@hybridlabor-api/aos 4.8.0 → 4.10.0
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.
- package/.agents/{agents.md → AGENTS.md} +3 -1
- package/.agents/nodes.json +3 -1
- package/.agents/vendor-manifest.json +23 -1
- package/.claude/agents/godmode-media-eventtech.md +1 -1
- package/.claude/hooks/conventional-commits.mjs +125 -0
- package/.claude/hooks/env-file-protection.mjs +105 -0
- package/.claude/hooks/go-gate.mjs +101 -81
- package/.claude/hooks/memb-inject.mjs +29 -1
- package/.claude/settings.json +13 -0
- package/.claude/workflows/startcycle-dispatch.mjs +23 -1
- package/.opencode/agents/godmode-media-eventtech.md +1 -1
- package/.opencode/plugins/bdb-aos.js +31 -4
- package/CLAUDE.md +0 -571
- package/README.de.md +1 -1
- package/README.md +1 -1
- package/README.pt.md +1 -1
- package/THIRD_PARTY_NOTICES.md +126 -0
- package/bin/aos-doctor.mjs +1 -1
- package/docs/skills_table.md +1 -1
- package/installer.js +187 -55
- package/package.json +7 -3
- package/packages/aos-cli/README.md +80 -0
- package/packages/aos-cli/bin/aos-cli.mjs +134 -0
- package/packages/aos-cli/core-skills.json +12 -0
- package/packages/aos-cli/extensions/aos.ts +321 -0
- package/packages/aos-cli/package-lock.json +1923 -0
- package/packages/aos-cli/package.json +29 -0
- package/packages/aos-cli/scripts/check-theme.mjs +63 -0
- package/packages/aos-cli/themes/aos.json +97 -0
- package/scripts/build-plugin-manifest.mjs +131 -0
- package/scripts/validate-skills.mjs +81 -6
- package/skills/basic/ao-orchestrator/SKILL.md +116 -0
- package/skills/basic/bdb-eventagency-skill/SKILL.md +252 -0
- package/skills/basic/bdb-shipping-skill/SKILL.md +161 -0
- package/skills/basic/godmode-eventtech/SKILL.md +4 -1
- package/skills/global_config/agenttrail/SKILL.md +6 -1
- package/skills/global_config/aos-project-init/SKILL.md +2 -0
- package/skills/global_config/aos-project-init/assets/AGENTS.template.md +1 -1
- package/skills/global_config/aos-project-init/scripts/aos-project-doctor.mjs +1 -1
- package/skills/global_config/aos-setup/SKILL.md +1 -1
- package/skills/global_config/aos-setup/scripts/aos-doctor.mjs +1 -1
- package/skills/global_config/ask-tim/SKILL.md +7 -7
- package/skills/global_config/bash-script-generator/SKILL.md +201 -0
- package/skills/global_config/bash-script-generator/assets/templates/standard-template.sh +96 -0
- package/skills/global_config/bash-script-generator/docs/bash-scripting-guide.md +729 -0
- package/skills/global_config/bash-script-generator/docs/generation-best-practices.md +193 -0
- package/skills/global_config/bash-script-generator/docs/script-patterns.md +566 -0
- package/skills/global_config/bash-script-generator/docs/text-processing-guide.md +437 -0
- package/skills/global_config/bash-script-generator/examples/log-analyzer.sh +92 -0
- package/skills/global_config/bash-script-generator/scripts/generate_script_template.sh +123 -0
- package/skills/global_config/bash-script-generator/scripts/run_ci_checks.sh +172 -0
- package/skills/global_config/bash-script-generator/scripts/test_generator.sh +413 -0
- package/skills/global_config/bash-script-validator/SKILL.md +249 -0
- package/skills/global_config/bash-script-validator/docs/awk-reference.md +449 -0
- package/skills/global_config/bash-script-validator/docs/bash-reference.md +468 -0
- package/skills/global_config/bash-script-validator/docs/common-mistakes.md +623 -0
- package/skills/global_config/bash-script-validator/docs/grep-reference.md +395 -0
- package/skills/global_config/bash-script-validator/docs/regex-reference.md +391 -0
- package/skills/global_config/bash-script-validator/docs/sed-reference.md +454 -0
- package/skills/global_config/bash-script-validator/docs/shell-reference.md +463 -0
- package/skills/global_config/bash-script-validator/docs/shellcheck-reference.md +399 -0
- package/skills/global_config/bash-script-validator/examples/bad-bash.sh +55 -0
- package/skills/global_config/bash-script-validator/examples/bad-shell.sh +54 -0
- package/skills/global_config/bash-script-validator/examples/good-bash.sh +71 -0
- package/skills/global_config/bash-script-validator/examples/good-shell.sh +69 -0
- package/skills/global_config/bash-script-validator/scripts/run_ci_checks.sh +23 -0
- package/skills/global_config/bash-script-validator/scripts/shellcheck_wrapper.sh +174 -0
- package/skills/global_config/bash-script-validator/scripts/test_validate.sh +446 -0
- package/skills/global_config/bash-script-validator/scripts/validate.sh +512 -0
- package/skills/global_config/ci-pipeline/SKILL.md +135 -0
- package/skills/global_config/deja-memory/SKILL.md +3 -1
- package/skills/global_config/dispatching-parallel-agents/SKILL.md +170 -0
- package/skills/global_config/dockerfile-generator/SKILL.md +1038 -0
- package/skills/global_config/dockerfile-generator/examples/example.dockerignore +95 -0
- package/skills/global_config/dockerfile-generator/examples/golang-distroless.Dockerfile +34 -0
- package/skills/global_config/dockerfile-generator/examples/java-springboot.Dockerfile +45 -0
- package/skills/global_config/dockerfile-generator/examples/nextjs-production.Dockerfile +49 -0
- package/skills/global_config/dockerfile-generator/examples/nodejs-multistage.Dockerfile +55 -0
- package/skills/global_config/dockerfile-generator/examples/python-fastapi.Dockerfile +48 -0
- package/skills/global_config/dockerfile-generator/references/language_specific_guides.md +510 -0
- package/skills/global_config/dockerfile-generator/references/multistage_builds.md +570 -0
- package/skills/global_config/dockerfile-generator/references/optimization_patterns.md +492 -0
- package/skills/global_config/dockerfile-generator/references/security_best_practices.md +375 -0
- package/skills/global_config/dockerfile-generator/scripts/generate_dockerignore.sh +199 -0
- package/skills/global_config/dockerfile-generator/scripts/generate_golang.sh +172 -0
- package/skills/global_config/dockerfile-generator/scripts/generate_java.sh +185 -0
- package/skills/global_config/dockerfile-generator/scripts/generate_nodejs.sh +279 -0
- package/skills/global_config/dockerfile-generator/scripts/generate_python.sh +218 -0
- package/skills/global_config/dockerfile-generator/scripts/test_generator.sh +210 -0
- package/skills/global_config/dockerfile-validator/SKILL.md +300 -0
- package/skills/global_config/dockerfile-validator/examples/.dockerignore.example +34 -0
- package/skills/global_config/dockerfile-validator/examples/bad-example.Dockerfile +37 -0
- package/skills/global_config/dockerfile-validator/examples/golang-distroless.Dockerfile +50 -0
- package/skills/global_config/dockerfile-validator/examples/good-example.Dockerfile +52 -0
- package/skills/global_config/dockerfile-validator/examples/python-optimized.Dockerfile +53 -0
- package/skills/global_config/dockerfile-validator/examples/security-issues.Dockerfile +42 -0
- package/skills/global_config/dockerfile-validator/references/docker_best_practices.md +348 -0
- package/skills/global_config/dockerfile-validator/references/optimization_guide.md +473 -0
- package/skills/global_config/dockerfile-validator/references/security_checklist.md +208 -0
- package/skills/global_config/dockerfile-validator/scripts/dockerfile-validate.sh +699 -0
- package/skills/global_config/dockerfile-validator/scripts/test_validate.sh +35 -0
- package/skills/global_config/dockerfile-validator/tests/fixtures/copy-before-yarn-lock-read.Dockerfile +6 -0
- package/skills/global_config/dockerfile-validator/tests/fixtures/copy-before-yarn.Dockerfile +6 -0
- package/skills/global_config/dockerfile-validator/tests/fixtures/from-platform-nonroot.Dockerfile +7 -0
- package/skills/global_config/dockerfile-validator/tests/test_regression.sh +164 -0
- package/skills/global_config/finishing-a-development-branch/SKILL.md +228 -0
- package/skills/global_config/github-actions-generator/SKILL.md +353 -0
- package/skills/global_config/github-actions-generator/assets/templates/action/composite/action.yml +82 -0
- package/skills/global_config/github-actions-generator/assets/templates/action/docker/Dockerfile +25 -0
- package/skills/global_config/github-actions-generator/assets/templates/action/docker/action.yml +42 -0
- package/skills/global_config/github-actions-generator/assets/templates/action/docker/entrypoint.sh +27 -0
- package/skills/global_config/github-actions-generator/assets/templates/action/javascript/action.yml +33 -0
- package/skills/global_config/github-actions-generator/assets/templates/action/javascript/index.js +50 -0
- package/skills/global_config/github-actions-generator/assets/templates/action/javascript/package.json +27 -0
- package/skills/global_config/github-actions-generator/assets/templates/workflow/basic_workflow.yml +242 -0
- package/skills/global_config/github-actions-generator/assets/templates/workflow/reusable_workflow.yml +106 -0
- package/skills/global_config/github-actions-generator/examples/README.md +147 -0
- package/skills/global_config/github-actions-generator/examples/actions/setup-node-cached/action.yml +93 -0
- package/skills/global_config/github-actions-generator/examples/caching/docker-buildkit.yml +256 -0
- package/skills/global_config/github-actions-generator/examples/security/dependency-review.yml +62 -0
- package/skills/global_config/github-actions-generator/examples/security/sbom-attestation.yml +119 -0
- package/skills/global_config/github-actions-generator/examples/triggers/chatops-commands.yml +475 -0
- package/skills/global_config/github-actions-generator/examples/triggers/repository-dispatch.yml +418 -0
- package/skills/global_config/github-actions-generator/examples/triggers/workflow-orchestration.yml +404 -0
- package/skills/global_config/github-actions-generator/examples/workflows/docker-build-push.yml +68 -0
- package/skills/global_config/github-actions-generator/examples/workflows/go-ci.yml +161 -0
- package/skills/global_config/github-actions-generator/examples/workflows/monorepo-ci.yml +340 -0
- package/skills/global_config/github-actions-generator/examples/workflows/multi-environment-deploy.yml +406 -0
- package/skills/global_config/github-actions-generator/examples/workflows/nodejs-ci.yml +122 -0
- package/skills/global_config/github-actions-generator/examples/workflows/python-ci.yml +157 -0
- package/skills/global_config/github-actions-generator/examples/workflows/scheduled-tasks.yml +376 -0
- package/skills/global_config/github-actions-generator/references/advanced-triggers.md +917 -0
- package/skills/global_config/github-actions-generator/references/best-practices.md +755 -0
- package/skills/global_config/github-actions-generator/references/common-actions.md +715 -0
- package/skills/global_config/github-actions-generator/references/custom-actions.md +320 -0
- package/skills/global_config/github-actions-generator/references/expressions-and-contexts.md +688 -0
- package/skills/global_config/github-actions-generator/references/modern-features.md +421 -0
- package/skills/global_config/github-actions-generator/scripts/test_generator.sh +344 -0
- package/skills/global_config/github-actions-templates/SKILL.md +7 -0
- package/skills/global_config/github-actions-validator/SKILL.md +576 -0
- package/skills/global_config/github-actions-validator/examples/README.md +88 -0
- package/skills/global_config/github-actions-validator/examples/outdated-versions.yml +76 -0
- package/skills/global_config/github-actions-validator/examples/valid-ci.yml +79 -0
- package/skills/global_config/github-actions-validator/examples/with-errors.yml +47 -0
- package/skills/global_config/github-actions-validator/references/act_usage.md +233 -0
- package/skills/global_config/github-actions-validator/references/action_versions.md +122 -0
- package/skills/global_config/github-actions-validator/references/actionlint_usage.md +343 -0
- package/skills/global_config/github-actions-validator/references/common_errors.md +512 -0
- package/skills/global_config/github-actions-validator/references/modern_features.md +384 -0
- package/skills/global_config/github-actions-validator/references/runners.md +317 -0
- package/skills/global_config/github-actions-validator/scripts/install_tools.sh +113 -0
- package/skills/global_config/github-actions-validator/scripts/validate_workflow.sh +910 -0
- package/skills/global_config/github-actions-validator/tests/test_validate_workflow.sh +237 -0
- package/skills/global_config/makefile-generator/SKILL.md +614 -0
- package/skills/global_config/makefile-generator/assets/templates/.gitkeep +1 -0
- package/skills/global_config/makefile-generator/docs/makefile-structure.md +530 -0
- package/skills/global_config/makefile-generator/docs/optimization-guide.md +784 -0
- package/skills/global_config/makefile-generator/docs/patterns-guide.md +642 -0
- package/skills/global_config/makefile-generator/docs/security-guide.md +361 -0
- package/skills/global_config/makefile-generator/docs/targets-guide.md +642 -0
- package/skills/global_config/makefile-generator/docs/variables-guide.md +596 -0
- package/skills/global_config/makefile-generator/scripts/add_standard_targets.sh +539 -0
- package/skills/global_config/makefile-generator/scripts/generate_makefile_template.sh +690 -0
- package/skills/global_config/makefile-generator/test/test_helper_scripts.sh +190 -0
- package/skills/global_config/makefile-validator/SKILL.md +244 -0
- package/skills/global_config/makefile-validator/docs/bake-tool.md +1000 -0
- package/skills/global_config/makefile-validator/docs/best-practices.md +858 -0
- package/skills/global_config/makefile-validator/docs/common-mistakes.md +944 -0
- package/skills/global_config/makefile-validator/examples/bad-makefile.mk +77 -0
- package/skills/global_config/makefile-validator/examples/good-makefile.mk +103 -0
- package/skills/global_config/makefile-validator/scripts/test_validate.sh +382 -0
- package/skills/global_config/makefile-validator/scripts/validate_makefile.sh +712 -0
- package/skills/global_config/{MCP_Manage → mcp-manage}/SKILL.md +3 -3
- package/skills/global_config/plan-canvas/SKILL.md +9 -2
- package/skills/global_config/plan-canvas/scripts/lib/plan-canvas/ui.js +10 -2
- package/skills/global_config/plan-canvas/scripts/plan-canvas.js +1 -1
- package/skills/global_config/read-the-damn-docs/SKILL.md +175 -0
- package/skills/global_config/requesting-code-review/SKILL.md +98 -0
- package/skills/global_config/requesting-code-review/code-reviewer.md +198 -0
- package/skills/global_config/using-git-worktrees/SKILL.md +170 -0
- package/skills/global_config/verification-before-completion/SKILL.md +123 -0
- package/skills/global_config/writing-plans/SKILL.md +126 -46
- package/skills/global_config/writing-plans-legacy/SKILL.md +152 -0
- package/.claude/CLAUDE.md +0 -12
- package/mcps/RhinoMCP/docs/content/docs/getting-started/gemini.md +0 -61
- /package/{GEMINI.md → RULES.md} +0 -0
|
@@ -5,7 +5,7 @@ model: opencode/muse-spark-1.3-contributor-free
|
|
|
5
5
|
---
|
|
6
6
|
Creative-Tech & Show-Control Specialist. Governs 3D modeling, spatial design, TouchDesigner networks, Unreal Engine scenes, DaVinci Resolve color/edit, DMX/grandMA3 lighting, and Resolume media servers.
|
|
7
7
|
|
|
8
|
-
**Primary skills:** godmode-eventtech, godmode-media-creation, godmode-3d-creation, bdbmediastorm,
|
|
8
|
+
**Primary skills:** godmode-eventtech, godmode-media-creation, godmode-3d-creation, bdbmediastorm, mcp-manage, threejs-skills, spline-3d-integration
|
|
9
9
|
|
|
10
10
|
**MCP servers used:** bdb_td_minddesigner, bdb_unreal_mcp, bdb_davinci_mcp, bdb_grandma3_mcp, bdb_resolume_mcp, adobe_uxp_mcp, bdb_after_effects_mcp, bdb_rhino_mcp
|
|
11
11
|
|
|
@@ -19,10 +19,14 @@ import os from 'node:os';
|
|
|
19
19
|
import path from 'node:path';
|
|
20
20
|
|
|
21
21
|
const GUARDED_PATTERNS = [
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
22
|
+
/(?:^|[;&|]\s*)git\s+push\b/i,
|
|
23
|
+
/(?:^|[;&|]\s*)npm\s+publish\b/i,
|
|
24
|
+
/(?:^|[;&|]\s*)npm\s+version\b/i,
|
|
25
|
+
/(?:^|[;&|]\s*)gh\s+pr\s+merge\b/i,
|
|
26
|
+
/(?:^|[;&|]\s*)gh\s+release\s+create\b/i,
|
|
27
|
+
/(?:^|[;&|]\s*)git\s+reset\s+--hard\b/i,
|
|
28
|
+
/(?:^|[;&|]\s*)git\s+clean\s+-[a-zA-Z]*f\b/i,
|
|
29
|
+
/(?:^|[;&|]\s*)rm\s+(?:-\w*[rR]\w*|--recursive)\b/i,
|
|
26
30
|
];
|
|
27
31
|
|
|
28
32
|
let lastHumanPrompt = '';
|
|
@@ -176,6 +180,29 @@ export default async function bdbAosPlugin(input) {
|
|
|
176
180
|
});
|
|
177
181
|
}
|
|
178
182
|
} catch {}
|
|
183
|
+
|
|
184
|
+
// Recognize and wire /startcycle-graph workflow
|
|
185
|
+
const graphMatch = fullText.match(/^\/startcycle-graph(?:\s+(.*))?$/is);
|
|
186
|
+
if (graphMatch) {
|
|
187
|
+
const goal = (graphMatch[1] || '').trim();
|
|
188
|
+
const graphInstructions = [
|
|
189
|
+
`[AOS Autonomous Graph Workflow Engine - Active]`,
|
|
190
|
+
`Goal: "${goal || 'Execute planned architecture cycle'}"`,
|
|
191
|
+
`State Schema: .agents/state.schema.json`,
|
|
192
|
+
`Persisted State: production_artifacts/state.json`,
|
|
193
|
+
`Available Nodes: .agents/nodes.json (Architect -> TechLead -> Build [UI_UX, Engineering, Media] -> Reviewer -> Shipping)`,
|
|
194
|
+
`Rule: Nodes never call each other. Dispatch each node turn in sequence, evaluate state.json, and enforce Reviewer & Shipping gates before completion.`
|
|
195
|
+
].join('\n');
|
|
196
|
+
|
|
197
|
+
msgOutput.parts.unshift({
|
|
198
|
+
id: `startcycle-${Date.now()}`,
|
|
199
|
+
sessionID: msgInput.sessionID,
|
|
200
|
+
messageID: msgInput.messageID || '',
|
|
201
|
+
type: 'text',
|
|
202
|
+
text: graphInstructions,
|
|
203
|
+
synthetic: true,
|
|
204
|
+
});
|
|
205
|
+
}
|
|
179
206
|
},
|
|
180
207
|
|
|
181
208
|
'tool.execute.before': async (toolInput, toolOutput) => {
|
package/CLAUDE.md
CHANGED
|
@@ -50,574 +50,3 @@ npm test # selftest + validate
|
|
|
50
50
|
```
|
|
51
51
|
|
|
52
52
|
|
|
53
|
-
# AOS — Antigravity / Gemini
|
|
54
|
-
|
|
55
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
56
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
57
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
58
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
59
|
-
rule that will eventually disagree with itself.
|
|
60
|
-
|
|
61
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
62
|
-
|
|
63
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
64
|
-
|
|
65
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
66
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
67
|
-
and neither is anything else.
|
|
68
|
-
|
|
69
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
70
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
71
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
72
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
73
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
74
|
-
this the harness where the rule matters most, not least.
|
|
75
|
-
|
|
76
|
-
## Model routing
|
|
77
|
-
|
|
78
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
79
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
80
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
81
|
-
effort suffix.
|
|
82
|
-
|
|
83
|
-
| Work | Model |
|
|
84
|
-
|---|---|
|
|
85
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
86
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
87
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
88
|
-
|
|
89
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
90
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
91
|
-
source.
|
|
92
|
-
|
|
93
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
94
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
95
|
-
must not do.
|
|
96
|
-
|
|
97
|
-
## Headless timeouts
|
|
98
|
-
|
|
99
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
100
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
101
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
102
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
103
|
-
"broken".
|
|
104
|
-
|
|
105
|
-
# AOS — Antigravity / Gemini
|
|
106
|
-
|
|
107
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
108
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
109
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
110
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
111
|
-
rule that will eventually disagree with itself.
|
|
112
|
-
|
|
113
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
114
|
-
|
|
115
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
116
|
-
|
|
117
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
118
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
119
|
-
and neither is anything else.
|
|
120
|
-
|
|
121
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
122
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
123
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
124
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
125
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
126
|
-
this the harness where the rule matters most, not least.
|
|
127
|
-
|
|
128
|
-
## Model routing
|
|
129
|
-
|
|
130
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
131
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
132
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
133
|
-
effort suffix.
|
|
134
|
-
|
|
135
|
-
| Work | Model |
|
|
136
|
-
|---|---|
|
|
137
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
138
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
139
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
140
|
-
|
|
141
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
142
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
143
|
-
source.
|
|
144
|
-
|
|
145
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
146
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
147
|
-
must not do.
|
|
148
|
-
|
|
149
|
-
## Headless timeouts
|
|
150
|
-
|
|
151
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
152
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
153
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
154
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
155
|
-
"broken".
|
|
156
|
-
|
|
157
|
-
# AOS — Antigravity / Gemini
|
|
158
|
-
|
|
159
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
160
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
161
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
162
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
163
|
-
rule that will eventually disagree with itself.
|
|
164
|
-
|
|
165
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
166
|
-
|
|
167
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
168
|
-
|
|
169
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
170
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
171
|
-
and neither is anything else.
|
|
172
|
-
|
|
173
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
174
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
175
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
176
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
177
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
178
|
-
this the harness where the rule matters most, not least.
|
|
179
|
-
|
|
180
|
-
## Model routing
|
|
181
|
-
|
|
182
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
183
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
184
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
185
|
-
effort suffix.
|
|
186
|
-
|
|
187
|
-
| Work | Model |
|
|
188
|
-
|---|---|
|
|
189
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
190
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
191
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
192
|
-
|
|
193
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
194
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
195
|
-
source.
|
|
196
|
-
|
|
197
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
198
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
199
|
-
must not do.
|
|
200
|
-
|
|
201
|
-
## Headless timeouts
|
|
202
|
-
|
|
203
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
204
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
205
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
206
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
207
|
-
"broken".
|
|
208
|
-
|
|
209
|
-
# AOS — Antigravity / Gemini
|
|
210
|
-
|
|
211
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
212
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
213
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
214
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
215
|
-
rule that will eventually disagree with itself.
|
|
216
|
-
|
|
217
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
218
|
-
|
|
219
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
220
|
-
|
|
221
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
222
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
223
|
-
and neither is anything else.
|
|
224
|
-
|
|
225
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
226
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
227
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
228
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
229
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
230
|
-
this the harness where the rule matters most, not least.
|
|
231
|
-
|
|
232
|
-
## Model routing
|
|
233
|
-
|
|
234
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
235
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
236
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
237
|
-
effort suffix.
|
|
238
|
-
|
|
239
|
-
| Work | Model |
|
|
240
|
-
|---|---|
|
|
241
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
242
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
243
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
244
|
-
|
|
245
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
246
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
247
|
-
source.
|
|
248
|
-
|
|
249
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
250
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
251
|
-
must not do.
|
|
252
|
-
|
|
253
|
-
## Headless timeouts
|
|
254
|
-
|
|
255
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
256
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
257
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
258
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
259
|
-
"broken".
|
|
260
|
-
|
|
261
|
-
# AOS — Antigravity / Gemini
|
|
262
|
-
|
|
263
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
264
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
265
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
266
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
267
|
-
rule that will eventually disagree with itself.
|
|
268
|
-
|
|
269
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
270
|
-
|
|
271
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
272
|
-
|
|
273
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
274
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
275
|
-
and neither is anything else.
|
|
276
|
-
|
|
277
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
278
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
279
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
280
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
281
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
282
|
-
this the harness where the rule matters most, not least.
|
|
283
|
-
|
|
284
|
-
## Model routing
|
|
285
|
-
|
|
286
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
287
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
288
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
289
|
-
effort suffix.
|
|
290
|
-
|
|
291
|
-
| Work | Model |
|
|
292
|
-
|---|---|
|
|
293
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
294
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
295
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
296
|
-
|
|
297
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
298
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
299
|
-
source.
|
|
300
|
-
|
|
301
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
302
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
303
|
-
must not do.
|
|
304
|
-
|
|
305
|
-
## Headless timeouts
|
|
306
|
-
|
|
307
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
308
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
309
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
310
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
311
|
-
"broken".
|
|
312
|
-
|
|
313
|
-
# AOS — Antigravity / Gemini
|
|
314
|
-
|
|
315
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
316
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
317
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
318
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
319
|
-
rule that will eventually disagree with itself.
|
|
320
|
-
|
|
321
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
322
|
-
|
|
323
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
324
|
-
|
|
325
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
326
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
327
|
-
and neither is anything else.
|
|
328
|
-
|
|
329
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
330
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
331
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
332
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
333
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
334
|
-
this the harness where the rule matters most, not least.
|
|
335
|
-
|
|
336
|
-
## Model routing
|
|
337
|
-
|
|
338
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
339
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
340
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
341
|
-
effort suffix.
|
|
342
|
-
|
|
343
|
-
| Work | Model |
|
|
344
|
-
|---|---|
|
|
345
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
346
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
347
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
348
|
-
|
|
349
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
350
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
351
|
-
source.
|
|
352
|
-
|
|
353
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
354
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
355
|
-
must not do.
|
|
356
|
-
|
|
357
|
-
## Headless timeouts
|
|
358
|
-
|
|
359
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
360
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
361
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
362
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
363
|
-
"broken".
|
|
364
|
-
|
|
365
|
-
# AOS — Antigravity / Gemini
|
|
366
|
-
|
|
367
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
368
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
369
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
370
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
371
|
-
rule that will eventually disagree with itself.
|
|
372
|
-
|
|
373
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
374
|
-
|
|
375
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
376
|
-
|
|
377
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
378
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
379
|
-
and neither is anything else.
|
|
380
|
-
|
|
381
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
382
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
383
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
384
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
385
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
386
|
-
this the harness where the rule matters most, not least.
|
|
387
|
-
|
|
388
|
-
## Model routing
|
|
389
|
-
|
|
390
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
391
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
392
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
393
|
-
effort suffix.
|
|
394
|
-
|
|
395
|
-
| Work | Model |
|
|
396
|
-
|---|---|
|
|
397
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
398
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
399
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
400
|
-
|
|
401
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
402
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
403
|
-
source.
|
|
404
|
-
|
|
405
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
406
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
407
|
-
must not do.
|
|
408
|
-
|
|
409
|
-
## Headless timeouts
|
|
410
|
-
|
|
411
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
412
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
413
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
414
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
415
|
-
"broken".
|
|
416
|
-
|
|
417
|
-
# AOS — Antigravity / Gemini
|
|
418
|
-
|
|
419
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
420
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
421
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
422
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
423
|
-
rule that will eventually disagree with itself.
|
|
424
|
-
|
|
425
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
426
|
-
|
|
427
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
428
|
-
|
|
429
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
430
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
431
|
-
and neither is anything else.
|
|
432
|
-
|
|
433
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
434
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
435
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
436
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
437
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
438
|
-
this the harness where the rule matters most, not least.
|
|
439
|
-
|
|
440
|
-
## Model routing
|
|
441
|
-
|
|
442
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
443
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
444
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
445
|
-
effort suffix.
|
|
446
|
-
|
|
447
|
-
| Work | Model |
|
|
448
|
-
|---|---|
|
|
449
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
450
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
451
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
452
|
-
|
|
453
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
454
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
455
|
-
source.
|
|
456
|
-
|
|
457
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
458
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
459
|
-
must not do.
|
|
460
|
-
|
|
461
|
-
## Headless timeouts
|
|
462
|
-
|
|
463
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
464
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
465
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
466
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
467
|
-
"broken".
|
|
468
|
-
|
|
469
|
-
# AOS — Antigravity / Gemini
|
|
470
|
-
|
|
471
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
472
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
473
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
474
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
475
|
-
rule that will eventually disagree with itself.
|
|
476
|
-
|
|
477
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
478
|
-
|
|
479
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
480
|
-
|
|
481
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
482
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
483
|
-
and neither is anything else.
|
|
484
|
-
|
|
485
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
486
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
487
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
488
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
489
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
490
|
-
this the harness where the rule matters most, not least.
|
|
491
|
-
|
|
492
|
-
## Model routing
|
|
493
|
-
|
|
494
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
495
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
496
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
497
|
-
effort suffix.
|
|
498
|
-
|
|
499
|
-
| Work | Model |
|
|
500
|
-
|---|---|
|
|
501
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
502
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
503
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
504
|
-
|
|
505
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
506
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
507
|
-
source.
|
|
508
|
-
|
|
509
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
510
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
511
|
-
must not do.
|
|
512
|
-
|
|
513
|
-
## Headless timeouts
|
|
514
|
-
|
|
515
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
516
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
517
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
518
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
519
|
-
"broken".
|
|
520
|
-
|
|
521
|
-
# AOS — Antigravity / Gemini
|
|
522
|
-
|
|
523
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
524
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
525
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
526
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
527
|
-
rule that will eventually disagree with itself.
|
|
528
|
-
|
|
529
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
530
|
-
|
|
531
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
532
|
-
|
|
533
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
534
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
535
|
-
and neither is anything else.
|
|
536
|
-
|
|
537
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
538
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
539
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
540
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
541
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
542
|
-
this the harness where the rule matters most, not least.
|
|
543
|
-
|
|
544
|
-
## Model routing
|
|
545
|
-
|
|
546
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
547
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
548
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
549
|
-
effort suffix.
|
|
550
|
-
|
|
551
|
-
| Work | Model |
|
|
552
|
-
|---|---|
|
|
553
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
554
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
555
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
556
|
-
|
|
557
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
558
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
559
|
-
source.
|
|
560
|
-
|
|
561
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
562
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
563
|
-
must not do.
|
|
564
|
-
|
|
565
|
-
## Headless timeouts
|
|
566
|
-
|
|
567
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
568
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
569
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
570
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
571
|
-
"broken".
|
|
572
|
-
|
|
573
|
-
# AOS — Antigravity / Gemini
|
|
574
|
-
|
|
575
|
-
**Read [AGENTS.md](AGENTS.md) first.** It holds every rule that applies to all
|
|
576
|
-
harnesses: the non-negotiables, the release gate, Conventional Commits, the
|
|
577
|
-
skill contract and category routing, the pipeline variants, and the delegation
|
|
578
|
-
policy. Nothing from it is repeated here — a rule that lives in two files is a
|
|
579
|
-
rule that will eventually disagree with itself.
|
|
580
|
-
|
|
581
|
-
This file covers only what is specific to running AOS under Antigravity.
|
|
582
|
-
|
|
583
|
-
## Nothing enforces the gate here — you are the enforcement
|
|
584
|
-
|
|
585
|
-
`.claude/hooks/go-gate.mjs` is a Claude Code hook and does not run on this
|
|
586
|
-
harness. Not one of the four release commands is mechanically blocked for you,
|
|
587
|
-
and neither is anything else.
|
|
588
|
-
|
|
589
|
-
So the full rule in AGENTS.md applies here in its entirety, not just its
|
|
590
|
-
hook-enforced subset: when the user asks for a plan, review, audit or
|
|
591
|
-
multi-step action, you are read-only until they answer with the literal **GO**
|
|
592
|
-
— including file writes and `git commit`. On Claude Code a hook would catch the
|
|
593
|
-
four worst cases if discipline failed. Here there is no backstop, which makes
|
|
594
|
-
this the harness where the rule matters most, not least.
|
|
595
|
-
|
|
596
|
-
## Model routing
|
|
597
|
-
|
|
598
|
-
The tier names in `agy-delegate` map to models that go stale — the built-in
|
|
599
|
-
`flash` still points at Gemini 3.7 while 3.8 ships. Re-check against
|
|
600
|
-
`agy models` after every upgrade; the id carries both the version and the
|
|
601
|
-
effort suffix.
|
|
602
|
-
|
|
603
|
-
| Work | Model |
|
|
604
|
-
|---|---|
|
|
605
|
-
| Media, fast or mechanical coding, boilerplate | `Gemini 3.8 Flash (Medium)` |
|
|
606
|
-
| Trivial one-liners | `Gemini 3.8 Flash (Low)` |
|
|
607
|
-
| Review, architecture, hard reasoning | `Claude Sonnet 4.6 (Thinking)` |
|
|
608
|
-
|
|
609
|
-
Set these once via `CLAUDE_PLUGIN_OPTION_TIER_{FLASH,FLASH_LO,PRO}` in
|
|
610
|
-
`~/.zshenv` — **not** `~/.zshrc`, which non-interactive tool shells never
|
|
611
|
-
source.
|
|
612
|
-
|
|
613
|
-
Adversarial review is the case that most repays the stronger model: a Flash
|
|
614
|
-
tier tends to agree with what it is shown, which is the one thing a reviewer
|
|
615
|
-
must not do.
|
|
616
|
-
|
|
617
|
-
## Headless timeouts
|
|
618
|
-
|
|
619
|
-
A trivial headless `agy` prompt was measured at **605s** in 2026-09.
|
|
620
|
-
`agy-delegate` defaults to `--print-timeout 5m`, so it aborts at 300s and
|
|
621
|
-
reports an empty body while the answer is still arriving. Pass `--timeout 15m`
|
|
622
|
-
for anything non-trivial. A short timeout does not read as "slow", it reads as
|
|
623
|
-
"broken".
|
package/README.de.md
CHANGED
|
@@ -164,7 +164,7 @@ Nachfolgend finden Sie die vollständige Übersicht aller in diesem Paket enthal
|
|
|
164
164
|
#### 🎨 Frontend & UI/UX
|
|
165
165
|
| Skill-Name | Beschreibung |
|
|
166
166
|
|------------|-------------|
|
|
167
|
-
| `
|
|
167
|
+
| `mcp-manage` | Manages the BDB specialized MCP servers including Unreal Engine, Rhino 7/8, DaVinci Resolve, grandMA3, Resolume, GitHub, Chrome DevTools, and TouchDesigner. |
|
|
168
168
|
| `api-design-principles` | Master REST and GraphQL API design principles to build intuitive, scalable, and maintainable APIs that delight developers and stand the test of time. |
|
|
169
169
|
| `api-patterns` | API design principles and decision-making. REST vs GraphQL vs tRPC selection, response formats, versioning, pagination. |
|
|
170
170
|
| `database-design` | Database design principles and decision-making. Schema design, indexing strategy, ORM selection, serverless databases. |
|
package/README.md
CHANGED
|
@@ -377,7 +377,7 @@ While heavily optimized for the creative tech industry, these skills are deeply
|
|
|
377
377
|
|
|
378
378
|
| Skill | Description |
|
|
379
379
|
|-------|-------------|
|
|
380
|
-
| `
|
|
380
|
+
| `mcp-manage` | Manages the BDB specialized MCP servers including Unreal Engine, Rhino 7/8, DaVinci Resolve, grandMA3, Resolume, GitHub, Chrome DevTools, and TouchDesigner. |
|
|
381
381
|
| `n8n-code-javascript` | Write JavaScript code in n8n Code nodes. `$input`/`$json`/`$node` syntax, HTTP requests, DateTime handling. |
|
|
382
382
|
| `n8n-code-python` | Write Python code in n8n Code nodes. `_input`/`_json`/`_node` syntax and standard library. |
|
|
383
383
|
| `n8n-expression-syntax` | Validate n8n expression syntax and fix common errors. `{{}}` syntax, `$json`/`$node` variables. |
|
package/README.pt.md
CHANGED
|
@@ -167,7 +167,7 @@ Abaixo está a visão geral completa de todas as skills curadas de agentes inclu
|
|
|
167
167
|
#### 🎨 Frontend & UI/UX
|
|
168
168
|
| Nome da Skill | Descrição |
|
|
169
169
|
|------------|-------------|
|
|
170
|
-
| `
|
|
170
|
+
| `mcp-manage` | Gerencia os servidores MCP especializados da BDB incluindo Unreal Engine, Rhino 7/8, DaVinci Resolve, grandMA3, Resolume, GitHub, Chrome DevTools e TouchDesigner. |
|
|
171
171
|
| `api-design-principles` | Domine princípios de design de API REST e GraphQL para construir APIs intuitivas, escaláveis e de fácil manutenção que encantam os desenvolvedores e resistem ao teste do tempo. |
|
|
172
172
|
| `api-patterns` | Princípios de design de API e tomada de decisões. Seleção entre REST vs GraphQL vs tRPC, formatos de resposta, versionamento, paginação. |
|
|
173
173
|
| `database-design` | Princípios de design de banco de dados e tomada de decisões. Design de esquema, estratégia de indexação, seleção de ORM, bancos de dados serverless. |
|