Install every plugin settings.json enables, and prove the bench's private files stay off the remote

This commit is contained in:
2026-10-03 00:13:51 +02:00
parent a60cefd283
commit 835ffe422c
6 changed files with 40 additions and 11 deletions
+1
View File
@@ -2,6 +2,7 @@
## Unreleased
- README → Install adds the official plugin marketplace and installs all three plugins `settings.json` enables: `claude-security`, `code-review` and `rust-analyzer-lsp`, which needs `rust-analyzer` on `PATH`. Claude Code installs none of them by itself.
- `loop-claude` waits out a usage limit and retries: 2 minutes past the reset, or every 10 minutes when the reset time is unknown. A limit is not a failure, and restarts the count of three failures in a row. It prints what claude said about the limit and when the next try is due.
- Install the `claude-security` plugin (README → Install): larv-review's full security scan now runs once you confirm its cost; an unattended larv-review reports the scan as not run. `settings.json` silences the plugin's scan tip after `git push`.
- systems-architect and product-owner-reviewer can no longer edit what they review; stability-reviewer keeps its shell to run tests, and loses Claude Code's file-editing tools.
+10 -1
View File
@@ -20,7 +20,16 @@ Me, on my own machines, which take tagged releases. Nobody else depends on it.
## Install
Check out the latest tag in `~/.claude`, with Python 3.11 or newer as `python3` on `PATH`. Install the plugin behind larv-review's full security scan: `claude plugin install claude-security@claude-plugins-official`. Link `loop-claude` onto `PATH`: `ln -s ~/.claude/bin/loop-claude ~/.local/bin/loop-claude`. Run `loop-claude` only in a sandbox with no direct access to a production environment, such as a dedicated VM: every claude it starts, the interrupt included, skips every permission prompt.
Check out the latest tag in `~/.claude`, with Python 3.11 or newer as `python3` on `PATH`, and `rust-analyzer` too for Rust projects. Install every plugin `settings.json` enables, `claude-security` among them for larv-review's full security scan:
```
claude plugin marketplace add anthropics/claude-plugins-official
claude plugin install claude-security@claude-plugins-official
claude plugin install code-review@claude-plugins-official
claude plugin install rust-analyzer-lsp@claude-plugins-official
```
Link `loop-claude` onto `PATH`: `ln -s ~/.claude/bin/loop-claude ~/.local/bin/loop-claude`. Run `loop-claude` only in a sandbox with no direct access to a production environment, such as a dedicated VM: every claude it starts, the interrupt included, skips every permission prompt.
## Release
+4
View File
@@ -25,6 +25,10 @@ PRIVATE = [
"sessions/a.json",
"shell-snapshots/a.sh",
"tasks/a/b.json",
"workflow-bench/a-1-1.md",
"workflow-bench/a-1-1/README.md",
"workflow-bench/judge/a/1/fixes.diff",
"workflow-bench/logs/a.jsonl",
]
+15
View File
@@ -0,0 +1,15 @@
import json
import unittest
from test_python_floor import ROOT
class Install(unittest.TestCase):
def test_the_readme_installs_every_enabled_plugin_from_a_declared_marketplace(self):
settings = json.loads((ROOT / "settings.json").read_text(encoding="utf-8"))
readme = (ROOT / "README.md").read_text(encoding="utf-8")
for plugin, enabled in settings["enabledPlugins"].items():
if enabled:
self.assertIn(f"claude plugin install {plugin}\n", readme)
marketplace = settings["extraKnownMarketplaces"][plugin.split("@")[1]]["source"]["repo"]
self.assertIn(f"claude plugin marketplace add {marketplace}\n", readme)
+8
View File
@@ -25,6 +25,14 @@
"rust-analyzer-lsp@claude-plugins-official": true,
"code-review@claude-plugins-official": true
},
"extraKnownMarketplaces": {
"claude-plugins-official": {
"source": {
"source": "github",
"repo": "anthropics/claude-plugins-official"
}
}
},
"effortLevel": "xhigh",
"modelSettings": {
"claude-fable-5-1": {
+2 -10
View File
@@ -25,9 +25,7 @@
| ID | Release | Exempt | Item | R | S | A | G | Goals | Score |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.1.0 | question | **Say in README → Install how the transcript archive is backed up, once bender-infra backs up ai-sandbox's data volume.** | 3 | 5 | 8 | 9 | 2, 3 | 27.8 |
| 37 | 0.1.0 | | **Make README → Install's steps install every plugin `settings.json` enables, or confirm Claude Code installs an enabled plugin itself.** | 1 | 1 | 5 | 7 | 1 | 22.8 |
| 7 | 0.1.0 | decision | **Give the judge the project's recorded decisions, and ask it whether anything the project records settles a choice.** | 3 | 4 | 8 | 8 | 4.2 | 19.0 |
| 38 | 0.1.0 | | **Make `test_gitignore.py` prove the bench's logs, fixes and clones stay off the remote.** | 1 | 1 | 4 | 6 | 3 | 16.1 |
| 8 | 0.1.0 | | **Judge a cycle on more than the transcript's last 40,000 characters.** | 4 | 5 | 7 | 8 | 4.2 | 15.8 |
| 13 | 0.1.0 | decision | **Make larv-review §3 applicable on its own, cite its rules by name, and say where a pass's round count is recorded.** | 3 | 5 | 6 | 7 | 4.1 | 15.6 |
| 23 | 0.1.0 | | **Launch a principles auditor only for a principles file the project shows applies, and rerun only the reviewer whose finding a fix answers.** | 3 | 4 | 7 | 8 | 4.4 | 13.5 |
@@ -48,18 +46,10 @@ It lives under `~/.claude`, which `.gitignore` keeps from the remote, so a lost
2026-10-02, the human: ai-sandbox is the VM `loop-claude` will run on, hosted on the server bender. bender-infra, the infrastructure-as-code project that manages bender, will back up ai-sandbox's data volume once ai-sandbox is rebuilt. The item waits for the human to say the rebuild has landed. Before writing the README line, check that `~/.claude` lives on that volume and that the backup is private and encrypted at rest.
### 37. Make README → Install's steps install every plugin `settings.json` enables, or confirm Claude Code installs an enabled plugin itself.
`settings.json` enables `code-review@claude-plugins-official` and `rust-analyzer-lsp@claude-plugins-official`, and Install names only `claude-security`; `rust-analyzer-lsp` also needs `rust-analyzer` on `PATH`. If an enabled plugin is not installed on its own, a new machine set up from Install lacks both, and the changelog's claim that `settings.json` runs unedited on any machine with Install's prerequisites is false.
### 7. Give the judge the project's recorded decisions, and ask it whether anything the project records settles a choice.
Today it reads only the README's Goals and Audience (`premises.py`) and cannot see an answer recorded in `docs/decisions.md` or `AGENTS.md`, so it can interrupt the human again for a question already answered. Revise the decision "loop-claude: a judge decides each interrupt from the transcript and the README's premises; the worker's own account of which goal settles a choice decides nothing" to match.
### 38. Make `test_gitignore.py` prove the bench's logs, fixes and clones stay off the remote.
Its private-file list lacks `workflow-bench/logs/*.jsonl` (unattended runs' transcripts), `workflow-bench/judge/<case>/*/fixes.diff` (private projects' code) and the clones `workflow-bench/<case>-<version>-<n>/`. `.gitignore` keeps them out today, but a broader rule such as `!workflow-bench/*` would pass the test and publish them.
### 8. Judge a cycle on more than the transcript's last 40,000 characters.
`judge_prompt` cuts the transcript there. The session spans every cycle of the chunk, so a decision made early in a long run is never judged, and the human is not interrupted for it. Larger chunks make runs longer.
@@ -76,6 +66,8 @@ On this repo the frontend, REST and logging auditors each read the whole project
Claude Code rewrites `"model"` in this tracked file whenever a model is picked, so one `/model` dirties the checkout and overrides the synced choice.
Any write Claude Code makes to the file does it, `claude plugin install` included: on 2.1.288 it turned `"opus"` into `"opus[1m]"`. Each write also puts the keys back in Claude Code's own order, so a sorted file stays sorted only while nothing writes it.
### 36. Ask the user once before a full security scan, folding larv-review's ask into the plugin's cost prompt.
When the full scan fires, "Record, then ask" rule 3 asks whether to run it, and `claude-security` then asks the user to confirm its cost.