# How much of GitHub's most-starred AI tooling is written by coding agents?

**AI authorship, churn and self-published results in 52 of GitHub's most-starred AI-tooling repositories**

Rafael Frazao · data as of 2026-11-08 · [short version](https://www.soufraz.com/famous-repos-research/) · [data and evidence](https://github.com/Soufraz/famous-repos-research-data)

## Summary

We audited the full git history of 52 of the most-starred repositories in GitHub's AI-tooling topics (324,088 authored commits). 38 carry explicit AI signatures on at least 5% of their commits (all-time or in their last 180 days of activity), and 9 more commit instructions for coding agents (CLAUDE.md, AGENTS.md and similar) while signing fewer than 5%. Only 5 show neither signal, which means *unknown*, not *no AI*. In the median repository, 16.9% of the commits in its last 180 days of activity are AI-signed, a lower bound; 21 of 52 are at 25% or more. How much AI a repository uses does not predict how famous it is (Spearman ρ = -0.05, p = 0.72). Compared with 12 famous pre-Copilot projects at the same age (six months), the cohort labels 2.0 times as many of its commits as fixes (median 34.1% vs 17.3%) yet keeps more of the code it writes (median survival 67.1% vs 41.7%). The fix share, churn and survival gaps are concentrated in the 31 projects started in 2025 or later: among them, repositories that sign at least 5% of their commits and those that sign less show no detectable difference, and the 10 cohort projects started earlier do not differ detectably from the baseline. Heavy AI use is concentrated in the newest projects (AI share vs start date: ρ = 0.72), so this sample cannot separate AI use from recency. At one year (20 cohort projects old enough), the fix share gap remains but the churn and survival gaps are no longer detectable. 15 repositories publish outcome benchmarks with a runnable harness and committed results, 11 of them AI-signed, and several publish results that did not go their way.

The claim is descriptive. Heavy AI use, frequent fixing and published evidence of results coexist in today's famous AI-tooling projects. This does not show that AI use causes popularity or quality, and the published results are the authors' own, not re-run here.

## 1. What was measured

- **Cohort (52 repositories).** The top 10 by stars in each of 7 GitHub topics (`agent-harness`, `ai-agents`, `coding-agents`, `context-engineering`, `llm`, `prompt-engineering`, `spec-driven-development`) as listed on 2026-11-08, minus repositories that are reading material or collections rather than tools (books, courses, guides, prompt collections, awesome-lists), plus the 4 repositories the study started from (OmniRoute, context-mode, rtk, superpowers). No other repository was left out, whatever its age or size. 39 of the 52 were created in 2025 or later; 11 before 2024 (AutoGPT, anything-llm, dify, go-zero, langchain, lobehub, ollama, open-webui, ragflow, transformers, yao). Half of them have between 15,632 and 126,250 stars (median 65,126), fetched from the GitHub API on 2026-11-08.
- **Baseline (12 repositories).** Famous projects (over 40k stars) whose first year ended before GitHub Copilot became generally available (2022-06-21): denoland/deno, excalidraw/excalidraw, fastapi/fastapi, junegunn/fzf, gin-gonic/gin, gradio-app/gradio, hoppscotch/hoppscotch, jesseduffield/lazygit, n8n-io/n8n, microsoft/playwright, BurntSushi/ripgrep, huggingface/transformers. Both groups are chosen by today's fame. Cohort snapshots taken before that date (go-zero, transformers, yao at day 180) are left out of the age comparison, so no repository counts on both sides.
- **Commits.** Full history of the default branch; merge commits (38,229) and maintenance bots such as dependabot and release bots (10,346) are left out of every percentage. get-shit-done and gsd-core share 2,554 commits of history (up to 2026-05-22), which count in both.
- **AI signals.** (1) *Signed*: a commit carries an AI co-author trailer, an AI agent as author or committer, or a tool's "Generated with" footer, matched on the identities the tools write. (2) *Configured*: the repository commits agent instructions (`CLAUDE.md`, `AGENTS.md`, `.cursor/`, `.github/copilot-instructions.md`, ...) or runs a coding agent in CI.
- **Fix %.** Share of commits whose subject contains fix, bug, hotfix or regression. **Churn ratio**: lines added plus deleted over all history, divided by lines at HEAD. **Survival %**: lines at HEAD divided by lines ever added (source code and Markdown; vendored, generated, translated, lock and changelog files excluded).
- **Published results.** Each repository's own claims about its performance, located by file and line in the audited commit and classified by evidence type. No benchmark was re-run.

## 2. AI authorship is common

Across the cohort, the median repository has 16.9% AI-signed commits in its last 180 days of activity, counted back from its latest commit (all-time median 13.6%). 21 of 52 are at or above 25%, and 10 at or above 50%. 44 commit agent instructions, and 7 run a coding agent in CI.

| Repository | Stars | AI-signed, last 180 days of activity | Main tool |
| --- | ---: | ---: | --- |
| [holaOS](https://github.com/holaboss-ai/holaOS) | 11,434 | 85.2% | claude |
| [caveman](https://github.com/JuliusBrussee/caveman) | 118,218 | 82.3% | claude |
| [failproofai](https://github.com/FailproofAI/failproofai) | 8,230 | 79.3% | claude |
| [AutoGPT](https://github.com/Significant-Gravitas/AutoGPT) | 187,937 | 73.6% | claude |
| [munder-difflin](https://github.com/HarnessMD/munder-difflin) | 9,425 | 72.6% | claude |
| [academic-research-skills](https://github.com/Imbad0202/academic-research-skills) | 55,793 | 70.4% | claude |
| [claude-mem](https://github.com/thedotmack/claude-mem) | 114,633 | 66.9% | claude |
| [OpenSpec](https://github.com/Fission-AI/OpenSpec) | 74,248 | 63.9% | claude |
| [graphify](https://github.com/Graphify-Labs/graphify) | 144,135 | 52.7% | claude |
| [get-shit-done](https://github.com/gsd-build/get-shit-done) | 64,157 | 50% | claude |

Some projects have switched recently, older ones included. These repositories' last 180 days of activity are at least 15 points more AI-signed than their history as a whole:

| Repository | Created | AI-signed, all time | AI-signed, last 180 days |
| --- | --- | ---: | ---: |
| [go-zero](https://github.com/zeromicro/go-zero) | 2020-08-07 | 1.3% | 23.1% |
| [yao](https://github.com/YaoApp/yao) | 2021-09-06 | 3.7% | 33% |
| [AutoGPT](https://github.com/Significant-Gravitas/AutoGPT) | 2023-03-16 | 20.4% | 73.6% |
| [lobehub](https://github.com/lobehub/lobehub) | 2023-05-21 | 22% | 40.7% |
| [firecrawl](https://github.com/firecrawl/firecrawl) | 2024-04-15 | 13.1% | 42.1% |
| [OpenSpec](https://github.com/Fission-AI/OpenSpec) | 2025-08-05 | 30.6% | 63.9% |

### Unsigned does not mean "no AI"

Every AI percentage here is a lower bound. Agents can be configured not to add attribution, squash merges can drop trailers, and autocomplete or pasted chat output never signs anything. Claude Code adds a trailer by default, which is partly why Claude accounts for 85.3% of the 36,892 signed commits; tools that sign less often are under-counted. Detecting agent commits without signatures is close to guesswork ([Agent or Human?, 2026](https://arxiv.org/abs/2610.36229)), so we do not try.

Agent configuration is a second, independent signal. These 9 repositories sign fewer than 5% of their commits but commit instructions for coding agents:

| Repository | Stars | AI-signed | Agent configuration in the repository |
| --- | ---: | ---: | --- |
| [deepseek-harness](https://github.com/deepseek-ai/deepseek-harness) | 276,014 | 0% | `.agents/`, `.claude/`, `AGENTS.md`, `CLAUDE.md` |
| [hermes-agent](https://github.com/NousResearch/hermes-agent) | 261,582 | 2.9% | `AGENTS.md` |
| [ollama](https://github.com/ollama/ollama) | 186,319 | 0.1% | `AGENTS.md`, `CLAUDE.md` |
| [langchain](https://github.com/langchain-ai/langchain) | 149,190 | 0.3% | `AGENTS.md` |
| [ragflow](https://github.com/infiniflow/ragflow) | 93,749 | 2.4% | `.agents/`, `.github/copilot-instructions.md`, `AGENTS.md` |
| [OmniRoute](https://github.com/cheaper-inference/OmniRoute) | 85,531 | 3% | `.github/copilot-instructions.md`, `AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, `.github/workflows/claude.yml` |
| [herdr](https://github.com/herdrdev/herdr) | 49,219 | 0.3% | `.agents/`, `AGENTS.md` |
| [elephant-goldfish-model](https://github.com/google/elephant-goldfish-model) | 14,254 | 0.7% | `.agents/`, `.gemini/`, `AGENTS.md` |
| [gentle-ai](https://github.com/Gentleman-Programming/gentle-ai) | 8,586 | 0.3% | `.claude/`, `AGENTS.md` |

The remaining 5 (anything-llm, PageIndex, QwenPaw, conductor, open-webui) sign fewer than 5% of their commits and commit no agent configuration. Their AI use is unknown.

## 3. AI share does not track popularity

Stars and AI-signed share show no detectable association in this cohort: Spearman ρ = -0.15 all-time (p = 0.30) and -0.05 for the last 180 days of activity (p = 0.72). More-starred repositories churn more (ρ = 0.33, p = 0.017). The ten most-starred repositories:

| Repository | Stars | AI-signed, all time | Last 180 days of activity | AI evidence |
| --- | ---: | ---: | ---: | --- |
| [superpowers](https://github.com/obra/superpowers) | 314,290 | 13.3% | 6.1% | signed |
| [ECC](https://github.com/affaan-m/ECC) | 299,342 | 8.1% | 7.9% | signed |
| [deepseek-harness](https://github.com/deepseek-ai/deepseek-harness) | 276,014 | 0% | 0% | agent config, <5% signed |
| [hermes-agent](https://github.com/NousResearch/hermes-agent) | 261,582 | 2.9% | 2.8% | agent config, <5% signed |
| [firecrawl](https://github.com/firecrawl/firecrawl) | 199,609 | 13.1% | 42.1% | signed |
| [AutoGPT](https://github.com/Significant-Gravitas/AutoGPT) | 187,937 | 20.4% | 73.6% | signed |
| [ponytail](https://github.com/DietrichGebert/ponytail) | 186,540 | 24.1% | 24.1% | signed |
| [ollama](https://github.com/ollama/ollama) | 186,319 | 0.1% | 0.2% | agent config, <5% signed |
| [transformers](https://github.com/huggingface/transformers) | 168,646 | 1% | 10% | signed |
| [dify](https://github.com/langgenius/dify) | 160,649 | 5.3% | 6.7% | signed |

So the observation is "AI-heavy development is common among famous AI-tooling projects", not "the more AI, the more famous".

## 4. Fixing and churn at the same age

At HEAD the cohort's median fix share is 36.3% (interquartile range 27.6%-47.2%), its median churn ratio 2.08x and its median survival 65%.

Churn depends on a project's age, so the comparison with the baseline is made at the same age: each repository is checked out at its last mainline commit 180 and 365 days after its first commit and measured the same way.

| Metric (median) | Cohort, day 180 | Baseline, day 180 | | Cohort, day 365 | Baseline, day 365 | |
| --- | ---: | ---: | --- | ---: | ---: | --- |
| Commits | 1318 | 678.5 | p = 0.082 | 2013.5 | 1119.5 | p = 0.045 |
| Fix % | 34.1 | 17.3 | p < 0.001 | 27.3 | 17.1 | p < 0.001 |
| Churn ratio | 1.92 | 3.77 | p = 0.008 | 2.47 | 3.56 | p = 0.41 |
| Survival % | 67.1 | 41.7 | p = 0.011 | 57.55 | 43.65 | p = 0.46 |
| Repositories | 41 | 12 | | 20 | 12 | |

p-values: two-sided Mann-Whitney U, 10,000 permutations, not corrected for the number of tests. Baseline repositories show 0% AI-signed commits at both ages, as they should, which is a check on the detector's false-positive rate. At day 180, 41 of the 52 cohort repositories are compared: 8 have less than 180 days of history (deepseek-harness, elephant-goldfish-model, get-shit-done, gsd-2, holaOS, munder-difflin, ponytail, thinkrail) and 3 snapshots predate Copilot. Day 180 is the main comparison: at day 365 only 20 cohort repositories qualify, and they were far less AI-heavy at that age (median 0.8% AI-signed).

The comparison is between eras (2022-2026 projects vs 2013-2020 projects), not between AI and no AI. Within the cohort, AI share tracks fix share (Spearman ρ = 0.46 at day 180, p = 0.003; ρ = 0.34 at HEAD, p = 0.017) but not churn (Spearman ρ = -0.22 at day 180, p = 0.17; ρ = -0.23 at HEAD, p = 0.096) or survival (Spearman ρ = 0.21 at day 180, p = 0.18; ρ = 0.24 at HEAD, p = 0.087). But AI share also rises with how recently a project started (ρ = 0.72, p < 0.001): 23 of the 31 projects started in 2025 or later signed at least 5% of their commits by day 180, against 0 of the 10 started earlier. Splitting by start year:

| Metric (median, day 180) | Started 2025+, signed ≥5% | Started 2025+, signed <5% | Started 2022-2024 | Baseline | 2025+: signed vs not | 2025+ vs baseline | 2022-2024 vs baseline |
| --- | ---: | ---: | ---: | ---: | --- | --- | --- |
| Fix share | 36.2 | 34.65 | 22 | 17.3 | p = 0.47 | p < 0.001 | p = 0.18 |
| Churn | 1.89 | 1.86 | 2.79 | 3.77 | p = 0.57 | p = 0.003 | p = 0.37 |
| Survival | 69.2 | 69.25 | 52.95 | 41.7 | p = 0.60 | p = 0.003 | p = 0.38 |
| Repositories | 23 | 8 | 10 | 12 | | | |

In short: the fix share, churn and survival gaps are concentrated in the 31 projects started in 2025 or later: among them, repositories that sign at least 5% of their commits and those that sign less show no detectable difference, and the 10 cohort projects started earlier do not differ detectably from the baseline. Heavy AI use is concentrated in the newest projects (AI share vs start date: ρ = 0.72), so this sample cannot separate AI use from recency. A less-signed project is not necessarily one without agents (unsigned is unknown, not no AI), so the comparison inside the newest projects is between more and less visible AI use.

Two cautions. Fix % is read from commit subjects, and recent projects write conventional subjects ("fix: ...") far more consistently than people did in 2013-2020, so part of the gap may be writing style rather than more bugs. And churn counts every committed line: vendored or generated files committed under ordinary paths (gradio's copy of `bootstrap.js`, playwright's generated protocol files) inflate churn in either group, which is why medians are reported. What the data supports: these projects label many commits as fixes, and they do not discard more of their code than famous projects did at the same age.

## 5. What they publish about their own results

Strongest evidence each repository publishes about its own performance: harness + committed results: 24; harness, no committed results: 12; results, no harness: 8; claim without data: 3; no quantitative results: 3; third-party only: 2. Of the 24 with a harness and committed results, 15 measure whether the tool does its job (tokens saved, accuracy, cost, task success) and 9 measure only the speed of the software itself (AutoGPT, Paradev, agent-orchestrator, cmux, deepseek-harness, happy, orca, transformers, yao).

Outcome benchmarks with a harness and committed results, AI-signed repositories first. Summaries are the verified evidence notes, with links to the exact files at the audited commit:

- **[caveman](https://github.com/JuliusBrussee/caveman)** (82.3% AI-signed, last 180 days of activity). Caveman publishes three results with a runnable harness and committed data: a 10-prompt, single-run output snapshot on claude-opus-5-5 in which, on top of an 'Answer concisely.' control, `/caveman` cuts output a further 3% at the median and `/ultracave` 35%; the caveman-compress fixtures (6,198 → 4,138 tokens across five memory files); and a browser-snapshot benchmark (121 vs 15,704 tokens on a 200-row table, while on a small form caveman's full result is 2.34x larger than Playwright's). The headline input-token result (885,793 → 591,673 tokens, 33.2%, 18 of 18 answers right, with the HTML case 9.9% worse) and the Auto result (43.7% vs 44.3% resolved at 0.46x the cost, an interim run on 85 of 169 tasks) are committed as written reports: docs/WRAP-BENCHMARK.md states the first cannot be independently reproduced from this checkout, and docs/AUTO-BENCHMARK.md states the raw trajectories are not in the repository. The 63.6% (Elastic), 1.4 to 2.4x cost cut (Adobe Research) and p = 0.82 (JetBrains) figures are third-party results, and docs/HONEST-NUMBERS.md lists output reduction against default verbose replies as 'Not published' and states that no quality claim is published because no reviewed fidelity snapshot is committed. Evidence: [`README.md:116`](https://github.com/JuliusBrussee/caveman/blob/1e91f130/README.md?plain=1#L116), [`README.md:247`](https://github.com/JuliusBrussee/caveman/blob/1e91f130/README.md?plain=1#L247), [`README.md:246`](https://github.com/JuliusBrussee/caveman/blob/1e91f130/README.md?plain=1#L246)
- **[academic-research-skills](https://github.com/Imbad0202/academic-research-skills)** (70.4% AI-signed, last 180 days of activity). academic-research-skills commits held-out measurement suites with preregistrations, run records and raw outputs: on seeded-defect manuscripts the reviewer panel reached recall 1.00 on both defective fixtures with 0 clean-control false findings while severity agreement fell from 0.672 to 0.607 (and from 0.663 to 0.536 after the #581 batch), and a preregistered audit put the panel's false-claim rate between 12/805 (1.5%) and 104/805 (12.9%). A guard prompt lowered judged claim-strength drift from 7/16 to 1/16 without eliminating it, the Wording-Pattern Advisory's held-out miss rate went from 0.375/0.344 to 0.094 with 0/16 false fires (judge protocol, no runner script), and the gpt-5.6-sol citation-verifier bakeoff recorded recall 0.90 vs 0.80 on 10 non-existent references (p95 latency 25.0 s vs 49.6 s) from committed receipts and an offline scorer. Smaller checks are reported as session counts, including a routing fix that passed 3 of 6 sessions and did not meet its preregistered acceptance rule; the per-mode token and cost table in docs/PERFORMANCE.md is an estimate measured on Opus 4.x, and several release notes mark prompt-level changes as unmeasured. Evidence: [`evals/heldout/reviewer_seeded_defects/README.md:287-291`](https://github.com/Imbad0202/academic-research-skills/blob/6d938867/evals/heldout/reviewer_seeded_defects/README.md?plain=1#L287), [`evals/heldout/reviewer_seeded_defects/README.md:295-296`](https://github.com/Imbad0202/academic-research-skills/blob/6d938867/evals/heldout/reviewer_seeded_defects/README.md?plain=1#L295), [`evals/heldout/panel_claim_factuality/PREREGISTRATION.md:10-14`](https://github.com/Imbad0202/academic-research-skills/blob/6d938867/evals/heldout/panel_claim_factuality/PREREGISTRATION.md?plain=1#L10)
- **[claude-mem](https://github.com/thedotmack/claude-mem)** (66.9% AI-signed, last 180 days of activity). The read-gate eval (scripts/eval-read-gate.ts, `npm run eval:read-gate`) has a runnable harness and committed per-case tables: with the gate on, mean cost per run on a 49 KB file was $0.077 vs $0.124 (-38%) on Sonnet and $0.171 vs $0.243 (-30%) on Opus, and -1% and 0% on a 19 KB file, where an earlier version of the gate had cost Opus 20% more. The work-view SQLite benchmark (scripts/benchmark-work-view.ts) has its synthetic-fixture and approximately 3 GB real-database latencies summarized in docs/development/viewer-sync-recovery.md; the real database is not in the repository, raw reports from both benchmarks go to a gitignored reports/ directory, and the read-gate eval needs paid model calls. The Smart Explore comparison (17.8x fewer tokens for discovery, 19.4x for targeted reads) is a written A/B report with no script in the repository, and the README and CHANGELOG figures (~10x, 90%, 200x) are stated without data. Evidence: [`tests/evals/read-gate/README.md:87`](https://github.com/thedotmack/claude-mem/blob/3ab6e1a21/tests/evals/read-gate/README.md?plain=1#L87), [`tests/evals/read-gate/README.md:41`](https://github.com/thedotmack/claude-mem/blob/3ab6e1a21/tests/evals/read-gate/README.md?plain=1#L41), [`docs/development/viewer-sync-recovery.md:126`](https://github.com/thedotmack/claude-mem/blob/3ab6e1a21/docs/development/viewer-sync-recovery.md?plain=1#L126)
- **[graphify](https://github.com/Graphify-Labs/graphify)** (52.7% AI-signed, last 180 days of activity). graphify's BENCHMARKS.md reports results against named baselines with cost and judge-agreement figures (90.6% agreement, Cohen's kappa 0.81): LOCOMO recall@10 0.497 vs BM25 0.362 and mem0 0.048, LongMemEval-S 76% tied with dense RAG, and ERPNext key-fact coverage 82.0% vs 70.8% for a grep-and-read baseline (n=6); on LOCOMO QA accuracy supermemory scores higher (49.7% vs 45.3%) at about 11x the ingest cost ($15.67 vs $1.40). The harness these results point to (memory/runner.py, crosstool/run.py) is not in the repository, and no per-run outputs or spend ledgers are committed. The 71.5x token-reduction figure is backed by in-repo code (graphify benchmark) and a committed graph.json, with corpus tokens estimated from word counts rather than measured. The Graphify Cloud claim of cutting agent token spend by 80% is stated without data. Evidence: [`worked/karpathy-repos/review.md:11-28`](https://github.com/Graphify-Labs/graphify/blob/6c8002ec/worked/karpathy-repos/review.md?plain=1#L11), [`BENCHMARKS.md:19`](https://github.com/Graphify-Labs/graphify/blob/6c8002ec/BENCHMARKS.md?plain=1#L19), [`BENCHMARKS.md:20`](https://github.com/Graphify-Labs/graphify/blob/6c8002ec/BENCHMARKS.md?plain=1#L20)
- **[gsd-core](https://github.com/open-gsd/gsd-core)** (48% AI-signed, last 180 days of activity). gsd-core publishes no results on task success, output quality or end-to-end agent cost. Its measured numbers are static prompt-size reductions from two offline token-count scripts with committed JSON baselines: with the opt-in `workflow.compact_content` mode (off by default), the six spine/detail workflow splits come to 15.5% fewer tokens (113,676 to 96,055) and the 32 compact variant files to 21.91% fewer (122,467 to 95,637), counted with gpt-tokenizer as a stand-in for Claude's tokenizer and without a check that output quality holds; both rerun deterministically with `npm run benchmark:compact-content` and `npm run benchmark:compact-content-variants`. The '~40%' static-overhead reduction from prompt thinning and the '~75%' smaller command files are stated without a measurement. Evidence: [`tests/fixtures/compact-content-benchmark-baseline.json:3`](https://github.com/open-gsd/gsd-core/blob/a00b42e65/tests/fixtures/compact-content-benchmark-baseline.json?plain=1#L3), [`tests/fixtures/compact-content-variant-benchmark-baseline.json:3`](https://github.com/open-gsd/gsd-core/blob/a00b42e65/tests/fixtures/compact-content-variant-benchmark-baseline.json?plain=1#L3), [`docs/features/context-window-aware-prompt-thinning.md:7`](https://github.com/open-gsd/gsd-core/blob/a00b42e65/docs/features/context-window-aware-prompt-thinning.md?plain=1#L7)
- **[headroom](https://github.com/headroomlabs-ai/headroom)** (32.3% AI-signed, last 180 days of activity). Headroom's main token-savings table (21–57% fewer input tokens on four seeded scenarios, 178,024 → 104,261 tokens in total) is reproducible offline: benchmarks/index_proof_table.py and its committed output benchmarks/results/index_proof_table.txt give the same numbers. The latency, accuracy (GSM8K 0.870 → 0.870; TruthfulQA 0.530 → 0.560, which the README calls no detectable difference), HTML extraction (F1 0.919), tool-search and Vertex AI (76.7% fewer prompt tokens) results have harness code with results recorded in markdown; no raw run files are committed for them, and several need paid API calls. Some README figures differ from the docs: the '0.21 ms p50' matches the docs table's p95 (p50 0.20 ms), and the docs tables from benchmarks/bench_latency.py show a repeated 500-string JSON array at 4.9K → 228 tokens but structured logs at 53% and 54%, against the README's '90%+ on repeated JSON arrays and log lines'. The 40–90% image-compression figure is stated without a benchmark. Evidence: [`README.md:38`](https://github.com/headroomlabs-ai/headroom/blob/bb8d18a5/README.md?plain=1#L38), [`README.md:38-40`](https://github.com/headroomlabs-ai/headroom/blob/bb8d18a5/README.md?plain=1#L38), [`README.md:202-203`](https://github.com/headroomlabs-ai/headroom/blob/bb8d18a5/README.md?plain=1#L202)
- **[context-mode](https://github.com/mksglu/context-mode)** (28.8% AI-signed, last 180 days of activity). context-mode ships an offline benchmark suite (bench/run.mjs, 66 scenarios) with a committed results file, from which the README's 99.4% median byte saving recomputes (36 rows with 10 KB or more of raw output); these are byte counts on fixtures with tokens estimated as bytes / 4, not model runs. Live A/B runs in Claude Code with claude-sonnet-5-5 (scripts and committed aggregates; rerunning needs the `claude` CLI and paid model calls) found v2 cut the median tool_result bytes on large tasks by 84% (734.5 vs 4,694.5) with 12 of 12 large-task sessions passing in every arm, while session tokens were 5.7% worse than no plugin on small tasks and 19.6% worse on large ones; a later run put v2's session tokens 3.8% below no plugin on large tasks and 5.6% above on small ones. A paired cost study with committed run files and an analysis script found the open-source plugin cost more than Claude Code alone on six short tasks (v1 34.6% more, v2 2.2% more) and the hosted Context Mode Gateway's published rows 24.5% less, and the repo's own check of 13 numbers from the README's Try It section (commit a92cf6a7, measured on v1) found that 2 hold and 11 differ. Evidence: [`README.md:29`](https://github.com/mksglu/context-mode/blob/2000abba/README.md?plain=1#L29), [`README.md:415`](https://github.com/mksglu/context-mode/blob/2000abba/README.md?plain=1#L415), [`bench/RESULTS.md:11`](https://github.com/mksglu/context-mode/blob/2000abba/bench/RESULTS.md?plain=1#L11)
- **[book-to-skill](https://github.com/virgiliojr94/book-to-skill)** (24.4% AI-signed, last 180 days of activity). book-to-skill's headline, 24×–51× fewer tokens than placing the whole book in context to answer one question, comes from tools/discovery_tax.py run on three real books whose texts are not committed (119,264 / 175,253 / 256,287 context-dump tokens vs ~5,000 for book-to-skill, a ~4K core plus ~1K chapter that the tool takes as a design cap unless a generated skill is passed), and the 2.4×–15.6× discovery-loop figures are a model, as the tool's docstring states. Extraction counts, Docling vs pdftotext (164s vs 0.1s; 48 vs 0 tables), a ~$1-per-book generation cost and a one-chapter before/after table are published as results without a committed harness, and docs/performance.md lists 133 auto-detected chapters for Moby-Dick where docs/how-it-works.md lists none. The tools/evals harness for routing and answer accuracy has no live results yet, and docs/research/progressive-disclosure-evals.md summarizes an external study of a book-to-skill-based recipe in which a strong navigator on a single book reduced the accuracy benefit of the pre-built pack. Evidence: [`README.md:107`](https://github.com/virgiliojr94/book-to-skill/blob/e7bb659d/README.md?plain=1#L107), [`docs/performance.md:42-55`](https://github.com/virgiliojr94/book-to-skill/blob/e7bb659d/docs/performance.md?plain=1#L42), [`docs/performance.md:14-25`](https://github.com/virgiliojr94/book-to-skill/blob/e7bb659d/docs/performance.md?plain=1#L14)
- **[ponytail](https://github.com/DietrichGebert/ponytail)** (24.1% AI-signed, last 180 days of activity). Ponytail commits an agentic harness (benchmarks/agentic/run.py with judge, completeness and compare scripts) and results with per-task medians, 95% bootstrap intervals and sign tests: on 39 tasks with 5 runs per task and arm (585 sessions), Ponytail 5 reports -53% lines of code, -45% output tokens, -26% cost and -41% time versus no skill, and a test for 98% of risky logic (no skill: 68%). The same results report where it did not come out ahead: hidden checks are level (87/90 vs 86/90 for no skill, and all three arms passed 30 of 30 security runs), no-skill tests catch more injected bugs where a test exists (77% vs 69%), and a blind judge leaned toward the no-skill replies (82 : 106, p=0.09, not significant). An earlier single-shot benchmark found ponytail 26.2% and 38.7% more expensive than no skill on gpt-5.4-mini and gpt-5.5. Raw run workspaces are not committed, so checking the results requires a paid live rerun, and the /ponytail-review detection rates (100% vs 87% for no skill) appear in an image caption without data. Evidence: [`README.md:42`](https://github.com/DietrichGebert/ponytail/blob/96608e0/README.md?plain=1#L42), [`benchmarks/README.md:13`](https://github.com/DietrichGebert/ponytail/blob/96608e0/benchmarks/README.md?plain=1#L13), [`benchmarks/results/2026-10-07-agentic.md:66`](https://github.com/DietrichGebert/ponytail/blob/96608e0/benchmarks/results/2026-10-07-agentic.md?plain=1#L66)
- **[archify](https://github.com/tt-a1i/archify)** (14.4% AI-signed, last 180 days of activity). Archify publishes three committed 15-run matrices (3 models x 5 tasks) from its ordinary-model-floor benchmark, with per-run receipts and a runnable verifier and report harness. The first run recorded 10/15 first-pass usable; under the current verifier the original, post-fix and quality-first matrices each score 8/15, and the README states that the fixes do not demonstrate an overall uplift (lifecycle improved from 0/3 to 1/3 and architecture from 2/3 to 3/3, while workflow and data-flow regressed). Candidate generation uses an external runner and model providers that the repository does not ship, so the numbers can be re-verified from the committed candidates but not regenerated end to end from this checkout. A second harness, repair-rounds, has no published results. Evidence: [`benchmarks/ordinary-model-floor/README.md:96-100`](https://github.com/tt-a1i/archify/blob/a4065941/benchmarks/ordinary-model-floor/README.md?plain=1#L96), [`benchmarks/ordinary-model-floor/README.md:102-110`](https://github.com/tt-a1i/archify/blob/a4065941/benchmarks/ordinary-model-floor/README.md?plain=1#L102), [`benchmarks/ordinary-model-floor/README.md:111`](https://github.com/tt-a1i/archify/blob/a4065941/benchmarks/ordinary-model-floor/README.md?plain=1#L111)
- **[ECC](https://github.com/affaan-m/ECC)** (7.9% AI-signed, last 180 days of activity). ECC publishes GateGuard evaluations with runnable harnesses and committed aggregate results. Its pre-registered confirmatory run (run-intent.js, 96 trials) reports the gate's benefit as not confirmed (25/48 passed with the gate against 18/48 without, p = 0.210) at about 37% more turns and 40% more cost per trial; an earlier exploratory run gave 15/24 against 8/24 (p = 0.065). The replay and latency scripts report denials falling from 157 to 132 and must-deny bypasses from 8 to 0 on a 184-step corpus, and a fresh-process hook p50 of 23.68 ms against 11.58 ms on upstream/main; row-level results and transcripts of the hidden-intent runs are not committed. Outside GateGuard, the README's '~60% cost reduction' for defaulting to Sonnet is stated without data, and the mgrep '~2x fewer tokens' figure is attributed to Mixedbread. Evidence: [`docs/gateguard/question-effectiveness.md:451`](https://github.com/affaan-m/ECC/blob/d2ef2680/docs/gateguard/question-effectiveness.md?plain=1#L451), [`docs/gateguard/question-effectiveness.md:370`](https://github.com/affaan-m/ECC/blob/d2ef2680/docs/gateguard/question-effectiveness.md?plain=1#L370), [`docs/gateguard/question-effectiveness.md:471`](https://github.com/affaan-m/ECC/blob/d2ef2680/docs/gateguard/question-effectiveness.md?plain=1#L471)
- **[hermes-agent](https://github.com/NousResearch/hermes-agent)** (2.8% AI-signed, last 180 days of activity). hermes-agent commits per-PR A/B evals under evals/ with runnable harnesses and scorecards that report token savings next to measured accuracy: lean+recovery compaction recall 68.3% vs 45.8% for the current policy at 3.3x fewer tokens (49K vs 162K; 96.7% uncompacted), core-tool deferral −7% to −23% tokens with grand accuracy 0.916 vs 0.923, and browser_exec −60% to −66% tokens with 18/18 runs succeeding on both arms. The scorecards also record results that went against the change, such as gpt-5.6-terra accuracy falling from 0.938 to 0.879 and the clarify tool being used in 7/18 runs instead of 18/18 under core-tool deferral. The runs need live paid models, the compaction transcripts are not committed, and per-run rows for the deferral, read-tool and browser suites are gitignored, so those results are summary tables; the session_search eval commits all 108 per-run rows (52/54 vs 49/54 passed). The toolperf figures (−21% turns, −29% tool calls on qwen3-coder-30b) are stated in a README next to the harness, without committed results. Evidence: [`evals/compaction/results/SCORECARD-2026-08-15.md:18-21`](https://github.com/NousResearch/hermes-agent/blob/21ae0e1238/evals/compaction/results/SCORECARD-2026-08-15.md?plain=1#L18), [`evals/compaction/results/SCORECARD-2026-08-15.md:67`](https://github.com/NousResearch/hermes-agent/blob/21ae0e1238/evals/compaction/results/SCORECARD-2026-08-15.md?plain=1#L67), [`evals/core_tool_deferral/results/SUMMARY.md:1-20`](https://github.com/NousResearch/hermes-agent/blob/21ae0e1238/evals/core_tool_deferral/results/SUMMARY.md?plain=1#L1)
- **[OmniRoute](https://github.com/cheaper-inference/OmniRoute)** (1.7% AI-signed, last 180 days of activity). OmniRoute's headline '15–95% tokens (~89% avg)' is arithmetic on third-party README figures: the docs combine RTK's 60-90% and Caveman's ~46% input savings as 1 − (1 − RTK) × (1 − Caveman_input), giving 78.4–94.6% and an 89.2% average. The per-mode figures (Lite ~15%, Standard ~30%, Aggressive ~50%, Ultra ~75%) and the 69 → 19 token (72%) example are stated without data. The harness-backed numbers are narrower: a committed CI floor of compressed tokens per task over a 5-case corpus (`npm run check:compression-budget`), which records no uncompressed counts and lists the same values (129 / 127 / 160) for lite, aggressive, rtk, session-dedup and headroom; a unit test that runs the single input behind the quoted 95.93% (300 identical error lines) and asserts savings above 90%; and a routing-event microbenchmark (`npm run bench:routing-events`, ~0.12 µs of event-dispatch overhead per request). Other harnesses (bench:compression, eval:compression, perf:lts-baseline) have no committed outputs. Evidence: [`docs/compression/COMPRESSION_GUIDE.md:161`](https://github.com/cheaper-inference/OmniRoute/blob/2a8dc2741f/docs/compression/COMPRESSION_GUIDE.md?plain=1#L161), [`scripts/check/compression-budget-baseline.json:2-8`](https://github.com/cheaper-inference/OmniRoute/blob/2a8dc2741f/scripts/check/compression-budget-baseline.json?plain=1#L2), [`docs/architecture/ADAPTIVE_ROUTING.md:128`](https://github.com/cheaper-inference/OmniRoute/blob/2a8dc2741f/docs/architecture/ADAPTIVE_ROUTING.md?plain=1#L128)
- **[PageIndex](https://github.com/VectifyAI/PageIndex)** (0% AI-signed, last 180 days of activity). The repo ships a real, re-runnable comparison notebook (63 vendor-written NVIDIA 10-K questions, LLM judge, saved outputs) showing PageIndex at 100.0% accuracy vs 42.9-92.1% for vector RAG configurations, though the vector answers and judge grades are not committed and the run depends on paid APIs. The headline README numbers (FinanceBench 98.7%, 97% token cut, financial-QA benchmark) point to external vendor repos/blogs, while the indexing cost/time and native-PDF cost ratios are plots with no committed harness. Evidence: [`cookbook/pageindex-vs-vector-10k/pageindex-vs-vector-10k.ipynb:603`](https://github.com/VectifyAI/PageIndex/blob/2ac8694/cookbook/pageindex-vs-vector-10k/pageindex-vs-vector-10k.ipynb?plain=1#L603), [`cookbook/pageindex-vs-vector-10k/pageindex-vs-embeddings-10k.ipynb:574`](https://github.com/VectifyAI/PageIndex/blob/2ac8694/cookbook/pageindex-vs-vector-10k/pageindex-vs-embeddings-10k.ipynb?plain=1#L574), [`cookbook/pageindex-vs-vector-10k/pageindex-vs-vector-10k.ipynb:640`](https://github.com/VectifyAI/PageIndex/blob/2ac8694/cookbook/pageindex-vs-vector-10k/pageindex-vs-vector-10k.ipynb?plain=1#L640)
- **[gentle-ai](https://github.com/Gentleman-Programming/gentle-ai)** (0.1% AI-signed, last 180 days of activity). gentle-ai ships a deterministic, offline Go harness (bench/, its own go.mod, run in CI at .github/workflows/ci.yml:100-163) that measures the friction of its review lifecycle, such as human prompts, commands to completion and blocks, with no model calls and no timing, and commits one full run (bench/results.json: 64 journeys counted, 0 failed); a second results.json at the repo root records 0 counted and 1 failed journey. The README's 14-journey totals and its comparison with v2.1.2 (blocks 10 to 3 on the 5 comparable journeys) come from results-after.json and results-before.json, which are gitignored and not in the repo. The README's agent-performance table reports single (n=1) Gentle Shell runs with no data or harness in this repo; in it the data-migration run cost $0.555 against $0.403 before, and was accepted on the first round. The spec row and the comparison with Codex CLI link to gentle-shell PRs and an issue (#1718, #1723, #1794, #1713); the comparison reports that the large task took about three times longer (17.5–17.9 min vs 6.1 min). Evidence: [`bench/results.json:12558`](https://github.com/Gentleman-Programming/gentle-ai/blob/55240afb/bench/results.json?plain=1#L12558), [`bench/README.md:409-431`](https://github.com/Gentleman-Programming/gentle-ai/blob/55240afb/bench/README.md?plain=1#L409), [`bench/README.md:439-446`](https://github.com/Gentleman-Programming/gentle-ai/blob/55240afb/bench/README.md?plain=1#L439)

Several of them publish results that did not favour them, among them ECC, archify, context-mode and caveman (details above). 3 repositories (agents, conductor, elephant-goldfish-model) publish no quantitative results at all.

## 6. What this does not show

- **No causation.** Famous projects were selected by today's stars, so they are survivors. Projects that used AI the same way and failed are not in the sample.
- **AI-tooling projects, not famous projects in general.** People who build tools for coding agents are the likeliest to use them, so these shares say nothing about famous repositories in other domains.
- **A snapshot of moving lists.** Topic rankings change daily; the cohort is the top 10 of each topic on 2026-11-08.
- **AI share is a lower bound.** Signatures measure attribution, not authorship, and Claude Code is over-represented because it signs by default.
- **Agent configuration can be product, not process.** Only root-level paths are matched, but in AI-tooling repositories a root `CLAUDE.md`, `.claude/` or `.cursor/` can be a template the product ships rather than instructions for the project's own development.
- **Signed is not unreviewed.** A signature shows a coding agent wrote or co-wrote a commit; it says nothing about how carefully a person reviewed it.
- **AI use and recency go together.** Almost all heavy AI use is in projects started in 2025 or later, so differences from older projects cannot be attributed to AI rather than to when a project started, or to the tools and conventions of its time.
- **Stars are a noisy popularity measure** ([Six Million (Suspected) Fake Stars, ICSE 2026](https://arxiv.org/abs/2412.13459)).
- **Fix % is keyword-based** and sensitive to commit-message conventions; it is not a defect rate.
- **Churn mixes everything that is committed**, including vendored or generated files that escape the exclusion rules.
- **Published results are self-reported.** They were located and classified, not reproduced; the classification of outcome vs engineering benchmarks is manual.
- **Small baseline.** 12 pre-Copilot projects, chosen by fame, language and clone size.

## 7. Related work

**AI adoption and code quality**

- [Speed at the Cost of Quality: How Cursor AI Increases Short-Term Velocity and Long-Term Complexity in Open-Source Projects](https://arxiv.org/abs/2511.04427), He, Miller, Agarwal, Kästner, Vasilescu (Carnegie Mellon), MSR 2026. Compares GitHub projects that adopted Cursor (first commit touching Cursor configuration files such as .cursorrules) with matched non-adopters. Lines added rose 3-5x in the first month and faded after two; static-analysis warnings (+30%) and code complexity (+41%) rose persistently and predicted later slowdowns. *Relation to this study:* The closest study design. It measures the average effect of adoption; we describe the most-starred projects, which are survivors by construction. Its use of agent config files as an adoption signal is the same idea as our agent-config signal.
- [AI Copilot Code Quality: 2025 Look Back at 12 Months of Data (and the 2024 'Coding on Copilot' report)](https://www.gitclear.com/ai_assistant_code_quality_2025_research), GitClear, Industry report, 2024-2025. 211M changed lines (2020-2024): a spike in duplicated code blocks (copy/pasted lines rose from 8.3% to 12.3% of changed lines between 2021 and 2024 and exceeded moved lines for the first time), more short-term churn, and refactoring-associated lines falling from 25% of changed lines in 2021 to under 10% in 2024. Not peer reviewed; GitClear sells code-analysis tooling. *Relation to this study:* GitClear's churn means lines revised within two weeks; ours is lifetime churn relative to the code at HEAD. Both find churn rising with AI; neither measures whether the software works.
- [Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild](https://arxiv.org/abs/2603.28592), Liu, Widyasari, Zhao, Irsan, Chen, Lo, arXiv 2603.28592, 2026. 302.6k AI-authored commits from 6,299 repositories and five assistants; runs static analysis before and after each commit and tracks introduced issues to HEAD. *Relation to this study:* Uses the same kind of commit attribution. It asks what defects AI commits introduce; we ask how heavily famous projects use AI and what results they publish.
- [To What Extent Does Agent-generated Code Require Maintenance? An Empirical Study](https://arxiv.org/abs/2605.06464), Sawada, Shirai, Kashiwa, Yamaguchi, Iwata, Iida, EASE 2026 (arXiv 2605.06464). Compares the later maintenance of 508 AI-generated and 508 human-written files across 100 repositories. *Relation to this study:* File-level maintenance comparison; our churn and survival metrics are repository-level and do not separate AI-written lines from human-written ones.

**Measuring agent activity on GitHub**

- [On the Use of Agentic Coding: An Empirical Study of Pull Requests on GitHub](https://arxiv.org/abs/2509.14745), Watanabe, Li, Kashiwa, Reid, Iida, Hassan, ACM TOSEM (accepted); arXiv 2509.14745, 2025. 567 Claude Code pull requests across 157 projects; 34.1% of the commits made while revising them were co-authored by Claude. *Relation to this study:* Finds Claude Code PRs by the "Generated with Claude Code" description marker and measures AI co-authorship of revision commits with the same "Co-Authored-By: Claude" trailer that dominates our AI-signed counts.
- [How Do AI Coding Agents Contribute to Software Development? An Empirical Study of Agentic Pull Requests](https://arxiv.org/abs/2607.21832), Mazloomzadeh, Morovati, Khomh, arXiv 2607.21832, 2026. Compares pull requests from OpenAI Codex, GitHub Copilot, Cursor, Devin and Claude Code (AIDev dataset) with human PRs in Python repositories with more than 100 stars, over time. Agents are recognised by branch prefixes (codex/, copilot/, cursor/), the Devin bot account, and "Co-Authored-By: Claude" in the PR body. *Relation to this study:* Pull-request level view of the same tools whose signatures we count at commit level.
- [Detecting AI Coding Agents in Open Source: A Validated Multi-Method Census of 180 Million Repositories](https://arxiv.org/abs/2606.24429), Khosravani, Mockus, arXiv 2606.24429, 2026. Multi-method detection (configuration-file scanning, commit-message analysis, author-identity matching, bot-signature lookup) across World of Code's 180M+ Git repositories. Bot-account lookup alone recovers only 3.3% of the 850,157 Claude Code commits found in one snapshot, so single-signal estimates are biased low. *Relation to this study:* Ecosystem-scale detection that already combines commit signatures with agent-config files; we apply the same two kinds of signal to the cohort.
- [Agent or Human? Detecting Commits by AI Coding Agents in the Absence of Authorship Signatures](https://arxiv.org/abs/2610.36229), Horikawa, Li, Kashiwa, Ouatiti, Sayagh, Adams, Iida, Hassan, arXiv 2610.36229, 2026. With authorship signatures masked, detectors reach ROC AUC up to 0.859, but the best one has a precision of only 0.332 at the observed agent share, and detection transfers poorly to held-out agents (TF-IDF AUC only 0.524 for Claude Code). The authors conclude signature-based labels are a lower bound on agent activity. *Relation to this study:* Why an unsigned commit cannot be reliably attributed from its style, and why we report signed AI as a lower bound instead of guessing.

**Productivity**

- [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/), METR, arXiv 2507.09089, 2025. Randomized trial: 16 experienced developers, 246 tasks in their own mature repositories. With early-2025 AI tools (mainly Cursor with Claude 3.5/3.7 Sonnet) tasks took 19% longer, while developers believed they were about 20% faster. *Relation to this study:* Measures individual speed on mature codebases; our cohort is mostly young projects and we do not measure speed.
- [2025 State of AI-assisted Software Development (DORA)](https://dora.dev/dora-report-2025/), Google Cloud DORA, Industry survey, 2025. Survey of nearly 5,000 professionals: AI acts as an amplifier of existing team strengths and weaknesses; adoption is now linked to higher throughput but still to lower delivery stability. *Relation to this study:* Practice-level counterpart to our observation that some AI-heavy projects pair fast iteration with their own evaluation harnesses.

**Vibe coding**

- [Vibe Coding in Practice: Motivations, Challenges, and a Future Outlook - a Grey Literature Review](https://arxiv.org/abs/2510.00328), Fawzy, Tahir, Blincoe, ICSE 2026 (SEIP). 101 practitioner sources (518 firsthand accounts): vibe coders are motivated by speed and accessibility, most perceive the resulting code as fast but flawed, and QA practices are frequently overlooked. *Relation to this study:* Defines the practice qualitatively; we look for its footprint in the git history of famous projects.
- [Evaluation-Driven Development and Operations of LLM Agents: A Process Model and Reference Architecture](https://arxiv.org/abs/2411.13768), Xia, Lu, Zhu, Xing, Zhao, Zhang, arXiv 2411.13768, 2024-2025. From a multivocal literature review, proposes evaluation-driven development and operations (EDDOps): offline and online evaluation in a closed feedback loop, building on the iterative principles of TDD and BDD for LLM agents. *Relation to this study:* Names the practice visible in several cohort repos that commit their own eval harnesses and results, including null results.

**Churn and software evolution**

- [Use of Relative Code Churn Measures to Predict System Defect Density](https://www.microsoft.com/en-us/research/?p=151819), Nagappan, Ball, ICSE 2005. On Windows Server 2003, relative churn measures predicted defect density (89% accuracy separating fault-prone binaries); absolute churn did not. *Relation to this study:* Why we report churn relative to code size, and why high churn is a known risk signal rather than a sign of success.
- [Towards a Better Understanding of Software Evolution: An Empirical Study on Open-Source Software](https://www.cs.kent.edu/~jmaletic/cs63902/Papers/jsme11neamtiu.pdf), Neamtiu, Xie, Chen, Journal of Software: Evolution and Process, 2013. Across nine open-source projects, only Lehman's laws of continuing change and continuing growth held for every program. *Relation to this study:* Change is normal in living software, which is why we compare the cohort with famous pre-AI projects at the same age instead of judging churn in the abstract.

**Popularity signals**

- [Six Million (Suspected) Fake Stars in GitHub: A Growing Spiral of Popularity Contests, Spams, and Malware](https://arxiv.org/abs/2412.13459), He, Yang, Burckhardt, Kapravelos, Vasilescu, Kästner, ICSE 2026. 6.0M suspected fake stars across 18,617 repositories (July 2019-December 2024); fake stars give only a short-lived boost (less than two months) and become a liability in the long term. *Relation to this study:* Stars are our popularity measure and a noisy one; we use them to select famous projects, not as evidence of quality.

## 8. Data

Every number in this report, the per-repository metrics, the age-matched snapshots and the evidence records are published at [github.com/Soufraz/famous-repos-research-data](https://github.com/Soufraz/famous-repos-research-data) (CC BY 4.0), with the metric definitions in its README. The evidence records were collected by reading each repository at the audited commit, and every number shown was then checked against the cited file and line. The analysis code is not published yet.

## Appendix: scorecard

| Repository | Stars | Created | Selection | Commits | AI evidence | AI-signed % | AI-signed % (last 180d) | Main AI tool | Agent config since | Fix % | Churn ratio | Survival % | Code LOC | Docs LOC | Benchmark evidence |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| superpowers | 314,290 | 2025-10-09 | hand-picked | 661 | AI-signed | 13.3 | 6.1 | claude | 2025-11-22 | 24.5 | 3.86 | 41.2 | 14,874 | 20,464 | results_only |
| ECC | 299,342 | 2026-01-18 | topic top-10 | 3,507 | AI-signed | 8.1 | 7.9 | claude | 2026-01-22 | 43.7 | 1.57 | 77.8 | 406,712 | 526,066 | harness+results |
| deepseek-harness | 276,014 | 2026-08-13 | topic top-10 | 16,192 | Agent config, <5% signed | 0.0 | 0.0 |  | 2026-06-10 | 38.4 | 2.43 | 57.9 | 1,557,314 | 461,046 | harness+results |
| hermes-agent | 261,582 | 2025-07-22 | topic top-10 | 59,687 | Agent config, <5% signed | 2.9 | 2.8 | claude | 2026-02-02 | 53.5 | 2.73 | 53.8 | 3,880,145 | 364,751 | harness+results |
| firecrawl | 199,609 | 2024-04-15 | topic top-10 | 5,679 | AI-signed | 13.1 | 42.1 | claude | 2025-08-07 | 32.3 | 2.16 | 63.2 | 425,175 | 15,918 | results_only |
| AutoGPT | 187,937 | 2023-03-16 | topic top-10 | 8,306 | AI-signed | 20.4 | 73.6 | claude | 2025-05-22 | 36.7 | 2.55 | 56.2 | 1,794,028 | 74,871 | harness+results |
| ponytail | 186,540 | 2026-06-12 | topic top-10 | 361 | AI-signed | 24.1 | 24.1 | claude | 2026-06-12 | 41.8 | 1.66 | 75.2 | 11,859 | 4,965 | harness+results |
| ollama | 186,319 | 2023-06-26 | topic top-10 | 5,164 | Agent config, <5% signed | 0.1 | 0.2 | claude | 2026-06-07 | 15.1 | 4.45 | 36.7 | 491,336 | 11,124 | harness_only |
| transformers | 168,646 | 2018-10-29 | topic top-10 | 23,982 | AI-signed | 1.0 | 10.0 | claude | 2025-06-10 | 31.7 | 3.38 | 45.7 | 1,814,070 | 253,856 | harness+results |
| dify | 160,649 | 2023-04-12 | topic top-10 | 14,499 | AI-signed | 5.3 | 6.7 | claude | 2025-08-27 | 42.6 | 3.55 | 43.9 | 2,632,857 | 72,348 | harness_only |
| open-webui | 158,080 | 2023-10-06 | topic top-10 | 16,789 | Neither signal (unknown) | 1.3 | 2.6 | claude |  | 22.2 | 3.37 | 45.8 | 411,657 | 3,205 | claim_only |
| langchain | 149,190 | 2022-10-17 | topic top-10 | 16,136 | Agent config, <5% signed | 0.3 | 2.1 | copilot | 2025-09-24 | 21.7 | 5.63 | 30.1 | 427,420 | 36,854 | harness_only |
| graphify | 144,135 | 2026-04-03 | topic top-10 | 2,840 | AI-signed | 50.4 | 52.7 | claude | 2026-04-15 | 56.6 | 1.4 | 83.2 | 261,339 | 64,511 | harness+results |
| browser-use | 120,289 | 2024-10-31 | topic top-10 | 7,081 | AI-signed | 4.9 | 5.6 | claude | 2025-05-25 | 22.2 | 5.56 | 30.5 | 114,385 | 10,358 | results_only |
| orca | 118,822 | 2026-03-17 | topic top-10 | 19,084 | AI-signed | 14.7 | 13.2 | orca | 2026-03-17 | 46.6 | 3.63 | 43.2 | 4,955,851 | 25,498 | harness+results |
| caveman | 118,218 | 2026-04-04 | topic top-10 | 1,716 | AI-signed | 78.7 | 82.3 | claude | 2026-04-11 | 55.5 | 1.8 | 71.5 | 347,821 | 17,947 | harness+results |
| claude-mem | 114,633 | 2025-08-31 | topic top-10 | 3,677 | AI-signed | 61.7 | 66.9 | claude | 2025-10-21 | 46.6 | 3.38 | 47.0 | 551,913 | 31,016 | harness+results |
| archify | 94,563 | 2026-04-15 | topic top-10 | 1,203 | AI-signed | 14.0 | 14.4 | devin | 2026-08-31 | 35.8 | 1.43 | 81.5 | 163,385 | 28,330 | harness+results |
| ragflow | 93,749 | 2023-12-12 | topic top-10 | 10,788 | Agent config, <5% signed | 2.4 | 4.3 | claude | 2025-12-09 | 49.3 | 3.03 | 49.6 | 1,316,195 | 37,318 | results_only |
| rtk | 85,862 | 2026-01-22 | hand-picked | 1,826 | AI-signed | 34.8 | 35.2 | claude | 2026-01-28 | 60.1 | 1.79 | 71.8 | 224,237 | 23,900 | harness_only |
| OmniRoute | 85,531 | 2026-02-13 | hand-picked | 11,214 | Agent config, <5% signed | 3.0 | 1.7 | claude | 2026-02-18 | 61.2 | 2.3 | 51.9 | 2,903,134 | 90,767 | harness+results |
| lobehub | 84,366 | 2023-05-21 | topic top-10 | 12,870 | AI-signed | 22.0 | 40.7 | claude | 2025-04-23 | 34.9 | 1.87 | 69.7 | 3,381,118 | 127,297 | harness_only |
| headroom | 79,105 | 2026-01-07 | topic top-10 | 3,183 | AI-signed | 31.0 | 32.3 | claude | 2026-06-12 | 63.2 | 1.58 | 77.4 | 809,016 | 38,146 | harness+results |
| OpenSpec | 74,248 | 2025-08-05 | topic top-10 | 891 | AI-signed | 30.6 | 63.9 | claude | 2025-09-06 | 42.9 | 1.66 | 75.1 | 146,228 | 81,341 | claim_only |
| anything-llm | 68,053 | 2023-06-04 | topic top-10 | 2,506 | Neither signal (unknown) | 0.4 | 0.5 | claude |  | 25.3 | 2.0 | 66.7 | 225,999 | 2,207 | claim_only |
| humanizer | 66,094 | 2026-01-18 | topic top-10 | 71 | AI-signed | 32.4 | 35.8 | claude | 2026-05-26 | 25.4 | 6.07 | 28.3 | 146 | 660 | external |
| get-shit-done | 64,157 | 2025-12-14 | topic top-10 | 2,558 | AI-signed | 50.0 | 50.0 | claude | 2026-04-03 | 43.3 | 1.82 | 70.5 | 253,185 | 116,736 | results_only |
| academic-research-skills | 55,793 | 2026-02-26 | topic top-10 | 835 | AI-signed | 75.8 | 70.4 | claude | 2026-02-26 | 27.9 | 1.16 | 92.7 | 289,031 | 198,254 | harness+results |
| BMAD-METHOD | 55,118 | 2025-04-13 | topic top-10 | 2,085 | AI-signed | 14.7 | 8.5 | claude | 2026-03-14 | 31.7 | 35.84 | 5.4 | 43,788 | 21,867 | harness_only |
| herdr | 49,219 | 2026-03-27 | topic top-10 | 1,736 | Agent config, <5% signed | 0.3 | 0.3 | claude | 2026-03-27 | 55.1 | 1.84 | 70.4 | 344,596 | 210,298 | harness_only |
| agents | 42,294 | 2025-07-24 | topic top-10 | 503 | AI-signed | 11.9 | 20.5 | copilot | 2026-03-26 | 29.8 | 2.26 | 58.5 | 22,094 | 204,525 | none |
| PageIndex | 40,181 | 2025-04-01 | topic top-10 | 500 | Neither signal (unknown) | 0.6 | 0.0 | copilot |  | 25.0 | 1.39 | 83.6 | 53,287 | 343 | harness+results |
| QwenPaw | 37,473 | 2026-02-24 | topic top-10 | 2,685 | Neither signal (unknown) | 0.3 | 0.2 | gemini |  | 43.3 | 1.58 | 77.5 | 1,544,633 | 95,765 | results_only |
| book-to-skill | 37,242 | 2026-05-01 | topic top-10 | 232 | AI-signed | 27.2 | 24.4 | claude | 2026-08-23 | 42.7 | 1.28 | 87.5 | 16,876 | 3,458 | harness+results |
| context-mode | 33,662 | 2026-02-23 | hand-picked | 3,496 | AI-signed | 36.3 | 28.8 | claude | 2026-03-03 | 31.6 | 3.5 | 44.2 | 260,095 | 22,446 | harness+results |
| go-zero | 33,420 | 2020-08-07 | topic top-10 | 2,621 | AI-signed | 1.3 | 23.1 | copilot | 2025-09-20 | 26.7 | 2.29 | 60.7 | 156,948 | 7,928 | external |
| cmux | 28,745 | 2026-01-28 | topic top-10 | 18,358 | AI-signed | 24.5 | 27.2 | claude | 2026-01-22 | 35.0 | 1.65 | 74.4 | 4,546,734 | 85,464 | harness+results |
| happy | 24,536 | 2025-07-18 | topic top-10 | 4,090 | AI-signed | 13.9 | 5.2 | claude | 2026-03-06 | 24.2 | 2.27 | 61.2 | 711,528 | 39,429 | harness+results |
| Paradev | 16,090 | 2025-10-21 | topic top-10 | 5,359 | AI-signed | 5.1 | 4.1 | claude | 2025-10-21 | 40.6 | 2.34 | 59.8 | 1,152,619 | 114,274 | harness+results |
| agent-orchestrator | 14,256 | 2026-02-13 | topic top-10 | 3,670 | AI-signed | 40.4 | 34.9 | claude | 2026-06-03 | 58.6 | 1.92 | 68.5 | 1,314,260 | 25,709 | harness+results |
| elephant-goldfish-model | 14,254 | 2026-09-13 | topic top-10 | 139 | Agent config, <5% signed | 0.7 | 0.7 | claude | 2026-10-21 | 26.6 | 1.46 | 81.4 | 6,333 | 15,212 | none |
| holaOS | 11,434 | 2026-03-22 | topic top-10 | 155 | AI-signed | 85.2 | 85.2 | claude | 2026-07-26 | 42.6 | 1.04 | 98.2 | 502,684 | 42,349 | results_only |
| gsd-core | 11,334 | 2026-05-22 | topic top-10 | 5,681 | AI-signed | 48.5 | 48.0 | claude | 2026-04-03 | 49.3 | 1.82 | 70.6 | 1,073,385 | 261,901 | harness+results |
| munder-difflin | 9,425 | 2026-05-31 | topic top-10 | 1,112 | AI-signed | 72.6 | 72.6 | claude | 2026-08-19 | 34.5 | 4.37 | 37.2 | 162,667 | 40,786 | harness_only |
| Orkas | 9,098 | 2026-04-29 | topic top-10 | 735 | AI-signed | 5.2 | 2.2 | claude |  | 26.7 | 1.18 | 91.5 | 1,848,732 | 38,371 | results_only |
| openrig | 8,737 | 2026-04-01 | topic top-10 | 4,815 | AI-signed | 26.7 | 19.5 | claude | 2026-10-01 | 49.1 | 1.34 | 85.4 | 931,203 | 61,761 | harness_only |
| gentle-ai | 8,586 | 2026-02-27 | topic top-10 | 3,589 | Agent config, <5% signed | 0.3 | 0.1 | claude | 2026-03-15 | 55.1 | 2.34 | 59.8 | 545,632 | 36,536 | harness+results |
| yao | 8,280 | 2021-09-06 | topic top-10 | 2,974 | AI-signed | 3.7 | 33.0 | cursor |  | 11.6 | 2.51 | 57.0 | 456,876 | 53,667 | harness+results |
| failproofai | 8,230 | 2026-04-06 | topic top-10 | 337 | AI-signed | 83.7 | 79.3 | claude | 2026-04-07 | 32.6 | 2.32 | 60.3 | 423,095 | 213,716 | harness_only |
| thinkrail | 7,921 | 2026-07-08 | topic top-10 | 684 | AI-signed | 6.1 | 6.1 | claude | 2026-07-14 | 31.0 | 1.57 | 77.9 | 363,343 | 31,714 | harness_only |
| gsd-2 | 7,750 | 2026-03-11 | topic top-10 | 5,145 | AI-signed | 25.7 | 25.7 | claude |  | 64.6 | 1.79 | 71.1 | 715,737 | 91,818 | harness_only |
| conductor | 3,777 | 2025-12-17 | topic top-10 | 81 | Neither signal (unknown) | 0.0 | 0.0 |  |  | 33.3 | 1.92 | 68.6 | 54 | 3,121 | none |