An honest separation of what Kazma does today from what is planned, aspirational, or partially wired. Every row re-checked against the code on 2026-09-27.
Items are marked:
- ✅ Implemented & wired — works in the default runtime, verified in code.
- 🟡 Implemented but not fully wired — code exists but isn’t connected in the default path.
- 🔴 Planned / Roadmap — not in the codebase, or declared as a goal in
ROADMAP.md.
| Capability | Status | Notes |
|---|
| LangGraph supervisor ReAct loop | ✅ | graph_builder.py. |
| Tool calling with OpenAI-compatible providers | ✅ | httpx, no SDK. |
| NVIDIA NIM tool-fallback | ✅ | llm_provider.py:285-300. |
| Strict tool JSON Schema | ✅ | additionalProperties: false always; KAZMA_STRICT_TOOLS=1 for OpenAI function.strict. |
Structured outputs (response_format) | ✅ | Opt-in on LLMProvider.chat / chat_stream; not forced on supervisor turns. |
| Pre/Post tool hooks | ✅ | agent/tool_hooks.py. Deny/rewrite/observe. Cannot skip HITL. KAZMA_TOOL_HOOKS=0. |
| First-class plan mode | ✅ | /plan on · /plan go. Structural read-only, then execute. KAZMA_PLAN_MODE=0. |
| Streaming (SSE) | ✅ | chat_stream() + invoke_llm_chat(); SSE/WS consume synthetic on_chat_model_stream. |
| Context compaction (LLM summarise) | ✅ | compaction.py. |
| Compaction with memory retrieval + checkpoint | 🟡 | Memory adapter wired on main paths; checkpoint_manager still optional. |
| Rate-limit (429) handling | ✅ | Exponential backoff + Retry-After in llm_provider.py and native Anthropic /messages. Exhausted 429 is transient=True + kind=rate_limit_exhausted (no same-provider re-retry). |
| Cost breaker auto-wired | ✅ | CostCircuitBreaker instantiated per-agent (agent_runner.py) and driven on the live loop — record_user_interaction() on each inbound message, should_halt() gate, and record_cost() after each LLM call (graph_builder.py). Exposed on the dashboard via .status(). |
Updated 2026-09-27 — every memory findable, evidence-ranked recall, weekly summaries, About me, forgetting, the Postgres mirror kept whole.
Backlog: docs/plans/MEMORY_REMAINING.md. Full guide: Memory & RAG.
| Capability | Status | Notes |
|---|
| V2 cognitive engine | ✅ | Bi-temporal belief graph + 4-tier episodes + procedural DAGs. Single memory stack (V1 removed). |
| Per-turn RAG | ✅ | V2 recall() (beliefs + episodes + PPR) every user turn. |
| Compaction memory inject | ✅ | V2 recall feeds a compaction. Its summary is not stored in memory (2026-09-27): every turn already is a memory, and the weekly summaries hold the gist. |
| Swarm memory bridge | ✅ | Worker results + SoulEvolution written to V2. |
| Bi-temporal belief graph | ✅ | Functional/set/state predicates; valid_until/invalidated_at. |
| Local Ego-Graph PPR | ✅ | 2-hop, N≤200, α=0.15 recall boost. |
| Durable consolidation queue | ✅ | memory_ops.db task queue + 6h macro_sleep + 6h backup/export + 24h reconsolidation. |
| Procedural action DAGs | ✅ | Laplace-smoothed skill confidence C(d)=(S+1)/(N+2). |
| Backup + export | ✅ | Native sqlite3.backup() + JSONL/GraphML every 6 h; restic snapshots, local and offsite. |
| Arabic tokenizer (FTS5) | ✅ | V2 episode FTS5 + symmetric normalization. |
| Multi-install memory | 🟡 | Memories and facts are mirrored to Postgres and kept whole by an anti-entropy sync; recall reads what other installs wrote. Vectors and the graph stay local to each install. |
| Evidence-ranked recall | ✅ | A memory is shown only when its meaning and words clear per-kind thresholds set on a benchmark (memory/benchmark.py); nothing is injected when nothing matches. |
| Weekly topic summaries | ✅ | One summary per topic per week, written by the model, recalled after the history. |
| About me | ✅ | Settings → Memory: text the user writes, read at the start of every reply. |
| Forget, don’t remember a chat, export | ✅ | Tombstones and a forget ledger every writer asks; /memory off on chat platforms; GET /api/memory/v2/export. |
| Memory used, per answer | ✅ | Each turn stores which memories, facts and summaries the model was given, shown in its activity. |
checkpoint_manager in compaction | 🟡 | Still optional — LangGraph checkpointer covers turns. |
| Capability | Status | Notes |
|---|
| Six dispatch patterns | ✅ | dispatch/broadcast/pipeline/fan-out/consult/conditional. |
| Aggregation (collect/first_valid/merge_all/vote/synthesize) | ✅ | aggregator.py. |
| Circuit breakers (half-open single-probe) | ✅ | reliability.py. |
| Retry / timeout / output validation / bounded concurrency | ✅ | reliability.py. |
| Pipeline HITL checkpoints with auto-reject timeout | ✅ | checkpoint_manager.py. |
| Handoff cycle detection (depth 5, visits 2) | ✅ | handoff_guards.py. |
| Worker autoscaling | ✅ | dispatch_inner spawns a worker from swarm_templates.json when none matches; idle workers are reaped after 5 minutes. |
| Prometheus metrics | ✅ | Optional prometheus-client extra; /metrics endpoint in kazma_ui/metrics.py. |
| Capability | Status | Notes |
|---|
| Graph HITL gate (interrupt) | ✅ | Active on all production build sites. |
| Swarm bus HITL gate (fail-closed) | ✅ | swarm/safety.py. |
| Pipeline checkpoint HITL | ✅ | checkpoint_manager.py. |
| Skill HMAC signing + verification | ✅ | hub/cli.py + hub/loader.py. |
| Delegation Ed25519 + AES-GCM | 🔴 | Not in the code. Skills are HMAC-verified at load; there is no signed delegation between agents. |
| MCP SSE bearer auth | ✅ | mcp/manager.py:461-466. |
| MCP stdio auth | ✅ | auth.type: env / arg injection supported on stdio servers. |
| Vault-backed ConfigStore secrets | ✅ | Sensitive keys → AES vault when KAZMA_VAULT_KEY set (2026-07 audit remediations). |
/undo / /edit checkpoint mutation | ✅ | Live graph path via aget_state / aupdate_state. |
| Remote secret login page | ✅ | /login + POST /api/auth/login. |
| Cryptographic “trust tiers” | 🔴 | Only the hub’s boolean certified flag and an MCP server’s trust: trusted (skips its approval card). |
kazma-security.yaml hardening checks | 🔴 | No code reads that file; it is a declaration. What enforces security lives in the code and CI: the approval gates, the commitment layer, the vault, the static gates in tests/, bandit in CI and Dependabot. |
| Capability | Status | Notes |
|---|
| Telegram adapter (full-featured) | ✅ | Long-poll + optional webhook, voice, reactions, keyboards. |
| Discord adapter | ✅ | Gateway WebSocket. |
| Slack adapter | ✅ | Socket Mode / polling. |
| Web UI (SSE) | ✅ | /api/chat/stream. |
| WebSocket chat | ✅ | /ws/chat/{session_id} is telemetry / cursor resume; SSE /api/chat/stream is the graph transport (the WS second graph client was removed 2026-09-30). |
| TUI | ✅ | Textual, read-mostly. |
| EN/AR i18n + RTL | ✅ | Catalog-merged dict, IBM Plex Sans / IBM Plex Sans Arabic, shared 14px root. |
X Studio (/x) | ✅ | Composer + X-only planner (Post now, Schedule, reschedule, threads, delete). Chat x_post stays always-HITL. Official API only. |
| Majlis protocol | 🟡 | A greeting and farewell fast path on the chat platforms; ordinary conversation does not enter its phase machine (majlis.py). |
| Voice on Discord/Slack/Web | ✅ | STT + TTS wired into all platforms via voice_helpers.py (was Telegram-only). |
| Media / attachments (photo/doc/video) | ✅ | Attachment contract on IncomingMessage/OutboundMessage; inbound+outbound on all platforms + Web /api/chat/upload. |
/undo, /edit slash commands | ✅ | Handled by the graph (_handle_undo/_handle_edit mutate checkpoint state). |
| Time-travel replay & fork | ✅ | SnapshotRecorder wired into all graph-build sites; /replay restore + /fork branch + Web UI /replay timeline panel + live SSE snapshot events + /api/replay/*. |
| Capability | Status | Notes |
|---|
| Provider presets (20: OpenAI-compatible, native, local, custom) | ✅ | providers.py — incl. Mistral/Together/Cohere/Fireworks/Perplexity/AI21/Groq/xAI/OpenRouter/NVIDIA. |
| Native non-OpenAI providers | ✅ | AnthropicProvider (/messages), AzureProvider (api-key+api-version), BedrockProvider (SigV4 + Converse), GeminiProvider (ADC). See LLM Providers. |
| Google Vertex AI (ADC) | ✅ | google_llm.py. |
| Local servers (Ollama/LM Studio) | ✅ | Dummy-key handling. |
| MCP (stdio + SSE + Streamable HTTP) | ✅ | mcp/manager.py — Streamable HTTP (MCP 2025-03-26 spec) with Mcp-Session-Id resumption. Resources/prompts/sampling/roots client surfaces (2026-08-25); sampling is HITL fail-closed. Kazma is also an MCP server: kazma mcp (stdio) offers its tools to other agents, danger tools behind the same approval gate. |
| Skill Hub (registry, signing, certification) | ✅ | hub/. |
| Langfuse tracing | ✅ | KazmaTracer; logging.langfuse.enabled: auto when keys exist. KAZMA_LANGFUSE=0. |
| OpenTelemetry | ✅ opt-in | GenAI spans for every LLM call and tool execution once opentelemetry-sdk is installed and an OTLP endpoint is set (ops guide). |
| Cloudflare Pages / edge | 🔴 | Not applicable — stateful Python service. |
| PostgreSQL (main agent) | ✅ | First-class backend for ConfigStore/sessions/swarm/checkpoints; HITL pending-approvals enumerate Postgres threads (hitl_approval.py). |
| Capability | Status | Notes |
|---|
| Structured JSON logs | ✅ | logging.format: json. |
| Swarm metrics (in-memory + SQLite) | ✅ | MetricsCollector. |
| In-house tracing spans | ✅ | TraceStore (dashboard) + TracingEmitter (swarm). |
| SSE telemetry events | ✅ | /api/chat/stream + telemetry router. |
| Langfuse tracing | ✅ | Wired via KazmaTracer; auto-on when keys exist (logging.langfuse.enabled: auto). |
| Prometheus scrape endpoint | ✅ | /metrics + /api/metrics in kazma_ui/metrics.py, mounted in app.py (gateway-active block). Emits text/plain; version=0.0.4 with inbound/outbound/error counters, active threads, adapter, queue-depth, and swarm gauges. |
| OpenTelemetry export | ✅ opt-in | GenAI semantic-convention spans over OTLP (ops guide). Langfuse and the console remain. |
Memory (2026-07): strengthen + L2 graph + consolidator + graph UI are done.
Remaining memory polish/scale only in MEMORY_REMAINING.md.
Other open items:
- 429 backoff — done 2026-08-25 (generic + Anthropic; see leftover GOAL).
- OpenTelemetry — resolved: GenAI spans ship, opt-in (ops guide).
- Vectors — memory’s meaning search is exact and local (sqlite-vec, or NumPy), measured at about 1 ms per 1,000 memories. pgvector is used when the Postgres server has the extension (picked automatically) and Qdrant is a Settings choice; neither is needed. The Knowledge Library’s vectors live in Chroma and are rebuilt from its chunks.
- IDE chrome — CodeMirror 5
fromTextArea + file_apply_patch / file_apply_patch_set (Hands 0.11). Monaco was tried and reverted (hang / empty tabs). Codebase index (ripgrep + symbols) shipped 2026-08-25. /api/ide/lsp still exists; the Web editor is syntax-only. kazma ask + ACP stdio (live tokens, TTY HITL, session/request_permission, structured diffs, session/cancel).
- E2B + Temporal — opt-in adapters (industry stack part 8). Untrusted
python_exec via Firecracker; durable swarm steps via Temporal. Planner and HITL stay Kazma.
- This file intentionally resists over-promising. Where README/marketing copy describes a feature that is only partially wired, the status column says 🟡 with the specific reason.
- The “Suggested next steps” are the audit’s opinionated recommendations, prioritized by impact-to-effort ratio. They are not commitments.
- This file reflects code reality as of v0.11+, not marketing futures.
