A pair of self-contained Claude Code plugins: conductor-kernel supplies 19 domain-agnostic agents and 14 skills; conductor-dev adds 18 development agents and the /conduct command. Together they provide tiered quality gates, 5-signal classification, BRD-driven development, capability-based handoffs, intent engineering, hooks and schemas.
Installs as a Claude Code plugin. The repositories are currently private, so marketplace add cannot resolve until they are published:
/plugin marketplace add bulletproofsoftware-ai/bulletproof-conductor-kernel /plugin install conductor-kernel@conductor-kernel /plugin list
bulletproof-conductor-kernel/docs/media/Complex software projects need more than a single AI agent. Individual agents optimize for their narrow scope while the overall system drifts — no shared state, no quality gates, no traceability back to requirements. You need orchestrated agents that enforce standards at every checkpoint.
But standalone agent files are fragile. They can be accidentally moved, archived, or broken by directory changes. There is no lifecycle management, no hook system, no bundled state tracking. A plugin solves this: agents, hooks, skills, schemas, and commands are packaged as a self-contained unit. Install the plugin, get the entire orchestration system.
Single agent does everything. No quality gates, no traceability. Scope creep goes undetected. Security review is optional at best.
Separate .md files scattered in a directory. No hooks, no state persistence, no lifecycle management. Easily broken by file moves or archives.
Self-contained unit. Hooks for lifecycle events. Built-in state tracking. Bundled skills and schemas. Single install = complete system.
Each signal is scored 1–4, then combined into a single weighted score that selects the tier.
File count affected: 1 file = 1, 2–5 = 2, 6+ = 3, new repo = 4.
Typo/config/bugfix = 1, enhancement = 2, new feature = 3, greenfield = 4.
Easily reversible = 1, reversible with effort = 2, hard to reverse = 3, irreversible = 4.
Crystal clear = 1, mostly clear = 2, needs discovery = 3, significant unknowns = 4.
No intent overlap = 1, touches objectives = 2, touches trade-offs = 3, touches hard limits = 4. A score of 4 forces STANDARD tier or above.
(scope × 0.25) + (type × 0.20) + (risk × 0.20) + (ambiguity × 0.15) + (intent_sensitivity × 0.20) — the result maps directly onto the four tier bands above. The tier may be promoted mid-workflow as new information arrives, but never demoted automatically.
No code is written without a BRD ID. The state file conductor-state.json tracks the BRD ↔ spec ↔ implementation chain. QA closes the loop by verifying each BRD ID has a passing test. Drift between BRD and code is impossible to hide because the conductor's state file makes it visible.
Split across the two plugins: conductor-kernel ships 19 domain-agnostic agents, conductor-dev ships 18 development agents. Names below are short forms; kernel files are <name>.md and dev files are conductor-<name>.md.
All 14 ship in conductor-kernel; conductor-dev ships agents and the /conduct command, and inherits these skills from the kernel.
conductor-state.json lifecycle, validation, recovery from corruption.
BRD ID extraction, spec linkage, gap analysis between requirements and implementation.
Context budget thresholds, spec-per-session limits, PreCompact coordination and handoff.
SOP retrieval from Qdrant for non-deterministic procedures.
Capability-based handoff schema; agents declare what they need + what they produce.
The capability matrix the director consults to validate every agent assignment.
Canonical per-tier workflow templates the director consults at every checkpoint.
Structured behavior reflection — agents reflect on outputs and revise before handoff.
Bounded retry and escalation rules applied when an agent dispatch fails.
Failure classification and automated recovery playbooks.
Event taxonomy and routing rules that connect workflow events to handlers.
Completion, rework and quality metrics collected across finished workflows.
Workload forecasting used to pre-select model tiers and resource budgets.
Surfaces live workflow state and metrics to external dashboards.
Capability-based handoffs replace named-agent calls. The director declares what capability it needs (e.g., "spec-decomposer", "security-reviewer"), not which agent. The plugin's manifest registry resolves the capability to a current implementation. This means agents can be replaced or renamed without breaking the director.
Build a Claude Code multi-agent orchestration plugin:
1. PLUGIN STRUCTURE (two self-contained plugins — see §3):
conductor-kernel (domain-agnostic):
- agents/ (19 kernel agents)
- skills/ (14 skills — state-management, brd-tracking,
context-management, process-knowledge, agent-interop,
agent-capabilities, workflow-reference, sbr, retry-policy,
self-healing, event-automation, outcome-measurement,
predictive-scaling, dashboard-integration)
- hooks/ (SessionStart + PostToolUse with schema validation)
- schemas/ (conductor-state.json schema)
conductor-dev (development domain, built on the kernel):
- agents/ (18 dev agents)
- commands/ (/conduct slash command)
- hooks/ (mirrors of the kernel hooks, own PLUGIN_ROOT)
2. 4-TIER QUALITY GATES (by weighted score):
- TRIVIAL : 1.0–1.5 typo, config, single-file bugfix
- MINOR : 1.6–2.3 small enhancement, validation addition
- STANDARD : 2.4–3.2 new feature, multi-file change, API additions
- MAJOR : 3.3–4.0 greenfield, architectural change, new service
3. 5-SIGNAL CLASSIFICATION (scored 1–4, weighted):
- scope 0.25
- type 0.20
- risk 0.20
- ambiguity 0.15
- intent sensitivity 0.20 (a 4 forces STANDARD or above)
4. BRD-DRIVEN: every change traces to a BRD ID. State file
conductor-state.json tracks BRD ↔ spec ↔ impl ↔ test.
5. CAPABILITY-BASED HANDOFFS: director dispatches by capability,
not by agent name. Manifest registry resolves capabilities.
6. HOOKS:
- SessionStart → detect conductor-state.json, surface status
- PostToolUse (Write|Edit) → validate state file, enforce gates
(block phase transitions on STANDARD/MAJOR)
7. RECOVERY ENGINE: classify failures (transient/model/data/permission/
logic/infrastructure/external) and apply playbook retries before escalating.
8. ADVERSARIAL REVIEW: Gemini second-opinion at every checkpoint for
STANDARD+. Dual-AI code review with consensus engine.
9. INTEGRATION: writes trajectories + learnings + task outcomes
to the memory system. Reads governance policy decisions.
Bundling makes the orchestration system installable, versionable, and resistant to accidental file moves. The plugin manifest is the contract.
Names couple the director to specific implementations. Capabilities decouple — agents can be replaced, renamed, or upgraded without breaking the director.
A one-size-fits-all workflow is either too heavy for trivial changes or too light for major ones. Tiered gates match ceremony to risk.
Every change traces back to a requirement. Drift between requirements and code is impossible to hide once the state file records the chain.
Trivial changes shouldn't be blocked on every nit, so their gates are skipped entirely and MINOR gates only advise. MAJOR changes carry enough risk that every gate blocks; STANDARD blocks at the pre-release and post-pentest gates only.
Single-model bias is real. Adversarial review with a different model surfaces issues a same-model reviewer would miss.
The conductor ecosystem is split across two installable plugins, each with its own hooks.json — both auto-discovered by Claude Code when installed under ~/.claude/plugins/local/. The hooks share the same script names but run from different CLAUDE_PLUGIN_ROOT values, so they execute independently and never race.
| Plugin | Event | Matcher | Script | Purpose |
|---|---|---|---|---|
| conductor-dev | SessionStart | * | session-start.sh | Detect conductor-state.json in CWD, surface active-workflow status as a systemMessage. Silent if no state file present. |
| conductor-dev | PostToolUse | Write|Edit | post-state-write.sh | When Write/Edit targets conductor-state.json: validate against bundled JSON schema AND enforce phase transition gates. Returns exit-code 1 to block illegal transitions for STANDARD/MAJOR tier workflows. Advisory (exit 0) for trivial-tier. |
| conductor-kernel | SessionStart | * | session-start.sh | Mirror of conductor-dev SessionStart for the kernel's own state surface. Different PLUGIN_ROOT, same script content. |
| conductor-kernel | PostToolUse | Write|Edit | post-state-write.sh | Mirror of conductor-dev PostToolUse for the kernel's state schema and gates. |
conductor-dev ships the developer-facing agents (architect, builder, qa, devops, etc.); conductor-kernel ships the system primitives (state management, event router, recovery engine, compliance, retrospective). They are deployed independently so the kernel can iterate without churning the dev pack. The hooks are duplicated because each plugin needs to know about its own state files and schemas.
Conductor writes trajectories, learnings, and task outcomes to Qdrant via MCP tools. The director consults process-knowledge SOPs and heuristics at SessionStart. See the Memory System PRD.
The governance plugin's policy engine evaluates every conductor agent dispatch. Trust-mediated delegation enforces breadth/depth budgets. Audit events flow through the governance bus. See the Governance PRD.
Context guard's 60% budget rule and max-3-specs constraint feed into the conductor's context-management skill. PreCompact telemetry tells the conductor when to defer follow-up work. See the Context-Guard PRD.
Conductor is the canonical example of the plugin architecture: agents + skills + hooks + schemas + commands packaged as a single installable unit, deployed under ~/.claude/plugins/local/.