File a bug or enhancement on the agentuity/infra repo when the SDK's behavior is being broken or made worse by something on the platform side (Catalyst, Sandbox runtime, Ion, Gluon, Hadron, region routing, auth surfaces). Use this when an SDK debugging session ends with "this is the platform's fault, not ours" — unhelpful error messages from the API, opaque 5xx responses, behavior that contradicts the docs, missing capabilities the SDK is documented to expose. Always checks for duplicates first and refuses to file if the issue is actually an SDK bug.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
File a bug or enhancement on the agentuity/infra repo when the SDK's behavior is being broken or made worse by something on the platform side (Catalyst, Sandbox runtime, Ion, Gluon, Hadron, region routing, auth surfaces). Use this when an SDK debugging session ends with "this is the platform's fault, not ours" — unhelpful error messages from the API, opaque 5xx responses, behavior that contradicts the docs, missing capabilities the SDK is documented to expose. Always checks for duplicates first and refuses to file if the issue is actually an SDK bug.
File Infra Issue
When the SDK is breaking because of an infra-side problem (bad error responses, missing endpoints, misrouting, unauthorized where authorized was expected, etc.), file a clean, deduplicated issue on agentuity/infra.
When to use this skill
Use it when all of the following are true:
You hit a real failure (CI red, smoke test red, manual repro red), and
The root cause sits on the platform side — Catalyst HTTP API, sandbox runtime images, Ion, Gluon, Hadron, region routing, auth/permission surfaces — not in SDK code, and
The fix has to happen in agentuity/infra (or its downstream services), not in this repo.
Concrete examples that qualify:
The API returns 500 Internal Server Error for a user-input error that should be a structured 4xx.
An endpoint returns 401 Unauthorized when the resource simply doesn't exist (should be 404).
The SDK is documented to use SDK-key auth, but the endpoint silently rejects SDK keys and only accepts CLI tokens.
A sandbox runtime image is missing a binary the docs claim it ships.
An endpoint requires the caller to supply a piece of state (region, project id) that the platform should be able to resolve from existing context.
Error responses are missing context the user needs to debug (no session id, empty stderr, no message body).
Cross-region routing falls back in a confusing way (e.g., wrong host returns 401 instead of 421 misdirected).
When NOT to use this skill
Refuse / suggest a different action if:
The bug is in SDK code (wrong types, bad URL construction, wrong header name on the SDK side, schema mismatch in our zod definitions, etc.) — fix it in this repo instead.
The "issue" is actually a feature request for the SDK that the platform already supports — file it on agentuity/sdk (or just implement it).
The behavior is as documented in the platform's own contract — the SDK is the wrong layer to be surprised by it; either update the SDK to match or update SDK docs.
You only suspect it's infra but haven't confirmed (e.g., didn't curl the endpoint directly, didn't compare against CLI behavior, didn't check whether SDK code is constructing a bad request). Confirm first.
The platform behavior is annoying but intentional (e.g., SDK keys can't list all orgs by design — that's an auth-model decision, not a bug).
If unsure, prefer not filing. Duplicates and noise in agentuity/infra cost the team more than missing a marginal report.
Workflow
1. Confirm it's actually infra
Before doing anything else, answer all of these in writing (in your reasoning, not in the issue):
What does the SDK code do? (Read it. Don't guess.)
What does the platform return? (Reproduce with curl or a minimal script — not via the SDK — so the SDK isn't a confounding variable.)
Does the same call work from another client (CLI, browser, raw curl with a different token)? If yes, what's different?
Is there a documented contract (in docs/, the openapi spec, the platform's own README) that the platform is violating? Or are you the one who made an assumption?
If after answering these you can't write one sentence describing what the platform should do differently, stop. The issue isn't ready.
2. Check for duplicates
Run all three searches before drafting. Different keyword sets catch different existing issues.
# Open issues, broad keyword
gh issue list --repo agentuity/infra --state open --search "<short-keyword-from-symptom>" --limit 20 \
--json number,title,state,labels,updatedAt
# Closed issues, same keyword (catches "fixed but still broken" cases)
gh issue list --repo agentuity/infra --state closed --search "<short-keyword-from-symptom>" --limit 10 \
--json number,title,state,labels,closedAt
# Component-scoped search (catalyst | sandbox | ion | gluon | hadron)
gh issue list --repo agentuity/infra --state all --search "<component>" --limit 30 \
--json number,title,state,updatedAt
For each candidate, run gh issue view <number> --repo agentuity/infra and decide:
Exact dupe — don't file. Add a comment on the existing issue with the new repro, the new failing CI run link, or any extra detail you have. Use gh issue comment <number> --repo agentuity/infra --body-file ....
Related but distinct — file the new one and reference the related issue with Related: agentuity/infra#NNN near the top of the body.
Closed but you're hitting it again — comment on the closed issue with the fresh repro and ask whether it should be reopened. Don't open a new one until that's been answered.
3. Pick a label
Match the symptom to the closest existing label (run gh label list --repo agentuity/infra if unsure):
Symptom
Label
Bad / unhelpful behavior, wrong status codes, broken endpoints
bug
Missing endpoint / capability the SDK needs
enhancement
Sandbox runtime, sandbox API, runtime images
(no specific label exists yet — use bug or enhancement and put Sandbox: in the title)
Catalyst HTTP routing, region routing, auth surface
bug (or enhancement for new endpoints)
Ion-specific
Ion
Gluon-specific
gluon
Hadron-specific
hadron
Production outage right now
outage (only if it's actively broken in prod, not just CI)
If multiple apply, use multiple. Don't invent new labels.
4. Write the issue
Use this structure. Keep it short — engineers who triage these issues are busy. Aim for under 60 lines of body.
<One short paragraph or a numbered list. Concrete: status code, response shape, behavior. Don't write essays — bullet points beat prose here.>
Why this matters
<One short paragraph. What SDK code path / user workflow is broken because of this? Link the failing CI run, the SDK file, the docs page that contradicts the current behavior.>
References
SDK code: <link to file on github.com/agentuity/sdk at the current branch>
Failing run / repro:
Related: agentuity/infra#NNN (only if there's a related-but-distinct issue)
Style rules:
- **No essays.** If the body is over ~60 lines, you're padding.
- **No SDK debugging narrative.** They don't need to read your debugging session — they need the platform-facing facts.
- **No "this is urgent" / "P0" framing** unless it's an actual prod outage. Use the `outage` label instead.
- **Mask tokens.** Replace real `ck_live_...` values with `<ck_live_...>` placeholders.
- **Title format**: `<Component>: <short symptom>`. Examples:
- `Catalyst: DB endpoints should resolve region server-side`
- `Sandbox: unhelpful errors when spawning a missing binary (500 from createJob, empty stderr from execute)`
- `Catalyst: 401 Unauthorized for missing resources should be 404`
### 5. File it
Write the body to a tempfile, then create:
```bash
gh issue create \
--repo agentuity/infra \
--title "<Component>: <short symptom>" \
--body-file /tmp/infra-issue.md \
--label bug # or enhancement, or both
Capture the URL from the response. Reference it in:
The CI workflow comment that gates the failing test off (so future readers see the link)
The relevant SDK source file's comments (so future maintainers see why something looks weird)
Any tracking memory / project doc
6. Loop back into the SDK
If you gated a test off / added a workaround / hardcoded a value because of the platform issue, leave a code comment that:
Explains what's currently broken on the platform side.
Links the infra issue.
Describes how to revert the workaround once the issue is fixed.
Example:
# Disabled: SDK keys can't currently auth against DB /resource endpoints.# See agentuity/infra#120. Re-enable by removing this `if: false` once the# issue is closed.if:false
Scope guard
Before filing, ask one last time: "If a backend engineer reads this issue, will they immediately know what's broken, where it is, and what to change?" If not, the issue isn't ready — keep iterating before filing.
