.NET: [Bug]: skills prompt causes reads of nonexistent example resources

[KOTerra]

Description

AgentSkillsProvider injects a concrete fictional resource path into the model's system instructions:

- Use `read_skill_resource` to read any referenced resources, using the name exactly as listed
   (e.g. `"style-guide"` not `"style-guide.md"`, `"references/FAQ.md"` not `"FAQ.md"`).

In repeated .NET agent runs, the model sometimes interprets references/FAQ.md as an available resource after loading a skill, even though that skill neither references nor contains that file. It then calls read_skill_resource with the example path and receives a missing-resource error.

This occurred across unrelated skills, indicating that the framework-generated system instruction is the common primer rather than any individual SKILL.md.

Expected behavior
Framework instructions should not introduce fictional resource names. They should direct the model to:

1. Read only resources explicitly referenced by the loaded skill.
2. Pass the resource path exactly as written.
3. Never infer or guess resource paths.

Steps to reproduce

1. Configure AgentSkillsProvider with skills that enable resource support.
2. Include a skill that does not reference or contain references/FAQ.md.
3. Run a task that loads that skill and may benefit from supporting resources.
4. Inspect the model's tool calls.
   Depending on model sampling, it may copy references/FAQ.md from the system prompt into a read_skill_resource call.

The model behavior is nondeterministic, but the ungrounded concrete path is deterministically included in every generated prompt when resource instructions are enabled.

Code Sample

var skillsProvider = new AgentSkillsProvider(skillsPath);

// The generated system prompt advertises the fictional path:
// references/FAQ.md

// Observed model-generated call:
// read_skill_resource(
//     skillName: "some-unrelated-skill",
//     resourceName: "references/FAQ.md")

Error Messages / Stack Traces

Error: Resource 'references/FAQ.md' not found in skill 'some-unrelated-skill'.

Package Versions

Microsoft.Agents.AI: 1.2.0

.NET Version

.NET 10

Additional Context

The same concrete example currently appears in both implementations:

• dotnet/src/Microsoft.Agents.AI/Skills/AgentSkillsProvider.cs
• python/packages/core/agent_framework/_skills.py

This is distinct from #7008 / #7031. Those fixed the Python example incorrectly dropping the .md extension. The remaining problem is that any concrete fictional resource name in executable model instructions can be mistaken for an actually available resource.

A safer rule-only instruction would be:

Use `read_skill_resource` only for a resource explicitly referenced by the loaded skill.
Pass its resource path exactly as written in that skill.
Never infer or guess resource paths.

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

The RESOURCE_INSTRUCTIONS constant in python/packages/core/agent_framework/_skills.py (lines 1811-1814) and dotnet/src/Microsoft.Agents.AI/Skills/AgentSkillsProvider.cs (lines 161-162) contains fictional concrete resource paths ("references/FAQ.md", "style-guide") that are injected into every system prompt via the {resource_instructions} placeholder in SKILLS_PROMPT_TEMPLATE. Minimal repro: assert "references/FAQ.md" in RESOURCE_INSTRUCTIONS — passes, confirming the fictional paths are present. Fix requires replacing the concrete examples with rule-only instructions that direct the model to use only resources explicitly listed by the loaded skill.

• Failing test: python/packages/core/tests/core/test_issue_7663_fictional_resource_paths.py
• Files examined: python/packages/core/agent_framework/_skills.py, dotnet/src/Microsoft.Agents.AI/Skills/AgentSkillsProvider.cs, python/packages/core/tests/core/test_skills.py
• Tests run: python/packages/core/tests/core/test_issue_7663_fictional_resource_paths.py
• Current version: 1.14.0

[samrusani]

I can take this as a focused Python/.NET parity fix. I’ll replace the fictional resource-name examples with rule-only guidance that permits only resources explicitly referenced by the loaded skill, preserves the referenced path exactly, and forbids inferred paths. I’ll add focused assertions in the existing skills test suites for both implementations and keep the public API unchanged.

[MilkyWay008]

Good catch, the example paths in RESOURCE_INSTRUCTIONS are the common primer, and since your .NET 1.2.0 report matches the Python side verbatim it's clearly the shared template. Quick way to confirm on your side: dump the rendered system prompt and look at the resource_instructions block, then log the model's read_skill_resource calls to separate hallucinated paths from ones the skill actually lists. Upgrading won't clear it either, the earlier fix (#7008 / PR #7031) only touched the Python .md example. Looks like a contributor already claimed the rule-only rewrite, so tracking that PR is probably the fastest path.

[KOTerra]

Confirmed on .NET 1.2.0. Traces also showed the model treating an allowed-tools identifier as a resource path. The rewrite in #7664 matches our successful workaround; explicitly stating that tool identifiers are not resource paths may also be worth adding. I’ll track the PR and review it to suggest a better way of writing those instructions as we test it further.