fix(mcp ux): even more error handling

who asked blud to conventional commits

docs/durable_run_operations.md

@@ -52,6 +52,7 @@ The compact response includes:
 - `run_id`: durable handle for this execution attempt
 - `status`: runtime status such as `completed`, `failed`, or `interrupted`
 - `outcome`: terminal workflow outcome when available
+- `error`: failed-run error text when execution failed before a terminal outcome
 - `output`: projected workflow output when available
 - `diagnostics`: dependency/runtime diagnostics
 - `trace_count`: total trace entry count
@@ -70,8 +71,8 @@ Reads one stopped run by `run_id` without returning trace entries:
 ```
 
 Use this when a client already has a `run_id` and needs the current durable
-summary: status, outcome/output if available, diagnostics, and checkpoint
-metadata.
+summary: status, outcome/output if available, failed-run error text,
+diagnostics, and checkpoint metadata.
 
 ## `read_run_trace`
 
@@ -129,6 +130,8 @@ the core rule that external callers resume by `run_id`.
 - Prefer `inspect_run` before reading trace detail.
 - Use `read_run_trace` with explicit small ranges.
 - Treat `trace_count` as metadata, not an instruction to fetch the entire trace.
+- If a run failed with `trace_count: 0`, read the top-level `error` first; the
+  failure may have happened before any trace entry could be emitted.
 - If `resume_run` is blocked, repair the reported dependency issue and retry
   with the same `run_id`.
 - If the run failed because a live source errored during execution, do not retry
@@ -141,4 +144,3 @@ the core rule that external callers resume by `run_id`.
 - No protocol-native MCP Tasks integration yet.
 - No dynamic saved-workflow-as-tool projection requirement.
 - No automatic pause on disconnected sources; source failures are failed runs.
-

docs/wf_mcp_operator_manual.md

@@ -282,7 +282,8 @@ top-level input/output field names, while full JSON schemas stay behind
 `wf.workflow.call_capability` is the REPL-style test step. Its result is
 self-describing: `kind` is either `node_spec` or `wrapper_artifact`,
 `source_id` identifies the owner when applicable, and `diagnostics` is empty for
-successful calls.
+successful calls. Failed test calls return a structured diagnostic with
+`outcome="runtime_error"` instead of leaking raw transport exceptions.
 
 ### 4. Manage Saved Workflows
 
@@ -354,9 +355,10 @@ brand-new MCP tools. Many LLM harnesses do not reliably refresh callable tool
 schemas mid-session.
 
 The default `run_deployment` response is intentionally compact. It includes run
-status, output, diagnostics, and `trace_count`, where `trace_count` is the total
-number of trace entries in the original run. If the caller needs node-level
-debug detail, pass an explicit `trace_range` object such as
+status, terminal outcome when available, failed-run `error` text when available,
+output, diagnostics, and `trace_count`, where `trace_count` is the total number
+of trace entries in the original run. If the caller needs node-level debug
+detail, pass an explicit `trace_range` object such as
 `{"start": 0, "limit": 10}`; otherwise trace entries stay out of the normal
 response. Trace entries may include resolved node inputs, node outputs, and
 state changes, so treat them as debug payloads rather than ordinary list/summary

docs/workflow_drafts.md

@@ -600,11 +600,12 @@ Workspaces are mutable and revisioned. Artifacts are immutable and versioned.
 Patch calls must include the current `revision`; stale revisions return
 `revision_conflict` and do not mutate the workspace.
 
-`create_minimal_draft_workspace` is intentionally only a bootstrapper. It wires
-an `error` outcome for naive MCP wrappers only when `error_message_source` is
-provided or a state path can be derived from canonical `output` bindings or the
-compatibility `output_map`. Provider-specific
-error envelopes still belong in saved wrapper artifacts or follow-up patches.
+`create_minimal_draft_workspace` is intentionally only a bootstrapper. For
+naive MCP wrappers with an `error` outcome, it wires `wf.std.runtime_error` with
+a static default message unless `error_message_source` is explicitly provided.
+It does not guess that a normal output state path is also an error message.
+Provider-specific error envelopes still belong in saved wrapper artifacts or
+follow-up patches.
 
 In MCP Inspector, workspace mutation tools accept a single `request` object.
 This is deliberate: the request object carries descriptions and validation for