4take 0.2.0__tar.gz → 0.2.2__tar.gz

{4take-0.2.0 → 4take-0.2.2}/AGENTS.md

@@ -118,11 +118,13 @@ usable review; read `agreement` against `summary.distinct_models`.
 
 ```json
 {
-  "version": "0.2.0",
+  "version": "0.2.2",
   "timestamp": "2026-08-16T14:20:00Z",
+  "config": {"models_requested": 12, "min_responses": 4, "timeout_seconds": 600, "strict": false},
   "summary": {
     "task_hash": "25d297f773c3",
     "responses_received": 4,
+    "distinct_models": 4,
     "responses_failed": 0,
     "total_latency_ms": 3200,
     "context": {"files": ["src/mailer.py", "src/users.py (lines 40-90)"], "tokens": 812, "dropped": []}

{4take-0.2.0 → 4take-0.2.2}/CHANGELOG.md

@@ -7,6 +7,53 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.2.2] - 2026-10-26
+
+One fix to what the review shows; how findings are grouped and counted is unchanged.
+
+### Changed
+- **The README leads with what the council found, not with what an agent did with it.** On the
+  weak-primary test's 50 PRs the council's findings caught 37 bugs and Gemini 3.8 Flash alone 24
+  (15 against 2 discordant, exact McNemar p ≈ 0.002, not pre-registered; `docs/benchmark` gives the
+  strict figures). The model given those findings caught 26, which is now the reason for
+  `report_for_user` rather than the section's headline. The MCP section no longer opens with the
+  20-bug Opus test, which stays in the ROADMAP.
+- The README's JSON example shows `served_model`, `served_provider`, `finish_reason`, `truncated`,
+  `duplicate_of`, `config` and `summary.distinct_models`, which the CLI has emitted all along and
+  the agent guides tell readers to rely on.
+- `uv.lock` pins the patched PyJWT, certifi, cryptography, mcp and urllib3. It is used by CI and
+  `uv sync`; installs from PyPI resolve their own versions and already had the fixed ones.
+
+### Fixed
+- **A consensus entry no longer takes a bare symbol as its title.** Cohere's north-mini-code
+  sometimes writes the symbol where the title belongs (`[CRITICAL] INJECTION find_user (def
+  find_user)`, 11 of its 75 recorded findings). Among equally severe findings the first in order set
+  the title, so in a live review on 26 October all five entries read "find_user", in the CLI and in
+  `report_for_user` alike. An entry now takes a title that names the defect when one of its findings
+  has one, and otherwise the start of the issue, cut at a word. Grouping and agreement are
+  unchanged: `evals/score.py` prints the same report before and after.
+
+## [0.2.1] - 2026-10-26
+
+Documentation and packaging only; the review behaves exactly as in 0.2.0.
+
+### Changed
+- **The benchmark comparison says when each side was recorded, and what the models may have
+  seen.** The commercial reviewers' comments on Greptile's 50 PRs date from late July 2025;
+  4take's runs from October 2026. The PRs and their bug descriptions have been public since 2025,
+  so a model trained later may have seen them. Both favour 4take, and the README, `docs/benchmark`
+  and the 0.2.0 release notes now say so.
+- **Claims the measurements do not support are gone.** The agent diagram no longer says the
+  council "dismisses hallucinations" and yields "final clean code", `CLAUDE.md` no longer says it
+  eliminates single-model hallucinations, and "universal", "seamlessly" and "beautiful" are out of
+  the README, `llms.txt`, the CLI help and the PyPI summary.
+- **Two README badges carried other projects' logos** (FastAPI's for FastMCP, OpenAI's for
+  OpenRouter). They are replaced by the PyPI version.
+
+### Removed
+- `evals/backfill_duplicates.py`, a one-off migration of recordings made before 0.2.0. It stays in
+  the git history.
+
 ## [0.2.0] - 2026-10-26
 
 ### Added

{4take-0.2.0 → 4take-0.2.2}/CLAUDE.md

@@ -2,7 +2,8 @@
 
 ## What is this?
 4take is a CLI and MCP server that queries 4 independent LLMs in parallel to review code.
-You pipe code into it, and it returns multi-model opinions to eliminate single-model hallucinations.
+You pipe code into it, and it returns four models' reviews with their findings grouped by how
+many raised each one.
 
 It reviews four axes, not just defects: **correctness**, **security**, **structure**
 (responsibility in the wrong layer, coupling that spreads one behavioural change across

{4take-0.2.0 → 4take-0.2.2}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.5
 Name: 4take
-Version: 0.2.0
-Summary: Universal LLM Council CLI — get 4 independent AI opinions on your code in one command
+Version: 0.2.2
+Summary: Code review by a council of four free LLMs on OpenRouter, with findings ranked by how many reviewers raised them
 Project-URL: Homepage, https://github.com/aleksbuss/4take
 Project-URL: Repository, https://github.com/aleksbuss/4take
 Project-URL: Issues, https://github.com/aleksbuss/4take/issues
@@ -40,16 +40,15 @@ Description-Content-Type: text/markdown
 
 [![Python](https://img.shields.io/badge/Python-3.10+-3776AB?style=flat&logo=python&logoColor=white)](https://www.python.org/)
 [![License: BUSL-1.1](https://img.shields.io/badge/License-BUSL--1.1-yellow.svg)](https://github.com/aleksbuss/4take/blob/main/LICENSE)
-[![FastMCP](https://img.shields.io/badge/FastMCP-Ready-6366f1?style=flat&logo=fastapi&logoColor=white)](https://modelcontextprotocol.io/)
-[![Powered by OpenRouter](https://img.shields.io/badge/API-OpenRouter-000000?style=flat&logo=openai&logoColor=white)](https://openrouter.ai/)
+[![PyPI](https://img.shields.io/pypi/v/4take)](https://pypi.org/project/4take/)
 [![CI Status](https://github.com/aleksbuss/4take/actions/workflows/ci.yml/badge.svg)](https://github.com/aleksbuss/4take/actions)
 [![Linted with Ruff](https://img.shields.io/badge/linting-ruff-D7FF64.svg)](https://github.com/astral-sh/ruff)
 
 **Multi-Model AI Council for Code Review.**  
-*Get 4 independent AI opinions on your code via OpenRouter in one command.*
+*Four free models from four vendors review your code in one command, and their findings are ranked by how many raised each one.*
 
 On Greptile's own 50-PR bug benchmark, a council of free models found 37 of 50 bugs:  
-level with Greptile (41), ahead of Copilot and CodeRabbit. [How it was measured](#-against-the-paid-reviewers-greptiles-benchmark)
+level with Greptile (41), ahead of Copilot and CodeRabbit, whose reviews date from July 2025. [How it was measured](#-against-the-paid-reviewers-greptiles-benchmark)
 
 [⚡ 30-Second Quickstart](#-30-second-quickstart) •
 [🤖 AI Agent & MCP Integration](#-ai-coding-agents--mcp-integration) •
@@ -114,33 +113,33 @@ git diff --cached | 4take
 
 ## 🤖 AI Coding Agents & MCP Integration
 
-`4take` is built for **AI Coding Agents** (Claude Code, Cursor, Antigravity IDE, Windsurf, Cline, Roo Code, Codex). It provides agents with a native "Council of LLMs" to audit their code in parallel before presenting it to the user.
+`4take` works with AI coding agents (Claude Code, Cursor, Antigravity IDE, Windsurf, Cline, Roo Code, Codex) as an MCP server or a CLI: the agent sends its code to the council and gets the findings back ranked by agreement, to show the user.
 
-Whether the council improves what a given coding agent ships has not been measured. The eval below scores the council against single reviewers on a fixed corpus, and the one test with a frontier model as the primary (20 bugs) found no gain. The [ROADMAP](https://github.com/aleksbuss/4take/blob/main/ROADMAP.md) lists what is planned.
+An agent tends to keep the findings most reviewers agree on and drop the rest, which loses much of what the council catches ([measured](#-read-the-council-yourself)). So the MCP payload opens with every finding and its agreement, for the agent to show you.
 
 ```text
-                       ┌─────────────────────────┐
-                       │    AI Coding Agent      │
-                       │ (Chairman / Synthesizer)│
-                       └────────────┬────────────┘
-                                    │
-                                    │ 1. Submits code draft
-                                    ▼
-                         ┌────────────────────┐
-                         │   4take Council    │
-                         └──────────┬─────────┘
-                                    │
-       ┌────────────────────────────┼────────────────────────────┐
-       ▼                            ▼                            ▼
-  Model 1 (Nemotron)           Model 2 (Laguna)            Model 3 (Cohere)
-  (Flags Race Condition)      (Flags Race Condition)       (Suggests Type Hint)
-       └────────────────────────────┬────────────────────────────┘
-                                    │ 2. Synthesizes consensus
-                                    │ 3. Dismisses hallucinations
-                                    ▼
-                       ┌─────────────────────────┐
-                       │ Final Clean Code & PR   │
-                       └─────────────────────────┘
+                    ┌──────────────────────┐
+                    │   AI coding agent    │
+                    └──────────┬───────────┘
+                               │ 1. sends its code or diff
+                               ▼
+                    ┌──────────────────────┐
+                    │    4take council     │
+                    └──────────┬───────────┘
+          ┌────────────┬───────┴────┬────────────┐
+          ▼            ▼            ▼            ▼
+     reviewer 1   reviewer 2   reviewer 3   reviewer 4
+     (vendor A)   (vendor B)   (vendor C)   (vendor D)
+          └────────────┴───────┬────┴────────────┘
+                               │ 2. findings grouped and ranked
+                               ▼    by how many raised each one
+                    ┌──────────────────────┐
+                    │   report_for_user    │
+                    │  e.g. "3/4 HIGH ..." │
+                    └──────────┬───────────┘
+                               │ 3. the agent shows every line and
+                               ▼    checks each against the code
+                             user
 ```
 
 ### Option A: FastMCP Server (`4take-mcp`)
@@ -224,11 +223,13 @@ cat modified_file.py | 4take --format json --accept-risk --quiet
 
 ```json
 {
-  "version": "0.2.0",
+  "version": "0.2.2",
   "timestamp": "2026-08-16T14:20:00Z",
+  "config": {"models_requested": 12, "min_responses": 4, "timeout_seconds": 600, "strict": false},
   "summary": {
     "task_hash": "25d297f773c3",
     "responses_received": 4,
+    "distinct_models": 4,
     "responses_failed": 0,
     "total_latency_ms": 3200,
     "context": {"files": ["src/mailer.py", "src/users.py (lines 40-90)"], "tokens": 812, "dropped": []}
@@ -251,12 +252,22 @@ cat modified_file.py | 4take --format json --accept-risk --quiet
   "responses": [
     {
       "model": "openrouter/nvidia/nemotron-3-super-120b-a12b:free",
+      "served_model": "nvidia/nemotron-3-super-120b-a12b:free",
+      "served_provider": "NVIDIA",
       "content": "### Code Review\n1. Found SQL Injection on line 8...",
+      "finish_reason": "stop",
+      "truncated": false,
+      "duplicate_of": null,
       "latency_ms": 1150
     },
     {
       "model": "openrouter/poolside/laguna-s-2.1:free",
+      "served_model": "poolside/laguna-s-2.1:free",
+      "served_provider": "Poolside",
       "content": "### Security & Performance\n1. Use parameterized query for user_id...",
+      "finish_reason": "stop",
+      "truncated": false,
+      "duplicate_of": null,
       "latency_ms": 1420
     }
   ],
@@ -299,7 +310,7 @@ council took about 8½ minutes, and 6 ran to the 600 s limit. On the 28 short sn
 corpus the median was about 50 s, and 27 of 28 councils seated four.
 
 ### Ready-to-Use Agent Configuration Files:
-- **Universal Agent Instructions**: [`AGENTS.md`](https://github.com/aleksbuss/4take/blob/main/AGENTS.md)
+- **Instructions for any agent**: [`AGENTS.md`](https://github.com/aleksbuss/4take/blob/main/AGENTS.md)
 - **Claude Code**: [`CLAUDE.md`](https://github.com/aleksbuss/4take/blob/main/CLAUDE.md)
 - **Antigravity Skill**: [`SKILL.md`](https://github.com/aleksbuss/4take/blob/main/SKILL.md)
 - **Machine Summary**: [`llms.txt`](https://github.com/aleksbuss/4take/blob/main/llms.txt)
@@ -308,15 +319,15 @@ corpus the median was about 50 s, and 27 of 28 councils seated four.
 
 ## ✨ Features
 
-- 🚀 **1-Command Multi-Model Council**: Pipe any code, diff, or file directly into `4take`.
+- 🚀 **One Command**: Pipe a file or a diff into `4take`; four free models review it in parallel.
 - 🧠 **Auto-Discovery Ranked on Evidence**: Fetches and filters the free models on OpenRouter, then ranks them by the success rate and median latency they have actually delivered on your machine. The name-based heuristic (parameter count, context window) survives only as the cold-start prior for models you have not run yet.
 - 🔎 **Four Review Axes**: Correctness, security, structure and simplicity — not just defects. Every structural finding has to name its cost, so the breadth does not become a flood of taste.
 - 🤝 **Consensus Grouping**: Findings several models raised independently are grouped and ranked by how many raised them, so the cross-referencing is done for you rather than left in four parallel reviews.
-- 🔄 **Fault-Tolerant Dynamic Rotation**: If a model hits a rate limit (HTTP 429) or fails, the engine seamlessly rotates in replacement candidates from a pre-ranked 12-model pool with exponential backoff.
+- 🔄 **Fault-Tolerant Dynamic Rotation**: A rate-limited model (HTTP 429) is retried with exponential backoff; one that keeps failing gives its seat to the next model in a pre-ranked pool of up to 12.
 - 🆓 **Free Models Only**: Requests go only to OpenRouter's free (`:free`) models, and a paid model is rejected before anything is sent. Free has limits: roughly 4 to 8 reviews a day on an account that has never been topped up, roughly 80 to 160 once it has ([details](#-openrouter-free-tier--rate-limits)).
 - 🛡️ **Secret Redaction**: Automatically scrubs API keys, passwords, and authorization tokens before sending payloads over the wire. The code itself still goes to third-party free endpoints ([privacy](#%EF%B8%8F-privacy--security)).
 - 🚦 **Cross-Process Rate Limiter**: Shared file-locked token bucket ensures concurrent terminal sessions or background agents never trigger rate limit bans.
-- 🎨 **Rich Terminal UI & JSON Output**: Beautiful ANSI Markdown rendering in terminal, plus `--format json` for automated scripts & CI pipelines.
+- 🎨 **Terminal and JSON Output**: Markdown rendered in the terminal, or `--format json` for scripts and CI.
 
 ---
 
@@ -324,9 +335,9 @@ corpus the median was about 50 s, and 27 of 28 councils seated four.
 
 Greptile published a benchmark of 50 pull requests from five open-source projects (Cal.com,
 Discourse, Grafana, Keycloak, Sentry), each with one labelled bug, together with the comments five
-commercial AI reviewers left on them. 4take reviewed the same 50 diffs using only OpenRouter's free
-models (commit 2f07956, 8 October 2026). Its reviews were graded blind, against a rubric and an
-analysis written down before the run.
+commercial AI reviewers left on them in late July 2025. Fifteen months later 4take reviewed the
+same 50 diffs using only OpenRouter's free models (commit 2f07956, 8 October 2026). Its reviews were
+graded blind, against a rubric and an analysis written down before the run.
 
 | reviewer | benchmark standard, 50 PRs | strict, 42 verified labels |
 | :--- | ---: | ---: |
@@ -355,6 +366,10 @@ analysis written down before the run.
   4take. A strong signal, not proof. The pool has moved on since: two of the four most-seated
   models of that run (apodex-1.1-mini, ling-3.0-flash-sante) have left the free tier, so the same
   councils cannot be seated again.
+- **Fifteen months apart, on public data.** The commercial reviews were written in July 2025 and
+  4take's in October 2026, and models improved in between: those tools would probably score higher
+  today. The PRs and the benchmark's bug descriptions have been public since 2025, so a model trained
+  later may have seen them, and nothing here rules that out. Both effects favour 4take.
 - **Where it is weakest.** Large Java diffs: Keycloak 5 of 10. On two of them, and on one Sentry
   diff, only one free vendor reasoned through to an answer. A council took a median of six
   minutes on these diffs and up to the 600 s timeout.
@@ -365,30 +380,30 @@ diffs are other projects' code, so the recordings and the graded packets are not
 
 ---
 
-## 🧪 When the Council Helps, and When It Doesn't
+## 🧪 Read the Council Yourself
 
-**Agreement is the signal.** On the eval corpus, over three recordings with four reviewers seated,
-a finding all four raised held a real, labelled defect 96% of the time; one raised by two or three
-held one 60% of the time; one raised by a single reviewer 41%. A single reviewer's finding is a
-pointer to check, not a verdict, and four in ten of them are real.
-
-**Read the council's output; don't only hand it to another model.** A
+**Four free models found more than a commercial model alone.** A
 [pre-registered test](https://github.com/aleksbuss/4take/blob/main/docs/benchmark/README.md#does-the-council-help-a-weaker-model)
-on the same 50 PRs gave Gemini 3.8 Flash each diff alone, and again with the council's output as
-notes (three runs each, graded blind):
+on the same 50 PRs, in October 2026, gave Gemini 3.8 Flash (thinking high) each diff alone, and
+again with the council's findings as notes; three runs each, graded blind:
 
-| | benchmark standard, 50 PRs |
+| | bugs found, of 50 |
 | :--- | ---: |
+| **the council's findings, read directly** | **37** |
 | Gemini 3.8 Flash alone | 24 |
-| Gemini 3.8 Flash with the council's notes | 26 (no difference shown, p = 0.69) |
-| the council's output, read directly | 37 |
+| Gemini 3.8 Flash given the council's findings | 26 |
+
+The council caught 15 bugs the model missed, and the model 2 that the council missed (exact McNemar
+p ≈ 0.002). This comparison was not pre-registered, and it credits the council's best reviewer per
+PR against the model's majority of three runs. Both sides were recorded in the same month on the
+same diffs, so the time gap of the commercial comparison does not arise here, and any exposure of
+the PRs in training data applies to both.
 
-In an exploratory breakdown, the model took up 3 of the 4 bugs only the council caught that three or
-four reviewers shared, and none of the 11 that one or two raised: it kept the consensus and dropped
-the long tail with the noise. (The council's row credits its best reviewer per PR against the
-model's majority of three runs, which favours the council.) An earlier test with Claude Opus as the
-primary model showed no gain either (15 of 20 real bugs from 4take's own history alone, 14 with the
-council). 4take does not make a model a better reviewer by being pasted into its context.
+**An agent that reads the findings keeps only part of them.** Given the council's findings as
+notes, the model caught 26, not 37; against the model alone that is no difference shown (p = 0.69).
+In an exploratory breakdown it took up 3 of the 4 bugs only the council caught that three or four
+reviewers shared, and none of the 11 that one or two raised: it kept the consensus and dropped the
+long tail.
 
 **So 4take puts every finding in front of you.** Over MCP the user sees only what the agent
 relays, so the payload opens with `report_for_user`: one line per finding with its agreement
@@ -396,6 +411,11 @@ relays, so the payload opens with `report_for_user`: one line per finding with i
 act on agreement 3 or more, and to check agreement 1 or 2 against the code instead of dropping it.
 The CLI prints the same ranking to your terminal.
 
+**Agreement tells you where to look first.** On the eval corpus, over three recordings with four
+reviewers seated, a finding all four raised held a real, labelled defect 96% of the time; one
+raised by two or three held one 60% of the time; one raised by a single reviewer 41%. A single
+reviewer's finding is a pointer to check, not a verdict, and four in ten of them are real.
+
 ---
 
 ## 🔎 What the Council Reviews

{4take-0.2.0 → 4take-0.2.2}/README.md

@@ -4,16 +4,15 @@
 
 [![Python](https://img.shields.io/badge/Python-3.10+-3776AB?style=flat&logo=python&logoColor=white)](https://www.python.org/)
 [![License: BUSL-1.1](https://img.shields.io/badge/License-BUSL--1.1-yellow.svg)](https://github.com/aleksbuss/4take/blob/main/LICENSE)
-[![FastMCP](https://img.shields.io/badge/FastMCP-Ready-6366f1?style=flat&logo=fastapi&logoColor=white)](https://modelcontextprotocol.io/)
-[![Powered by OpenRouter](https://img.shields.io/badge/API-OpenRouter-000000?style=flat&logo=openai&logoColor=white)](https://openrouter.ai/)
+[![PyPI](https://img.shields.io/pypi/v/4take)](https://pypi.org/project/4take/)
 [![CI Status](https://github.com/aleksbuss/4take/actions/workflows/ci.yml/badge.svg)](https://github.com/aleksbuss/4take/actions)
 [![Linted with Ruff](https://img.shields.io/badge/linting-ruff-D7FF64.svg)](https://github.com/astral-sh/ruff)
 
 **Multi-Model AI Council for Code Review.**  
-*Get 4 independent AI opinions on your code via OpenRouter in one command.*
+*Four free models from four vendors review your code in one command, and their findings are ranked by how many raised each one.*
 
 On Greptile's own 50-PR bug benchmark, a council of free models found 37 of 50 bugs:  
-level with Greptile (41), ahead of Copilot and CodeRabbit. [How it was measured](#-against-the-paid-reviewers-greptiles-benchmark)
+level with Greptile (41), ahead of Copilot and CodeRabbit, whose reviews date from July 2025. [How it was measured](#-against-the-paid-reviewers-greptiles-benchmark)
 
 [⚡ 30-Second Quickstart](#-30-second-quickstart) •
 [🤖 AI Agent & MCP Integration](#-ai-coding-agents--mcp-integration) •
@@ -78,33 +77,33 @@ git diff --cached | 4take
 
 ## 🤖 AI Coding Agents & MCP Integration
 
-`4take` is built for **AI Coding Agents** (Claude Code, Cursor, Antigravity IDE, Windsurf, Cline, Roo Code, Codex). It provides agents with a native "Council of LLMs" to audit their code in parallel before presenting it to the user.
+`4take` works with AI coding agents (Claude Code, Cursor, Antigravity IDE, Windsurf, Cline, Roo Code, Codex) as an MCP server or a CLI: the agent sends its code to the council and gets the findings back ranked by agreement, to show the user.
 
-Whether the council improves what a given coding agent ships has not been measured. The eval below scores the council against single reviewers on a fixed corpus, and the one test with a frontier model as the primary (20 bugs) found no gain. The [ROADMAP](https://github.com/aleksbuss/4take/blob/main/ROADMAP.md) lists what is planned.
+An agent tends to keep the findings most reviewers agree on and drop the rest, which loses much of what the council catches ([measured](#-read-the-council-yourself)). So the MCP payload opens with every finding and its agreement, for the agent to show you.
 
 ```text
-                       ┌─────────────────────────┐
-                       │    AI Coding Agent      │
-                       │ (Chairman / Synthesizer)│
-                       └────────────┬────────────┘
-                                    │
-                                    │ 1. Submits code draft
-                                    ▼
-                         ┌────────────────────┐
-                         │   4take Council    │
-                         └──────────┬─────────┘
-                                    │
-       ┌────────────────────────────┼────────────────────────────┐
-       ▼                            ▼                            ▼
-  Model 1 (Nemotron)           Model 2 (Laguna)            Model 3 (Cohere)
-  (Flags Race Condition)      (Flags Race Condition)       (Suggests Type Hint)
-       └────────────────────────────┬────────────────────────────┘
-                                    │ 2. Synthesizes consensus
-                                    │ 3. Dismisses hallucinations
-                                    ▼
-                       ┌─────────────────────────┐
-                       │ Final Clean Code & PR   │
-                       └─────────────────────────┘
+                    ┌──────────────────────┐
+                    │   AI coding agent    │
+                    └──────────┬───────────┘
+                               │ 1. sends its code or diff
+                               ▼
+                    ┌──────────────────────┐
+                    │    4take council     │
+                    └──────────┬───────────┘
+          ┌────────────┬───────┴────┬────────────┐
+          ▼            ▼            ▼            ▼
+     reviewer 1   reviewer 2   reviewer 3   reviewer 4
+     (vendor A)   (vendor B)   (vendor C)   (vendor D)
+          └────────────┴───────┬────┴────────────┘
+                               │ 2. findings grouped and ranked
+                               ▼    by how many raised each one
+                    ┌──────────────────────┐
+                    │   report_for_user    │
+                    │  e.g. "3/4 HIGH ..." │
+                    └──────────┬───────────┘
+                               │ 3. the agent shows every line and
+                               ▼    checks each against the code
+                             user
 ```
 
 ### Option A: FastMCP Server (`4take-mcp`)
@@ -188,11 +187,13 @@ cat modified_file.py | 4take --format json --accept-risk --quiet
 
 ```json
 {
-  "version": "0.2.0",
+  "version": "0.2.2",
   "timestamp": "2026-08-16T14:20:00Z",
+  "config": {"models_requested": 12, "min_responses": 4, "timeout_seconds": 600, "strict": false},
   "summary": {
     "task_hash": "25d297f773c3",
     "responses_received": 4,
+    "distinct_models": 4,
     "responses_failed": 0,
     "total_latency_ms": 3200,
     "context": {"files": ["src/mailer.py", "src/users.py (lines 40-90)"], "tokens": 812, "dropped": []}
@@ -215,12 +216,22 @@ cat modified_file.py | 4take --format json --accept-risk --quiet
   "responses": [
     {
       "model": "openrouter/nvidia/nemotron-3-super-120b-a12b:free",
+      "served_model": "nvidia/nemotron-3-super-120b-a12b:free",
+      "served_provider": "NVIDIA",
       "content": "### Code Review\n1. Found SQL Injection on line 8...",
+      "finish_reason": "stop",
+      "truncated": false,
+      "duplicate_of": null,
       "latency_ms": 1150
     },
     {
       "model": "openrouter/poolside/laguna-s-2.1:free",
+      "served_model": "poolside/laguna-s-2.1:free",
+      "served_provider": "Poolside",
       "content": "### Security & Performance\n1. Use parameterized query for user_id...",
+      "finish_reason": "stop",
+      "truncated": false,
+      "duplicate_of": null,
       "latency_ms": 1420
     }
   ],
@@ -263,7 +274,7 @@ council took about 8½ minutes, and 6 ran to the 600 s limit. On the 28 short sn
 corpus the median was about 50 s, and 27 of 28 councils seated four.
 
 ### Ready-to-Use Agent Configuration Files:
-- **Universal Agent Instructions**: [`AGENTS.md`](https://github.com/aleksbuss/4take/blob/main/AGENTS.md)
+- **Instructions for any agent**: [`AGENTS.md`](https://github.com/aleksbuss/4take/blob/main/AGENTS.md)
 - **Claude Code**: [`CLAUDE.md`](https://github.com/aleksbuss/4take/blob/main/CLAUDE.md)
 - **Antigravity Skill**: [`SKILL.md`](https://github.com/aleksbuss/4take/blob/main/SKILL.md)
 - **Machine Summary**: [`llms.txt`](https://github.com/aleksbuss/4take/blob/main/llms.txt)
@@ -272,15 +283,15 @@ corpus the median was about 50 s, and 27 of 28 councils seated four.
 
 ## ✨ Features
 
-- 🚀 **1-Command Multi-Model Council**: Pipe any code, diff, or file directly into `4take`.
+- 🚀 **One Command**: Pipe a file or a diff into `4take`; four free models review it in parallel.
 - 🧠 **Auto-Discovery Ranked on Evidence**: Fetches and filters the free models on OpenRouter, then ranks them by the success rate and median latency they have actually delivered on your machine. The name-based heuristic (parameter count, context window) survives only as the cold-start prior for models you have not run yet.
 - 🔎 **Four Review Axes**: Correctness, security, structure and simplicity — not just defects. Every structural finding has to name its cost, so the breadth does not become a flood of taste.
 - 🤝 **Consensus Grouping**: Findings several models raised independently are grouped and ranked by how many raised them, so the cross-referencing is done for you rather than left in four parallel reviews.
-- 🔄 **Fault-Tolerant Dynamic Rotation**: If a model hits a rate limit (HTTP 429) or fails, the engine seamlessly rotates in replacement candidates from a pre-ranked 12-model pool with exponential backoff.
+- 🔄 **Fault-Tolerant Dynamic Rotation**: A rate-limited model (HTTP 429) is retried with exponential backoff; one that keeps failing gives its seat to the next model in a pre-ranked pool of up to 12.
 - 🆓 **Free Models Only**: Requests go only to OpenRouter's free (`:free`) models, and a paid model is rejected before anything is sent. Free has limits: roughly 4 to 8 reviews a day on an account that has never been topped up, roughly 80 to 160 once it has ([details](#-openrouter-free-tier--rate-limits)).
 - 🛡️ **Secret Redaction**: Automatically scrubs API keys, passwords, and authorization tokens before sending payloads over the wire. The code itself still goes to third-party free endpoints ([privacy](#%EF%B8%8F-privacy--security)).
 - 🚦 **Cross-Process Rate Limiter**: Shared file-locked token bucket ensures concurrent terminal sessions or background agents never trigger rate limit bans.
-- 🎨 **Rich Terminal UI & JSON Output**: Beautiful ANSI Markdown rendering in terminal, plus `--format json` for automated scripts & CI pipelines.
+- 🎨 **Terminal and JSON Output**: Markdown rendered in the terminal, or `--format json` for scripts and CI.
 
 ---
 
@@ -288,9 +299,9 @@ corpus the median was about 50 s, and 27 of 28 councils seated four.
 
 Greptile published a benchmark of 50 pull requests from five open-source projects (Cal.com,
 Discourse, Grafana, Keycloak, Sentry), each with one labelled bug, together with the comments five
-commercial AI reviewers left on them. 4take reviewed the same 50 diffs using only OpenRouter's free
-models (commit 2f07956, 8 October 2026). Its reviews were graded blind, against a rubric and an
-analysis written down before the run.
+commercial AI reviewers left on them in late July 2025. Fifteen months later 4take reviewed the
+same 50 diffs using only OpenRouter's free models (commit 2f07956, 8 October 2026). Its reviews were
+graded blind, against a rubric and an analysis written down before the run.
 
 | reviewer | benchmark standard, 50 PRs | strict, 42 verified labels |
 | :--- | ---: | ---: |
@@ -319,6 +330,10 @@ analysis written down before the run.
   4take. A strong signal, not proof. The pool has moved on since: two of the four most-seated
   models of that run (apodex-1.1-mini, ling-3.0-flash-sante) have left the free tier, so the same
   councils cannot be seated again.
+- **Fifteen months apart, on public data.** The commercial reviews were written in July 2025 and
+  4take's in October 2026, and models improved in between: those tools would probably score higher
+  today. The PRs and the benchmark's bug descriptions have been public since 2025, so a model trained
+  later may have seen them, and nothing here rules that out. Both effects favour 4take.
 - **Where it is weakest.** Large Java diffs: Keycloak 5 of 10. On two of them, and on one Sentry
   diff, only one free vendor reasoned through to an answer. A council took a median of six
   minutes on these diffs and up to the 600 s timeout.
@@ -329,30 +344,30 @@ diffs are other projects' code, so the recordings and the graded packets are not
 
 ---
 
-## 🧪 When the Council Helps, and When It Doesn't
+## 🧪 Read the Council Yourself
 
-**Agreement is the signal.** On the eval corpus, over three recordings with four reviewers seated,
-a finding all four raised held a real, labelled defect 96% of the time; one raised by two or three
-held one 60% of the time; one raised by a single reviewer 41%. A single reviewer's finding is a
-pointer to check, not a verdict, and four in ten of them are real.
-
-**Read the council's output; don't only hand it to another model.** A
+**Four free models found more than a commercial model alone.** A
 [pre-registered test](https://github.com/aleksbuss/4take/blob/main/docs/benchmark/README.md#does-the-council-help-a-weaker-model)
-on the same 50 PRs gave Gemini 3.8 Flash each diff alone, and again with the council's output as
-notes (three runs each, graded blind):
+on the same 50 PRs, in October 2026, gave Gemini 3.8 Flash (thinking high) each diff alone, and
+again with the council's findings as notes; three runs each, graded blind:
 
-| | benchmark standard, 50 PRs |
+| | bugs found, of 50 |
 | :--- | ---: |
+| **the council's findings, read directly** | **37** |
 | Gemini 3.8 Flash alone | 24 |
-| Gemini 3.8 Flash with the council's notes | 26 (no difference shown, p = 0.69) |
-| the council's output, read directly | 37 |
+| Gemini 3.8 Flash given the council's findings | 26 |
+
+The council caught 15 bugs the model missed, and the model 2 that the council missed (exact McNemar
+p ≈ 0.002). This comparison was not pre-registered, and it credits the council's best reviewer per
+PR against the model's majority of three runs. Both sides were recorded in the same month on the
+same diffs, so the time gap of the commercial comparison does not arise here, and any exposure of
+the PRs in training data applies to both.
 
-In an exploratory breakdown, the model took up 3 of the 4 bugs only the council caught that three or
-four reviewers shared, and none of the 11 that one or two raised: it kept the consensus and dropped
-the long tail with the noise. (The council's row credits its best reviewer per PR against the
-model's majority of three runs, which favours the council.) An earlier test with Claude Opus as the
-primary model showed no gain either (15 of 20 real bugs from 4take's own history alone, 14 with the
-council). 4take does not make a model a better reviewer by being pasted into its context.
+**An agent that reads the findings keeps only part of them.** Given the council's findings as
+notes, the model caught 26, not 37; against the model alone that is no difference shown (p = 0.69).
+In an exploratory breakdown it took up 3 of the 4 bugs only the council caught that three or four
+reviewers shared, and none of the 11 that one or two raised: it kept the consensus and dropped the
+long tail.
 
 **So 4take puts every finding in front of you.** Over MCP the user sees only what the agent
 relays, so the payload opens with `report_for_user`: one line per finding with its agreement
@@ -360,6 +375,11 @@ relays, so the payload opens with `report_for_user`: one line per finding with i
 act on agreement 3 or more, and to check agreement 1 or 2 against the code instead of dropping it.
 The CLI prints the same ranking to your terminal.
 
+**Agreement tells you where to look first.** On the eval corpus, over three recordings with four
+reviewers seated, a finding all four raised held a real, labelled defect 96% of the time; one
+raised by two or three held one 60% of the time; one raised by a single reviewer 41%. A single
+reviewer's finding is a pointer to check, not a verdict, and four in ten of them are real.
+
 ---
 
 ## 🔎 What the Council Reviews

{4take-0.2.0 → 4take-0.2.2}/SECURITY.md

@@ -8,7 +8,7 @@ Only the latest release receives security fixes. Today that is 0.2.x.
 
 Please do not open a public issue for a vulnerability.
 
-- Use GitHub's **Report a vulnerability** button (Security tab, once the repository is public), or
+- Use GitHub's **Report a vulnerability** button on the Security tab, or
 - write to **takelicensing@gmail.com** with the subject `4take security`.
 
 Include the output of `4take --version`, the steps to reproduce, and what an attacker gains. A

{4take-0.2.0 → 4take-0.2.2}/SKILL.md

@@ -7,7 +7,7 @@ description: Get multiple independent AI opinions on your code using the 4take C
 ## When to use
 - After writing a new feature or complex algorithm
 - Before committing critical security, database, or async logic
-- When you want multiple independent AI reviews to catch hallucinations and blindspots
+- When you want several independent AI reviews to catch what one model misses
 - When code works and you want to know whether its structure will survive the next change
 
 Reviews four axes: correctness, security, structure and simplicity. Takes code, a diff or

{4take-0.2.0 → 4take-0.2.2}/evals/README.md

@@ -19,4 +19,3 @@ carries.
 | `recordings_ctx/`, `runs_ctx/` | the cross-file corpus recorded without context, with it, and with `--auto-context` |
 | `record.py` | re-records against live models; the only step that uses the network |
 | `score.py` | scores recordings offline |
-| `backfill_duplicates.py` | one-off migration that restored `duplicate_of` in older recordings |

{4take-0.2.0 → 4take-0.2.2}/llms.txt

@@ -1,6 +1,6 @@
-# 4take — Universal LLM Council CLI (OpenRouter)
+# 4take — LLM council for code review (OpenRouter free models)
 
-> Universal CLI and FastMCP server to query 4 independent LLMs in parallel via OpenRouter for code review, with findings grouped by how many reviewers raised them. Source-available under the Business Source License 1.1: free for individuals, solo developers and volunteer projects; organizations need a commercial license.
+> CLI and MCP server that sends code to 4 free LLMs from different vendors via OpenRouter for review, with findings grouped by how many reviewers raised them. Source-available under the Business Source License 1.1: free for individuals, solo developers and volunteer projects; organizations need a commercial license.
 
 ## Quickstart
 

{4take-0.2.0 → 4take-0.2.2}/pyproject.toml

@@ -10,7 +10,7 @@ name = "4take"
 # Single source of truth: `src/fourtake/__init__.py`. Two hand-maintained version strings
 # drift, and the one that ships on PyPI is not the one `4take --version` prints.
 dynamic = ["version"]
-description = "Universal LLM Council CLI — get 4 independent AI opinions on your code in one command"
+description = "Code review by a council of four free LLMs on OpenRouter, with findings ranked by how many reviewers raised them"
 readme = "README.md"
 license = {text = "BUSL-1.1"}
 requires-python = ">=3.10"

4take-0.2.2/src/fourtake/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.2.2"