Feature request: a light-ship mode (or per-diff toggles) for subscription-capped users
/ship currently has one ceremony level. The <50-line specialist skip is great, but three costs remain unconditional on every ship:
- The Claude adversarial subagent ("always-on for every review... no configuration needed") — a fresh full-context agent even for one-line copy tweaks.
- The document-release pass (+ its verifier agent).
- The workflow turn count itself (~50 tool steps in /ship + ~40 in /land-and-deploy), where every >5-min CI wait expires the Anthropic prompt cache, so each wake-up re-uploads the session's full context at cache-WRITE price.
On Claude Max (weekly usage caps), measured on a real project over 7 days: ~93M input-token-equivalents, of which a one-line wording tweak shipped via /ship + /land-and-deploy cost 324 turns, 3 subagents, ~1.4M cache-write tokens, and a lint-cleanup ship cost 528 turns, 7 subagents, ~4.6M cache-writes. "Boil the ocean" is the right ethos when marginal cost ≈ 0, but subscription caps make the marginal cost real; today the only knobs are the global skip_eng_review (all-or-nothing, "not recommended") and codex_reviews.
Ask (any of these would help):
- A
light_ship_max_lines config: below the threshold AND no user-designated risky paths, skip the adversarial + doc agents and the deep canary — keep tests, checklist review, version/changelog, PR.
- Or per-feature toggles:
adversarial_review: always|risky-only|off, doc_release: always|risky-only.
- Independently: batch the long CI waits in /land-and-deploy into a single blocking script call (one model wake-up instead of one per phase) — this alone removes ~1M cache-write tokens per shipped PR for cap-limited users, with zero behavior change.
Workaround we've deployed locally (happy to share): a "two-lane" CLAUDE.md rule that routes small non-risky diffs around /ship to a native commit/push/PR flow with an inline checklist review, gated by a PreToolUse hook that deterministically re-verifies diff size + risky paths before allowing the push, plus a single-wake watcher script for the CI chain.
Feature request: a light-ship mode (or per-diff toggles) for subscription-capped users
/shipcurrently has one ceremony level. The <50-line specialist skip is great, but three costs remain unconditional on every ship:On Claude Max (weekly usage caps), measured on a real project over 7 days: ~93M input-token-equivalents, of which a one-line wording tweak shipped via /ship + /land-and-deploy cost 324 turns, 3 subagents, ~1.4M cache-write tokens, and a lint-cleanup ship cost 528 turns, 7 subagents, ~4.6M cache-writes. "Boil the ocean" is the right ethos when marginal cost ≈ 0, but subscription caps make the marginal cost real; today the only knobs are the global
skip_eng_review(all-or-nothing, "not recommended") andcodex_reviews.Ask (any of these would help):
light_ship_max_linesconfig: below the threshold AND no user-designated risky paths, skip the adversarial + doc agents and the deep canary — keep tests, checklist review, version/changelog, PR.adversarial_review: always|risky-only|off,doc_release: always|risky-only.Workaround we've deployed locally (happy to share): a "two-lane" CLAUDE.md rule that routes small non-risky diffs around /ship to a native commit/push/PR flow with an inline checklist review, gated by a PreToolUse hook that deterministically re-verifies diff size + risky paths before allowing the push, plus a single-wake watcher script for the CI chain.