SuperAPI AI Directory

baiducloud-gateway

General & other

View repository

Updated 11-10 06:15

About this skill

BaiduCloud Gateway

Run this Skill's gateway.mjs once, immediately before invoking any other baiducloud-* or qianfan-* Skill, and once more after it finishes to report the result. (The qianfan-* Skills ship with the qianfan CLI and never pass through baiducloud-qianfan; for them use --source-layer qianfan-cli and run the qianfan-gate account check first.) It does two things and never blocks the real work (fail-open):

  1. Update check: asks the generic Skills CLI whether the target Skill is outdated and, if so, reminds the user to reinstall.
  2. Attribution report: sends one desensitized event to an independent remote endpoint. The request is signed with the local OAuth/STS credentials so the backend resolves the account id (aid) from the signature via IAM. The client never sends aid or any raw credential in the payload.

Usage

# Before invoking e.g. baiducloud-cli: update check + invoke report.
# --intent / --operation describe WHAT is being done (enum / identifier only).
node scripts/gateway.mjs preflight \
  --skill baiducloud-cli --source-layer bce-cli \
  --caller-type ai_agent --caller-name comate \
  --intent query --operation bcc.ListInstances
# -> output contains attribution.event_id / session_id / turn_id
# (first call of a conversation: omit --session-id; first call of a turn: omit --turn-id;
#  later calls pass --session-id <attribution.session_id> --turn-id <attribution.turn_id>)

# After the Skill finishes (success or failure): result report, joined by event_id
# success: carry the response's http status and request id when available
node scripts/gateway.mjs report \
  --skill baiducloud-cli --source-layer bce-cli \
  --event-id <attribution.event_id from preflight> \
  --session-id <attribution.session_id> --turn-id <attribution.turn_id> \
  --outcome success --http-status 200 --request-id <requestId from the response>

# failure: add the category / code as well
node scripts/gateway.mjs report \
  --skill baiducloud-cli --source-layer bce-cli \
  --event-id <attribution.event_id from preflight> \
  --session-id <attribution.session_id> --turn-id <attribution.turn_id> \
  --outcome failure --error-category permission --error-code AccessDenied \
  --http-status 403 --request-id <requestId from the error response>

# Report only / update check only
node scripts/gateway.mjs report --skill baiducloud-openapi --source-layer openapi-runtime
node scripts/gateway.mjs check-update --skill baiducloud-cli

# CLI binary update check (--source-layer bce-cli or qianfan-cli): read-only,
# cached for 24h, adds `cli_update` {tool, current, latest, available,
# upgrade_command}. Set GATEWAY_CLI_UPDATE_DISABLE=1 to skip. The script itself
# never upgrades; the Agent runs upgrade_command per the gateway rule.
node scripts/gateway.mjs check-update --skill qianfan-chat --source-layer qianfan-cli

# Inspect the signed request without sending it
node scripts/gateway.mjs report --skill baiducloud-cli --source-layer bce-cli --dry-run

Reported fields

  • aid: not sent by the client; resolved server-side via IAM from the signed request's STS token.
  • caller_type: human when the user explicitly chose and invoked the Skill, otherwise ai_agent (default). Pass --caller-type or set BAIDUCLOUD_SKILL_CALLER_TYPE.
  • caller_name: host Agent such as claude, codex, comate. Pass --caller-name or set BAIDUCLOUD_SKILL_CALLER_NAME (default unknown).
  • skill_name: the Skill carrying the capability, e.g. baiducloud-cli.
  • source_layer: execution layer — bce-cli / openapi-runtime / qianfan-cli (also bcecmd).
  • event_id: auto-generated, format bcs-e-{uuid}; one per Skill invocation, joins the before-invocation and after-invocation reports. Pass the preflight value back with --event-id on the result report.
  • session_id: format bcs-s-{uuid}; one per conversation. Generated by the first gateway call when --session-id is omitted; the Agent remembers it and passes it on every later preflight and report.
  • turn_id: format bcs-t-{uuid}; one per user turn. Generated when --turn-id is omitted; the Agent passes it on later calls in the same turn and omits it again after the next user message. Links shown to the user in a turn that invoked a baiducloud-* / qianfan-* Skill carry track=<turn_id> in their query (see step 5 in references/gateway-rule.md).
  • event, intent, operation, outcome, error_category, error_code, http_status, request_id (optional): power intent-distribution and failure-reason analysis. http_status (100-599) and request_id (the service-side request id) go on every result report, success or failure, so the backend can join the event to its own logs and, on failures, find the full error. error_category / error_code are failure-only. Enums or short identifier tokens only — never the user's original text, error messages, resource ids, IPs, or parameter values. Invalid values are dropped and listed by field name in ignored_fields.

These are structured in the JSON payload. The request User-Agent carries only the product token (BaiduCloudSkill/1.0). See references/attribution.md.

Safety and controls

  • Reporting goes only to an https Baidu Cloud host (allowlisted, no bare IPs, no user:pass@ userinfo); a bad endpoint is skipped, not forced. Before sending, the host is resolved and refused if it points at a private/reserved address.
  • BAIDUCLOUD_SKILLS_TELEMETRY_ENDPOINT overrides the endpoint; the shipped default is https://cloud.baidu.com/api/baiducloud_skill/behavior.
  • BAIDUCLOUD_SKILL_REPORT_DISABLE=1 disables sending.
  • BAIDUCLOUD_SKILL_REPORT_DEV=1 (dev only) additionally allows an http(s) loopback endpoint (localhost / 127.0.0.0/8 / ::1) so you can post to the local scripts/report_sink.mjs for self-testing; it bypasses the DNS global-IP guard only for that loopback target. Off by default.
  • Never reads, copies, or prints the OAuth credential file contents; credentials ride only in the standard signing headers.
  • All failures (no login state, network error, missing Skills CLI) are fail-open and never block the downstream Skill.

To enforce "always run the gateway first", install the always-apply rule in references/gateway-rule.md into your host Agent's rule directory. Run node scripts/install_rule.mjs from this Skill's directory: it writes the rule to every detected host (Comate ~/.comate/rules/, Claude Code ~/.claude/rules/, Codex ~/.codex/AGENTS.md) without asking for confirmation, is idempotent, and never writes through a symlink or over a file it did not create. Add --project <dir> for the Cursor project rule, --host to pick hosts explicitly, --dry-run to preview, --uninstall to remove.