agent-smoke-e2e
End-to-end smoke check of the moraine MCP server. Exercises the public retrieval workflow against a live moraine stack and prints…
Help prepare and, only with explicit user approval, file a sanitized Moraine bug report for reproducible defects in Moraine itself.
$ npx -y skills add eric-tramel/moraine --skill bug-report --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/bug-reportContext preview
The summary Claude sees to decide when to auto-load this skill.
Help prepare and, only with explicit user approval, file a sanitized Moraine bug report for reproducible defects in Moraine itself.
name: bug-report description: Help prepare and, only with explicit user approval, file a sanitized Moraine bug report for reproducible defects in Moraine itself.
Use this skill when the user hit a likely Moraine defect and wants help turning it into a useful GitHub issue. A Moraine bug is a reproducible problem in the Moraine CLI, ingest, monitor, MCP server, setup flow, packaging, documentation, or bundled plugins.
Do not use this skill for general support, one-off local setup help, unrelated agent-harness issues, or private debugging that cannot be generalized into a Moraine repository issue. Troubleshoot those locally instead, and file a report only if the evidence points to a fixable Moraine defect.
The Moraine repository is:
https://github.com/eric-tramel/moraine
New bug reports go through GitHub Issues:
https://github.com/eric-tramel/moraine/issues/new
If a GitHub issue tool, browser, or `gh` is available, use it only after the user explicitly confirms the final redacted report. If no posting tool is available, provide the issue URL and the prepared report body.
Never silently post an issue on the user's behalf. Even if the user says "file it", first show the final sanitized title and body, then ask for explicit confirmation before creating the issue. If the user declines, stop after providing the draft.
Do not upload logs, configs, screenshots, session transcripts, or command output without the same explicit confirmation.
All bug reports must anonymize PII and host-specific details before they leave the user's machine.
Redact or generalize:
internal project names.
and temp directories.
the user explicitly approves a minimal sanitized excerpt.
labels like "macOS on Apple Silicon", "Linux x86_64", "Homebrew install", or "uv tool install" instead of full hardware, hostname, local account, or full directory layouts.
Use placeholders such as `<home>`, `<repo>`, `<project>`, `<host>`, `<workspace>`, `<token>`, and `<private transcript omitted>`.
Before drafting a report, determine whether the issue belongs in the Moraine repo:
1. Identify the affected Moraine surface: CLI, setup, ingest source, monitor, MCP search, ClickHouse schema, packaging, documentation, Codex plugin, Claude Code plugin, Hermes plugin, or another bundled integration. 2. Check whether the behavior is reproducible or has enough evidence to debug. 3. Separate Moraine defects from user-specific environment problems. A missing local dependency, local permissions issue, custom shell setup, broken external harness install, or private repo configuration is usually support unless Moraine handles it incorrectly or gives a misleading failure. 4. Prefer a small sanitized reproducer over broad environment dumps.
Draft reports in this shape:
## Summary One or two sentences describing the Moraine bug and affected surface. ## Affected Area - Moraine component: CLI | setup | ingest | monitor | MCP | packaging | docs | plugin - Harness, if relevant: Codex | Claude Code | Hermes | OpenCode | Cursor | Kimi CLI | Pi Coding Agent | other - Install method, if relevant: release bundle | uv tool | Homebrew | source checkout | unknown ## Version - Moraine version: output of `moraine --version`, if available - Broad platform: sanitized OS and architecture category only ## Steps To Reproduce 1. Minimal first step. 2. Minimal second step. 3. Command or action that triggers the bug, with paths and private names replaced. ## Expected Behavior What Moraine should have done. ## Actual Behavior What happened instead, including the sanitized error message or symptom. ## Sanitized Evidence Short redacted excerpts only. Replace private values with placeholders. ## Workaround Known workaround, if any. ## Notes Any suspected regression range, related issue, or additional context that does not expose private data.
Ask before running commands that may read local state. Prefer narrow commands whose output can be summarized and redacted, such as:
moraine --version moraine status moraine logs --lines 100
When logs are useful, include only the shortest sanitized excerpt that explains the failure. Do not include full session content or full config files. Summarize config shape instead of posting exact values.
If prior agent context matters and Moraine MCP tools are available, use the current harness's Moraine search tools to recover the relevant session or error, then apply the same redaction rules before drafting the issue.
1. Draft a sanitized title and body using the report structure. 2. Point out any assumptions or missing reproduction details. 3. Ask the user whether they want to post it to GitHub. 4. If they say yes, show the final redacted report and ask for explicit posting confirmation. 5. Create the GitHub issue only after that confirmation, then return the issue URL. 6. If they do not confirm, leave them with the draft and the new-issue URL.
Moraine is a local trace stack for agent work. It indexes sessions from agent harnesses such as Codex, Claude Code, Qwen Code, Kiro CLI, Kimi CLI, NAC, OpenCode, Hermes, Pi Coding Agent, and Prime Agent into ClickHouse, serves a monitor UI, and exposes MCP
End-to-end smoke check of the moraine MCP server. Exercises the public retrieval workflow against a live moraine stack and prints…
Rewrite a moraine GitHub release body in the house format — one-sentence pitch, install block, usage-focused "what's new"…
Use Moraine realtime session data to inspect active agent work across harnesses when the user asks what is happening now or what…
Use Moraine MCP tools to recover prior agent-session context when the user refers to earlier work, decisions, failures, branches,…
Review a PR through the CodeReviewCompleteness persona. Use when asked to check whether a PR fully addresses the objectives it…
Review a PR through the CodeReviewCorrectness persona. Use when asked to check whether the implementation does the job correctly,…