Extract production host lessons into Agentic framework shape

[tnez]

Goal

Extract concrete lessons from production personal-assistant host runtimes into the public Agentic framework shape.

The evidence base now includes long-lived assistant state, tool/capability discovery, auth and human-approval boundaries, artifact iteration, workflow lifecycle, and operational dogfood. Use that evidence to decide what belongs in Agentic versus what remains host/runtime-specific.

Why now

The package and repository are now public under the Agentic identity. Before adding new features, we should turn recent production usage into a clear public framework thesis instead of accumulating one-off implementation details from downstream host applications.

Workstreams to reconcile

• Public identity first: package/repo rename work is complete, so new work can land under Agentic rather than the legacy Spores identity.
• Workflow lifecycle: decide whether feat(workflow): pin runs to graph versions #49 graph-version pinning is a central long-lived-run primitive or a host-specific operational workaround.
• Operational task source: revisit v0.2+: TaskAdapter implementation over GitHub Issues #23 after the extraction frame; stale local dogfood task queues show the pain, but the GitHub TaskAdapter should be sequenced deliberately.
• Composition object: revisit Design: v0.2 composition object (informed by task_next LIFO observation) #16 with newer production-host evidence, not only the older local task-filter observation.
• Capability/tool discovery: fold in the capability-contract milestone and real host tool-discovery lessons to clarify what Agentic declares versus what host runtimes execute.
• Boundary discipline: explicitly document what Agentic should not absorb: sessions, transport, scheduling, identity, webhooks, hosting, and provider-specific auth flows unless a primitive boundary is clear.

Acceptance criteria

• A written extraction memo or issue comment names the Agentic primitives validated by production host usage and the host/runtime responsibilities that stay out.
• feat(workflow): pin runs to graph versions #49, v0.2+: TaskAdapter implementation over GitHub Issues #23, and Design: v0.2 composition object (informed by task_next LIFO observation) #16 are either reactivated with sharper scopes or deliberately left held with rationale.
• New implementation issues are small, independently shippable slices tied to public-safe production evidence.
• README/package positioning can explain Agentic using host-runtime experience as proof without depending on or naming any private downstream product.

Initial sequencing

1. Write the public-safe extraction memo.
2. Reactivate only the slices that survive the memo.
3. File new focused implementation issues where the memo identifies missing primitives or docs.

[dottie-weaver]

Framework boundary memo merged in #88. The doc is now in docs/framework-boundaries.md.

Condensed thesis: Agentic is the primitive layer, not the host runtime. It owns declarations, data shapes, pure matching logic, local filesystem implementations, and CLI dogfood. Host runtimes own durable orchestration, reliability, identity, auth, approval flows, queues, and transport-specific execution.

Provisional decisions from the memo:

• feat(workflow): pin runs to graph versions #49 — Reactivate with sharper scope: graph identity and runtime resolution for long-lived runs. No scheduling, migrations, or deployment management.
• v0.2+: TaskAdapter implementation over GitHub Issues #23 — Do not add a GitHub Issues adapter to core. If useful, it belongs in examples/ or an external integration package. Core can still improve the task adapter interface.
• Design: v0.2 composition object (informed by task_next LIFO observation) #16 — Keep design-only, reframe around Scope: one active hat at a time, persona metadata supplies defaults, adapter-backed primitives query through scoped defaults, identity stays outside. Needs a resolved design comment before implementation.
• Lifecycle events — File a focused issue for lifecycle-event vocabulary (artifact, persona, workflow, memory, capability events). Agentic owns the vocabulary; hook execution belongs to harnesses and hosts.
• README positioning — Update after boundary decisions settle.

Next shippable slices: reactivate #49, rewrite/close #23, update #16, file lifecycle-event issue.

[dottie-weaver]

Follow-up status:

• Framework boundary memo merged (docs: draft framework boundary memo #88)
• Extract production host lessons into Agentic framework shape #87 comment with condensed memo and provisional decisions
• feat(workflow): pin runs to graph versions #49 — Commented with revised scope (reactivate: graph identity + runtime resolution)
• v0.2+: TaskAdapter implementation over GitHub Issues #23 — Closed. GitHub adapter does not belong in core.
• Design: v0.2 composition object (informed by task_next LIFO observation) #16 — Commented with Scope-centered design frame. Stays design-only.
• feat: lifecycle event vocabulary for primitive operations #89 — Filed: lifecycle event vocabulary for primitive operations
• README positioning update (after boundary decisions settle)

[tnez]

Closing this extraction umbrella as complete.

Completed follow-through:

• docs: draft framework boundary memo #88 merged the framework boundary memo into docs/framework-boundaries.md.
• feat(workflow): pin runs to graph versions #49 was re-scoped around graph identity and runtime resolution for long-lived workflow runs.
• v0.2+: TaskAdapter implementation over GitHub Issues #23 was closed because GitHub Issues adapter work does not belong in core Agentic.
• Design: v0.2 composition object (informed by task_next LIFO observation) #16 was reframed around Scope and remains design-only.
• feat: lifecycle event vocabulary for primitive operations #89 was filed for lifecycle event vocabulary.
• docs: reposition README for agentic-as-primitive-layer story #90 merged the README positioning update.

The README is improved, but not final; future README revisions can happen as separate follow-up PRs before additional releases.