Skip to content
    ↑↓ select↵ openesc close
    English中文
    jabworks

    Condux

    v0.12.0Agent Orchestration
    @jabworks/condux

    OpenCode plugin for the condux workflow toolkit — bundles the condux skills and injects the specialist agents (coder, explorer, planner, researcher) plus an opt-in plan-review listener.

    GitHub stars

    0

    Monthly installs

    2,832

    1.1k in 7 days

    Composite scoreSCORE

    41.2

    Multi-signal model

    Last commit

    2 hours ago

    2026-08-20

    Install and configure

    opencode.json

    Writes to this project's opencode.json — applies to this repository only.

    opencode.json

    {
      "$schema": "https://opencode.ai/config.json",
      "plugin": ["@jabworks/condux@0.12.0"]
    }

    opencode loads npm dependencies through its embedded runtime on startup and caches them locally — no manual global install needed.

    CI condux plugin npm Trigger routing

    Personal collection of agentic coding skills. Compatible with Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and 40+ other tools via npx skills add.

    Contents

    Install

    npx skills add auto-detects the running agent and installs to the right directory:

    npx skills add jabworks/agentic-toolkit
    

    This works inside Claude Code, Codex, OpenCode, Cursor, and most other agentic tools — no flags needed.

    Ask an agent to install it

    The marketplace install is only half the job for condux, docket and concord — each finishes with a host-specific step no marketplace can perform. Every one of them ships an INSTALL.md written to be followed by an agent, so the shortest path is to hand the whole thing to the agent you already have open. Paste this:

    Install plugins from the jabworks/agentic-toolkit marketplace on this machine.
    
    1. Detect the host you are running in — Claude Code (~/.claude), Codex
       ($CODEX_HOME or ~/.codex), or OpenCode (~/.config/opencode).
    
    2. Add the marketplace and install the plugins I named. If I did not name any,
       ask me first — do not install all of them.
         Claude Code   /plugin marketplace add jabworks/agentic-toolkit
                       /plugin install <plugin>@jabworks-agentic-toolkit
         Codex         codex plugin marketplace add jabworks/agentic-toolkit
                       codex plugin add <plugin>@jabworks-agentic-toolkit
         OpenCode      add "@jabworks/condux" to the plugin array in opencode.json
    
    3. Three plugins need a step the marketplace cannot do. Read the INSTALL.md
       next to each one you installed, then run its installer:
         condux    <plugin-root>/install.mjs
         docket    <plugin-root>/skills/docket/record/server/install.sh
         concord   <plugin-root>/skills/concord/remember/references/install-codex-hook.sh
       Show me the --dry-run plan before writing anything.
    
    4. Verify rather than assume — run the doctor for each plugin and show me the
       report as it printed:
         /condux:condux-doctor    /docket:docket-doctor    /concord:concord-doctor
    
    5. Report one row per host: what you changed, what was already correct, and
       what any doctor flagged. If one warns that another installed skill library
       conflicts, show me its removal command but do not remove anything.
    

    Every step is safe to re-run, and each installer is read-modify-write on one key — none of them rewrites config another plugin owns.

    Compatibility with other skill libraries

    condux conflicts with obra/superpowers — the library it reworks. Both register a SessionStart hook on the same matcher, each injecting a router claiming every dev task, and 11 of superpowers' 14 skills overlap 8 of condux's. Run one or the other. install.mjs and /condux:condux-doctor both detect it and print the removal command without running it; the overlap table and reasoning are in the condux README, and the registry driving the check is skills/condux-doctor/conflicts.json.

    No other conflict is currently detected. The rest of the toolkit's skills — git, release, spec, session — occupy their own ground.

    Claude Code — plugin marketplace

    Alternatively, register as a plugin marketplace to install individual skills:

    Via CLI:

    claude plugin marketplace add jabworks/agentic-toolkit
    claude plugin install session-handoff@jabworks-agentic-toolkit # claude plugin install session-handoff
    claude plugin install toolkit-ops@jabworks-agentic-toolkit # claude plugin install toolkit-ops
    

    Via Claude Code CLI:

    /plugin marketplace add jabworks/agentic-toolkit
    /plugin install session-handoff@jabworks-agentic-toolkit # /plugin install session-handoff
    /plugin install toolkit-ops@jabworks-agentic-toolkit # /plugin install toolkit-ops
    

    Via Claude Code CLI plugin menu:

    /plugin
    -> Marketplaces tab -> Add Marketplace -> jabworks/agentic-toolkit
    -> Discover tab -> Search plugin name
    

    Codex — plugin manifests

    This repo is also Codex plugin compatible:

    Via CLI:

    codex plugin marketplace add jabworks/agentic-toolkit
    codex plugin add session-handoff@jabworks-agentic-toolkit
    codex plugin add toolkit-ops@jabworks-agentic-toolkit
    

    Via Codex CLI plugins menu:

    /plugins
    -> Add Marketplace -> jabworks/agentic-toolkit
    -> Select jabworks/agentic-toolkit tab to add plugins
    

    OpenCode — condux plugin + merged-trigger skill variants

    For the condux workflow, one plugin line is the whole install. The @jabworks/condux plugin bundles the 15 condux skills, injects the specialist agents (coder / explorer / planner / researcher), and wires an opt-in plan-review listener:

    // opencode.json
    {
      "plugin": ["@jabworks/condux"]
    }
    

    The bundled skills register themselves on config.skills.paths, so no separate skills install is needed for condux.

    For the rest of the toolkit skills (git-commit, session-handoff, release, spec-browser, coding-directive, …), install the OpenCode-facing variants. These fold each skill's when_to_use trigger conditions into its description — OpenCode surfaces only description when deciding which skill to load, so the plain skills/ tree would lose them:

    npx skills add https://github.com/jabworks/agentic-toolkit/tree/main/dist/opencode/skills -a opencode
    

    This copies skills into ~/.config/opencode/skills/. It also includes the condux skills, but the plugin already provides those — OpenCode dedupes by name, so installing both is harmless.

    Cursor — merged-trigger skill variants

    Cursor loads SKILL.md natively (since Cursor 2.4) but surfaces only description when deciding which skill to load — it never reads when_to_use. Install the Cursor-facing variants, which fold the trigger conditions in (same transform as the OpenCode tree):

    npx skills add https://github.com/jabworks/agentic-toolkit/tree/main/dist/cursor/skills -a cursor
    

    Run it in your project — skills land in .agents/skills/, which Cursor picks up per-project. Prefer project installs over --global: the CLI currently writes global skills to ~/.agents/skills/ against its own docs (vercel-labs/skills#421), and on a WSL setup that's the WSL home while Windows-side Cursor reads C:\Users\<you>\.agents\skills\ — the install silently vanishes.

    Every plugin is also a spec-conformant Agent Plugin (root plugin.json, flat skills/, docket MCP via spec mcp.json) — Cursor loads them directly. To try one without a marketplace: copy dist/plugins/<name> into ~/.cursor/plugins/local/<name> and fully restart Cursor. Plugin installs ship raw skill frontmatter, so the trigger-quality caveat below applies; the dist/cursor/skills/ install remains the best path for auto-triggering.

    Verified end-to-end 2026-08-14 (Cursor on Windows, WSL remote):

    Status
    Skills install, list, and invoke (/workflow etc.) ✅ works — project scope, from dist/cursor/skills/
    Plugins load as Agent Plugins (root plugin.json, flat skills/) ✅ works — verified via ~/.cursor/plugins/local
    Docket MCP server ✅ works — auto-imported from an existing Claude Code plugin install, or manually via .cursor/mcp.json (see docket INSTALL.md)
    Docket CLI fallback ✅ works as-is (dependency-free)
    Condux /workflow routing ⚠️ degrades — no SessionStart hook on Cursor, so routing relies on the skill descriptions instead of the injected routing rule
    plan-review auto-capture, named agents (explorer/researcher/planner/coder) ❌ absent — Cursor has no ExitPlanMode/Stop hook or custom-subagent surface

    Bonus: if Claude Code is installed on the same machine, Cursor picks up its plugin ecosystem by itself — already-installed plugins (skills and MCP servers, docket included) just appear, and registered marketplaces surface in Cursor's Customize panel with one-click Add. Marketplace installs ship the raw SKILL.md, whose when_to_use Cursor ignores — fine for MCP-bearing plugins like docket, but for reliable skill triggering prefer the dist/cursor/skills/ install above. Cursor reads the marketplace clone on its own side of a WSL split, which can lag — update the marketplace in Claude Code if the listing looks stale.

    Manual fallback

    If running outside an agent environment, clone and copy to your tool's skills directory:

    Tool Skills directory
    Claude Code ~/.claude/skills/
    Codex ~/.codex/skills/
    OpenCode ~/.config/opencode/skills/
    Cursor ~/.cursor/skills/
    Gemini CLI ~/.gemini/skills/
    Most others .agents/skills/
    git clone https://github.com/jabworks/agentic-toolkit /tmp/agentic-toolkit
    cp -r /tmp/agentic-toolkit/skills/<name> <skills-directory>/
    

    Skills

    Skill Description
    session-report Generate an explorable HTML report of session usage — tokens, cache, cost, subagents, skills
    session-handoff Preserve and restore session context across agentic coding sessions
    adapting-skills Stack and style priors for adapting skills to the jabworks conventions (opinionated — fork the priors for your own stack)
    git-commit Conventional-commit message from the diff, run safely — review before staging, no blind git add .
    git-operations Decision router for git — pick the right operation (undo/discard/stash/merge/push) with an undo path
    git-worktree Decision router for git worktrees — isolated workspaces, native-first, with cleanup and recovery paths
    spec-browser Catalog and browse a specs/ tree — markdown index + folder-grouped doc site
    release Cut releases safely — machinery detection (changesets/toolkit/GitHub), dry-run first, rollback paths
    coding-directive House style for @jabworks repos — TypeScript, React, formatting, imports, naming, CSS (personal)
    remember Continuous memory for Codex (concord plugin) — captures each session from the rollout, ages it into tiers, recalls it
    concord-doctor Health check for the concord plugin — probes its Codex hooks and memory store on this host
    condux-doctor Health check for the condux plugin — runs the SessionStart routing hook and checks the agents shipped
    docket-doctor Health check for the docket plugin — probes the MCP registration on every host and the CLI fallback

    Condux

    Lean agentic workflow plugin (condux). Install the full workflow bundle as a unit:

    /plugin install condux@jabworks-agentic-toolkit
    

    Inspired by obra/superpowers — condux reworks its skill-orchestration ideas around a lean, proportional-effort philosophy (tiered routing, lazy loading, soft gates) rather than always-on maximalism.

    Installing is the whole setup: condux ships a SessionStart hook (Claude Code and Codex both) that puts its routing rule in context before your first prompt, so /workflow is reached as the entry point instead of being inferred from the skill catalog. That inference sits at roughly 80% on its own, and the misses are condux's own siblings winning the query — root-cause-analysis on a crash report, draft-plan on "write the plan" — which no description can fix without stealing their trigger space. The payload is ~390 tokens per session and lives in skills/workflow/hooks/routing.md; delete or edit that file to change or disable it.

    Why condux

    Most workflow frameworks are maximal — every gate runs on every task, so a typo fix pays the same ceremony as a cross-cutting refactor. Condux inverts that with five operating rules:

    1. Proportional, not maximal — the process tier matches the task size.
    2. Lazy loading — downstream skills load only when their step is reached, keeping context (and cost) small.
    3. Soft gates, not hard walls — a gate can be skipped, but only consciously: the agent asks before bypassing, never silently.
    4. The user is in control — explicit instructions always beat the framework; "treat this as LARGE" or "skip the plan" is always valid.
    5. Implement yourself by default — subagents are the exception and need a concrete justification, never spawned to fill time.

    The payoff: no over-engineered button changes, no under-planned refactors, predictable checkpoints where you drive every transition, and a small context footprint because only the skills a tier actually needs ever load.

    The flow

    Every dev task enters through /workflow. It infers a tier from the task (file count, design clarity, boundary crossings), confirms it with you, then runs the matching pipeline:

    Tier Looks like Flow
    SMALL isolated change, 1–3 files, clear requirements implement → /preflight/finalize
    MEDIUM multi-file, some design needed, known boundaries inline plan → implement → /preflight/finalize
    LARGE cross-cutting, unclear scope, multiple subsystems /discovery/draft-plan → implement → /preflight/finalize
    flowchart TD
        T([dev task]) --> W{"/workflow — infer tier, confirm"}
        W -->|SMALL| S[implement]
        W -->|MEDIUM| M["inline plan"] --> M2[implement]
        W -->|LARGE| L["/discovery"] --> L2["/draft-plan"] --> L3[implement]
        S --> F["/preflight → /finalize"]
        M2 --> F
        L3 --> F
    

    Every tier ends with /preflight (an "am I actually done?" checklist) and /finalize (typecheck → lint → format → test, in order, once, stopping on the first failure) — quality checks run at the end, not scattered mid-implementation.

    Checkpoints (MEDIUM/LARGE only): at each phase boundary the agent stops and presents a menu — after the plan (start implementing / tests-first / spawn agents / revise), after implementation (verify & finalize / code review / keep building), and after everything is green (review / commit / release / done). The agent never auto-advances past a checkpoint; SMALL runs linear with no menus.

    Named agents: four specialists ship with the bundle — explorer (read-only codebase navigation), researcher (external API/library verification), planner (design → executable plan), and coder (executes a provided plan). Pipeline: explorer/researcher gather → planner plans → coder executes → finalize validates. The default is still to implement directly — agents are opt-in at checkpoints or justified by genuinely parallel work.

    The skills

    Skill Description
    Skill Description
    /workflow Tier router — infer Small/Medium/Large, confirm with user, load only the skills the tier needs
    /discovery Design gate — goal-round questions, alternatives, then a post-approach detail round (contracts, mappings, edge cases); sign-off saves the design doc and, default-on, a structured tech spec
    /blueprint Design-time visual clarity — grayscale HTML wireframes for UI surfaces, inline-SVG diagrams (data models, flows, architecture, state machines) for backend work; dependency-free, structure not styling
    /draft-plan Lean task-card plan (what/why/gotchas/deps), Markdown, LARGE tasks only
    /test-first-development Opt-in tests-first — one upfront consent, then red-green-refactor; asks before editing existing specs
    /subagent-execution Named specialist agents for LARGE plans, only when justified, never to fill time
    /subagent-deployment Fan out independent tasks across named agents in one message — ad-hoc, not a formal plan
    /finalize End-of-task quality gate — typecheck → lint → format → test, once, stop on first failure
    /live-verification Run the change and watch it work — drives the real UI or endpoint after finalize, light mode then dark, reports claim → evidence → verdict and names what it couldn't verify
    /code-review On-request diagnostic report (Critical/Important/Minor), never auto-triggers, never fixes
    /preflight "Am I actually done?" checklist before finalize — catches skipped steps and regressions, and drift-checks the implementation against the task's spec
    /root-cause-analysis Root-cause-first bug investigation — enforces the 4-phase sequence before any fix
    /technical-spec Scaffold and persist feature specs (decisions, API, fields, quirks) with a live HTML preview
    /plan-review Annotate a plan in a local browser with a categorized comment toolbar, then return approve/revise/deny to the agent — auto via a Claude Code ExitPlanMode hook or a Codex Stop hook, or manually. Self-contained, no egress

    Concord

    Continuous memory for Codex (concord) — the remember skill plus its doctor:

    /plugin install concord@jabworks-agentic-toolkit
    

    Three Codex hooks share one idempotent operation — sync the rollout forward from its recorded position — so capture is exactly-once regardless of hook order, and a hard-killed session is recovered at next start rather than lost. Memory ages through buffer, daily, recent, and archive tiers; explicitly pinned facts are never auto-compressed. references/install-codex-hook.sh wires the hooks; concord-doctor tells you whether they are still wired.

    Docket

    File-based project backlog plugin (docket) — two skills plus machinery:

    /plugin install docket@jabworks-agentic-toolkit
    

    Open items live in docket/DOCKET.md, closed items move to docket/archive/<year>.md with verification records, and ids are never reused — #N in a commit subject refers to the docket, not GitHub. The record skill captures and closes items; groom sweeps for stale and ghost work and recommends what to pick next. A dependency-free CLI (server/docket.mjs) keeps id allocation and archive moves byte-exact, a bundled MCP server exposes the same ops as tools (auto-registered on Claude Code; server/install.sh or the agent-followable server/INSTALL.md registers it for Codex/OpenCode), and docket.mjs browse renders a self-contained HTML board. Legacy root BACKLOG.md layouts are detected and work in place — migration is offered, never forced.

    Toolkit Ops

    Repo-maintenance bundle (toolkit-ops) for working on this toolkit itself:

    /plugin install toolkit-ops@jabworks-agentic-toolkit
    
    Skill Description
    Skill Description
    toolkit-orientation Zero-context map of the repo — trees, bundles, manifest pairing, docs trust order
    toolkit-foundry Create and maintain skills — scaffold, register, sync, publish (formerly standalone plugin-foundry)
    toolkit-change-control Classify a change, pick the version bump, gate on the publish checklist
    toolkit-skill-standards Frontmatter budgets, trigger contract, progressive disclosure, collision scan
    toolkit-debugging-playbook Symptom → discriminating command → root cause for skill/plugin problems
    toolkit-failure-archaeology Git-evidenced incident ledger — don't re-fight settled battles
    toolkit-plugin-reference Verified plugin.json / marketplace.json schema, Claude↔Codex divergences
    toolkit-research-frontier Open problems, assets, next steps, and the library-health campaign

    Where artifacts land

    Skills that write files into your project follow one rule: durable output goes to a normal project path; working state goes to .<plugin-name>/ at the git root, gitignored.

    <git-root>/
      specs/              # durable — tech specs, committed, browsable via spec-browser
      .condux/            # working state — designs/ plans/ progress/ scratch/
      .session-handoff/   # working state — handoff docs
      .session-report/    # working state — generated HTML usage reports
    

    The spec is what you keep; the design and plan are scaffolding it was built from. Directories are named for the owning plugin, not the skill and not the artifact — condux has 12 skills and one .condux/, so uninstalling condux tells you exactly which directory stops appearing.

    Nothing is written to your repo root or CWD, and nothing lands in docs/ — that belongs to your project, not to the toolkit. The first time a skill writes to a repo it checks git check-ignore and asks once before adding the entry to .gitignore (or .git/info/exclude if you'd rather not touch a tracked file). An AGENTS.md path override always wins.

    Evals

    The skills are eval-tested for trigger routing — given a user query, does the right skill activate (and does nothing activate when nothing should)? Latest run (2026-07-09, claude-haiku-4-5, 394 cold-trigger cases, 3 trials):

    • 88.6% ± 3.8pp mean routing accuracy (95% CI, t-distribution); per-run 87.1% / 90.1% / 88.6%, 37 flaky cases recorded
    • Per-skill breakdown and the full miss list live in the trial record

    This measures whether skills activate correctly — not end-to-end workflow quality. Structural invariants (dist mirror, manifests, frontmatter budgets, no-egress guarantees) are separately enforced by node --test in CI.

    Acknowledgments

    The plan-review skill is inspired by Plannotator — its interactive plan-review workflow served as the design reference. plan-review is an independent in-house reimplementation with no shared code and no third-party runtime dependency.

    condux reworks obra/superpowers' skill-orchestration ideas around proportional effort. Reworking a library that thoroughly makes the two direct alternatives rather than companions — see Compatibility before installing both.

    Structure

    skills/<name>/          # Editable source — also what `npx skills add` installs
    dist/plugins/<name>/    # Built plugin dirs — the plugin-marketplace install source
    dist/opencode/skills/   # OpenCode-facing variants (when_to_use folded into description)
    dist/cursor/skills/     # Cursor-facing variants (same fold, separate tree)
    packages/condux-opencode/ # OpenCode plugin (npm) — bundles condux skills + agents + plan-review listener
    .claude-plugin/
      marketplace.json      # Plugin registry (Claude Code / Codex marketplace channel)
    

    Four distribution channels read different trees:

    • npx skills add (vercel-labs/skills, 68+ agents) scans the top-level skills/<name>/SKILL.md layout — it installs straight from skills/ and ignores dist/.
    • /plugin install …@jabworks-agentic-toolkit (Claude Code / Codex native) reads .claude-plugin/marketplace.json, whose source paths point at the assembled plugin dirs under dist/.
    • OpenCode gets condux from the @jabworks/condux npm plugin, which bundles the condux skills (under packages/condux-opencode/skills/) and the agents; the rest of the toolkit installs from dist/opencode/skills/ (see the OpenCode install section). Both trees are generated by scripts/build-opencode.mjs, which scripts/sync.sh runs automatically.
    • Cursor installs from dist/cursor/skills/ (see the Cursor install section) — generated by scripts/build-cursor.mjs, which reuses the OpenCode fold transform; the trees are byte-identical today but kept separate so the channels can drift deliberately.

    dist/ is a build artifact mirrored from skills/ via scripts/sync.sh — never edit it by hand; node --test (run in CI) fails if it drifts from source. After cloning, run pnpm install (the test suite needs it) and bash scripts/install-hooks.sh, which installs a pre-commit hook that validates SKILL.md frontmatter, then syncs and stages dist/ automatically.

    SKILL.md frontmatter follows a deliberately narrow grammar — every line is key: value, values are plain when safe and double-quoted otherwise, and single quotes are banned. Four separate releases shipped frontmatter that strict YAML parsers reject (breaking the skill in Codex), so the grammar is enforced by node scripts/check-frontmatter.mjs (which also has --fix) and by a real strict parse in the test suite.