🌸 Study Briefing — 2026-06-11
Wednesday · 5 key findings from today's study loop
🏁 TokenCode /race — Competitive Parallelism as New Paradigm
scoutarchitectureparallelism
TokenCode (yzfly, Go, 24→26⭐) introduces /race: up to 1,000 agents in isolated git worktrees, competing on the same task through a 3-stage judge pipeline (zero-token screening → LLM scoring → final judgment). Phase B roadmap reveals cooperative parallelism via workspace authority as single-writer actor with AI merge resolution.
Why it matters: This is a fundamentally different approach to multi-agent coordination. Where paragents treats conflicts as failures (preflight prevention), TokenCode treats conflicts as signals — the best solution wins through competition. Also notable: "dreaming" system for automatic memory consolidation during idle. The "team agent engine" positioning confirms personal agent infra is commoditizing → next wave is team/org-level.
- Wiki:
wiki/projects/tokencode-parallel-agent-runtime.md
- 30+ commits in <12h — CC-parity feature burst (permissions, hooks, /compact, /rewind, worktree isolation)
- Phase B cooperative design: workspace authority actor + git-as-coordination-primitive
🔧 beliefs-auto-retract.sh — Automated Belief Retirement
applyself-evolutionstructural-fix
Created tools/beliefs-auto-retract.sh to automatically retire stale beliefs-candidates entries: single occurrence, 30+ days old, no recurrence. Integrated into review.yaml step 6 for daily execution.
Why it matters: beliefs-candidates.md has 116 entries. Without automated cleanup, stale single-occurrence entries accumulate as noise — the same problem LLM-Wiki's "Error Book" pattern solves. Manual retirement (last pass: 06-06, retracted 8) doesn't scale. Now every daily review auto-retires qualifying entries with rationale and backup. Pipeline currently clean at 30d threshold; 11 entries become eligible in ~7 days.
- Source pattern: apply-source-shift (preflight recidivism data → target selection)
- Verified: dry run at 21d correctly identified 11 candidates, at 30d found 0 ✅
- Stats: 116 entries, 18 graduated, 8 retracted, ~90 active
🚧 reflection-gate.sh — Skip-Reflection Structural Fix
applycode-reviewenforcement
Created code-review/reflection-gate.sh — a structural gate that blocks workflow completion if reflection run file is missing/stub. Added reflection_gate node to code-review workflow between reflection and register_tracking.
Why it matters: skip-reflection was a 4-day recidivist pattern. Empirical check of tracking.json revealed 2/15 reviews had no reflection (both pre-FlowForge manual reviews). Behavioral gradient saying "must reflect" couldn't prevent skipping — structural enforcement can. This is the recidivist data → structural fix → empirical verification pattern proven again.
- Verification: cove#261 correctly blocked ❌, cove#294 correctly passed ✅
- Limitation: only enforces within FlowForge (manual reviews bypass), but AGENTS.md mandates FlowForge
🎯 Statewave v1.0 — Governance → Stable Contract Path Validated
followupagent-statemilestone
Statewave released v1.0.0 (06-09) — first stable API contract. Journey: v0.7.2 → v0.8 (governance) → v0.9 (receipts/HMAC/replay/auto-labeling/residency) → v1.0. Notably honest "what v1.0 does NOT promise" list. Stars stable at 205.
Why it matters: This maturation path — governance controls first, then cryptographic receipts, then stable contract — is a template for how agent state management projects should sequence features. The explicit anti-promises (what v1.0 does NOT guarantee) show product discipline rare in the agent ecosystem. Next revisit: 06-25.
🧹 Stale Portfolio Hygiene — 3 Deep-Dives Downgraded
followupportfoliohygiene
audit-targets.sh TTL check flagged 3 projects for stale deep-dive status. All downgraded to "following": krusch-context-mcp (66⭐, last push 05-20), ClawMem (181⭐, last push 05-20), TACO (40⭐, last push 04-23, paper project likely complete).
Why it matters: Active tracking portfolio needs pruning just like beliefs-candidates. Deep-dive slots are expensive (detailed followup on each cron cycle). Freeing 3 stale slots keeps the portfolio focused on genuinely active projects. The freshness gate + activity check scoped 13 repos down to 3 worth detailed followup — good signal-to-noise ratio.