tare-cli 0.3.1__tar.gz → 0.4.1__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (52) hide show
  1. {tare_cli-0.3.1 → tare_cli-0.4.1}/CONCEPT.md +62 -48
  2. {tare_cli-0.3.1 → tare_cli-0.4.1}/PKG-INFO +13 -5
  3. {tare_cli-0.3.1 → tare_cli-0.4.1}/README.md +12 -4
  4. {tare_cli-0.3.1 → tare_cli-0.4.1}/pyproject.toml +1 -1
  5. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/agents.py +21 -4
  6. tare_cli-0.4.1/src/tare/answer.py +412 -0
  7. tare_cli-0.4.1/src/tare/assets/fonts/IBMPlexMono-Regular.woff2 +0 -0
  8. tare_cli-0.4.1/src/tare/assets/fonts/IBMPlexMono-SemiBold.woff2 +0 -0
  9. tare_cli-0.4.1/src/tare/assets/fonts/IBMPlexSans-Regular.woff2 +0 -0
  10. tare_cli-0.4.1/src/tare/assets/fonts/IBMPlexSans-SemiBold.woff2 +0 -0
  11. tare_cli-0.4.1/src/tare/assets/fonts/OFL.txt +93 -0
  12. tare_cli-0.4.1/src/tare/assets/logo-dark.svg +260 -0
  13. tare_cli-0.4.1/src/tare/assets/logo-light.svg +260 -0
  14. tare_cli-0.4.1/src/tare/assets/tare.css +228 -0
  15. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/calibrate.py +1 -1
  16. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/cli.py +23 -14
  17. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/cliff.py +11 -11
  18. tare_cli-0.4.1/src/tare/dashboard.html +359 -0
  19. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/dashboard.py +16 -2
  20. tare_cli-0.4.1/src/tare/overview.html +65 -0
  21. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/probe.py +11 -11
  22. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/room.py +14 -0
  23. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/swap.py +21 -21
  24. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/trail.py +6 -1
  25. tare_cli-0.4.1/tests/test_agents.py +55 -0
  26. tare_cli-0.4.1/tests/test_answer.py +117 -0
  27. tare_cli-0.4.1/tests/test_cli.py +56 -0
  28. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_cliff.py +7 -7
  29. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_dashboard.py +30 -3
  30. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_probe.py +2 -2
  31. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_room.py +29 -0
  32. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_swap.py +35 -7
  33. tare_cli-0.3.1/src/tare/dashboard.html +0 -269
  34. tare_cli-0.3.1/src/tare/overview.html +0 -78
  35. tare_cli-0.3.1/tests/test_agents.py +0 -9
  36. {tare_cli-0.3.1 → tare_cli-0.4.1}/.gitignore +0 -0
  37. {tare_cli-0.3.1 → tare_cli-0.4.1}/LICENSE +0 -0
  38. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/__init__.py +0 -0
  39. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/capsule.py +0 -0
  40. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/fake.py +0 -0
  41. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/journal.py +0 -0
  42. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/judge.py +0 -0
  43. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/recipe.py +0 -0
  44. {tare_cli-0.3.1 → tare_cli-0.4.1}/src/tare/usage.py +0 -0
  45. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/conftest.py +0 -0
  46. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_agy.py +0 -0
  47. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_calibrate.py +0 -0
  48. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_fake.py +0 -0
  49. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_judge.py +0 -0
  50. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_pi.py +0 -0
  51. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_seatbelt.py +0 -0
  52. {tare_cli-0.3.1 → tare_cli-0.4.1}/tests/test_usage.py +0 -0
@@ -7,8 +7,10 @@ until the reading is `tare: 0.00`, unless you pass `--allow-dirty`. It supports
7
7
  and the Antigravity CLI (`agy`) on Linux and WSL, and Claude Code, Codex and Pi on macOS.
8
8
 
9
9
  This document describes the design and the measurements behind it. The documentation explains the
10
- words this document uses: room, probe, reading, leak, declared and dirty twin in
11
- [docs/words.md](docs/words.md), check, tail and handoff in [docs/measuring.md](docs/measuring.md).
10
+ words this document uses: room, probe, reading, leak, declared and control run in
11
+ [docs/words.md](docs/words.md), check, restart and handover note in [docs/measuring.md](docs/measuring.md).
12
+ The experiment records under `experiments/` call the control run the control run and a restart a
13
+ tail.
12
14
  Every claim marked **measured** was observed on WSL2. The section [Measurements](#measurements)
13
15
  lists the agent versions and links each result.
14
16
 
@@ -72,11 +74,11 @@ your instructions, memories, plugins, hooks and skills. That is not enough.
72
74
  server keeps the request, which is the context the agent CLI put together, and answers "ok".
73
75
  The idea is the blank run from analytical chemistry, which runs the whole procedure without a
74
76
  sample to see what the procedure itself brings in.
75
- - **Control:** the same agent run in your real setup, the dirty twin, must show your context. The
77
+ - **Control:** the same agent run in your real setup, the control run, must show your context. The
76
78
  risk was that features that come with the login (connected accounts, account skills) would
77
79
  not load against a fake server, so the probe would read clean while a real run is not.
78
80
  **Measured:** for Claude Code this risk did not occur. Against the fake model server, the
79
- dirty twin still loaded the connected accounts, account skills, plugin skills, hook output,
81
+ control run still loaded the connected accounts, account skills, plugin skills, hook output,
80
82
  global instructions, memories and the email. The room loaded none of them except the email.
81
83
  Pointing Claude Code at a custom base URL turns its tool search off and puts every tool schema
82
84
  into the request. So the probe and the runs set `ENABLE_TOOL_SEARCH=true` to see the context as
@@ -96,7 +98,10 @@ A bubblewrap room is a sandbox for one agent run: an empty `/home` with the room
96
98
  directory at `/home/tare`, a fresh copy of the login, the agent's executable, the project at
97
99
  `/work`, and the environment allowlist. `tare claude` and the other agent commands mount your
98
100
  project directly, so the agent changes your real files. Calibrate, Swap and Cliff give every run
99
- its own copy of the project. The room has network access and a working login. It keeps your setup out of the measurement. It does not protect your machine from a hostile
101
+ its own copy of the project. The room has network access and a working login. Where `/etc/resolv.conf` is a link into `/run`
102
+ (systemd-resolved, NetworkManager), the room mounts the one file the link names; without it no host
103
+ name would resolve in the room, and the probe would not notice, because it only talks to the fake
104
+ model server by address. It keeps your setup out of the measurement. It does not protect your machine from a hostile
100
105
  agent.
101
106
 
102
107
  A Seatbelt room (macOS) cannot mount anything. It is one directory under `/private/tmp` that
@@ -134,19 +139,19 @@ the copy does not start inside the renewal window (#126).
134
139
  ### Probe and reading
135
140
 
136
141
  The probe has the three parts of layer 3: the agent in the room against the fake model server,
137
- the dirty twin, and the reach script. It sorts what it finds into **kinds of context**,
142
+ the control run, and the reach script. It sorts what it finds into **kinds of context**,
138
143
  each checked on its own: instructions, memories, home path, MCP servers (connected accounts count
139
144
  here), skills, plugins, extensions, instruction files above the project, the account email, and
140
145
  reach.
141
146
 
142
147
  - A kind of context counts only when you have it. tare learns that from your files, for example
143
148
  a global instructions file or a skill folder.
144
- - If you have a kind of context and the dirty twin does not show it, the probe could not see
145
- it. tare then runs the dirty twin once more after 10 seconds. If the second dirty twin still
149
+ - If you have a kind of context and the control run does not show it, the probe could not see
150
+ it. tare then runs the control run once more after 10 seconds. If the second control run still
146
151
  does not show that kind of context, the reading is `tare: not proven (blind)`. **Measured:**
147
- with six runs starting together, a dirty twin sent its request before its MCP servers were
152
+ with six runs starting together, a control run sent its request before its MCP servers were
148
153
  connected.
149
- - Every probe makes no paid model call, because both the room and the dirty twin talk to the
154
+ - Every probe makes no paid model call, because both the room and the control run talk to the
150
155
  fake model server.
151
156
 
152
157
  ### Agents
@@ -154,7 +159,11 @@ reach.
154
159
  Each agent needs its own way to the fake model server, its own login copy and its own markers:
155
160
  lines of your own files that the probe looks for in a request.
156
161
 
157
- | Agent | How the room is set up | Way to the fake model server | Kinds of context the dirty twin showed on the measuring machine |
162
+ tare mounts the agent's own program into the room. Where the name on `PATH` is a mise shim, a link
163
+ to mise that picks the tool by the name it was called as, tare asks mise for the program behind it
164
+ (`mise which`). Mounted in its place, mise would find no installed tool in the room.
165
+
166
+ | Agent | How the room is set up | Way to the fake model server | Kinds of context the control run showed on the measuring machine |
158
167
  |---|---|---|---|
159
168
  | 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 |
160
169
  | 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 |
@@ -181,15 +190,15 @@ What each agent needed:
181
190
  Codex's default model. Pass `-m` when you compare runs.
182
191
  - **Pi:** a fresh agent directory is clean on its own and keeps the project's `AGENTS.md`, so the
183
192
  room uses no flags. `--no-context-files` would drop the project's `AGENTS.md`. tare finds Pi's
184
- extension tools by comparing the dirty twin's tools with a run that uses `--no-extensions`.
193
+ extension tools by comparing the control run's tools with a run that uses `--no-extensions`.
185
194
  With standard input left open, Pi waits for piped input and never answers, so tare closes it.
186
195
  - **Antigravity CLI:** the fake model server speaks Google's Cloud Code `v1internal` API and
187
196
  offers one model. It lists that model in the groups the CLI shows models in
188
197
  (`agentModelSorts`), because the CLI rejects a model that no group lists. The
189
- CLI writes into `~/.gemini` on every run, so its dirty twin runs with an overlay over
198
+ CLI writes into `~/.gemini` on every run, so its control run runs with an overlay over
190
199
  `~/.gemini`. A user namespace mounts the overlay, and every write goes to a temporary file
191
200
  system that disappears after the run. **Measured:** `~/.gemini` kept its fingerprint (paths,
192
- sizes, modification times) through every dirty twin. The CLI writes each skill's path into
201
+ sizes, modification times) through every control run. The CLI writes each skill's path into
193
202
  its skill list, so the probe compares skill lines without the path. Without that, a skill that
194
203
  leaked into the room would not match yours, because its path differs.
195
204
  - **Instruction files above the project:** Claude Code's one-shot mode, which the probe uses, did
@@ -199,7 +208,7 @@ What each agent needed:
199
208
  - **Long prompts:** Linux limits a single argument to about 128 KiB. tare passes the task to
200
209
  every agent as an argument, so a longer task fails. Pi could read it from a file with `@file`.
201
210
 
202
- ### Capsules, trail and handoff
211
+ ### Capsules, trail and handover note
203
212
 
204
213
  Calibrate, Swap and Cliff need a run they can stop and continue. During a recorded run, a hook
205
214
  archives the workspace after every tool call. Each archive is a **capsule**. The hook comes from
@@ -213,16 +222,16 @@ A run continues from a capsule in one of two ways:
213
222
  resumes a session file cut after the step, with `codex exec resume <id>`. Pi resumes with
214
223
  `--session`. **Measured:** a resumed Codex session kept its id and finished the remaining plan
215
224
  from step 1.
216
- - **A handoff.** The **trail** is a record of a run in the same format for every agent: the task,
217
- the messages, the tool calls and their results. tare writes a handoff from any prefix of it, so
225
+ - **A handover note.** The **trail** is a record of a run in the same format for every agent: the task,
226
+ the messages, the tool calls and their results. tare writes a handover note from any prefix of it, so
218
227
  any agent can continue any run without reading another agent's session format. The Antigravity
219
228
  CLI stores its sessions in a format that tare cannot cut, so it always continues with a
220
- handoff. Its trail comes from `transcript_full.jsonl`.
229
+ handover note. Its trail comes from `transcript_full.jsonl`.
221
230
 
222
231
  ### Calibrate
223
232
 
224
233
  `tare calibrate` starts each side (an agent with its arguments) several times from a fresh copy
225
- of the project and reports each side's pass rate with a 95% interval (Wilson). A side can have
234
+ of the project and reports each side's pass rate with a likely range (a 95% Wilson interval). A side can have
226
235
  its own prompt, so two skills that start with different commands compare in one call. Without a
227
236
  check, a run counts as finished when its agent exits with 0, and tare keeps its workspace for
228
237
  scoring elsewhere. **Measured:** in a field test that compared two ideation skills, 18 rooms all
@@ -243,31 +252,32 @@ judge found a real one-cent rounding bug and an overflow.
243
252
 
244
253
  ### Swap
245
254
 
246
- Swap cuts two runs at several points. At each cut, each agent continues each run's workspace
247
- by handoff. Swap reports a state effect (how much the state of the workspace matters), a model
248
- effect (which agent does better), and, where an agent can continue its own saved session, the
249
- handoff cost. At cut 0 both workspaces are the untouched project, so the state effect there must
250
- be zero. Swap checks this, and the check is called the null check. When the state effect and the
251
- model effect are within 0.1 of each other, Swap says they explain the failure about equally.
255
+ Swap stops two runs at several handover points. At each one, each agent continues each run's
256
+ workspace from a handover note. Swap reports a workspace effect (how much the state of the
257
+ workspace matters), an agent effect (which agent does better), and, where an agent can continue
258
+ its own saved session, the note cost. At the handover point 0 both workspaces are the untouched
259
+ project, so the workspace effect there must be zero. Swap checks this, and the check is called the
260
+ start check. When the workspace effect and the agent effect are within 0.1 of each other, Swap
261
+ says they explain the failure about equally.
252
262
 
253
- **Measured:** against scripted models with a known truth, the null check passed and the blame
254
- passed from the model to the workspace between cut 0 and cut 0.5, as the scripts were written
255
- to cause.
263
+ **Measured:** against scripted models with a known truth, the start check passed and the blame
264
+ passed from the agent to the workspace between the handovers at 0 and 0.5, as the scripts were
265
+ written to cause.
256
266
 
257
267
  ### Cliff
258
268
 
259
269
  Cliff restarts a failed run from its capsules and names the step after which it no longer
260
- succeeds. The baseline is the set of tails from the start. A step counts as good while its
261
- tails pass at least half as often as the baseline. Cliff halves the range between the last good
262
- and the first bad step, then adds tails to the two neighbours until their intervals separate or
263
- the budget is spent. It reports a model gap instead of a step only when the upper end of the
264
- baseline's 95% interval lies below 0.2.
270
+ succeeds. Fresh starts are the restarts from the beginning. A step counts as good while its
271
+ restarts pass at least half as often as a fresh start. Cliff halves the range between the last
272
+ good and the first bad step, then adds restarts to the two neighbours until their likely ranges
273
+ separate or the budget is spent. It reports that the task is too hard for the model instead of a
274
+ step only when the upper end of the fresh starts' likely range (a 95% interval) lies below 0.2.
265
275
 
266
276
  **Measured:** against a scripted model whose wrong step is known, Cliff found step 4 with
267
- separated intervals in 18 tails. Cliff stays experimental, because neither of two real test
277
+ separated likely ranges in 18 restarts. Cliff stays experimental, because neither of two real test
268
278
  rounds had a failure that sat in one step (see [Measurements](#measurements)). Swap already
269
- shows, at three cuts, how often an agent passes when it continues its own workspace. When a real
270
- Swap shows that rate dropping between two cuts, Cliff can narrow the drop down to one step. Only
279
+ shows, at three handover points, how often an agent passes when it continues its own workspace. When a real
280
+ Swap shows that rate dropping between two handover points, Cliff can narrow the drop down to one step. Only
271
281
  then does Cliff get new experiments.
272
282
 
273
283
  ### Tokens and cost
@@ -288,30 +298,34 @@ judge read 186.4k tokens in and $0.093 in total, the Codex part estimated from L
288
298
 
289
299
  Calibrate, Swap and Cliff write everything into a run directory: a journal (one JSON line per
290
300
  event), a `recipe.json`, a `usage.json` with the totals, and the live output of every agent. The dashboard reads only the run
291
- directory, so `tare watch` shows a finished run the same way as a live one. The recipe holds the
301
+ directory, so `tare watch` shows a finished run the same way as a live one. Each run page opens with
302
+ the answer: a sentence from fixed templates per verdict, worked out from the run directory the way
303
+ Calibrate, Swap and Cliff decide, so a run recorded by an older tare reads like a new one. The recipe holds the
292
304
  command, a fingerprint of the project and the agent versions. `tare rerun` repeats a run and
293
305
  names what has changed since.
294
306
 
295
307
  ## Measurements
296
308
 
297
309
  All measurements ran on WSL2 on 2026-10-04 and 2026-10-05. The rows that name macOS ran on
298
- macOS 27.0 (arm64) on 2026-10-05 and 2026-10-06.
310
+ macOS 27.0 (arm64) on 2026-10-05 and 2026-10-06. The row that names Omarchy ran on native Linux on
311
+ 2026-11-16.
299
312
 
300
313
  | What | Agent and version | Result | Source |
301
314
  |---|---|---|---|
302
- | 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) |
303
- | Dirty twin and room | Codex CLI 0.160.0 | the room read `tare: 0.00` | [results](experiments/dirty-twin-codex/RESULTS.md) |
304
- | Dirty twin and room | Pi 1.0.2 | the room read `tare: 0.00` without flags | [results](experiments/dirty-twin-pi/RESULTS.md) |
305
- | 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) |
315
+ | 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) |
316
+ | Control run and room | Codex CLI 0.160.0 | the room read `tare: 0.00` | [results](experiments/dirty-twin-codex/RESULTS.md) |
317
+ | Control run and room | Pi 1.0.2 | the room read `tare: 0.00` without flags | [results](experiments/dirty-twin-pi/RESULTS.md) |
318
+ | 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) |
306
319
  | 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 |
307
- | 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` |
320
+ | 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` |
308
321
  | 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) |
309
322
  | 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) |
323
+ | Native Linux with systemd-resolved: a room that resolves host names | Claude Code 2.1.325, bubblewrap 0.13.1, Omarchy 4.5.0 (Arch Linux) | before the fix the room read `tare: 0.00` and the agent's first request failed with a DNS error; with the resolver file mounted the probe still read `tare: 0.00` and the agent answered one prompt | [#152](https://github.com/a-ratz/tare/issues/152) |
310
324
  | 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 |
311
325
  | 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 |
312
326
  | 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 |
313
- | 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) |
314
- | 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 |
327
+ | 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) |
328
+ | 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 |
315
329
 
316
330
  ## Open questions and ideas
317
331
 
@@ -321,11 +335,11 @@ macOS 27.0 (arm64) on 2026-10-05 and 2026-10-06.
321
335
  a fresh room never did. Cached account data is the suspected cause. Not tested.
322
336
  - **Claude Code login renewal:** whether a renewal inside the room logs out the real session. Not
323
337
  tested, so tare avoids it.
324
- - **Codex handoffs:** Codex can run several commands in one tool call, which makes its handoffs
338
+ - **Codex handover notes:** Codex can run several commands in one tool call, which makes its handover notes
325
339
  long.
326
- - **Cost:** every tail is a full agent run, so a Cliff search or a Swap costs as much as its tails.
340
+ - **Cost:** every restart is a full agent run, so a Cliff search or a Swap costs as much as its restarts.
327
341
  - **macOS:** the room is built on Seatbelt, released in 0.3.0 (epic #85). Not done: the
328
- Antigravity CLI, whose dirty twin needs Linux's overlay, so tare stops with a message there.
342
+ Antigravity CLI, whose control run needs Linux's overlay, so tare stops with a message there.
329
343
  Not tested: whether a program started through launchd or XPC runs outside the room's profile.
330
344
  - **Other probes, not built:** weighing the first request's input tokens before and after adding
331
345
  junk to the real setup, and file tripwires that report which secret files a run opened.
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.5
2
2
  Name: tare-cli
3
- Version: 0.3.1
3
+ Version: 0.4.1
4
4
  Summary: Prove your coding agent starts without your personal setup before you measure it.
5
5
  Project-URL: Homepage, https://github.com/a-ratz/tare
6
6
  Project-URL: Repository, https://github.com/a-ratz/tare
@@ -68,7 +68,7 @@ WSL, and with the first three on macOS.
68
68
  ```text
69
69
  $ tare probe claude
70
70
  tare probe · claude 2.1.289 · /home/me/tare
71
- control dirty twin shows: instructions, home path, mcp, skills, plugins, email, reach
71
+ control your real setup shows: instructions, home path, mcp, skills, plugins, email, reach
72
72
  leaks none
73
73
  declared
74
74
  email account email subscription login
@@ -93,13 +93,21 @@ On Linux and WSL, tare also needs bubblewrap (`sudo apt install bubblewrap`). Th
93
93
  platform, and [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md) explains
94
94
  every line of the output above.
95
95
 
96
+ To compare agents, skills or models, `tare calibrate`, `tare swap` and `tare cliff` start many runs
97
+ and open a live dashboard that leads with the answer:
98
+
99
+ <picture>
100
+ <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/docs/assets/dashboard-calibrate-dark.png">
101
+ <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">
102
+ </picture>
103
+
96
104
  ## Features
97
105
 
98
106
  - **Measures what the model receives.** The probe records the request the agent sends to a fake
99
107
  model server on your machine. It makes no paid calls and never asks the agent about itself,
100
108
  because an agent asked about its own context can refuse or be wrong.
101
109
  [How it works](https://github.com/a-ratz/tare/blob/main/docs/how-it-works.md)
102
- - **Knows when it cannot see.** A dirty twin runs the agent in your real setup. If the probe could
110
+ - **Knows when it cannot see.** A control run starts the agent in your real setup. If the probe could
103
111
  not see your setup there, the reading is `tare: not proven (blind)` instead of a `0.00` that
104
112
  proves nothing. [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md)
105
113
  - **Keeps your files out and your login working.** The room has no home directory of yours, no
@@ -118,8 +126,8 @@ every line of the output above.
118
126
  - **Finds the step where a failed run was lost** (experimental). `tare cliff` restarts the run
119
127
  from its saved steps and names the step after which it no longer passes.
120
128
  [Cliff](https://github.com/a-ratz/tare/blob/main/docs/cliff.md)
121
- - **Live dashboard, cost and reruns.** A dashboard follows every run, each report ends with tokens
122
- and cost, and `tare rerun` repeats a run from its recipe.
129
+ - **Live dashboard, cost and reruns.** The dashboard leads with each run's answer and follows it
130
+ live, each report ends with tokens and cost, and `tare rerun` repeats a run from its recipe.
123
131
  [Dashboard and rerun](https://github.com/a-ratz/tare/blob/main/docs/dashboard.md) ·
124
132
  [Tokens and cost](https://github.com/a-ratz/tare/blob/main/docs/tokens-and-cost.md)
125
133
 
@@ -43,7 +43,7 @@ WSL, and with the first three on macOS.
43
43
  ```text
44
44
  $ tare probe claude
45
45
  tare probe · claude 2.1.289 · /home/me/tare
46
- control dirty twin shows: instructions, home path, mcp, skills, plugins, email, reach
46
+ control your real setup shows: instructions, home path, mcp, skills, plugins, email, reach
47
47
  leaks none
48
48
  declared
49
49
  email account email subscription login
@@ -68,13 +68,21 @@ On Linux and WSL, tare also needs bubblewrap (`sudo apt install bubblewrap`). Th
68
68
  platform, and [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md) explains
69
69
  every line of the output above.
70
70
 
71
+ To compare agents, skills or models, `tare calibrate`, `tare swap` and `tare cliff` start many runs
72
+ and open a live dashboard that leads with the answer:
73
+
74
+ <picture>
75
+ <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/a-ratz/tare/main/docs/assets/dashboard-calibrate-dark.png">
76
+ <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">
77
+ </picture>
78
+
71
79
  ## Features
72
80
 
73
81
  - **Measures what the model receives.** The probe records the request the agent sends to a fake
74
82
  model server on your machine. It makes no paid calls and never asks the agent about itself,
75
83
  because an agent asked about its own context can refuse or be wrong.
76
84
  [How it works](https://github.com/a-ratz/tare/blob/main/docs/how-it-works.md)
77
- - **Knows when it cannot see.** A dirty twin runs the agent in your real setup. If the probe could
85
+ - **Knows when it cannot see.** A control run starts the agent in your real setup. If the probe could
78
86
  not see your setup there, the reading is `tare: not proven (blind)` instead of a `0.00` that
79
87
  proves nothing. [Reading a probe](https://github.com/a-ratz/tare/blob/main/docs/probe.md)
80
88
  - **Keeps your files out and your login working.** The room has no home directory of yours, no
@@ -93,8 +101,8 @@ every line of the output above.
93
101
  - **Finds the step where a failed run was lost** (experimental). `tare cliff` restarts the run
94
102
  from its saved steps and names the step after which it no longer passes.
95
103
  [Cliff](https://github.com/a-ratz/tare/blob/main/docs/cliff.md)
96
- - **Live dashboard, cost and reruns.** A dashboard follows every run, each report ends with tokens
97
- and cost, and `tare rerun` repeats a run from its recipe.
104
+ - **Live dashboard, cost and reruns.** The dashboard leads with each run's answer and follows it
105
+ live, each report ends with tokens and cost, and `tare rerun` repeats a run from its recipe.
98
106
  [Dashboard and rerun](https://github.com/a-ratz/tare/blob/main/docs/dashboard.md) ·
99
107
  [Tokens and cost](https://github.com/a-ratz/tare/blob/main/docs/tokens-and-cost.md)
100
108
 
@@ -1,6 +1,6 @@
1
1
  [project]
2
2
  name = "tare-cli"
3
- version = "0.3.1"
3
+ version = "0.4.1"
4
4
  description = "Prove your coding agent starts without your personal setup before you measure it."
5
5
  readme = "README.md"
6
6
  requires-python = ">=3.11"
@@ -68,7 +68,24 @@ def _which(name: str) -> Path:
68
68
  found = shutil.which(name)
69
69
  if not found:
70
70
  raise TareError(f"{name} is not on PATH")
71
- return Path(found).resolve()
71
+ real = Path(found).resolve()
72
+ if real.name == "mise" and Path(found).name != "mise":
73
+ return _behind_mise(name, found, real)
74
+ return real
75
+
76
+
77
+ def _behind_mise(name: str, shim: str, mise: Path) -> Path:
78
+ """The program behind a mise shim.
79
+
80
+ A mise shim is a link to mise itself, which picks the tool by the name it was called as. Mounted into
81
+ the room in place of the agent, mise finds no installed tool there and tries to fetch one.
82
+ """
83
+ out = subprocess.run([str(mise), "which", name], capture_output=True, text=True)
84
+ tool = Path(out.stdout.strip()) if out.returncode == 0 and out.stdout.strip() else None
85
+ if tool is None or not tool.is_file():
86
+ raise TareError(f"{name} on PATH is a mise shim ({shim}), and `mise which {name}` did not name the program. "
87
+ f"Put the directory of the program itself first on PATH.")
88
+ return tool.resolve()
72
89
 
73
90
 
74
91
  def _absolute_links(original: Path, copy: Path):
@@ -526,7 +543,7 @@ class Pi:
526
543
 
527
544
  @contextmanager
528
545
  def twin(self, real: Real, fake_url: str):
529
- """The dirty twin runs on a copy of the user's agent directory; the real one is never written."""
546
+ """The control run runs on a copy of the user's agent directory; the real one is never written."""
530
547
  with tempfile.TemporaryDirectory(prefix="tare-pi-twin-") as copy:
531
548
  agent_dir = Path(copy) / "agent"
532
549
  shutil.copytree(real.config, agent_dir, symlinks=True, ignore=shutil.ignore_patterns("sessions"))
@@ -641,7 +658,7 @@ class Antigravity:
641
658
 
642
659
  def discover(self) -> Real:
643
660
  if sys.platform == "darwin":
644
- raise TareError("the Antigravity CLI is not supported on macOS yet: its dirty twin needs Linux's overlay "
661
+ raise TareError("the Antigravity CLI is not supported on macOS yet: its control run needs Linux's overlay "
645
662
  "over ~/.gemini (epic #85)")
646
663
  home = Path.home()
647
664
  return Real(home, home / ".gemini" / "antigravity-cli", _which("agy"))
@@ -679,7 +696,7 @@ class Antigravity:
679
696
 
680
697
  @contextmanager
681
698
  def twin(self, real: Real, fake_url: str):
682
- """The dirty twin sees the real setup, but agy writes into ~/.gemini on every run (its
699
+ """The control run sees the real setup, but agy writes into ~/.gemini on every run (its
683
700
  conversations, caches, onboarding state). So it runs with an overlay over ~/.gemini
684
701
  whose writes go to a tmpfs that dies with the run; see twin_argv."""
685
702
  mount = tempfile.mkdtemp(prefix="tare-agy-twin-")