base-action: persist execution log when SDK throws (e.g. max-turns)

[STRd6]

Problem

When the Agent SDK's query() iterator throws — most commonly on --max-turns exhaustion — the partial transcript is silently lost.

base-action/src/run-claude-sdk.ts:

try {
  for await (const message of query({ prompt, options: sdkOptions })) {
    messages.push(message);
    // ...
  }
} catch (error) {
  console.error("SDK execution error:", error);
  throw new Error(`SDK execution error: ${error}`);  // re-throws here
}

// never reached on SDK error:
try {
  await writeFile(EXECUTION_FILE, JSON.stringify(messages, null, 2));
  result.executionFile = EXECUTION_FILE;
} catch (error) { ... }

The rethrow bubbles to index.ts, whose catch calls core.setFailed and sets only conclusion — execution_file and session_id outputs are never set. So:

• $RUNNER_TEMP/claude-execution-output.json is never written
• steps.<id>.outputs.execution_file is empty
• any downstream step gated on that output (e.g. an S3 upload of the transcript) silently skips
• the partial transcript is unrecoverable

Repro

A workflow that hits --max-turns N fails with:

SDK execution error: Error: Claude Code returned an error result: Reached maximum number of turns (N)

Exit code 1, no execution_file output, no log persisted — even though 1000 turns' worth of data was in memory moments earlier.

Fix

Capture the SDK error, write the transcript, publish the outputs, then re-throw so the step still fails. Outputs are now set from inside the SDK runner because index.ts re-throws before reaching its own setOutput calls on the failure path.

On the success path the behavior is unchanged: the outputs get set in run-claude-sdk.ts and again in index.ts with identical values (last-write-wins, no observable difference).

Why this matters

Max-turns failures are precisely the runs where the transcript is most valuable — long, expensive runs that the user wants to inspect and recover partial work from. Losing them silently is a bad default.