Skip to content

feat(fabric)!: accept a skill folder or a folder of skills in skills.paths - #14296

Merged
cv merged 3 commits into
v1from
chore/fabric-pin-skill-paths
Nov 7, 2026
Merged

cv merged 3 commits into
v1from
chore/fabric-pin-skill-paths

Conversation

@cv

@cv cv commented Nov 7, 2026

Copy link
Copy Markdown
Collaborator

OpenClaw and Hermes agents can now list both kinds of folder in skills.paths: a folder that is one skill, and a folder of skills. A deployment can mix them, for example a shared volume of team skills plus one skill mounted on its own with its own mount settings. NemoClaw pins a Fabric revision that applies this rule.

Problem

Skills on a Mounted Volume mounts a volume on /sandbox/skills and lists that path for an OpenClaw agent. At Fabric 934ab7ec, OpenClaw accepted a path only if it contained SKILL.md itself. With a new, empty volume or a volume holding a folder per skill, the agent failed to start (start_failed, cause openclaw_skill_invalid). Hermes searched every listed folder recursively, at any depth, and skipped a missing path without an error. #14271 left the OpenClaw side open.

Change

The Fabric pin moves to ec77dcc0132bb02198f7aac736671bd0effa87eb, the head of NVIDIA/NeMo-Fabric#615, which is stacked on NVIDIA/NeMo-Fabric#576. It is kept reachable as branch integration/nemoclaw-pin-skills on cv/NeMo-Fabric. The pin now equals the Fabric stack's top pull request exactly, with no extra merges or reverts. NVIDIA/NeMo-Fabric#615's base already contains Fabric main, NVIDIA/NeMo-Fabric#573, NVIDIA/NeMo-Fabric#572 and NVIDIA/NeMo-Fabric#576.

For OpenClaw and Hermes, each skills.paths entry is now resolved like this:

  • a folder that contains SKILL.md is one skill;
  • otherwise, each folder directly inside it that contains SKILL.md is a skill, and other files and folders are ignored;
  • a folder with no skills, such as a new, empty volume, adds none;
  • a path that doesn't exist, or two skills with the same name, stops the agent from starting, and the error names the path or both skill folders.

What Hermes users lose: skills more than one folder deep inside a listed folder no longer load. A listed path that doesn't exist now stops the agent from starting.

Other Fabric changes between the two pins:

In this repository:

  • FABRIC_REVISION and FABRIC_SHA256 in image/fabric/Dockerfile, the nemo-fabric-core revisions in the three crates and Cargo.lock, the regenerated catalog, and the pinned links in the docs. The Hermes bundled-plugin recipe keeps 934ab7ec, because it records the binding the recipe was written against.

  • Skills on a Mounted Volume states the rule and shows two agents. An OpenClaw agent and a Hermes agent share a read-only team-skills volume that holds a folder per skill, and only the OpenClaw agent also mounts a writable review-skill volume that is one skill. The page has YAML and HCL versions of the example.

  • The docker-openclaw-hermes-interfaces live case gives each sandbox two skills volumes and lists both mounts in skills.paths:

    • OpenClaw's folder holds two skills and a file that isn't one;
    • Hermes's folder is empty;
    • each agent's second volume is one read-only skill.

    The tests/e2e/hcl/volume module takes an optional files map and writes those files before the agent starts. cargo ci live-docker names the volumes after the run's UID. Whether an agent uses a skill isn't checked live: Fabric's tests check what each harness loads.

  • The verify-fabric-boundary skill no longer says OpenClaw needs each path to be one skill folder. It still lists the supplied skill's own folder.

  • The fixture image that the HCL cases name, and the registration names in openshift-external-two-providers, follow the new catalog.

Beyond CI

  • Before the pin change, the extended interfaces case failed at apply, with OpenClaw's start_failed, cause openclaw_skill_invalid. After it, the case passed.
  • On linux_arm64:
    • cargo images build built all 11 production agent images and the reference image, and cargo images qualify passed for all 12;
    • catalog.py, run with the built Claude Code image's interpreter, produced the committed catalog.
  • GitHub's codeload archives of ec77dcc0 under NVIDIA/NeMo-Fabric and cv/NeMo-Fabric have the pinned SHA-256.
  • Fabric ec77dcc0's tests are reported in feat: accept a skill folder or a folder of skills in skills.paths NeMo-Fabric#615.

Not run: cargo ci live-kind, and the images on linux/amd64. The Kubernetes images change only by the Fabric pin.

cv added 3 commits November 6, 2026 18:33
…paths

Pin Fabric ec77dcc0, the head of NVIDIA/NeMo-Fabric#615, so OpenClaw and
Hermes resolve each skills.paths entry as one skill (it contains
SKILL.md) or a folder of skills (each folder directly inside it that
contains SKILL.md). A folder without skills adds none; a missing path or
a duplicate skill name stops the agent from starting.

The interfaces live case mounts a folder of skills and a separate skill
for each agent, and Skills on a Mounted Volume shows both forms mixed.

Signed-off-by: Carlos Villela <cvillela@nvidia.com>
…abric-boundary

Signed-off-by: Carlos Villela <cvillela@nvidia.com>
@cv

cv commented Nov 7, 2026

Copy link
Copy Markdown
Collaborator Author
Authoring session — opening
Field Status Value Evidence or limit
Harness recorded Pi coding agent 1.5.0 (session format 3), TUI, as a Herdr helper agent Session header and installed release
Skills recorded nemoclaw-contributor-create-pr (complete read); test-driven-development (complete read) Tool results to cutoff; NeMo Fabric's prepare-pr was read only in part, so it isn't counted
Instruction files recorded NeMo-Fabric AGENTS.md (injected); NeMo-Fabric CONTRIBUTING.md, NemoClaw AGENTS.md, WRITING.md, NeMo-Fabric .github/pull_request_template.md, NemoClaw .github/PULL_REQUEST_TEMPLATE.md and .github/SESSION_REPORT_TEMPLATE.md (complete reads) Project-context injection and tool results to cutoff
Tool calls partial typescript: 263 Attempted calls to cutoff, including 6 that returned errors; the shell, git, gh, and file calls inside each program aren't recorded separately
Session window recorded 2026-11-06T23:18Z to 2026-11-07T01:38Z; 2 h 20 min First event to cutoff, including idle time; the session also authored NVIDIA/NeMo-Fabric#615 before this change
Compactions recorded 0 Complete record to cutoff; 48 context edits elided or summarized earlier tool results
Models recorded nvidia-internal aws/anthropic/bedrock-claude-opus-5-5 Model change at session start; no later changes
Reasoning levels recorded high Thinking-level change at session start; no later changes

@cv
cv enabled auto-merge (squash) November 7, 2026 01:40
@github-actions

github-actions Bot commented Nov 7, 2026

Copy link
Copy Markdown
Contributor

@cv
cv merged commit 47d6498 into v1 Nov 7, 2026
68 checks passed
@cv
cv deleted the chore/fabric-pin-skill-paths branch November 7, 2026 02:00
cv added a commit that referenced this pull request Nov 11, 2026
…ms (#14413)

The release notes now list every user-visible change on `v1` since the
`v0.1.0-rc9` tag, including the breaking ones, and four example pages
drop claims that later changes made false.

## Problem

`docs/about/release-notes.md` listed three breaking changes and none of
the features, fixes, and onboarding changes that reached `v1` after
`v0.1.0-rc9`; it didn't name the configuration shapes the parser now
rejects or the intent record version `v0.1.0-rc9` writes.
Some example claims went stale:

- `examples/spark/README.md` said the pinned OpenShell has the Podman
TLS initialization blocker, NVIDIA/OpenShell#3427; its fix merged before
OpenShell 0.1.0, and `v1` pins 0.1.7.
- `examples/station/README.md` and `examples/gtc-dc/README.md` required
exactly one visible GPU on the host; since #13929,
`container.gpuDevices` lets a service on a host with more GPUs request
one.
- `examples/hcl/travel-team` said Fabric's Deep Agents adapter can't
turn reasoning off, NVIDIA/NeMo-Fabric#433; NVIDIA/NeMo-Fabric#439 fixed
that, and the pinned catalog's Deep Agents `reasoning_effort` accepts
`none`.
- The travel team cited #13313 for the booking key its service doesn't
receive, but #13313 closed when `openshell_api` landed.

## Change

- `docs/about/release-notes.md` gains "Changes Since v0.1.0-rc9",
grouped by area, from the commits in `v0.1.0-rc9..origin/v1`; where a
later commit changed an earlier one, the entry describes the current
code, and changes added and removed after the tag aren't listed:
- Agents and interfaces: #14037, #14038, #14269, #13981, #14024, #14186,
#14052, #14385, #13971, #13858, #13947, #14039, #14042, #14044, #14343.
- Inference: #13929, #14391, #14396, #14398, #14400, #13963, #13912,
#13964, #14388, #14390, #14397.
- Gateways and Kubernetes: #13909, #13970, #13922 (an existing
namespace, refused at plan by #13756, is now bound; cluster-wide reads
dropped), #14014, #14169, #13921.
- Plan, apply, and diagnostics: #14224, #13756, #14141, #14133, #14135,
#13775, #13744, #13701.
- Onboarding: #13693, #13695, #13741, #13757, #13851, #13978, #14228,
#14233, #14334, #14336.
- Examples and builds: #14166, #14184, #14188, #14191, #14192, #14200,
#13889, #14267, #14312, #14315, #14355, b30e9b4, #13781, #13813,
#14246, and the pins in `versions.json` and `image/fabric/Dockerfile`.
- Its breaking changes since the tag:
- the intent record moving from 15, which `v0.1.0-rc9` writes, through
16 (#14266), 17 (#14343), 18 (#13929), 19 (#14391) and 20 (#14394);
  - name-keyed `sandboxes` and `inferenceProviders` (#13703, #13705);
- a recipe `compatibility.gpu` outside every hardware profile (#13912);
- model tuning the adapter doesn't accept (#14111), and Hermes's
`max_tokens` losing its default (#14296);
  - `skills.paths` resolution (#14296);
  - endpoint and `networkCIDR` validation (#13847, #14081);
- the HCL `gateway_values` attribute (#14014), and Kubernetes ownership
receipts from before #13922 and #14014.
  The two earlier items stay, under their own heading.
- `examples/spark/README.md` replaces the closed OpenShell#3427 blocker
with the Podman and remote-model limits in `docs/resources/limits.md`.
- `examples/station/README.md` and `examples/gtc-dc/README.md` require
one GPU visible in the model container, and point to
`container.gpuDevices` for a host with more.
- `examples/hcl/travel-team/README.md` and `WALKTHROUGH.md` drop the
NeMo-Fabric#433 limitation, and the travel team's booking-key statements
link #14417 instead of #13313.

Not listed, because both changes landed after the tag: Deep Agents Code
(#14043, #14250), `compatibility.gpuCount` (#14391, #14394), and the
"Creating infrastructure state" progress step (#14169, #14248).

Checked and left unchanged: the root `README.md`,
`crates/nemoclaw-onboarding/README.md`,
`crates/nemoclaw-cli/assets/README.md`, `internal-docs/README.md`, the
other example READMEs, and every example YAML header comment. The GPU
autoscaling README and `load.sh` read the JSON fields `OperationResult`
serializes, and NVIDIA/OpenShell#4287 is fixed only after the pinned
0.1.7. The two pre-tag release-note items still match the schema and
SDK. `docs/README.md` is left to #14412, which retires the stale field
reference snapshot it describes.

Not verified: whether Hermes sends the booking key's placeholder to MCP
servers; #14417 tracks the live check.

## Beyond CI

Each relative link and anchor in the changed files resolves to an
existing file and heading.
@wscurran wscurran added area: e2e End-to-end tests, nightly failures, or validation infrastructure area: skills Skills, agent behaviors, prompts, or skill packaging feature PR adds or expands user-visible functionality integration: hermes Hermes integration behavior integration: openclaw OpenClaw integration behavior v1 NemoClaw v1 branch labels Nov 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

area: e2e End-to-end tests, nightly failures, or validation infrastructure area: skills Skills, agent behaviors, prompts, or skill packaging feature PR adds or expands user-visible functionality integration: hermes Hermes integration behavior integration: openclaw OpenClaw integration behavior v1 NemoClaw v1 branch

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants