@tnezdev/agentic-runtime-local

Local runtime package for Agentic workspaces.

The public runtime target is local:

agentic serve examples/case-review-bundle --clean --json
agentic run
agentic runtime status local

This package is intentionally thin. It exports the Agentic runtime manifest, creates local runtime glue, and prepares inspectable local workspace runs using Agentic primitives.

Pi can power this runtime under the hood. In this project, Pi means https://pi.dev/, not Raspberry Pi. pi is not a public runtime target.

Initial Scope

• init creates .agentic/runtime/local/runtime.json, targets/, and invocations/.
• serve is the preferred local lifecycle command for authored bundles. It delegates to this package, processes declared surfaces, schedules, hooks, actions, capabilities, handlers, approvals, and writes inspectable local state.
• run resolves a workspace/workflow target, loads the matching persona, skills, ready task, and workflow, records a runtime invocation, creates a workflow run, and writes a finalized local-runtime-run artifact as the durable run packet. It initializes local runtime glue on first run. For non-bundle contexts this is prepare-only by default. If no target is provided, ready task metadata (persona and workflow) is used before falling back to an unambiguous workspace default.
• run --harness pi also invokes the Pi CLI as the local harness driver after the Agentic run packet is prepared. Add --interactive when you want to land in a live Pi terminal session instead of one-shot print mode.
• status reports whether the local runtime glue has been initialized and summarizes the latest invocation.

Invocation records live under .agentic/runtime/local/invocations/<id>.json. Bundle run records live under the bundle state dir, usually .agentic/.data/. Both are runtime-owned execution facts, not transcripts or sessions. Harness integrations attach an optional harness_ref without importing that harness's session model into Agentic core.

Pi Harness Mode

Default integration mode for the first Pi bridge: CLI print mode.

agentic run <workspace-or-workflow> --harness pi

For conversational assistant-style starts, attach the terminal to Pi instead:

agentic run <workspace-or-workflow> --harness pi --interactive

You can also make Pi the default harness for this runtime target with config:

[runtime.local]
harness = "pi"

The local runtime still owns the public command and target name. When Pi is enabled, agentic run:

1. Prepares the same Agentic invocation record, workflow run, and run-packet artifact as prepare-only mode.
2. Writes generated Pi prompt files beside the invocation JSON: .agentic/runtime/local/invocations/<id>.pi-system.md and .agentic/runtime/local/invocations/<id>.pi-user.md.
3. Invokes Pi from the target workspace. Print mode uses pi --print --mode text --session-id <invocation-id> --session-dir .agentic/runtime/local/pi-sessions ...; interactive mode drops --print --mode text and inherits the current terminal stdio.
4. Records harness_ref: { provider: "pi", id: "<invocation-id>" } on the invocation.

Agentic exposes personas, skills, the selected task, workflow definition, and run-packet IDs to Pi through the generated prompt files. Pi owns the model loop, tools, compaction, and session tree. Agentic records only the stable harness reference and durable primitive IDs.

The Pi command can be overridden for local development:

agentic run research-loop --harness pi --pi-command /path/to/pi

or with runtime config:

[runtime.local]
harness = "pi"
pi_command = "/path/to/pi"

Pi Limitations

• Pi must be installed on PATH; the local runtime does not depend on a Pi npm SDK.
• The integration relies on Pi CLI guarantees for --session-id, --session-dir, --name, --append-system-prompt, and @file prompt inputs; print mode also relies on --print and --mode text.
• Agentic does not parse Pi transcripts or continue Pi sessions itself. The invocation id is reused as the Pi session id so the reference is deterministic.
• Pi mode may call a model and use tools. Interactive mode attaches your terminal to Pi until that process exits. Prepare-only mode remains the default so agentic run is safe as an inspectable setup step.

Non-goals for this skeleton: daemon/service mode, scheduling, cloud-provider code, model/tool execution inside Agentic core, and a session primitive.