v2.18.0

v2.18.0 — 2026-06-27

Changed — ~48% fewer tokens loaded per memex:run (behavior-preserving)

A token-footprint optimization across the three tiers that enter a context
window, with no change to what any skill does (one bounded, documented exception
below). Measured with a fixed tokenizer:

• Always-loaded routing skill −47.7% (skills/run/SKILL.md, paid on every
  memex:run): the Step 0 cold path — platform-specific Python-install blocks,
  the plugin-root discovery cascade, and the bootstrap consent/error prompts —
  moves to a new skills/run/STEP0.md, read on demand only when a preflight
  check actually fails. On a healthy install it is never loaded. The happy-path
  checks (Python probe, config.json plugin-root read, five-path existence) and
  the failure-branch semantics stay in the always-loaded skill; every verbatim
  install/bootstrap/error string survives byte-for-byte in STEP0.md.
• Per-operation cost −35% to −41%: per-procedure prose trims, and the
  per-tier model-rationale blockquotes collapse to one-line pointers (the
  canonical ENFORCED table already lives in CLAUDE.md). The repeated
  EmbeddingUnavailable try/except is centralized into
  embeddings.encode_or_skip().
• Per-dispatch −30% on the two LLM-dispatched agent profiles, plus compacted
  in-prompt JSON (separators=(",", ":")) in the Librarian and Community-Reporter
  prompts.

Changed — bounded large-document recall trade (the one behavior delta)

The Librarian classification window (build_prompt char_budget) drops 80000 →
40000 chars and the synthesize source budget drops 50000 → 32000. Documents
longer than the window are classified/synthesized on a head window — already
surfaced honestly via the existing body_truncated / truncated markers; the
citation graph is unaffected (the synthesizes / relation edges are added
independently of the prompt body). Anti-revert test ceilings were tightened in
lockstep.

Internal

• tests/test_skill_run_preflight.py repointed to assert the moved Step 0
  verbatim blocks against STEP0.md (the verbatim-survival guard follows the
  text); added tests/test_embeddings.py coverage for encode_or_skip.
• Deterministic-agent seed profiles (Archivist, DBA, Data Steward) trimmed for
  symmetry with the Librarian / Reference-Librarian profiles.
• Shipped via a 3-cycle kaizen run (#15); no model: tier line, the M3
  single-write-path, or any untrusted-input guard was altered.