# CyberBrain26 – Complete Diagnostic Plan **Plan ID:** `1793687720757-cyberbrain-complete-diagnostic` **Target:** Orchestrator autonomous execution **Authoritative Plan:** `.kilo/plans/CyberBrain26cil-031026-v2.md` (approved 2026-10-03) **Gate Target:** Gate 1 verified → Gate 2 E2E flow evidence --- ## Goal Produce a complete, evidence-based diagnostic of CyberBrain26 current state: - What works, what doesn't, with [OVĚŘENO]/[NÁVRH]/[HYPOTÉZA] markers - How the security chain (identity → policy → permission → approval → sandbox → audit) actually operates - Model gateway routing (local vs cloud), EgressGate status - Kilo integration status (config, agents, permissions) - All open security gaps per PROJECT_STATUS.md - Artifacts: markdown report, JSON health_check, structured evidence index --- ## Tasks (ordered, with dependencies) ### T0 – Baseline Capture - **Action:** Read `PROJECT_STATUS.md`, `gate_report.txt`, `.kilo/plans/CyberBrain26cil-031026-v2.md`, `AGENTS.md`, `.kilo/kilo.jsonc` - **Output:** Baseline summary (HEAD commit, uncommitted files, test counts, phase position) - **Validation:** Matches `gate_report.txt:2` and `PROJECT_STATUS.md:83-109` ### T1 – Test Suite Execution & Evidence Collection - **Action:** Run full test suite, capture output - **Commands (via cyberbrain run):** - `PYTHONPATH=src python3 -m pytest -v --tb=short 2>&1 | head -200` - `PYTHONPATH=src python3 -m pytest -q tests/test_kilo_agent_policy.py` - `python3 -m ruff check --line-length=110 src/cyberbrain26/ tests/ .kilo/` - `PYTHONPATH=src python3 -m mypy --cache-dir=/tmp/cb26 src/cyberbrain26` - **Output:** Test counts, pass/fail, coverage, lint/type status - **Validation:** ≥ 591 tests pass, ruff clean, mypy 0 errors on 70+ files ### T2 – Security Chain Verification - **Action:** Read security plane modules and their tests - **Files:** - `src/cyberbrain26/security_plane/identity/__init__.py` - `src/cyberbrain26/security_plane/policy/__init__.py` - `src/cyberbrain26/security_plane/permission/__init__.py` - `src/cyberbrain26/security_plane/approval/__init__.py` - `src/cyberbrain26/execution_plane/secure_gateway.py` (AuthorizedExecutionService) - `src/cyberbrain26/execution_plane/bubblewrap_provider.py` - `src/cyberbrain26/security_plane/audit/__init__.py` - **Tests:** `tests/test_security_*.py`, `tests/test_security_e2e.py`, `tests/test_execution_provider.py` - **Output:** Evidence map for each link in chain, fail-closed behavior proof - **Validation:** Each step has positive + negative test; `ExecutionRefused` on missing approval ### T3 – Model Gateway & Local Model Status - **Action:** Read model gateway, router, clients; run health_check - **Files:** - `src/cyberbrain26/control_plane/model_gateway/__init__.py` - `src/cyberbrain26/model_layer/router.py` - `src/cyberbrain26/model_layer/llama_cpp_client.py` - `src/cyberbrain26/control_plane/model_gateway/ollama_client.py` - `src/cyberbrain26/control_plane/model_gateway/litellm_client.py` - **Command:** `PYTHONPATH=src python3 -c "from cyberbrain26.control_plane.model_gateway import ModelGateway; g=ModelGateway(); import json; print(json.dumps(g.health_check(), indent=2))"` - **Output:** JSON health_check, route decision for local-first, EgressGate wiring status - **Validation:** `prefer_local=True`, `free_only=True`, Ollama/llama.cpp providers registered ### T4 – Kilo Integration Status - **Action:** Read Kilo config and agent declarations - **Files:** - `.kilo/kilo.jsonc` (subagent_depth, default_agent, permissions, MCP disabled) - `.kilo/agent/promptmaster.md`, `.kilo/agent/orchestrator.md` - `.kilo/agent/verifier.md`, `.kilo/agent/code-reviewer.md`, etc. - **Output:** Config verification, delegation matrix, MCP status - **Validation:** `subagent_depth: 1`, `default_agent: "promptmaster"`, `bash: deny`, `task: deny` globally, MCP `enabled: false` ### T5 – Lifecycle & Persistence - **Action:** Read control plane (task_queue, approval), audit, SQLite - **Files:** - `src/cyberbrain26/control_plane/task_queue/__init__.py` - `src/cyberbrain26/control_plane/approval/__init__.py` - `src/cyberbrain26/security_plane/audit/__init__.py` - `src/cyberbrain26/data_plane/sqlite_store.py` - **Tests:** `tests/test_control_plane.py`, `tests/test_sqlite_persistence.py`, `tests/test_audit_chain.py` - **Output:** Durable job handling, approval flow, audit chain verification, prune_expired safety - **Validation:** 12/12 mutation tests fail, `prune_expired` writes chained event ### T6 – Open Security Gaps Inventory - **Action:** Extract from `PROJECT_STATUS.md` lines 203-218 and cross-reference with code - **Output:** Structured list with code references for each gap - **Validation:** Each gap has file:line or test showing why it's open ### T7 – Phase Position & Gate Status - **Action:** Read `.kilo/plans/CyberBrain26cil-031026-v2.md` §5 and §8 - **Output:** Current phase (F2), Gate 1/2/3 criteria, what's blocking next gate - **Validation:** Matches `corrections.md` memory: "Currently at F2 (minimal complete agent flow...)" ### T8 – Produce Diagnostic Artifacts - **Action:** Write three files 1. `YYYYMMDD-cyberbrain-diagnostic.md` – full report per template in prompt 2. `YYYYMMDD-model-gateway-health.json` – ModelGateway.health_check() output 3. `YYYYMMDD-evidence-index.txt` – structured list: component | status | file:line | test - **Validation:** All [OVĚŘENO] claims have file:line or test name; [HYPOTÉZA]/[NÁVRH] clearly separated --- ## Dependencies ``` T0 → T1, T2, T3, T4, T5, T6, T7 (all depend on baseline) T1, T2, T3, T4, T5, T6, T7 → T8 (artifacts need all evidence) ``` --- ## Validation Criteria (Exit Gates) | Gate | Criterion | How Verified | |------|-----------|--------------| | Gate 1 | All quality gates green | T1 output matches `gate_report.txt` | | Gate 1 | Security chain fail-closed | T2: each link has negative test | | Gate 1 | Model gateway local-first | T3: health_check shows local providers | | Gate 1 | Kilo config enforced | T4: subagent_depth=1, bash=deny, MCP disabled | | Gate 2 | E2E diagnostic complete | T8: artifacts exist with evidence refs | | Gate 2 | No silent cloud fallback | T3: free_only=True, EgressGate opt-in | --- ## Risk / Open Questions | # | Question | Recommended Resolution | |---|----------|------------------------| | R1 | Is Ollama/llama.cpp actually running on ports 11434/8080? | T3 health_check will show; if not running, mark as [HYPOTÉZA] "configured but not verified live" | | R2 | Should we attempt a live model call to prove routing? | No – diagnostic is read-only; live call would need approval + EgressGate | | R3 | Are there uncommitted changes in src/ that affect diagnostic? | T0 captures git status; T1 runs against current working tree | | R4 | Does `cyberbrain run` work for executing the health_check command? | Yes – it's the authorized execution path (D4). Use it for T1, T3 commands. | --- ## Out of Scope - Fixing any discovered gaps (that's F3+ work) - Implementing EgressGate wiring (F4) - Customer discovery (F5 blocker) - Adding cgroup/memory quotas (F6) - Any source code modifications --- ## Execution Notes for Orchestrator - Use `cyberbrain run` for all command execution (authorized path) - Delegate read_file tasks to appropriate subagents (verifier, security-auditor, deep-research) - Each subagent produces evidence; orchestrator aggregates into T8 artifacts - Maintain [OVĚŘENO]/[NÁVRH]/[HYPOTÉZA] discipline throughout - Final report must be readable by owner without code access – include file:line refs for every claim