Skillscript

Phase 3 cold-adopter dogfood (AMP-backed MemoryStore + AmpSkillStore vault
migration) wrapped successfully. The cold agent delivered an implementer's-eye
summary surfacing several contract-design questions Perry locked in thread
4b58c283; v0.13.8 ships the narrow set (bug fix + docs + small additive),
v0.14.0 will absorb the contract evolution (Q2 vault op-syntax + Q3
strict_filters).

Added

• memory_read(id, store?) -> PortableMemory | null MCP tool (Perry Q1C).
  Direct memory lookup by substrate-assigned id; mirrors skill_read pattern.
  Returns null on miss (matches MemoryStore's empty-set convention — distinct
  from SkillStore's throw-on-miss; cold agent flagged this asymmetry as
  load-bearing in their "credit where due" summary). No memory_write MCP
  by design — writes are skill-context-only ($ memory_write op inside a
  skill body) to preserve intent-tracking. Reaches via the registered
  MemoryStore; errors cleanly when no store is registered.
• MemoryStore.get(id): Promise<PortableMemory | null> — new contract
  method backing memory_read. SqliteMemoryStore impl is a simple
  SELECT WHERE id = $id. Adopters forking MemoryStoreTemplate add a
  matching impl; null semantics must hold.
• MemoryWrite + MemoryWriteRecord re-exported from
  skillscript-runtime/connectors. Phase 3 cold agent caught v0.13.7's F2
  fix was incomplete — MemoryStoreTemplate.ts imported these two types but
  they weren't re-exported. Template now compiles standalone after fork
  (same as the SkillStore/LocalModel/McpConnector/AgentConnector templates
  did post-v0.13.7).
• README.md Memory row + tool count bumped 15 → 16 (memory_read in its
  own row; matches dogfood-t7's table-vs-runtime drift guard added below).
• docs/connector-contract-reference.md § "Storage-layer conventions"
  (~120 lines). Documents the SkillStore + MemoryStore conventions that live
  in bundled reference impls but aren't in the typed contracts:
  content_hash = sha256(approval-stamped body), version = first 12 hex,
  store() auto-stamps approval token (runtime-semantic leak callout),
  summary/detail split is convention, get(id) returns null vs SkillStore
  load() throws, durability stance, filter-scope discipline. Cold-agent
  framing: "the contracts are shape-portable, not semantics-portable —
  hidden coupling to the reference impl, documented only in its source."
  Now documented.
• docs/adopter-playbook.md § "Notable gaps" — three new bullets covering
  (1) SkillStore vs MemoryStore lifecycle asymmetry as design stance, (2)
  durability-is-implementer's-responsibility, (3) filter-scope advisory-
  today + v0.14.0 strict_filters preview. Includes Phase 3's AMP TTL example
  as cautionary tale.

Fixed

• Doc drift in MCP-surface table (Perry's a8ac822f). README listed 3
  trigger tools; actual is 4. v0.13.3 fixed the total count but missed
  per-row update for Triggers. set_trigger_enabled now in the Triggers
  row.
• docs/ERD.md was silently shipping in the tarball since v0.13.3 — the
  .npmignore was theater (npm ignores .npmignore when files array is
  set in package.json). Fixed via explicit files list (5 user-facing
  docs enumerated, not a docs/*.md glob). .npmignore deleted as
  redundant + misleading.
• tests/dogfood-t7.test.ts test #17 — README-vs-runtime tool-list
  drift guard. Closes the doc-drift class structurally: every tool the
  README claims exists must exist in the runtime, every runtime tool must
  be claimed somewhere in the MCP-surface table. Different class than
  v0.13.3's link-guard (which validates path resolution, not table
  content); same belt-and-braces pattern.

Meta-lesson banked

"Defaults matter more than knobs for security-relevant surfaces"
(Perry, thread 4b58c283). All three Phase 3 contract design questions
laddered to the same axis: Q1 (writes skill-context-only by default — no
memory_write MCP; preserves intent-tracking), Q2 (vault choice explicit
per-write; intent-driven not implicit), Q3 (strict_filters default-on for
filter enforcement). The "we'll let adopters opt in" pattern is seductive
and consistently wrong for scope/visibility/security shapes — aware
adopters opt out, naive adopters get protection. Banked as a feedback
memory for future contract decisions. Sibling to v0.13.7's skill_status
silent-corruption finding: "the contract declares a discipline, but
enforcement is opt-in or absent" is the recurring class.

Note on v0.14.0

Two contract-evolution items locked but deferred:

• Q2 — $ memory_write vault="..." op-syntax kwarg (writes-only, not
  on $ memory query). Real signal from Phase 3 cold agent who hit
  "the op can't express it" for per-call team override.
• Q3 — strict_filters: true default + lint-time filter check +
  UnsupportedFilterError class. Cold agent's "single most valuable fix
  for non-SQLite adopters." Belt-and-braces (runtime + lint) per Perry.

Both are language-grammar / behavior-change scope; bundling together in
v0.14.0 keeps the contract evolution narrative clean.