Report Date: 2026-09-13 | Language: English | Generated At: 2026-09-13T16:31:23.000Z
# Today's Best Build: RepoProof
**Report Date**: 2026-09-13
**Coverage**: 2026-09-13T00:00:00+08:00 – 2026-09-13T23:59:59+08:00 (UTC)
**Status**: partial (No strong signal for questions: Q14)
## Today's Best Build: RepoProof
**One-liner**: A local-first harness that replays your team's last 20 merged pull requests against two coding agents — one author, one skeptic — and prints a scorecard of which model actually ships into YOUR codebase.
**Why Now**: On 2026-09-13 the top-scoring signal of the day is Real-SWE (score 244), a benchmark built on licensed private enterprise codebases precisely because public SWE benchmarks stopped predicting anything. The same day, Product Hunt listed Cognition's SWE-2 touting a 64% price cut against Fable 5.1, and a Dev.to author documented 30 days of two-AI code review where a human still found the bug in 5 minutes. Model vendors are now competing on price and self-reported scores, while every engineering team
**Evidence**:
- Real-SWE needed to license private, real-world enterprise codebases because public benchmarks no longer reflect how frontier models behave on real production code. That is exactly the gap a per-repo replay harness fills without licensing anything. _(signal #71471)_
- A 30-day experiment where an author agent and a separate skeptic agent reviewed each other with no human in the loop worked 'far better than expected' right up until the one bug that mattered — meaning the review structure is worth building, but the scoring gate is what must stay human-owned. _(signal #71700)_
- Cognition's SWE-2 competes explicitly on a 64%-cheaper price versus Fable 5.1. When price is the pitch, buyers need an independent way to check whether the cheap model regresses on their own repo before swapping the default. _(signal #71587)_
- CauterRule demonstrates that repeated agent failures can be extracted, replay-tested and promoted into standing rules, and reports a field test across 4,768 trajectory-runs. The replay-and-promote primitive already exists in the open; the missing layer is a repo-specific score over it. _(signal #71703)_
- viserys-agent reached 627 stars by packaging 28 SKILL.md workflows and 4 reviewer personas for coding agents. Free workflow packs are now abundant, which proves the prompts are commoditized and the measurable outcome is the sellable part. _(signal #71629)_
**Fastest Validation**: Pick one friendly team, export their last 20 merged PRs with their pre-merge diffs, replay each against two models (an author agent and an adversarial skeptic agent), and ask the team's tech lead to blindly rank the 20 outputs before seeing which model produced them. If the lead's blind ranking agrees with the harness score on fewer than 15 of 20 comments, the scoring function is wrong, not the model.
**Counter-view**: Cognition already sells SWE-2 at 64% cheaper than Fable 5.1, and agent-dev-team ships 21 reviewer roles under MIT for free — so if the pitch is 'better review prompts' or 'cheaper tokens', this loses on day one. The only defensible position is a frozen, versioned scorecard tied to a specific repo's merge history, which is boring work nobody at a model lab wants to run for your codebase. And the hard ceiling is real: Real-SWE had to license production code from real companies to get its tasks, so
## Top Signals
### Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases
**Source**: hackernews | **Metric**: Score: 244 / Comments: 136
A benchmark built on licensed private production code is an admission that public SWE scores no longer predict enterprise behavior — the strongest possible argument for repo-local evaluation tooling.
### I made two AIs review each other's code for 30 days. A human still caught the bug in 5 minutes.
**Source**: devto | **Metric**: Comments: 8
A documented 30-day author/skeptic agent loop that mostly worked and then failed at the one moment that mattered — validating both the two-agent structure and the need for a human-owned scoring gate.
### Cognition's SWE-2
**Source**: producthunt | **Metric**: 64% cheaper than Fable 5.1
The coding-agent market is now competing on price, which forces buyers to ask whether the cheaper model regresses on their own repo — a question only a local replay harness can answer.
### 4,768 LLM Runs, Zero Lost Sweeps: Hardening a Field-Test Runner for Timeouts, Hangs, and Cost
**Source**: devto | **Metric**: Comments: 4
CauterRule turns repeated agent failures into replay-tested standing rules across 4,768 trajectory-runs, showing the extract-replay-promote primitive is proven and can be wrapped with a repo-specific score.
### rizqinrr/viserys-agent
**Source**: github-trending | **Metric**: Stars: 627
28 SKILL.md workflows and 4 reviewer personas going viral for free proves agent workflow prompts are commoditized, so the monetizable layer must be measurement, not markdown.
### Why are AI agents lying, cheating and coordinating?
**Source**: hackernews | **Metric**: Score: 428 / Comments: 501
Bengio's piece on agents escaping containment and gaming assigned tasks is the reputational backdrop: teams are being pushed to adopt agents faster than they can verify them, which is demand for verification.
### Homebrew 7.0.0
**Source**: hackernews | **Metric**: Score: 317 / Comments: 132
Homebrew 7.0.0 adds built-in vulnerability checks and an advisory database — evidence that the package-manager layer itself is absorbing verification duties, a distribution channel a replay gate could plug into.
## Discovery
### Q1. What solo-founder products launched today?
_No strong signal found today. Possible reasons: no relevant discussion in the collection window, or signals scattered below actionable threshold._
### Q2. Which search terms or discussion threads are suddenly rising?
**Signal**: Hacker News: 'Why are AI agents lying, cheating and coordinating?' has Score 428 / Comments 501; 'Everyone should slow down AI development except for me' has Score 654 / Comments 381; 'LG denies TV spying claims, says tracking and snooping concerns not true' has Score 533 / Comments 416.
**Analysis**: The rising threads cluster around AI agent misalignment, AI development pace, and consumer privacy/surveillance. The AI-agent lying/cheating thread has unusually high comment volume (501) relative to its score, indicating active debate rather than passive upvoting. The AI slowdown thread and LG TV spying thread both have 380+ comments, showing that policy and privacy distrust are drawing sustained discussion.
**Takeaway**: Watch agent-alignment and AI-policy threads for narrative shifts; build audit and explanation tooling for agents before enterprises turn these debates into procurement requirements.
**Counter-view**: LG denies the spying claims with an official statement (Score 533 / Comments 416), showing that privacy complaints can be publicly deflected by vendor messaging, so rising discussion may not convert into near-term product demand.
### Q3. Which open-source projects are growing fast but lack a commercial offering?
**Signal**: GitHub Trending: rizqinrr/viserys-agent has Stars 627; DefiLeoo/YOINK has Stars 347; eternityspring/reelbench-skills has Stars 454.
**Analysis**: These repositories are accumulating stars quickly and present themselves as open-source projects without an obvious paid tier, hosted service, or commercial license in the excerpts. viserys-agent and YOINK in particular look like agent/note-taking infrastructure that could support a hosted product or support plan. reelbench-skills appears to be a benchmark/skills project, also without a visible commercial wrapper.
**Takeaway**: Watch viserys-agent and YOINK for sustained star growth; if they keep climbing, build a hosted or managed layer that solves deployment, sync, or compliance before the maintainers do.
**Counter-view**: DeepSeek's DeepSelect has 321 stars and is backed by a major AI lab that may ship a hosted API or integrated commercial offering, so open-source momentum alone does not guarantee a durable commercial gap.
### Q4. What are developers complaining about today?
**Signal**: Hacker News: 'Why are AI agents lying, cheating and coordinating?' Score 428 / Comments 501; Dev.to: 'I made two AIs review each other's code for 30 days. A human still caught the bug in 5 minutes.' Comments 8; Hacker News: 'Linux Zoom client proactively reading everything written to X11 clipboard' Score 220 / Comments 67.
**Analysis**: Developer complaints today center on three areas: AI agents behaving unpredictably or deceptively, AI code review missing real bugs that humans catch quickly, and desktop clients invading privacy by reading clipboard or system data. The AI agent thread has the highest comment count (501), indicating strong frustration or curiosity about agent reliability. The Dev.to post about AI reviewers shows a concrete trust failure: two AIs reviewed each other for 30 days, yet a human found the bug in five
**Takeaway**: Build audit logs and human-in-the-loop review for agent workflows; pass on relying solely on AI-to-AI review for production code quality.
**Counter-view**: Cognition's SWE-2 claims to be 64% cheaper than Fable 5.1 on Product Hunt, but the Dev.to thread with 8 comments shows a human still caught the critical bug in 5 minutes, so cheaper AI coding models do not automatically solve developer distrust.
## Tech Radar
### Q5. What is the fastest-growing developer tool this week?
**Signal**: GitHub Trending: rizqinrr/viserys-agent tops the board with 627 stars, ahead of eternityspring/reelbench-skills (454 stars) and yifanzhang-pro/recurrent-looped-tranformer (405 stars); on Hacker News, JetKVM Mini scored 331 points / 127 comments and AgentsDock, an IDE for agentic AI research, reached 57 points / 29 comments.
**Analysis**: Star velocity this week is concentrated in agent-runner plumbing rather than editors: viserys-agent at 627 stars is roughly 1.4x the next repo, and the surrounding activity (AgentsDock 57/29, Real-SWE 244/136) shows teams shopping for agent harnesses and evaluation rigs instead of another IDE. JetKVM Mini's 331 points / 127 comments is the only hardware developer tool with comparable pull, making remote-KVM access a second growth pocket alongside agent scaffolding.
**Takeaway**: Watch viserys-agent's star curve for one more week before adopting it; build your next agent harness behind a stable interface (MCP or a CLI contract) so swapping in the eventual winner costs a config change rather than a rewrite.
**Counter-view**: GitHub star spikes can be launch-driven noise — Cognition's commercial SWE-2, pitched on Product Hunt as 64% cheaper than Fable 5.1, already ships a supported alternative, so betting a team stack on a 627-star repo with no published benchmark may be the wrong trade.
### Q6. Which AI models, frameworks, or infrastructure deserve attention?
**Signal**: Product Hunt: Cognition's SWE-2, a coding model marketed as 64% cheaper than Fable 5.1; Hacker News: Real-SWE scored 244 points / 136 comments benchmarking frontier models on private enterprise codebases; GitHub Trending: deepseek-ai/DeepSelect reached 321 stars with a TopK kernel for DeepSeek Sparse Attention used in V3.2/V4/V4.1, and recurrent-looped-tranformer hit 405 stars.
**Analysis**: Two directions deserve attention: cost-per-token engineering (SWE-2 at -64% versus Fable 5.1, DeepSelect's kernel-level optimization at 321 stars) and evaluation realism (Real-SWE's 244/136 thread arguing that private enterprise code remains the real test). Perplexity Hybrid Compute, splitting research to the cloud and privacy-sensitive work to a Mac, points to a third direction: hybrid local/cloud routing as a product primitive rather than an afterthought.
**Takeaway**: Ship a private-repo eval harness this week and defer model swaps until a candidate clears it; re-run vendors' cost claims such as -64% against your own token mix before committing budget.
**Counter-view**: Real-SWE (244 points / 136 comments) plus 'Astra and Fable still hack on simple variants of alignment evals from 2025' (47 points / 7 comments) both suggest headline capability claims decay fast — a 64%-cheaper model that fails on private code simply fails 64% cheaper.
### Q7. Which platforms, products, or technologies are declining?
**Signal**: Hacker News: 'LG denies TV spying claims, says tracking and snooping concerns not true' scored 533 points / 416 comments, the day's largest thread, with the follow-up 'LG Says We're Fake News' at 199 points / 91 comments; a separate report that the Linux Zoom client proactively reads everything written to the X11 clipboard scored 220 points / 67 comments.
**Analysis**: The decline is in trust for always-on consumer platforms, not in raw usage: LG's 533/416 thread is automatic-content-recognition backlash against smart-TV tracking, and Zoom's X11 clipboard behavior (220/67) shrinks the viability of desktop clients that quietly read the whole session. Both push demand toward local-first, permission-scoped products — Perplexity Hybrid Compute's 'Mac for privacy' split and Epilude's '100% private meeting notes' are the same pressure expressed as product positionin
**Takeaway**: Defer any telemetry-by-default client design; audit your app for session-wide data access and ship local-first defaults before the next privacy thread names your product.
**Counter-view**: LG's own rebuttal still pulled 199 points / 91 comments and the ACR ad model is not shrinking commercially — competitors such as Roku keep monetizing the identical tracking pipeline, so this is trust erosion rather than platform decline.
### Q8. What tech stacks are successful Show HN / GitHub projects using?
**Signal**: Hacker News: buildprof scored 134 points / 24 comments profiling Rust (ripgrep) and Bun compile times; Dev.to: the phone PR inbox shipped on Expo + Supabase at $0 cost, the hourly e-ink newspaper uses AWS Lambda with ~100 RSS feeds and a small model exposed as an MCP server, and CauterRule turned 4,768 LLM runs into standing rules on GitHub/PyPI; Show HN: 'Durable execution without history replay' scored 19 points / 6 comments, Gonc's Go P2P netcat scored 10 points, and a multi-tenant SaaS login
**Analysis**: Winning side projects converge on a thin stack: managed serverless (Lambda, Cloudflare Free, Supabase) plus one small model for the single hard step, with MCP as the integration surface and Expo/React Native covering mobile cheaply. Native and tracing tools still cluster in Rust and Go (buildprof 134/24, Gonc 10 points), while agent frameworks such as agent-dev-team advertise zero runtime dependencies and MIT licensing to lower adoption friction.
**Takeaway**: Build your next project as Cloudflare/Supabase + one small model + an MCP endpoint, keep the CLI in Go or Rust when you need native tracing, and pass on bespoke auth/email infrastructure when Turnstile plus OAuth already covers it.
**Counter-view**: The durable-execution Show HN (19 points / 6 comments) argues that serverless glue hides replay and history failure modes — the same reason CauterRule exists after 4,768 LLM runs — so the cheapest stack may not survive production traffic or long-running agents.
## Competitive Intel
### Q9. What pricing and revenue models are indie developers discussing?
_No strong signal found today. Possible reasons: no relevant discussion in the collection window, or signals scattered below actionable threshold._
### Q10. What migration, replacement, or "X is dead" trends are emerging?
**Signal**: Dev.to article 'I stopped asking my AI agents to read the project memory. Now the server does it for them.' (Comments: 3) documents a replacement pattern: moving from agent-managed project memory to server-side memory injection.
**Analysis**: The replacement trend is not about abandoning AI agents but relocating an unreliable step: agents no longer read project memory manually; a server layer supplies context. This reduces repeated prompt overhead and token cost, and it matches broader migration from prompt hacks to infrastructure for agent context. The signal is small (3 comments), so treat it as an early workflow pattern, not a proven category shift.
**Takeaway**: Build a server-side memory/context service or library that injects project state before agent turns; ship a compatibility layer for Claude Code, Cursor, and Codex rather than another prompt-only skill.
**Counter-view**: Claude Code's native project memory and Cursor's rules files may be enough for many repos, and with only 3 comments the server-side pattern lacks proof of broad adoption; a failed migration case would be teams adding infrastructure without measuring token or recall gains.
### Q11. Which old projects or legacy needs are suddenly coming back?
**Signal**: Hacker News 'I fixed a tractor using John Deere's self-repair service. Farmers aren't sold' (Score: 113 / Comments: 115) and 'I Added a Non-Wi-Fi Mitsubishi AC to Home Assistant' (Score: 128 / Comments: 63) show legacy hardware repair and non-connected appliance integration resurfacing.
**Analysis**: Both threads point to a comeback of legacy physical needs: repairing locked-down tractors and integrating old AC units into modern smart-home stacks. The demand is driven by right-to-repair frustration and by owners wanting to keep functioning hardware instead of replacing it. The HN scores (113 and 128) suggest this is a lively, recurring pain point, not a one-off nostalgia post.
**Takeaway**: Build right-to-repair guides, parts/diagnostic marketplaces, or Home Assistant integrations for legacy appliances; ship narrow tools for John Deere, Mitsubishi AC, and similar locked-down hardware before generic repair platforms.
**Counter-view**: John Deere's authorized dealer network still controls parts and service, and Apple iPod Engraver nostalgia at 259 points shows HN attention can reward retro curiosity rather than durable willingness to pay for repair software; a failed case is farmers reverting to dealer service once warranty or parts friction appears.
## Trends
### Q12. What are the highest-frequency keywords this week?
**Signal**: Hacker News 'Why are AI agents lying, cheating and coordinating?' scored 428 with 501 comments; GitHub trending viserys-agent has 627 stars; Dev.to '4,768 LLM Runs' post has 4 comments; Real-SWE on Hacker News scored 244 with 136 comments.
**Analysis**: Across the top 90 signals, the dominant keyword cluster is agent/agentic: control gateway for AI agents, agent-dev-team for Claude Code/Cursor/Codex, AgentsDock, agent memory, agentic loops, and TDD for AI agents. The second cluster is verification and review: two AIs reviewing each other's code, PR inbox, LiveReview, TDD handoff, and fault-injection tests. Memory/context and MCP also recur as infrastructure terms. Benchmark/SWE language is concentrated around Real-SWE, Cognition SWE-2, and priv
**Takeaway**: Build agent observability, review gates, and memory/context plumbing now; ship narrow agent workflows with human checkpoints before expanding autonomy.
**Counter-view**: Cognition's SWE-2 claims it is 64% cheaper than Fable 5.1, so model-layer price cuts could reduce willingness to pay for custom agent-control layers.
### Q13. Which concepts are cooling down?
**Signal**: Dev.to 'I made two AIs review each other's code for 30 days' has 8 comments; Dev.to 'When Plan Mode Was No Longer Enough' has 1 comment; Real-SWE on Hacker News scored 244 with 136 comments.
**Analysis**: The cooling concept is not AI coding broadly, but single-pass generation or vibe coding as a sufficient workflow. The strongest signals increasingly treat code generation as commodity and shift attention to review, verification, TDD, plan-mode limits, and human checkpoints. RAG-only boilerplate also looks cooler: only one Reddit post explicitly centers Docker, vector database configuration, embeddings, and starter APIs, while memory APIs, context apps, MCP servers, and agent memory get separate
**Takeaway**: Defer standalone codegen wrappers and RAG-boilerplate CLIs; build verification, human review, and memory/control layers instead.
**Counter-view**: Cognition's SWE-2 still markets a coding model 64% cheaper than Fable 5.1, and JetKVM Mini drew 331 Hacker News score, so cost-optimized model and hardware layers can keep generation-centric products competitive.
### Q14. Which new terms or categories are emerging from zero?
_No strong signal found today. Possible reasons: no relevant discussion in the collection window, or signals scattered below actionable threshold._
## Action
### Q15. What is most worth spending 2 hours on today?
**Signal**: Hacker News: Real-SWE has Score 244 / Comments 136; Dev.to: AI code-review cross-check post has Comments: 8; Product Hunt: Cognition's SWE-2 is 64% cheaper than Fable 5.1.
**Analysis**: The strongest actionable cluster is verification of AI-generated code. Real-SWE shows demand for private, real-world enterprise codebase benchmarking (244 score / 136 comments). The Dev.to post about two AIs reviewing each other for 30 days (8 comments) highlights the failure mode: AI reviewers miss bugs that a human catches in 5 minutes. Cognition's SWE-2 at 64% cheaper increases the volume of AI-written code, which increases the review bottleneck. Two hours should go into a narrow cross-model
**Takeaway**: Build a throwaway GitHub Action that runs two different model reviewers plus a 5-minute human checklist on one real AI-generated PR, then measure what it catches.
**Counter-view**: Cognition's SWE-2 is 64% cheaper than Fable 5.1, so coding-model vendors may bundle review into the IDE and make a standalone review gate hard to monetize.
### Q16. Why not the other two candidate directions?
**Signal**: Reddit: control gateway for AI agents has no metric; Reddit: agent-dev-team has 21 roles, 4 seniority tiers, 26 skills, MIT license; Hacker News: Real-SWE has Score 244 / Comments 136; Dev.to: AI review post has Comments: 8.
**Analysis**: The two alternate directions are an AI agent control gateway and a drop-in AI engineering team. The control gateway is infrastructure-heavy: it needs deep API/tool integrations and a trust/security sales cycle before it can prove value in two hours. The agent-dev-team is a prompt/workflow pack with 21 roles, 4 tiers, 26 skills, MIT license, and zero runtime deps; it is easy to clone and hard to defend as a paid product. The cross-model review gate has a tighter wedge: one PR, two model reviewers
**Takeaway**: Defer the control gateway and the agent-dev-team wrapper; build the review gate because it has a faster testable loop and a clearer failure metric.
**Counter-view**: Cognition's SWE-2 is 64% cheaper than Fable 5.1, which shows the model layer is deflating fast and may commoditize a review gate too.
### Q17. What is the fastest validation step?
**Signal**: Dev.to: AI code-review post has Comments: 8; Hacker News: Real-SWE has Score 244 / Comments 136; Reddit: control gateway for AI agents has no metric.
**Analysis**: The fastest validation is not a survey or landing page. It is a controlled bug-catch test. Take five real AI-generated pull requests, inject or identify one known bug in each, run two different model reviewers, then give a human five minutes to review the AI output. Record whether the combined gate catches the bug and how long it takes. The Dev.to post already shows the pain: two AIs reviewed code for 30 days, but a human caught the bug in 5 minutes (8 comments). Real-SWE's 244 score / 136 comme
**Takeaway**: Run five AI-PR review tests today and ship the results as a public benchmark post to recruit design partners.
**Counter-view**: If AgentsDock (Score 57 / Comments 29) already ships review inside an IDE, a separate five-PR test may not convince users to switch workflows.
### Q18. What product should this become over the weekend?
**Signal**: Dev.to: PR Inbox for phone has Comments: 2; Dev.to: AI code-review post has Comments: 8; Hacker News: Real-SWE has Score 244 / Comments: 136.
**Analysis**: The weekend product should be MergeGate, a GitHub App plus mobile approval inbox for AI-generated PRs. It runs two different model reviewers, shows a diff summary, flags disagreements, and requires a human sign-off before merge. The Dev.to PR Inbox post (2 comments) shows appetite for phone-based PR triage. The AI code-review post (8 comments) shows the need for a human final check. Real-SWE (244 score / 136 comments) gives the enterprise benchmark angle for later.
**Takeaway**: Ship MergeGate as a weekend MVP: GitHub App, cross-model review diff, and one-tap phone approval with an evidence log.
**Counter-view**: Cognition's SWE-2 is 64% cheaper than Fable 5.1; if GitHub or Cognition bundles a PR inbox, MergeGate's mobile approval becomes a feature rather than a product.
### Q19. How should initial pricing and packaging look?
**Signal**: Product Hunt: Cognition's SWE-2 is 64% cheaper than Fable 5.1; Hacker News: Real-SWE has Score 244 / Comments 136; Reddit: agent-dev-team has 21 roles, 4 seniority tiers, 26 skills, MIT license.
**Analysis**: Use open-core packaging because the agent-dev-team alternative is MIT with 21 roles and zero runtime deps. Free: public repos and the OSS GitHub Action. Team: $29 per developer per month for private repos, cross-model review, and a 30-day evidence log. Enterprise: $499 per month for private-code benchmark runs, audit export, and SSO. The paid hook is not prompts; it is the evidence trail and benchmark against real private code, which Real-SWE validates with 244 score / 136 comments. Cognition SW
**Takeaway**: Ship the free OSS action now and charge $29/dev/mo for private evidence dashboards; defer enterprise pricing until five teams ask for audit exports.
**Counter-view**: agent-dev-team is MIT with zero runtime deps, so free prompt packs can undercut simple paid review wrappers unless MergeGate owns the evidence and benchmark layer.
### Q20. What is the strongest counter-view?
**Signal**: Product Hunt: Cognition's SWE-2 is 64% cheaper than Fable 5.1; Hacker News: AgentsDock has Score 57 / Comments: 29; Hacker News: Real-SWE has Score 244 / Comments: 136.
**Analysis**: The strongest counter-view is that AI code review is not a standalone product. Cognition's SWE-2 at 64% cheaper than Fable 5.1 makes the coding model itself cheaper and more capable, and AgentsDock at Score 57 / Comments: 29 already embeds agentic review inside an IDE. If the model vendor or IDE absorbs review, MergeGate becomes a thin workflow layer with little pricing power. Real-SWE at 244 score / 136 comments may also become a vendor-run benchmark rather than a neutral tooling wedge.
**Takeaway**: Watch this risk closely and validate five paid pilots before building any enterprise-only features beyond the weekend MVP.
**Counter-view**: Cognition's SWE-2 is 64% cheaper than Fable 5.1 and AgentsDock has Score 57 / Comments: 29, so the counter-view is that review is a bundled feature, not a standalone product.
## Action Plan
**2-Hour Build**: Write a single CLI that (1) reads `git log --merges` and exports 20 merged PRs with their base commit, diff and the tests that were touched, (2) spins up an author agent that reproduces the change on a clean checkout and a skeptic agent prompted with 'assume this is broken, produce the input that loses a customer money', (3) runs the repo's own test command in a container, and (4) prints one table: PR id, author-agent pass, skeptic-agent objection count, wall-clock, token cost. No UI, no databas
**Why This Wins**: It answers the only question a team actually has when a new coding model launches — 'does it regress on our code?' — using artifacts the team already owns. It cannot be copied by a model vendor because the value is the frozen 20-PR set and its history, not the prompt. And it fits the same shape as CauterRule's extract-replay-promote loop, which is already validated at 4,768 runs but has no repo-level scorecard on top.
**Why Not Alternatives**:
- agent-dev-team and viserys-agent already give away 21 reviewer roles and 28 SKILL.md workflows under MIT — packaging prompts is free and therefore not a product.
- Cognition's SWE-2 and other hosted agents sell you a model at a headline price cut (64% vs. Fable 5.1), not a verdict about your repository, so the buyer still has no answer at swap time.
- Generic LLM-as-a-judge tooling scores diffs in isolation and never replays against your merged history, so it cannot detect a regression introduced by an agent upgrade.
- Public SWE leaderboards are structurally unable to help here — Real-SWE had to license private enterprise codebases precisely because public tasks stopped discriminating between frontier models.
**Fastest Validation**: Run the 20-PR replay on one friendly team's repo this week and hand the tech lead a blind ranking sheet: 20 diffs, no model labels, ranked 1-20. Compare that ranking to the harness score. Ship publicly only if the harness agrees on at least 15 of 20 — otherwise publish the disagreement itself as the first post, because that failure analysis is the product's real credibility asset.
**Weekend Expansion**: Add a second and third model so the report becomes a selection matrix ('model A: 17/20 pass at $0.42/PR; model B: 15/20 at $0.27/PR'), a per-PR cost column computed from real token counts, and a `--gate` mode that exits non-zero when a new agent version drops below the frozen baseline. Finish with a scheduled weekly run that posts the scorecard to a Slack webhook so the number stays alive instead of rotting in a README.