Load before implementing anything that runs continuously or per-item at scale — animation and render loops, timers, pollers, per-request middleware, large-collection iteration. Do not load for one-shot scripts, setup code, or paths that run a bounded handful of times.
Load before hand-writing or hand-editing files a toolchain normally generates or owns — Xcode project files, lockfiles, generated schemas or migrations, IDE workspace files. Do not load for ordinary source code.
Load when entering an unfamiliar codebase, or before proposing an architecture-level change. Do not load for tiny edits, known local fixes, post-implementation summaries, or tasks already scoped by a current state-map.md.
Load on strike 3, a tangled run, or a contradiction that cannot be localized. Do not load for first failures, ordinary test red/green loops, or known environment outages.
Load for cross-session orchestration — dispatching work to an Albion/GLM worker over tmux, polling a worker completion manifest, or reviewing worker output. Do not load for in-session subagent delegation; use delegation instead.
Load when the stop gate blocks twice on the same task. Do not load for ordinary unfinished work, first gate blocks, or reports where no completion claim is being made.
Load when writing a subagent brief for non-trivial dispatched work. Do not load for direct edits, trivial lookups, user-facing summaries, or delegation already covered by an existing brief.
Use when explicitly invoked as "fable-mode" for GLM-5.2, or when GLM-5.2 is handling complex long-horizon coding, debugging, refactoring, architecture, migration, performance, research, or proof-like work where correctness depends on sustained reasoning, tool evidence, state definitions, memory, and verification. Do not use for trivial edits, simple Q&A, copyediting, or tasks where the user only wants a brief answer.