Skillscript

First connector contract audit + lock for v1.0 per Perry's thread b722bbf4.
Closes Q1-Q12 across three pressure-test rounds (sender/substrate lens →
agent-as-receiver lens → full code-audit reconciliation). Opens the
"connector house in order" gate-3 sequence; remaining contracts (McpConnector,
SkillStore, MemoryStore, LocalModel) audit + lock in subsequent v0.9.x slots.

Contract changes (src/connectors/agent.ts)

• Added DeliveryMeta envelope on both DeliveryPayload variants (Q8).
  Runtime auto-fills dispatch_id (UUID per emit), sent_at (emit-clock),
  origin.skill_name / origin.trigger_kind (required); optional
  origin.entry_skill_name, origin.caller_agent_id, event_type,
  correlation_id populate when context provides them.
• Added health_check() required method (Q6). Bootstrap-throws on false
  via registerAgentConnector() — wiring failures surface at boot, not at
  first skill-fire.
• Added request_response() required method (Q1) with locked Response
  shape { correlation_id, content, sent_at, agent_id }. Impl deferred to
  v0.10 when exchange() op ships; adopters throw NotImplementedError
  until then.
• Added delivery_skipped?: boolean to DeliveryReceipt (Q7). Contract-
  level signal for "accepted but not pushed" (offline agent, rate-limit drop,
  etc.); runtime honors connector-set value, preserves NoOp fallback inference.
• Dropped manifest() (Q2). No production callers; folded into
  conformance test surface.
• Dropped TriggerProvenance interface (Q12) and its envelope fields
  (source_skill, triggered_by, delivery_context, templates, format).
  Each folded into meta per Q8-Q11 or dropped to trace-only surface.

Syntax changes

• # Delivery-context: → # Event-type: frontmatter rename (Q9). Vocab
  consistency between skill-author and receiver-agent surfaces; the field
  flows to meta.event_type as the frontmatter fallback (notify(event_type=...)
  kwarg takes precedence per-emit).
• notify() op gains event_type= and correlation_id= kwargs. Per-emit
  override of frontmatter; correlation_id is the reply-correlation primitive
  for the v0.10 exchange() shape.

Behavior changes

• Registry.registerAgentConnector() is now async — invokes health_check().
• triggerCtx.source enum migrated from {cron, session, event, agent-event, file-watch, sensor, manual} to Q8's {cron, session, webhook, agent, cli, dashboard, inline}. The pre-v0.9.6 "extra" values were parse-only (no production firing path emitted them). manual → inline is the only meaningful migration.
• # Templates: frontmatter still parsed for unknown-template-reference lint
  but no longer flows through DeliveryPayload (Q10 — vestigial removal).

Deliverables

• New docs/connector-contract-reference.md — canonical reference for adopters
  (audience: adopter agents). Includes the four footnotes Perry pinned during
  the audit (broadcast dispatch_id semantics, deeper-than-2-level chain elision,
  caller_agent_id general rule, sent_at vs delivered_at distinction).
• Updated docs/adopter-playbook.md AgentConnector method list.
• Updated docs/language-reference.md AgentConnector + DeliveryPayload sections.
• New tests/v0.9.6-agent-connector-audit.test.ts (15 tests covering Q1-Q12);
  includes Perry's plumbing-risk SHAPE test for entry_skill_name propagation
  (cites lesson 1bc9d7a2).

Pre-adoption migration

Pre-adoption rule applies (no external users); breaking changes are cheap.
Adopters who experimented against pre-v0.9.6 code:

• Drop manifest() from AgentConnector impls.
• Add health_check() + request_response() (throw NotImplementedError until v0.10).
• Replace source_skill / triggered_by / delivery_context / templates /
  format payload field reads with meta.origin.* / meta.event_type reads.
• Rename # Delivery-context: → # Event-type: in skill source.

Methodology lesson (per Perry, worth banking)

"Contract audits compound. Each pressure-test pass before lock has near-zero
cost; each pass after adoption has near-infinite cost. Spend pre-lock budget
liberally." Three pressure-test rounds (Q1-Q7 → Q8 → Q9-Q12 audit-reconciliation)
each caught real gaps the previous round missed. The pre-adoption rule made
the discipline cheap; the audit-as-methodology made the pass-count load-bearing.

LOC ceiling

Bumped narrow-core from 8300 → 8500. Third bump in the v0.9.x series — per §18
note, the "consolidate first" threshold. Most of the ~150 net LOC growth is
doc-comments on agent.ts; per the adopter-agent-as-author memory,
reference impls are the dominant signal source for adopter agents, so
docstring richness is load-bearing.