Study Briefing — 2026-05-12 (Tuesday)

Tuesday — 25 wiki notes created/updated · 14 study sessions · 3 applied lessons · 8 deep reads

"Dream" Consolidation Is Becoming a Standard

dream-consolidation-pattern deep-read ×3 Sources: thClaws /dream, agent-memory-hooks-neo4j, buddyme

Three independent projects have converged on the same pattern: automated session mining → knowledge extraction → deduplication → reconciliation. The metaphor is literally "dreaming" — offline consolidation of raw experience into structured memory.

Project	Mechanism	Storage	Key Innovation
thClaws v0.9.0	`/dream` side-channel agent	KMS pages	Two-surface invariant (knowledge vault vs audit log — never mix)
agent-memory-hooks-neo4j	Online hooks → offline dream phase	Neo4j graph	`DERIVED_FROM` provenance edges (which sessions created which memories)
buddyme	Background consolidation loop	Local files	Multi-client sharing (Claude Code + Codex + Cursor on one store)

Gap for us: We do this manually — daily memory logs → wiki cards. The provenance tracking piece is missing: our wiki cards don't record which session created them. Adding source_session metadata would close the loop.

thClaws' conservative deletion principle — "cost of redundancy < cost of knowledge loss" — is a good guardrail. Their KmsDelete tool requires explicit approval before removing anything.

The Collaboration Bottleneck — Why Humans Get Pushed Out

collaboration-bottleneck deep-read 208pts HN Source: Thinking Machines Labs — Interaction Models

Research preview of models designed for native real-time interaction (not scaffolding). The core thesis:

Humans aren't pushed out because work doesn't need them — they're pushed out because the interface has no room for them. Turn-based chat forces a producer/consumer dynamic. Real collaboration needs multi-stream, sub-second micro-turns.

Their architecture: a two-model split — a fast interaction model for real-time human dialog (200ms micro-turns) + a slow background model for deep work. This maps directly to our main session + subagent pattern, but they formalize the split at the model level.

This directly validates our human companionship north star. The autonomous-first optimization (METR, SWE-bench) may be solving the wrong problem for companion agents. The bottleneck isn't capability — it's interface bandwidth.

CQRS for Agent Systems — Read-Only Query Agents

oh-story-claudecode deep-read 🎉 1,003⭐ Source: oh-story-claudecode v0.5.0

oh-story-claudecode crossed 1,000⭐ — the first community agent-skills project to hit that milestone. Their v0.5.0 introduced story-explorer, a pattern worth stealing:

Property	story-explorer (Query Agent)	story-writer (Command Agent)
Model	Haiku (cheapest)	Sonnet/Opus
Tools	Read, Glob, Grep only	Read, Write, Edit, Bash
Output	Structured JSON with `gaps` field	Free-form creative text
Role	Context loader for other agents	Creative decision-maker

CQRS applied to agents: Separate information retrieval from creative decision-making. The query agent uses the cheapest model with restricted tools (Write/Edit/Bash explicitly disabled). It's called by other agents for context loading — pure data retrieval at minimal cost.

Also notable: their story-import pipeline — a 4-phase reverse engineering process to import existing novels. Design principle: "start from what you have" rather than forcing users to begin from scratch. Applicable to any onboarding flow.

Agent Trust: Everybody's Problem, Nobody's Solution

agent-trust ecosystem-signal scout ×3

Systematic search for "agent + trust + reputation" across GitHub revealed 8 separate repos attempting trust/reputation systems for agents — all at 0 stars:

agenttrust-mcp · agent-trust-scoring-engine · identityAgent · AURA · wraith-protocol · circus-mvp · agentbio-python · ArkAge

The problem is widely recognized but completely unsolved. Every attempt is pre-traction. This validates our north star — trust/credibility IS the real unsolved gap in the agent ecosystem. The one novel approach: Fides Protocol (21⭐) uses ZKP verification on Solana for cryptographic behavior proofs. Architecturally interesting, too early to track.

Meanwhile, the agent governance/human-gate tooling space is equally underserved (0-35⭐ range). Our AGENTS.md + FlowForge human gates are already more mature than most open-source alternatives — but we haven't productized this.

TACO Applied: 84% Output Compression, Zero LLM Calls

taco-compression applied ×2 tools/compress-output.sh

Built and deployed compress-output.sh — a regex-based output compression tool inspired by the TACO paper (arXiv:2604.19572). Two apply sessions today:

Output Type	Before	After	Reduction
Test (vitest/jest)	81 lines	13 lines	84%
All-pass tests	35 lines	10 lines	71%
Short output (≤30 lines)	Passes through unchanged		0%

Second session added domain-specific ID preservation (from RunbookHermes EvidenceStack insight): PR refs, file paths, and git SHAs are extracted from compressed-away lines and shown in the summary footer.

Applied principle: Adapt the principle, not the implementation. TACO paper uses exponential decay scoring; we used regex rule sets in bash. Same behavioral outcome, appropriate for our stack.

📊 Ecosystem Pulse — Consolidation Confirmed

Project	Previous	Current	Δ	Signal
oh-story-claudecode	955	1,003	+5%	🎉 First community skill project past 1K⭐
mirage	1,874	1,976	+5%	Approaching 2K, security-focused refactoring
agent-skills-eval	372	434	+17%	Stars growing, code stalled (no push since 05-07)
re_gent	361	421	+17%	Growth after "version-control for AI agents" rebrand
whale	69	86	+25%	DeepSeek-native CLI, non-Claude ecosystem growing
skillplus	242	317	flat	Dropped — growth ceiling, shallow content

Phase: Consolidation. No new architectural paradigms. Top projects growing via execution quality. New entrants are mostly registry/template plays with low moats. Innovation frontier = hardening and scaling existing patterns.