>>> [!note] Migrated issue <!-- Drupal.org comment --> <!-- Migrated from issue #3588759. --> Reported by: [ajv009](https://www.drupal.org/user/3653917) >>> <h3 id="summary-problem-motivation">Problem/Motivation</h3> <p>Every subagent dispatch in a recorded migrate-site session was sequential &mdash; no turn contained more than one Agent tool call. All 54 subagent dispatches ran serially. Measured waste: Phase 3 (component-builder) consumed 86 minutes serial vs. ~14 minutes parallel = 72 minutes wasted; Phase 7 (content-composer) consumed 77 minutes vs. ~20 minutes = 57 minutes wasted. Total: 129 minutes per full migration run lost to avoidable serial execution.</p> <p>The single apparent parallel case was fire-and-forget: an async media-handler dispatched without <code>SendMessage</code>, never consumed by the orchestrator, with no result handling.</p> <p>The fix was applied manually in a later session &mdash; 9 agents ran in parallel, with a background media-handler &mdash; and it worked. But the human orchestrator learned this from the analysis document; SKILL.md was not updated. The Claude model defaults to sequential dispatch unless explicitly instructed otherwise. The next fresh migration run from the unmodified skill will revert to serial dispatch.</p> <h4 id="summary-steps-reproduce">Steps to reproduce</h4> <ol> <li>Run the migrate-site skill end-to-end against any multi-section site.</li> <li>Inspect the session JSONL: search for turns containing more than one <code>"name":"Agent"</code> tool_use block.</li> <li>Observe: every turn has 0 or 1 Agent calls; never 2+.</li> <li>Compare wall-clock times for Phase 3 and Phase 7 against parallel-dispatch projections.</li> </ol> <h3 id="summary-proposed-resolution">Proposed resolution</h3> <p>Add mandatory parallel-dispatch blocks to <code>.claude/skills/migrate-site/SKILL.md</code>:</p> <p><strong>Before Phase 3:</strong></p> <pre>CRITICAL: Phase 3 dispatches are INDEPENDENT. You MUST batch 4-6 Agent tool calls in a SINGLE turn &mdash; NOT one at a time. Every component-builder dispatch in a separate turn adds ~10 min of serial overhead. Dispatch them ALL at once.</pre><p><strong>Before Phase 7:</strong></p> <pre>CRITICAL: Phase 7 dispatches are INDEPENDENT (each page writes to its own JSON file). You MUST batch all content-composer + phase-verifier dispatches in parallel. Process homepage FIRST &mdash; if the session runs out of time, the homepage must be composed.</pre><p>Also add a Phase 5 (media-handler) note: media uploads can dispatch in parallel with Phase 3 component-builder work since they write to different directories.</p> <p>Audit SKILL.md for sequential-looking example patterns (pseudocode showing <code>dispatch agent-A</code> on one line then <code>dispatch agent-B</code> on the next) and convert them to parallel examples.</p> <h3 id="summary-remaining-tasks">Remaining tasks</h3> <ul> <li>Add CRITICAL parallel-dispatch blocks before Phase 3 and Phase 7 in SKILL.md.</li> <li>Add the Phase 5 media-handler parallel note.</li> <li>Audit SKILL.md for any sequential-looking example patterns; convert to parallel.</li> <li>Confirm with project owner that the L630 session-time deferral has expired before writing.</li> <li>Verify on a fresh run: the first Phase 3 turn must contain at least 3 Agent tool calls in a single turn.</li> <li>Measure Phase 3 wall-clock time: should drop from ~86 min to ~20 min on the same component count.</li> </ul> <h3 id="summary-ui-changes">User interface changes</h3> <p>None.</p> <h3 id="summary-api-changes">API changes</h3> <p>None.</p> <h3 id="summary-data-model-changes">Data model changes</h3> <p>None.</p> <h3 id="summary-internal-references">Internal references</h3> <ul> <li>Source markdown: <code>RESEARCH-2026-04-28/99-final/ISSUES/Sprint-1-Foundations/04-P0-4-orchestrator-dispatches-serially.md</code> (id: P0-4)</li> <li>CRITICAL finding: <code>RESEARCH-2026-04-28/00-existing-issues/prior-deep-analysis-key-points.md</code> (54 sequential dispatches; waste table)</li> <li>Workaround evidence: <code>RESEARCH-2026-04-28/06-jamie-demo-old/95q5i-77cf5bb9-half2/what-was-fixed-in-second-session.md</code> (FULLY FIXED in second session via human orchestrator)</li> <li>Theme analysis: <code>RESEARCH-2026-04-28/00-existing-issues/issue-themes.md</code> (Theme 5; L630 deferral)</li> <li>Recommended verbatim text: <code>RESEARCH-2026-04-28/99-final/migrate-site-skill-recommendations.md</code> (mandatory parallel dispatch blocks)</li> <li>Pattern: <code>RESEARCH-2026-04-28/07-cross-cutting/patterns.md</code> &sect; A2 (Ralph Loop Creates Compaction Storms)</li> <li>Anti-pattern: <code>RESEARCH-2026-04-28/07-cross-cutting/session-anti-patterns.md</code> &sect; AP1 (Retry Storm)</li> </ul> <h3 id="summary-open-questions">Open questions</h3> <ul> <li>The user said "Deeper issues to ignore: ... parallelism in skill docs" at session-time L630. Consolidated analysis says this deferral has expired. Confirm before writing.</li> <li>Does current SKILL.md have any parallel-dispatch mention? Determines whether the fix is adding "MUST" emphasis or a new addition.</li> <li>Phase 5 media-handler: currently dispatched before Phase 3 completes, or after? Confirm intended ordering when parallelism is added.</li> </ul>