Configuration Persistence
Persist stable nonsecret configuration and validate it when loading.
Persist stable nonsecret configuration and validate it when loading.
Load a named reference at the step that needs it, match an explicit key and cite the record; keep missing-data work pending.
Embed compact domain facts with scope and version next to their use.
Write a self-contained brief with task, constraints and artifact pointers.
Persist a versioned progress ledger so a new session can resume work.
Record a verified failure lesson with applicability and disconfirming conditions.
Keep the entrypoint small and load branch-specific material only when needed.
Separate persistent facts, preferences and resource pointers with freshness rules.
Save task identity, constraints and verified progress before compressing history.
Transfer binding global values without paraphrasing their meaning.
Track dependency fingerprints so regenerated upstream inputs invalidate derived artifacts.
Define canonical terms and explicitly disambiguate near-synonyms.
Carry a server-issued change token between recurring reads.
Preserve service cursors for paginated APIs and complete-unit boundaries for generated documents; distinguish partial results from completion.
Separate human requests from parent-agent dispatches, injected text and tool results.
A fresh resource validator does not make a previously transformed answer relevant to a different question.
Preserve attempted ideas and their rejection evidence even when the implementation is reverted.
Verify referenced material in the receiver installed layout rather than only the author repository.
Avoid stacking a full meta-router on top of a host that already selects the same skills natively.
Keep protected content in an authoritative store while the editing view contains explicit opaque references.
Retain structured response state for continuation and derive human-readable text as a separate view.
Recommend a concrete resource only when the user wants learning material and a current catalog supports it.
Compare the scope and load conditions of two instructions before treating different wording as a conflict.
For a host with a verified prefix-cache contract, place stable inputs before volatile inputs and check the rendered overlapping prefix when reuse regresses.
Keep shared identity fixed while varying individual artifact purpose and composition.
Break dense references into units that preserve the details needed downstream.
Express design tokens as semantic names, exact values and functional roles.
Describe behavior and named interfaces instead of volatile file locations in delayed task handoffs.
Store demonstrated knowledge separately from material merely covered.
Keep one concept definition, decision rules and exceptions together after selecting its disclosure tier.
Recall substantive rejected concepts without poisoning that memory with already implemented requests.
Give a mined behavior a stable identity tied to its actual enforcement point and retain explicit test and uncertainty references.
Assign each durable fact to one document role and reference it from other views.
Anchor project conventions to current concrete exemplar files, with explicit differences and anti-patterns.
Report incomplete retrieval separately from a complete search with no matches.
Refine retrieved context using explicit missing-information notes instead of choosing a fixed file set upfront.
Check current authority and explicit supersession before reusing a saved decision.
Admit only information available at the decision time when replaying or evaluating a historical task.
Store original statements separately from interpretations, candidate formulations and unresolved contradictions.
Bind memory recall to the current project, recipient and permitted scopes.
Resolve the exact library and requested version before using retrieved API guidance.
Carry accepted requirements into later revisions while preserving the amendment history.
Give hosts that can read instruction files but cannot run the full skill suite a small, versioned behavioral core.
Measure always-loaded metadata separately from the full content eagerly loaded for one invocation.
Verify persisted context in the intended account, source and storage target before reporting a successful save.