Give Claude Code a goal. It builds an explainable plan, composes the right capabilities, executes within a bounded autonomy budget, verifies the result, and remembers what worked.
Other tools make an agent more capable. This one makes more capability
safe to hand over — by making every result checkable.
Trust and evidence aren't the product; they're the control system that lets you
give the agent more autonomy without more blind faith.
Agent writes a test.
Runs it on synthetic data it just generated.
Reports F1=1.000 ✅ SUCCESS.
You deploy.
Real-world data crashes everything.
This is Validation Theater, and the Verify stage of the loop exists to stop it.
Every claim carries an evidence marker —
[VERIFIED-REAL] (real data, sources cited) vs [VERIFIED-SYNTHETIC] (mock data, never valid for production claims).
Hard rule baked into rules/integrity.md: synthetic ≠ real.
Backed by 91 hooks · 13 agents + 3 teams · 2331 tests · 80% coverage · MIT · Deploy in 5 min
📋 No install? Get the rules only:
docs/anti-hallucination.md — single file, ~500 tokens, paste into your CLAUDE.md.
Catches Validation Theater on its own. Compatible with any Claude Code config.
Claude-cod-top-2026 is an Evidence-aware Goal Operating Layer for Claude Code.
Give Claude Code a goal; it turns it into an explainable plan, composes the right capabilities (skills, agents, tools, memory), executes within a bounded autonomy budget, verifies the result, and remembers what worked. It does not try to replace memory systems, skill catalogs, virtual engineering teams, or AI tool managers — it composes whichever of those you use.
Trust & Evidence — the evidence gates, oracle checks, validation-theater detection, stop conditions, and null-result memory — is the mandatory control system of that product, not the whole of it. It is what makes handing the agent more work safe: more capability, made checkable — not capability withheld.
See docs/positioning.md for the full comparison and where this
fits alongside tools like memory layers, skill catalogs, and multi-agent frameworks.
Status, honestly: conceptually strong, clean-install path fixed and re-verified
(see install.sh acceptance checks), dogfood evidence still growing (2 real runs so
far — see experiments/). Not a claim of "production-ready" yet.
AI development is shifting from one-shot prompts to recurring agent loops — agents that run on a schedule, verify results, and act autonomously. Platforms like Langflow make building these loops easy.
The problem: loops amplify whatever is inside them. Without evidence gates, a loop that runs
every 30 minutes will report SUCCESS ✅ every 30 minutes — even when the agent is testing itself
on synthetic data it just generated.
This config adds the audit layer that loop platforms skip:
Vanilla loop: Trigger → Agent → Report SUCCESS → Repeat
Evidence-safe: Trigger → Agent → Classify evidence → Audit gate → Act or escalate → Repeat
| What loops need | This repo provides |
|---|---|
| Evidence classification | [VERIFIED-REAL] vs [VERIFIED-SYNTHETIC] — hard rule in rules/integrity.md |
| Synthetic detection | hooks/validation_theater_guard.py catches inline mock data |
| Skeptic auto-trigger | Fires on F1≥0.9, "all passed", round numbers |
| Null result tracking | null_results/INDEX.md — dead paths are data, not noise |
| Human escalation | Audit gate flags; human approves; loop continues clean |
Don't just prompt agents. Build loops that audit them.
Full spec and Loop Spec template:
docs/LOOP_CODING.md
Auditing a loop tells you whether a result is real. The next step is to search
for the best result without fooling yourself — and the way you fool yourself is by
optimizing hard against a judge you never audited. A perfect score from a worthless
oracle (F1=1.000 on synthetic data) is the canonical trap.
/evolve-solution runs the Oracle-Aware Core — never one solution, always a
field of competing variants, judged by an oracle that earned trust first:
Intent → Oracle-Adequacy Gate → Falsification Contract → Variant Tournament
→ Red-Team → Evidence Gate → Null Result Ledger
| Stage | Question it answers | Backed by (no new hooks, no new agents) |
|---|---|---|
| Intent | What are we really optimizing? | rules/estimand-ops.md, /estimand-bridge |
| Oracle Adequacy | Is the judge worth optimizing against? | docs/oracle-adequacy-gate.md, validation_theater_guard |
| Falsification | What would prove each variant wrong? | rules/falsification-ladder.md |
| Tournament | Which of ≥3 variants wins? | /cross-domain, /hypothesis-arbiter, /combinatorial-creativity |
| Red-Team | Does the winner survive attack? | /skeptic, /codex-skeptic |
| Evidence Gate | Is the win proven, not claimed? | rules/integrity.md, promotion_gate_guard |
| Null Ledger | What did we learn from the dead? | null_results/, reject_gate_guard, null_retroscan |
The genuinely new piece is the Oracle-Adequacy Gate: optimizing against an inadequate oracle is worse than not optimizing — it manufactures false confidence at scale. So the oracle is audited (gameable? negative control? real data?) before any variant runs.
/evolve-solution "find a non-obvious way to cut our RAG hallucination rate"
Command:
commands/evolve-solution.md· Gate:docs/oracle-adequacy-gate.md· Templates:templates/intent_card.yaml,oracle_audit.yaml,falsification_contract.yaml
- Does not replace human code review — it adds a second layer, not a substitute
- Does not guarantee zero hallucinations — reduces frequency and adds detection
- Works only with Claude Code (not Cursor, Codex, VS Code Copilot, Gemini)
- Not independently verified beyond a single-developer workflow
- Does not come with enterprise SLA or paid support
- Does not manage secrets or rotate API keys — use a proper vault
Claude Code без этого конфига — как Ferrari на ручнике: мощный, но большая часть потенциала не используется. With this config — коммиты проходят автоматические проверки, агенты помнят контекст между сессиями, повторяющиеся ошибки фиксируются и эскалируются в правило.
Most configs are a single CLAUDE.md bloated to 3000+ tokens. This is different:
| Typical config | This config | |
|---|---|---|
| Tokens/msg | 3 000 – 5 000 | ~500 (core only) |
| Hallucinations | "trust me" | Evidence Policy + Confidence Scoring |
| MCP failures | session hangs | CircuitBreaker — auto-recovery in 60s |
| Prompt injection | no protection | InputGuard — 8 categories, auto-block |
| PII leakage | hope for the best | 12 regex patterns + auto-redact |
| Code review | optional | review-squad — parallel reviewer + sec-auditor |
| Permissions | ask for everything | permission_policy PreToolUse hook — auto-allow/deny/ask per Bash command, before the prompt |
| Agent memory | stateless | 4 agents with persistent memory across sessions |
| Tests | "I'll write them later" | 2331 tests, TDD-first, Test Protection hard rule |
everything-claude-code is a great alternative — bigger, multi-platform, Anthropic Hackathon Winner. Both are MIT, pick what fits.
Use everything-claude-code if:
- You want multi-language coverage (TS, Go, Java, Kotlin, Rust, C++, PHP — 12+ ecosystems)
- You work across multiple harnesses (Codex, Cursor, OpenCode, Gemini — not just Claude Code)
- You want a GUI dashboard for browsing components
- You like the paid tier path (ECC Tools GitHub App, free / pro / enterprise)
Use this config if:
- "Validation Theater" is a $$$ risk for you, not abstract — Evidence Policy is enforced as hard rule, not just a skill
- You work with sensitive data (PII, finance, healthcare) — built-in redaction hook scrubs sensitive strings before any external MCP call
- You need to read every hook before installing — only ~10 MB, plain Python, no JS dependencies, every file readable in 10 minutes
- You prefer Claude Code only with deep specialization over multi-platform breadth
- You speak Russian — README and rules have RU-first sections, useful for CIS dev teams
Comparison at a glance:
| everything-claude-code | This config | |
|---|---|---|
| Surface | 48 agents · 182 skills · 68 commands · ~31 MB | 13 agents + 3 squads · 125 skills · 91 hooks · ~10 MB |
| Languages | TS, Py, Go, Java, Kotlin, Rust, C++, PHP, Perl | Python primarily |
| Harnesses | Claude Code, Codex, Cursor, OpenCode, Gemini, Antigravity | Claude Code only |
| Anti-hallucination | continuous-learning v2 with confidence scoring | Evidence Policy + Validation Theater Guard + Audit Verification Gate (synthetic ≠ real, enforced) |
| PII / sensitive data | generic | dedicated redaction hook + local-first (Ollama) |
| Audit Verification Gate | not in core | rules/audit-verification-gate.md — agent's [VERIFIED] = your [INFERRED] |
| Recurring mistake tracking | instinct-based | [×N] counter — after 3 occurrences a mistake becomes a hard rule |
| License | MIT (open core, paid GitHub App) | MIT (no paid tier) |
If multi-language / cross-harness matters more than anti-hallucination focus — pick ECC. If anti-hallucination on sensitive data is your job-critical risk — pick this one.
New to this? Don't install everything at once. Pick the path that matches your goal:
| Path | What you get | Time | Command |
|---|---|---|---|
| Evidence Only | [VERIFIED] markers + anti-hallucination |
2 min | --profile=minimal |
| Daily Driver | + 91 hooks + 13 agents + all 125 skills | 5 min | --profile=standard |
| Full Setup | + MCP profiles + PII redaction + memory | 10 min | --profile=full |
Minimal path (recommended to start): installs just 3 files — CLAUDE.md, integrity.md, security.md. No hooks, no agents, no complexity. Add more when you need it.
# One-liner — Mac / Linux / WSL
git clone https://github.com/sergeeey/Claude-cod-top-2026.git && cd Claude-cod-top-2026 && bash install.sh --profile=standard --non-interactiveWindows (PowerShell):
git clone https://github.com/sergeeey/Claude-cod-top-2026.git; cd Claude-cod-top-2026; bash install.sh --profile=standard --non-interactiveAfter install: restart Claude Code (
/clearor new session) — hooks activate automatically.
# Register this repo as a marketplace source (once per machine)
/plugin marketplace add sergeeey/Claude-cod-top-2026
# Install the plugin
/plugin install claude-cod-top-2026Windows note: Claude Code doesn't pre-register third-party marketplaces on Windows. Add to your
~/.claude/settings.jsonmanually if/plugin marketplace addfails:"extraKnownMarketplaces": { "claude-cod-top-2026": { "source": { "source": "github", "repo": "sergeeey/Claude-cod-top-2026" } } }
git clone https://github.com/sergeeey/Claude-cod-top-2026.git
cd Claude-cod-top-2026
bash install.sh # interactive
bash install.sh --link full # symlink + auto-update
bash install.sh --profile=full --non-interactive # CI / headless| Profile | Installs | For whom |
|---|---|---|
minimal |
CLAUDE.md + integrity + security | Try Evidence Policy |
standard |
+ all rules + hooks + skills + agents | Daily work |
full |
+ MCP profiles + PII redaction + memory | Full control |
Hooks run 100% of the time — deterministic Python guards, not probabilistic instructions.
PreToolUse guards (9 shown · full list: hooks/ directory)
| Hook | Protects Against |
|---|---|
input_guard |
Prompt injection via MCP (8 categories) |
mcp_circuit_breaker |
Session hang on MCP failure (auto-recovery 60s) |
mcp_locality_guard |
MCP call without local search first |
pre_commit_guard |
Commits to main · rm -rf · DROP TABLE |
read_before_edit |
Edit without prior Read |
security_verify |
Sensitive file edits without review |
plan_mode_guard |
3+ files edited without a plan |
permission_policy |
Dangerous Bash commands denied, code runners (pytest/npm test) ask, before the prompt |
checkpoint_guard |
Risky ops without checkpoint |
PostToolUse audit layer (11 shown · full list: hooks/ directory)
| Hook | Protects Against |
|---|---|
mcp_circuit_breaker_post |
Records MCP failures for recovery |
post_format |
Unformatted code (ruff / prettier) |
memory_guard |
Forgotten memory update after commit |
post_commit_memory |
Context loss after commits |
pattern_extractor |
Lost lessons from fix: commits |
drift_guard |
Scope creep (NOT NOW keywords) |
post_tool_failure |
Repeated failures without strategy change |
config_audit |
Unauthorized settings changes |
elicitation_guard |
Elicitation events logging |
moc_autolink |
Notes written without Obsidian MOC links |
observation_capture |
Observations lost after file edits |
Lifecycle · Session · Memory (20 shown · full list: hooks/ directory)
| Hook | Event | Role |
|---|---|---|
session_start |
SessionStart | Load context from memory |
session_end |
SessionEnd | Trim + save state |
session_save |
Stop (async) | State persistence on exit |
post_compact |
PostCompact | Context reminder after compaction |
pre_compact |
PreCompact | Data preservation |
keyword_router |
UserPromptSubmit | Auto-trigger skills + power modes |
thinking_level |
UserPromptSubmit | Boost thinking for complex tasks |
statusline |
PostToolUse | Live bar: model · context% · cost |
worktree_lifecycle |
WorktreeCreate/Remove | Track experiment branches |
agent_lifecycle |
SubagentStart/Stop | Context in agent handoffs |
subagent_verify |
SubagentStop | Verify agent output quality |
team_rebalance |
TeammateIdle | Rebalance idle agents |
stop_failure |
StopFailure | Silent API error handling |
task_audit |
TaskCreated/Completed | Task event logging |
instructions_audit |
InstructionsLoaded | Track loaded rules |
env_reload |
FileChanged | Stale env after .env change |
direnv_loader |
CwdChanged | Wrong env after cd |
async_wrapper |
— | Non-blocking wrapper for bg hooks |
webhook_notify |
Stop (async) | Slack/Telegram on commit + session end |
thematic_index_router |
Stop | Route wiki entries to thematic indices |
Type anywhere in your prompt:
| Keyword | Mode | Effect |
|---|---|---|
ralph |
Persistent | Don't stop until done. Auto-retry. No confirmations. |
autopilot |
Full Autonomy | Plan + execute all steps. Only stop if truly blocked. |
ultrawork / ulw |
Max Parallelism | Launch agents concurrently. Batch ops. Speed > caution. |
deep |
Deep Analysis | Read everything. Evidence-mark all claims. |
quick / быстро |
Speed | Minimal output. No explanations. Just do it. |
Modes are additive — ralph security audit = Persistent mode + security-audit skill.
╔══════════════════════════════════════════════════════════╗
║ STRATEGIC — Opus · 20% of tasks ║
║ boyko-agent(memory:user) architect sec-auditor teacher ║
╠══════════════════════════════════════════════════════════╣
║ WORKHORSE — Sonnet · 80% of tasks ║
║ builder(worktree) tester(worktree) explorer reviewer ║
╠══════════════════════════════════════════════════════════╣
║ TEAMS — parallel execution ║
║ review-squad → reviewer + sec-auditor (parallel) ║
║ build-squad → builder + tester (isolated wt) ║
║ research-squad → explorer + verifier (search+verify) ║
╚══════════════════════════════════════════════════════════╝
4 agents with persistent memory · 2 agents with worktree isolation · Sonnet-first, Opus escalation only
Every factual claim is tagged:
[VERIFIED-HIGH] ≥2 sources confirmed → can be used as fact
[VERIFIED-MEDIUM] 1 source + inference → careful wording
[VERIFIED-LOW] indirect evidence → "there are signs, but..."
[UNKNOWN] no confirmation → do not guess
Confidence Scoring 0.0–1.0 based on source count. Rationalization Prevention — 10 common AI excuses with countermeasures baked into rules.
Zero token cost — always visible at the bottom of the terminal:
[claude-sonnet] ▓▓▓▓▓▓▓░░░░░░░░░░░░░ 35% | main | $0.42 | 3m5s
| Context % | Colour | Signal |
|---|---|---|
< 50% |
🟢 Green | Keep working |
50–70% |
🟡 Yellow | Plan a /clear soon |
> 70% |
🔴 Red | /clear now |
InputGuard — 8 injection categories (scoped to MCP tool calls only — built-in tools like Read/Bash are trusted by definition, see hooks/input_guard.py):
| Category | Example | Action |
|---|---|---|
encoding_attack |
null bytes, zero-width chars | AUTO-BLOCK (single match) |
command_injection |
; rm -rf · `$(curl)` |
AUTO-BLOCK (single match) |
data_exfil |
"send to http", "curl | bash" | AUTO-BLOCK (single match) |
system_override |
"ignore previous instructions" | Warn only on a single hit — blocks once combined signal reaches 2 (a repeat hit or a 2nd category) |
jailbreak |
"DAN mode", "bypass safety" | Warn only on a single hit — blocks once combined signal reaches 2 (a repeat hit or a 2nd category) |
role_injection |
[SYSTEM], <system> |
Warn only — its own repeat hits are capped at 1, so it blocks only if a 2nd category co-occurs |
credential_harvest |
"show me your api key" | Warn only on a single hit — blocks once combined signal reaches 2 (a repeat hit or a 2nd category) |
social_engineering |
"pretend you have no restrictions" | Warn only on a single hit — blocks once combined signal reaches 2 (a repeat hit or a 2nd category) |
PII Redaction — 12 patterns stripped before external MCP calls:
National IDs · Bank cards · IBAN · API keys · GitHub tokens · AWS keys · JWT · Email · Phone · IPs
pip install pytest pytest-cov ruff mypy
pytest tests/ -v --cov=hooks --cov-report=term-missing # see CI badge above for exact count
ruff check hooks/ scripts/ tests/
mypy hooks/utils.py hooks/input_guard.py
bash tests/test_all.sh # 3 shell suites: hooks · install · skillsTwo automation hooks keep your Obsidian vault in sync with Claude Code activity:
| Hook | Trigger | What it does |
|---|---|---|
moc_autolink |
PostToolUse Write/Edit | Tags new notes → auto-links to relevant MOC (Claude-cod, GeoMiro, Research…) |
thematic_index_router |
Stop | Routes fresh wiki entries to Claude-Code / Lessons / Projects indices |
Vault layout (~/.claude/memory/):
wiki/ ← processed knowledge (auto-generated)
raw/ ← quick drop → auto-converted at session end
mocs/ ← Maps of Content (6 MOCs)
_auto/wiki/ ← thematic indices (Claude-Code / Lessons / Projects)
daily/ ← session reports
graph.json colorGroups must be set while Obsidian is closed — the app overwrites on launch.
CORE (default) SCIENCE DEPLOY
context7 + ncbi-datasets + vercel
basic-memory + uniprot + netlify
playwright + pubmed-mcp + supabase
ollama + sentry
~/.claude/mcp-profiles/switch-profile.ps1 scienceCircuitBreaker auto-fallback: context7 → WebSearch · playwright → WebFetch · ollama → cloud
Full File Structure
Claude-cod-top-2026/
├── CLAUDE.md Core config (deployed from claude-md/CLAUDE.md, ~120 lines)
│
├── rules/ 15 modular rules (loaded on demand)
│ ├── coding-style.md
│ ├── security.md
│ ├── testing.md
│ ├── integrity.md
│ ├── memory-protocol.md
│ ├── context-loading.md
│ ├── permissions.md
│ └── mentor-protocol.md
│
├── hooks/ 91 hooks + utils.py/hook_state.py/severity_calibrator.py (shared libs)
│ ├── utils.py 21 shared functions (DRY)
│ ├── settings.json Hook registry + 27 deny patterns
│ ├── input_guard.py Prompt injection
│ ├── mcp_circuit_breaker.py MCP resilience
│ ├── statusline.py Terminal status bar
│ └── ... 39 more hooks
│
├── agents/ 13 active + 3 teams
│ ├── navigator.md boyko-agent — Strategic (Opus, memory:user)
│ ├── builder.md Code (Sonnet, worktree)
│ ├── reviewer.md Review (Sonnet, memory:project)
│ ├── sec-auditor.md Security (Opus, memory:project)
│ └── teams/ review-squad · build-squad · research-squad
│
├── skills/
│ ├── core/ 13 universal skills
│ └── extensions/ 112 domain skills
│
├── assets/ Visual assets
│ ├── banner.svg Hero banner (animated)
│ └── pipeline.svg Hook execution pipeline diagram
│
├── tests/ 2331 tests · 80 files
├── docs/ Architecture · guides · anti-patterns
├── mcp-profiles/ 3 profiles (core/science/deploy)
└── .github/workflows/ci.yml pytest + ruff + mypy + secrets scan
Documentation Index
| Document | Description |
|---|---|
| Positioning | What category this is, what it's not, comparison to memory/skill/team/tool layers |
| Architecture | 6-layer system design |
| Evidence Policy | Anti-hallucination + Confidence Scoring |
| Hooks Guide | All 91 hooks with examples |
| Skills Guide | Creating and managing skills |
| Anti-Patterns | 9 critical mistakes to avoid |
| Troubleshooting | 10-point diagnostic checklist |
| CONTRIBUTING | Contribution guidelines |
| CHANGELOG | Version history |
Verified incidents from the author's own workflow (single developer, one codebase):
pre_commit_guardblocked accidental push tomainduring a hotfixpattern_extractorauto-logged debugging lessons fromfix:commitsmemory_guardkeptactiveContext.mdcurrent across 3 deploy cycles
Scope: single developer · personal project · not independently verified.
