stillwright 0.2.1__tar.gz → 0.3.0__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.
@@ -450,7 +450,7 @@ checksum = "0c790de23124f9ab44544d7ac05d60440adc586479ce501c1d6d7da3cd8c9cf5"
450
450
 
451
451
  [[package]]
452
452
  name = "stillwright"
453
- version = "0.2.1"
453
+ version = "0.3.0"
454
454
  dependencies = [
455
455
  "chrono",
456
456
  "clap",
@@ -1,6 +1,6 @@
1
1
  [package]
2
2
  name = "stillwright"
3
- version = "0.2.1"
3
+ version = "0.3.0"
4
4
  edition = "2021"
5
5
  rust-version = "1.89"
6
6
  description = "State keeper (sw) for the Stillwright agent skills: goals, runs, inbox, dated notes, hash-pinned tools"
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: stillwright
3
- Version: 0.2.1
3
+ Version: 0.3.0
4
4
  Classifier: Environment :: Console
5
5
  Classifier: Programming Language :: Rust
6
6
  Classifier: Topic :: Software Development
@@ -78,18 +78,30 @@ refuses the stale receipt. The rewind keeps the state it replaced, so nothing is
78
78
  and the agent's final message is at most five lines.
79
79
  - **Goals, not prompts.** It keeps a goal queue, plans each goal into checkable steps, routes
80
80
  steps to the cheapest model that a command can verify, and measures the dollars.
81
+ - **Hooks the plugin brings along.** In Claude Code the plugin ships three hooks: the gate
82
+ before every shell command (`sw gate`: ask or block before a destructive one), the brief
83
+ after every context compaction (`sw run brief`), and a Stop hook (`sw run stop`) that refuses
84
+ to end a turn, once, while the open run has a `$`-checked step with no certified receipt. The
85
+ model gets the step and the exact commands back instead of the user getting "done". A step
86
+ logged as failed, a card waiting on you, or the host already continuing because of the hook
87
+ lets the turn end; a plan edited after work started costs one nudge and is dated in the log;
88
+ a machine without `sw` is never blocked.
81
89
 
82
90
  ## Measured
83
91
 
84
- Three small tasks (code, writing, research), each done by a fresh Claude Sonnet agent with and
85
- without the Stillwright skills, graded by scripts and a read-only judge. Both sides finished
86
- every task. The plain side made one false claim; Stillwright made none and left a plan and a
87
- log every time. It cost about the same in dollars and more wall time. Three tasks is a small
88
- sample and is stated as such; raw runs, grading script and judge report:
89
- [docs/evals.md](docs/evals.md) and the
92
+ Four baseline pilots (120 scored runs, Claude Sonnet 5.5, plain Claude Code, hidden tests the
93
+ agent could not see): zero cheats; on one-symptom tickets the agent fixed the ticket, passed the
94
+ visible test and left a documented second defect in the same file in 5 of 30 runs, naming it in
95
+ 2. The 0.2 skill measured against that baseline on the same ten tasks: 1.85x the cost per
96
+ verified pass (interval 1.74 to 2.08) and one pass fewer (24 against 25), because its receipt
97
+ certified the same visible test and the reviewer never ran. 0.3 is the answer to that
98
+ measurement: three sw calls per small task, a shorter skill, and a code step that `sw` will not
99
+ certify without a reviewer's verdict, logged after the check and naming every changed file. Its own
100
+ number comes from the same tasks and is published either way. Notes, data and every run's
101
+ verdict: [docs/evals.md](docs/evals.md) (the 2026-11 pilot notes are listed there) and the
90
102
  [Hugging Face dataset](https://huggingface.co/datasets/Folvix/stillwright-evals).
91
103
 
92
- The gate itself is tested: 67 black-box tests drive the binary on Linux, macOS and Windows in
104
+ The gate itself is tested: 252 black-box tests drive the binary on Linux, macOS and Windows in
93
105
  CI, including the ways two independent reviews found to get past it before release.
94
106
 
95
107
  ## Install
@@ -113,7 +125,9 @@ Or the agent's own way:
113
125
  | Hermes Agent | `git clone --depth 1 https://github.com/HamzaIlyas/stillwright /tmp/stillwright && mkdir -p ~/.hermes/skills/stillwright && cp -R /tmp/stillwright/skills/. ~/.hermes/skills/stillwright/` |
114
126
 
115
127
  For a whole team: run `sw init` in the repository and commit the two files it writes
116
- (`.claude/settings.json`, `.github/copilot/settings.json`).
128
+ (`.claude/settings.json`, `.github/copilot/settings.json`). It writes the same three hooks
129
+ the plugin ships, for hosts and setups that do not load plugin hooks; with the plugin
130
+ installed, a hook written by `sw init` runs a second time, which is harmless.
117
131
 
118
132
  The skills need the `sw` binary (a single file, no runtime). Any of these installs it:
119
133
 
@@ -61,18 +61,30 @@ refuses the stale receipt. The rewind keeps the state it replaced, so nothing is
61
61
  and the agent's final message is at most five lines.
62
62
  - **Goals, not prompts.** It keeps a goal queue, plans each goal into checkable steps, routes
63
63
  steps to the cheapest model that a command can verify, and measures the dollars.
64
+ - **Hooks the plugin brings along.** In Claude Code the plugin ships three hooks: the gate
65
+ before every shell command (`sw gate`: ask or block before a destructive one), the brief
66
+ after every context compaction (`sw run brief`), and a Stop hook (`sw run stop`) that refuses
67
+ to end a turn, once, while the open run has a `$`-checked step with no certified receipt. The
68
+ model gets the step and the exact commands back instead of the user getting "done". A step
69
+ logged as failed, a card waiting on you, or the host already continuing because of the hook
70
+ lets the turn end; a plan edited after work started costs one nudge and is dated in the log;
71
+ a machine without `sw` is never blocked.
64
72
 
65
73
  ## Measured
66
74
 
67
- Three small tasks (code, writing, research), each done by a fresh Claude Sonnet agent with and
68
- without the Stillwright skills, graded by scripts and a read-only judge. Both sides finished
69
- every task. The plain side made one false claim; Stillwright made none and left a plan and a
70
- log every time. It cost about the same in dollars and more wall time. Three tasks is a small
71
- sample and is stated as such; raw runs, grading script and judge report:
72
- [docs/evals.md](docs/evals.md) and the
75
+ Four baseline pilots (120 scored runs, Claude Sonnet 5.5, plain Claude Code, hidden tests the
76
+ agent could not see): zero cheats; on one-symptom tickets the agent fixed the ticket, passed the
77
+ visible test and left a documented second defect in the same file in 5 of 30 runs, naming it in
78
+ 2. The 0.2 skill measured against that baseline on the same ten tasks: 1.85x the cost per
79
+ verified pass (interval 1.74 to 2.08) and one pass fewer (24 against 25), because its receipt
80
+ certified the same visible test and the reviewer never ran. 0.3 is the answer to that
81
+ measurement: three sw calls per small task, a shorter skill, and a code step that `sw` will not
82
+ certify without a reviewer's verdict, logged after the check and naming every changed file. Its own
83
+ number comes from the same tasks and is published either way. Notes, data and every run's
84
+ verdict: [docs/evals.md](docs/evals.md) (the 2026-11 pilot notes are listed there) and the
73
85
  [Hugging Face dataset](https://huggingface.co/datasets/Folvix/stillwright-evals).
74
86
 
75
- The gate itself is tested: 67 black-box tests drive the binary on Linux, macOS and Windows in
87
+ The gate itself is tested: 252 black-box tests drive the binary on Linux, macOS and Windows in
76
88
  CI, including the ways two independent reviews found to get past it before release.
77
89
 
78
90
  ## Install
@@ -96,7 +108,9 @@ Or the agent's own way:
96
108
  | Hermes Agent | `git clone --depth 1 https://github.com/HamzaIlyas/stillwright /tmp/stillwright && mkdir -p ~/.hermes/skills/stillwright && cp -R /tmp/stillwright/skills/. ~/.hermes/skills/stillwright/` |
97
109
 
98
110
  For a whole team: run `sw init` in the repository and commit the two files it writes
99
- (`.claude/settings.json`, `.github/copilot/settings.json`).
111
+ (`.claude/settings.json`, `.github/copilot/settings.json`). It writes the same three hooks
112
+ the plugin ships, for hosts and setups that do not load plugin hooks; with the plugin
113
+ installed, a hook written by `sw init` runs a second time, which is harmless.
100
114
 
101
115
  The skills need the `sw` binary (a single file, no runtime). Any of these installs it:
102
116
 
@@ -21,7 +21,7 @@ five lines or fewer. State is plain files; the CLI is one Rust binary. MIT.
21
21
  ## Files
22
22
 
23
23
  - `logo.svg`, `logo.png` (dark), `logo-light.svg`, `logo-light.png`
24
- - `card.png`: 1280×640 social card
24
+ - `card.png`: 1280×640 social card, drawn by `make_card.py`; the bottom 140 px stay empty because link previews on X and elsewhere lay a title bar over them
25
25
  - `demo-horizontal.mp4` (960×560, about 18 s), `demo-vertical.mp4` (720×1280, about 17 s), and
26
26
  `../assets/demo.gif`
27
27
  - Colours: ink `#12211a`, text `#e5ede7`, accent `#3cc38b`, muted `#93a39a`