context-police

context-police — skills-catalog audit, curation & measurement in agent harnesses

(formerly skills-catalog-context-cost-skilloverrides-scoping. v2.0.0 reframe: the problem + method are harness-agnostic; the Claude Code levers are ONE implementation. Earlier versions were Claude-Code-only, and a big chunk of that work was superseded when harnesses added native budgets — see "History" at the bottom.)

The problem (any harness with an always-on skill catalog)

An agent harness injects the listing of available skills/tools (names + descriptions) into context every turn, and into every sub-agent's base context. An auto-minting loop (claudeception-style: ~1 new skill/session) grows that catalog unboundedly, and every entry is force-loaded forever. Two real effects: (a) per-turn and per-subagent token cost that multiplies on fan-out (a trivial subagent was observed carrying ~30k tokens of base context for a one-word reply); (b) small-context agent types overflow on launch ("Prompt is too long", 0 tokens). Skill bodies lazy-load on use on every modern harness — it's the always-on listing that's the tax.

The portable core — this is what travels to ANY harness

The levers further down are platform-specific; these ideas are not — they're information architecture + arithmetic.

1. Catalog cost is real and multiplies per sub-agent. Measure it (appendix recipe), don't hand-wave.
2. Most auto-minted "skills" are episodic lessons, not skills — single-incident gotchas (flask-flash-silently-dropped-…). A lesson belongs in a searchable archive surfaced on demand, not the always-loaded catalog. The bloat is a knowledge base in the wrong substrate.
3. Curate by description INTENT, never name shape. Warning-shaped names are often real traps; command-shaped "lessons" exist. A hyphen-count heuristic mislabeled 171/886 skills. The discriminator: does the agent go LOOKING for it BY NAME (procedure → keep) or does it only help if SURFACED REACTIVELY to warn of a specific mistake (trap → curation candidate)?
4. Conservative, asymmetric bias. Hiding a real procedure (or restoring a user command) is the silent, costly error; failing to hide a trap is harmless (a few unrealized tokens). When in doubt, take the harmless side.
5. A "hide from auto-invocation" flag has TWO roles — don't conflate them (the reverse-audit footgun, below). One is context-saving; the other is the correct config for a user slash-command.
6. Retrieval can't replace force-load for a dense trap corpus — proven base-rate wall (History, below). Curation + the agent's own grep-lessons-on-task-start discipline is the lever; an on-demand hook can assist, not replace.
7. Once a harness has a native budget, the forward hide-sweep is largely played out — it then reclaims only bare names. The durable value shifts to relevance-scoping + reading the diagnostics right + NOT over-hiding.