Graceful Tool Failure Handling skill

[Tim (Timziito)]

When a tool, connector, or API call fails or misbehaves, states exactly what happened and offers a real fallback, instead of continuing as if it succeeded or fabricating the result it would have returned.

[Copilot]

Pull request overview

Adds a new submission to the skills gallery that instructs agents to explicitly acknowledge tool/connector/API failures and provide honest, actionable fallbacks instead of fabricating results.

Changes:

• Introduces the graceful-tool-failure-handling skill (SKILL.md) with concrete instructions and guardrails for handling failed tool calls.
• Adds a human-facing overview (README.md) describing when to use the pattern and how it differs from tool-tracer.
• Registers the submission in the gallery via metadata.json (name/description/platforms/tags/author/version/dates).

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
submissions/graceful-tool-failure-handling/SKILL.md	Adds the agent-facing runtime instructions for handling tool failures honestly and providing fallbacks.
submissions/graceful-tool-failure-handling/README.md	Adds a human-facing explanation and cross-references to related patterns/skills.
submissions/graceful-tool-failure-handling/metadata.json	Adds gallery metadata for discovery, attribution, and platform targeting.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[Adriana Trujillo (adrianatruji)]

Appreciate this, Tim (@Timziito). Transparency into the agent's tool trace is definitely a demanded capability.

Two comments. The README notes in other skills such as it-support-ticket-agent and company-memory-builder (they already do it for their domains) suggest that this could be baked into the relevant skill/process at hand rather than as a separate process. The broad scope of the trigger ("whenever a tool call fails/acts unexpected/... before deciding how to respond") gets in between the agent's own reasoning loop on every hiccup, which could be counterproductive.

The challenge behind the idea (CoT/tool trace visibility) is totally relevant though, so a suggestion can be to reframe it around that (e.g. triggered on a specific intent: "describe what failed / what did you actually do"). Thank you!