NyxID Service Doctor
Use this when a tool call failed with 403 / "not connected" / "service not found", when a skill
skipped a data source, or when someone asks "bot 现在绑了哪些服务 / 为什么搜索用不了".
You (the agent) inspect the real state with typed tools — primarily nyxid_services (list /
discover connected services), nyxid_status (runtime + connection overview), and where visible
nyxid_catalog, nyxid_providers, nyxid_llm_status, nyxid_api_keys. All inspection is
read-only. Quote errors verbatim; never fail the run, never guess.
How to run it
Reproduce the claim. If the user reports a failing skill/tool, identify which service it
needed (e.g. a search/enrichment service, a Lark proxy slug, an LLM provider). If they pasted an
error, extract the service slug and status code from it.
Inspect what is actually connected. nyxid_services list — match by endpoint/label
semantics, NOT by guessing exact slugs (slugs vary; discovery-first is the house rule). Report:
service label, slug, auth kind, and whether it is personally bound or org-shared.
Apply the known diagnosis table:
- Chat works, cron 403s → scheduled runs use a minted key with a narrower service grant
than interactive chat. The fix is binding-side (grant the service to scheduled runs), not the
skill. Say exactly which service needs the grant.
Provider 'X' not connected. Connect at /providers → the LLM upstream is disconnected at
NyxID, not an aevatar bug. Fix at the NyxID dashboard, or switch the preferred LLM route.
- Service visible to you but 403 for another user via the bot → tool credentials are
SENDER-scoped in channel chats: each sender needs the service personally, or the service must
be org-shared. Personal binding by the bot owner does not cover teammates.
- Service simply absent → it was never connected. Prescribe:
nyxid service add --custom
with the right auth method (or the catalog connect flow), then org-share it if the whole team
needs it. Creating an org copy requires re-entering the key (existing keys are not exposed).
- Slug mismatch → the service exists under a different slug than the skill expected;
recommend the skill switch to discovery by label/endpoint instead of hardcoding.
Verify after the fix. Re-run the cheapest read that exercises the same service (or re-run
the original failing skill) and report before/after.
Report. One diagnosis block: symptom → evidence (verbatim) → root cause from the table →
exact next action and WHO can do it (user self-serve vs org admin).
Guardrails
- Inspection only: never create, rebind, or delete services/keys yourself; prescribe the command
and let a human run it. Never echo secrets or full API keys (ids and prefixes only).
- Only use real returned data; if a typed tool is missing in this runtime, note the absence (that
is itself a finding) and fall back to reporting what the error text alone proves.
- Never ask the user for tokens — your NyxID-brokered tools handle all credentials.
1---2name: nyxid-service-doctor3version: "1.0"4description: Diagnose NyxID service and credential problems from chat — "工具 403 / Provider not connected / 服务没绑 / 找不到搜索服务" — using the typed nyxid_* tools to inspect what is actually connected, who owns the credential, and which grant the current run is using, then prescribe the exact fix (connect, org-share, or rebind).5metadata:6 source: https://skillmd.com/skills/chronoaiproject/nyxid-service-doctor7 category: plain8 tag:9 - aevatar10 - s-capability11 - nyxid12 - credentials13 - diagnostics14 - self-service15---1617# NyxID Service Doctor1819Use this when a tool call failed with 403 / "not connected" / "service not found", when a skill20skipped a data source, or when someone asks "bot 现在绑了哪些服务 / 为什么搜索用不了".2122**You (the agent) inspect the real state with typed tools** — primarily `nyxid_services` (list /23discover connected services), `nyxid_status` (runtime + connection overview), and where visible24`nyxid_catalog`, `nyxid_providers`, `nyxid_llm_status`, `nyxid_api_keys`. All inspection is25read-only. Quote errors verbatim; never fail the run, never guess.2627## How to run it28291. **Reproduce the claim.** If the user reports a failing skill/tool, identify which service it30 needed (e.g. a search/enrichment service, a Lark proxy slug, an LLM provider). If they pasted an31 error, extract the service slug and status code from it.32332. **Inspect what is actually connected.** `nyxid_services` list — match by endpoint/label34 semantics, NOT by guessing exact slugs (slugs vary; discovery-first is the house rule). Report:35 service label, slug, auth kind, and whether it is personally bound or org-shared.36373. **Apply the known diagnosis table:**38 - **Chat works, cron 403s** → scheduled runs use a minted key with a narrower service grant39 than interactive chat. The fix is binding-side (grant the service to scheduled runs), not the40 skill. Say exactly which service needs the grant.41 - **`Provider 'X' not connected. Connect at /providers`** → the LLM upstream is disconnected at42 NyxID, not an aevatar bug. Fix at the NyxID dashboard, or switch the preferred LLM route.43 - **Service visible to you but 403 for another user via the bot** → tool credentials are44 SENDER-scoped in channel chats: each sender needs the service personally, or the service must45 be org-shared. Personal binding by the bot owner does not cover teammates.46 - **Service simply absent** → it was never connected. Prescribe: `nyxid service add --custom`47 with the right auth method (or the catalog connect flow), then org-share it if the whole team48 needs it. Creating an org copy requires re-entering the key (existing keys are not exposed).49 - **Slug mismatch** → the service exists under a different slug than the skill expected;50 recommend the skill switch to discovery by label/endpoint instead of hardcoding.51524. **Verify after the fix.** Re-run the cheapest read that exercises the same service (or re-run53 the original failing skill) and report before/after.54555. **Report.** One diagnosis block: symptom → evidence (verbatim) → root cause from the table →56 exact next action and WHO can do it (user self-serve vs org admin).5758## Guardrails5960- Inspection only: never create, rebind, or delete services/keys yourself; prescribe the command61 and let a human run it. Never echo secrets or full API keys (ids and prefixes only).62- Only use real returned data; if a typed tool is missing in this runtime, note the absence (that63 is itself a finding) and fall back to reporting what the error text alone proves.64- Never ask the user for tokens — your NyxID-brokered tools handle all credentials.