strategy-repertoire.md

The strategy repertoire — bank the method, not only the fix

The idea. Memory & recall remembers the problem and its fix. The repertoire is that idea one layer up: it remembers the method used to solve the problem, so the next time a similar problem appears the agent picks the best-fit strategy from an accumulated toolbox instead of re-deriving an approach from scratch.

This is a human skill — a person carries a lifetime of strategies and matches the right one by instinct — rebuilt in a form an AI can actually run.

Three ways it is model-native, not a human imitation:

1. Externalized, not internalized. A person's repertoire lives in their head; an agent starts each session blank, so the repertoire lives in files (docs/strategies/) — one card per strategy, each keyed by the problem-signature it fits.
2. Auto-surfaced, not recalled-by-will. A person decides to think "what do I know here?"; an agent must not rely on that, because "remember to consult the repertoire" is exactly the invited control that decays. So the SessionStart hook pushes a tiny name — trigger index into context every session — the poka-yoke applied to the learning system itself.
3. Explicit breadth-match, not intuition. A person's intuition prunes many candidates in a flash; an agent's strength is the inverse — it can hold all candidates as text and score each against the problem in the open, which also leaves an evidence trail for free.

Capture is soft first. After a non-trivial problem, the method is banked as a new card in docs/strategies/; a session watchdog makes a stale repertoire visible rather than blocking. Harden to a required artifact only if the log shows capture is being skipped — build soft, prove it's un-annoying, then harden.

Provenance. A strategy earns its rank seed → used → proven by real use, so the repertoire can't fill with plausible-sounding methods that never earned their place.