Skillscript

Opens v0.17 line. Closes the write-side gap that mirrored v0.16.9's
read-side fix: when an agent calls skill_write via the MCP transport,
the calling agent's identity now reaches the runtime's handler and
stamps SkillMeta.author. Combined with v0.16.9's owner-as-default
dispatch, MCP-authored skills execute under the calling agent's
substrate scope.

Charter: Perry's 3f47b16e Gap-1 finding + warm-adopter's 6ce97894
prototype (validated against live AMP: caller=perry→author=perry
distinct from caller=alice→author=alice through the HTTP/MCP boundary)

• Perry's 2a9c234a 4-point spec ack. Same substrate-author-prototype-in-parallel
  pattern as v0.16.8/9; the prototype's measured shape became the runtime
  spec verbatim.

1. Inbound caller-identity threading (the load-bearing fix)

DashboardServer reads a configured HTTP header (e.g. X-Agent-Id)
on every /rpc request, captures it as McpRequestCtx.callerIdentity,
threads it into McpServer.handle(req, ctx). The skill_write handler
uses it as store({author}). Header lookup is case-insensitive (Node
lowercases inbound header names).

New types:

• McpRequestCtx — request-scoped ctx passed from transport into
  McpServer.handle() and through to tool handlers. Inbound mirror of
  v0.16.9's outbound McpDispatchCtx, one layer up. Re-exported from
  the package index.
• DashboardServerConfig.mcpCallerIdentityHeader?: string — opt-in
  header name. Default unset; existing v0.16.8 behavior preserved.
• McpServerDeps.mcpCallerIdentityHeader?: string — companion field
  for non-DashboardServer embeddings.
• SkillscriptConfig.dashboard.mcpCallerIdentityHeader?: string —
  configurable via skillscript.config.json.

2. Opt-in for multi-agent hosts; invisible to simple-substrate adopters

Header unset = existing v0.16.8 behavior. Single-user CLI, single-tenant
deployments, hobby installs — none configure this surface. The
SkillStore.store() author fallback (e.g. userInfo().username for
FilesystemSkillStore) continues to capture the right author for these
deployments. Multi-agent hosts (where one runtime serves multiple
authenticated agents via MCP) configure the header.

3. Inbound vs outbound X-Agent-Id — two layers, two semantics

Same header name as v0.16.9's outbound HttpMcpConnector.identityHeader
by convention. They are NOT the same value in general:

• Inbound (this ring) = request-scoped caller (current caller of the runtime)
• Outbound (v0.16.9) = dispatch-scoped owner (derived from SkillMeta.author)

They MEET at SkillMeta.author — that's the bridge. Forwarding inbound
straight to outbound would be a setuid hazard (anyone invoking alice's
skill would dispatch as the caller, not alice). The runtime keeps them
separate; outbound is always derived from author at dispatch.

4. Three-test discipline + multi-agent empirical lane

• Runtime unit: McpServer.handle(req, ctx) threads callerIdentity to
  skill_write → store({author}). First-write-locking still wins.
• E2E: DashboardServer reads header end-to-end; case-insensitive lookup;
  absent header → fallback; non-configured runtime ignores the header.
• Multi-agent lane: two callers with distinct X-Agent-Id values →
  distinct stored authors (the test (C) per-host couldn't pass).

5. Docs ring (bundled from v0.16.9)

• docs/adopter-playbook.md HttpMcpConnector section now documents
  identityHeader + maxPoolSize config (v0.16.9 surfaces that lacked
  adoption docs).
• New section forthcoming on the multi-agent inbound-identity setup.
• examples/skillscripts/skill-store-roundtrip.skill.md +
  youtrack-morning-sweep.skill.md — dropped stale version-history
  parentheticals per feedback_playbook_is_timeless.
• examples/connectors/McpConnectorTemplate/McpConnectorTemplate.ts —
  expanded supports_identity_propagation comment to point at the
  Level 1 + Level 2 probe-keying convention so adopters who flip the
  flag don't hit the v0.16.9 auto-coverage gate empty-handed.

Operational note

Existing skills authored via the old write path (e.g., authored as the
runtime's own identity rather than the calling agent's) stay attributed
to the runtime until they're re-written via skill_write under the new
header convention, OR until an adopter-side admin operation transfers
ownership. The first-write-locking discipline still holds: subsequent
overwrites by a different caller don't re-author.