Tell the agent how to use catalogs and where outputs go

[AshishKumar4]

What does this change?

Two things the agent is given but never told how to use.

Outputs. writeFile takes a workpiece, so every write lands inside one specific Gadget as one
of its source files. There is no free-standing file path, and nothing in the prompt says so. When an
instruction says "save it to notes.md" the agent has no defined behaviour: sometimes it answers in
chat, sometimes it writes a .md into a Gadget's source tree where it renders as code rather than a
document, sometimes it creates the right output. This states the rule as its own paragraph in the
section that already explains the deployment's formats, directly above the list it refers to.

Catalogs. A resource's discovery catalog is inlined into the system prompt, and truncated is
emitted as a bare JSON key with nothing explaining it. The agent is told to describeBinding a
resource when one is relevant, but never to consult the entries first, and never that a partial list
has a remainder reachable through the binding's own API. So a listed item goes unused unless the user
names it, and a truncated list looks complete.

Why is this obviously correct and trivially verifiable?

Prompt text only: 13 added lines across two functions, no code path changes. The complete effect is
the wording, read in place.

Checklist

Checking every item does not guarantee acceptance. Maintainers determine whether
a pull request meets the contribution policy.

• This is a small, concrete change; it is not a feature, refactor, or low-value cleanup.
• I understand that maintainers decide whether the change is obviously correct and trivially verifiable.
• I have read and followed the contribution guidelines.

[ask-bonk]

github run

[ask-bonk]

@AshishKumar4 Bonk workflow failed. Check the logs for details.

View workflow run · To retry, trigger Bonk again.

[AshishKumar4]

Closing — #208 removed the basis for both halves of this.

The catalog legend explained "truncated":true, but #208 landed the cap at 1000. The library is 116 skills + 1 collection = 117 entries, so that flag can no longer fire for this deployment. Documenting a field that cannot appear isn't worth prompt space.

The output-path paragraph doesn't clear the bar either, as discussed: the section above it already tells the agent to instantiate a format rather than build an equivalent, the motivation came from an external skill corpus that doesn't live in this repo, and there's no eval here to show the agent actually gets it wrong. Arguing from a plausible failure mode isn't enough for the most contended text in the product.