Skip to content

Agent Plane

Peter Pak edited this page Jul 26, 2026 · 8 revisions

Agent Plane

One MCP tool surface, four harnesses, everything recorded. This is the instrument of the thesis: because the tools, database, and printer are identical across harnesses, any difference in behavior is attributable to the harness/model. Design overview: Architecture §B.

The MCP server (services/plugins/mcp/)

TypeScript MCP server, npm @ppak10/agentic-sls-mcp. 21 tools in two groups; full ground rules in services/plugins/AGENTS.md.

Control (live printer, via Inova API Plugin on :5001):

Tool Effect
printer_status Read-only: is_printing + job/phase/progress + recording build id (via the recorder), layer progress, temperatures, power, positions, override state
recoater_passes_get/set Staged powder delivery within one layer (see Hardware Reference — not N full recoats)
recoater_full_passes_get/set Expand each layer into N full recoat cycles (optionally dry)
layer_overrides_get/set Runtime knob registry: recoater speeds, shake, powder dosing, Z clearance, delays
job_list / job_get Stored jobs (queued layouts, not print history): metadata, print profile, nesting instances; works while idle
job_set Rename a stored job / change its print profile — metadata-only, inert; refuses [TEMPLATE] jobs
job_create_from_template Clone an operator-blessed [TEMPLATE] job with a different print profile (2026-07-18); the clone is never started by the agent
profile_list / profile_get Live print profiles; merged=true returns effective values (profile names lie — read fields)
profile_set Upsert a profile (create from base / partial update); bounds enforced in schema + plugin (the firmware model has NO validation of its own); system Default refused

Knowledge (recorder Postgres, forced read-only pool):

Tool Effect
builds_list / build_get Build catalog and per-build card (linkage, profile params, telemetry summary, ASTM outcomes)
astm_query D638 tensile / D790 flex results by batch/profile/material/specimen
telemetry_summary Thermal tracking error, recoat timeline, laser duty for one build
db_query Single read-only SELECT escape hatch
reference_list/get Curated corpus in the Agentic-SLS-Knowledge dataset submodule (source/): TDS values, Fuse 1 ↔ Inova translation, profile conventions — the submodule pin records which corpus version agents saw

Every *_set call is logged server-side to agent_actions at the MCP choke point — this is how interventions become queryable evidence (and how reflections avoid trusting agent self-reports). It is expected behavior, not something to bypass.

The four harnesses

Claude Code, OpenCode, Codex, Antigravity (agy) — all verified against the same tools. Packaging is one core + thin manifests per harness (Claude plugin marketplace, .opencode/plugins/, Codex config.toml snippet, gemini-extension.json).

Harness quirks worth knowing:

  • codex exec needs --dangerously-bypass-approvals-and-sandbox for MCP (openai/codex#16685).
  • agy reads MCP config only from ~/.gemini/config/mcp_config.json.
  • Claude's built-in Bash/Edit/Write aren't restricted by --allowedTools (worth remembering if tool gating ever returns).
  • Evaluate how a harness answered, not just the answer — agy once improvised its own MCP client script and still got the right number.

Presets — removed

The analyst/operator preset split was removed 2026-07-18: every chat gets the full MCP tool surface (knowledge tools are read-only by construction; *_set calls are logged to agent_actions). Reintroduce deliberately if tool gating is ever needed again.

Runner, broker, GUI

  • services/harness/run.ts — uniform headless driver: npx tsx services/harness/run.ts <harness> --prompt '…'.
  • services/harness/server.ts — chat broker on :3100: SSE turns, session resume. Interactive chats are keyed by conversation id (POST /agent/conversations/:id/messages) — the broker rehydrates its in-memory chat from the conversation row after a restart (harness, cli_session_id, turn counter), so GUI conversations survive broker cycles.
  • services/web/ (:5173, docker compose) — Builds registry pages, conversation pages (chat-style transcript rendering: markdown assistant turns, collapsed tool rows, right-aligned user bubbles; sidebar recents + interactive chat, /conversations/new draft creates nothing until first send), Mission Control PanelConsole. (The right-docked chat panel and the Agents session-browser page are both gone — 2026-07-16 / 2026-07-18.)
  • agent_conversations.debug — dev/test flag, settable at creation or via the header badge / PATCH /agent/conversations/:id (cascades to panel candidates). Debug conversations are excluded from the Conversations dataset export; everything created before 2026-07-18 is flagged debug (tooling-era test data).

The profile-experiment workflow (2026-07-18)

The conversation-driven loop for testing a new print profile: research prior builds (astm_query / build_get / telemetry_summary) → read effective values from a live profile (profile_get merged=true) → create a tuned variant (profile_set, bounds-enforced) → stage it on an operator-blessed [TEMPLATE] job (job_create_from_template) → the inert job appears in the GUI Jobs page (3D preview works while idle) → the operator reviews and starts the print. Agents never start prints, never modify templates, never touch the system Default profile, and every mutation lands in agent_actions.

Everything is recorded

Every session/turn lands twice: normalized in Postgres (agent_conversations → agent_sessions → agent_messages; one parser for all four harnesses in services/harness/store.ts) and raw in the Conversations dataset repo (datasets/Agentic-SLS-Conversations/source/sessions/ — commit it often). Failed runs are data too.

Session rows carry (harness, model, role, build). The comparison design: harnesses bundle models, so OpenCode×Claude vs Claude Code×Claude isolates the harness variable, and OpenCode×{models} isolates the model variable — OpenCode being the strongest local-model harness.


See also: Roadmap for the learning loop that consumes all of this

Clone this wiki locally