI reference.

Increment 1 smoke (2026-10-24)

ADR 0291 decisions 1 and 3 for social rooms, behind CLANKIE_ROOM_LOOP (off by default). Scripted local model, two repetitions per arm:

The --room-loop smoke also sends two speakers to one social room at once. Both settle in one loop, and the second is accepted after the first completes. With serialization removed, the same check fails: only one turn settles as its own. The existing room-handoff, burst, hosted, native and room-skill integration suites pass with the switch off (23 tests, 2 skipped).

Not shown yet: anything about quality or latency. That needs the real-model campaign on both arms, still waiting on James's spend decision, and the fork tool for long work.