You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
SkillForge helps Codex extension authors scaffold, lint, smoke-test, inspect, and package skills/plugins before they publish or submit them to a marketplace.
9
+
SkillForge helps agent extension authors scaffold, lint, smoke-test, inspect, and package skills/plugins before they publish or submit them to a marketplace.
10
+
11
+
Codex is the first-class target today. Portable skill compatibility for Claude-style skills and cross-agent packages is now starting with `skillforge compat`.
10
12
11
13
Listed in [awesome-codex-plugins](https://github.com/hashgraph-online/awesome-codex-plugins) under "Validate Before You Ship."
Codex skills and plugins are small, powerful folders. They are also easy to get subtly wrong:
30
+
Agent skills and plugins are small, powerful folders. They are also easy to get subtly wrong:
29
31
30
32
- weak skill descriptions that Codex will not trigger well
31
33
- stale `agents/openai.yaml` shapes
32
34
- plugin paths that are not `./`-relative
33
35
- missing bundled skills, hooks, MCP, app, or asset files
34
-
- plugins that work locally but are not marketplace-ready
36
+
- packages that work locally but are not marketplace-ready
37
+
- skills that claim cross-agent compatibility but still contain agent-specific assumptions
35
38
36
-
SkillForge is not a marketplace. It is the publish-readiness check you run before sharing a Codex skill/plugin repo.
39
+
SkillForge is not a marketplace. It is the publish-readiness check you run before sharing an agent skill/plugin repo.
37
40
38
41
## Security Model
39
42
40
-
SkillForge is a CLI linter, not a Codex runtime plugin. Running `lint`, `doctor`, and `smoke` reads local files and reports issues; it does not install skills, load plugins into Codex, or execute scripts from the target project.
43
+
SkillForge is a CLI linter, not an agent runtime plugin. Running `lint`, `compat`, `doctor`, and `smoke` reads local files and reports issues; it does not install skills, load plugins into Codex, or execute scripts from the target project.
41
44
42
45
Commands that write files are explicit:
43
46
@@ -47,32 +50,34 @@ Commands that write files are explicit:
47
50
For cautious use, pin the npm version, review the source, and start with read-only commands:
Default lint mode focuses on deterministic publish-readiness problems. Use `--strict` to include advisory checks such as trigger-description quality, large skill bodies, unreferenced scripts, and plugin name/folder mismatch.
149
156
157
+
## Compatibility
158
+
159
+
Use `compat` to check whether a skill/package is likely to work in a specific agent ecosystem.
160
+
161
+
```bash
162
+
skillforge compat . --target codex
163
+
skillforge compat . --target claude
164
+
skillforge compat . --target portable
165
+
```
166
+
167
+
`codex` runs Codex-oriented lint checks.
168
+
169
+
`claude` checks the shared `SKILL.md` basics expected by Claude-style skills.
170
+
171
+
`portable` combines both and adds warnings for agent-specific wording, `.claude/`, `.codex/`, `.agents/`, and obvious OS-specific script assumptions.
172
+
173
+
This is an analysis command, not an auto-converter.
174
+
150
175
## Profiles
151
176
152
177
Use profiles to match where the plugin is being checked:
@@ -178,7 +203,7 @@ on:
178
203
branches: [main]
179
204
180
205
jobs:
181
-
lint-codex-extensions:
206
+
lint-agent-extensions:
182
207
runs-on: ubuntu-latest
183
208
steps:
184
209
- uses: actions/checkout@v4
@@ -281,7 +306,7 @@ Verify from a clean directory after npm publish:
Default mode now focuses on higher-confidence publish-readiness problems. Strict mode includes advisory checks such as trigger-description quality, large skill bodies, unreferenced scripts, and plugin folder/name mismatch.
16
+
Default mode now focuses on deterministic publish-readiness problems. Strict mode includes advisory checks such as trigger-description quality, large skill bodies, unreferenced scripts, and plugin folder/name mismatch.
17
17
18
18
## Most Common Default Findings
19
19
@@ -40,7 +40,7 @@ The initial scan showed that some checks were too noisy for default CI usage. In
40
40
41
41
-`--strict` was added for advisory checks.
42
42
- Default mode became lower-noise.
43
-
-Findings gained confidence metadata.
43
+
-Later releases replaced confidence metadata with deterministic `blocking` and `advisory` impact labels.
44
44
-`reference.missing` was tightened so placeholder words and external URLs are not treated as missing files.
45
45
- A regression fixture was added for reference-noise cases.
0 commit comments