2.16.1: plan-mode receipts, and a bound on what a receipt may cost - #94

Documentation and receipts. The shipped instructions under plugins/skiphow/ are byte-identical to 2.16.0; nothing a project installs behaves differently.

What this settles

The 2026-09-02 dogfood audit of two installed CI-cost sessions named one candidate for why methods go unread in long installed sessions: both began in plan mode, whose host reminder opens by telling the run to dispatch Explore agents in parallel with the agent type already chosen. Eleven isolated sessions on exact unmodified v2.16.0, plan mode the only variable, show the two arms behaving identically — all eleven opened delegation before their first spawn, and all eighty-nine spawns named a level. The only asymmetry favours plan mode, which opened execution-health in four of its six sessions and none of the five without.

Because the clean fixture never reproduced the failure, this bounds plan mode as a sole cause, not as one term in an interaction. The cause stays UNVERIFIED and no candidate replaces it. The audit's own recommendation — change nothing in the package — is followed.

What changed

AGENTS.md now bounds what a receipt may cost. This pass spent forty sessions across fifty-one invocations, $165 and nearly five hours on one yes-or-no question, while the four sessions that carried the answer cost under six dollars between them. Name the observable and stop when it lands; pilot one session per arm and run one more; keep one fixture and fix an unproductive prompt from the record rather than by running another batch; cap spend and sessions in flight; give every run its own fixture directory and log.

It also records that disabling every setting source keeps the maintainer's own instruction files out of a session's context but does not stop the session from reading them with a shell command — one run in this pass did exactly that fifteen seconds in, and was voided.

Checks

• python scripts/check.py passed, git diff --check clean.
• scripts/check_hosts.py: Claude package validation and isolated install PASS; Codex UNVERIFIED (no Codex validator on this machine).
• Cross-host review: two rounds in Codex. Round one found the receipts lacked budget controls, which is why the deciding pair was rerun under a per-session limit, plus five factual corrections. Round two found three more, all applied: the refutation was overstated, the session count conflated invocations with sessions, and the stop-at-the-dispatch rule would have forbidden measuring post-dispatch delegation behaviour the repository leaves UNVERIFIED.

🤖 Generated with Claude Code