✅ Done · 1 turn · 10095 in / 155 out · cache: 0cr+0rd+0ct · 2s — tools: none
+
✅ Done · 1 turn · 10095 in / 155 out · cache: 8143 write / 0 read / 0 total · 2s — tools: none
13:10
@@ -677,6 +736,52 @@
odek
+
+
+
+
+
What's new in v1.15
+
+ The biggest release since Extended Memory: the agent now anticipates, and cleans up after itself.
+ Full notes in the v1.15.0 release.
+
+
+
+
+
💡 Proactive Engagement v1.15.0
+
The agent ends answers with follow-up suggestions it already predicted, tracks your open loops (unanswered questions, stated goals), and can send opt-in proactive nudges — with hard anti-annoyance caps so it never becomes noise. Scheduled jobs are memory-wired, so a cron task can analyze your memory and sessions. docs ↗
+
+
+
🧹 Storage Maintenance v1.15.0
+
A janitor goroutine keeps long-lived agents (Telegram bot, serve) clean forever: session/audit/log/plan/media retention driven by an operator-only maintenance config. odek cleanup --dry-run previews exactly what would be deleted before anything is touched. docs ↗
+
+
+
🧠 Memory overhaul v1.15.0
+
Semantic dedup of stored atoms, relevance-preserving recall ranking, atom consolidation, index staleness detection, and odek memory extended stats for operators. Guard-rejected memories go to quarantine for human review instead of vanishing.
+
+
+
🛡️ PIGuard + guard fixes v1.15.x
+
The prompt-injection sidecar got its score semantics fixed (confident-benign memories no longer rejected), a JSON-faithful HTTP gateway, NDJSON socket transport, and Apple-Silicon-safe builds. Skill auto-learn is hardened against project-repo injection.
+
+
+
🗣️ Multilingual voice (STT) v1.15.2
+
Docker images ship the multilingual ggml-small whisper model — usable German, English, and Spanish voice transcription out of the box (was effectively English-only).
+
+
+
📊 Honest cache metrics v1.15.3
+
DeepSeek's native cache fields are parsed, and providers that report nothing show cache: n/a instead of misleading zeros.
+
+
+
+
The Five Whys
@@ -691,7 +796,7 @@
The Five Whys
Why?
- Because locking into one LLM provider is a strategic mistake. odek is provider-agnostic — Deepseek, OpenAI, Anthropic, Ollama, vLLM, Groq, Together, Fireworks. Any OpenAI-compatible endpoint.
+ Because locking into one LLM provider is a strategic mistake. odek is provider-agnostic — DeepSeek, OpenAI, Anthropic, Ollama, vLLM, Groq, Together, Fireworks. Any OpenAI-compatible endpoint.
Why?
@@ -817,6 +922,21 @@
See it in action
🏆 80.3% — highest published agent score
+
+
+ # Proactive engagement — the agent anticipates
+ $ odek run --session "I plan to migrate Postgres 14 → 17 next quarter"
+ Good call to plan early. Watch out for: logical replication
+ slots, removed GUCs, and driver upgrades…
+ ── You might also want to ──
+ • What are the breaking changes between PG 14 and 17?
+ • How do I test the migration with minimal downtime?
+ # … days later, via Telegram — an opt-in nudge:
+ 💡 Your Postgres migration goal has been quiet for a week —
+ want a pg_upgrade dry-run checklist?
+ # capped: max/day + per-kind cooldown. Never noisy.
+
+
@@ -848,8 +968,26 @@
Self-Learning Skills
💾
-
Persistent Memory
-
Three tiers: facts (markdown on disk), session buffer, and searchable episodes. Merge-on-write via go-vector cosine similarity saves ~80% LLM calls. go-vector RP is ephemeral — rebuilt from text on every write, no embedding state to persist.
+
Extended Memory
+
Semantic, atomic memory: facts, preferences, goals extracted per turn, deduplicated by meaning, ranked by relevance and recency. Quarantine for untrusted content, associations between atoms, and a stats surface for operators. Legacy three-tier facts/buffer/episodes included too.
+
+
+
+
💡
+
Proactive Engagement NEW
+
The agent anticipates: follow-up suggestions after answers, open-loop tracking of your questions and goals, and opt-in proactive nudges with anti-annoyance caps. Scheduled jobs are memory-wired, so cron tasks can analyze memory and sessions.
+
+
+
+
🧹
+
Storage Maintenance NEW
+
Long-lived agents stay clean forever: a janitor handles session, audit-log, log, plan, and media retention on an operator-only config. odek cleanup --dry-run shows exactly what would be removed, first.
+
+
+
+
🛡️
+
PIGuard Sidecar
+
Optional ML prompt-injection guard (ONNX, local) as a semantic second opinion on top of the rule-based scan — guarding memory writes, skill bodies, system prompts, and MCP descriptions. Docker Compose sidecar included.
@@ -891,7 +1029,7 @@
Parallel Tool Execution
💬
Interaction Modes
-
Three modes: engaging (emoji-rich LLM narration), enhance (narration persists after response), and verbose (raw tool traces). Per-tool trace messages with LLM reasoning content. Works across CLI, Telegram, and Web UI.
+
Four modes: engaging (emoji-rich LLM narration), enhance (narration persists after response), verbose (raw tool traces), and off (clean, machine-friendly). Per-tool trace messages with LLM reasoning content. Works across CLI, Telegram, and Web UI.
@@ -1013,7 +1151,13 @@
Commands
odek mcpStart MCP server (stdio)
odek mcp --sse-addr :8081Start MCP server (SSE)
odek telegramStart Telegram bot
-
odek run --deliver Run task, deliver to Telegram
+
odek schedule add <task>Schedule a recurring agent task (cron)