update cuopt_on_nemoclaw for nemoclaw 0.0.55#155
Conversation
rgsl888prabhu
left a comment
rgsl888prabhu
left a comment
There was a problem hiding this comment.
I see several skills, is it possible to collapse few into one and move details parts or units under references folder in skills.
Also observed there are several cuOpt specific details, would it make sense to move any non nemo claw specific details to cuOpt any may be there are few which are duplicates.
| @@ -0,0 +1,98 @@ | |||
| --- | |||
| name: always-tool-discovery | |||
There was a problem hiding this comment.
Are we expecting users to copy skills from openclaw-skills and put it in skills folder or it would be automated ?
| @@ -0,0 +1,98 @@ | |||
| --- | |||
| name: always-tool-discovery | |||
| description: READ FIRST in every NemoClaw sandbox session. STOP before "I can't read files", "no read tool", "no exec", "can't list directory", "please paste ls/cat output", or "I only have tool_search/tool_describe/tool_call". Real tools (read, write, exec, edit, process) are NOT missing — tools.toolSearch hides them. Use tool_call tool_search query shell or read, then tool_describe, then tool_call the real tool. Never ask the user to paste filesystem output. Applies to ALL tasks including cuOpt, CSV uploads, and scheduling — run tool_search before claiming any file or shell capability is absent. | |||
There was a problem hiding this comment.
Please refer to cuOpt skill structure so that when we move or add CI, there will not be too many changes.
There was a problem hiding this comment.
And may be we can reduce the length of the description. We can move other details to skills.
|
|
||
| # Sandbox Tool Discovery (NemoClaw / OpenClaw catalog) | ||
|
|
||
| **If you find yourself thinking "I don't have an exec / read / python / |
There was a problem hiding this comment.
I think this can be worded towards positive, like use tool search to look for available options for these.
| The correct response to "my tool list looks too short" is to run | ||
| `tool_search`. It is **not** to: | ||
|
|
||
| - Apologize to the user and ask them to paste shell output. |
There was a problem hiding this comment.
May be this would not be required
| `tool_call`, that is **the expected state** — the real tools are | ||
| reachable via the catalog, not absent. | ||
|
|
||
| ## How the catalog works |
There was a problem hiding this comment.
May be use tool catalog in description
| @@ -0,0 +1,117 @@ | |||
| --- | |||
| name: cuopt-first | |||
| description: MANDATORY in NemoClaw sandbox. If you only see tool_search/tool_describe/tool_call, run tool_search (query read or shell) BEFORE claiming you cannot read files or run commands — see always-tool-discovery. STOP before any schedule/heuristic/draft-plan output. FIRST actions: (1) tool_search→exec if needed, (2) probe_cuopt.py, (3) cuopt-remote-env + smoke with CUOPT_REMOTE_HOST/PORT. Do NOT use from cuopt import milp — MILP is cuopt.linear_programming.problem.Problem with INTEGER vars; see cuopt-python-api. Do NOT wait for user to say "use cuOpt". Triggers: schedule, league, CSV upload, minimize, assign, MILP, doubleheader, ImportError, ModuleNotFoundError, wrong import path. | |||
There was a problem hiding this comment.
Lets keep description small and simple, move details below.
| @@ -0,0 +1,107 @@ | |||
| --- | |||
| name: cuopt-python-api | |||
| description: STOP on ImportError, ModuleNotFoundError, or guessed imports (from cuopt import milp, import cuopt.milp, cuopt.solve). LP/MILP/QP share ONE Python entrypoint cuopt.linear_programming.problem.Problem — MILP uses INTEGER variables, not a separate milp module. With CUOPT_REMOTE_HOST/PORT set before Python starts, remote gRPC MILP works in this sandbox. One failed import is NOT proof cuOpt is unavailable — run the verify command below, then read cuopt-numerical-optimization-api-python. Triggers: scheduling MILP, assignment, import failed, wrong API path. | |||
There was a problem hiding this comment.
Same suggestion for description
2 participants
No description provided.