What does this skill do, and when should you use it?
vibe-replay is a local-first CLI that converts sessions from Claude Code, Cursor, Codex, OpenCode, Hermes, Pi, and Grok Bot into a single self-contained HTML replay, with dashboards, heatmaps, cost tracking, and token analytics. The companion replay skill lets an agent discover and search local sessions, generate GitHub-ready PR summaries, and run opt-in cleanup steps — credential review, translation, tone softening — before anything is pasted into a PR. Everything runs locally by default; data only leaves the machine when you explicitly publish. It ships as both a Claude Code plugin and a portable Agent Skills package.
The skill: 1) locates the current Claude Code session via grep on ${CLAUDE_SESSION_ID}, or searches past sessions with npx vibe-replay sessions (by project, provider, keywords; supports --, --scan, --brief, --dedupe); 2) runs npx vibe-replay --provider <provider> --session <PATH> --github, producing github-summary.md, redactions., and optional GIF/SVG previews in ~/.vibe-replay/<slug>/; 3) reviews flagged-but-not-auto-redacted credentials from redactions. and applies edits only to a copy file, github-summary.clean.md; 4) optionally translates non-English prompts and softens frustrated tone into Professional/Neutral/Friendly styles; 5) appends (recommended default) or replaces the summary in a PR body via gh pr edit --body-file; 6) for retros, uses --scan output to analyze prompt efficiency, tool calls, token cost, and compactions.
- A developer creating a PR wants reviewers to understand the AI session's decision process and attaches a replay summary to the PR description
- A user vaguely remembers a past debugging session ('how did I fix that auth bug?') and asks the agent to search and locate it
- A team runs a prompt-quality retro, analyzing tool calls per prompt, token spend, API errors, and compactions from a session
- A session contains non-English prompts or frustrated wording, and the user wants it redacted, translated, and softened before sharing
- Teaching or onboarding: use animated replays of real sessions as walkthrough material
How do you install this skill?
- The skill runs the third-party npm package vibe-replay via npx; supply-chain trust rests on the npm registry. Pin a version before relying on it.
- The PR write step uses gh pr edit; the replace mode overwrites the entire PR body. Prefer the default append mode and preserve existing content.
- Secret redaction is regex-based; sensitive data outside leftoverFindings may still slip through — manually review the final markdown before sharing.
- Core functionality depends on npm registry reachability; npx installation may fail or be unstable from mainland-China networks.
- This is a static source review; no commands were executed, so output quality and script robustness are unverified.
- Shell / CLI
- Network access
- Local filesystem
Node.js >= 22.19.0npx (vibe-replay)GitHub CLI (gh, for PR steps)Claude Code harness ($CLAUDE_SESSION_ID)
Three options:
- Agent Skills CLI (recommended for Cursor users):
bash
npx skills add tuo-lei/vibe-replay --skill replay -g
Cursor discovers global skills from ~/.agents/skills/ and ~/.cursor/skills/.
- Claude Code plugin: run /plugin in Claude Code and search vibe-replay in the Discover tab, or:
bash
/plugin marketplace add tuo-lei/vibe-replay
/plugin install vibe-replay@vibe-replay
- Manual single-file install:
bash
mkdir -p ~/.claude/skills/replay
curl -o ~/.claude/skills/replay/SKILL.md \
https://raw.githubusercontent.com/tuo-lei/vibe-replay/main/skills/replay/SKILL.md
The CLI requires Node.js >= 22.19.0; PR steps require the GitHub CLI (gh).
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/tuo-lei/vibe-replay.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/replay" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- generate a replay of this session
- attach session context to this PR
- find the session where I debugged that bug
- run a retro on this session
Trigger it with natural language after install: 'generate a replay of this session', 'attach session context to this PR', 'find the session where I debugged that bug', 'run a retro on this session'. The skill's argument hint is [session-path-or-search-query]. You can also use the CLI directly: npx vibe-replay for the current session, npx vibe-replay -d for the local dashboard, npx vibe-replay sessions --query "..." -- to search history.
What are this skill's strengths and limitations?
- Local-first: a single self-contained HTML file works offline and makes no automatic external requests
- Nine providers supported (Claude Code/Desktop/Cowork, Codex, Cursor, OpenCode, Hermes, Pi, Grok Bot), with multi-file and resumed session auto-discovery
- Built-in secret redaction plus agent-side credential review before PR writes; the original summary stays read-only and all edits go to a .clean.md copy
- Clean agent/CLI division of labor: the CLI does mechanical parsing, the agent ranks search results, judges credential hits, and handles translation/tone — with user confirmation at every cleanup step
- PR writes default to appending under a '## Session Replay' heading instead of replacing, avoiding silent wipes of hand-written PR bodies
- Requires Node.js >= 22.19.0 (a relatively high bar); PR features depend on the gh CLI
- ${CLAUDE_SESSION_ID} relies on Claude Code harness interpolation — other platforms need a manual session path
- Session search is metadata/subsequence matching, not semantic search; fuzzy queries may need several retries
- Broad --scan is expensive by design and the skill explicitly advises against it over many sessions
- Cursor token snapshots are marked estimated and compaction counts are lower bounds, so metric precision varies by provider
- The CLI sends opt-out pseudonymous telemetry (no prompts or content) that privacy-sensitive users must disable manually
- Remote SSH sources have no Live mode; Grok Bot transcripts do not support SSH remote indexing yet
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Vibe Replay this page | 56 · Use with care | ★ 38 | 11d ago | MIT |
| Obsidian Knowledge Brain v4.0 | 56 · Use with care | ★ 100 | 3mo ago | MIT |
| Agents Observe | 54 · Use with care | ★ 696 | 1mo ago | MIT |
| Agent Skills Manager — manage-skills Skill | 50 · Use with care | ★ 41 | 6mo ago | MIT |
| Engram Persistent Memory Protocol | 68 · Recommended | ★ 7.2k | 11d ago | MIT |
The source names no direct competitors. Compared with reading raw session JSONL files manually, its value is automating parsing, redaction, analytics, and HTML packaging; compared with plain session browsers, it adds PR workflow integration and retro analytics.
How did FollowSkills review this skill?
SKILL.md demonstrates least privilege (scoped allowed-tools), mandatory user confirmation (translation/tone/PR writes are opt-in), no overwriting of originals, a credential-redaction review flow, and an explicit recovery path (original github-summary.md stays read-only). Deductions: the underlying dependency npx vibe-replay is a remotely fetched third-party npm package whose supply-chain trust is not declared in the skill; shell command interpolation carries a minor injection surface; publisher is unverified.
Instructions are self-consistent with error branches (empty-result --version check, oldest-file tiebreak for /resume, degradation when ${CLAUDE_SESSION_ID} is empty) and clear failure feedback. Deductions: static review only; CI and tests cover the CLI and symlink integrity but no end-to-end reproduction of the skill's key path (search → generate → redact → PR write), so the anchor caps this at 10.
Trigger scenarios are concrete (replay, find session, retro, PR context) with declared boundaries (metadata not semantic search, avoid broad --scan, Cursor metrics estimated). Deductions: core function depends on npx fetching from the npm registry, a mainland-China reachability risk that is not disclosed; no Chinese-language support noted.
Well-layered docs (SKILL.md for agents, README for humans), multiple install paths, a dedicated agent-instructions test guarding symlinks, clear MIT license and maintenance path. Deductions: the skill file itself has no version/changelog; version governance rests on the repo-level package. (0.0.3); parameter stability is not committed.
The value claim is specific (delegate hard-to-write session stats/structured summaries to the CLI, agent handles ranking and judgment) and output is directly usable markdown. Deductions: static review cannot verify output completeness; robustness of the PR-append script (e.g., bodies with special characters) is unproven; comparative benefit is asserted in docs without independent verification.
The repo has real CI (lint/test/build/windows-smoke/npm-pack simulation/audit) and committed test suites — auditable primary material. Deductions: this evidence covers the CLI/website, not the skill's claimed key paths; demo links and sample outputs are not statically re-checkable, so multiple independent corroboration is not met.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →