salitronic/eda-agent/autodesign

Autonomous PCB design (eda-agent)

Drive a full spec-to-board design by looping the server-side state machine. The server owns sequencing, gates, and durable state, so you never memorize the workflow: you call design_next_action, do what it says, log the result, and repeat. A weaker model produces a plainer board, never a broken pipeline, because the integrity lives server-side.

Before you start

• Confirm the editor is actually answering, with app_ping on Altium or

easyeda_ping on EasyEDA. Ping, not app_get_status: status reports that the process exists and that something once called attach, and neither of those proves the bridge replies. If it does not answer, tell the user how to start it; don't guess.

• Read design_get_discipline once: the hard rules and the DesignPlan

schema. Or call design_autonomy_guide for the full protocol + the 13 stages with their tools and exit gates.

The loop

1. design_session_start(requirement) opens the durable journal. Keep the

returned session_id; every later call takes it. Then design_live_view so the user can watch the board change, and record each decision with its reason before acting on it (design_live_note).

1. If a project is open or will be modified, checkpoint first so the whole

run is revertible in one step: app_checkpoint("before autonomous run") on Altium, easyeda_checkpoint on EasyEDA.

1. Loop: design_next_action(session_id) and act on status:

• proceed / retry: do the stage using its suggested_tools until the

exit_gate is met, then design_session_log(event="stage_result", stage=<stage>, status="ok"). Where the gate names numbers (placement, routing, pours, verification), measure them with pcb_layout_audit and pass the result as data=: the harness checks them and sends the stage back if one fails. If you cannot finish without the user, log status="blocked" with a question and stop.

• blocked: put open_question to the user; when answered,

design_session_log(event="resolved", text=<answer>) and continue.

• complete: the 13 stages are done; write the design report with

design_session_report and review outputs with the user.

1. Checkpoint again before each high-risk mutating stage: sch_to_pcb,

routing, pours_tuning.

1. Long engine runs (routing a dense board) can exceed the tool timeout;

start them with design_job_start and poll design_job_status / design_job_result.

Bounded retries: a stage that fails 3 times escalates to a human question automatically. Don't loop past it; surface it.

How the board is laid out

The discipline's "Board layout method" is the standard: floorplan before placement (fixed skeleton, critical corridors, sections, one rectangle per sub-block in signal-flow order), each block laid out to a pattern, the legaliser's moves read rather than accepted, routing in stages (pairs, plane fan-out with dog-bones, buses as a whole, then the rest), and one small checked step at a time. A scattered placement means the method is wrong; no parameter fixes it.

Resuming

A run survives context loss. In a fresh session, call design_session_resume(session_id) (or design_next_action) and pick up from recorded state: the journal, not the chat history, is the source of truth.

Hard constraints (non-negotiable)

• Datasheet-first: every device fact fetched and cited from the

manufacturer datasheet; never fabricated. Use WebSearch/WebFetch.

• NDA isolation: never mine or reference other client designs.
• No third-party routing engines or account-gated APIs in the design

loop: the in-house router is the only routing engine.

• Verify render-and-look, not by score alone; the visual rubric is the

shipping bar.

• No unverifiable safety tables: ship only sourced/verified values.

Discovering tools

The surface is several hundred tools. Use tool_catalog(category=, maturity=, interaction=, query=) to find the right one without loading every schema; interaction="modal" and "partial" flag tools that need a human or leave work incomplete. Run a discovered tool by name with tool_invoke.