Skip to content

README.md

Latest commit

 

History

History
193 lines (160 loc) · 13.4 KB

README.md

File metadata and controls

193 lines (160 loc) · 13.4 KB
Raw

omp-dev-team — an Oh-My-Pi marketplace

🌐 English · Français

Six independent Oh-My-Pi (OMP) plugins. Install any subset — they share nothing. A global installer sets up OMP and walks you through them.

Plugin What it does
dev-team Agentic dev team — a verbatim port of bdfinst/agentic-dev-team v10.20.0 (Bryan Finster): orchestrator + 45 specialist/critic agents, 93 skills including /ship (spec → plan → build → review → PR as one idempotent command), a 58-file knowledge corpus, and upstream's 38 Python hook invocations run unmodified. 456 of 567 upstream files land byte-identical; the rest carry only the Claude Code → OMP format conversion, applied by scripts/port-upstream-dev-team.mjs so the next upstream bump is one command. Model-open: agents route through OMP roles (@smol/@plan/@slow), never a vendor id. Requires python3.
copilot-preset GitHub Copilot model preset — route OMP (and the dev-team tiers) through github-copilot to run on a Copilot license. Config-only: tier→model mapping, post-June-2026 AI-credit pricing comparison, and MAI-Code-1-Flash wired in.
token-diet Context compression via lean-ctx (MIT) — bash, read, grep, find and ls are routed through a local Rust binary that compresses command output and file reads, search results and project context, recognises 75+ tools, and caches per session so an unchanged re-read is nearly free. Plus the caveman skill for terse agent output, the one axis nothing else addresses. Replaces the retired ctx-wire filter pack.
azure-devops-fs Azure DevOps as a filesystem — read repos/files/PRs/diffs via ado:// URIs (paginated), PR gates/policies + CI (builds/logs/run), create/checkout/push/complete PRs, comment/vote. Backed by the Azure CLI (az + the azure-devops extension), PAT auth, SQLite read cache; works behind corporate TLS proxies.
openai-compatible Any OpenAI-compatible provider — point it at a LiteLLM, Ollama, vLLM, or LocalAI endpoint (name + URL + API key); the installer lists the models and writes the provider into ~/.omp/agent/models.yml with runtime discovery, usable as <name>/<model-id>. API key stored chmod 600, never in env.
datadog Datadog observability from the terminal — via the Datadog pup CLI (logs, metrics, traces/APM, monitors, incidents, dashboards, SLOs, RUM, security/audit, CI test visibility, LLM observability). One broad datadog skill drives pup; installer sets up pup + auth.

Quick start (recommended)

The global installer installs OMP, registers this marketplace, and interactively offers each plugin + its config, fixing your PATH at the end.

git clone https://github.com/outofrange-consulting/omp-dev-team
cd omp-dev-team
bash install.sh                 # Linux/macOS   (-y for non-interactive)
#   pwsh -File install.ps1      # Windows

Works out of the box / defaults: the global installer reinstalls the selected plugins to the latest and brings runtimes/OMP/tooling up to date by default (pass --no-update / -NoUpdate to keep tools already installed). It then merges the managed model-roles + skills defaults into ~/.omp/agent/config.yml and the team MCP servers into ~/.omp/agent/mcp.json — anything you've already set is preserved (no clobber, no backup needed). With copilot-preset the tiers wire, via GitHub Copilot: smol/task → Haiku, default/plan → Sonnet 5 (runs the dev-team orchestrator + architecture/domain design — non-trivial work goes research → plan → implement → review), slow → Opus (high-stakes security verdicts); without it, the same tiers on Anthropic ids. token-diet's lean-ctx routing and the skills are enabled too. --no-config leaves your config + mcp.json untouched.

Three official remote MCP servers are merged into ~/.omp/agent/mcp.json, all type: http:

server endpoint auth
github https://api.githubcopilot.com/mcp/ PAT (enabled once one is provided / $GITHUB_TOKEN is set)
atlassian https://mcp.atlassian.com/v1/mcp/authv2 OAuth 2.1, on first tool call — nothing stored, enabled by default
context7 https://mcp.context7.com/mcp CONTEXT7_API_KEY header (free tier); registered disabled without one

Atlassian and Context7 used to ship here as CLI wrappers (acli, ctx7) on the argument that an MCP server's tool schemas inline into every system prompt. OMP 17 retired that argument: tools.xdev (default on) mounts MCP tools as xd:// devices behind a doc budget, so a server costs a mount, not a schema dump. Both are back on their official endpoints and the wrappers are gone.

Lean startup context (out of the box). Every enabled tool's JSON schema is loaded into the system prompt on every request, so the generated config drops what the team never calls:

  • dev-team turns off the DAP debug tool and the Python/JS eval tool — no dev-team agent or command uses them (they go through bash + the systematic-debugging skill).

Correction (OMP 17). Earlier versions of this README advertised a tools.discoveryMode: all + tools.essentialOverride trim as the headline win ("~29K → ~20K, −31%"). Those settings were removed in OMP 17.0.0 and are now deleted from your config on load, so the trim no longer exists — and the installers no longer write them. OMP's replacement needs no configuration: tools.xdev (default on) mounts non-essential and MCP tools as xd:// devices behind a 48k/10k-char doc budget.

Corporate proxies (Zscaler / Trend Micro under WSL): if a TLS-intercepting proxy breaks certificate checks, the UNIX installers give you two options:

  • Preferred — trust the corporate CA: --ca-file=/path/to/corp-root-ca.pem (or OMP_CA_FILE=…). Verification stays on; node/bun, git, curl/wget, Python and Go tools (Ollama model pulls) are pointed at your CA via NODE_EXTRA_CA_CERTS/SSL_CERT_FILE/CURL_CA_BUNDLE/GIT_SSL_CAINFO, and it's persisted to your shell profile so omp and ollama pull trust it later too. On WSL you don't even need the .pem — --ca-from-windows exports the Windows trust store (incl. the corporate roots) automatically, and the global installer offers it when it detects WSL. To instead install the corporate CA into WSL's system trust store (so curl/git/node trust it natively, no env vars), run scripts/wsl-trust-zscaler.ps1 from Windows PowerShell — it uses wsl --user root (no sudo) and update-ca-certificates.
  • Escape hatch — bypass: --insecure-tls (or OMP_INSECURE_TLS=1) disables verification for that run (curl/wget incl. piped installers, git, node/bun/npm). It can't bypass Go/libcurl tools (Ollama, etc.) — use --ca-file for those.

The global installer propagates either choice to the plugin installers. Run the installers without sudo (everything is per-user: ~/.bun, ~/.local/bin, ~/.omp).

Manual install

omp plugin marketplace add outofrange-consulting/omp-dev-team   # or:  add ./
omp plugin install dev-team@omp-dev-team
omp plugin install copilot-preset@omp-dev-team
omp plugin install token-diet@omp-dev-team
omp plugin install azure-devops-fs@omp-dev-team
omp plugin install openai-compatible@omp-dev-team
omp plugin install datadog@omp-dev-team

Important — extension modules. OMP does not load extension modules (a plugin's package.json omp.extensions) from a marketplace cache install — only skills/commands/agents/rules/MCP surface that way. The plugins whose core behavior is an extension — azure-devops-fs (the ado tool), dev-team (the blocking guards), openai-compatible (the provider), token-diet (read-dedup/context-dedup/context-compress/cache-meter

  • path-inject, which puts ~/.local/bin on OMP's own PATH), and datadog (path-inject, same reason — pup lands in ~/.local/bin) — therefore need their installer to run too. The global install.sh and each plugin's install.sh/install.ps1 mirror those modules into OMP's native ~/.omp/agent/extensions/<plugin>/ dir (always discovered, survives config resets), so the ado tool / guards / provider / PATH fix actually load. A bare omp plugin install <name>@omp-dev-team alone will show the skill but not register the extension.

Each plugin ships its own install.sh + install.ps1 (installs that plugin's tools at their latest versions) — see its README:

  • dev-team → bash plugins/dev-team/install.sh --apply-config (prereq check + config). All-cloud; no local model backend. C# navigation uses OMP's native lsp tool, which auto-detects omnisharp.
  • copilot-preset → bash plugins/copilot-preset/install.sh --apply-config, then omp → /login → GitHub Copilot.
  • token-diet → bash plugins/token-diet/install.sh (installs lean-ctx + its OMP extension).
  • azure-devops-fs → bash plugins/azure-devops-fs/install.sh (installs the Azure CLI + azure-devops extension, prompts for org/project/PAT, runs az devops login), then restart omp so the ado tool loads.
  • openai-compatible → bash plugins/openai-compatible/install.sh --name=litellm --url=http://localhost:4000 --api-key=… (lists the endpoint's models, writes the provider to ~/.omp/agent/models.yml), then restart omp.
  • datadog → bash plugins/datadog/install.sh (installs the Datadog pup CLI + sets up auth; --with-skills to also add pup's domain skills).

Layout

install.sh · install.ps1               # global installer (OMP + marketplace + per-plugin prompts)
.claude-plugin/marketplace.json        # catalog (pluginRoot ./plugins)
plugins/
  dev-team/         agents/ skills/ commands/ rules/ extensions/ .mcp.json
                    config.snippet.yml · install.sh · install.ps1
                    skills/dev-team-knowledge/  (registries, rubrics, pricing)
  copilot-preset/   config.snippet.yml · pricing.md · skills/ · install.{sh,ps1}
  token-diet/       .mcp.json · rules/ · skills/ · install.{sh,ps1}
  azure-devops-fs/  extensions/ (ado.ts + lib/az.ts) commands/ skills/ rules/ knowledge/ · install.{sh,ps1}
  openai-compatible/  extensions/ (openai-provider.ts) · skills/ · install.{sh,ps1}
  datadog/          extensions/ (path-inject.ts) · skills/ (umbrella) · install.{sh,ps1}

Each plugin's extensions load from its own package.json omp.extensions; the guard extensions resolve their data relative to the plugin (so they work whatever the consuming project's cwd is). Runtime state is written under the consuming project's .omp/state/.

Agent Package Manager (APM) & duplicate definitions

If you use Agent Package Manager and run apm compile --all, the same agents/skills/rules get written into .claude, .copilot, .cursor, .agents, … next to .omp. OMP does not load them multiple times — it de-duplicates by name (first match wins) before loading, so duplicates cost no extra tokens and are not double-registered:

  • agents / commands / skills are scanned only from .omp › .claude › .codex › .gemini (project before user); identical skill files are additionally realpath-deduped. .copilot and .cursor are not scanned for these.
  • rules are name-deduped across providers native › agents › cursor › windsurf › cline; shadowed same-name rules are excluded from the active set.

Notes:

  • Don't delete the other directories. apm compile --all creates .claude / .copilot / .cursor on purpose for Claude Code / Copilot / Cursor, which need them. Removing them to "de-dupe" would break those tools — and OMP already ignores the extras.
  • OMP silently uses the highest-precedence copy and shadows the rest. To be sure OMP uses a specific variant, keep the canonical one in .omp/ (or .claude/). For skills you can also pin/exclude with skills.includeSkills / skills.ignoredSkills in your config.

Tested

Verified end-to-end and in CI (Linux/macOS/Windows — see .github/workflows/installers.yml): all install.sh pass bash -n; all install.ps1 parse under PowerShell 7; all manifests are valid JSON; the 8 dev-team extensions (plus the token-diet, azure-devops-fs, openai-compatible, and datadog extension modules) compile under bun and pass tsc --noEmit against OMP's published types; lean-ctx and OMP install via the exact commands the scripts use; and all six plugins install through real OMP on each OS (omp plugin marketplace add ./ → omp plugin install <name>@omp-dev-team).

Credits