Lane orders: size lanes and finish phases from measured lane cost

[erikdarlingdata]

Why

We measured all 34 lane transcripts from 2026-09-23. Four results matter here:

• 40% of lane cost came after the watchdog's 200k warning. Nearly every lane went past 200k.
• The lane orders told lanes to hand off at 150k, which a lane cannot see. Its only signal is the watchdog.
• Lanes re-ran the plain-English checker on their PR body several times, at their largest context. One lane spent
  37 calls at 200-250k tokens after its first warning, on the finish phase alone.
• Every lane read an 8-10k-character PR section to see one example it needed.

What changes

.claude/skills/lane-orders/SKILL.md only:

• For the coordinator, at dispatch: size each lane to finish under 200k. At the measured growth rate, that is about
  80 calls. Paste the lines a lane needs into its brief instead of pointing it at a PR or file.
• For the coordinator, after each wave: measure it with lane-audit.py and record the result and the defects fixed
  per PR in the handoff.
• For the lane: plan to finish before the 200k warning, and start no new item past about 150k. This replaces the
  150k handoff rule.
• For the lane: write the PR body once, run the plain-English checker once, and fix its hits in one pass.

The matching machine-level changes are in the user-level lane.md agent and token guardrails, outside this repo.
lane.md now sets effort: high, because lanes were inheriting the session's max. It also sets
omitClaudeMd: true.

Test plan

• Documentation only. No code, tests or build inputs change.
• The next wave's lane-audit.py numbers, compared with the wave B baseline, show whether these changes worked.

🤖 Generated with Claude Code

https://claude.ai/code/session_016tFt8hAVG5mcWEGEUngTTw