Integrate AgentTerm with SourceOS Agent Machine workspaces

[mdheller]

Context

AgentTerm should expose the operator-facing shell for SourceOS Agent Machine workspaces.

Target flow:

AgentTerm room -> Agent Registry resolution -> SourceOS Agent Machine profile -> Podman-backed workspace -> AgentPlane evidence.

Scope

Add an Agent Machine adapter and command surface:

/agent-machine doctor
/agent-machine init <profile> <repo>
/agent-machine attach <workspace>
/agent-machine run <tool> <workspace>
/agent-machine fingerprint <workspace>
/agent-machine evidence <workspace>

Requirements

• Preserve Matrix-first event log semantics.
• Resolve non-human participants through Agent Registry before dispatch.
• Treat Codex, Claude Code, Hermes, OpenCLAW/OpenClaw, local shell agents, and GitHub bots as governed participants.
• Record events for profile render, workspace attach, run request, policy decision, execution start, execution complete, evidence produced, and revocation.
• Governed execution should hand off to AgentPlane.
• Direct operator attach should be logged.
• Keep platform-specific Podman logic behind an adapter boundary.

Acceptance criteria

• Docs reference SourceOS Agent Machine.
• Stub commands may land first, but they must record governance-preserving events.
• Example event log includes an agent-machine lifecycle trace.
• No credentials or device-local private material are committed.

Non-goals

• Do not implement the Podman workspace engine here.
• Do not create another agent registry.
• Do not redefine AgentPlane evidence formats.

[mdheller]

Status update 2026-07-18: No dedicated agent-term PR has landed the Agent Machine adapter and command surface (/agent-machine doctor/init/attach/run/fingerprint/evidence). The agent-machine side has advanced significantly (PRs #92-114 in agent-machine), but agent-term has not yet consumed an Agent Machine adapter behind the AgentTermAdapter boundary. This requires coordination between the two repos and stays open.

[mdheller]

Implemented in commits pushed to feat/state-integrity-graph-reasoning-stubs (PR #89, merged).

Issue #18 addressed by:

• tui.py: Textual App — sidebar with all 14 PANE_ORDER panes, event log, approval affordance bar, governed command input line. Wired as agent-term tui. Import-guarded with clear error message if textual not installed.
• agent_machine_adapter.py: 6 ChatOps commands (doctor/init/attach/run/fingerprint/evidence), all producing governed AgentTermEvents. Non-human participants resolved through Agent Registry before dispatch. EXAMPLE_LIFECYCLE_TRACE documents the full init→attach→run→evidence lifecycle.
• cli.py: /agent-machine* wired into interactive shell.