fix(sdd): ensure sub-agent responses are not lost when ending with tool calls

internal/assets/generic/sdd-orchestrator.md

@@ -29,6 +29,16 @@ Anti-patterns — these ALWAYS inflate context without need:
 - Running tests or builds inline → delegate
 - Reading files as preparation for edits, then editing → delegate the whole thing together
 
+### Empty Sub-Agent Response Recovery
+
+When a sub-agent returns empty, minimal, or tool-result-only output (e.g., "Observation saved", a raw tool response, or blank):
+
+1. Do NOT launch a new sub-agent — this wastes the completed work and full context.
+2. RESUME the same task using the existing `task_id`, adding a prompt: "Your work completed but I received no text response. Please provide your findings/summary as text now."
+3. If the second attempt also returns empty, THEN report the issue to the user with the task ID and last output received.
+
+**Why**: Sub-agents sometimes end with a tool call (e.g., `mem_save`) as their last action, causing the orchestrator to receive only the tool result instead of the analysis. Resuming preserves the sub-agent's full context; re-launching discards all work done.
+
 ## SDD Workflow (Spec-Driven Development)
 
 SDD is the structured planning layer for substantial changes.

internal/assets/skills/_shared/persistence-contract.md

@@ -131,6 +131,14 @@ After completing your work, you MUST call:
 If you return without calling mem_save, the next phase CANNOT find your artifact and the pipeline BREAKS.
 ```
 
+## Sub-Agent Response Ordering
+
+When a sub-agent persists artifacts (via `mem_save` or file writes), the persistence call MUST happen BEFORE the final text response. The sub-agent's absolute last output must be text, never a tool call.
+
+**Why**: The parent agent only receives the sub-agent's final output. If the last action is a tool call (e.g., `mem_save`), the parent sees the tool result ("Observation saved") instead of the sub-agent's actual analysis and findings.
+
+Order: do work → persist artifact (`mem_save`) → return text summary as final action.
+
 ## Skill Registry
 
 The orchestrator pre-resolves compact rules from the skill registry and injects them as `## Project Standards (auto-resolved)` in your launch prompt. Sub-agents do NOT read the registry or individual SKILL.md files — rules arrive pre-digested.

internal/assets/skills/_shared/sdd-phase-common.md

@@ -66,6 +66,8 @@ Return result inline only. Do not write any files or call `mem_save`.
 
 ## D. Return Envelope
 
+> **CRITICAL — Response ordering**: Your FINAL output MUST be text (the return envelope), NOT a tool call. If you need to save to Engram (`mem_save`), do it BEFORE your final text response. Do NOT call `mem_session_summary` — that's for top-level agents only. **Why**: When a sub-agent's last action is a tool call, the parent agent receives only the tool result — your text response (the actual analysis) is lost.
+
 Every phase MUST return a structured envelope to the orchestrator:
 
 - `status`: `success`, `partial`, or `blocked`