Skip to content
swarm

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

1.4.0 - 2026-02-26

Section titled “1.4.0 - 2026-02-26”

Added

Section titled “Added”
  • JSONL Log Analyzer skill (skills/jsonl-log-analyzer/SKILL.md): Automated schema-aware JSONL log analysis — discovers field schema via jq, generates tailored extraction recipes, and orchestrates the RLM fan-out/fan-in pipeline with swarm:rlm-json-analyzer analysts
  • JSONL Log Analysis examples in docs/rlm-examples.md: Copy-paste prompts for error investigation, traffic analysis, and incident timeline reconstruction with comparison to standard JSON/JSONL RLM
  • Cross-references in docs/agent-types.md and skills/agent-types/SKILL.md linking swarm:rlm-json-analyzer to the new JSONL Log Analyzer skill
  • docs/concepts.md (explanation): Team lifecycle, pattern mechanics, RLM content processing internals, cross-file analysis, task dependencies
  • docs/reference.md (reference): Consolidated lookup tables — primitives, pattern quick reference, message types, task statuses, task fields, configuration, spawn backends, RLM tables, known limitations, debugging commands

Changed

Section titled “Changed”
  • Restructured docs/ along the Diataxis framework (tutorial, how-to, reference, explanation separation)
  • Refocused docs/patterns.md as how-to: extracted “How it works” blocks to concepts.md, pattern table and agent recs to reference.md, absorbed “Choosing Agents for Teams” from agent-types.md
  • Cleaned docs/rlm-examples.md as how-to: extracted “What Claude does” blocks to concepts.md, reference tables to reference.md, added cross-reference links
  • Cleaned docs/troubleshooting.md: extracted Known Limitations and Debugging Commands to reference.md
  • Cleaned docs/getting-started.md: extracted “What happens” lifecycle to concepts.md, added reference link to Next Steps
  • Cleaned docs/agent-types.md: moved “Choosing Agents for Teams” to patterns.md, added See Also footer

1.3.0 - 2026-02-13

Section titled “1.3.0 - 2026-02-13”

Added

Section titled “Added”
  • Anti-pattern guardrail (“Critical: What RLM is NOT”): New section in skills/rlm-pattern/SKILL.md that prevents thematic decomposition — explicitly warns against creating per-analysis-goal tasks instead of per-data-partition tasks
  • Pre-processing for compressed inputs: New section handles .zip, .gz, .tar.gz archives before content-type detection, with routing to single-file or multi-file mode based on archive contents
  • Mandatory CSV partitioning procedure: Concrete step-by-step bash commands (wc -l → divide → write chunk files) in the Structured Data section — eliminates ambiguity about how to partition CSVs
  • Multi-file partitioning sanity check: Warning in Multi-File Directory Analysis section with concrete validation: “if task count = file count, you haven’t partitioned”
  • Large CSV Directory walkthrough: New abbreviated walkthrough showing 11 CSV files → 88 chunks → 88 analysts with staged spawning — mirrors real-world Jira export analysis scenario

Changed

Section titled “Changed”
  • 1:1 analyst-per-partition model: Replaced ceil(partition_count / 4) formula with analyst_count = partition_count across all sizing guidance (3 locations in SKILL.md, design doc, and all referencing docs)
  • Staged spawning: Analysts spawn in batches of ~15 instead of all at once; pre-assigned chunks instead of shared task pool claiming
  • Fresh context per analyst: Each analyst gets a clean context window for one chunk — no cross-chunk contamination from reusing analyst contexts
  • Wide CSV chunk range widened to ~500-1500 rows: ~500 for dense wide data, ~1500 for sparse wide data (e.g., Jira exports with many empty columns)
  • Removed ~15 analyst cap: No hard ceiling on analyst count; Haiku analysts are lightweight
  • Aggressive partitioning guidance: Replaced “practical ceiling” language with “no ceiling when using findings-in-task-descriptions mode and staged spawning”
  • Updated Step 1 and Step 2 lifecycle code examples to show 1:1 pre-assigned model
  • Updated “Why Teams” table to show pre-assigned distribution model
  • Updated multi-file sections: analyst mix text, walkthrough analyst counts, context management
  • Synchronized all referencing documentation: docs/patterns.md, docs/rlm-examples.md, skills/orchestration-patterns/SKILL.md, skills/agent-types/SKILL.md, docs/design/content-aware-rlm.md, docs/design/multi-file-rlm.md, skills/orchestration-patterns/examples/complete-workflows.md

1.2.2 - 2026-02-13

Section titled “1.2.2 - 2026-02-13”

Changed

Section titled “Changed”
  • Replaced hard caps (6 analysts, 30 partitions) in RLM pattern with data-driven scaling guidance
    • Partition count now scales with data size using content-type chunk targets instead of a fixed 30-partition cap
    • Analyst count scales to task volume (1 per 3-5 tasks) instead of a fixed 6-analyst cap
    • CSV partitioning uses concrete row targets (~2000 rows narrow, ~500 rows wide) instead of “target 5-10 partitions”
    • Practical ceilings framed as cost/context trade-offs (~50 partitions, ~15 analysts) rather than hard limits
  • Updated sizing guidance across skills/rlm-pattern/SKILL.md, skills/orchestration-patterns/SKILL.md, skills/agent-types/SKILL.md, docs/rlm-examples.md, docs/design/multi-file-rlm.md, and CLAUDE.md

1.2.1 - 2026-02-11

Section titled “1.2.1 - 2026-02-11”

Added

Section titled “Added”
  • RLM Examples Guide (docs/rlm-examples.md): Copy-paste example prompts for all 6 RLM modes (basic, code-aware, CSV, JSON, directory, multi-type) with feature comparison table and sizing guidelines

1.2.0 - 2026-02-11

Section titled “1.2.0 - 2026-02-11”

Added

Section titled “Added”
  • Content-Aware RLM: Automatic content-type detection (extension mapping + content sniffing) and type-specific chunking strategies. Design document at docs/design/content-aware-rlm.md
  • New agents (agents/):
    • swarm:rlm-code-analyzer — Haiku-powered source code analyzer with function/class scope context, severity levels, and configurable analysis focus (general, security, architecture, performance)
    • swarm:rlm-data-analyzer — Haiku-powered CSV/TSV analyzer with column-aware distributions, statistics, and aggregatable frequency counts
    • swarm:rlm-json-analyzer — Haiku-powered JSON/JSONL analyzer with schema pattern detection, field distributions, and type consistency checks
  • Multi-File Directory RLM: Extension of single-file RLM to process entire directories with mixed content types in one team session. Design document at docs/design/multi-file-rlm.md
    • Per-file content-type detection and mixed analyst routing
    • Tiered partition budget (small/medium/large) with 30-partition global cap
    • Small file batching by content type
    • Two-phase synthesis: per-type (parallel) then cross-type (sequential) via task dependencies
    • Findings-in-task-descriptions pattern to protect Team Lead context
    • Max 6 analysts across all types
  • Workflow 5: Multi-File Directory Analysis (skills/orchestration-patterns/examples/complete-workflows.md): End-to-end example with mixed analyst types, two-phase synthesis, and task dependency wiring

Changed

Section titled “Changed”
  • Updated skills/rlm-pattern/SKILL.md with content-type detection, type-specific partitioning strategies, agent routing matrix, content-aware team composition, and multi-file directory analysis section
  • Updated skills/agent-types/SKILL.md with 3 new content-aware RLM agents and multi-file mixed-type usage note
  • Updated skills/orchestration-patterns/SKILL.md Pattern 7 with content-aware agent recommendations and multi-file variant
  • Updated agents/rlm-chunk-analyzer.md with role scope note (general-purpose analyzer for logs, prose, config) and multi-file mode workflow
  • Updated agents/rlm-synthesizer.md with heterogeneous findings support, TaskGet/TaskUpdate tools, and multi-file synthesis modes (per-type and cross-type)
  • Updated all 4 analyst agents with multi-file mode reporting (findings to task descriptions via TaskUpdate)
  • Updated docs/agent-types.md, docs/patterns.md to reflect content-aware and multi-file capabilities
  • Updated docs/design/content-aware-rlm.md Future Considerations with forward reference to multi-file design
  • Added analyst model override guardrails to skills/rlm-pattern/SKILL.md and skills/agent-types/SKILL.md

1.1.0 - 2026-02-10

Section titled “1.1.0 - 2026-02-10”

Added

Section titled “Added”
  • RLM Pattern (Pattern 7): Fan-out/fan-in chunked parallel analysis for files exceeding context limits, based on arXiv:2512.24601. Uses role-based workflow descriptions that naturally trigger team orchestration (TeamCreate, Task, SendMessage, TeamDelete)
  • New agents (agents/):
    • swarm:rlm-chunk-analyzer — Haiku-powered chunk-level analysis agent with structured JSON output; reads files directly via Read tool with offset/limit
    • swarm:rlm-synthesizer — Sonnet-powered synthesis agent for aggregating multi-chunk findings into consolidated reports with file:line references
  • New skill (skills/rlm-pattern/):
    • swarm:rlm-pattern — Complete RLM workflow with partitioning strategy guide (CSV, logs, source code, prose, time-series), team composition table, analyst report format, synthesis output structure, context management guardrails, and comparison with rlm-rs plugin
  • Workflow 4: RLM Document Analysis (skills/orchestration-patterns/examples/complete-workflows.md): End-to-end scenario-based example with Grep scouting, targeted chunk analysis, and structured synthesis
  • Example analysis prompt (docs/RLM-Prompt-Example.md): Reference prompt for CSI Jira export analysis using the RLM pattern

Changed

Section titled “Changed”
  • Updated orchestration pattern count from 6 to 7 across all documentation
  • Added RLM agents (swarm:rlm-chunk-analyzer, swarm:rlm-synthesizer) to agent type selection guides
  • Added best practice “Use Pass-by-Reference for Large Content” to orchestration patterns
  • Updated README skills reference table and verify-installation skill list
  • Updated plugin version from 1.0.0 to 1.1.0

1.0.0 - 2026-02-10

Section titled “1.0.0 - 2026-02-10”

Added

Section titled “Added”
  • Plugin manifest (.claude-plugin/plugin.json): Packaged as standalone Claude Code plugin named “swarm” with proper metadata, version, homepage, and license
  • Skills (8 modular skills under skills/):
    • orchestrating — Primitives overview, lifecycle diagram, quick reference
    • team-management — Team creation, teammate spawning, delegate mode, permissions, graceful shutdown, cleanup
    • task-system — Task CRUD, dependencies, claiming with file locking, status tracking
    • messaging — Direct messages, broadcasts, shutdown requests/responses, plan approval protocol, idle notifications
    • agent-types — Built-in agents (Bash, Explore, Plan, general-purpose, claude-code-guide) and plugin agents (review, research, refactoring, CI/CD, test generation) with selection guide
    • orchestration-patterns — Six proven patterns: parallel specialists, pipeline, swarm, research+implementation, plan approval, multi-file refactoring
    • spawn-backends — In-process, tmux, and iTerm2 backends with auto-detection logic and configuration
    • error-handling — Common errors, quality gate hooks (TeammateIdle, TaskCompleted), known limitations, recovery strategies
  • Complete workflow examples (skills/orchestration-patterns/examples/complete-workflows.md): Full end-to-end examples for parallel code review, research-plan-implement pipeline, and self-organizing swarm
  • Documentation (docs/):
    • getting-started.md — End-to-end walkthrough from install to first team run
    • agent-types.md — Agent selection reference with capabilities and use cases
    • patterns.md — Pattern selection guide with when-to-use guidance
    • troubleshooting.md — Common failures, debugging commands, and known limitations
  • README.md: Adoption-focused guide with prerequisites (Claude Code, agent teams flag, tmux), installation, quick start example, skills reference table, and attribution

Attribution

Section titled “Attribution”

Based on the Claude Code Swarm Orchestration Skill by @kieranklaassen. Decomposed from a monolithic gist into modular, maintainable plugin skills.