GLM-5 and DeepSeek 3.2 are blocked in Kiro's agentic mode because Kiro gates
tool-calling and image support together and those models don't pass the check.
Amazon Nova Lite (CAT 2) and Amazon Nova Pro (CAT 3) are Bedrock-native, support
full tool-calling in Kiro's agent flow, and maintain the same cost-control intent.
Updated: model-routing-protocol.md, cat-router-check.mjs, PROJECT_OVERVIEW.md,
MASTER_PROMPT.md, wo1 tasks.md (completed tasks, for accuracy).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Pipeline 1: AWS-native digital human — batteries-included first workflow
hunt+1: Claude Sonnet 5
6933ec1
Add .kiro/steering/pipeline-1-aws-native.md: full node mapping (Transcribe →
Bedrock → Polly → Sumerian Hosts → KVS WebRTC), Polly viseme→BlendshapeFrame
adapter table, Sumerian Hosts SDK setup, credential access via AMANDA vault MCP,
and Miranda's architectural purpose as the harness that proves any workflow —
including a research paper — can be brought to life without rebuilding the infra.
Update PROJECT_OVERVIEW.md: Pipeline 1 table at the top, node-role-label
clarification (labels are roles, not deployments), reference to pipeline-1 doc.
Update MASTER_PROMPT.md: Pipeline 1 steering doc reference, node-label
explanation appended so any Kiro session opening this prompt understands
the Riva/Nemotron/Hive names before touching the topology.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Diversify CAT-5 model routing across providers, not all-Claude
hunt
6a139bc
Claude reserved for CAT 4-5 only, per explicit direction. New mapping
across Kiro's confirmed model roster:
CAT 1 -> Qwen3 Coder Next (coding-specialized, cheapest)
CAT 2 -> GLM-5
CAT 3 -> DeepSeek 3.2 (new workhorse tier)
CAT 4 -> Claude Sonnet 5, escalate to Opus 5 after 2 failed verifications
CAT 5 -> Claude Opus 5 only
28 of 36 total tasks (CAT 1-3) now never touch Claude at all -- a real
cost improvement beyond the original single-provider design. Generalized
the recruit-and-relieve procedure to cover any tier transition, not just
the climb to Opus. Updated: steering doc, PROJECT_OVERVIEW.md,
MASTER_PROMPT.md, cat-router-check.mjs's model table. Verified the
scanner script still runs correctly against the new mapping.
Add the CAT-5 Model Routing Protocol and clarify Miranda as workflow-agnostic harness
hunt
6c1cf40
- .kiro/steering/model-routing-protocol.md: mandatory, global model-routing
rule. CAT 1-5 difficulty scale per task, mapped to Haiku 4.5 / Sonnet 5 /
Opus 5. Opus 5 is recruited only for the 3 true CAT 5 tasks in the whole
project (WO-1 ring buffer, WO-3 SIMD blendshape regression, WO-5 WebGPU
viewport), with evidence-triggered escalation for CAT 4 and a documented
recruit-and-relieve procedure. Honest about the real constraint this is
built around: Kiro has no API for automatic mid-session model switching,
so this is a decision protocol for the human/agent to follow via the
chat model dropdown, not an invisible auto-router.
- All 5 tasks.md files: every task now tagged [CAT n] inline.
- scripts/cat-router-check.mjs: real, tested scanner reporting the pending
CAT backlog across all Work Orders (verified output: 3 CAT5, 5 CAT4,
10 CAT3, 7 CAT2, 11 CAT1 across 36 tasks).
- nobility-posh-framework skill + PROJECT_OVERVIEW.md: clarified that
Miranda-Engine is the harn…
