Claudinite growth: extract lessons

[missingbulb]

Daily growth capture for the 2026-07-28 window (dispatch #533). Three lessons, all prose — each is in-flight judgment or a machinery contract with no repo-state signature a check could carry confidently.

Added

.claudinite/local/packs/claudinite/RULES.md

• A scheduled task is gated twice — prove a change through planRun, never through its own precondition. planRun computes the due list before any precondition runs, so a task whose most-recent slot already ran is never evaluated and never reads its signals. Scheduler: an opaque override bag only the interested task reads #513's forced-run override landed with every test driving the precondition or the collector wire directly, and was inert on exactly the run it existed for (evaluations: [], fixed in Forcing a task must clear the SLOT gate too, not just the precondition #516). Drive the real planRun from a deliberately non-due slot.
• Every due task in a run shares ONE checkout — an agentless worker must not touch it. A worker that checks a branch out hands the next task a tree it did not expect. Regenerated output lands through engine/scheduler/deliver-generated.mjs (git plumbing against the fetched base tip); baselining's own deliver is the deliberate exception (Skill-usage metrics: enriched capture, per-repo fold, fleet aggregation #524).

.claudinite/local/packs/canon-curation/item-routing.md (worthiness bar)

• Readable through the channel it lands in. A rule whose premise is that the delivery channel itself failed cannot be carried by that channel — the "SessionStart hook didn't fire" tell was admitted as pack prose in Growth: promote 5 lessons from the fleet's local packs #497 and removed on review, because pack prose is what that hook delivers. Detecting a channel's own failure belongs in code; reject such an item rather than rewording it.

Considered and dropped on dedupe

• file-placement: exempt the pack -> engine-surface reach the contract mandates #517's "a waiver every adopter must write is a rule flagging a constant" — already landed in the file-placement skill by that same PR.
• Pack badges: a mark per pack, and a README row bootstrap and baselining maintain #525's "the engine is what every consumer vendors, so a maintainer-side image tool lives in dev/tools/" — already stated in packs/README.md.
• loadConfig dropped the mount stamp, so baselining self-skipped fleet-wide #518 (loadConfig dropping a declared key) and Scheduler: fail closed on an unreadable run ledger, narrow the first-run fan-out #523 (null conflated with "never ran") — the silent-degradation family the existing fail-soft steps hide their own breakage rule already covers, each now pinned by a general-form test.
• Executor: read the CCR trigger, and never guess an issue #511's CCR_TRIGGER_* transport, and the one-checkout rationale — documented at their own implementation sites, which is where a usage-site gotcha belongs.
• Drop vestigial agent_instructions from agent_model: 'none' tasks #530, Remove grow_with_claudinite's retention_days adoption question #527, Two instructions that stopped being true: adoption, and executor-routine setup #507, growth-extract: rule out lessons that just narrate a settled, already-tested decision #512, Claudinite growth: conversation extract #501 — settled, already-enforced choices, or lessons already captured in the window itself.

Verification

node --test $(git ls-files '*.test.mjs') — 891 pass, 0 fail. check_the_world — 0 blocking, 8 advisory (all pre-existing engine-internal file-placement reaches).

Refs #533. Logged on the standing tracker #370.

🤖 Generated with Claude Code

https://claude.ai/code/session_016PyeCgQGF3qMHdziP38f2C

---

Generated by Claude Code