diff --git a/docs/automation/DEVELOPING.md b/docs/automation/DEVELOPING.md
new file mode 100644
index 000000000..dae5f5da8
--- /dev/null
+++ b/docs/automation/DEVELOPING.md
@@ -0,0 +1,98 @@
+# Developing the MCP bridge
+
+Operator-facing setup lives in [README.md](README.md). This note is for
+contributors iterating on the bridge itself: the stdio JSON-RPC layer in
+`src/mcp.rs` and the operation validators in `src/mcp_ops.rs`.
+
+## Fast iteration: no version bumps needed
+
+- Run the bridge straight from a local build: `target/debug/OpenCADStudio --mcp`
+  for speed, `target/release/OpenCADStudio --mcp` when close to shipping.
+  Both report the same handshake; only the `build_profile` field differs.
+- Tool-surface changes need no version-number change. `serverInfo` carries a
+  content `tool_schema` digest plus the `build_rev` revision hash, so staleness
+  is observable without a release.
+- If you change any tool schema, the `tool_schema_digest_is_pinned` test
+  fails: review the diff, then update `TOOL_SCHEMA_DIGEST` deliberately
+  (never blindly).
+
+## Sessions are per client connection
+
+Clients bind the tool surface once when they connect, so after every
+rebuild, restart the client session (or reconnect its MCP servers) to pick
+up the new schema. The bridge helps: it watches its own executable and, once
+it detects a rebuild, finishes the in-flight request and exits, so the next
+call spawns a fresh bridge. No in-flight call ever fails because of this;
+at worst the following call pays one respawn.
+
+## Session discovery
+
+`ocs_sessions` reads GUI descriptor files and verifies each one is alive
+before probing it: process-liveness checks come first (cached snapshots
+where the OS has no process table to read), dead descriptors are deleted,
+and every remaining TCP probe uses a bounded timeout. Any step that cannot
+run fails open toward treating the session as alive, so a broken check can
+slow discovery but never wedge it.
+
+Concurrent launches are serialized with `automation/starting.lock`
+(`{"pid":..,"started":unix_secs}`, 60s TTL): the first `launch_if_none: true`
+caller claims it and spawns the GUI, later callers with a fresh live claim
+wait for the same descriptor instead of opening another window. Stale claims
+and claims from dead pids are reclaimed; an unusable directory fails open so
+a broken lock can never wedge launching. Covered by
+`startup_lock_serializes_concurrent_launches` (claim, wait, stale-reclaim,
+dead-pid-reclaim, release).
+
+## Cancellation
+
+`run()` reads stdin on a feeder thread into a channel (`CancelPump`); the
+main loop stays the only stdout writer. Wait loops harvest the channel in
+short quanta instead of sleeping blind. A `notifications/cancelled` naming
+the in-flight request dismisses the GUI prompt with a second `{"op":"cancel"}`
+exchange, drops the late result, and sends nothing back. Cancels arriving
+between polls resolve through a pending-operation map; re-polls for dismissed
+operations answer `cancelled` without touching the GUI. Interactive picks
+(`user_select`, `getpoint`) have a ten-minute single-call ceiling ending in
+an explicit timeout error; everything else keeps the one-minute clamp.
+
+## Error routing
+
+Tool-call failures are classified at the source (`CallError`): tool-domain
+failures (bad arguments, failed operations) carry model-actionable guidance
+and stay `isError` results; bridge and GUI infrastructure failures (no live
+session, dead transport) become `-32603`; cancelled requests get no response.
+Resource-not-found follows the negotiated era (`-32602` modern, `-32002`
+legacy). Error results omit `structuredContent`: strict clients validate it
+against `outputSchema`, which only describes success.
+
+## Safety model
+
+Never assume the client prompts before destructive calls. `ocs_execute` is
+marked `destructiveHint: true`, but confirmation behavior is client-defined
+and unobservable from here — server-side validation is the only safety net.
+Validate every argument before touching the drawing, keep destructive and
+read-only paths in separate ops, and prefer reversible effects.
+
+## Rebuilding while the bridge runs
+
+If the linker refuses to replace the binary while a bridge process runs
+from it, move the running binary aside under a new name and rebuild. The
+live bridge keeps serving from the moved file; restart the client
+afterwards so it respawns from the fresh binary, and delete the moved file
+once nothing runs from it anymore.
+
+## Conformance gates
+
+These tests hold the spec line; keep them green and extend them with every
+protocol change: `tool_schema_digest_is_pinned`,
+`initialize_is_spec_pure_with_experimental_build`,
+`malformed_requests_get_jsonrpc_errors`,
+`legacy_list_results_carry_sep2549_ttl`,
+`resource_not_found_code_follows_era`,
+`non_object_input_is_invalid_request`,
+`unknown_tool_is_a_protocol_error_with_names`,
+`cancel_matching_and_wait_deadlines`.
+For end-to-end proof, run the official MCP Inspector CLI against a local
+build (headless-safe methods only: `initialize`, `tools/list`,
+`resources/list`, `resources/read` — anything touching the drawing needs a
+live desktop session and stays in unit tests).