Synthesis
What should I build today?
The Phase E synthesis forced a specific answer: if forced to merge frameworks into a flawless reference design, which five do you pick? The answer is the same as Phase C — because Phase D did not surface a framework that replaces any of the five on its own dimension. But Phase D added two more for sub-agent context injection and the security model, yielding a 7-framework defensible baseline.
Source: _index/PHASE-E-EXECUTIVE-SUMMARY.md §5
The 5-framework synthesis baseline
Same as Phase C. Phase D confirmed rather than replaced these.
Iron-law SKILL.md + lifecycle hooks. Required baseline for behavioral enforcement.
Spec-as-source-of-truth, delta-diff specs, multi-tool mirror. Required baseline for spec lifecycle.
Most complete PR-lifecycle pipeline; tracker poll + tmux PTY + reaction system. Required baseline for production daemon.
Spec auto-backpropagation + worktree-per-task. Most novel living spec mechanism.
HTTP-proxy memory layer with WITNESS hallucination certificates. Only substrate-level memory framework.
Phase D additions — the 7-framework defensible set
Add these two to the Phase C baseline to cover all 25 rubric dimensions including the 5 new Phase D dimensions.
PreToolUse-on-Task sub-agent context injection. Specs injected, not remembered. Solves context drift in multi-agent hierarchies.
GitHub Next's safe-outputs separation, SHA-pinned deps, network isolation. Most security-conscious framework in the corpus.
Top 15 frameworks worth deep-dive
Selected to maximize archetype diversity, primitive novelty, and engineering rigor — not stars.
From PHASE-E-EXECUTIVE-SUMMARY.md §5.
A1, 207k★ — iron-law SKILL.md + lifecycle-hook formula. Required baseline for behavioral enforcement.
A2, 51k★ — spec-as-source-of-truth, delta-diff specs, multi-tool mirror. Required baseline for spec lifecycle.
A6, 55k★ — hive-mind queen+workers + 305 MCP tools + memory. Required baseline for orchestration scale.
A7 — most complete PR-lifecycle pipeline; tracker poll + tmux PTY + reaction system. Required baseline for production daemon.
A6 — 16 agents, typed confirmation gates, 1,660 passing tests, cost-tier executor mapping. Best mid-size operational rigor.
A11 — spec auto-backprop + worktree-per-task. Most novel living spec mechanism.
A21 — LLM-evaluated hooks + Bug Council 5-analyst escalation + SQLite per-task model history.
A11 — adversarial duel review + Zod-validated typed completion attestations. Best self-review pattern.
A10 — HTTP-proxy memory layer with WITNESS hallucination certificates. Only substrate-level memory framework.
A14 — JIT semantic context loading; cleanest answer to context window bottleneck.
A21, 8.4k★ — PreToolUse-on-Task sub-agent context injection. Specs injected, not remembered.
A16, 5.4k★ — Cross-vendor Claude+Codex+Gemini routing + BREAK-LOOP PROTOCOL meta-cognitive recovery.
A11, 4.5k★ — GitHub Next safe-outputs separation, SHA-pinned deps, network isolation. Most security-conscious.
A19, 54.6k★ — PreToolUse-Bash interception + verbose-CLI compression. Highest-starred Phase D framework.
A13, 10.5k★ — Forensic observatory of Claude Code's internal system prompts. Reveals 8-stage multi-agent review + dream-memory consolidation already inside the tool.