OpenCode Swarm

Your AI writes the code. Swarm proves it works.

Closing the trust gap between "the model said it's done" and "this actually works in production."

npm version License: MIT Tests

Website · Getting Started · Configuration · Architecture

OpenCode Swarm is a plugin for OpenCode that turns a single AI coding session into an architect-led team of specialized core, optional, and conditional agents. Run /swarm agents for the live roster; it is generated from the current plugin configuration. One agent writes the code. A different agent reviews it. Another writes and runs tests. Another checks security. Nothing ships until every required gate passes.

bunx opencode-swarm install

This single command installs the package, registers it as an OpenCode plugin, disables conflicting default agents, and creates a ready-to-edit config at ~/.config/opencode/opencode-swarm.json. Requires Bun (bun --version to check). If you must use npm: npm install -g opencode-swarm && opencode-swarm install.

First-run note: the installer registers the plugin, writes the global plugin config, creates a project override when missing, and disables the native explore and general agents in opencode.json. If you are not using a Swarm architect, the Swarm gates, reviewers, and test agents are bypassed. Open the OpenCode agent or mode picker and choose the Swarm architect when needed.

Why Swarm?

Most AI coding tools let one model write code and ask that same model whether the code is good. That misses too much. Swarm separates planning, implementation, review, testing, and documentation into specialized internal roles — and enforces gated execution so agents never mutate the codebase in parallel.

Key Features

  • 🏗️ Specialized core, optional, and conditional agents — architect, coder, reviewer, test_engineer, critic, explorer, sme, docs, designer, critic_oversight, critic_sounding_board, critic_drift_verifier, critic_hallucination_verifier, curator_init, curator_phase, council_generalist, council_skeptic, council_domain_expert. Run /swarm agents for the live roster — that is the source of truth, not this list.
  • 🔒 Gated pipeline — code never ships without reviewer + test engineer approval
  • 🔍 DEEP_DIVE Protocol — High-rigor, on-demand read-only codebase audit via specialized skills
  • 🔬 External Skill Curation Pipeline — Opt-in discovery, quarantine, evaluation, and promotion of external skill candidates from configured sources (disabled by default; enable via external_skills.curation_enabled: true in config). Includes 7 tools: external_skill_discover, external_skill_list, external_skill_inspect, external_skill_promote, external_skill_reject, external_skill_delete, external_skill_revoke. Candidates pass through a 3-gate validation pipeline before evaluation: prompt injection scan (12 regex patterns), unsafe instruction scan (25 patterns), and provenance integrity check (SHA-256, timestamp, URL, publisher, and hash verification).
  • 🔄 Phase completion gates — completion-verify and drift verifier gates enforced before phase completion
  • 🔁 Resumable sessions — all state saved to .swarm/; pick up any project any day
  • 🖥️ PR Monitor — GitHub PR subscription and background polling via gh CLI; delivers real-time CI, review, and merge status updates via the AutomationEventBus (FR-001, opt-in via pr_monitor.enabled: true). Subscribe with /swarm pr subscribe <pr-url|owner/repo#N|N>; unsubscribe with /swarm pr unsubscribe <pr-url|owner/repo#N|N>; check status with /swarm pr status. Enable auto_pr_feedback: true in pr_monitor config to inject [MODE: PR_FEEDBACK pr="URL"] on CI failures and merge conflicts automatically.
  • 🌐 13 full language profiles (TypeScript, JavaScript, Python, Go, Rust, Java, Kotlin, C/C++, C#, Ruby, Swift, Dart, PHP) with tree-sitter parse validation across 20 grammars (adds CSS, Bash, PowerShell, INI, Regex — and .tsx / .c aliases) — extending: see docs/adding-a-language.md
  • 🛡️ Built-in security — SAST, secrets scanning, dependency audit per task
  • 🔒 Scope enforcement — Validates write targets against declared scope with cross-process persistence, TTL expiry, and scope-aware destructive command blocking. Handles both single-string and array-based path arguments (files[], paths[], targetFiles[]) to prevent scope bypass via multi-file tool calls.
  • 📝 Shell write detection — Static analysis of POSIX/PowerShell/cmd commands to detect file writes (redirects, builtins, in-place editors, network downloads, archive extraction, git destructive ops) before execution
  • 🆓 Free tier — works with OpenCode Zen's free model roster
  • ⚙️ Fully configurable — override any agent's model, disable agents, tune guardrails

The Swarm architect coordinates all internal agents automatically. You never manually switch between internal roles. If the active OpenCode agent is not a Swarm architect, the plugin workflow is bypassed.

Shell Write Detection

Swarm includes comprehensive static analysis for shell commands to detect and intercept file write operations before execution.

Shell Write Detection Features

  • POSIX shell detection — Parses commands with bash-parser AST for accurate detection of:

    • Redirect operators (>, >>, >|, <<, <<-)
    • Here-documents and here-strings
    • Write-effect builtins (cp, mv, install, ln, truncate, dd)
    • In-place editors (sed -i, perl -i, awk -i)
    • Interpreter eval (python -c, node -e, bun -e, ruby -e, php -r)
    • Network downloaders (curl -o, wget -O, scp)
    • Archive extraction (tar -x, unzip, gunzip)
    • Git destructive operations (git clean -fd, git reset --hard)

  • Windows shell detection — Uses regex heuristics for PowerShell and cmd.exe:

    • PowerShell cmdlets: Out-File, Set-Content, Add-Content, Copy-Item, Move-Item
    • cmd.exe builtins: copy, move, ren, del, rd, md
    • Redirect operators (>, >>)

  • Interactive session denial — Blocks commands that create persistent or open-ended sessions:

    • POSIX: watch, screen, tmux new-session
    • PowerShell: Start-Process

  • Cross-process scope enforcement — Declared scope is persisted to .swarm/scopes/scope-{taskId}.json with:

    • TTL expiry (default 24 hours)
    • Symlink guards (O_NOFOLLOW + realpath containment)
    • Schema versioning and fail-closed validation
    • Scope-aware destructive command blocking — Recursive delete patterns (rm -rf, rmdir /s, del /s, Remove-Item -Recurse, rsync --delete) are blocked unless ALL target paths are within the declared scope (coder agents only)

Security Patterns

The guardrails system blocks destructive shell commands targeting:

  • System paths (/root, /etc, C:\Windows, etc.)
  • Symlink/junction creation with external targets
  • File operations under .swarm/ directory
  • Fork bombs and infinite loops
  • Disk wiping and ransomware-grade operations

What Actually Happens

You say:

Build me a JWT auth system.

Swarm then:

  1. Clarifies only what it cannot infer.
  2. Scans the codebase to understand what already exists.
  3. Consults domain experts when needed and caches the guidance.
  4. Writes a phased implementation plan.
  5. Sends that plan through a critic gate before coding starts.
  6. Executes one task at a time through the QA pipeline:
  • coder writes code
  • automated checks run
  • reviewer checks correctness
  • test engineer writes and runs tests
  • architect runs regression sweep
  • failures loop back with structured feedback
  1. After each phase, docs and retrospectives are updated.

All project state lives in .swarm/ — plans, evidence, context, knowledge, and telemetry. Resumable by design. If .swarm/ already exists, the architect goes straight into RESUME → EXECUTE instead of repeating discovery.

Execution Modes

Swarm has two independent mode systems:

Session modes — toggle per session with a slash command:

Mode Safety Speed When to Use
Balanced (default) High Medium Everyday development
Turbo Medium Fast Rapid iteration; skips Stage B gates for non-Tier-3 files
Lean Turbo High Fast Parallel lanes for non-conflicting tasks (up to max_parallel_coders coders)
Full-Auto Deterministic policy + critic oversight Fast Unattended multi-interaction runs

Auto-proceed — a session toggle (/swarm auto-proceed [on|off]) that skips the "Ready for Phase N+1?" prompt at phase boundaries. Defaults to false; set as a plan default via execution_profile.auto_proceed during QA GATE SELECTION. Session override always wins. Independent of Full-Auto.

Full-Auto reduces approval friction by deterministically allowing safe operations (read-only tools, in-scope writes, safe shell) and routing every ambiguous or high-risk action (writes to plugin/build/guardrail paths, network, dependency changes, plan/phase mutations, subagent delegation) through the read-only critic_oversight agent before it executes. Denials are returned to the agent as structured signals so it can choose a safer path; repeated denials pause the run; phase completion requires an APPROVED oversight record. See docs/modes.md for mode, permission_policy, denials, and oversight config keys, fail-closed semantics, and recovery from a paused run.

Project mode — persistent via execution_mode config key:

Value Effect
strict Maximum safety — adds slop-detector and incremental-verify hooks
balanced (default) Standard hooks
fast Skips compaction service — for short sessions under context pressure

Switch session modes with /swarm turbo [on|off] or /swarm full-auto [on|off]. Control phase-boundary auto-proceed with /swarm auto-proceed [on|off]. Set project mode in config. Lean Turbo is configured in turbo.lean.* in config and composes with all session modes. See docs/modes.md.

Quick Start

→ For a complete first-run walkthrough, see Getting Started.

The 15-minute guide covers:

  • Installation (bunx opencode-swarm install)
  • First-run auto-configuration (architect selected automatically)
  • Running your first task
  • Troubleshooting common issues

The installer automatically:

  • Creates a project config at .opencode/opencode-swarm.json when missing so project-level overrides have a place to live
  • Adds opencode-swarm to the OpenCode plugin list
  • Disables the native explore and general agents to reduce routing conflicts

30-Second Demo

No animated GIF is shipped in the repo — instead, here is the exact terminal session you can record yourself with asciinema rec demo.cast (or any screen recorder). Every command below is real and runs against this repo as published.

Recording script (copy/paste-able, ~30 seconds):

# 1. Install the plugin (5s)
bunx opencode-swarm install

# 2. Open opencode and select a Swarm architect if it is not already active
opencode

# 3. Inside the OpenCode session, verify Swarm is live (5s)
/swarm help
/swarm agents

# 4. Kick off a task — the architect plans, then gates fire automatically (15s)
Build me a JWT auth helper with tests.

# 5. Watch the gates land in real time (5s)
/swarm status
/swarm evidence

ASCII storyboard of what a viewer should see:

┌──────────────────────────────────────────────────────────────┐
│ $ bunx opencode-swarm install                                │
│ ✓ installed opencode-swarm                                   │
│ ✓ created .opencode/opencode-swarm.json                     │
│                                                              │
│ $ opencode                                                   │
│ [Swarm] Welcome! Architect auto-selected. Type /swarm help  │
│                                                              │
│ > /swarm help                                                │
│ Available commands: status, plan, agents, help, diagnose... │
│                                                              │
│ > Build me a JWT auth helper with tests.                     │
│ [architect]  PLAN → critic gate → APPROVED                   │
│ [coder]      task 1.1 implementing…                          │
│ [reviewer]   correctness OK                                  │
│ [test_eng.]  3 tests written, 3 pass                         │
│ [architect]  regression sweep clean → phase_complete         │
│                                                              │
│ > /swarm evidence                                            │
│ task 1.1: review ✓  tests ✓  sast ✓  secrets ✓  drift ✓      │
└──────────────────────────────────────────────────────────────┘

Each row corresponds to a real gate documented further down this README — none are simulated.

Upgrading

OpenCode caches plugins indefinitely. A normal OpenCode restart does not pull newer versions from npm — once a plugin is cached, OpenCode keeps using that exact copy on every subsequent launch (issue #675). The cache lives in several places depending on your platform and OpenCode version:

  • Current OpenCode cache layout: ~/.cache/opencode/node_modules/opencode-swarm/
  • Legacy Linux / devcontainers / GitHub Codespaces: ~/.config/opencode/node_modules/opencode-swarm/
  • Package cache layout: ~/.cache/opencode/packages/opencode-swarm@latest/
  • Platform-specific macOS / Windows cache roots: ~/Library/Caches/opencode/..., %LOCALAPPDATA%\opencode\..., or %APPDATA%\opencode\...

The updater also clears known OpenCode lock files (bun.lock, bun.lockb, and package-lock.json) so the next start resolves the latest package.

To upgrade to the latest published version (clears both layouts automatically):

bunx opencode-swarm update     # cache-only refresh, then restart opencode
# or
bunx opencode-swarm install    # full reinstall (re-asserts config), then restart opencode

/swarm diagnose shows the running version and, when available, the latest version on npm so you can tell at a glance whether your cache is stale.

To disable the background staleness check entirely, set version_check: false in your opencode-swarm.json.

Commands

Common subcommands at a glance:

/swarm help [command]      # List all commands or get detailed help for a specific command
/swarm status              # Current phase and task
/swarm show-plan [N]       # Full plan or filtered by phase
/swarm agents              # Registered agents and models
/swarm diagnose            # Health check
/swarm evidence [task]     # Test and review results
/swarm reset --confirm     # Clear swarm state

Use /swarm help to see all available commands categorized by function. Use /swarm help <command> for detailed usage information on a specific command.

Nine commands display a ⚠️ warning in help output because they share names with Claude Code built-in slash commands (e.g., /plan, /reset, /status). The warning reminds you to always use /swarm <command> — the bare CC command does something different and sometimes destructive. See docs/commands.md#claude-code-command-conflicts for the full conflict registry.

See docs/commands.md for the full reference. The live source of truth is src/commands/registry.ts, which includes canonical commands, compound commands, and deprecated aliases.

Command Aliases

Some commands are available under deprecated names for backwards compatibility. Using the canonical name is recommended:

Alias (deprecated) Canonical command
/swarm config-doctor /swarm config doctor
/swarm diagnosis /swarm diagnose
/swarm evidence-summary /swarm evidence summary
/swarm doctor /swarm config doctor
/swarm info /swarm status
/swarm list-agents /swarm agents
/swarm health /swarm diagnose
/swarm plan /swarm show-plan
/swarm close /swarm finalize
/swarm check /swarm preflight
/swarm clear /swarm reset-session

Aliases are hidden from help output but still function. The canonical command should be used in scripts and documentation.

The Agents

Swarm registers a roster of specialized core, optional, and conditional agents. The exact count shifts as agents are added or feature-flagged, so treat /swarm agents as the live source of truth — that command lists what is actually registered in your session. You don't manually switch between them — the architect coordinates automatically.

Agent Role Badge
architect Orchestrates workflow, writes plans, enforces gates Core
explorer Scans codebase, gathers context, maps facts Core
coder Implements one task at a time Core
reviewer Checks correctness and security Core
test_engineer Writes and runs tests, adversarial testing Core
critic Reviews plans before implementation begins Core
critic_oversight Sole quality gate in full-auto autonomous mode Core
sme Provides domain expertise guidance Core
docs Updates documentation to match implementation Core
designer Generates UI scaffolds and design tokens Conditional
critic_sounding_board Pre-escalation pushback to the architect Optional
critic_drift_verifier Verifies implementation matches spec Optional
critic_hallucination_verifier Verifies APIs and citations against real sources Optional
curator_init Consolidates prior knowledge at session start Optional
curator_phase Consolidates phase outcomes, detects workflow drift Optional
council_generalist Broad analytical voice in the General Council (uses reviewer model) Conditional
council_skeptic Adversarial stress-tester voice in the General Council (uses critic model) Conditional
council_domain_expert Technical-depth voice in the General Council (uses SME model) Conditional

Legend: Core = always available, Optional = available by default (can be disabled), Conditional = requires specific feature config (ui_review or council)

Run /swarm status and /swarm agents to see what's active.

graph TB
    subgraph Orchestration
        A[Architect]
    end

    subgraph Discovery
        E[Explorer]
    end

    subgraph Execution
        C[Coder]
        T[Test Engineer]
        D[Designer]
    end

    subgraph Quality
        R[Reviewer]
        CR[Critic]
        CO[Critic Oversight]
        CSB[Critic Sounding Board]
        CDV[Critic Drift Verifier]
        CHV[Critic Hallucination Verifier]
    end

    subgraph Support
        S[SME]
        DOC[Docs]
        CI[Curator Init]
        CP[Curator Phase]
    end

    subgraph Council
        CM[Council Member]
        CMO[Council Moderator]
    end

    A -->|Discover| E
    E -->|Context| A
    A -->|Consult| S
    S -->|Guidance| A
    A -->|Design| D
    D -->|Scaffold| C
    A -->|Implement| C
    C -->|Review| R
    R -->|Verify| A
    C -->|Test| T
    T -->|Results| A
    A -->|Deliberate| CM
    CM -->|Verdict| A

How It Compares

Feature Swarm oh-my-opencode get-shit-done
Multiple specialized agents ✅ Core + optional + conditional roster (/swarm agents) ❌ ❌
Plan reviewed before coding ✅ ❌ ❌
Every task reviewed + tested ✅ ❌ ❌
Different model for review vs. code ✅ ❌ ❌
Shell write detection (POSIX/PowerShell/cmd) ✅ ❌ ❌
Scope enforcement with cross-process persistence ✅ ❌ ❌
Interactive session detection and blocking ✅ ❌ ❌
Resumable sessions ✅ ❌ ❌
Built-in security scanning ✅ ❌ ❌
Learns from mistakes ✅ ❌ ❌

LLM Provider Guide

Swarm works with any provider supported by OpenCode.

Free Tier (OpenCode Zen)

No API key required. Excellent starting point:

{
  "agents": {
    "coder": { "model": "opencode/minimax-m2.5-free" },
    "reviewer": { "model": "opencode/big-pickle" },
    "explorer": { "model": "opencode/big-pickle" }
  }
}

Paid Providers

For production, mix providers by role:

Agent Recommended Why
architect OpenCode UI selection Needs strongest reasoning
coder minimax-coding-plan/MiniMax-M2.5 Fast, accurate code generation
reviewer zai-coding-plan/glm-5 Different training from coder
test_engineer minimax-coding-plan/MiniMax-M2.5 Same strengths as coder
explorer google/gemini-2.5-flash Fast read-heavy analysis
sme kimi-for-coding/k2p5 Strong domain expertise

Provider Formats

Provider Format Example
OpenCode Zen opencode/<model> opencode/big-pickle
Anthropic anthropic/<model> anthropic/claude-sonnet-4-20250514
Google google/<model> google/gemini-2.5-flash
Z.ai zai-coding-plan/<model> zai-coding-plan/glm-5
MiniMax minimax-coding-plan/<model> minimax-coding-plan/MiniMax-M2.5
Kimi kimi-for-coding/<model> kimi-for-coding/k2p5

Model Fallback

Automatic fallback to a secondary model on transient errors:

{
  "agents": {
    "coder": {
      "model": "anthropic/claude-sonnet-4-20250514",
      "fallback_models": ["opencode/gpt-5-nano"]
    }
  }
}

See docs/configuration.md for full configuration reference.

Process Remediation Model (PRM)

Swarm monitors agent trajectories and injects course-correction guidance before loops form. Detects five failure patterns:

  1. Repetition Loop — Same agent performs the same action repeatedly
  2. Ping-Pong — Agents hand off back and forth without progress
  3. Expansion Drift — Plan scope grows beyond original task
  4. Stuck-on-Test — Coder and tests fail in a loop
  5. Context Thrashing — Agent requests increasingly large file sets

When detected, escalation levels trigger:

  • Level 1: Advisory guidance injected
  • Level 2: Architect alert sent
  • Level 3: Hard stop directive

Configure via:

{
  "prm": {
    "enabled": true,
    "pattern_thresholds": {
      "repetition_loop": 2,
      "ping_pong": 4,
      "expansion_drift": 3,
      "stuck_on_test": 3,
      "context_thrash": 5
    }
  }
}

Note: Some configuration fields (max_trajectory_lines, escalation_enabled) are defined in schema but not yet enforced at runtime.

Persistent Memory

.swarm/plan-ledger.jsonl — authoritative source of truth (v6.44 durability model)

.swarm/context.md — technical decisions and cached SME guidance

.swarm/evidence/ — review/test results per task

.swarm/telemetry.jsonl — session observability events (fire-and-forget, never blocks execution)

.swarm/curator-summary.json — phase-level intelligence and drift reports

Guardrails & Circuit Breakers

Every agent runs inside a circuit breaker that prevents runaway behavior:

Signal Default Limit
Tool calls 200
Duration 30 min
Same tool repeated 10x
Consecutive errors 5

Limits reset per task. Per-agent overrides available in config.

Note: Built-in per-agent profiles override the global max_consecutive_errors to 8 for architect, coder, test_engineer, explorer, and reviewer (src/config/schema.ts DEFAULT_AGENT_PROFILES). The 5 shown above is the GuardrailsConfigSchema global default that applies only when no per-agent profile sets the value. See docs/installation.md § Guardrails Configuration for the full schema and docs/configuration.md § Per-agent override fields for how to override.

File Authority (Per-Agent Write Permissions)

Each agent can only write to specific paths:

  • architect — Everything (except .swarm/plan.md, .swarm/plan.json)
  • coder — src/, tests/, docs/, scripts/
  • reviewer — .swarm/evidence/
  • test_engineer — tests/, .swarm/evidence/
  • explorer, sme — Read-only

Override via authority.rules in config.

Quality Gates

Built-in tools verify every task before it ships:

  • syntax_check — Tree-sitter validation across the configured language grammar map
  • placeholder_scan — Catches TODOs, stubs, incomplete code
  • sast_scan — 68 security rules across 8 languages (offline)
  • sbom_generate — Dependency tracking (CycloneDX)
  • quality_budget — Complexity, duplication, test ratio limits

These quality gates run locally. No Docker is required. Config-gated features such as General Council web search can make external API calls only when you enable them and provide a Tavily or Brave Search key.

Context Budget Guard

The Context Budget Guard monitors how full the model's context window is getting across the whole conversation. It helps prevent context overflow before it becomes a problem.

Default Behavior

  • Enabled automatically — No setup required. Swarm starts tracking context usage right away.
  • What it measures — The estimated total tokens across all messages in the conversation (every text part of every message — your prompts, the model's responses, and injected content alike), divided by the model's context window. The window comes from context_budget.model_limits (default 128000), resolved per model/provider. It does not measure only swarm-injected content, and it does not read max_injection_tokens.
  • Warning threshold (0.7 ratio) — When total conversation tokens reach 70% of the model context window (e.g. ~89,600 of 128,000), the architect receives a one-time advisory warning. This is informational — execution continues normally.
  • Critical threshold (0.9 ratio) — When total conversation tokens reach 90% of the model context window (e.g. ~115,200 of 128,000), the architect receives a critical alert with a recommendation to run /swarm handoff. This is also one-time only.
  • Non-nagging — Alerts fire once per session, not repeatedly. You won't be pestered every turn.
  • Who sees warnings — Only the architect receives these warnings. Other agents are unaware of the budget.

To disable entirely, set context_budget.enabled: false in your swarm config.

On-demand status check: architects can invoke the context_status tool at any time to read current context-window headroom (tokens used, model limit, usage percent, threshold state, model ID, provider) without triggering advisory warnings or mutating state. Works even when context_budget.enabled is false.

Skill Propagation

Swarm includes an intelligent skill propagation system that tracks, validates, and scores skill usage across agent delegations. When the architect delegates to subagents (coder, reviewer, test_engineer, etc.), the system:

  • Logs skill usage to .swarm/skill-usage.jsonl with session-scoped entries (agent, task ID, skill paths, timestamp)
  • Scores skill relevance based on frequency, compliance, recency, task ID diversity, and context matching
  • Provides recommendations — when delegating without a SKILLS: field, the system suggests relevant skills with relevance scores
  • Enforces compliance — optional enforcement mode can block delegations missing the SKILLS: field
  • Auto-populates skill index — maintains a ## Available Skills section in .swarm/context.md with usage counts and compliance rates
  • Supports explicit routing — .opencode/skill-routing.yaml maps agent types to specific skill paths with optional keyword descriptions

Guardrails:

  • Relevance scoring threshold: 0.5 (skills below this are not recommended)
  • Maximum recommendations per delegation: 5
  • Scoring budget safeguard: Skipped when session exceeds 500 skill-usage entries to prevent unbounded file reads
  • Graceful degradation: Zero installed skills = zero friction — no warnings, no blocks, no auto-population

Configuration:

{
  "skillPropagation": {
    "enabled": true,
    "enforce": false,
    "audiences": []
  }
}
  • enabled (boolean, default true) — turns skill-propagation scoring, recommendations, and warnings on or off. The mandatory explicit-skill-reference integrity check still runs even when this is false.
  • enforce (boolean, default false) — when true, blocks delegations that omit the SKILLS: field instead of only warning.
  • audiences (string array, default [], max 16 entries, de-duplicated) — extra audience tokens to broaden which skills are considered in scope for this project, in addition to the active runner's own audience. Each entry must be a lowercase domain token matching ^[a-z0-9]+(?:[._-][a-z0-9]+)*$; the reserved value swarm-plugin and any runner:* prefix are rejected.

Skill routing file format (.opencode/skill-routing.yaml):

version: 1
routing:
  coder:
    - path: .claude/skills/writing-tests/SKILL.md
      keywords: ["test", "testing", "writing tests"]
    - path: .claude/skills/engineering-conventions/SKILL.md
      keywords: ["engineering", "conventions", "invariants"]
  reviewer:
    - path: .claude/skills/swarm-pr-review/SKILL.md
      keywords: ["review", "security", "audit"]

Routing skills are merged with scored recommendations, with explicitly routed skills receiving a boosted score (0.9) to prioritize them.

Skill Lifecycle Management

Seven skill-management tools (skill_generate, skill_list, skill_apply, skill_inspect, skill_regenerate, skill_retire, skill_improve) are opt-in via the skills.enabled config flag (default false). With the flag off, the architect does not see them; with the flag on, they reappear. Tools remain exported and registered — only the merged architect tool map is gated.

Swarm provides tools for managing generated skill lifecycles:

  • skill_retire — Retires an active generated skill by creating a retired.marker file in its directory. Retired skills are excluded from discovery, scoring, and injection. The SKILL.md file is preserved for auditability. Use skill_retire(slug, reason?) to retire a skill, or pass a reason for tracking purposes.

  • skill_regenerate — Re-clusters the source knowledge entries for an active skill and updates the SKILL.md in place. Reads the existing SKILL.md frontmatter to identify source knowledge IDs, resolves current entries from knowledge stores, and re-renders the skill content. Falls back to fuzzy slug-based re-clustering if source IDs yield no matches. Skills whose source knowledge is entirely archived are automatically retired during regeneration.

  • Auto-retire health check — During each curator phase run, skills are automatically evaluated for retirement when:

    • The skill's violation rate exceeds 30% (based on skill-usage.jsonl compliance verdicts)
    • All of the skill's source knowledge entries have been archived

    Auto-retired skills are noted in the phase digest summary. The check is non-blocking — errors are caught and logged without failing the curator.

  • Proposal cleanup — When a draft skill proposal is activated via skill_apply, the source proposal file is deleted as part of the activation process (best-effort; permission errors are logged but do not block activation).

External Skill Curation

Swarm provides an opt-in, quarantine-first pipeline for discovering, validating, and promoting external skills. Disabled by default — no network calls are made until explicitly enabled.

Enabling

external_skills:
  curation_enabled: true
  sources:
    - type: url
      location: https://example.com/skills/
      trust_level: medium

Validation Pipeline

Every candidate passes a 3-gate pipeline before entering quarantine:

Gate Name Description
1 Prompt Injection Scan 12 regex patterns plus oversized field, invisible character, and suspicious formatting checks detect system instruction injection, role hijacking, and instruction override attempts
2 Unsafe Instruction Scan 25 patterns detect shell commands, file system attacks, network exfiltration, and privilege escalation
3 Provenance Integrity SHA-256 content hash, timestamp validation, URL format checks, and publisher presence validation

Trust level modulation: low trust promotes warning-severity findings to errors (stricter); medium and high trust levels keep warnings advisory. Error-severity findings always block regardless of trust level.

Tool Reference

Tool Description
external_skill_discover Fetch and validate a skill from a configured source
external_skill_list List candidates with status filters
external_skill_inspect View full candidate details
external_skill_promote Promote validated candidate to active skill (user approval required)
external_skill_reject Reject candidate with reason
external_skill_delete Remove candidate from quarantine store
external_skill_revoke Retire a previously promoted skill

Security Guarantees

  • Disabled by default — no network calls until explicitly enabled
  • All candidates quarantined until human review and promotion
  • TOCTOU re-validation at promotion time
  • Content hash verification prevents tampering
  • Bounded concurrent fetches (5 simultaneous) and discovery limits (50 candidates per invocation)
  • Max candidate size and count bounds
  • Source origin validation (URLs must match configured sources)

Limitations

  • Static regex patterns only (no LLM-based detection)
  • No cryptographic signing (deferred)
  • No batch import (deferred)
  • No auto-promotion (human approval always required)

Configuration Reference

Key Type Default Description
context_budget.enabled boolean true Enable or disable the context budget guard entirely
context_budget.max_injection_tokens number 4000 Separate per-turn cap on system-enhancer injection. It is NOT the guard's budget — the guard measures total conversation tokens against model_limits, not this value
context_budget.warn_threshold number 0.7 Ratio (0.0-1.0) of the model context window (model_limits, default 128000) at which total conversation tokens trigger a warning advisory
context_budget.critical_threshold number 0.9 Ratio (0.0-1.0) of the model context window (model_limits, default 128000) at which total conversation tokens trigger a critical alert with handoff recommendation
context_budget.enforce boolean true When true, enforces budget limits and may trigger handoffs
context_budget.prune_target number 0.7 Ratio (0.0-1.0) of context to preserve when pruning occurs
context_budget.preserve_last_n_turns number 4 Number of recent turns to preserve when pruning
context_budget.recent_window number 10 Number of turns to consider as "recent" for scoring
context_budget.tracked_agents string[] ['architect'] Agents to track for context budget warnings
context_budget.enforce_on_agent_switch boolean true Enforce budget limits when switching agents
context_budget.model_limits record { default: 128000 } Per-model token limits (model name -> max tokens)
context_budget.unified_injection_tokens number undefined Opt-in unified ceiling (tokens) for combined system-enhancer + knowledge-injector injection per turn. When set, both hooks share this budget with proportional split
context_budget.tool_output_mask_threshold number 2000 Threshold for masking tool outputs (chars)
skills.enabled boolean false Gates the 7 skill-management tools (skill_generate, skill_list, skill_apply, skill_inspect, skill_regenerate, skill_retire, skill_improve) behind an opt-in flag. When false (default), these tools are hidden from the architect's tool map.
context_budget.scoring.enabled boolean false Enable context scoring/ranking
context_budget.scoring.max_candidates number 100 Maximum items to score (10-500)
context_budget.scoring.weights object { recency: 0.3, ... } Scoring weights for priority
context_budget.scoring.decision_decay object { mode: 'exponential', half_life_hours: 24 } Decision relevance decay
context_budget.scoring.token_ratios object { prose: 0.25, code: 0.4, ... } Token cost multipliers

Example Configurations

Minimal (disable):

{
  "context_budget": {
    "enabled": false
  }
}

Default (reference):

{
  "context_budget": {
    "enabled": true,
    "max_injection_tokens": 4000,
    "warn_threshold": 0.7,
    "critical_threshold": 0.9,
    "enforce": true,
    "prune_target": 0.7,
    "preserve_last_n_turns": 4,
    "recent_window": 10,
    "tracked_agents": ["architect"],
    "enforce_on_agent_switch": true,
    "model_limits": { "default": 128000 },
    "tool_output_mask_threshold": 2000,
    "scoring": {
      "enabled": false,
      "max_candidates": 100,
      "weights": { "recency": 0.3, "relevance": 0.4, "importance": 0.3 },
      "decision_decay": { "mode": "exponential", "half_life_hours": 24 },
      "token_ratios": { "prose": 0.25, "code": 0.4, "json": 0.6, "logs": 0.1 }
    }
  }
}

Aggressive (for long-running sessions):

{
  "context_budget": {
    "enabled": true,
    "max_injection_tokens": 2000,
    "warn_threshold": 0.5,
    "critical_threshold": 0.75,
    "enforce": true,
    "prune_target": 0.6,
    "preserve_last_n_turns": 2,
    "recent_window": 5,
    "tracked_agents": ["architect"],
    "enforce_on_agent_switch": true,
    "model_limits": { "default": 128000 },
    "tool_output_mask_threshold": 1500,
    "scoring": {
      "enabled": true,
      "max_candidates": 50,
      "weights": { "recency": 0.5, "relevance": 0.3, "importance": 0.2 },
      "decision_decay": { "mode": "linear", "half_life_hours": 12 },
      "token_ratios": { "prose": 0.2, "code": 0.35, "json": 0.5, "logs": 0.05 }
    }
  }
}

What This Does NOT Do

  • Does NOT prune chat history — Your conversation with the model is untouched
  • Does NOT modify tool outputs