Skillscript

Agent-as-author hardening. Closes structural gaps surfaced by the v0.7.3
roadmap review (076bdeac, Perry/CC thread, May 26). The language shape
locked at v0.7.2; v0.7.3 hardens the adopter surface so agents (and the
humans wiring them up) have a substrate-neutral, merge-friendly, honestly-
documented surface to build against. $ memory_write is deferred to v0.8.x
bundled with the passthrough auth model — see the auth design thread
43178c86 for the settled framing.

Added — adopter-extensible connector class registry

• registerConnectorClass(name, entry) public API. Adopters with custom
  McpConnector classes call this from their bootstrap before
  loadConnectorsConfig runs. Closes the merge-conflict bait of editing
  the bundled KNOWN_CONNECTOR_CLASSES Map directly. Adopter overrides
  take precedence over bundled set on name collision (lets adopters swap
  bundled classes with hardened variants). Companion: unregisterConnectorClass,
  getConnectorClass. See examples/custom-bootstrap.example.ts.

Added — canonical runtime config

• skillscript.config.json. Externalizes runtime knobs (skillsDir,
  traceDir, memoryDbPath, dashboard port + host, pollIntervalSeconds,
  enableUnsafeShell, mode, ollamaBaseUrl, triggersFilePath,
  connectorsConfigPath) into one declarative file. ${VAR} substitution
  matches connectors.json. Loader is graceful on missing file. CLI's
  dashboard / serve commands accept --config <path>; CLI flags
  override file values; file values override defaults.
• Driver: two-instance posture. Running dev-skillscript + adopter-wiring
  instance on the same machine requires independent ports/paths. The
  config file makes this a copy-and-tweak operation rather than threading
  CLI flags.
• skillscript.config.json.example ships at repo root as the template.

Added — onboarding scaffold (examples/onboarding-scaffold/)

• Complete adopter deployment with file-backed memory + OpenAI LLM +
  tmux-shell agent delivery. ~200 LOC across three adapter files plus
  bootstrap. Case-1 typed-contract wiring end-to-end — skills authored
  against this scaffold use canonical $ llm / $ memory and run
  unchanged against any other Case-1 substrate.
• file-memory-store.ts — MemoryStore impl over a JSON file with
  simple substring FTS
• openai-local-model.ts — LocalModel impl over OpenAI Chat
  Completions API
• tmux-shell-agent-connector.ts — AgentConnector impl via tmux send-keys (mirrors what nanoclaw-style harnesses do internally)
• bootstrap.ts — wiring example tying all three together with the
  v0.7.2 bridges
• README.md walking through quick-start, two-instance posture, and
  what to modify for production

Added — adopter playbook (docs/adopter-playbook.md)

• Case-1 typed-contract vs Case-2 MCP-tools wiring tradeoff (the
  load-bearing decision)
• Joe-Programmer setup walkthrough
• Conventions for upstream-merge-friendly modifications (dedicated
  adopter files, // ADOPTER:org — sentinels, public registration APIs)
• Substrate ship-status table with honest v0.7.x gap callouts
• Skill discovery + cross-agent composition patterns under Case-1 memory

Changed — OutputKind cleanup (substrate-neutrality)

• Dropped slack and card from OutputKind. Substrate-specific
  delivery names in what's supposed to be a substrate-neutral language.
  Same anti-pattern v0.7.0 removed for LocalModel and MemoryStore
  dispatch shapes. OutputKind now: text / prompt-context: <agent> /
  template: <agent> / file: <path> / none. Adopters wanting
  Slack / WhatsApp / Discord / etc. use either $ slack.post ... MCP
  dispatch inside the skill body OR # Output: prompt-context: <agent>
  letting the receiving agent decide. The bundled EmissionConnector
  v1.x backlog item is also dropped — MCP-dispatch handles the use case.
• Substrate-neutrality sweep of the language surface (parser
  enums, lint code names, ambient refs, frontmatter values, help
  content). Clean post-cleanup; no other substrate-specific leaks
  found in the language contract. Internal AST field ampParams
  flagged for renaming in a future pass (not user-facing, no
  contract impact).

Changed — reference bootstrap framing

• src/bootstrap.ts reframed as reference wiring, not canonical.
  File header + bootstrap() docstring explicitly direct adopters with
  custom substrates to write their own bootstrap importing the public
  APIs. bootstrap() remains part of the v0.7.x+ stable public surface
  for default deployments. See examples/custom-bootstrap.example.ts
  • examples/onboarding-scaffold/bootstrap.ts.

Changed — public exports

• New top-level exports from skillscript-runtime:
  loadSkillscriptConfig, bootstrap, defaultRegistry,
  wireDeclarativeTriggers, registerConnectorClass,
  unregisterConnectorClass, getConnectorClass,
  listKnownConnectorClasses, loadConnectorsConfig.
• New skillscript-runtime/connectors exports:
  LocalModelMcpConnector, MemoryStoreMcpConnector (so adopter
  bootstraps can wire bridges over their own typed-contract impls
  without reaching into bundled bootstrap()).

Deferred to v0.8.x

• $ memory_write with the MemoryStore.write({content, tags?, recipients?, expires_at?, metadata?}) contract extension. Ships
  bundled with the passthrough auth model so the credibility-crack of
  documented-but-unshipped delivery closes alongside the auth design.
• Authorization model. Mixed: passthrough for op-level mutations
  (substrate enforces; runtime threads credentials); runtime-enforced
  for skill-promotion (# Status: Draft → Approved requires non-author
  signer). # Autonomous: true becomes documentation; unconfirmed-mutation
  lint drops (substrate enforcement makes it theater). New-version-of-
  same-name returns to Draft on content change. See thread 43178c86
  for the settled design.

Deferred (post-v0.7.3 ship)

• R7 cold-author harness. Tests NFR-6 agent-modifiability empirically:
  cold agent adds a new filter, a new lint rule, and a new runtime-
  intrinsic op against the v0.7.3 file layout. If they succeed, NFR-6 is
  satisfied empirically; if they fail, the failure mode targets which
  extension point to refactor first. Evidence-driven, no major-version
  refactor slot.

Notes for cold authors

• Run help({topic: "frontmatter"}) for the updated 5-kind OutputKind set.
• Adopters integrating new substrates: see docs/adopter-playbook.md
  for the Case-1 vs Case-2 framing — this is the most important
  decision in your wiring.
• ${VAR} interpolation, the canonical $ llm / $ memory surfaces,
  bare-form bridge dispatch, the v0.7.2 typed contracts — all unchanged
  in v0.7.3. Existing skills run as-is unless they used # Output: slack:
  or # Output: card: (rewrite to prompt-context: or use MCP dispatch).