Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/open-software-network/os-clovy/collaborationgit clone --depth 1 https://github.com/open-software-network/os-clovyWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00000 | $0.00852 |
| Opus 5 | $0.00000 | $0.00426 |
| Sonnet 5 | $0.00000 | $0.00170 |
| Haiku 4.5 | $0.00000 | $0.00085 |
Grade A, and why
collaboration scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 64 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Agent collaboration: build, delegate, review
How coding agents (Claude Code, Codex, or any future harness) collaborate on this repo: who orchestrates, who implements, who reviews, and what each layer may touch. This is the map; the skills own the mechanics — never restate a skill's commands here or anywhere else (single owner per fact).
The skill family
repo-build-pr is the entry point. It orchestrates and defers to owners:
| Skill | Altitude | Git contract |
|---|---|---|
repo-build-pr |
the whole workflow: intake, tracker lifecycle, plan, chunking, validation, review, publish | commits and publishes |
repo-review |
one review axis per run (Standards / Spec / Adversarial), any harness | read-only |
repo-delegate |
one scoped implementation brief on another harness | edits + gate, never commits |
repo-orchestrate |
the whole workflow handed to another harness | commits; push/PR only with --publish |
agent-e2e-qa |
live QA process: surfaces, recording, upload, evidence format | read-only on code |
browser-test-tauri-fe |
browser-surface technique (fake Tauri IPC bridge, Playwright/CDP) | read-only on code |
Entry points by situation:
- Build a task in this session:
/repo-build-pr <task>. - Build with the other harness implementing per chunk:
/repo-build-pr <task> with codex(orwith claudefrom Codex). - Hand the entire build to the other harness:
repo-orchestrate. - Review a branch/PR/diff:
repo-review(sized per its Sizing section). - Ship one precise fix prescription to an implementer:
repo-delegate.
Principles
- The reviewer is never the author (default). The adversarial review
axis runs on a harness that did not write the diff. This is
evidence-based: the two harnesses' finding sets are measurably disjoint
(
.agents/skills/repo-review/CALIBRATION.md). Deliberate exception: a cross-harness implementer build (with codex/with claude) runs implementation and all review axes on the implementer harness by convention — self-review accepted; the counterweights are regression-gated fixture tests per fix and the orchestrating session's own verification. (PR #615: Codex adversarial rounds still surfaced two real defects in largely Codex-written code.) - Judgment stays with the orchestrator. Plans, contracts, triage, go/no-go never delegate. A delegate's report is a claim; the diff and real gate output are the evidence.
- Trust is two-level and stated honestly. Codex-side runners get OS
sandboxes (
read-only/workspace-write); Claude-side runners are policy-level (plan mode / acceptEdits + allowlists + detection guards). Policy-level runners are for branches authored in-session only — never unvetted third-party diffs. Escalations (publish, untracked files) are explicit flags, never defaults. - Docs are load-bearing. Plans are grilled against CONTEXT.md, ADRs,
and
spec/before the user is asked anything; doc contradictions found while building become doc fixes in the same PR. - The battery calibrates itself. Every review cycle ends by appending reviewer precision rows to CALIBRATION.md and folding new lessons into the skill that missed them.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 64 lines · 0 tokens per session scan A 0ad2f498b901
collaboration is an agent published in the GitHub repository open-software-network/os-clovy (357 stars, last pushed 8d ago), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 852 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other agents, from other repositories
ai-systems-architect
Use as the cross-cutting AI/systems DESIGN authority to decide WHERE a seam goes across the Rust core, Angular FE, and brain — provider/model abstraction, agentic tool-use loop bounds, tool/ACI design, the egress-consent-redaction-ledger firewall, routing, context assembly, a new ingest source or consumption surface.…
demo-agents
We need a steady flow of demo videos — X/Twitter, Reddit, YouTube Shorts, blog posts, App Store-style screen recordings — and we'd like to scale that without scheduling Daniel to record every clip by hand. This doc maps the content-creation agent landscape, names the gap between "synthetic video hype" and "scripted…
qa-agents
Every quarter someone pitches an "AI does QA" tool. Most are web-first or mobile-first. MacParakeet is a menu-bar macOS app with a non-activating KeylessPanel overlay, global dictation hotkeys, and TCC-gated microphone/screen-recording flows. The general AI-QA frontier doesn't speak our shape yet. This doc tracks…
flow-verifier
Runs the VZT Flow end-to-end verification ladder (build, tests, TTS-transcribe checks, clean-test latency, paste-test, daemon socket checks, overlay states) and reports real measured numbers — never estimates. Use before claiming a change works, before a release, or when asked to verify VZT Flow.
flow-feature-builder
Implements VZT Flow features observing this repo's boundaries — additive edits, the gotcha list from CLAUDE.md, pull-rebase discipline in a shared worktree, never touching the user's running daily-driver app without relaunching it. Use for routine feature/bugfix work in this repo.
adversarial-verifier
Anti-false-positive verifier for Murmur. Use AFTER a change is written, to TRY TO BREAK it — run the real gates (cargo test --lib / ng build / ng lint), live-reproduce in a browser against the zoneless Angular app with a mocked Tauri IPC, and hunt for the failure modes this app has actually shipped (sealed-content…