Skillscript

Three small but high-DX fixes surfaced by the cold-adopter Phase 1 dogfood
against shipped v0.15.0. None are runtime correctness bugs — every probe
passed end-to-end. They're UX seams that read as "something's wrong with my
install" to a cold adopter discovering the package for the first time.

skillfile init now seeds the three Phase 1 demos into skills/

Pre-v0.15.1, init copied only hello-world.skill.md and only into
examples/. Adopters wanting to dispatch via execute_skill({name}) had
to manually cp from node_modules/skillscript-runtime/examples/skillscripts/
into $SKILLSCRIPT_HOME/skills/. Now init seeds all three Phase 1
demos directly into skills/:

• hello-world.skill.md — baseline runtime check, no substrate
• skill-store-roundtrip.skill.md — Lisp-shape $ skill_write demo
• data-store-roundtrip.skill.md — $ data_write + $ data_read round-trip

All three ship pre-stamped (# Status: Approved v1:<token>). After
skillfile init, execute_skill({skill_name: "hello-world"}) works
immediately — no manual seeding required.

skillfile init Next: hint points at the canonical MCP/dashboard path

Pre-v0.15.1, init printed:

Next:
  skillfile run examples/skillscripts/hello-world.skill.md

— a file-path execution that bypasses the SkillStore + the v0.9.0 hash-token
approval gate. Cold adopters following the hint ended up evaluating a
different code path than the canonical production flow. Now init prints:

Next:
  skillfile dashboard --host 0.0.0.0 --port 7878
  # Then dispatch a seeded demo via the MCP wire:
  #   curl -s -X POST http://localhost:7878/rpc \
  #     -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"execute_skill","arguments":{"skill_name":"hello-world"}}}' \
  #     | jq -r '.result.content[0].text' | jq

Aligns the hint with the documented production workflow: dashboard launches
the MCP server; in-skill dispatch goes through execute_skill({skill_name}).

node:sqlite ExperimentalWarning suppressed at the CLI entry

(node:XXXX) ExperimentalWarning: SQLite is an experimental feature ...
fired on every CLI invocation that touched sqlite (dashboard, serve,
execute against substrate-resident skills). To a cold adopter it read as
"something's wrong with my install." SqliteDataStore + SqliteSkillStore
are intentional substrate choices; the warning was noise, not signal.

Filtered at process.emitWarning at the CLI entry point — only the specific
SQLite ExperimentalWarning is suppressed; all other warnings pass through
to stderr unchanged. CLI-only — programmatic library consumers see the
warning unchanged so they can choose their own logging discipline.

Tests

• tests/cli.test.ts — two new cases verifying init's post-v0.15.1 shape:
  seeds three demos with valid Approved+token headers; Next: hint matches
  the canonical dashboard + execute_skill pattern (and the old
  skillfile run examples/... shape is regression-guarded against
  reappearance).

Files changed

• src/cli.ts — cmdInit seeds three demos into skills/; rewrites
  Next: hint; adds process.emitWarning filter for the SQLite warning.
• scaffold/skills/{hello-world, skill-store-roundtrip, data-store-roundtrip}.skill.md
  (NEW) — pre-stamped scaffold payload mirroring examples/skillscripts/.
• scaffold/examples/ (REMOVED) — folded into scaffold/skills/ since
  init now seeds the runnable location, not the browsable one.
• tests/cli.test.ts — two new test cases (init seeds + Next: hint).
• tests/dogfood-t7.test.ts — version assertion bumped 0.15.0 → 0.15.1.
• package.json — version 0.15.0 → 0.15.1.
• CHANGELOG.md — this entry.

Cold-adopter Phase 1 collapses

Post-v0.15.1, the Phase 1 cold-adopter briefing loses the manual cp step
entirely. The flow becomes: npm install → skillfile init →
skillfile dashboard → execute_skill({name}). Four commands, no manual
file shuffling, no spurious experimental warnings.