arch-lookup

star 1.5k

Look up Tensix architecture, instruction, or LLK implementation details across architectures. Orchestrates sage agents in parallel.

tenstorrent By tenstorrent schedule Updated 4/20/2026

name: arch-lookup description: Look up Tensix architecture, instruction, or LLK implementation details across architectures. Orchestrates sage agents in parallel. user_invocable: true

/arch-lookup — Architecture Documentation Lookup

Usage

/arch-lookup "How does SFPMAD work?"
/arch-lookup "How does unpack handle Float16 on Blackhole?"
/arch-lookup "What is BroadcastType?"
/arch-lookup "How do T0/T1/T2 synchronize?"

Step 1: Analyze the Query

Determine which architectures are relevant:

  • Specific architecture mentioned → launch ONLY that sage
  • No architecture specified → launch ALL relevant sages in parallel (up to 3)
  • Pure ISA question with no arch context → launch sage-wormhole + sage-blackhole only (they have DeepWiki access). Skip sage-quasar unless Quasar is mentioned — tt-isa-documentation has no Quasar content.

Step 1b: Classify Question Type (MANDATORY)

Before writing the sage prompt, classify the question. This determines which source the sage must consult first — otherwise the sage will default to grep and miss authoritative HW information.

Question Type Examples Required Primary Source
HW capability / spec "What formats does X support?", "What SFPU instructions exist?", "How wide is the FPU?", "How many DEST rows?" Confluence (QSR) / DeepWiki (WH,BH) FIRST, code second
HW behavior / semantics "How does SFPMAD handle NaN?", "When does the pipeline stall?", "What does instruction X do in HW?" Confluence / DeepWiki FIRST, code second
LLK implementation "How is matmul implemented?", "What does llk_unpack_AB do?", "Where is X wired up?" Code FIRST (tt_llk_*/), docs second
Mixed / end-to-end "How is Float16 handled from unpack to pack?" Both required — docs for HW limits, code for the software path

Critical distinction: LLK code shows what is wired up in software. It can lag or diverge from what the hardware actually supports. For HW-capability / HW-behavior questions, a grep-only answer is incomplete — the authoritative source is Confluence (Quasar) or DeepWiki / tt-isa-documentation (WH, BH).

Step 2: Launch Sage Agents

Launch up to 3 agents IN PARALLEL using the Agent tool:

Agent Scope Primary ISA Source
sage-wormhole tt_llk_wormhole_b0/ DeepWiki + assembly.yaml
sage-blackhole tt_llk_blackhole/ DeepWiki + assembly.yaml
sage-quasar tt_llk_quasar/ Confluence + assembly.yaml (NO DeepWiki)

Prompt construction rules:

  1. Start with the user's exact question, verbatim.
  2. If the question is HW capability / behavior (per Step 1b), you MUST explicitly instruct the sage to consult docs first. Example phrasing to include:

    "This is a HW-capability question. Consult Confluence first (sage-quasar) / DeepWiki tt-isa-documentation first (sage-wormhole, sage-blackhole) to get the authoritative HW answer. Use code grep only to cross-verify what is actually wired up in the LLK. If docs and code disagree, call out the conflict — do not silently prefer one."

  3. Do NOT over-constrain the sage prompt with code-only framing (e.g. "grep for the enum", "cite file:line for the definition") on HW-capability questions — that will push the sage into grep-only mode and bypass its doc-consultation protocol.
  4. Always allow the sage to return both doc-sourced facts (with staleness caveats) and code-sourced facts (with file:line).

Step 3: Aggregate

After all sages return:

  1. Commonalities — what's consistent across architectures
  2. Differences — architecture-specific variations (highlight these clearly)
  3. Synthesize — unified explanation with per-arch sections where they diverge

Quality Checklist

Before responding, verify:

  • Question was classified per Step 1b; sage prompt matched the classification
  • For HW-capability / HW-behavior questions, the sage cited Confluence (QSR) or DeepWiki / tt-isa-documentation (WH, BH) — not just code grep
  • Where HW docs and LLK code disagree, the conflict is called out (code lagging HW is common)
  • WHY explained — hardware constraints documented, not just code description
  • Default path identified — baseline vs variants distinguished
  • All data format paths covered
  • Code references include file:line; doc references include Confluence URL + last-updated date (Quasar)
  • Architecture differences highlighted where they exist
Install via CLI
npx skills add https://github.com/tenstorrent/tt-metal --skill arch-lookup
Repository Details
star Stars 1,507
call_split Forks 483
navigation Branch main
article Path SKILL.md
More from Creator