stillwright 0.1.4__tar.gz → 0.2.0__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.
- {stillwright-0.1.4 → stillwright-0.2.0}/Cargo.lock +1 -1
- {stillwright-0.1.4 → stillwright-0.2.0}/Cargo.toml +1 -1
- stillwright-0.2.0/PKG-INFO +201 -0
- stillwright-0.2.0/README.md +183 -0
- stillwright-0.2.0/src/gate.rs +1006 -0
- stillwright-0.2.0/src/main.rs +3355 -0
- stillwright-0.1.4/PKG-INFO +0 -155
- stillwright-0.1.4/README.md +0 -137
- stillwright-0.1.4/src/main.rs +0 -1120
- {stillwright-0.1.4 → stillwright-0.2.0}/LICENSE +0 -0
- {stillwright-0.1.4 → stillwright-0.2.0}/examples/quickstart-run/README.md +0 -0
- {stillwright-0.1.4 → stillwright-0.2.0}/press/README.md +0 -0
- {stillwright-0.1.4 → stillwright-0.2.0}/pyproject.toml +0 -0
|
@@ -0,0 +1,201 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: stillwright
|
|
3
|
+
Version: 0.2.0
|
|
4
|
+
Classifier: Environment :: Console
|
|
5
|
+
Classifier: Programming Language :: Rust
|
|
6
|
+
Classifier: Topic :: Software Development
|
|
7
|
+
License-File: LICENSE
|
|
8
|
+
Summary: sw, the state keeper for the Stillwright agent skills: goals, runs, inbox, dated notes, hash-pinned tools
|
|
9
|
+
Keywords: agent,orchestrator,ai,skills,cli
|
|
10
|
+
Home-Page: https://github.com/HamzaIlyas/stillwright
|
|
11
|
+
Author: Hamza Ilyas
|
|
12
|
+
License-Expression: MIT
|
|
13
|
+
Requires-Python: >=3.8
|
|
14
|
+
Description-Content-Type: text/markdown; charset=UTF-8; variant=GFM
|
|
15
|
+
Project-URL: Changelog, https://github.com/HamzaIlyas/stillwright/blob/main/CHANGELOG.md
|
|
16
|
+
Project-URL: Homepage, https://github.com/HamzaIlyas/stillwright
|
|
17
|
+
|
|
18
|
+
# Stillwright
|
|
19
|
+
|
|
20
|
+
[](https://github.com/HamzaIlyas/stillwright/actions/workflows/ci.yml)
|
|
21
|
+
[](https://github.com/HamzaIlyas/stillwright/releases/latest)
|
|
22
|
+
[](LICENSE)
|
|
23
|
+
|
|
24
|
+
**Your agent says "tests pass". Stillwright makes it prove it.** A step is done only when
|
|
25
|
+
its check ran and passed on the files as they are right now, with a receipt in the log, and
|
|
26
|
+
any step can be undone. For Claude Code, Codex, Gemini CLI, Cursor, OpenCode and any agent
|
|
27
|
+
that reads [Agent Skills](https://agentskills.io).
|
|
28
|
+
|
|
29
|
+

|
|
30
|
+
|
|
31
|
+
## See it work
|
|
32
|
+
|
|
33
|
+
A plan step with a command as its check. The agent cannot write "pass" by hand; it has to
|
|
34
|
+
run the check, and the receipt records what ran, the exit code and a hash of the working
|
|
35
|
+
tree:
|
|
36
|
+
|
|
37
|
+
```
|
|
38
|
+
$ sw run freeze latest tests
|
|
39
|
+
frozen: tests; any change fails the step's check until frozen again
|
|
40
|
+
|
|
41
|
+
$ sw run log latest step=1 result=pass
|
|
42
|
+
step 1 has a command check; a pass written by hand is not evidence.
|
|
43
|
+
Run `sw run check latest 1`, then `sw run done latest 1`.
|
|
44
|
+
|
|
45
|
+
$ sw run check latest 1
|
|
46
|
+
step 1: check failed (exit 1); output in ~/.stillwright/runs/20261029-031220-discount-floor/checks/1-2026-10-29031220.log
|
|
47
|
+
|
|
48
|
+
(the agent fixes pricing.py)
|
|
49
|
+
|
|
50
|
+
$ sw run check latest 1
|
|
51
|
+
step 1: check passed (exit 0, 0.1s, tree 64c2003852c2); `sw run done` when the step's work is complete
|
|
52
|
+
|
|
53
|
+
$ sw run done latest 1
|
|
54
|
+
step 1: done, certified (`python3 -m unittest -q tests.test_pricing` exit 0 on tree 64c2003852c2)
|
|
55
|
+
|
|
56
|
+
(something goes wrong later)
|
|
57
|
+
|
|
58
|
+
$ sw run rewind latest 1
|
|
59
|
+
step 1: files restored from snapshot 8bcbaf39015a (0 newer files removed);
|
|
60
|
+
the state before is kept at refs/sw/20261029-031220-discount-floor/before-rewind-2026-10-29031220
|
|
61
|
+
```
|
|
62
|
+
|
|
63
|
+
Real output from v0.2, unedited except for the home folder. If the agent edits a test after
|
|
64
|
+
`sw run freeze`, the check fails as tampered. If files change after the check, `sw run done`
|
|
65
|
+
refuses the stale receipt. The rewind keeps the state it replaced, so nothing is lost.
|
|
66
|
+
|
|
67
|
+
## What it does
|
|
68
|
+
|
|
69
|
+
- **Proof, not claims.** Every step of a plan has a check written before the work. A command
|
|
70
|
+
check is run by `sw`, with a receipt; anything else is judged by a fresh reviewer that never
|
|
71
|
+
saw the worker's reasoning.
|
|
72
|
+
- **Undo.** Every certified step leaves a git snapshot (untracked files included) under
|
|
73
|
+
`refs/sw/`; `sw run rewind` brings it back.
|
|
74
|
+
- **Vetted tools.** Thousands of skills and MCP servers, some malicious. Stillwright picks by
|
|
75
|
+
evidence, reads the code before trusting it, asks you once, and pins it by hash so it cannot
|
|
76
|
+
change silently.
|
|
77
|
+
- **Quiet.** One inbox with four card types (ask, review, notify, idea). Silence is the default,
|
|
78
|
+
and the agent's final message is at most five lines.
|
|
79
|
+
- **Goals, not prompts.** It keeps a goal queue, plans each goal into checkable steps, routes
|
|
80
|
+
steps to the cheapest model that a command can verify, and measures the dollars.
|
|
81
|
+
|
|
82
|
+
## Measured
|
|
83
|
+
|
|
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
|
|
90
|
+
[Hugging Face dataset](https://huggingface.co/datasets/Folvix/stillwright-evals).
|
|
91
|
+
|
|
92
|
+
The gate itself is tested: 67 black-box tests drive the binary on Linux, macOS and Windows in
|
|
93
|
+
CI, including the ways two independent reviews found to get past it before release.
|
|
94
|
+
|
|
95
|
+
## Install
|
|
96
|
+
|
|
97
|
+
One line for most agents:
|
|
98
|
+
|
|
99
|
+
```sh
|
|
100
|
+
npx plugins add HamzaIlyas/stillwright # as a plugin: Claude Code, Cursor, Codex, Copilot CLI, VS Code, Grok Build, Kimi Code
|
|
101
|
+
npx skills add HamzaIlyas/stillwright # as skills: 75+ agents, including all of the above
|
|
102
|
+
```
|
|
103
|
+
|
|
104
|
+
Or the agent's own way:
|
|
105
|
+
|
|
106
|
+
| Agent | Command |
|
|
107
|
+
|---|---|
|
|
108
|
+
| Claude Code | `claude plugin marketplace add HamzaIlyas/stillwright` then `claude plugin install stillwright@stillwright` |
|
|
109
|
+
| Codex | `codex plugin marketplace add HamzaIlyas/stillwright`, then install Stillwright from `/plugins` |
|
|
110
|
+
| Gemini CLI | `gemini extensions install https://github.com/HamzaIlyas/stillwright` |
|
|
111
|
+
| Cursor | loads the root `plugin.json` (Agent Plugins standard); marketplace listing pending |
|
|
112
|
+
| OpenCode, or any Agent Skills agent | `npx skills add HamzaIlyas/stillwright` |
|
|
113
|
+
| 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
|
+
|
|
115
|
+
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`).
|
|
117
|
+
|
|
118
|
+
The skills need the `sw` binary (a single file, no runtime). Any of these installs it:
|
|
119
|
+
|
|
120
|
+
```sh
|
|
121
|
+
pip install stillwright # any OS with Python
|
|
122
|
+
brew install HamzaIlyas/tap/stillwright # macOS / Linux with Homebrew
|
|
123
|
+
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/HamzaIlyas/stillwright/releases/latest/download/stillwright-installer.sh | sh # Linux / macOS
|
|
124
|
+
powershell -ExecutionPolicy Bypass -c "irm https://github.com/HamzaIlyas/stillwright/releases/latest/download/stillwright-installer.ps1 | iex" # Windows
|
|
125
|
+
cargo install --git https://github.com/HamzaIlyas/stillwright --locked # from source
|
|
126
|
+
```
|
|
127
|
+
|
|
128
|
+
## Use
|
|
129
|
+
|
|
130
|
+
New here? The [quickstart](docs/quickstart.md) takes ten minutes, and CI runs every command in it.
|
|
131
|
+
Also: [troubleshooting](docs/troubleshooting.md), [Windows notes](docs/windows.md),
|
|
132
|
+
[coming from OpenClaw or Hermes](docs/migrate.md), a [real example run](examples/quickstart-run/README.md),
|
|
133
|
+
and a [press kit](press/README.md).
|
|
134
|
+
|
|
135
|
+
Tell your agent a goal, or just say **"stillwright"**. The first time, it asks a few short
|
|
136
|
+
questions to get to know you. After that:
|
|
137
|
+
|
|
138
|
+
```
|
|
139
|
+
sw status what's going on, in a few lines
|
|
140
|
+
sw inbox list the only things that need you
|
|
141
|
+
sw goal add "..." give it more work
|
|
142
|
+
sw run status the current plan, step by step, with what is proven
|
|
143
|
+
sw run cost dollars a run spent, by model
|
|
144
|
+
```
|
|
145
|
+
|
|
146
|
+
To keep it working while you're away, schedule one loop per hour, for example
|
|
147
|
+
`claude -p "Use the stillwright skill: run one loop, then stop."` from cron. Unattended runs
|
|
148
|
+
never push, deploy, publish or install anything new; those wait in your inbox.
|
|
149
|
+
|
|
150
|
+
## What's inside
|
|
151
|
+
|
|
152
|
+
| Skill | Does |
|
|
153
|
+
|---|---|
|
|
154
|
+
| `stillwright` | the loop: pick goal → plan → get tools → execute → verify → report |
|
|
155
|
+
| `sw-onboard` | learns who you are and what you want, in three short rounds at most |
|
|
156
|
+
| `sw-scout` | finds, vets, trials and pins the best existing tool for a job |
|
|
157
|
+
| `sw-forge` | builds and tests a new skill when nothing good exists |
|
|
158
|
+
| `sw-verify` | independent, default-fail proof for every step |
|
|
159
|
+
| `sw-radar` | spots what matters to you (updates, risks, new tools) as rare idea cards |
|
|
160
|
+
| `sw-deep-search` | research that ends in verified, sourced answers |
|
|
161
|
+
| `sw-context-diet` | keeps long sessions lean and cheap |
|
|
162
|
+
| `sw-launch` | pitch, README, posts and launch plan for what you build |
|
|
163
|
+
|
|
164
|
+
All state is plain files in `~/.stillwright` (profile, goals, inbox, notes, pinned tools, run
|
|
165
|
+
logs, receipts). Nothing leaves your machine except what your agent already sends. `sw` is a
|
|
166
|
+
single Rust binary.
|
|
167
|
+
|
|
168
|
+
## Why these rules
|
|
169
|
+
|
|
170
|
+
Each one answers a failure that was measured or reported, not a hunch:
|
|
171
|
+
|
|
172
|
+
- Agents mark work done without testing it, and judges that read the agent's prose cannot
|
|
173
|
+
tell: Anthropic's [notes on long-running harnesses](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents);
|
|
174
|
+
[arXiv 2607.14890](https://arxiv.org/abs/2607.14890) (evidence-gated loops cut hidden
|
|
175
|
+
failures from 31 to 2 per 1,800). Hence receipts, and a reviewer that is a gate.
|
|
176
|
+
- Agents weaken tests to make them pass and pipe test output so failures exit 0
|
|
177
|
+
([anthropics/claude-code #319](https://github.com/anthropics/claude-code/issues/319)). Hence
|
|
178
|
+
frozen check files and `pipefail`.
|
|
179
|
+
- Undo of agent edits is the most-requested feature no host ships
|
|
180
|
+
([openai/codex #9203](https://github.com/openai/codex/issues/9203)). Hence snapshots and rewind.
|
|
181
|
+
- Agents rarely ask at the right moment and fill gaps with wrong assumptions:
|
|
182
|
+
[HiL-Bench](https://arxiv.org/abs/2604.09408). Hence the typed inbox, silence by default.
|
|
183
|
+
- Hundreds of malicious skills were found on a public registry in early 2026 (341 of 2,857):
|
|
184
|
+
[Snyk](https://snyk.io/articles/skill-md-shell-access/),
|
|
185
|
+
[Unit 42](https://unit42.paloaltonetworks.com/openclaw-ai-supply-chain-risk/). Hence vetting,
|
|
186
|
+
approval and hash-pinning before any third-party tool runs.
|
|
187
|
+
- Agent success falls as tasks get longer: [METR time horizons](https://metr.org/time-horizons/).
|
|
188
|
+
Hence small, checkable steps.
|
|
189
|
+
|
|
190
|
+
Cost is measured, not assumed: the harness, the price table and the cache linter are in
|
|
191
|
+
[docs/cost.md](docs/cost.md). Known limits are listed in the [CHANGELOG](CHANGELOG.md).
|
|
192
|
+
|
|
193
|
+
## Who runs it
|
|
194
|
+
|
|
195
|
+
One person, [Hamza Ilyas](https://github.com/HamzaIlyas), building it at
|
|
196
|
+
[Folvix](https://huggingface.co/Folvix). Questions and ideas:
|
|
197
|
+
[Discussions](https://github.com/HamzaIlyas/stillwright/discussions). Bugs:
|
|
198
|
+
[issues](https://github.com/HamzaIlyas/stillwright/issues/new?template=idea.yml), read daily.
|
|
199
|
+
|
|
200
|
+
MIT.
|
|
201
|
+
|
|
@@ -0,0 +1,183 @@
|
|
|
1
|
+
# Stillwright
|
|
2
|
+
|
|
3
|
+
[](https://github.com/HamzaIlyas/stillwright/actions/workflows/ci.yml)
|
|
4
|
+
[](https://github.com/HamzaIlyas/stillwright/releases/latest)
|
|
5
|
+
[](LICENSE)
|
|
6
|
+
|
|
7
|
+
**Your agent says "tests pass". Stillwright makes it prove it.** A step is done only when
|
|
8
|
+
its check ran and passed on the files as they are right now, with a receipt in the log, and
|
|
9
|
+
any step can be undone. For Claude Code, Codex, Gemini CLI, Cursor, OpenCode and any agent
|
|
10
|
+
that reads [Agent Skills](https://agentskills.io).
|
|
11
|
+
|
|
12
|
+

|
|
13
|
+
|
|
14
|
+
## See it work
|
|
15
|
+
|
|
16
|
+
A plan step with a command as its check. The agent cannot write "pass" by hand; it has to
|
|
17
|
+
run the check, and the receipt records what ran, the exit code and a hash of the working
|
|
18
|
+
tree:
|
|
19
|
+
|
|
20
|
+
```
|
|
21
|
+
$ sw run freeze latest tests
|
|
22
|
+
frozen: tests; any change fails the step's check until frozen again
|
|
23
|
+
|
|
24
|
+
$ sw run log latest step=1 result=pass
|
|
25
|
+
step 1 has a command check; a pass written by hand is not evidence.
|
|
26
|
+
Run `sw run check latest 1`, then `sw run done latest 1`.
|
|
27
|
+
|
|
28
|
+
$ sw run check latest 1
|
|
29
|
+
step 1: check failed (exit 1); output in ~/.stillwright/runs/20261029-031220-discount-floor/checks/1-2026-10-29031220.log
|
|
30
|
+
|
|
31
|
+
(the agent fixes pricing.py)
|
|
32
|
+
|
|
33
|
+
$ sw run check latest 1
|
|
34
|
+
step 1: check passed (exit 0, 0.1s, tree 64c2003852c2); `sw run done` when the step's work is complete
|
|
35
|
+
|
|
36
|
+
$ sw run done latest 1
|
|
37
|
+
step 1: done, certified (`python3 -m unittest -q tests.test_pricing` exit 0 on tree 64c2003852c2)
|
|
38
|
+
|
|
39
|
+
(something goes wrong later)
|
|
40
|
+
|
|
41
|
+
$ sw run rewind latest 1
|
|
42
|
+
step 1: files restored from snapshot 8bcbaf39015a (0 newer files removed);
|
|
43
|
+
the state before is kept at refs/sw/20261029-031220-discount-floor/before-rewind-2026-10-29031220
|
|
44
|
+
```
|
|
45
|
+
|
|
46
|
+
Real output from v0.2, unedited except for the home folder. If the agent edits a test after
|
|
47
|
+
`sw run freeze`, the check fails as tampered. If files change after the check, `sw run done`
|
|
48
|
+
refuses the stale receipt. The rewind keeps the state it replaced, so nothing is lost.
|
|
49
|
+
|
|
50
|
+
## What it does
|
|
51
|
+
|
|
52
|
+
- **Proof, not claims.** Every step of a plan has a check written before the work. A command
|
|
53
|
+
check is run by `sw`, with a receipt; anything else is judged by a fresh reviewer that never
|
|
54
|
+
saw the worker's reasoning.
|
|
55
|
+
- **Undo.** Every certified step leaves a git snapshot (untracked files included) under
|
|
56
|
+
`refs/sw/`; `sw run rewind` brings it back.
|
|
57
|
+
- **Vetted tools.** Thousands of skills and MCP servers, some malicious. Stillwright picks by
|
|
58
|
+
evidence, reads the code before trusting it, asks you once, and pins it by hash so it cannot
|
|
59
|
+
change silently.
|
|
60
|
+
- **Quiet.** One inbox with four card types (ask, review, notify, idea). Silence is the default,
|
|
61
|
+
and the agent's final message is at most five lines.
|
|
62
|
+
- **Goals, not prompts.** It keeps a goal queue, plans each goal into checkable steps, routes
|
|
63
|
+
steps to the cheapest model that a command can verify, and measures the dollars.
|
|
64
|
+
|
|
65
|
+
## Measured
|
|
66
|
+
|
|
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
|
|
73
|
+
[Hugging Face dataset](https://huggingface.co/datasets/Folvix/stillwright-evals).
|
|
74
|
+
|
|
75
|
+
The gate itself is tested: 67 black-box tests drive the binary on Linux, macOS and Windows in
|
|
76
|
+
CI, including the ways two independent reviews found to get past it before release.
|
|
77
|
+
|
|
78
|
+
## Install
|
|
79
|
+
|
|
80
|
+
One line for most agents:
|
|
81
|
+
|
|
82
|
+
```sh
|
|
83
|
+
npx plugins add HamzaIlyas/stillwright # as a plugin: Claude Code, Cursor, Codex, Copilot CLI, VS Code, Grok Build, Kimi Code
|
|
84
|
+
npx skills add HamzaIlyas/stillwright # as skills: 75+ agents, including all of the above
|
|
85
|
+
```
|
|
86
|
+
|
|
87
|
+
Or the agent's own way:
|
|
88
|
+
|
|
89
|
+
| Agent | Command |
|
|
90
|
+
|---|---|
|
|
91
|
+
| Claude Code | `claude plugin marketplace add HamzaIlyas/stillwright` then `claude plugin install stillwright@stillwright` |
|
|
92
|
+
| Codex | `codex plugin marketplace add HamzaIlyas/stillwright`, then install Stillwright from `/plugins` |
|
|
93
|
+
| Gemini CLI | `gemini extensions install https://github.com/HamzaIlyas/stillwright` |
|
|
94
|
+
| Cursor | loads the root `plugin.json` (Agent Plugins standard); marketplace listing pending |
|
|
95
|
+
| OpenCode, or any Agent Skills agent | `npx skills add HamzaIlyas/stillwright` |
|
|
96
|
+
| 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
|
+
|
|
98
|
+
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`).
|
|
100
|
+
|
|
101
|
+
The skills need the `sw` binary (a single file, no runtime). Any of these installs it:
|
|
102
|
+
|
|
103
|
+
```sh
|
|
104
|
+
pip install stillwright # any OS with Python
|
|
105
|
+
brew install HamzaIlyas/tap/stillwright # macOS / Linux with Homebrew
|
|
106
|
+
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/HamzaIlyas/stillwright/releases/latest/download/stillwright-installer.sh | sh # Linux / macOS
|
|
107
|
+
powershell -ExecutionPolicy Bypass -c "irm https://github.com/HamzaIlyas/stillwright/releases/latest/download/stillwright-installer.ps1 | iex" # Windows
|
|
108
|
+
cargo install --git https://github.com/HamzaIlyas/stillwright --locked # from source
|
|
109
|
+
```
|
|
110
|
+
|
|
111
|
+
## Use
|
|
112
|
+
|
|
113
|
+
New here? The [quickstart](docs/quickstart.md) takes ten minutes, and CI runs every command in it.
|
|
114
|
+
Also: [troubleshooting](docs/troubleshooting.md), [Windows notes](docs/windows.md),
|
|
115
|
+
[coming from OpenClaw or Hermes](docs/migrate.md), a [real example run](examples/quickstart-run/README.md),
|
|
116
|
+
and a [press kit](press/README.md).
|
|
117
|
+
|
|
118
|
+
Tell your agent a goal, or just say **"stillwright"**. The first time, it asks a few short
|
|
119
|
+
questions to get to know you. After that:
|
|
120
|
+
|
|
121
|
+
```
|
|
122
|
+
sw status what's going on, in a few lines
|
|
123
|
+
sw inbox list the only things that need you
|
|
124
|
+
sw goal add "..." give it more work
|
|
125
|
+
sw run status the current plan, step by step, with what is proven
|
|
126
|
+
sw run cost dollars a run spent, by model
|
|
127
|
+
```
|
|
128
|
+
|
|
129
|
+
To keep it working while you're away, schedule one loop per hour, for example
|
|
130
|
+
`claude -p "Use the stillwright skill: run one loop, then stop."` from cron. Unattended runs
|
|
131
|
+
never push, deploy, publish or install anything new; those wait in your inbox.
|
|
132
|
+
|
|
133
|
+
## What's inside
|
|
134
|
+
|
|
135
|
+
| Skill | Does |
|
|
136
|
+
|---|---|
|
|
137
|
+
| `stillwright` | the loop: pick goal → plan → get tools → execute → verify → report |
|
|
138
|
+
| `sw-onboard` | learns who you are and what you want, in three short rounds at most |
|
|
139
|
+
| `sw-scout` | finds, vets, trials and pins the best existing tool for a job |
|
|
140
|
+
| `sw-forge` | builds and tests a new skill when nothing good exists |
|
|
141
|
+
| `sw-verify` | independent, default-fail proof for every step |
|
|
142
|
+
| `sw-radar` | spots what matters to you (updates, risks, new tools) as rare idea cards |
|
|
143
|
+
| `sw-deep-search` | research that ends in verified, sourced answers |
|
|
144
|
+
| `sw-context-diet` | keeps long sessions lean and cheap |
|
|
145
|
+
| `sw-launch` | pitch, README, posts and launch plan for what you build |
|
|
146
|
+
|
|
147
|
+
All state is plain files in `~/.stillwright` (profile, goals, inbox, notes, pinned tools, run
|
|
148
|
+
logs, receipts). Nothing leaves your machine except what your agent already sends. `sw` is a
|
|
149
|
+
single Rust binary.
|
|
150
|
+
|
|
151
|
+
## Why these rules
|
|
152
|
+
|
|
153
|
+
Each one answers a failure that was measured or reported, not a hunch:
|
|
154
|
+
|
|
155
|
+
- Agents mark work done without testing it, and judges that read the agent's prose cannot
|
|
156
|
+
tell: Anthropic's [notes on long-running harnesses](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents);
|
|
157
|
+
[arXiv 2607.14890](https://arxiv.org/abs/2607.14890) (evidence-gated loops cut hidden
|
|
158
|
+
failures from 31 to 2 per 1,800). Hence receipts, and a reviewer that is a gate.
|
|
159
|
+
- Agents weaken tests to make them pass and pipe test output so failures exit 0
|
|
160
|
+
([anthropics/claude-code #319](https://github.com/anthropics/claude-code/issues/319)). Hence
|
|
161
|
+
frozen check files and `pipefail`.
|
|
162
|
+
- Undo of agent edits is the most-requested feature no host ships
|
|
163
|
+
([openai/codex #9203](https://github.com/openai/codex/issues/9203)). Hence snapshots and rewind.
|
|
164
|
+
- Agents rarely ask at the right moment and fill gaps with wrong assumptions:
|
|
165
|
+
[HiL-Bench](https://arxiv.org/abs/2604.09408). Hence the typed inbox, silence by default.
|
|
166
|
+
- Hundreds of malicious skills were found on a public registry in early 2026 (341 of 2,857):
|
|
167
|
+
[Snyk](https://snyk.io/articles/skill-md-shell-access/),
|
|
168
|
+
[Unit 42](https://unit42.paloaltonetworks.com/openclaw-ai-supply-chain-risk/). Hence vetting,
|
|
169
|
+
approval and hash-pinning before any third-party tool runs.
|
|
170
|
+
- Agent success falls as tasks get longer: [METR time horizons](https://metr.org/time-horizons/).
|
|
171
|
+
Hence small, checkable steps.
|
|
172
|
+
|
|
173
|
+
Cost is measured, not assumed: the harness, the price table and the cache linter are in
|
|
174
|
+
[docs/cost.md](docs/cost.md). Known limits are listed in the [CHANGELOG](CHANGELOG.md).
|
|
175
|
+
|
|
176
|
+
## Who runs it
|
|
177
|
+
|
|
178
|
+
One person, [Hamza Ilyas](https://github.com/HamzaIlyas), building it at
|
|
179
|
+
[Folvix](https://huggingface.co/Folvix). Questions and ideas:
|
|
180
|
+
[Discussions](https://github.com/HamzaIlyas/stillwright/discussions). Bugs:
|
|
181
|
+
[issues](https://github.com/HamzaIlyas/stillwright/issues/new?template=idea.yml), read daily.
|
|
182
|
+
|
|
183
|
+
MIT.
|