VUH-1456 — standing instructions cut to identity, boundaries and where things live
Date: 2026-09-30. Issue. ADR 0203, run guide, baseline. No push, deployment or live-service restart was part of this work.
What changed
apps/clankie/src/captain/instructions.md went from 14 sections to five: Identity,
Trust, Remembering, Where things live, Honesty. The text before the cut is frozen
as instructions-pre-1456.md
(eval arm pre-1456); the candidate that was tested is
trimmed.md, byte-identical to the new file.
| Old section | Where it went |
|---|---|
| Leading agents | Generic judgment was already in the lead skill (delegation, one harvest owner, fixed done, tracker identity). Clankie tool usage is in the hire_agent, message_seat, herdr_watch, work_items and Swarm connection descriptions. Kept: trust in delegation, hire_agent over bare herdr agent start, clankie herdr agent list, pane naming, the ownership boundaries. |
| Connected work (Linear inbox protocol) | Already complete in the this-machine skill; the wake text now points there. Kept: tracker identity, mail and account boundaries, mcp_tool_search. |
| Skills, Looking things up | Computer-use delegation is on the reach card; skill-upkeep reminder dropped (no reflection ritual). Kept: $skill-name and disabled-skill rule, clankie doctor / this-machine. |
| Initiative | Goal rules are in create_goal, note_goal_decision, schedule_wake and the goal-turn prompt. Kept: propose in conversation. |
| Remembering | Cut from ~480 to ~130 tokens; retain, corrects and status-receipt guidance were already in remember_episode. |
| Your other rooms, songs, making things, diagrams, PokeAgent | Already in the tool descriptions; gaps added there (host recovery via pokeagents, voice handoffs in observe_room, narration in get_self_state, draw only what is true). |
| Showing what you saw | A new # In Discord paragraph in the reach section, Discord lanes only. |
| Long code in files (VUH-1391) | The # Machine access paragraph, shell-holding lanes only. |
Hired workers never received the standing instructions: they get their brief, the work-tracking contract and the selected skills. That is unchanged.
Tokens per lane
Counted with o200k_base (Codex's tokenizer) on the real assembleLanePrompt
output, default persona, no fleet notes. Claude's tokenizer counts roughly 10%
more (the lead's 6.2k for the old file).
| Lane | Before | After |
|---|---|---|
instructions.md alone | 5,584 | 1,194 |
| Operator (pi console), all sections | 5,870 | 1,519 |
| Discord text, social | 5,915 | 1,608 |
| Discord text, machine grant | 5,880 | 1,612 |
| Discord voice | 5,915 | 1,608 |
| Gameplay | 5,832 | 1,442 |
| Claude seat output style | 6,045 | 1,655 |
| Codex seat instructions | 5,795 | 1,405 |
Tool descriptions gained about 130 tokens net across five tools. Pi lanes send them only where those tools exist; the Claude seat defers MCP tools, so they cost nothing there until searched. The 35-skill catalog (~1.1k) is unchanged (VUH-1457).
Gate
Codex (the gate). pre-1456 (the prompt before the cut, which current
meant until this change) against trimmed, on codex-cli 0.159.1 with
gpt-6-astra, all 16 cases (incidents, social, held-out) × 3 reps, split into
two halves run in parallel on the main account (~/.codex) after James reset
it: 96 calls, no errors, default usage guard. Halves:
A, B,
summary.
| Arm | Pass rate | 95% CI | Tokens/trial (95% CI) | Wall/trial |
|---|---|---|---|---|
pre-1456 | 48/48 (100%) | 93% to 100% | 64k (57k to 70k) | 26 s |
trimmed | 48/48 (100%) | 93% to 100% | 47k (42k to 52k) | 25 s |
Paired by case: pass rate +0 (within noise, every slice); tokens −17k per trial
(−27k to −11k), outside noise. Every incident case and voice-interruption
passed 3/3 in both arms; voice-interruption fell from 112k to 31k per trial.
A first three-arm Codex run (bare,pre-1456,trimmed, 5 reps) was stopped at 56
trials as too slow; all 56 passed, bare averaging 36k, pre-1456 67k and
trimmed 48k (report). On Codex even
bare passed voice-interruption, so there the suite is at ceiling on every case.
Earlier attempts stopped at the usage guard after one call each: ~/.codex at 91%
weekly before its reset (report) and
~/.codex-jamescvolpe at 92% weekly (resets 2026-10-03 12:03 CDT). No guard was raised.
Claude, small, partial. Following James's fallback, current against trimmed
on Claude Code 2.1.285 with claude-sonnet-5-5, planned as 16 cases × 2 reps.
The five-hour window, shared with live workers, reached the 80% guard after 20
calls; the run was stopped rather than resumed after the 12:50 reset, to leave
that window to the workers. It covers every incident case and five of the six
social cases once. social-injection and the held-out slice did not run.
Report, summary.
| Arm | Pass rate | Tokens/trial (95% CI) |
|---|---|---|
current | 10/10 | 50k (41k to 58k) |
trimmed | 10/10 | 31k (23k to 40k) |
Paired by case, tokens −19k per trial (−23k to −14k), outside noise. The
baseline's bare arm averaged 24k on the same suite, so the trimmed layer costs
about 7k per trial where the old one cost 27k. All five incident regressions and
voice-interruption passed in trimmed (it answered clankie metrics, the case
bare failed 0/5 in the baseline).
This is a weak gate. The suite is at ceiling for both models, and three repetitions (one on Claude) cannot show a small pass-rate difference. It shows that the cut does not break the known cases and that it cuts tokens; it cannot show that nothing was lost. The seat and lead evals (VUH-1473, VUH-1474) are the real gates.
Validation
pnpm exec vitest run on the lane-prompt, both plugin, eval-runner, Linear
webhook, operator-context, herdr-seat and lane-MCP test files: 8 files, 96 tests.
node integrations/{claude,codex}-plugin/build.mjs --check pass. In pnpm check,
lint, typecheck and infra passed. Formatting failed only on another agent's
in-progress scripts/evals/isolation.mjs and run.mjs, and knip only on its
untracked scripts/evals/seat*.mjs.