{"id":50919066,"url":"https://github.com/skoppisetty/paxel-skill","last_synced_at":"2026-06-16T18:01:30.054Z","repository":{"id":363667031,"uuid":"1264379181","full_name":"skoppisetty/paxel-skill","owner":"skoppisetty","description":"Score your own Claude Code sessions on YC Paxel's 5-axis builder rubric — locally, with Claude Haiku as the scorer. No upload, no backend.","archived":false,"fork":false,"pushed_at":"2026-06-09T20:43:14.000Z","size":26,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-09T22:20:03.443Z","etag":null,"topics":["agent-skills","anthropic","claude","claude-code","claude-skill","code-review","developer-tools","llm","paxel","self-assessment","ycombinator"],"latest_commit_sha":null,"homepage":"https://github.com/skoppisetty/paxel-skill","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/skoppisetty.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-06-09T20:40:50.000Z","updated_at":"2026-06-09T20:43:17.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/skoppisetty/paxel-skill","commit_stats":null,"previous_names":["skoppisetty/paxel-skill"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/skoppisetty/paxel-skill","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/skoppisetty%2Fpaxel-skill","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/skoppisetty%2Fpaxel-skill/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/skoppisetty%2Fpaxel-skill/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/skoppisetty%2Fpaxel-skill/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/skoppisetty","download_url":"https://codeload.github.com/skoppisetty/paxel-skill/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/skoppisetty%2Fpaxel-skill/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34417416,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-16T02:00:06.860Z","response_time":126,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["agent-skills","anthropic","claude","claude-code","claude-skill","code-review","developer-tools","llm","paxel","self-assessment","ycombinator"],"created_at":"2026-06-16T18:01:29.002Z","updated_at":"2026-06-16T18:01:30.034Z","avatar_url":"https://github.com/skoppisetty.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# paxel-skill\n\nA local **self-assessment** that scores your own Claude Code and Codex CLI sessions on YC Paxel's exact 5-axis rubric — **without uploading anything to YC**.\n\nIt bundles Paxel's **verbatim** scoring and narrative prompts (extracted from the public `ghcr.io/yc-software/paxel-client` image) and its band-cut thresholds. The one change: every narrative/scoring call is forced to **Claude Haiku 4.5** (`model: haiku`), not YC's gpt-5.5-none proxy. Nothing leaves your machine.\n\n## What it does\n\n```\nyour ~/.claude/projects/**/*.jsonl  +  ~/.codex/sessions/**/*.jsonl  +  your git repo\n  → condense.py        (deterministic: drop tool/file bodies, scrub secrets, chunk;\n                        Codex rollouts auto-detected \u0026 normalized to the same form)\n  → events.py          (deterministic: events, session_signals, user highlights,\n                        plan files, dispatch/return pairs — EventExtractor port)\n  → gitdata.py         (deterministic: commit grouping + session→episode linking\n                        with real git LOC — CommitGrouper/EpisodeLinker port)\n  → decisions.py       (extract → Haiku classifier, verbatim v5 prompt + 49-law\n                        catalog → finalize with in-session outcome analysis)\n  → narrative prompt   (Haiku, verbatim Paxel SYSTEM_PROMPT)    → per-session notes\n  → episodes.py        (deterministic: byte-faithful build_episode_input — all\n                        blocks, caps, Code volume, dispatch ratio)\n  → scoring prompt     (Haiku, verbatim Paxel rubric)           → 5-axis scores\n  → aggregate.py       (verbatim band cuts + approx rollup)     → overall + band\n  → analytics.py       (optional: the upload-only payload Paxel's server would\n                        receive — 14 code-quality dims, velocity, steering traces)\n```\n\nAxes: **Execution Leverage, Steering, Engineering Quality, Product Thinking, Planning** (1.0–10.0). Bands: WEAK / LIMITED / STRONG / ELITE / EXEMPLAR.\n\n## What's faithful vs approximate (read this)\n\n| Layer | Fidelity |\n|---|---|\n| Scoring rubric + calibration | **Verbatim** from the client image |\n| Narrative prompt | **Verbatim** |\n| Condensing (drop bodies, scrub, caps) | **Faithful port** of the client pipeline |\n| Codex CLI ingestion | **Faithful port** of `CodexNormalizer` (boilerplate filter, `apply_patch` → per-file edits, `update_plan` → plan signal), plus one deliberate extension: the newer `custom_tool_call` apply_patch shape, which the archived client misses |\n| Event extraction + `session_signals` | **Faithful port** of `EventExtractor` / `SessionSignalExtractor` (verbatim regexes, caps, Ruby string semantics) |\n| Commit grouping + episode linking | **Faithful port** of `CommitGrouper` / `EpisodeLinker` / `LinkingStrategy` (pr 1.0 / sha 0.9 / branch 0.7 / ±1h timestamp 0.5; first-match-wins; real numstat LOC) |\n| Decision exchanges | **Faithful port** of extractor + classifier (verbatim v5 prompt + 49-law catalog, Haiku-classified) + deterministic in-session outcome analysis; **chains not ported** (need YC's embedding service — the client itself degrades identically when embeddings fail) |\n| Episode input (`build_episode_input`) | **Byte-faithful port** — all blocks, caps (50K/30K/10K/5K), Code volume gating, committed-return ratio. `## Code Reviews` absent exactly as in Paxel's own local pipeline (server-side only) |\n| Upload-only analytics (14 code-quality dims, velocity, steering, parallelism) | **Faithful port**, exposed as a transparency report (`analytics.py`) — never feeds scores, same as in Paxel |\n| Band cut thresholds (\u003c4/\u003c6/\u003c8/\u003c9/≥9) | **Verbatim** |\n| Per-axis scores | **Faithful** (model differs — see below) |\n| **Overall score + band** | **APPROXIMATION** — YC's rollup is server-side, not in the image |\n| Scorer model | **Claude Haiku 4.5** (forced via `model: haiku`), not gpt-5.5-none (Paxel's actual model) — directional, not identical |\n\nLLM scoring is nondeterministic; re-runs vary. The overall number is a calibration target, not YC's verdict. Full gap analysis: [`reference/GAPS.md`](reference/GAPS.md).\n\n**Known local deviations:** decision-exchange chains (need embeddings); `pr_number` recovered from in-session `gh pr create` output instead of Paxel's gh-CLI sidecar; committed-return ratio via the parent-commit-after-return branch (no child subagent session records locally); Gemini CLI / Cursor / opencode ingestion not ported (only Claude Code + Codex CLI here).\n\n## Install (as a Claude Code skill)\n\nClone straight into your Claude Code skills directory:\n\n```bash\ngit clone https://github.com/skoppisetty/paxel-skill.git ~/.claude/skills/paxel-skill\n```\n\nStart a **new** Claude Code session (skills load at startup), then invoke `/paxel-skill` — or just ask *\"score my Claude Code sessions on the Paxel rubric.\"* Update later with `git -C ~/.claude/skills/paxel-skill pull`.\n\n**Hacking on it?** Clone anywhere and symlink instead, so your edits are live:\n\n```bash\ngit clone https://github.com/skoppisetty/paxel-skill.git\nln -s \"$(pwd)/paxel-skill\" ~/.claude/skills/paxel-skill\n```\n\n**Requirements:** Claude Code + Python 3 (standard library only — nothing to `pip install`).\n\n\u003e **Note:** scoring is **per-project** — it defaults to the logs for the repo you're currently in (`~/.claude/projects/\u003cencoded-cwd\u003e/`, plus Codex sessions whose `cwd` matches), not your entire history. Point it at another project's log dir to score that one instead.\n\n## Run the scripts directly (no skill)\n\n```bash\npython3 scripts/condense.py ~/.claude/projects/\u003cencoded-cwd\u003e \u003e sessions.jsonl\n# narrate + score each session with the prompts/ files (any LLM), collect to episodes.json\npython3 scripts/aggregate.py episodes.json\n```\n\nTests (stdlib only): `python3 scripts/test_scripts.py`.\n\n## Provenance \u0026 IP\n\nPrompts and thresholds are extracted verbatim from YC's **publicly pullable** client image for **personal, interoperability/assessment** use. They are YC's text, not original work here — do not redistribute this as your own rubric. If you publish or share a derivative, reimplement the rubric in your own words.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fskoppisetty%2Fpaxel-skill","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fskoppisetty%2Fpaxel-skill","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fskoppisetty%2Fpaxel-skill/lists"}