Why: The user asked for a researched design covering subagent use,
tiered model routing (Claude via dmxapi for complex work, MiniMax for
the rest), a millisecond/sub-second task sandbox (record-only), and a
verdict on why the agent still feels weaker than codex/claude-code/kimi.
Why: Four research threads (external frontier survey, subagent/routing
evidence, harness-vs-model attribution, local code audit and probes)
converge on role/stage routing instead of whole-task routing, and on
three corrections to the sandbox assumptions (overlay is not a snapshot;
snapshots defer conflicts to merge time; destroy must follow sealing).
Evidence: Local measurements are SELF receipts in the Dev task card
(bwrap namespace ~7ms, overlay same, docker ~142ms; MiniMax reasoning
split; dmxapi /v1/messages cache_read; EXT-16 sandbox missing the
python-docx layer). External claims carry per-item source URLs in the
proposal's section 8.
Impact: Documents-only addition: new proposal 2026-10-10 (E1-E5/S0
increments, D/R workstreams, sandbox tiers L0-L2) plus INDEX rows for it
and the previously unlisted G1 package. No product code changes; all
implementation increments stay unauthorized until approved.
Refs: proposals/2026-10-09-agent-quality-improvement-plan.md (P4/P5,
model A/B), 2026-10-09-ext-g1-b1-b3-exec-backend.md (EXT-16),
Dev task feature/agent-exec-routing-sandbox-plan.
weichen