Docs: repo agent skills are unreachable from every documented entry point

[jkzhang7]

Request type

Add missing documentation

Documentation location

• AGENTS.md — "Repo Skills" section
• website/docs/agent-guide.md — https://nvidia.github.io/TensorRT-Model-Connect/agent-guide
• website/docs/extend/add-model-family.md — step 5
• website/static/llms.txt

Problem or missing content

The repository ships 11 agent skills under plugins/trtmc-agent-skills/skills/.
Two of them cover the most common contributor task end to end:

• transform-model (196 lines) — onboarding a Hugging Face model: the
  three-descriptor ownership map, the evidence gates, and the validation
  commands.
• debug-trt-mismatch (203 lines) — "find the first boundary where TensorRT and
  the declared reference disagree", then escalate to the smallest responsible
  family-owned operation.

Nothing outside plugins/trtmc-agent-skills/ references either by name.
AGENTS.md names the directory and routes exactly one skill
($write-git-messages), including the "if it is not listed in the active
runtime skills, load the SKILL.md directly" fallback — both bullets are scoped
to that one skill. The skills cross-reference each other, so the reference graph
is closed: a reader reaches it only by listing the directory on spec.

This matters more for non-Codex agents. The skills are registered through
.agents/plugins/marketplace.json (the Codex plugin path). There is no
.claude/skills/ and no CLAUDE.md, so for a Claude Code session the skills
never appear in the runtime skill list at all — and the natural check ("does a
skill cover this?") consults a list that structurally cannot contain them.

Concrete cost, from onboarding a new family in #1123: roughly fifteen rounds of
whole-pipeline iteration on a numerical mismatch (regenerate audio, score it,
change one thing, repeat). What finally worked was bisecting to the first
divergent stage against independently computed baselines — which is the opening
instruction of debug-trt-mismatch. The skill was correct, present, and unread.

Verification and search performed

# Every mention of these skill names outside their own directory:
for s in transform-model debug-trt-mismatch fp16-trt-network setup-trtmc-environment; do
  grep -rl -- "$s" --exclude-dir=node_modules --exclude-dir=.git . \
    | grep -v "^./plugins/trtmc-agent-skills/skills/$s/"
done
# transform-model and setup-trtmc-environment: no results.
# The other two: only other SKILL.md files and tests/tools/test_test_impact.py.

grep -c "skill" website/docs/extend/add-model-family.md   # 0
grep -n "AGENTS\|skill\|plugins/" CONTRIBUTING.md          # no matches
grep -in "skill" website/static/llms.txt                   # no matches
grep -c "transform-model" website/docs/agent-guide.md      # 0

Also checked: website/docs/wiki/ (all pages are archived redirect stubs, so
this is deliberately not a request to restore TRT-Internals.md),
website/docs/architecture/runtime-lifecycle.md, and
website/docs/models-recipes/overview.md.

Suggested correction

Three small, additive edits. No new pages, no duplication of skill content.

1. AGENTS.md "Repo Skills" — extend from naming the directory to a
   one-line-per-skill routing table, and generalize the existing load-by-path
   fallback beyond $write-git-messages. Roughly:

   Task	Skill
   Onboard an HF model / extend a family	transform-model
   TRT output disagrees with the reference	debug-trt-mismatch
   FP16/BF16 dtype and FP32 boundaries	fp16-trt-network
   Prepare a build or validation host	setup-trtmc-environment
   CI triage and rebases	pr-babysitter
   Commit and PR text	write-git-messages

2. website/docs/agent-guide.md — one sentence pointing at the skills
   directory, so the page that exists to orient agents actually names them.

3. website/docs/extend/add-model-family.md step 5 — name the local CI
   reproducer next to the existing commands:
   python3 -m tools.community_ci source-quality --base <base-ref> and
   python3 -m tools.community_ci unit --scope all. In feat(minimax_music3): onboard MiniMax-Music3 as a text-to-music family #1123 these two reproduce
   the Community CPU / Source quality and Unit / C++ and Python jobs exactly
   and would have caught a complexity-ceiling violation, a C++ compile break, and
   a test-fixture OOM before they reached CI.

Optional, smaller: append to fp16-trt-network a short list of TensorRT
network-definition calls that fail silently rather than raising — e.g.
add_concatenation defaults to axis = 0, and ActivationType.SILU is absent
at the pinned version so SwiGLU must be built as x * sigmoid(x). The
discriminator worth documenting is "silently computes the wrong thing", not
"raises an error", since the latter is self-diagnosing.

Possible follow-up, not proposed here: a shared stage-dump / stage-inject
convention for pipelines. In #1123 an ad-hoc environment-variable hook that
injected reference hidden states into the downstream stages was the single
highest-yield debugging step, because it separated "the codes are drawn wrong"
from "what happens after them is wrong". That is a capability question rather
than a documentation one, so it would need a maintainer design decision.

Happy to open a PR for items 1–3 if the approach looks right.

Submission checks

• I searched open and closed documentation issues and found no duplicate.
• I removed secrets, private/internal evidence, personal paths, and restricted artifacts.

[yifeif-nv]

Thanks for the feedback. Let me take a look and see if we can make the agent scale more explicit. The general direction looks good to me. Please do open the PR for them, and we can get those merged.

Especially on the third point, I think this is important. We need to make sure this is more explicit so people know:
• What steps need to be taken
• What tests will be run in the pre-merge, and what tests are expected to be run and verified locally first before the merge

For the third point specifically, I actually started a discussion thread to talk about how to improve the PR merging experience. Feel free to leave high-level feedback there as well. #1128