Report Issue

Report an upstream issue for maxcompute-semantic

File a bug report or feature request upstream. There are three destinations; the skill auto-selects based on the environment:

• Internal (Aone) — when the a1 CLI is installed and can authenticate (Alibaba-internal machines). Reports land in the maintainer's Aone "My workitems" view.
• GitHub — when gh is installed and authenticated. The public upstream tracker.
• Manual hand-off — when neither CLI can actually submit (sandbox, no browser for BUC / OAuth login, no network, auth failure). The skill prints the finished issue for the user to copy-paste. This path is always available and must never be a dead end.

Always (1) redact sensitive information — see "Redact sensitive information" below — then (2) show the drafted title + body + destination to the user and get confirmation before submitting or printing. Never submit silently.

Choose the destination

Presence of a CLI is not the same as being able to submit. A sandboxed or headless environment often has a1 / gh installed but cannot complete the browser-based BUC / OAuth login, so submission will fail.

1. Probe both CLIs:

   command -v a1; command -v gh
   

2. Route:

   • a1 present → internal Aone flow (default for internal users).
   • else gh present → GitHub flow.
   • neither present → manual hand-off (below).

3. If the chosen CLI cannot authenticate or the submit step fails — e.g. a sandbox with no browser for a1 auth login --buc, no network, or an auth error — do not stop at an error. Fall through to manual hand-off: the drafted body is already done, so hand it to the user to paste.

4. The user may override the auto-choice (e.g. "file it on GitHub" even on an internal box, or "just give me the text to paste"). Honor the explicit request.

When to trigger

• Manual: the user says "file an issue / report a bug / file this upstream / 提 issue".
• Suggested: when the conversation has clearly localized a defect, proactively offer to file one and wait for user confirmation. Never submit silently.

Shared context to gather

Collect from the conversation regardless of destination:

• title — short, imperative, in the user's language. Do not include the [maxcompute-semantic] prefix in {title} itself — each flow adds its own.
• category — bug (default), feature request, or question/task.
• summary — one-line description.
• repro — failing command + observable symptom (for bugs).
• root_cause — if localized this session; otherwise Unknown — needs triage.
• code_refs — path:line entries touched this session.
• plugin_version — output of mcs --version. Required.
• doctor_snapshot — output of mcs doctor (plain form, not -f json). Required in every bug report. If mcs doctor errors, include whatever it produced plus the error. Redact it first — see below.

Redact sensitive information

The SQL and mcs doctor output carry data only you can judge as safe. The agent's masking is help, not a guarantee — the binding safeguard is your confirmation (next section). Two rules:

1. The agent always strips, no exceptions — access-key secrets, tokens, passwords, full credentials. These never belong in an issue, on any destination; remove them before the draft is even shown.

2. The agent flags and proposes masking for anything else that looks sensitive, and points each one out so you can decide at confirmation time — never silently. Common cases:

• SQL: real table / project / schema names; literal values in WHERE / IN / CASE WHEN (event keys, ids, phone numbers, account names); inline sample data or pasted result rows.
• mcs doctor: profile names (can encode tenant / account ids); absolute paths that reveal the OS user or home; internal endpoints / bucket URLs / hostnames; access-key ids / account / principal strings.

Proposed masking keeps the diagnostic shape but drops the specifics — e.g. a table FQN → <project>.<table>, a literal → '<val>', a path → /home/<user>, a profile name → <name>_***. Keep what makes the bug actionable: the error message, which doctor checks failed, the SQL structure, versions, and tier. The agent can't know your business context, so it must propose, not decide.

Confirm before submitting — hard gate

Submitting or pasting is never automatic. For every destination — Aone, GitHub, and manual hand-off:

1. Show the user the final, exact text that will be sent: the [maxcompute-semantic] title and the full body, after redaction.
2. Ask explicitly — e.g. "OK to submit this? Anything else to redact?"
3. Submit / paste only after the user explicitly approves. If they want changes, edit and re-confirm; earlier approval never covers edited content.

Internal Aone flow

The workitem is created in an Aone project the user has write access to, with --related-space pointing at the upstream project 871418 ([Agent special project]) — the maintainer sees it in their personal "My workitems" view. This lets internal teams report issues without needing access to 871418.

Use category bug for defects, req for feature requests, task for miscellaneous tasks.

Steps

0. Preflight: confirm a1 CLI is installed. Run command -v a1. If missing, stop and ask the user to install it:

   curl -fsSL https://git.cn-hangzhou.oss-cdn.aliyun-inc.com/aone-cli/install.sh | sh
   

   Docs: https://a1.io.alibaba-inc.com/. After install, the user must run a1 auth login --buc once before submitting.

1. Pick the target project:

   • User already named a project → use it.
   • Otherwise run a1 project list --quiet (add --keyword <kw> to filter), show the list, and ask which to file under. The ID is <target_project>.

2. Redact (see "Redact sensitive information"), then get the user's explicit confirmation of the final content (see "Confirm before submitting").

3. Write the body to a temp file first, then submit. Write straight into /tmp/mcs-report-issue-<timestamp>.md, not via a bash heredoc (heredocs are unreliable under noclobber — we've seen silent empty-body submissions).

   Body template:

   ## Summary
   {summary}
   
   ## Reproduction
   - Plugin: maxcompute-semantic v{plugin_version}
   - Profile: {profile_name} (tier {2-level|3-level}, compute={compute_project})
   - Command: {command_or_NA}
   - Symptom: {observable}
   
   ### `mcs doctor` snapshot
   {doctor_snapshot}
   
   ## Root cause
   {root_cause}
   
   ## Code references
   - {file:line} — {what}
   

   Both the Plugin line and the mcs doctor block are required.

   Then submit:

   a1 project workitem create \
     --project <target_project> \
     --category {bug|req|task} \
     --related-space 871418 \
     --assignee 292165 \
     --title "[maxcompute-semantic] {title}" \
     --body-file /tmp/mcs-report-issue-<timestamp>.md \
     --quiet
   

   cat the temp file once before submitting to confirm it isn't empty.

4. Report back: workitem ID + URL https://project.aone.alibaba-inc.com/v2/project/<target_project>/{bug|req|task}/{id}.

Notes

• --related-space 871418 is this skill's identity. It does not require access to the target project, so the link lands even if the user can't read
  871418. The target ID must reference a real Aone project (invalid IDs are silently discarded).
• Default --assignee 292165 (jiexian.hc). Override when the user names someone else. Employee IDs avoid nickname ambiguity.
• No --tag: the [maxcompute-semantic] title prefix is the universal cross-project identifier and filter.
• Maintainer searches cross-project: a1 project workitem list --scope personal --title "[maxcompute-semantic]" (add --category bug for bugs only).
• a1 cannot delete workitems — only state transitions (Invalid / Won'tfix / Closed). Hard delete via web UI.
• Auth error → tell the user to run a1 auth login --buc.

GitHub flow

Prerequisites

The gh CLI must be installed and authenticated: gh auth status. If not authenticated, ask the user to run gh auth login first.

Steps

1. Classify: bug, feature request, or question.

2. Redact (see "Redact sensitive information"), then get the user's explicit confirmation of the final content (see "Confirm before submitting").

3. Create:

   gh issue create \
     --repo aliyun/maxcompute-semantic \
     --title "[bug] {title}" \
     --body "## Description
   {summary}
   
   ## Steps to reproduce
   1. ...
   
   ## Expected behavior
   <expected>
   
   ## Environment
   - mcs version: {plugin_version}
   - OS / Python version
   
   ## Additional context
   {doctor_snapshot}, error envelopes, code references, etc." \
     --label bug
   

   For feature requests: --label enhancement, title prefix [feat].

4. Report back: show the issue URL to the user.

Manual hand-off (cannot submit)

Use this whenever neither CLI can submit — sandbox, no browser for BUC / OAuth, no network, a missing or unauthenticated CLI, or a submit that errored. Never dead-end on "run a1 auth login --buc" when that login is impossible in the environment.

1. Redact the draft (see "Redact sensitive information"). This matters even more here — the user may paste it into a chat or ticket you don't control.
2. Print the finished issue as a single copy-paste block: the [maxcompute-semantic] title plus the full body (Summary / Reproduction / redacted SQL / redacted mcs doctor snapshot / Root cause / Code references).
3. Tell the user how to deliver it — for example:
   • paste it to the maintainer directly (in this conversation, IM, or email);
   • or, from a machine that can authenticate, file it via the Aone or GitHub steps above (the body is ready for --body-file / --body);
   • or paste it into the Aone / GitHub web UI by hand.

Do not attempt the BUC / OAuth login yourself in a restricted environment — just produce the clean, redacted text and hand it off.