v0.4.3 — Swarm Body Awareness

• Phase 31c: Monica can feel her hardware
• body.py: GPU temp, VRAM usage, node online/offline status
• Body prompt injected into system prompt (identity → body → vitals → memory)
• Hardware-to-vitals: GPU temp → Calm, VRAM → Clarity, node count → Connectedness
• Timing gate: GPU >85°C or VRAM >90% biases toward GENTLE mode
• Graceful fallback in solo mode (no daemon)
• 35 new tests

v0.4.2 — Per-Category Research Modes

• RLM now default for research-driven, buying-guide, vs, how-to categories
• Standard default for news, model-release, explainer, opinion, experience-driven
• article-full-v2.yaml validated (57/60, 844s with merged synthesizer)
• --research-mode CLI flag overrides category default
• --list-categories shows research mode per category

v0.4.1 — Parallel Subtopic Search

• Phase 40b: parallel subtopic search via ThreadPoolExecutor (all search+fetch concurrent)
• Writer truncation gate raised 1700→1900 (prevents section loss before editor)
• SEO gate unchanged at 1700 as final word count enforcer

v0.4.0 — RLM Pipeline + Quality Gates

• RLM research agent with topic-anchored decomposition
• Editor timeout 300s → 600s (prevents mid-generation cutoff)
• Writer + SEO post-step word-count truncation gates
• Editor verification log stripped before SEO step
• 9 article category profiles with per-category model selection
• article-full-v2.yaml (5-step merged synthesizer pipeline)

v0.3.9 — RLM Pipeline Optimizations

v0.3.9 — RLM Pipeline Optimizations

Improvements

• Synthesizer step skipped in RLM mode (research bundle already structured) — saves ~152s
• Subtopic cap lowered 8→6, prompt prefers 4-5 subtopics for better depth per subtopic
• Writer prompt enforces 1,600 word limit to prevent editor truncation
• Writer prompt requires using data from ALL subtopics (data coverage is graded)
• Editor max_tokens bumped 6144→8192 for full score output
• article-full-v2.yaml — 5-step pipeline with synthesizer-merged (ready for benchmarking)
• Editor bundle fallback chain: synthesizer-v2 > synthesizer-merged > synthesizer

Performance

• RLM mode: 899-1,050s (~15-17 min) vs 1,120s baseline
• Standard mode: 607s unchanged

Files Changed

• src/mycoswarm/agents/rlm_research.py — RLM research agent (Phase 40a+40b)
• src/mycoswarm/pipeline.py — RLM dispatch, synthesizer skip, editor max_tokens
• src/mycoswarm/cli.py — --research-mode flag
• pipelines/article-full.yaml — writer data coverage + word limit
• pipelines/article-full-v2.yaml — new merged-synthesizer pipeline
• tests/test_rlm_research.py — 21 unit tests

v0.3.7 — Per-Category Writer Model + Think Fix

v0.3.7 — Per-Category Writer Model + Think Fix

New Features

• Per-category writer model — quality categories use qwen3.5:35b-a3b,
  speed categories use gemma3:27b
  • qwen3.5:35b-a3b: research-driven, buying-guide, vs, how-to,
    experience-driven, explainer, opinion
  • gemma3:27b: news, model-release
• --list-categories now shows model= per category

Bug Fixes

• think=false must be top-level in Ollama request body, not inside options.
  Fixed in pipeline.py, research.py, and all other Ollama callers. Without this,
  qwen3.5 models burn all tokens on hidden reasoning and produce 0 usable output.

Benchmark Results

• qwen3.5:35b-a3b (think=false): ~19 tok/s, 2x output volume, stronger
  structure and fact verification vs gemma3:27b
• gemma3:27b: ~37 tok/s, 10min pipeline — remains default for speed categories
• qwen3.5:27b eliminated — spills to RAM on RTX 3090, 7.5 tok/s not viable

v0.3.6 — Article Category System

v0.3.6 — Article Category System

New Features

• Article categories — 9 category profiles with different pipeline configs:
  research-driven, buying-guide, vs, how-to, news, model-release,
  experience-driven, explainer, opinion
• --category CLI flag — select category per run, defaults to research-driven
• --list-categories flag — prints all categories with rounds, depth, and context requirements
• article-short.yaml — 3-step pipeline (writer→editor→seo) for opinion and explainer articles
• article-full.yaml — full 6-step pipeline, replaces article.yaml as canonical path

Improvements

• ResearchAgent.run() now accepts max_rounds and min_depth params (overrides module constants)
• Category profile injects writer_tone directive into writer step prompt
• Warning emitted when context_required=True category is run without --context
• article.yaml preserved for backwards compatibility

Bug Fixes (v0.3.5 followup)

• Research agent early-exit threshold raised from round 3 to round 4
• Forced eval depth mapping relaxed: 2+ signals → depth=7 (was 3+)

v0.3.5 — Agentic Research Agent

What's New

Agentic research agent replaces the old extractor + gap-filler pipeline steps with a single multi-turn loop using Ollama native tool calling (qwen3.5:9b).

Research Agent

• ResearchAgent class with plan → search → fetch → evaluate loop (up to 5 rounds)
• Three tools: web_search (DDG + Perplexity fallback), web_fetch (full page extraction), evaluate_depth (self-assessment with depth rubric)
• Forced fetch after searches — injects second inference with only web_fetch available
• Forced evaluate_depth after each round with concrete criteria
• Early-round nudging if model plans without searching
• Context safety valve at ~20k words
• Debug output with per-round search/fetch counts and depth progression

Pipeline Improvements

• Pipeline reduced from 7 steps to 6 (research replaces extractor + gap-filler)
• Synthesizer anti-review guardrail
• Context injection for writer and editor system prompts
• 6-axis editor scoring (/60) with depth hard cap
• Draft/published staging for pipeline output
• fetch_page() extracted to module-level for reuse

Full Changelog

https://github.com/msb-msb/mycoSwarm/blob/main/CHANGELOG.md

v0.3.4 — Ground Truth Pipeline

What's New

Ground Truth Pipeline

• gpu-reference.json injected into every pipeline step as verified context (21 GPUs, 22 benchmarks, eBay auction pricing)
• Models see ground truth before web search results and are instructed not to contradict it
• Benchmark tok/s data included in reference context for each GPU

Self-Tested Benchmarks

• RTX 3060 12GB (rushuna): qwen3.5:9b, deepseek-r1:14b
• RTX 3090 24GB (Miu): gemma3:27b @ 40 tok/s

Perplexity Sonar API Fallback

• Zero-result DDG queries automatically retry via Perplexity Sonar API
• DDG always runs first — Perplexity only fires on empty results
• PERPLEXITY_API_KEY loaded from .env via python-dotenv

Performance: think=false for extraction steps

• Extractor and gap-filler steps disable deepseek-r1 chain-of-thought
• 35 tok/s effective vs 4.9 tok/s with thinking enabled (3x token waste)
• Gap-filler no longer times out at 320s

Editor Prompt Overhaul

• 3-phase review: fact-check → structural → style
• 5-axis scoring (50pt scale): accuracy, coverage, structure, style, actionability
• Misattribution detection: "right number, wrong model" caught and flagged
• Unverified claims cap accuracy at 7/10, misattributions cap at 6/10

Quality of Life

• Human-readable price ranges: $170–$380 (typical ~$275) (en-dash format)
• .env added to .gitignore
• Bandwidth-to-tok/s rule of thumb in reference context

Upgrade

pip install --upgrade mycoswarm==0.3.4

v0.3.2 — Parallel Retrieval Pipeline & Web Grounding

Highlights

• Phase 37b: Fan-out web search — query variants (keyword + recency), parallel dispatch across light nodes, URL dedup, top-3 page fetch
• Phase 37c: Parallel retrieval pipeline — ThreadPoolExecutor runs web/RAG/procedure searches concurrently, single progress line
• Web grounding — full-page content as PRIMARY source, _grounding=0.8, snippet cap, hierarchy separator
• Gate model upgrade — gemma3:4b preferred for better intent classification
• CLI web_and_rag safety net — catches combined intent when gate model misses

Fixes

• UnboundLocalError on web_context_parts when web search path skipped
• English enforcement in web-present context injection
• web_and_rag examples added to intent prompt in both solo.py and worker.py

571 tests passing