tare-cli 0.3.0__tar.gz → 0.4.0__tar.gz

{tare_cli-0.3.0 → tare_cli-0.4.0}/CONCEPT.md

@@ -6,10 +6,13 @@ that it could not see your setup (`tare: not proven (blind)`). It refuses to sta
 until the reading is `tare: 0.00`, unless you pass `--allow-dirty`. It supports Claude Code, Codex, Pi
 and the Antigravity CLI (`agy`) on Linux and WSL, and Claude Code, Codex and Pi on macOS.
 
-This document describes the design and the measurements behind it. The [README](README.md)
-explains the words this document uses: room, probe, reading, leak, declared, dirty twin, check, tail and
-handoff. Every claim marked **measured** was observed on WSL2. The section
-[Measurements](#measurements) lists the agent versions and links each result.
+This document describes the design and the measurements behind it. The documentation explains the
+words this document uses: room, probe, reading, leak, declared and control run in
+[docs/words.md](docs/words.md), check, restart and handover note in [docs/measuring.md](docs/measuring.md).
+The experiment records under `experiments/` call the control run the control run and a restart a
+tail.
+Every claim marked **measured** was observed on WSL2. The section [Measurements](#measurements)
+lists the agent versions and links each result.
 
 ## Why
 
@@ -71,11 +74,11 @@ your instructions, memories, plugins, hooks and skills. That is not enough.
   server keeps the request, which is the context the agent CLI put together, and answers "ok".
   The idea is the blank run from analytical chemistry, which runs the whole procedure without a
   sample to see what the procedure itself brings in.
-- **Control:** the same agent run in your real setup, the dirty twin, must show your context. The
+- **Control:** the same agent run in your real setup, the control run, must show your context. The
   risk was that features that come with the login (connected accounts, account skills) would
   not load against a fake server, so the probe would read clean while a real run is not.
   **Measured:** for Claude Code this risk did not occur. Against the fake model server, the
-  dirty twin still loaded the connected accounts, account skills, plugin skills, hook output,
+  control run still loaded the connected accounts, account skills, plugin skills, hook output,
   global instructions, memories and the email. The room loaded none of them except the email.
   Pointing Claude Code at a custom base URL turns its tool search off and puts every tool schema
   into the request. So the probe and the runs set `ENABLE_TOOL_SEARCH=true` to see the context as
@@ -122,22 +125,30 @@ session. This has not been tested, so tare refuses to start when the copy would
 within the next hour. **Measured:** the Antigravity CLI's Google login keeps its refresh token
 when it renews, so a room may renew its own copy.
 
+**Measured** on Ubuntu on WSL2 with Claude Code 2.1.310: Claude Code renews its token only in
+the last minutes before it expires. A session started outside tare with 57 minutes left, and
+another with 7 minutes left, kept the old token; one started with under 6 minutes left renewed
+it, and the new token held for 8 hours. So for most of the refused hour no session outside tare
+can renew the login, and the refusal names the expiry time instead of asking for one. A run
+known to be shorter may ask for less with `TARE_MIN_LOGIN_MINUTES`, never below 10 minutes, so
+the copy does not start inside the renewal window (#126).
+
 ### Probe and reading
 
 The probe has the three parts of layer 3: the agent in the room against the fake model server,
-the dirty twin, and the reach script. It sorts what it finds into **kinds of context**,
+the control run, and the reach script. It sorts what it finds into **kinds of context**,
 each checked on its own: instructions, memories, home path, MCP servers (connected accounts count
 here), skills, plugins, extensions, instruction files above the project, the account email, and
 reach.
 
 - A kind of context counts only when you have it. tare learns that from your files, for example
   a global instructions file or a skill folder.
-- If you have a kind of context and the dirty twin does not show it, the probe could not see
-  it. tare then runs the dirty twin once more after 10 seconds. If the second dirty twin still
+- If you have a kind of context and the control run does not show it, the probe could not see
+  it. tare then runs the control run once more after 10 seconds. If the second control run still
   does not show that kind of context, the reading is `tare: not proven (blind)`. **Measured:**
-  with six runs starting together, a dirty twin sent its request before its MCP servers were
+  with six runs starting together, a control run sent its request before its MCP servers were
   connected.
-- Every probe makes no paid model call, because both the room and the dirty twin talk to the
+- Every probe makes no paid model call, because both the room and the control run talk to the
   fake model server.
 
 ### Agents
@@ -145,7 +156,7 @@ reach.
 Each agent needs its own way to the fake model server, its own login copy and its own markers:
 lines of your own files that the probe looks for in a request.
 
-| Agent | How the room is set up | Way to the fake model server | Kinds of context the dirty twin showed on the measuring machine |
+| Agent | How the room is set up | Way to the fake model server | Kinds of context the control run showed on the measuring machine |
 |---|---|---|---|
 | Claude Code | `CLAUDE_CONFIG_DIR` and `HOME` in the room, a copy of `.credentials.json`, `--strict-mcp-config --setting-sources project,local` | `ANTHROPIC_BASE_URL` | instructions, home path, MCP servers, skills, plugins, email, reach |
 | Codex | `CODEX_HOME` and `HOME` in the room, a copy of `auth.json` only, `--disable remote_plugin` | `-c openai_base_url=<server>/v1` keeps the ChatGPT login | instructions, memories, home path, skills, plugins, custom agents, reach |
@@ -172,15 +183,15 @@ What each agent needed:
   Codex's default model. Pass `-m` when you compare runs.
 - **Pi:** a fresh agent directory is clean on its own and keeps the project's `AGENTS.md`, so the
   room uses no flags. `--no-context-files` would drop the project's `AGENTS.md`. tare finds Pi's
-  extension tools by comparing the dirty twin's tools with a run that uses `--no-extensions`.
+  extension tools by comparing the control run's tools with a run that uses `--no-extensions`.
   With standard input left open, Pi waits for piped input and never answers, so tare closes it.
 - **Antigravity CLI:** the fake model server speaks Google's Cloud Code `v1internal` API and
   offers one model. It lists that model in the groups the CLI shows models in
   (`agentModelSorts`), because the CLI rejects a model that no group lists. The
-  CLI writes into `~/.gemini` on every run, so its dirty twin runs with an overlay over
+  CLI writes into `~/.gemini` on every run, so its control run runs with an overlay over
   `~/.gemini`. A user namespace mounts the overlay, and every write goes to a temporary file
   system that disappears after the run. **Measured:** `~/.gemini` kept its fingerprint (paths,
-  sizes, modification times) through every dirty twin. The CLI writes each skill's path into
+  sizes, modification times) through every control run. The CLI writes each skill's path into
   its skill list, so the probe compares skill lines without the path. Without that, a skill that
   leaked into the room would not match yours, because its path differs.
 - **Instruction files above the project:** Claude Code's one-shot mode, which the probe uses, did
@@ -190,7 +201,7 @@ What each agent needed:
 - **Long prompts:** Linux limits a single argument to about 128 KiB. tare passes the task to
   every agent as an argument, so a longer task fails. Pi could read it from a file with `@file`.
 
-### Capsules, trail and handoff
+### Capsules, trail and handover note
 
 Calibrate, Swap and Cliff need a run they can stop and continue. During a recorded run, a hook
 archives the workspace after every tool call. Each archive is a **capsule**. The hook comes from
@@ -204,16 +215,16 @@ A run continues from a capsule in one of two ways:
   resumes a session file cut after the step, with `codex exec resume <id>`. Pi resumes with
   `--session`. **Measured:** a resumed Codex session kept its id and finished the remaining plan
   from step 1.
-- **A handoff.** The **trail** is a record of a run in the same format for every agent: the task,
-  the messages, the tool calls and their results. tare writes a handoff from any prefix of it, so
+- **A handover note.** The **trail** is a record of a run in the same format for every agent: the task,
+  the messages, the tool calls and their results. tare writes a handover note from any prefix of it, so
   any agent can continue any run without reading another agent's session format. The Antigravity
   CLI stores its sessions in a format that tare cannot cut, so it always continues with a
-  handoff. Its trail comes from `transcript_full.jsonl`.
+  handover note. Its trail comes from `transcript_full.jsonl`.
 
 ### Calibrate
 
 `tare calibrate` starts each side (an agent with its arguments) several times from a fresh copy
-of the project and reports each side's pass rate with a 95% interval (Wilson). A side can have
+of the project and reports each side's pass rate with a likely range (a 95% Wilson interval). A side can have
 its own prompt, so two skills that start with different commands compare in one call. Without a
 check, a run counts as finished when its agent exits with 0, and tare keeps its workspace for
 scoring elsewhere. **Measured:** in a field test that compared two ideation skills, 18 rooms all
@@ -234,31 +245,32 @@ judge found a real one-cent rounding bug and an overflow.
 
 ### Swap
 
-Swap cuts two runs at several points. At each cut, each agent continues each run's workspace
-by handoff. Swap reports a state effect (how much the state of the workspace matters), a model
-effect (which agent does better), and, where an agent can continue its own saved session, the
-handoff cost. At cut 0 both workspaces are the untouched project, so the state effect there must
-be zero. Swap checks this, and the check is called the null check. When the state effect and the
-model effect are within 0.1 of each other, Swap says they explain the failure about equally.
+Swap stops two runs at several handover points. At each one, each agent continues each run's
+workspace from a handover note. Swap reports a workspace effect (how much the state of the
+workspace matters), an agent effect (which agent does better), and, where an agent can continue
+its own saved session, the note cost. At the handover point 0 both workspaces are the untouched
+project, so the workspace effect there must be zero. Swap checks this, and the check is called the
+start check. When the workspace effect and the agent effect are within 0.1 of each other, Swap
+says they explain the failure about equally.
 
-**Measured:** against scripted models with a known truth, the null check passed and the blame
-passed from the model to the workspace between cut 0 and cut 0.5, as the scripts were written
-to cause.
+**Measured:** against scripted models with a known truth, the start check passed and the blame
+passed from the agent to the workspace between the handovers at 0 and 0.5, as the scripts were
+written to cause.
 
 ### Cliff
 
 Cliff restarts a failed run from its capsules and names the step after which it no longer
-succeeds. The baseline is the set of tails from the start. A step counts as good while its
-tails pass at least half as often as the baseline. Cliff halves the range between the last good
-and the first bad step, then adds tails to the two neighbours until their intervals separate or
-the budget is spent. It reports a model gap instead of a step only when the upper end of the
-baseline's 95% interval lies below 0.2.
+succeeds. Fresh starts are the restarts from the beginning. A step counts as good while its
+restarts pass at least half as often as a fresh start. Cliff halves the range between the last
+good and the first bad step, then adds restarts to the two neighbours until their likely ranges
+separate or the budget is spent. It reports that the task is too hard for the model instead of a
+step only when the upper end of the fresh starts' likely range (a 95% interval) lies below 0.2.
 
 **Measured:** against a scripted model whose wrong step is known, Cliff found step 4 with
-separated intervals in 18 tails. Cliff stays experimental, because neither of two real test
+separated likely ranges in 18 restarts. Cliff stays experimental, because neither of two real test
 rounds had a failure that sat in one step (see [Measurements](#measurements)). Swap already
-shows, at three cuts, how often an agent passes when it continues its own workspace. When a real
-Swap shows that rate dropping between two cuts, Cliff can narrow the drop down to one step. Only
+shows, at three handover points, how often an agent passes when it continues its own workspace. When a real
+Swap shows that rate dropping between two handover points, Cliff can narrow the drop down to one step. Only
 then does Cliff get new experiments.
 
 ### Tokens and cost
@@ -279,7 +291,9 @@ judge read 186.4k tokens in and $0.093 in total, the Codex part estimated from L
 
 Calibrate, Swap and Cliff write everything into a run directory: a journal (one JSON line per
 event), a `recipe.json`, a `usage.json` with the totals, and the live output of every agent. The dashboard reads only the run
-directory, so `tare watch` shows a finished run the same way as a live one. The recipe holds the
+directory, so `tare watch` shows a finished run the same way as a live one. Each run page opens with
+the answer: a sentence from fixed templates per verdict, worked out from the run directory the way
+Calibrate, Swap and Cliff decide, so a run recorded by an older tare reads like a new one. The recipe holds the
 command, a fingerprint of the project and the agent versions. `tare rerun` repeats a run and
 names what has changed since.
 
@@ -290,19 +304,19 @@ macOS 27.0 (arm64) on 2026-10-05 and 2026-10-06.
 
 | What | Agent and version | Result | Source |
 |---|---|---|---|
-| Dirty twin and room, fake model server | Claude Code 2.1.289 | the dirty twin showed every kind of context, the room only the email | [results](experiments/dirty-twin/RESULTS.md) |
-| Dirty twin and room | Codex CLI 0.160.0 | the room read `tare: 0.00` | [results](experiments/dirty-twin-codex/RESULTS.md) |
-| Dirty twin and room | Pi 1.0.2 | the room read `tare: 0.00` without flags | [results](experiments/dirty-twin-pi/RESULTS.md) |
-| Dirty twin and room | Antigravity CLI 1.2.16 | the room read `tare: 0.00` without flags, and `~/.gemini` stayed unchanged | [results](experiments/dirty-twin-agy/RESULTS.md) |
+| Control run and room, fake model server | Claude Code 2.1.289 | the control run showed every kind of context, the room only the email | [results](experiments/dirty-twin/RESULTS.md) |
+| Control run and room | Codex CLI 0.160.0 | the room read `tare: 0.00` | [results](experiments/dirty-twin-codex/RESULTS.md) |
+| Control run and room | Pi 1.0.2 | the room read `tare: 0.00` without flags | [results](experiments/dirty-twin-pi/RESULTS.md) |
+| Control run and room | Antigravity CLI 1.2.16 | the room read `tare: 0.00` without flags, and `~/.gemini` stayed unchanged | [results](experiments/dirty-twin-agy/RESULTS.md) |
 | Rooms given copies of real settings on purpose | all four | Claude Code 4 leaks, Codex 6, Pi 5, Antigravity CLI 3, each named with its source | the results above |
-| Scripted Cliff and Swap | Claude Code against scripted models | Cliff found step 4. Swap passed the null check and found the blame passing from the model to the workspace between cut 0 and cut 0.5 | `experiments/cliff-scripted`, `experiments/swap-scripted` |
+| Scripted Cliff and Swap | Claude Code against scripted models | Cliff found step 4. Swap passed the start check and found the blame passing from the agent to the workspace between the handovers at 0 and 0.5 | `experiments/cliff-scripted`, `experiments/swap-scripted` |
 | Real task 1: a data migration | Claude Code with Sonnet and Haiku, Codex | Deleting the source file after a wrong-encoding import was planned as the point of no return. Swap showed that the workspace could still be repaired. Haiku passed about 1 in 14 runs, Codex with gpt-6.1-sol 4 of 4 | [results](experiments/migration/RESULTS.md) |
 | Real task 2: a web page judged by Sonnet | Claude Code with Haiku, Codex with gpt-6.1-sol | Swap confirmed all four predictions of the plan written before the runs. Cliff called a failed run unlucky, and a later rescoring of the same page confirmed it | [results](experiments/html/RESULTS.md) |
 | Field test: two ideation skills compared | Claude Code 2.1.289 with Opus 5.5 | all 18 rooms read `tare: 0.00`, including a skill that started `claude -p` inside the room | not published |
 | macOS: what a Seatbelt room has to handle | Claude Code 2.1.289, Codex CLI 0.159.0, Pi 0.99.1 | Claude Code's login is in the Keychain. With only the home directory denied, `/var/folders`, `/private/tmp` and the pasteboard stayed readable. Agents' own sandboxes fail inside a room (`sandbox_apply: Operation not permitted`) | comments on epic #85 |
 | macOS: a Seatbelt room with the real login | same | each agent answered one prompt. The room and its login copy were removed afterwards, and the Keychain item was unchanged (digest before and after). `tare probe` read `tare: 0.00` for all three | comments on epic #85 |
-| macOS: dirty twin, Seatbelt room, and parts of the real setup copied into the room | same | every room read `tare: 0.00`. Each copied part that reached the request was named with its source, and nothing under the home directory, under `/var/folders`, in `/private/tmp` or on the pasteboard was reachable from inside. Two kinds of Codex context are invisible to the probe on any system (#109) | [results](experiments/dirty-twin-macos/RESULTS.md) |
-| macOS: scripted Cliff and Swap, the judge | Claude Code 2.1.290 against scripted models; Microsoft Edge for the judge's screenshot | Cliff found step 4. Swap passed the null check and found the blame passing from the model to the workspace between cut 0 and cut 0.5. `tare judge` rendered a page and scored it. With a project that turns Claude Code's sandbox on, every Bash call failed in a Seatbelt room until tare turned it off | comments on epic #85 |
+| macOS: control run, Seatbelt room, and parts of the real setup copied into the room | same | every room read `tare: 0.00`. Each copied part that reached the request was named with its source, and nothing under the home directory, under `/var/folders`, in `/private/tmp` or on the pasteboard was reachable from inside. Two kinds of Codex context are invisible to the probe on any system (#109) | [results](experiments/dirty-twin-macos/RESULTS.md) |
+| macOS: scripted Cliff and Swap, the judge | Claude Code 2.1.290 against scripted models; Microsoft Edge for the judge's screenshot | Cliff found step 4. Swap passed the start check and found the blame passing from the agent to the workspace between the handovers at 0 and 0.5. `tare judge` rendered a page and scored it. With a project that turns Claude Code's sandbox on, every Bash call failed in a Seatbelt room until tare turned it off | comments on epic #85 |
 
 ## Open questions and ideas
 
@@ -312,11 +326,11 @@ macOS 27.0 (arm64) on 2026-10-05 and 2026-10-06.
   a fresh room never did. Cached account data is the suspected cause. Not tested.
 - **Claude Code login renewal:** whether a renewal inside the room logs out the real session. Not
   tested, so tare avoids it.
-- **Codex handoffs:** Codex can run several commands in one tool call, which makes its handoffs
+- **Codex handover notes:** Codex can run several commands in one tool call, which makes its handover notes
   long.
-- **Cost:** every tail is a full agent run, so a Cliff search or a Swap costs as much as its tails.
+- **Cost:** every restart is a full agent run, so a Cliff search or a Swap costs as much as its restarts.
 - **macOS:** the room is built on Seatbelt, released in 0.3.0 (epic #85). Not done: the
-  Antigravity CLI, whose dirty twin needs Linux's overlay, so tare stops with a message there.
+  Antigravity CLI, whose control run needs Linux's overlay, so tare stops with a message there.
   Not tested: whether a program started through launchd or XPC runs outside the room's profile.
 - **Other probes, not built:** weighing the first request's input tokens before and after adding
   junk to the real setup, and file tripwires that report which secret files a run opened.

tare_cli-0.4.0/PKG-INFO

@@ -0,0 +1,180 @@
+Metadata-Version: 2.5
+Name: tare-cli
+Version: 0.4.0
+Summary: Prove your coding agent starts without your personal setup before you measure it.
+Project-URL: Homepage, https://github.com/a-ratz/tare
+Project-URL: Repository, https://github.com/a-ratz/tare
+Project-URL: Issues, https://github.com/a-ratz/tare/issues
+Project-URL: Concept, https://github.com/a-ratz/tare/blob/main/CONCEPT.md
+Author: Andre Ratzenberger
+License-Expression: MIT
+License-File: LICENSE
+Keywords: agent-evaluation,ai-agents,claude-code,codex,coding-agents,evals,llm-evaluation,sandbox
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: Operating System :: MacOS
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Software Development :: Testing
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+
+<div align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/.github/logo-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/.github/logo-light.svg">
+    <img alt="tare" src="https://raw.githubusercontent.com/a-ratz/tare/main/.github/logo-light.svg" width="260">
+  </picture>
+
+  <p><b>Zero the scale before you weigh.</b><br>
+  Prove your coding agent starts without your personal setup, then measure it.</p>
+</div>
+
+<div align="center">
+
+[![License: MIT][license-shield]][license-url]
+[![PyPI][version-shield]][version-url]
+[![Python 3.11+][python-shield]][python-url]
+[![Linux | WSL | macOS][platform-shield]][platform-url]
+[![Claude Code | Codex | Pi | Antigravity][agent-shield]][agent-url]
+
+</div>
+
+<div align="center">
+  <a href="#quick-start">Quick start</a> &middot;
+  <a href="#features">Features</a> &middot;
+  <a href="https://github.com/a-ratz/tare/blob/main/docs/README.md">Documentation</a> &middot;
+  <a href="https://github.com/a-ratz/tare/blob/main/CONCEPT.md">Concept</a> &middot;
+  <a href="https://github.com/a-ratz/tare/issues/new?template=bug_report.md">Report Bug</a>
+</div>
+
+<br>
+
+---
+
+Your coding agent knows you. Every run loads your global instructions, memories, skills, MCP
+servers and connected accounts. At work that is the point. In an evaluation it is noise. Compare
+runs with and without a skill on your own machine, and both runs carry your setup. You measure
+your machine, not the skill.
+
+**tare** starts the agent in a room that holds nothing of yours and proves the room is clean
+before every run. It works with Claude Code, Codex, Pi and Google's Antigravity CLI on Linux and
+WSL, and with the first three on macOS.
+
+```text
+$ tare probe claude
+tare probe · claude 2.1.289 · /home/me/tare
+  control   your real setup shows: instructions, home path, mcp, skills, plugins, email, reach
+  leaks     none
+  declared
+    email        account email                                subscription login
+    credentials  /home/tare/.claude-config/.credentials.json  copy of the login, needed by the CLI
+tare: 0.00
+```
+
+`tare: 0.00` is the zeroed scale. Nothing of your setup reached the model except what the reading
+declares. A reading with leaks names each one and the file it came from, and the agent does not
+start.
+
+## Quick start
+
+```bash
+uv tool install tare-cli            # or: pipx install tare-cli
+tare probe claude                   # prove the room is clean
+tare claude -- -p "fix the tests"   # run Claude Code in it
+```
+
+On Linux and WSL, tare also needs bubblewrap (`sudo apt install bubblewrap`). The
+[install guide](https://github.com/a-ratz/tare/blob/main/docs/install.md) has the details per
+platform, and [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md) explains
+every line of the output above.
+
+To compare agents, skills or models, `tare calibrate`, `tare swap` and `tare cliff` start many runs
+and open a live dashboard that leads with the answer:
+
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/docs/assets/dashboard-calibrate-dark.png">
+  <img alt="tare's dashboard for a Calibrate run: Claude Code scored higher, 68 against 15 out of 100, with the scores of every run" src="https://raw.githubusercontent.com/a-ratz/tare/main/docs/assets/dashboard-calibrate-light.png">
+</picture>
+
+## Features
+
+- **Measures what the model receives.** The probe records the request the agent sends to a fake
+  model server on your machine. It makes no paid calls and never asks the agent about itself,
+  because an agent asked about its own context can refuse or be wrong.
+  [How it works](https://github.com/a-ratz/tare/blob/main/docs/how-it-works.md)
+- **Knows when it cannot see.** A control run starts the agent in your real setup. If the probe could
+  not see your setup there, the reading is `tare: not proven (blind)` instead of a `0.00` that
+  proves nothing. [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md)
+- **Keeps your files out and your login working.** The room has no home directory of yours, no
+  credentials of your other tools and no secret environment variables. It gets a fresh copy of
+  your login, which tare deletes after the run.
+  [Running agents in rooms](https://github.com/a-ratz/tare/blob/main/docs/running-agents.md)
+- **Compares skills, plugins and models.** `tare calibrate` starts each side several times from
+  scratch and reports pass rates with their intervals.
+  [Calibrate](https://github.com/a-ratz/tare/blob/main/docs/calibrate.md)
+- **Scores what no test can decide.** `tare judge` lets a second agent score a page by your
+  rubric, without agent or model names, and `tare judge-noise` shows how far its scores vary.
+  [Judge](https://github.com/a-ratz/tare/blob/main/docs/judge.md)
+- **Tells the workspace from the agent.** `tare swap` lets two agents continue each other's
+  unfinished work and reports whether the workspace or the agent decided the outcome.
+  [Swap](https://github.com/a-ratz/tare/blob/main/docs/swap.md)
+- **Finds the step where a failed run was lost** (experimental). `tare cliff` restarts the run
+  from its saved steps and names the step after which it no longer passes.
+  [Cliff](https://github.com/a-ratz/tare/blob/main/docs/cliff.md)
+- **Live dashboard, cost and reruns.** The dashboard leads with each run's answer and follows it
+  live, each report ends with tokens and cost, and `tare rerun` repeats a run from its recipe.
+  [Dashboard and rerun](https://github.com/a-ratz/tare/blob/main/docs/dashboard.md) ·
+  [Tokens and cost](https://github.com/a-ratz/tare/blob/main/docs/tokens-and-cost.md)
+
+## When to use
+
+| Use tare when | Look elsewhere when |
+|---|---|
+| you compare agent runs with and without a skill, plugin or prompt | you need protection from a hostile agent, because the room has network access and a working login |
+| you benchmark agents on your own machine | you use the Antigravity CLI on macOS (not supported there yet) |
+| you want a fresh-machine run without a fresh machine | you log in with an API key only (untested) |
+| an agent failed a task and you want to know which step lost it | you want a single run explained without running the agent again, because Cliff runs it many times |
+
+## Documentation
+
+| | |
+|---|---|
+| **Getting started** | [Install](https://github.com/a-ratz/tare/blob/main/docs/install.md) · [The words tare uses](https://github.com/a-ratz/tare/blob/main/docs/words.md) · [Running agents in rooms](https://github.com/a-ratz/tare/blob/main/docs/running-agents.md) · [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md) |
+| **Measuring** | [Overview](https://github.com/a-ratz/tare/blob/main/docs/measuring.md) · [Calibrate](https://github.com/a-ratz/tare/blob/main/docs/calibrate.md) · [Judge](https://github.com/a-ratz/tare/blob/main/docs/judge.md) · [Swap](https://github.com/a-ratz/tare/blob/main/docs/swap.md) · [Cliff](https://github.com/a-ratz/tare/blob/main/docs/cliff.md) · [Tokens and cost](https://github.com/a-ratz/tare/blob/main/docs/tokens-and-cost.md) · [Dashboard and rerun](https://github.com/a-ratz/tare/blob/main/docs/dashboard.md) |
+| **Background** | [How it works](https://github.com/a-ratz/tare/blob/main/docs/how-it-works.md) · [Concept: the design and every measurement](https://github.com/a-ratz/tare/blob/main/CONCEPT.md) · [Experiments](https://github.com/a-ratz/tare/tree/main/experiments) |
+
+## Status
+
+tare is on PyPI as `tare-cli`. Claude Code, Codex and Pi run on Linux, WSL and macOS; the
+Antigravity CLI runs on Linux and WSL, and stops with a message on macOS for now. What comes next
+is in the [open epics](https://github.com/a-ratz/tare/issues?q=is%3Aissue+is%3Aopen+label%3Atype%3Aepic).
+
+## Contributing
+
+Every change starts as an issue. Larger work is an epic, split into features, and one epic
+becomes one branch and one pull request. [AGENTS.md](https://github.com/a-ratz/tare/blob/main/AGENTS.md)
+describes the process, the checks and the conventions. Run the tests with `uv run pytest`.
+
+## License
+
+MIT. See [LICENSE](https://github.com/a-ratz/tare/blob/main/LICENSE).
+
+---
+
+Crafted with [Readme Craft](https://github.com/motiful/readme-craft)
+
+[license-shield]: https://img.shields.io/badge/license-MIT-blue.svg
+[license-url]: https://github.com/a-ratz/tare/blob/main/LICENSE
+[version-shield]: https://img.shields.io/pypi/v/tare-cli.svg
+[version-url]: https://pypi.org/project/tare-cli/
+[python-shield]: https://img.shields.io/badge/python-3.11%2B-3776AB.svg
+[python-url]: https://www.python.org/
+[platform-shield]: https://img.shields.io/badge/platform-Linux%20%7C%20WSL%20%7C%20macOS-555555.svg
+[platform-url]: https://github.com/a-ratz/tare/blob/main/docs/install.md
+[agent-shield]: https://img.shields.io/badge/agents-Claude%20Code%20%7C%20Codex%20%7C%20Pi%20%7C%20Antigravity-D97757.svg
+[agent-url]: https://github.com/a-ratz/tare/blob/main/docs/install.md

tare_cli-0.4.0/README.md

@@ -0,0 +1,155 @@
+<div align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/.github/logo-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/.github/logo-light.svg">
+    <img alt="tare" src="https://raw.githubusercontent.com/a-ratz/tare/main/.github/logo-light.svg" width="260">
+  </picture>
+
+  <p><b>Zero the scale before you weigh.</b><br>
+  Prove your coding agent starts without your personal setup, then measure it.</p>
+</div>
+
+<div align="center">
+
+[![License: MIT][license-shield]][license-url]
+[![PyPI][version-shield]][version-url]
+[![Python 3.11+][python-shield]][python-url]
+[![Linux | WSL | macOS][platform-shield]][platform-url]
+[![Claude Code | Codex | Pi | Antigravity][agent-shield]][agent-url]
+
+</div>
+
+<div align="center">
+  <a href="#quick-start">Quick start</a> &middot;
+  <a href="#features">Features</a> &middot;
+  <a href="https://github.com/a-ratz/tare/blob/main/docs/README.md">Documentation</a> &middot;
+  <a href="https://github.com/a-ratz/tare/blob/main/CONCEPT.md">Concept</a> &middot;
+  <a href="https://github.com/a-ratz/tare/issues/new?template=bug_report.md">Report Bug</a>
+</div>
+
+<br>
+
+---
+
+Your coding agent knows you. Every run loads your global instructions, memories, skills, MCP
+servers and connected accounts. At work that is the point. In an evaluation it is noise. Compare
+runs with and without a skill on your own machine, and both runs carry your setup. You measure
+your machine, not the skill.
+
+**tare** starts the agent in a room that holds nothing of yours and proves the room is clean
+before every run. It works with Claude Code, Codex, Pi and Google's Antigravity CLI on Linux and
+WSL, and with the first three on macOS.
+
+```text
+$ tare probe claude
+tare probe · claude 2.1.289 · /home/me/tare
+  control   your real setup shows: instructions, home path, mcp, skills, plugins, email, reach
+  leaks     none
+  declared
+    email        account email                                subscription login
+    credentials  /home/tare/.claude-config/.credentials.json  copy of the login, needed by the CLI
+tare: 0.00
+```
+
+`tare: 0.00` is the zeroed scale. Nothing of your setup reached the model except what the reading
+declares. A reading with leaks names each one and the file it came from, and the agent does not
+start.
+
+## Quick start
+
+```bash
+uv tool install tare-cli            # or: pipx install tare-cli
+tare probe claude                   # prove the room is clean
+tare claude -- -p "fix the tests"   # run Claude Code in it
+```
+
+On Linux and WSL, tare also needs bubblewrap (`sudo apt install bubblewrap`). The
+[install guide](https://github.com/a-ratz/tare/blob/main/docs/install.md) has the details per
+platform, and [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md) explains
+every line of the output above.
+
+To compare agents, skills or models, `tare calibrate`, `tare swap` and `tare cliff` start many runs
+and open a live dashboard that leads with the answer:
+
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/docs/assets/dashboard-calibrate-dark.png">
+  <img alt="tare's dashboard for a Calibrate run: Claude Code scored higher, 68 against 15 out of 100, with the scores of every run" src="https://raw.githubusercontent.com/a-ratz/tare/main/docs/assets/dashboard-calibrate-light.png">
+</picture>
+
+## Features
+
+- **Measures what the model receives.** The probe records the request the agent sends to a fake
+  model server on your machine. It makes no paid calls and never asks the agent about itself,
+  because an agent asked about its own context can refuse or be wrong.
+  [How it works](https://github.com/a-ratz/tare/blob/main/docs/how-it-works.md)
+- **Knows when it cannot see.** A control run starts the agent in your real setup. If the probe could
+  not see your setup there, the reading is `tare: not proven (blind)` instead of a `0.00` that
+  proves nothing. [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md)
+- **Keeps your files out and your login working.** The room has no home directory of yours, no
+  credentials of your other tools and no secret environment variables. It gets a fresh copy of
+  your login, which tare deletes after the run.
+  [Running agents in rooms](https://github.com/a-ratz/tare/blob/main/docs/running-agents.md)
+- **Compares skills, plugins and models.** `tare calibrate` starts each side several times from
+  scratch and reports pass rates with their intervals.
+  [Calibrate](https://github.com/a-ratz/tare/blob/main/docs/calibrate.md)
+- **Scores what no test can decide.** `tare judge` lets a second agent score a page by your
+  rubric, without agent or model names, and `tare judge-noise` shows how far its scores vary.
+  [Judge](https://github.com/a-ratz/tare/blob/main/docs/judge.md)
+- **Tells the workspace from the agent.** `tare swap` lets two agents continue each other's
+  unfinished work and reports whether the workspace or the agent decided the outcome.
+  [Swap](https://github.com/a-ratz/tare/blob/main/docs/swap.md)
+- **Finds the step where a failed run was lost** (experimental). `tare cliff` restarts the run
+  from its saved steps and names the step after which it no longer passes.
+  [Cliff](https://github.com/a-ratz/tare/blob/main/docs/cliff.md)
+- **Live dashboard, cost and reruns.** The dashboard leads with each run's answer and follows it
+  live, each report ends with tokens and cost, and `tare rerun` repeats a run from its recipe.
+  [Dashboard and rerun](https://github.com/a-ratz/tare/blob/main/docs/dashboard.md) ·
+  [Tokens and cost](https://github.com/a-ratz/tare/blob/main/docs/tokens-and-cost.md)
+
+## When to use
+
+| Use tare when | Look elsewhere when |
+|---|---|
+| you compare agent runs with and without a skill, plugin or prompt | you need protection from a hostile agent, because the room has network access and a working login |
+| you benchmark agents on your own machine | you use the Antigravity CLI on macOS (not supported there yet) |
+| you want a fresh-machine run without a fresh machine | you log in with an API key only (untested) |
+| an agent failed a task and you want to know which step lost it | you want a single run explained without running the agent again, because Cliff runs it many times |
+
+## Documentation
+
+| | |
+|---|---|
+| **Getting started** | [Install](https://github.com/a-ratz/tare/blob/main/docs/install.md) · [The words tare uses](https://github.com/a-ratz/tare/blob/main/docs/words.md) · [Running agents in rooms](https://github.com/a-ratz/tare/blob/main/docs/running-agents.md) · [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md) |
+| **Measuring** | [Overview](https://github.com/a-ratz/tare/blob/main/docs/measuring.md) · [Calibrate](https://github.com/a-ratz/tare/blob/main/docs/calibrate.md) · [Judge](https://github.com/a-ratz/tare/blob/main/docs/judge.md) · [Swap](https://github.com/a-ratz/tare/blob/main/docs/swap.md) · [Cliff](https://github.com/a-ratz/tare/blob/main/docs/cliff.md) · [Tokens and cost](https://github.com/a-ratz/tare/blob/main/docs/tokens-and-cost.md) · [Dashboard and rerun](https://github.com/a-ratz/tare/blob/main/docs/dashboard.md) |
+| **Background** | [How it works](https://github.com/a-ratz/tare/blob/main/docs/how-it-works.md) · [Concept: the design and every measurement](https://github.com/a-ratz/tare/blob/main/CONCEPT.md) · [Experiments](https://github.com/a-ratz/tare/tree/main/experiments) |
+
+## Status
+
+tare is on PyPI as `tare-cli`. Claude Code, Codex and Pi run on Linux, WSL and macOS; the
+Antigravity CLI runs on Linux and WSL, and stops with a message on macOS for now. What comes next
+is in the [open epics](https://github.com/a-ratz/tare/issues?q=is%3Aissue+is%3Aopen+label%3Atype%3Aepic).
+
+## Contributing
+
+Every change starts as an issue. Larger work is an epic, split into features, and one epic
+becomes one branch and one pull request. [AGENTS.md](https://github.com/a-ratz/tare/blob/main/AGENTS.md)
+describes the process, the checks and the conventions. Run the tests with `uv run pytest`.
+
+## License
+
+MIT. See [LICENSE](https://github.com/a-ratz/tare/blob/main/LICENSE).
+
+---
+
+Crafted with [Readme Craft](https://github.com/motiful/readme-craft)
+
+[license-shield]: https://img.shields.io/badge/license-MIT-blue.svg
+[license-url]: https://github.com/a-ratz/tare/blob/main/LICENSE
+[version-shield]: https://img.shields.io/pypi/v/tare-cli.svg
+[version-url]: https://pypi.org/project/tare-cli/
+[python-shield]: https://img.shields.io/badge/python-3.11%2B-3776AB.svg
+[python-url]: https://www.python.org/
+[platform-shield]: https://img.shields.io/badge/platform-Linux%20%7C%20WSL%20%7C%20macOS-555555.svg
+[platform-url]: https://github.com/a-ratz/tare/blob/main/docs/install.md
+[agent-shield]: https://img.shields.io/badge/agents-Claude%20Code%20%7C%20Codex%20%7C%20Pi%20%7C%20Antigravity-D97757.svg
+[agent-url]: https://github.com/a-ratz/tare/blob/main/docs/install.md

{tare_cli-0.3.0 → tare_cli-0.4.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "tare-cli"
-version = "0.3.0"
+version = "0.4.0"
 description = "Prove your coding agent starts without your personal setup before you measure it."
 readme = "README.md"
 requires-python = ">=3.11"

{tare_cli-0.3.0 → tare_cli-0.4.0}/src/tare/agents.py

@@ -526,7 +526,7 @@ class Pi:
 
     @contextmanager
     def twin(self, real: Real, fake_url: str):
-        """The dirty twin runs on a copy of the user's agent directory; the real one is never written."""
+        """The control run runs on a copy of the user's agent directory; the real one is never written."""
         with tempfile.TemporaryDirectory(prefix="tare-pi-twin-") as copy:
             agent_dir = Path(copy) / "agent"
             shutil.copytree(real.config, agent_dir, symlinks=True, ignore=shutil.ignore_patterns("sessions"))
@@ -641,7 +641,7 @@ class Antigravity:
 
     def discover(self) -> Real:
         if sys.platform == "darwin":
-            raise TareError("the Antigravity CLI is not supported on macOS yet: its dirty twin needs Linux's overlay "
+            raise TareError("the Antigravity CLI is not supported on macOS yet: its control run needs Linux's overlay "
                             "over ~/.gemini (epic #85)")
         home = Path.home()
         return Real(home, home / ".gemini" / "antigravity-cli", _which("agy"))
@@ -679,7 +679,7 @@ class Antigravity:
 
     @contextmanager
     def twin(self, real: Real, fake_url: str):
-        """The dirty twin sees the real setup, but agy writes into ~/.gemini on every run (its
+        """The control run sees the real setup, but agy writes into ~/.gemini on every run (its
         conversations, caches, onboarding state). So it runs with an overlay over ~/.gemini
         whose writes go to a tmpfs that dies with the run; see twin_argv."""
         mount = tempfile.mkdtemp(prefix="tare-agy-twin-")