fix(workflow): use Skill instead of Task for auto-advance phase transitions

[Tibsfox]

Summary

The auto-advance chain (discuss → plan → execute triggered by --auto) was spawning each subsequent phase as a Task(subagent_type="general-purpose"), creating 3-4 levels of nested agent sessions. At that depth, the Claude Code runtime hits resource limits or stdio contention, causing the execute-phase to either freeze indefinitely or attempt to shell out to claude as a subprocess — which is explicitly blocked with "Claude Code cannot be launched inside another Claude Code session."

The nesting looks like this:

discuss-phase (level 0)
  └─ Task(general-purpose) → plan-phase (level 1)
       └─ Task(general-purpose) → execute-phase (level 2)
            └─ Task(gsd-executor) → executor agents (level 3)
                 └─ ...potential further nesting (level 4)

By level 2-3, the general-purpose agent loses context about the execute-phase orchestration protocol and falls back to shelling out to claude --print, which triggers the nested session error.

The Fix

Replace Task(general-purpose) with Skill() invocations for phase transitions:

discuss-phase (level 0)
  └─ Skill("gsd:plan-phase") → plan-phase (level 0, same context)
       └─ Skill("gsd:execute-phase") → execute-phase (level 0, same context)
            └─ Task(gsd-executor) → executor agents (level 1)

The Skill tool runs in the same process context as the caller — it doesn't create a subprocess or new agent session. This keeps the entire orchestration chain flat at level 0, with only the actual worker agents (gsd-executor, gsd-verifier, etc.) spawning as Task subagents at level 1.

Changes

• workflows/plan-phase.md (step 14): Replace 30-line Task(general-purpose) block with Skill("gsd:execute-phase", args="${PHASE} --auto --no-transition")
• workflows/discuss-phase.md (auto_advance step): Replace 30-line Task(general-purpose) block with Skill("gsd:plan-phase", args="${PHASE} --auto")

Net effect: -67 lines of complex Task prompt engineering, +13 lines of clean Skill invocations.

Verification

1. Run /gsd:discuss-phase <phase> --auto
2. Complete the discussion flow
3. Auto-advance should trigger plan-phase (via Skill, not Task)
4. Plan-phase should auto-advance to execute-phase (via Skill, not Task)
5. Execute-phase should spawn gsd-executor agents normally
6. No "nested session" errors, no freezes

Why Skill Instead of Task?

The previous code had a comment: "do NOT use Skill tool — Skills don't resolve inside Task subagents". This was correct for the original architecture where each phase ran inside a Task. But the fix here changes the architecture: phases now chain via Skill at the top level, so the resolution concern doesn't apply. Each Skill invocation runs with full access to the command registry.

Closes #686

🤖 Generated with Claude Code