diff --git a/PhyAgentOS/skills/forge-skill/robodojo/README.md b/PhyAgentOS/skills/forge-skill/robodojo/README.md
new file mode 100644
index 000000000..dd3a4a09f
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/README.md
@@ -0,0 +1,152 @@
+# robodojo Skill
+
+`robodojo` is a Forge Skill with one official profile, `g05`. It runs the RoboDojo
+`put_bottles_into_dustbin` task through Isaac Sim and exposes two Gateway tools to
+PAOS: `bench.describe` and `bench.run`.
+
+This Skill ships the description, manifest, g05 profile, dataflow and install
+material only. The benchmark node and the G05 policy node are Forge Nodes whose
+source lives in `framework/forge_runtime`; the profile reaches them through the
+installed Node entrypoints under `${FORGE_RUNTIME_BIN}`.
+
+## Layout
+
+```text
+skill.yaml                     manifest (manifest_version 2, profile g05)
+SKILL.md                       Skill document referenced by skill_document
+profiles/g05/dataflow.yaml     gateway -> bench_endpoint -> robodojo_benchmark -> robodojo_policy
+profiles/g05/*.yaml            gateway / bench_endpoint / benchmark config
+profiles/g05/env.template.sh   host environment template
+scripts/install_robodojo.py    per-run TOS presign + `paos skill install --index`
+scripts/check_bundle_inventory.py  verify the template inventory against this tree
+docs/                          unsigned schema-v3 index template + 0.1.0 delivery record
+```
+
+There is no `nodes/` tree and no `profiles/g05/bin/` launcher here. An earlier
+revision of this branch vendored the benchmark/policy sources and two in-bundle shell
+launchers inside the Skill, which duplicated the Node sources and let the two copies
+drift (four files had already diverged through lint-only edits). The sources of
+truth are:
+
+| node | source of truth | build entry |
+| --- | --- | --- |
+| benchmark + G05 policy | `framework/forge_runtime`, `packages/nodes/benchmarks/robodojo/` | `scripts/build_node_artifacts.py` (`robodojo_benchmark`) |
+| gateway | `framework/forge_gateway` tag `v1.0.2` | same script (`forge_runtime`) |
+| bench endpoint | `framework/forge_runtime`, `packages/runtime/bench_endpoint/` | same script (`robodojo_endpoint`) |
+
+## Node bundles
+
+| node_id | artifact_id | entrypoints | node_digest |
+| --- | --- | --- | --- |
+| `forge_runtime` | `forge-runtime-0.1.0` | `gateway` | `cee927eb587f3f958c1803086e59692f6f7ae373477b3cfd15118d01d76c408e` |
+| `robodojo_endpoint` | `robodojo-endpoint-0.1.0` | `bench_endpoint` | `b9a66fb07b04ee1f76bb8221f68f96b35d1f1ce136d3d6cf967cb7a34da8e159` |
+| `robodojo_benchmark` | `robodojo-benchmark-0.1.0` | `robodojo_benchmark`, `robodojo_policy` | `e48cf4e838ba73512ac1eb0c144ea17ea177bb56ad41e524c8332f9a38c19ff2` |
+
+The benchmark node and the G05 policy node ship as **one** digest bundle with two
+entrypoints: they share `_core/`, `config.py` and `adapters/` and run in the same
+Python environment, so a second archive would only duplicate the same tree.
+
+## Source / artifact correspondence
+
+- The published `0.1.0` TOS objects (Skill archive sha256 `5c7806843e5cff16a8a407254ffe7736c9b1d5d13982bd1b82195c7e4f5bc907`
+  plus the `forge-runtime` / `robodojo-endpoint` Node archives) are frozen snapshots
+  of the **pre-split** layout: that Skill archive still contains `nodes/` and
+  `profiles/g05/bin/`. They stay untouched.
+- This branch is the **post-split** candidate. The Skill archive is rebuilt with the
+  official packager (`scripts/package_skill.py`) and needs a new object key or version
+  before it can be published. Its sha256 is recorded in the pull request description:
+  a file cannot carry the hash of the archive it is packed into.
+- Node implementations come from `framework/forge_runtime`. The recorded base is branch
+  `qinhan/robodojo-g05-node` at `1630e75a450464307cdab08de6e7d4536d04af86`; this revision adds
+  `93c2659` (lint-only port that makes the Node source byte-identical to the copy that had been
+  vendored in this Skill) and a `scripts/build_node_artifacts.py` fix that first sorts the
+  payload inventory by POSIX path, so one source tree yields one `node_digest` on Windows and on
+  Linux alike.
+- That branch (`qinhan/robodojo-node-bundles`) is committed locally only.
+  `framework/forge_runtime` has no personal fork and the GitLab project cannot be forked without
+  an API session, so pushing it — as a topic branch or as a fork — is a maintainer decision
+  under the fork + MR contribution flow.
+- `node_digest` is content-derived and reproducible across hosts; the Node *archive* sha256 is
+  not, because the tarball records file mtimes and ownership.
+
+## Runtime requirement
+
+`skill.yaml` uses two manifest features that must be present in the PAOS runtime:
+
+- `profiles.<name>.startup_timeout_s` — a per-profile Dora readiness timeout
+  (the generic LIBERO `lingbot_va` profile uses `900`).
+- digest-locked multi-file Node bundles — `artifacts.nodes.<id>.digest` pins the
+  installed `node-manifest.json` digest, and the installed Node manifest declares the
+  `entrypoints` that provide `gateway`, `bench_endpoint`, `robodojo_benchmark` and
+  `robodojo_policy`.
+
+`profiles.<name>.startup_timeout_s` is already supported by upstream `dev`.
+Digest-locked bundles come from `feature/skill-runtime-node-bundles` (PR #112); this
+branch depends on it. The legacy single-executable `artifact_type: executable_tar_gz`
+lock form keeps working unchanged.
+
+## Build
+
+```bash
+python scripts/package_skill.py PhyAgentOS/skills/forge-skill/robodojo \
+  --output-dir dist/skills
+```
+
+The Node bundles are built from `framework/forge_runtime`:
+
+```bash
+python scripts/build_node_artifacts.py --out dist/nodes --forge-src /path/to/forge_gateway
+```
+
+## Install
+
+Requires read permission for TOS bucket `phyagentos-resource-inner` and `tosutil`.
+
+```bash
+export PAOS_CONFIG=/abs/instance/config.json
+python3 scripts/install_robodojo.py --tosutil /path/to/tosutil --vp 1d
+```
+
+`install_robodojo.py` regenerates fresh presigned download URLs on every run, verifies
+each downloaded object against the sha256 recorded in the template, writes a private
+temporary schema-v3 index, and calls:
+
+```bash
+paos skill install robodojo --version 0.1.0 --index <generated-index.yaml>
+```
+
+The historical `paos-forge-packages.0.1.0.yaml` carries expired links and is not the
+recommended entry point.
+
+## Verify, start, stop
+
+```bash
+paos skill list
+paos skill inspect robodojo
+paos forge-node verify robodojo forge_runtime
+paos forge-node verify robodojo robodojo_endpoint
+paos forge-node verify robodojo robodojo_benchmark
+
+paos skill start robodojo --profile g05
+paos skill status robodojo
+paos skill stop robodojo
+```
+
+`paos skill start` needs `dora` on `PATH`. A running Skill leaves Dora up after the
+benchmark finishes; the pending item is that an Agent must actually stop and confirm
+when the user asks.
+
+## External environment
+
+The manifest declares `FORGE_GATEWAY_PYTHON`, `BENCH_ENDPOINT_PYTHON`, `ROBODOJO_ENV`,
+`ROBODOJO_ROOT`, `XPOLICYLAB_ROOT`, `G05_CKPT_PATH`, `G05_PROCESSOR_PATH`, `CUDA_HOME`,
+`CUDA_VISIBLE_DEVICES`, `ROBODOJO_CACHE_ROOT`, `ROBODOJO_DEVICE_ID` and
+`ROBODOJO_RESULT_DIR`. `profiles/g05/env.template.sh` lists the same values with
+descriptions.
+
+`TASK_NAME=put_bottles_into_dustbin` is also exported by `env.template.sh` and read by
+`policy_node.py`; it is not part of the frozen `skill.yaml` `required_environment` list,
+so the host must set it explicitly until a new candidate manifest is published.
+
+Checkpoint/processor and Isaac Sim are external and are not bundled here. The G05
+checkpoint is ~33.9 GB; do not copy it into the Skill package.
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/SKILL.md b/PhyAgentOS/skills/forge-skill/robodojo/SKILL.md
new file mode 100644
index 000000000..c9a164b08
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/SKILL.md
@@ -0,0 +1,144 @@
+---
+name: robodojo
+description: Run the RoboDojo G05 benchmark and report its aggregate score.
+metadata: {"PhyAgentOS":{"always":false,"requires":{"runtime":["robodojo"]}}}
+---
+
+# RoboDojo Benchmark
+
+Run a RoboDojo benchmark batch in Isaac Sim and report the score the benchmark
+produced. Use only these stable Tool IDs:
+
+- `bench.describe`: Query; return the live task inventory, defaults, and episode
+  budget limits. Does not touch the simulator state.
+- `bench.run`: Action; run one batch of episodes and return its aggregate result.
+  Long running — minutes per episode.
+
+Use the PAOS bridge tools `forge_tool_context`, `forge_tool_query`,
+`forge_tool_start_action`, `forge_tool_action_status`, `forge_tool_action_result`,
+and `forge_tool_cancel_action`. Always pass the stable Forge Tool ID explicitly.
+Do not use shell commands and do not construct Gateway HTTP requests directly.
+
+## What this Skill must not do
+
+Scoring belongs entirely to the benchmark node. This Skill selects and sequences
+Tools, and reports what came back. It must not:
+
+- compute, recompute, adjust, average, or extrapolate `success_rate` or any other
+  metric;
+- decide whether an individual episode succeeded, failed, or timed out;
+- aggregate results across separate `bench.run` invocations into one number;
+- convert a partial or failed batch into a score, or present one as if the batch
+  had completed;
+- estimate a score when the result is unavailable.
+
+If a number is needed and no terminal result exists, say so. Do not produce one.
+
+## 1. Always describe before running
+
+Call `bench.describe` first, every time. Never fill `bench.run` arguments from
+memory or from an earlier conversation — the installed suite, task count, and
+limits are properties of the running profile, not of this document.
+
+`bench.describe` returns:
+
+- `benchmark`, `suite`, `policy_id`, `env_cfg_type`, `action_type` — fixed by the
+  profile. They are read-only context, not parameters you can change.
+- `tasks[]` — each with `task_id`, `name`, `language`, `num_init_states`. The
+  init-state count differs per task; do not assume one task's count applies to
+  another.
+- `defaults` — `num_runs`, `seed`, `max_steps` to use when the user gave no value.
+- `limits` — `max_total_episodes`, `max_max_steps`. Hard caps; see §2.
+
+If `bench.describe` fails with `BENCH_NOT_READY`, the simulator is still starting.
+Wait and retry; do not proceed to `bench.run`.
+
+## 2. Build the bench.run arguments
+
+All five fields are required. There are no defaults on the wire.
+
+- `task_ids`: an array of ids taken from `bench.describe`, or `null` for every
+  task. `null` and `[]` are different — `[]` is always an error, never "all".
+- `init_state_ids`: an array, or `null` for every init state shared by all
+  selected tasks. Validated per task.
+- `num_runs`: repeats per (task, init_state) pair. Use `defaults.num_runs` unless
+  the user asked for more.
+- `seed`: base seed. The same seed and the same ids reproduce the same episodes.
+  Change it only when the user asks for a different sample.
+- `max_steps`: a **hard cap on control steps**. The episode is terminated as
+  `timed_out` if this many steps are reached before the environment ends. The
+  task native `step_lim` defines its normal end (success / partial / failed); set `max_steps` to at least the task horizon (from `bench.describe` defaults) to allow completion, or lower to force an explicit step-limit stop.
+
+Episode count is `len(task_ids) x len(init_state_ids) x num_runs`. Keep it at or
+below `limits.max_total_episodes`; exceeding it is rejected, not clamped. When the
+user asks for something larger, report the cap and the requested count and ask
+before splitting the work into several batches.
+
+Nothing else is selectable. Policy weights, suite, camera set, and GPU are fixed by
+the profile. If the user asks to change one of those, say it requires a different
+profile — do not try to express it through these arguments.
+
+## 3. Account for the Action lifecycle
+
+`bench.run` is an Action. Accepted is not completed.
+
+1. `forge_tool_start_action` returns an acknowledgement with an `invocation_id` and
+   `total_episodes`. The batch has not run yet.
+2. `forge_tool_action_status` reports `accepted`, `running`, or a terminal phase.
+   Status is **advisory** — never report a score from it. Poll at a human pace;
+   a single episode takes minutes.
+3. `forge_tool_action_result` is the only authoritative source of the outcome.
+
+Treat the four terminal outcomes as four different things:
+
+- `succeeded` — the batch ran to completion. Report per §4.
+- `failed` — the batch ended abnormally. Report the error code and the episode
+  counts; do not present `success_rate` as the headline.
+- `cancelled` — stopped on request. Partial counts only.
+- `unknown` — **the outcome cannot be recovered**. Side effects have already
+  happened: GPU time was consumed and a `result.json` may exist on disk. Do not
+  blindly retry. Report that the outcome is unknown, give the `invocation_id`, and
+  ask how to proceed.
+
+Use `forge_tool_cancel_action` only when the user asks to stop. Cancel is accepted,
+not immediate — keep polling until a terminal result appears.
+
+## 4. Reporting rules
+
+Report these together, always, in this order:
+
+1. `status` of the batch.
+2. `success_rate`, immediately followed by `successes` / `total_episodes`.
+3. `completed_episodes` / `total_episodes`.
+
+`success_rate` is computed over `total_episodes`, not over
+`completed_episodes`. When a batch ends early these differ, and `success_rate`
+alone reads as "the policy scored low" when the real story is "the batch did not
+finish". If `completed_episodes < total_episodes`, say that first, before the
+score.
+
+Also report:
+
+- `result_path` — the authoritative record on disk. Always include it.
+- `timed_out_episodes` when non-zero.
+- `obs_republished` when non-zero: observations were re-sent during the run. The
+  score is still valid, but mention it.
+- `episodes[]` per-episode outcomes when the user asks for detail. When
+  `episodes_truncated` is true the list was too large to inline — point at
+  `result_path` instead of summarising from an empty array.
+
+Never present a single number without its denominator.
+
+## 5. Errors and replanning
+
+| code | retryable | what to do |
+|---|---|---|
+| `BENCH_NOT_READY` | yes | Simulator still starting. Wait, retry `bench.describe`. |
+| `FORGE_BUSY` | yes | Another batch is running; one simulator per machine. Report which `invocation_id` holds it. Do not queue. |
+| `BENCH_INVALID_ARGUMENT` | no | Arguments are wrong. Re-read `bench.describe` and rebuild them. `details` carries the valid range and the requested vs allowed episode count. Never retry unchanged. |
+| `BENCH_NODE_UNAVAILABLE` | yes | The benchmark node stopped reporting. The outcome is unknown — see §3. Do not restart the batch without asking. |
+| `BENCH_UNKNOWN_INVOCATION` | no | The id is not known here. Do not invent one; start a new run if the user wants one. |
+| `FORGE_DEADLINE_EXCEEDED` | no | The gateway stopped tracking; the batch may still be running. Do not start another — that would collide. Report and ask. |
+
+When two errors are plausible, prefer the more conservative reading: an unknown
+outcome is not a failure, and a rejected batch is not a zero score.
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/docs/DELIVERY_0.1.0.md b/PhyAgentOS/skills/forge-skill/robodojo/docs/DELIVERY_0.1.0.md
new file mode 100644
index 000000000..3b29b2de0
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/docs/DELIVERY_0.1.0.md
@@ -0,0 +1,173 @@
+# RoboDojo 0.1.0 — Installation (release candidate)
+
+**Status**: candidate, post-split layout. The schema-v3 static-index route is verified
+(fresh-instance install on host 120 passed for the pre-split archive; a fresh isolated
+instance install of the post-split layout was also checked locally without GPU). No HTTP
+Resource Registry is used; no default Skill version changed. Final functional acceptance
+is pending.
+
+## What changed: node ownership
+
+This revision removes the vendored benchmark/policy sources and the two in-bundle shell
+launchers from the Skill. The nodes are maintained in `framework/forge_runtime` and are
+installed as digest Node bundles; the profile reaches them through
+`${FORGE_RUNTIME_BIN}/<entrypoint>`.
+
+- The published pre-split objects stay untouched: Skill `robodojo-0.1.0.tar.gz` sha256
+  `5c7806843e5cff16a8a407254ffe7736c9b1d5d13982bd1b82195c7e4f5bc907`, Node
+  `forge-runtime-0.1.0`, Node `robodojo-endpoint-0.1.0`.
+- This revision adds a third Node bundle, `robodojo-benchmark-0.1.0` (node_digest
+  `e48cf4e838ba73512ac1eb0c144ea17ea177bb56ad41e524c8332f9a38c19ff2`, entrypoints
+  `robodojo_benchmark` and `robodojo_policy`). It is **not uploaded yet**.
+- Because the Skill payload changed, its archive must be rebuilt and published under a new
+  object key or version before this revision can be installed from TOS. Versioning and
+  publishing are maintainer decisions; this revision overwrote nothing in TOS and did not
+  register anything in a public index.
+
+## Recommended entry point (credential-enabled, per-run presign)
+Use the **unsigned template + refresh helper** — NOT the historical index that carries
+pre-signed download URLs. The helper presigns fresh HTTPS download links on the target machine
+on every run, so it never depends on a link lifetime. The signed index is written to a private
+temp dir and is not tracked in git. This per-run presign route is **established**; a durable
+HTTP(S)-hosted index / HTTP Resource Registry is **not a blocker** for this route.
+
+## Verified baseline
+- Skill `robodojo-0.1.0.tar.gz` sha256 `5c7806843e5cff16a8a407254ffe7736c9b1d5d13982bd1b82195c7e4f5bc907`
+  (pre-split archive; the post-split candidate is rebuilt locally and needs a new object key)
+- Node `forge-runtime-0.1.0` node_digest `cee927eb587f3f958c1803086e59692f6f7ae373477b3cfd15118d01d76c408e` (entry `bin/gateway`, 60 files)
+- Node `robodojo-endpoint-0.1.0` node_digest `b9a66fb07b04ee1f76bb8221f68f96b35d1f1ce136d3d6cf967cb7a34da8e159` (entry `bin/bench_endpoint`, 27 files)
+- Node `robodojo-benchmark-0.1.0` node_digest `e48cf4e838ba73512ac1eb0c144ea17ea177bb56ad41e524c8332f9a38c19ff2` (entries `bin/robodojo_benchmark`, `bin/robodojo_policy`, 20 files) — candidate, not uploaded
+- Runtime source commit `aac600574a6d9dba225216e1f2271e069be2e4e2` plus the lint-port commit that makes the Node source byte-identical to the copy previously vendored in the Skill
+- TOS bucket `phyagentos-resource-inner`; Skill and Node bundles live under
+  `skill-bundles/robodojo/0.1.0/` and
+  `node-bundles/{forge_runtime,robodojo_endpoint,robodojo_benchmark}/0.1.0/`.
+
+The `node_digest` values above are reproducible from the source trees on any host (the builder
+sorts the payload inventory by POSIX path, so a Windows checkout cannot produce a different
+digest than an LF checkout); the *archive* sha256 is not, because the tarball records file
+mtimes and ownership. Pin and compare `node_digest`.
+
+## Post-split local verification (no GPU, this revision)
+
+Node sources come from `framework/forge_runtime` and are built with
+`scripts/build_node_artifacts.py` (branch `qinhan/robodojo-node-bundles`: base `1630e75`, lint
+port `93c2659`, plus the inventory-ordering fix). The Skill is packed with the official
+`scripts/package_skill.py`. On a Linux host, with no GPU, no Isaac Sim, no Dora and no Agent:
+
+- Fresh isolated instance, verified empty of Skill / Node / lock / cache before the run.
+- Installed with `install_robodojo.py --index <local schema-v3 index>` served over loopback; the
+  loopback server replaced TOS presigning because the rebuilt archives are still candidates.
+  The run exercised the `prepare_instance` fix (`config.json` + `py/sitecustomize.py`), the
+  digest bundle install path and the lock write, and ended with
+  `✓ Installed Skill robodojo 0.1.0`.
+- `paos skill list` reports `robodojo 0.1.0 / g05 / not started`.
+- `paos forge-node verify robodojo {forge_runtime,robodojo_endpoint,robodojo_benchmark}` all
+  report "verified against Skill lock".
+- The profile environment materialises `bin/{gateway,bench_endpoint,robodojo_benchmark,robodojo_policy}`
+  as relative symlinks into `<runtime>/nodes/<node_id>/versions/<artifact_id>/bin/...`, and the
+  rendered `launch/profiles/g05/dataflow.yaml` resolves every node `path` to that expanded
+  `${FORGE_RUNTIME_BIN}` directory — never to a copy inside the Skill tree. The installed Skill
+  contains no `nodes/` and no `profiles/g05/bin/`.
+- `python3 scripts/check_bundle_inventory.py --check` reports every Skill file matching the
+  template inventory. The inventory is not read at install time, so it drifts silently when a
+  Skill file changes; run the checker (or `--write`) on the pack host before registering or
+  publishing this candidate.
+- Still pending: a presign run against real TOS for this revision, and GPU / Isaac Sim /
+  Agent functional acceptance.
+
+## Prerequisites
+- Linux x86_64, CUDA GPU (>=16 GB for G05), Isaac Sim host, EGL/GL (`MUJOCO_GL=egl`,
+  `PYOPENGL_PLATFORM=egl`), `ffmpeg`, `dora` coordinator/daemon.
+- PAOS runtime: `startup_timeout_s` is already supported by upstream `dev`. Digest-locked
+  multi-file Node bundle support is provided by `feature/skill-runtime-node-bundles` (PR #112),
+  which this Skill revision depends on. `paos skill install --index` supports the schema-v3
+  static index. `forge-node install <artifact_id>` needs an HTTP Resource Registry URL and is
+  **not** configured — use the `--index` route.
+- **TOS read permission for `phyagentos-resource-inner`** plus `tosutil` (this is a
+  credential-required install, not a public credential-less one). If the target machine cannot
+  obtain this permission, "download-authorization method missing" is a delivery blocker.
+
+## Install helper (regenerates download links on each run)
+The template `paos-forge-packages.template.yaml` records a `bucket_key` and an
+`expected_sha256` per package. `install_robodojo.py` presigns a fresh HTTPS link for each key
+(via `tosutil presign`), downloads each object once to measure `size` and confirm
+`expected_sha256`, writes a temporary schema-v3 index with `direct_download_url` / `sha256` /
+`size` filled in (`paos skill install --index` rejects an index without them), and then runs the
+verified install command.
+
+The helper is **self-contained**: it finds the template next to itself (or in the repo layout
+`<script>/../docs/`), and it **prepares the isolated instance** on the target machine by
+creating `<config>.parent/config.json` and `<config>.parent/py/sitecustomize.py` (which anchors
+`get_data_dir()` to the instance via `PhyAgentOS.set_config_path`). No manually-created
+`sitecustomize.py` from a prior acceptance tree is required. `paos`/`tosutil` are resolved via
+`--paos`/`--tosutil`, then `PAOS_BIN`/`TOSUTIL`, then `PATH`; no hardcoded host path is assumed,
+and a missing entry exits with setup guidance.
+
+```bash
+# 1) put install_robodojo.py next to paos-forge-packages.template.yaml (e.g. from TOS support/)
+# 2) choose an isolated instance config path
+export PAOS_CONFIG=/your/instance/config.json
+
+# 3) run the helper (uses tosutil on PATH; override with --tosutil; --paos for paos CLI)
+python3 install_robodojo.py --tosutil /path/to/tosutil --vp 1d
+```
+The helper sets `PAOS_CONFIG` + `PYTHONPATH=<instance>/py` for the `paos` child and creates the
+isolation files if absent. Equivalent manual form after links are refreshed:
+```bash
+paos skill install robodojo --version 0.1.0 --index <generated-index.yaml>
+```
+
+## Expired links
+Re-run `install_robodojo.py` — it regenerates `direct_download_url` and re-verifies the content
+hash on every run and prints `generated_at` + the presign validity. You never edit the YAML by
+hand. The signed index is written to a private temp dir and is not tracked. Do **not** use the
+historical `paos-forge-packages.0.1.0.yaml` (its links are time-limited).
+
+## Verify the install
+
+> In-repo copy note: the frozen TOS support copy of this file lists the node verification
+> arguments in the older order; the CLI signature is `paos forge-node verify <skill> <node_id>`,
+> as shown below.
+```bash
+paos skill list
+paos skill inspect robodojo
+paos forge-node verify robodojo forge_runtime
+paos forge-node verify robodojo robodojo_endpoint
+paos forge-node verify robodojo robodojo_benchmark
+```
+Expected: `robodojo 0.1.0 / g05 / stopped`, Gateway `http://127.0.0.1:19011`,
+tools `bench.describe, bench.run`, all three nodes verified.
+
+Install does not load the model and does not start the simulator: it only resolves the Skill
+bundle, verifies the Node archives against the locked digests, and materialises
+`<env>/bin/{gateway,bench_endpoint,robodojo_benchmark,robodojo_policy}` plus the rendered
+dataflow. Those four entrypoints are relative symlinks into the installed Node bundles, so the
+Skill never falls back to a development checkout such as `FORGE_PKGS_ROOT` or
+`BENCH_ENDPOINT_SRC`.
+
+## External assets (not bundled)
+G05 checkpoint/processor + Isaac Sim are external (RoboDojo-sim-arx_x5-joint-0, action_source=fm,
+action_steps=32, continuous_action=True; model source OpenGalaxea/g05-robodojo analog). Configure
+`G05_CKPT_PATH`, `G05_PROCESSOR_PATH`, `ROBODOJO_ENV`, `ROBODOJO_ROOT`, `XPOLICYLAB_ROOT`,
+`CUDA_HOME`, `CUDA_VISIBLE_DEVICES`, `ROBODOJO_DEVICE_ID`, `ROBODOJO_CACHE_ROOT`,
+`ROBODOJO_RESULT_DIR`. The policy node also reads `TASK_NAME` (the dataflow passes it through
+from the host); the frozen manifest does not declare it, so export it explicitly —
+`profiles/g05/env.template.sh` sets `put_bottles_into_dustbin`.
+
+## Start a profile (later; not part of this install check)
+```bash
+paos skill start robodojo --profile g05
+paos skill status robodojo
+paos skill stop robodojo
+```
+
+## Known / pending
+- Agent auto-stop after a run is not implemented; benchmark finishing leaves the Skill running.
+  This is not treated as an error — the pending item is that the Agent actually stops and confirms
+  when the user explicitly asks.
+- No HTTP Resource Registry service is configured. The candidate static-index route is verified;
+  durable HTTPS hosting of the index + maintainer registration would be for public/long-lived
+  publishing and are **not required** for the per-run presign install route above.
+- The `robodojo-benchmark-0.1.0` object and the rebuilt Skill archive are not published yet; the
+  template records them as candidates.
+- `final_acceptance` is pending; no default-version change, no public index mutation.
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/docs/paos-forge-packages.template.yaml b/PhyAgentOS/skills/forge-skill/robodojo/docs/paos-forge-packages.template.yaml
new file mode 100644
index 000000000..dfef4f243
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/docs/paos-forge-packages.template.yaml
@@ -0,0 +1,574 @@
+# Unsigned schema-v3 static package index template for the robodojo Skill.
+#
+# Template-only fields (stripped by scripts/install_robodojo.py):
+#   bucket_key       object key inside TOS bucket phyagentos-resource-inner
+#   expected_sha256  archive sha256 of the published object; the helper re-downloads the
+#                    object, hashes it, and aborts on mismatch. `null` means the object is
+#                    not published yet (candidate).
+#
+# `paos skill install --index` requires `direct_download_url`, `sha256` and `size`; the
+# helper presigns `bucket_key`, measures size and sha256 itself and writes those three
+# fields into a generated index. Never feed this template to `--index` directly.
+#
+# Scope: benchmark and G05 policy node implementations are NOT part of the Skill bundle.
+# They are Forge Nodes built from framework/forge_runtime
+# (packages/nodes/benchmarks/robodojo, scripts/build_node_artifacts.py) and locked here by
+# node_digest. The published 0.1.0 objects are frozen pre-split snapshots and stay
+# untouched; the Skill entry below points at the rebuilt post-split candidate, which still
+# needs a new object key or version.
+schema_version: 3
+packages:
+- kind: skill_bundle
+  package_key: robodojo
+  artifact_id: robodojo
+  version: 0.1.0
+  platform: linux
+  arch: x86_64
+  description: Run the RoboDojo G05 benchmark and return its aggregate score.
+  dependencies:
+  - forge-runtime-0.1.0
+  - robodojo-endpoint-0.1.0
+  - robodojo-benchmark-0.1.0
+  bucket_key: skill-bundles/robodojo/0.1.0/robodojo-0.1.0.tar.gz
+  expected_sha256: null
+  inventory:
+  - path: README.md
+    sha256: 56381c08a3c46ce8e82bced526a29a7e331adb6a5e61a634b3826549d98087c9
+    size: 7350
+    category: file
+  - path: SKILL.md
+    sha256: 82aac86d637f634453ca94bbe61de112bff7dd4b63d5659db75ee29e0f1f3ec3
+    size: 7310
+    category: file
+  - path: docs/DELIVERY_0.1.0.md
+    sha256: d430abc97a74f974875b7bb59aba39eb92adc3a96b694c736d6aeccb31c93392
+    size: 10969
+    category: file
+  - path: profiles/g05/bench_endpoint.yaml
+    sha256: 9765a4828104c30b16ebe73f09ea0fafa57ba02c21dda75e7d48d9ef593459e4
+    size: 618
+    category: file
+  - path: profiles/g05/config.yaml
+    sha256: 3ccb45faf483fb08f1e7fbae534bf6e1cb2714223ab98b8ac4c8e6fd2e28372c
+    size: 740
+    category: file
+  - path: profiles/g05/dataflow.yaml
+    sha256: 050cc1df9a8e9db86cfa0053e0f0b53767daacf3224cc45c2f62b87285a4ae6a
+    size: 2957
+    category: file
+  - path: profiles/g05/env.template.sh
+    sha256: 7ae4e5bec354fe05eeae4f67e20ae91ac4d0b8901392c51f9777c3105d438e22
+    size: 2857
+    category: file
+  - path: profiles/g05/gateway.yaml
+    sha256: c5b49b11ee39d4df38c88514c9dc20f4e490efabec43d837bf2bbe8596ac4eb2
+    size: 6788
+    category: file
+  - path: scripts/check_bundle_inventory.py
+    sha256: 81218a85c909c0be95cd46f9d964a331f57570e7fa2db8fd3a3a26ce9c3d8bb3
+    size: 6330
+    category: file
+  - path: scripts/install_robodojo.py
+    sha256: 266eead1af8162a9c57f09f58da8a4a1efb1402571334b6d612ffc79caebf135
+    size: 10422
+    category: executable
+  - path: skill.yaml
+    sha256: 0c5bbdde16b91c10808ce49090486afe5caeb1ff63631ade484bfb03e43e95b8
+    size: 1635
+    category: file
+  backend_url: null
+  direct_download_url: null
+  archive:
+    format: tar.gz
+  todos: >-
+    Post-split candidate. The published robodojo-0.1.0.tar.gz (sha256
+    5c7806843e5cff16a8a407254ffe7736c9b1d5d13982bd1b82195c7e4f5bc907) still contains the
+    vendored nodes/ tree and profiles/g05/bin/; publish the rebuilt archive under a new
+    object key or version before pointing expected_sha256 at it. The inventory omits
+    docs/paos-forge-packages.template.yaml, which cannot hash itself.
+- kind: node_bundle
+  package_key: forge-runtime-0.1.0
+  artifact_id: forge-runtime-0.1.0
+  version: 0.1.0
+  platform: linux
+  arch: x86_64
+  install_root: ~/.PhyAgentOS/forge_runtime/nodes/forge_runtime
+  entrypoints:
+    gateway: bin/gateway
+  bucket_key: node-bundles/forge_runtime/0.1.0/forge-runtime-0.1.0-linux-x86_64.tar.gz
+  expected_sha256: c524ae7038775e48f2b0a6036e621913b35b327ad2c710f8a8e0c5cad9e7613b
+  node_digest: cee927eb587f3f958c1803086e59692f6f7ae373477b3cfd15118d01d76c408e
+  inventory:
+  - path: bin/gateway
+    sha256: cb134ebbc0f3ed0f804606dd07eda73fa74dad0a4f75f21fe135394fcd687339
+    size: 369
+    category: executable
+  - path: forge_gateway/__init__.py
+    sha256: 51035d2c0bb249b54edb472735e3bb20692df33dc41e99b8d0395e015bc74fe6
+    size: 58
+    category: file
+  - path: forge_gateway/__main__.py
+    sha256: 406b7188a391e7691b262bd978dbeeed3ac654264d533005b48cbad968543e37
+    size: 61
+    category: file
+  - path: forge_gateway/adapters/__init__.py
+    sha256: 779761d9503c1887b2b5d2afcf9642b9785248ceed65339ca422f86240caeeef
+    size: 41
+    category: file
+  - path: forge_gateway/adapters/dora_adapter.py
+    sha256: 237fc682eb211dcb20c0c0004e53e71844430eaf3e54c92b8ef72aa295c373dc
+    size: 11226
+    category: file
+  - path: forge_gateway/adapters/dora_runtime.py
+    sha256: 92c8afccfd8b4bbb3177f450f8b351bfbc17883dfaca5c5e2400bfcedfe2f8e1
+    size: 8145
+    category: file
+  - path: forge_gateway/adapters/http_app.py
+    sha256: f2aef6d34b20131e558415e5a778f19a067d54779cba88b5c47f3111190a5ce9
+    size: 1975
+    category: file
+  - path: forge_gateway/adapters/manifest_loader.py
+    sha256: 4897749c2db9455bef76282550d7d84037aac2c34401cf0a48eddb6242469d40
+    size: 3385
+    category: file
+  - path: forge_gateway/adapters/state_files.py
+    sha256: d4e97037f3a7183f700716c425c89daaf77ea5de284d5b679d9a9843ef3f557e
+    size: 2275
+    category: file
+  - path: forge_gateway/adapters/tool_dora.py
+    sha256: 8b4baa800f5d952e19a38120d34c403f66737e8272c074cb9f732bb0d3be9643
+    size: 4923
+    category: file
+  - path: forge_gateway/adapters/websocket.py
+    sha256: 1c0485bc225f34720177538bc53a6b8bc793dc6026bc89f6acba8d9f7e786b9b
+    size: 755
+    category: file
+  - path: forge_gateway/application.py
+    sha256: 55440507745ef30385699e73672f11d30a3448a6bd77dfbf35120ba4f819b8c2
+    size: 22323
+    category: file
+  - path: forge_gateway/cli.py
+    sha256: cb75a68e022c150b9c27c45c1697f0f79b3f4b98a7cc79f29f10da71bc23320e
+    size: 5049
+    category: file
+  - path: forge_gateway/config.py
+    sha256: a9e9e7885771b58e743b6661a9aaeb22e040d4c1d7f2deb9f1e99708a0bf719e
+    size: 28171
+    category: file
+  - path: forge_gateway/controllers/__init__.py
+    sha256: d55994c06ed732fec0708abd5cc0ddd75507ff4bbfb6203fd10d68690da3113b
+    size: 38
+    category: file
+  - path: forge_gateway/controllers/agent_controller.py
+    sha256: 1504f61d37ba87a89e68f513caa564e973a1659c8dd0f821c806c212f0bb72f6
+    size: 1947
+    category: file
+  - path: forge_gateway/controllers/playback_controller.py
+    sha256: 0fbc9368cfac7f30da67c77f7907806ac77c69e46098efb55e4615259cf1eee8
+    size: 1543
+    category: file
+  - path: forge_gateway/controllers/policy_controller.py
+    sha256: fd313e13690c7efe254795f36055ef3fdea45f6a8aab4590350753f67df00bbf
+    size: 1224
+    category: file
+  - path: forge_gateway/controllers/record_controller.py
+    sha256: c001662bbcbfc90eda2c751a808b443f66f3b418fa858cd52ad534847029ea6e
+    size: 2372
+    category: file
+  - path: forge_gateway/controllers/runtime_controller.py
+    sha256: a0ae571e36ca8ea6605aab32ce963ae1ce3c3e8170c72f03f3e77ab4ee9ca933
+    size: 3797
+    category: file
+  - path: forge_gateway/controllers/tool_controller.py
+    sha256: a95c9f0fe1f97777f99a6ecad904af68affc98fdc9d8ab63a497c6047ae5ebec
+    size: 16693
+    category: file
+  - path: forge_gateway/controllers/utils.py
+    sha256: ee72b155211db2f246d3849a7d1c80f1c3069db8826058008cef5ff9f9ee6687
+    size: 1015
+    category: file
+  - path: forge_gateway/controllers/websocket_controller.py
+    sha256: 7af3820b05e5edf997e6a8e5ff5f18a77076bf01b2c0466a8a17d6627092bdc8
+    size: 4641
+    category: file
+  - path: forge_gateway/domain/__init__.py
+    sha256: 729c6c3a6499d01f8b62640144680c26b7e80dec91b18847329a780f2fe26c81
+    size: 37
+    category: file
+  - path: forge_gateway/domain/action_invocations.py
+    sha256: 44d3626ef598184855c5df9d0625ffe617d285e09d20881cf7aebf767aeb3c6f
+    size: 15424
+    category: file
+  - path: forge_gateway/domain/action_manifest.py
+    sha256: c8a6849a35dd1fae860744c30fe9911066c07973f0b97128d533ef3942570b95
+    size: 4410
+    category: file
+  - path: forge_gateway/domain/commands.py
+    sha256: a76c576172697aa761f6760be1a7dce2ec1bd9817f4e5a95766b89f8d0ab6aa3
+    size: 1815
+    category: file
+  - path: forge_gateway/domain/node_status.py
+    sha256: 736cdb41ff11f867c7064bf6cc4654184cbfcefa58a56b770eb0c546d8054a80
+    size: 545
+    category: file
+  - path: forge_gateway/domain/sessions.py
+    sha256: 68a66d5b6d4114ba876a3942c00abf9ac7fa8061686055c73281961c2dadba21
+    size: 1251
+    category: file
+  - path: forge_gateway/domain/tool_catalog.py
+    sha256: 3af1304e13ef6b7663da6eeea6120e5677957ae5b68505bcd6d6d6de4369a70f
+    size: 7082
+    category: file
+  - path: forge_gateway/domain/tool_directory.py
+    sha256: f698e9e0456df2877ac57d80d53fe628b577a474412d6992ba45b2a7ae31ca07
+    size: 17271
+    category: file
+  - path: forge_gateway/resources/actions/piper/sam3.md
+    sha256: 07509d31d23d68944188eeb1674047492b07ff2569354e298bd8bfc4d64f6452
+    size: 1860
+    category: file
+  - path: forge_gateway/resources/actions/piper/sam3_policy_core.md
+    sha256: b186b3009ccd0eb2dccd89f6710f6513922d6ae6bdbf7005653427f9890055d6
+    size: 4605
+    category: file
+  - path: forge_gateway/resources/static/app.js
+    sha256: 7fe0f8a181f90a14da22c8f5272c21531f1e9d58edda9ddcf2e78d09b48b162d
+    size: 41895
+    category: file
+  - path: forge_gateway/resources/static/index.html
+    sha256: d84488345acd381fe31e28e1c4263341fad0ab66c831d0bc8c14fdf14cddb9d5
+    size: 10519
+    category: file
+  - path: forge_gateway/resources/static/styles.css
+    sha256: a2d45832b254a98cc8726a167b76001ffdd78b588503e7ff38b28028a4987432
+    size: 21987
+    category: file
+  - path: forge_gateway/services/__init__.py
+    sha256: 954a57e6961f9cf4eaab070cdd969c46c3c48ebbdd71a29647814cdcb6ddd1ba
+    size: 24
+    category: file
+  - path: forge_gateway/services/action_registry.py
+    sha256: a1c6ad1fae35adf0fb91938f332a3b6f7d22e567200988b4c76b3e6bb9c6f70b
+    size: 1298
+    category: file
+  - path: forge_gateway/services/agent_service.py
+    sha256: 41f52c0625dd1938e9023d83068afcc6e89c8db4877efab1fe141a1fb5a709b8
+    size: 14885
+    category: file
+  - path: forge_gateway/services/capability_service.py
+    sha256: befd539fcc4ae96bc98109809e5a387340d9aaab98576483767cf4cb47fa183a
+    size: 1712
+    category: file
+  - path: forge_gateway/services/command_service.py
+    sha256: ff520c966f716e44e915864a3276ae3c5dc38cc9dab046f37d566f71945b75a4
+    size: 4192
+    category: file
+  - path: forge_gateway/services/image_service.py
+    sha256: c8a9a031f7cdaebfd4141f8ab17b8c88d5a5e9b10aec258928722bdbb02713df
+    size: 6706
+    category: file
+  - path: forge_gateway/services/runtime_service.py
+    sha256: 08c8676c470af9c4bd01c722999ce5f8c49f73fff5d82f61f394821245d9bb43
+    size: 25318
+    category: file
+  - path: forge_gateway/services/tool_gateway_service.py
+    sha256: bb4a64ad0a6d23e79c1e8bf6dae99a7c4adb5ce3ffad36ad9ec95a7cb8d8982a
+    size: 90822
+    category: file
+  - path: forge_tool/__init__.py
+    sha256: 06c9c6b830c3771046a14b8d2f4b0dfa7d0b287e3c41cc3ae6ba4f4054dbac91
+    size: 5580
+    category: file
+  - path: forge_tool/_tool_message.py
+    sha256: afd99c1061cc9cfaa35386c2c954aafa00ccb22c5eac42222e5ffb1dcbf70e50
+    size: 14854
+    category: file
+  - path: forge_tool/dora.py
+    sha256: 3612f8bc27288898881f228405ac797058d11e20e48f93c16c94274d02bbb2b1
+    size: 16642
+    category: file
+  - path: forge_tool/endpoint/__init__.py
+    sha256: 2a46324051783e729e508536d43066414dbe866865e2fc4c37d44c8aa2b46f46
+    size: 1603
+    category: file
+  - path: forge_tool/endpoint/exceptions.py
+    sha256: 2dd2d6fe793a5308979d623a4e2e9f90de7861cef365091249526f16646e1408
+    size: 724
+    category: file
+  - path: forge_tool/endpoint/models.py
+    sha256: 5d6b0db2559b0cc513dfb3e592e995377b64ebbb90959a29ee5bb84492c610bc
+    size: 20673
+    category: file
+  - path: forge_tool/endpoint/protocol.py
+    sha256: 5243819dc9eadd561c719fea357c97dba1581eacf29de1bd99613923fbe93329
+    size: 2981
+    category: file
+  - path: forge_tool/handler.py
+    sha256: 59024c7edfb8be659d773293269a0826c8010ee1acc7f4bf39c0b453369f1d97
+    size: 34074
+    category: file
+  - path: forge_tool/py.typed
+    sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
+    size: 0
+    category: file
+  - path: forge_tool/wire/__init__.py
+    sha256: 033d32a508ea857433bb9b4f9f75e4aa5d5ea57b0092b1f182c9c0500bdd0501
+    size: 3960
+    category: file
+  - path: forge_tool/wire/codec.py
+    sha256: 4e80832a524355478ade051ac6dd1931ff374fea4514c29075ba49fafd53bd7d
+    size: 7820
+    category: file
+  - path: forge_tool/wire/descriptor.py
+    sha256: f65b1dd4f0b7689d5b833bc5894d4cf2096c254ade99b49afb281ee4e1b1f506
+    size: 6317
+    category: file
+  - path: forge_tool/wire/envelope.py
+    sha256: 18c9fb9b905238d0a915431f9accc14ff0107262cc3cfc31648251f4d3a8e0e6
+    size: 9584
+    category: file
+  - path: forge_tool/wire/errors.py
+    sha256: 96c938ec1f1084d7128bdeebf3f782c68899af6f4298cb685df6549873a2fd67
+    size: 507
+    category: file
+  - path: forge_tool/wire/factories.py
+    sha256: 24da5817bab5c97316f040b8456170d8eedee25875447cb4831e277d43898f8e
+    size: 14031
+    category: file
+  - path: forge_tool/wire/messages.py
+    sha256: 831862cb31a2f3d30562c407891bf71a1e5266ff22b06574662ddaea69356b41
+    size: 19549
+    category: file
+  backend_url: null
+  direct_download_url: null
+  archive:
+    format: tar.gz
+  todos: null
+  node_id: forge_runtime
+- kind: node_bundle
+  package_key: robodojo-endpoint-0.1.0
+  artifact_id: robodojo-endpoint-0.1.0
+  version: 0.1.0
+  platform: linux
+  arch: x86_64
+  install_root: ~/.PhyAgentOS/forge_runtime/nodes/robodojo_endpoint
+  entrypoints:
+    bench_endpoint: bin/bench_endpoint
+  bucket_key: node-bundles/robodojo_endpoint/0.1.0/robodojo-endpoint-0.1.0-linux-x86_64.tar.gz
+  expected_sha256: 80b7076b04b8abf21712f6f3b11bd9523fde91d5e49bbe3ed202bf17f238e46c
+  node_digest: b9a66fb07b04ee1f76bb8221f68f96b35d1f1ce136d3d6cf967cb7a34da8e159
+  inventory:
+  - path: bench_endpoint/__init__.py
+    sha256: 25a0bc14f808edadbfb636629ab28412009addd8aae96ec7abbf9031139524d2
+    size: 366
+    category: file
+  - path: bench_endpoint/__main__.py
+    sha256: 6aa1969dbad95243ba9dac5d9792e7f3ec4359f0ce27eed96458406dfb6f774f
+    size: 564
+    category: file
+  - path: bench_endpoint/config.py
+    sha256: 8f43dac26ac9401d5d7f1bd47ac36f037506b40126e9b2ec007dd4e2a46432e4
+    size: 1457
+    category: file
+  - path: bench_endpoint/descriptor.py
+    sha256: 4b0e1671bfdd9439a84da1acede00cdc9df36a804660b583ccd206d38f069469
+    size: 2802
+    category: file
+  - path: bench_endpoint/endpoint_lease.py
+    sha256: 0c4dcaf11b30e9ec17a5ef87ad958177e121455c65e18f94f4f1f4f24945e2d4
+    size: 2600
+    category: file
+  - path: bench_endpoint/endpoints.py
+    sha256: 88f103821a6f5bb65e2412bf3f4908b50029607f2ff5b40527073deeb25f8c94
+    size: 4851
+    category: file
+  - path: bench_endpoint/invocations.py
+    sha256: c086bfb19ad4e37ce638ca7fa790d9b2ac383b2ab7a16cc6f9354973b70dd04f
+    size: 7004
+    category: file
+  - path: bench_endpoint/peer.py
+    sha256: 46579c0806e9577594231ec95e3885e6d24f7852119c70b90dea910de1cba723
+    size: 9187
+    category: file
+  - path: bench_endpoint/runner.py
+    sha256: bb45ae890790e578d3e0b2483325383bc3a4452ea75e151ce8cbd56177c74c29
+    size: 5664
+    category: file
+  - path: bench_endpoint/wire.py
+    sha256: b7f1df988ca1373bbbf4d07aee79b230db7e86950a7b1efc8f5bf5a1d1c1a5fd
+    size: 3812
+    category: file
+  - path: bin/bench_endpoint
+    sha256: 37cabb6b35214691fcbb66c2354bdf965be3ca599a9004af6750853340459ba0
+    size: 373
+    category: executable
+  - path: forge_tool/__init__.py
+    sha256: 06c9c6b830c3771046a14b8d2f4b0dfa7d0b287e3c41cc3ae6ba4f4054dbac91
+    size: 5580
+    category: file
+  - path: forge_tool/_tool_message.py
+    sha256: afd99c1061cc9cfaa35386c2c954aafa00ccb22c5eac42222e5ffb1dcbf70e50
+    size: 14854
+    category: file
+  - path: forge_tool/dora.py
+    sha256: 3612f8bc27288898881f228405ac797058d11e20e48f93c16c94274d02bbb2b1
+    size: 16642
+    category: file
+  - path: forge_tool/endpoint/__init__.py
+    sha256: 2a46324051783e729e508536d43066414dbe866865e2fc4c37d44c8aa2b46f46
+    size: 1603
+    category: file
+  - path: forge_tool/endpoint/exceptions.py
+    sha256: 2dd2d6fe793a5308979d623a4e2e9f90de7861cef365091249526f16646e1408
+    size: 724
+    category: file
+  - path: forge_tool/endpoint/models.py
+    sha256: 5d6b0db2559b0cc513dfb3e592e995377b64ebbb90959a29ee5bb84492c610bc
+    size: 20673
+    category: file
+  - path: forge_tool/endpoint/protocol.py
+    sha256: 5243819dc9eadd561c719fea357c97dba1581eacf29de1bd99613923fbe93329
+    size: 2981
+    category: file
+  - path: forge_tool/handler.py
+    sha256: 59024c7edfb8be659d773293269a0826c8010ee1acc7f4bf39c0b453369f1d97
+    size: 34074
+    category: file
+  - path: forge_tool/py.typed
+    sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
+    size: 0
+    category: file
+  - path: forge_tool/wire/__init__.py
+    sha256: 033d32a508ea857433bb9b4f9f75e4aa5d5ea57b0092b1f182c9c0500bdd0501
+    size: 3960
+    category: file
+  - path: forge_tool/wire/codec.py
+    sha256: 4e80832a524355478ade051ac6dd1931ff374fea4514c29075ba49fafd53bd7d
+    size: 7820
+    category: file
+  - path: forge_tool/wire/descriptor.py
+    sha256: f65b1dd4f0b7689d5b833bc5894d4cf2096c254ade99b49afb281ee4e1b1f506
+    size: 6317
+    category: file
+  - path: forge_tool/wire/envelope.py
+    sha256: 18c9fb9b905238d0a915431f9accc14ff0107262cc3cfc31648251f4d3a8e0e6
+    size: 9584
+    category: file
+  - path: forge_tool/wire/errors.py
+    sha256: 96c938ec1f1084d7128bdeebf3f782c68899af6f4298cb685df6549873a2fd67
+    size: 507
+    category: file
+  - path: forge_tool/wire/factories.py
+    sha256: 24da5817bab5c97316f040b8456170d8eedee25875447cb4831e277d43898f8e
+    size: 14031
+    category: file
+  - path: forge_tool/wire/messages.py
+    sha256: 831862cb31a2f3d30562c407891bf71a1e5266ff22b06574662ddaea69356b41
+    size: 19549
+    category: file
+  backend_url: null
+  direct_download_url: null
+  archive:
+    format: tar.gz
+  todos: null
+  node_id: robodojo_endpoint
+- kind: node_bundle
+  package_key: robodojo-benchmark-0.1.0
+  artifact_id: robodojo-benchmark-0.1.0
+  version: 0.1.0
+  platform: linux
+  arch: x86_64
+  install_root: ~/.PhyAgentOS/forge_runtime/nodes/robodojo_benchmark
+  entrypoints:
+    robodojo_benchmark: bin/robodojo_benchmark
+    robodojo_policy: bin/robodojo_policy
+  bucket_key: node-bundles/robodojo_benchmark/0.1.0/robodojo-benchmark-0.1.0-linux-x86_64.tar.gz
+  expected_sha256: null
+  node_digest: e48cf4e838ba73512ac1eb0c144ea17ea177bb56ad41e524c8332f9a38c19ff2
+  inventory:
+  - path: bin/robodojo_benchmark
+    sha256: 6a4e075963c837a98d05f755f6d52a1837e79450b8a673e3430a98a24aa969fb
+    size: 1092
+    category: executable
+  - path: bin/robodojo_policy
+    sha256: 403261c5274b0c545fc97cd85ad5df698161f9d83fd233b68be3e157c61c77bb
+    size: 944
+    category: executable
+  - path: robodojo/.gitignore
+    sha256: 78902f790bbf2718606af332dfb97847652c5b38446c13607c1b27020926a29c
+    size: 124
+    category: file
+  - path: robodojo/README.md
+    sha256: 25c80e046c6aa6e749943370e6045a5861e1cc74250aa4c07aac6effa248a3fc
+    size: 5302
+    category: file
+  - path: robodojo/_core/__init__.py
+    sha256: ef1a5aa477c1cb5b1eb663e799fd08f09329c952b7f9dd41643bfe60e479db25
+    size: 53
+    category: file
+  - path: robodojo/_core/adapter.py
+    sha256: 93240ca820d7d356ae720ae41d5d31216dcc2e6bd7b3fa1230b42b51b8563237
+    size: 4433
+    category: file
+  - path: robodojo/_core/config.py
+    sha256: 9bb75d07bd64e7c26283af5da73fa6c7630079292ff7d6036b84afe514d5864c
+    size: 5687
+    category: file
+  - path: robodojo/_core/node.py
+    sha256: 031142bce97f56876c0e0ab1921991b171b8852a4a0d0b64cedd93edd8b65fb1
+    size: 12342
+    category: file
+  - path: robodojo/_core/result.py
+    sha256: 88ee7589b46c43a8365c4d63eec0ad5295e5fa7b168d0746b2ae2bd6327d42b9
+    size: 5173
+    category: file
+  - path: robodojo/_core/runner.py
+    sha256: f2a6482e50b0535df17914abd83b2958443032156bcfb71189fb5d2af4438bea
+    size: 18288
+    category: file
+  - path: robodojo/_core/session.py
+    sha256: 21c43500383464572873435db1d67403c8209e92b6d429960d34f20db7796cdc
+    size: 10534
+    category: file
+  - path: robodojo/_core/video.py
+    sha256: 0343107afc90108761ba775c1a815bff0cccb24482d5c563a9774d9ea495ba2f
+    size: 3261
+    category: file
+  - path: robodojo/adapters/__init__.py
+    sha256: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
+    size: 0
+    category: file
+  - path: robodojo/adapters/robodojo_adapter.py
+    sha256: 6d4f119784d6897a041c52c511d3765e57e022e9961ae2efa48869eeb34abd48
+    size: 28254
+    category: file
+  - path: robodojo/config.example.yaml
+    sha256: 41f37e10d76eb59bdb03c7a76e7dd0e51d036d9689720017f3f36860949023d2
+    size: 1313
+    category: file
+  - path: robodojo/config.py
+    sha256: 5a745ceff770876a87c727a069e901de983c2c71fb745740c92744b072f67be7
+    size: 3069
+    category: file
+  - path: robodojo/dataflow.example.yaml
+    sha256: 90ee1299028caef52d0dd2a275a26c8154adbde8ffb23923c0ba57c3f5d6cf2d
+    size: 2504
+    category: file
+  - path: robodojo/main.py
+    sha256: 918ba94b61c7fdba77e20ff80229d528d3f35edf07e359c1a6826ddaaef8175b
+    size: 4280
+    category: file
+  - path: robodojo/policy_node.py
+    sha256: 310569e1f4315119db126bbe69a2c10b1ea66d58e7435d94391d7311bb9857f0
+    size: 14638
+    category: file
+  - path: robodojo/pyproject.toml
+    sha256: e53220f19b584c83a38fa9f205df0f782ac1d7c6663d6f74545faf677ad9078f
+    size: 607
+    category: file
+  backend_url: null
+  direct_download_url: null
+  archive:
+    format: tar.gz
+  todos: "Candidate: built from framework/forge_runtime packages/nodes/benchmarks/robodojo; not uploaded yet."
+  node_id: robodojo_benchmark
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/bench_endpoint.yaml b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/bench_endpoint.yaml
new file mode 100644
index 000000000..1e094f55f
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/bench_endpoint.yaml
@@ -0,0 +1,13 @@
+# bench_endpoint 节点配置（g05 profile，与 build_tower 相同）。
+# 连续多久收不到 bench_report 就把在跑的 invocation 判成 unknown。
+#
+# ⚠️ 必须大于 benchmark 节点一次最长的阻塞调用（episode 启动时 reset_episode 要
+# 加载 Isaac Sim 场景，整个循环阻塞几十秒），不是心跳周期的几倍。太小会把正常的
+# 场景加载误判成掉线 → invocation 被写成终态 unknown，然后真结果到达时被丢弃，
+# 仿真器还在烧 GPU。
+heartbeat_timeout_s: 180.0
+
+# 保留多少条历史 invocation 供对账。
+retention: 8
+
+log_level: INFO
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/config.yaml b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/config.yaml
new file mode 100644
index 000000000..f5b581a53
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/config.yaml
@@ -0,0 +1,29 @@
+# benchmark 节点的服务模式配置(robodojo g05 profile) - 可移植版本。
+# result_dir / record_dir / device_id 允许被环境变量覆盖(见 main.py)。
+benchmark: robodojo
+suite: put_bottles_into_dustbin
+
+env_cfg_type: arx_x5
+device_id: 0          # 可被 ROBODOJO_DEVICE_ID 覆盖
+
+serve: true
+init_states_per_task: 1
+num_runs: 1
+seed: 0
+max_steps: 700
+
+max_total_episodes: 20
+max_max_steps: 2000
+
+action_type: joint
+policy_id: G05
+additional_info: forge
+
+camera_map:
+  image/head: cam_head
+  image/left_wrist: cam_left_wrist
+  image/right_wrist: cam_right_wrist
+
+# 默认相对路径(解析到 bundle 内源码目录); 正式运行请用环境变量覆盖。
+record_dir: results/robodojo/videos
+result_dir: results/robodojo
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/dataflow.yaml b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/dataflow.yaml
new file mode 100644
index 000000000..a5497d5c3
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/dataflow.yaml
@@ -0,0 +1,82 @@
+# 完整链路: gateway -> bench_endpoint -> robodojo_benchmark -> policy(G05) -> Isaac Sim
+# 职责边界: benchmark/policy 的实现由 forge_runtime 的 Node 包维护, 本 Skill 只带
+# 描述、清单、g05 profile、dataflow、安装脚本与文档。四个 node 一律走已安装 Node 的
+# 入口 (${FORGE_RUNTIME_BIN}/<entrypoint>), Skill 内不再存放节点源码或 launcher。
+# 运行环境经 required_environment 由宿主提供, 并在 dora start 时以 ${VAR} 展开写入
+# 各 node env。
+nodes:
+  - id: gateway
+    path: ${FORGE_RUNTIME_BIN}/gateway
+    args: --config ${PAOS_SKILL_ROOT}/profiles/g05/gateway.yaml
+    env:
+      FORGE_GATEWAY_PYTHON: ${FORGE_GATEWAY_PYTHON}
+    inputs:
+      tick: dora/timer/millis/50
+      bench_tool_in: bench_endpoint/tool_out
+    outputs:
+      - tool_response
+      - bench_tool_out
+  - id: bench_endpoint
+    path: ${FORGE_RUNTIME_BIN}/bench_endpoint
+    args: --config ${PAOS_SKILL_ROOT}/profiles/g05/bench_endpoint.yaml
+    env:
+      BENCH_ENDPOINT_PYTHON: ${BENCH_ENDPOINT_PYTHON}
+    inputs:
+      tick: dora/timer/millis/50
+      tool_in: gateway/bench_tool_out
+      bench_report: robodojo_benchmark/bench_report
+    outputs:
+      - tool_out
+      - bench_cmd
+  - id: robodojo_benchmark
+    path: ${FORGE_RUNTIME_BIN}/robodojo_benchmark
+    env:
+      BENCHMARK_CONFIG: ${PAOS_SKILL_ROOT}/profiles/g05/config.yaml
+      PATH: ${PATH}
+      ROBODOJO_ENV: ${ROBODOJO_ENV}
+      ROBODOJO_ROOT: ${ROBODOJO_ROOT}
+      ROBODOJO_CACHE_ROOT: ${ROBODOJO_CACHE_ROOT}
+      ROBODOJO_DEVICE_ID: ${ROBODOJO_DEVICE_ID}
+      ROBODOJO_RESULT_DIR: ${ROBODOJO_RESULT_DIR}
+      CUDA_HOME: ${CUDA_HOME}
+      CUDA_VISIBLE_DEVICES: ${CUDA_VISIBLE_DEVICES}
+    inputs:
+      action: policy/action
+      tick: dora/timer/millis/50
+      bench_cmd:
+        source: bench_endpoint/bench_cmd
+        queue_size: 16
+    outputs:
+      - proprio_state
+      - image/head
+      - image/left_wrist
+      - image/right_wrist
+      - policy_command
+      - benchmark_status
+      - benchmark_result
+      - bench_report
+  - id: policy
+    path: ${FORGE_RUNTIME_BIN}/robodojo_policy
+    env:
+      ROBODOJO_ENV: ${ROBODOJO_ENV}
+      XPOLICYLAB_ROOT: ${XPOLICYLAB_ROOT}
+      POLICY_NAME: ${POLICY_NAME}
+      ENV_CFG_TYPE: ${ENV_CFG_TYPE}
+      ACTION_TYPE: ${ACTION_TYPE}
+      TASK_NAME: ${TASK_NAME}
+      G05_CKPT_PATH: ${G05_CKPT_PATH}
+      G05_PROCESSOR_PATH: ${G05_PROCESSOR_PATH}
+      CUDA_HOME: ${CUDA_HOME}
+      CUDA_VISIBLE_DEVICES: ${CUDA_VISIBLE_DEVICES}
+      ROBODOJO_CACHE_ROOT: ${ROBODOJO_CACHE_ROOT}
+    inputs:
+      tick: dora/timer/millis/20
+      policy_command:
+        source: robodojo_benchmark/policy_command
+        queue_size: 32
+      proprio_state: robodojo_benchmark/proprio_state
+      image/head: robodojo_benchmark/image/head
+      image/left_wrist: robodojo_benchmark/image/left_wrist
+      image/right_wrist: robodojo_benchmark/image/right_wrist
+    outputs:
+      - action
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/env.template.sh b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/env.template.sh
new file mode 100644
index 000000000..7572d7eb9
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/env.template.sh
@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+# robodojo / g05 部署环境模板。
+#
+# 用法：
+#   cp env.template.sh env.<host>.sh
+#   把下面的 /path/to/... 换成目标机真实值（不要提交带个人路径的副本）
+#   source env.<host>.sh
+#   python scripts/precheck_robodojo_env.py   # GPU-free 预检
+#
+# 注意：只包含本 Skill 需要的必要变量，不要用 `env | sort` 导出全量环境。
+# 值里不要带引号包住路径（避免把路径当命令）。
+
+# ---------- 工具协议（forge_gateway / forge_tool 源码经 PYTHONPATH 使用） ----------
+# forge_bench 里 install 了版本化依赖 forge-msgs==1.0.1、dora-rs==0.4.1。
+export FORGE_GATEWAY_PYTHON=/path/to/forge_bench/python
+# NOTE(node bundles): forge_gateway / forge_tool are shipped inside the installed
+# forge_runtime + robodojo_endpoint Node bundles; FORGE_PKGS_ROOT is NOT needed.
+export BENCH_ENDPOINT_PYTHON=/path/to/forge_bench/python
+# NOTE(node bundles): bench_endpoint is shipped inside the robodojo_endpoint Node bundle;
+# BENCH_ENDPOINT_SRC is NOT needed.
+
+# ---------- RoboDojo / XPolicyLab ----------
+# ROBODOJO_ENV 需 python 3.11(3.11.x)、torch 2.7.0+cu128、isaaclab 0.54.3、
+# isaacsim 5.1.0.0、opencv-python(cv2)。benchmark 录像用 cv2，不需要 ffmpeg。
+export ROBODOJO_ENV=/path/to/robodojo/envs/RoboDojo
+export ROBODOJO_ROOT=/path/to/robodojo/repo              # 需含 env/ task/ utils/ src/
+export XPOLICYLAB_ROOT=/path/to/robodojo/repo/XPolicyLab # G05 adapter 在 policy/G05
+export ROBODOJO_CACHE_ROOT=/path/to/robodojo/cache       # 派生 HF_HOME/MODELSCOPE_CACHE/...；不设则跳过
+
+# ---------- G05 模型 / processor ----------
+export G05_CKPT_PATH=/path/to/.../checkpoints/checkpoint   # 约 33.9GB，可读
+export G05_PROCESSOR_PATH=/path/to/G05/hf_processor        # 需含 config.json/tokenizer.json/tokenizer_config.json 等
+
+# ---------- policy 选择（profile 固定） ----------
+export POLICY_NAME=G05
+export ENV_CFG_TYPE=arx_x5
+export ACTION_TYPE=joint
+export TASK_NAME=put_bottles_into_dustbin
+
+# ---------- CUDA / GPU ----------
+# CUDA_HOME 需含 bin/、lib64/。CUDA_VISIBLE_DEVICES 决定策略进程可见 GPU；
+# benchmark 的 ROBODOJO_DEVICE_ID 相对“当前可见集合”编号，必须 < 可见 GPU 个数。
+export CUDA_HOME=/path/to/cuda
+export CUDA_VISIBLE_DEVICES=0,1
+export ROBODOJO_DEVICE_ID=0
+export ROBODOJO_RESULT_DIR=/path/to/result_dir            # 结果落盘目录，需可写
+
+# 可选：录像目录；不设则用 profile 默认相对路径。
+# export ROBODOJO_RECORD_DIR=/path/to/record_dir
+# dora CLI lives in the bin dir above; paos skill start needs dora on PATH.
+export PATH="/path/to/forge_bench/bin:$PATH"
+export PAOS_SRC_ROOT=/path/to/paos_src/phyagentos
+# Which python runs the PAOS CLI (paos env); used by scripts/accept_robodojo_skill.sh.
+export PAOS_PY=/path/to/paos/bin/python
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/gateway.yaml b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/gateway.yaml
new file mode 100644
index 000000000..da0c2313a
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/profiles/g05/gateway.yaml
@@ -0,0 +1,188 @@
+# Forge Gateway 配置：把 bench_endpoint 暴露成两个 Tool。
+#
+# 与 build_tower 的区别：port 改 19011，避免端口冲突。specs / providers 的
+# 结构性约定完全相同（endpoint_id、input_id/output_id 与 dataflow 通道一一对应，
+# 对不上会静默失败，tool 永远 ready:false）。
+
+host: 127.0.0.1
+port: 19011
+
+# benchmark 场景没有关节流也没有图像流经过网关，两个都留空。
+joint_order: []
+image_input_ids: []
+
+readiness:
+  require_proprio_state: false
+  require_images: false
+
+agent:
+  enabled: false
+
+tools:
+  enabled: true
+  lease_ttl_ms: 3000
+  # ⚠️ 每次调用的 deadline，Query 与 Action 共用一个。按「一批最长跑多久」来定，
+  # 上限由 config.yaml 的 max_total_episodes 兜着。
+  invoke_timeout_ms: 3600000
+  request_input_id: tool_request
+  response_output_id: tool_response
+
+  specs:
+    - tool_id: bench.describe
+      implementation_id: robodojo.benchmark
+      endpoint_id: bench.runner
+      operation: describe
+      semantics: query
+      description: >-
+        Return the live benchmark inventory: available tasks, per-task init state
+        counts, default parameters, and episode budget limits. Call this before
+        bench.run; never assume task ids or counts.
+      input_schema:
+        type: object
+        additionalProperties: false
+        properties: {}
+      output_schema:
+        type: object
+        additionalProperties: false
+        required:
+          - benchmark
+          - suite
+          - policy_id
+          - num_tasks
+          - tasks
+          - defaults
+          - limits
+        properties:
+          benchmark: {type: string}
+          suite: {type: string}
+          policy_id: {type: string}
+          env_cfg_type: {type: string}
+          action_type: {type: string}
+          num_tasks: {type: integer, minimum: 1}
+          tasks:
+            type: array
+            items:
+              type: object
+              additionalProperties: false
+              required: [task_id, num_init_states]
+              properties:
+                task_id: {type: integer, minimum: 0}
+                name: {type: string}
+                language: {type: string}
+                num_init_states: {type: integer, minimum: 1}
+          joint_names:
+            type: array
+            items: {type: string}
+          camera_ids:
+            type: array
+            items: {type: string}
+          defaults:
+            type: object
+            additionalProperties: false
+            required: [num_runs, seed, max_steps]
+            properties:
+              num_runs: {type: integer, minimum: 1}
+              seed: {type: integer}
+              max_steps: {type: integer, minimum: 1}
+          limits:
+            type: object
+            additionalProperties: false
+            properties:
+              max_total_episodes: {type: integer, minimum: 0}
+              max_max_steps: {type: integer, minimum: 0}
+      readiness: []
+
+    - tool_id: bench.run
+      implementation_id: robodojo.benchmark
+      endpoint_id: bench.runner
+      operation: run
+      semantics: action
+      description: >-
+        Run one benchmark batch (task_ids x init_state_ids x num_runs) and return
+        the aggregate score. Long running: accepted immediately, poll status and
+        result. Policy weights, suite, and device (G05 / put_bottles_into_dustbin)
+        are fixed by the profile and are not selectable here.
+      input_schema:
+        type: object
+        additionalProperties: false
+        required: [task_ids, init_state_ids, num_runs, seed, max_steps]
+        properties:
+          task_ids:
+            description: Task ids from bench.describe; null means every task.
+            oneOf:
+              - type: "null"
+              - type: array
+                minItems: 1
+                items: {type: integer, minimum: 0}
+          init_state_ids:
+            description: Init state ids; null means every init state shared by all selected tasks.
+            oneOf:
+              - type: "null"
+              - type: array
+                minItems: 1
+                items: {type: integer, minimum: 0}
+          num_runs: {type: integer, minimum: 1}
+          seed: {type: integer}
+          max_steps:
+            description: >-
+              Hard cap on control steps. The episode is terminated as timed_out if
+              this many steps are reached before the environment terminates. For
+              RoboDojo the task's native step_lim defines its normal end (success /
+              partial / failed); set this to at least the task horizon (from
+              bench.describe defaults) to allow completion, or lower to force an
+              explicit step-limit stop. Both max_steps and num_steps count control
+              steps (policy actions).
+            type: integer
+            minimum: 1
+      output_schema:
+        type: object
+        additionalProperties: false
+        required:
+          - status
+          - success_rate
+          - successes
+          - total_episodes
+          - completed_episodes
+          - result_path
+        properties:
+          status: {enum: [succeeded, failed, cancelled]}
+          primary_metric:
+            type: object
+            properties:
+              name: {type: string}
+              value: {type: [number, "null"]}
+          success_rate: {type: [number, "null"], minimum: 0, maximum: 1}
+          successes: {type: integer, minimum: 0}
+          total_episodes: {type: integer, minimum: 0}
+          completed_episodes: {type: integer, minimum: 0}
+          valid_episodes: {type: integer, minimum: 0}
+          timed_out_episodes: {type: integer, minimum: 0}
+          num_steps: {type: integer, minimum: 0}
+          mean_policy_latency_ms: {type: [number, "null"]}
+          elapsed_s: {type: [number, "null"]}
+          obs_republished:
+            description: Non-zero means observations were re-sent; the score is still valid.
+            type: [integer, "null"]
+          result_path: {type: [string, "null"]}
+          episodes:
+            type: array
+            items:
+              type: object
+              additionalProperties: false
+              properties:
+                task_id: {type: integer}
+                init_state_id: {type: integer}
+                run_index: {type: integer}
+                seed: {type: [integer, "null"]}
+                success: {type: boolean}
+                termination: {type: string}
+                error_code: {type: [string, "null"]}
+          episodes_truncated:
+            description: True when the batch was too large to inline; read result_path instead.
+            type: boolean
+      readiness: []
+
+  providers:
+    - endpoint_id: bench.runner
+      input_id: bench_tool_in
+      output_id: bench_tool_out
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/scripts/check_bundle_inventory.py b/PhyAgentOS/skills/forge-skill/robodojo/scripts/check_bundle_inventory.py
new file mode 100644
index 000000000..62192dedd
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/scripts/check_bundle_inventory.py
@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+"""Verify the published index template against the files in this Skill tree.
+
+``docs/paos-forge-packages.template.yaml`` records a ``sha256`` and ``size`` for
+every Skill file so a maintainer can inspect or register the bundle without
+unpacking it.  Nothing reads those entries at install time, so editing any Skill
+file makes them stale silently.  Run this before publishing:
+
+    python3 scripts/check_bundle_inventory.py --check   # exit 1 on drift
+    python3 scripts/check_bundle_inventory.py --write   # rewrite the inventory
+
+``--write`` regenerates only the Skill inventory block; the header comments and
+the node entries keep their formatting.  Categories are re-derived from the file
+mode, so run ``--write`` on the pack host: a Windows checkout has no executable
+bit, and ``--write`` there keeps the categories already recorded instead.
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import os
+import sys
+from pathlib import Path
+
+SKILL_ROOT = Path(__file__).resolve().parent.parent
+TEMPLATE = SKILL_ROOT / "docs" / "paos-forge-packages.template.yaml"
+# The template cannot hash itself, and the packager generates archive-manifest.json.
+SKIPPED = {"docs/paos-forge-packages.template.yaml", "archive-manifest.json"}
+# Windows checkouts do not carry the executable bit.
+CHECK_CATEGORY = os.name != "nt"
+
+
+def sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(1 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def filesystem_category(path: Path) -> str:
+    return "executable" if path.stat().st_mode & 0o111 else "file"
+
+
+def actual_files(root: Path) -> list[dict[str, object]]:
+    found = [
+        {
+            "path": path.relative_to(root).as_posix(),
+            "sha256": sha256_file(path),
+            "size": path.stat().st_size,
+            "category": filesystem_category(path),
+        }
+        for path in root.rglob("*")
+        if path.is_file() and path.relative_to(root).as_posix() not in SKIPPED
+    ]
+    return sorted(found, key=lambda item: str(item["path"]))
+
+
+def skill_bounds(lines: list[str]) -> tuple[int, int]:
+    start = next(
+        (index for index, line in enumerate(lines) if line.strip() == "- kind: skill_bundle"),
+        None,
+    )
+    if start is None:
+        raise SystemExit(f"{TEMPLATE} has no skill_bundle entry")
+    end = next(
+        (index for index in range(start + 1, len(lines)) if lines[index].startswith("- kind:")),
+        len(lines),
+    )
+    return start, end
+
+
+def recorded_entries(lines: list[str], start: int, end: int) -> dict[str, dict[str, str]]:
+    entries: dict[str, dict[str, str]] = {}
+    current: str | None = None
+    for line in lines[start:end]:
+        stripped = line.strip()
+        if stripped.startswith("- path: "):
+            current = stripped[len("- path: ") :].strip()
+            entries[current] = {}
+        elif current and ":" in stripped:
+            key, _, value = stripped.partition(":")
+            if key in {"sha256", "size", "category"}:
+                entries[current][key] = value.strip()
+    return entries
+
+
+def render_block(expected: list[dict[str, object]], recorded: dict[str, dict[str, str]]) -> list[str]:
+    lines = ["  inventory:"]
+    for item in expected:
+        path = str(item["path"])
+        lines.append(f"  - path: {path}")
+        lines.append(f"    sha256: {item['sha256']}")
+        lines.append(f"    size: {item['size']}")
+        # Keep a recorded category on hosts that cannot report the mode; otherwise
+        # derive it from the file so a Linux pack host stays authoritative.
+        category = item["category"] if CHECK_CATEGORY else recorded.get(path, {}).get(
+            "category", item["category"]
+        )
+        lines.append(f"    category: {category}")
+    return lines
+
+
+def rewrite(lines: list[str], start: int, end: int, block: list[str]) -> list[str]:
+    output = list(lines[:start])
+    index = start
+    while index < end:
+        line = lines[index]
+        if line.strip() == "inventory:":
+            output.extend(block)
+            index += 1
+            while index < end:
+                current = lines[index]
+                if current.startswith("  - path:"):
+                    index += 1
+                    while index < end and lines[index].startswith("    "):
+                        index += 1
+                    continue
+                break
+            continue
+        output.append(line)
+        index += 1
+    return output + list(lines[end:])
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--check", action="store_true", help="report drift (default)")
+    parser.add_argument("--write", action="store_true", help="rewrite the Skill inventory")
+    args = parser.parse_args()
+
+    lines = TEMPLATE.read_text(encoding="utf-8").splitlines()
+    start, end = skill_bounds(lines)
+    recorded = recorded_entries(lines, start, end)
+    expected = actual_files(SKILL_ROOT)
+    by_path = {str(item["path"]): item for item in expected}
+
+    problems: list[str] = []
+    for item in expected:
+        path = str(item["path"])
+        entry = recorded.get(path)
+        if entry is None:
+            problems.append(f"missing from inventory: {path}")
+            continue
+        if entry.get("sha256") != item["sha256"] or entry.get("size") != str(item["size"]):
+            problems.append(f"stale hash/size: {path}")
+        elif CHECK_CATEGORY and entry.get("category") != item["category"]:
+            problems.append(f"stale category: {path}")
+    for path in sorted(set(recorded) - set(by_path)):
+        problems.append(f"inventory lists a missing file: {path}")
+
+    if not problems:
+        print(f"ok: {len(expected)} Skill files match {TEMPLATE.name}")
+        return 0
+    for problem in problems:
+        print(problem)
+    if not args.write:
+        print("run with --write to regenerate the Skill inventory")
+        return 1
+    block = render_block(expected, recorded)
+    TEMPLATE.write_text("\n".join(rewrite(lines, start, end, block)) + "\n", encoding="utf-8", newline="")
+    print(f"rewrote the Skill inventory in {TEMPLATE.name}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
\ No newline at end of file
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/scripts/install_robodojo.py b/PhyAgentOS/skills/forge-skill/robodojo/scripts/install_robodojo.py
new file mode 100755
index 000000000..84d8b0037
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/scripts/install_robodojo.py
@@ -0,0 +1,264 @@
+#!/usr/bin/env python3
+"""One-shot robodojo install helper (self-contained).
+
+Reads the unsigned schema-v3 index template shipped next to this script, presigns a fresh
+TOS download URL for every ``bucket_key``, downloads each object once to measure ``size``
+and confirm ``expected_sha256``, writes a generated schema-v3 index to a private temp dir,
+then calls::
+
+    paos skill install robodojo --version 0.1.0 --index <generated-index.yaml>
+
+``paos skill install --index`` rejects an index entry that has no ``sha256``/``size``, which
+is why the template carries ``bucket_key`` + ``expected_sha256`` instead of a link and a
+hand-written size: the helper derives the real values from the object it just downloaded and
+aborts on a hash mismatch.
+
+Requires TOS read permission for the inner bucket plus ``tosutil``; this is a
+credential-enabled route, not a credential-less public one. Requires PyYAML (already a PAOS
+runtime dependency).
+
+Self-contained preparation:
+  * locates the template next to this script (downloaded together from TOS) OR in the repo
+    layout (<script>/../docs/paos-forge-packages.template.yaml);
+  * if ``--paos-config`` is supplied, prepares the isolated instance so the install does NOT
+    depend on a previous acceptance tree: creates <config>.parent/config.json and
+    <config>.parent/py/sitecustomize.py (which calls PhyAgentOS set_config_path), and sets
+    PAOS_CONFIG + PYTHONPATH for the paos child.
+
+Never modifies PAOS integrity checks, never changes the frozen packages, never tracks the
+generated index in git.
+"""
+from __future__ import annotations
+
+import argparse
+import datetime
+import hashlib
+import os
+import shutil
+import subprocess
+import sys
+import tempfile
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+import yaml
+
+BUCKET = "phyagentos-resource-inner"
+TEMPLATE_FIELDS = ("bucket_key", "expected_sha256")
+
+# sitecustomize.py keeps get_data_dir() anchored to the instance dir (config.parent).
+SITECUSTOMIZE = '''import os
+_p = os.environ.get("PAOS_CONFIG")
+if _p:
+    try:
+        from PhyAgentOS.config.loader import set_config_path
+        from pathlib import Path
+        set_config_path(Path(_p).expanduser().resolve())
+    except Exception:
+        pass
+'''
+
+
+def resolve_template(explicit: str | None) -> Path:
+    """Locate the unsigned template. Same-directory (TOS download) first, repo layout second."""
+    if explicit:
+        p = Path(explicit)
+        if not p.is_file():
+            raise SystemExit(f"template not found: {p}")
+        return p
+    here = Path(__file__).resolve().parent
+    candidates = [
+        here / "paos-forge-packages.template.yaml",
+        here.parent / "docs" / "paos-forge-packages.template.yaml",
+    ]
+    for cand in candidates:
+        if cand.is_file():
+            return cand
+    raise SystemExit(
+        "template not found; place paos-forge-packages.template.yaml next to this script "
+        "or pass --template <path>"
+    )
+
+
+def find_tosutil(explicit: str | None) -> str:
+    if explicit:
+        return explicit
+    env = os.environ.get("TOSUTIL")
+    if env:
+        return env
+    found = shutil.which("tosutil")
+    if found:
+        return found
+    raise SystemExit(
+        "tosutil not found. Configure it with --tosutil <path>, "
+        "TOSUTIL=<path>, or put 'tosutil' on PATH. "
+        "It needs read access to the phyagentos-resource-inner bucket."
+    )
+
+
+def find_paos(explicit: str | None) -> str:
+    if explicit:
+        return explicit
+    env = os.environ.get("PAOS_BIN")
+    if env:
+        return env
+    found = shutil.which("paos")
+    if found:
+        return found
+    raise SystemExit(
+        "paos not found. Configure it with --paos <path>, "
+        "PAOS_BIN=<path>, or put 'paos' on PATH."
+    )
+
+
+def prepare_instance(config_path: Path) -> None:
+    """Make an isolated instance so install does not depend on a manually created tree."""
+    config_path = config_path.expanduser().resolve()
+    data_dir = config_path.parent
+    data_dir.mkdir(parents=True, exist_ok=True)
+    if not config_path.exists():
+        config_path.write_text("{}\n", encoding="utf-8")
+    py_dir = data_dir / "py"
+    py_dir.mkdir(parents=True, exist_ok=True)
+    sc = py_dir / "sitecustomize.py"
+    if not sc.exists():
+        sc.write_text(SITECUSTOMIZE, encoding="utf-8")
+
+
+def presign(tosutil: str, key: str, vp: str) -> str:
+    r = subprocess.run([tosutil, "presign", f"tos://{BUCKET}/{key}", f"-vp={vp}"],
+                       capture_output=True, text=True)
+    for line in r.stdout.splitlines():
+        line = line.strip()
+        if line.startswith("https://"):
+            return line
+    raise RuntimeError(f"presign failed for {key}: {r.stdout[:200]} {r.stderr[:200]}")
+
+
+def download_and_hash(url: str, key: str, dest: Path) -> tuple[str, int]:
+    """Download one object and return its (sha256, size); never trusts the link metadata."""
+    try:
+        with urllib.request.urlopen(url, timeout=300) as response, dest.open("wb") as out:
+            shutil.copyfileobj(response, out, length=1 << 20)
+    except urllib.error.HTTPError as exc:
+        raise SystemExit(
+            f"cannot download {key}: HTTP {exc.code}. If this is the post-split candidate, "
+            "the object has not been published yet."
+        ) from exc
+    except OSError as exc:
+        raise SystemExit(f"cannot download {key}: {exc}") from exc
+    digest = hashlib.sha256()
+    size = 0
+    with dest.open("rb") as fh:
+        for chunk in iter(lambda: fh.read(1 << 20), b""):
+            digest.update(chunk)
+            size += len(chunk)
+    return digest.hexdigest(), size
+
+
+def generate_index(document: dict, tosutil: str, vp: str, workdir: Path) -> Path:
+    packages = document.get("packages")
+    if not isinstance(packages, list) or not packages:
+        raise SystemExit("template has no packages list")
+    for item in packages:
+        if not isinstance(item, dict):
+            raise SystemExit("template package entry must be a mapping")
+        key = item.pop("bucket_key", None)
+        expected = item.pop("expected_sha256", None)
+        if not isinstance(key, str) or not key:
+            raise SystemExit(f"template package {item.get('kind')!r} has no bucket_key")
+        url = presign(tosutil, key, vp)
+        local = workdir / Path(key).name
+        sha256, size = download_and_hash(url, key, local)
+        if isinstance(expected, str) and expected:
+            if expected.lower() != sha256:
+                raise SystemExit(
+                    f"integrity check failed for {key}: template expected {expected.lower()}, "
+                    f"object has {sha256}"
+                )
+            print(f"[robodojo-helper] {key}: sha256 {sha256} matches the template")
+        else:
+            print(
+                f"[robodojo-helper] {key}: no expected_sha256 recorded yet "
+                f"(candidate object); recording {sha256}"
+            )
+        item["direct_download_url"] = url
+        item["sha256"] = sha256
+        item["size"] = size
+        local.unlink()
+    document["generated_at"] = datetime.datetime.now(datetime.timezone.utc).strftime(
+        "%Y-%m-%dT%H:%M:%SZ"
+    )
+    out = workdir / "index.yaml"
+    out.write_text(yaml.safe_dump(document, sort_keys=False, allow_unicode=True), encoding="utf-8")
+    return out
+
+
+def load_template(path: Path) -> dict:
+    try:
+        document = yaml.safe_load(path.read_text(encoding="utf-8"))
+    except (OSError, yaml.YAMLError) as exc:
+        raise SystemExit(f"cannot read template {path}: {exc}") from exc
+    if not isinstance(document, dict) or document.get("schema_version") != 3:
+        raise SystemExit(f"template {path} is not a schema_version 3 index")
+    return document
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--template", default=None, help="unsigned schema-v3 template (default: next to script, then repo docs)")
+    ap.add_argument("--tosutil", default=None, help="tosutil path (default: TOSUTIL/PATH)")
+    ap.add_argument("--paos", default=None, help="paos CLI path (default: PAOS_BIN/PATH)")
+    ap.add_argument("--vp", default="1d", help="presign validity, e.g. 1d/24h/1440min/86400s")
+    ap.add_argument("--out", default=None, help="generated index path (default: private temp dir)")
+    ap.add_argument("--gen-only", action="store_true", help="generate index only, do not install")
+    ap.add_argument("--index", default=None, help="use a pre-generated index file instead of generating")
+    ap.add_argument("--skill", default="robodojo")
+    ap.add_argument("--version", default="0.1.0")
+    ap.add_argument("--paos-config", default=os.environ.get("PAOS_CONFIG"))
+    args = ap.parse_args()
+
+    template = resolve_template(args.template)
+
+    if args.index:
+        idx_path = Path(args.index)
+        if not idx_path.is_file():
+            raise SystemExit(f"index not found: {idx_path}")
+    else:
+        document = load_template(template)
+        tosutil = find_tosutil(args.tosutil)
+        workdir = Path(args.out).expanduser().resolve().parent if args.out else Path(
+            tempfile.mkdtemp(prefix="robodojo-index-")
+        )
+        workdir.mkdir(parents=True, exist_ok=True)
+        generated = generate_index(document, tosutil, args.vp, workdir)
+        idx_path = Path(args.out).expanduser().resolve() if args.out else generated
+        if args.out:
+            shutil.move(str(generated), str(idx_path))
+        print(f"[robodojo-helper] index written: {idx_path}")
+        print(f"[robodojo-helper] generated_at={document['generated_at']} presign_vp={args.vp}")
+        print(f"[robodojo-helper] NOTE: download URLs expire after {args.vp}; re-run to refresh.")
+
+    if args.gen_only:
+        print(idx_path)
+        return 0
+
+    paos = find_paos(args.paos)
+    env = dict(os.environ)
+    if args.paos_config:
+        cfg = Path(args.paos_config)
+        prepare_instance(cfg)
+        env["PAOS_CONFIG"] = str(cfg.resolve())
+        instance_py = str(cfg.resolve().parent / "py")
+        existing_pypath = env.get("PYTHONPATH")
+        env["PYTHONPATH"] = instance_py + (os.pathsep + existing_pypath if existing_pypath else "")
+
+    cmd = [paos, "skill", "install", args.skill, "--version", args.version, "--index", str(idx_path)]
+    print(f"[robodojo-helper] running: {' '.join(cmd)}")
+    r = subprocess.run(cmd, env=env)
+    return r.returncode
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/PhyAgentOS/skills/forge-skill/robodojo/skill.yaml b/PhyAgentOS/skills/forge-skill/robodojo/skill.yaml
new file mode 100644
index 000000000..fe0815601
--- /dev/null
+++ b/PhyAgentOS/skills/forge-skill/robodojo/skill.yaml
@@ -0,0 +1,61 @@
+manifest_version: 2
+name: robodojo
+version: 0.1.0
+description: Run the RoboDojo G05 benchmark and return its aggregate score.
+skill_document: SKILL.md
+gateway_url: http://127.0.0.1:19011
+required_tools:
+- bench.describe
+- bench.run
+profiles:
+  g05:
+    dataflow: profiles/g05/dataflow.yaml
+    required_binaries:
+    - gateway
+    - bench_endpoint
+    - robodojo_benchmark
+    - robodojo_policy
+    required_assets:
+    - profiles/g05/gateway.yaml
+    - profiles/g05/bench_endpoint.yaml
+    - profiles/g05/config.yaml
+    required_environment:
+    - FORGE_GATEWAY_PYTHON
+    - BENCH_ENDPOINT_PYTHON
+    - ROBODOJO_ENV
+    - ROBODOJO_ROOT
+    - XPOLICYLAB_ROOT
+    - G05_CKPT_PATH
+    - G05_PROCESSOR_PATH
+    - CUDA_HOME
+    - CUDA_VISIBLE_DEVICES
+    - ROBODOJO_CACHE_ROOT
+    - ROBODOJO_DEVICE_ID
+    - ROBODOJO_RESULT_DIR
+    environment:
+      POLICY_NAME: G05
+      ENV_CFG_TYPE: arx_x5
+      ACTION_TYPE: joint
+      PYTHONNOUSERSITE: '1'
+      OMNI_KIT_ACCEPT_EULA: 'YES'
+artifacts:
+  resolver: registry
+  nodes:
+    forge_runtime:
+      artifact_id: forge-runtime-0.1.0
+      version: 0.1.0
+      platform: linux
+      arch: x86_64
+      digest: cee927eb587f3f958c1803086e59692f6f7ae373477b3cfd15118d01d76c408e
+    robodojo_endpoint:
+      artifact_id: robodojo-endpoint-0.1.0
+      version: 0.1.0
+      platform: linux
+      arch: x86_64
+      digest: b9a66fb07b04ee1f76bb8221f68f96b35d1f1ce136d3d6cf967cb7a34da8e159
+    robodojo_benchmark:
+      artifact_id: robodojo-benchmark-0.1.0
+      version: 0.1.0
+      platform: linux
+      arch: x86_64
+      digest: e48cf4e838ba73512ac1eb0c144ea17ea177bb56ad41e524c8332f9a38c19ff2