autoplan: "Run them simultaneously" causes Agent tool subagents to be abandoned

[ConejoCapital]

Bug Description

The /autoplan dual-voice feature instructs Claude to run the Agent tool (subagent) and Bash (Codex) simultaneously, with "All non-blocking" error handling. This causes every autoplan invocation to silently degrade to single-reviewer mode.

Root Cause

Two independent failure modes:

1. Permission Isolation (Critical)

Background subagents dispatched with run_in_background: true do NOT inherit the parent session's tool permissions. Both Read and Bash tools are denied. The subagent literally cannot read files or run commands — it's dead on arrival.

Tested output from background subagent:
```
status: completed
result: "I cannot complete this task. Both the Read and Bash tools have
been denied permission, preventing me from accessing the file..."
tool_uses: 2
duration_ms: 5898
```

2. Timing Race

Even if permissions worked, background task output arrives as a `` after the blocking Bash call returns. Claude has already moved on to present the consensus table, so subagent output is orphaned.

Tested foreground subagent (same prompt — works correctly):
```
Result: "SUBAGENT_COMPLETE: Line 198 specifies that dual voice reviews
(Claude subagent and Codex) must execute simultaneously..."
tool_uses: 1
duration_ms: 5857
```

Affected Lines

`autoplan/SKILL.md.tmpl` (source of truth):

• Line 198: `Run them simultaneously (Agent tool for subagent, Bash for Codex).`
• Line 220: `Error handling: All non-blocking.`
• Line 241: `Step 0.5 (Dual Voices): Run Claude subagent AND Codex simultaneously.`
• Line 341: `Step 0.5 (Dual Voices): Run Claude subagent AND Codex simultaneously.`
• Line 409: `Step 0.5 (Dual Voices): Run Claude subagent AND Codex simultaneously.`

Same text propagates to generated `autoplan/SKILL.md` at lines 541, 563, 584, 684, 752.

Reproduction

1. Run `/autoplan` on any plan file
2. Observe Claude launches a background Task and a Bash call in parallel
3. Background Task fails with permission denied OR completes after Claude has moved on
4. Consensus table shows N/A for the subagent voice
5. Degradation matrix triggers "single-reviewer mode" silently

Impact

• Every `/autoplan` invocation degrades to single-reviewer mode
• The dual-voice feature (a key differentiator) never actually works
• No error is surfaced — the degradation matrix masks the systematic failure
• Users see the consensus table but subagent voice is always missing

Proposed Fix

Change "simultaneously" → sequential dispatch:

1. Run Claude subagent via Agent tool in foreground (blocks, inherits permissions)
2. Run Codex via Bash (blocks)
3. Both outputs available → build real consensus table

Change "All non-blocking" → "Agent tool blocks in foreground; Codex blocks via Bash."

PR incoming: #fix/autoplan-subagent-parallelization

Environment

• Claude Code with Opus 4.6
• gstack v0.11.18.2
• macOS Darwin 25.3.0