Skip to content

Changelog

Only user-visible changes are listed here. For the full commit log see the gitea repo (internal) or the GitHub mirror for public releases.

v0.26.0 — Silent behaviour made visible; verified provider layer

The theme is telling you what happened. An import that discarded 74% of command outputs said nothing; an auto-export reached you only if the model chose to mention it; olav doctor passed 9/9 on an install whose every query path was dead.

Features

  • olav doctor grew from 9 to 11 checks. auth names the mode and the provider that will serve it — a clean install could pass 9/9 with every query path dead, because init had written auth.mode=token whose provider ships in olav-ent. context compares the agent's context budget against the window your server actually advertises, which is what turns into agent error, code 400 on a long run.
  • olav doctor --verify proves a service works rather than merely answers, and each registered service contributes its own readiness check.
  • Import report. Ingest writes exports/import_reports/<snapshot>.md naming how much parsed, what did not, and why — raw_only, no_parser_registered, parser_no_match, or a parser error. On a 339-device dataset, 8,761 command outputs produced 2,278 structured rows; that 74% was previously invisible. Where the learner could close a gap the report says so and stops — it never runs it, because that would turn a two-minute import into a long one.
  • Query exports are announced. When execute_sql caps rows to fit the model's context and writes the full result to a file, the CLI states the row count, the path, and how much you were actually shown.
  • Provider support is now measured, not assumed. Each provider has a first-class driver with its own function-call determinism defaults, pinned by an offline check on the exact request it receives plus an opt-in live call.

Fixes

  • Queries selecting stored config text failed with 'str' object has no attribute 'items'. A per-field size budget was being applied to whole rows, so any wide row stopped being a row.
  • olav init no longer overwrites an auth mode you chose. It only selects a mode on a config it created in that run.
  • Claude rejected requests with a 400 when temperature exceeded 1.0, which is its documented ceiling. Temperature is now clamped per provider.
  • Memory capture ran twice per query, costing a second LLM round-trip and writing duplicate memories.
  • A repeated query left duplicate CSVs. Exports are content-addressed, so an identical result set reuses its file.
  • devices.role stayed empty when no inventory supplied it; it is now derived from the hostname where the name says so.
  • Two co-located deployments could adopt each other's containers — compose project names are namespaced.

Removed

  • Mistral provider support and its [mistral] extra. The setup wizard was offering a provider whose driver a plain install did not ship.

v0.25.1 — Packaging fix: authored workspace files now ship

Fixes

  • 16 authored workspace reference files were never packaged. Guides and references that shape agent behaviour existed in the repo but not in the wheel, so a pip install behaved differently from a source checkout.
  • render_report initialised an LLM unconditionally, so a report that needed no model still required one to be reachable.

Changed

  • Developer-facing text in the workspace prompt surface is English-only, so a mixed-language prompt no longer reaches the model.

v0.25.0 — Tiering: team features move to olav-ent

Changed

  • The web API and WebGUI moved to olav-ent. The open-source package is the CLI, runtime and agents; team features (multi-user auth, the web surface) are the enterprise extension.
  • auth.mode=none grants the local OS user admin rights. A single-operator install needs no login at all.
  • A plain install is smaller — fastapi and the ldap3 extra are gone from the OSS dependency set.

Features

  • One write path to DuckDB (open_write_connection), so concurrent writers no longer corrupt state or race each other.

v0.24.1 — Audit and export fixes

Fixes

  • Audit postcheck prompts suggested a command that does not exist (olav -a ops); they now name a real agent.
  • Batfish export falls back to built-in per-platform backup commands when a device has no stored config, instead of exporting nothing.

v0.24.0 — Loop protection, declarative workflows, honest doctor

Features

  • Multi-step tasks can be declared rather than improvised: known-shape work routes across sub-agents by declaration.
  • A zero-LLM circuit breaker on every agent. A model repeating the same tool call is stopped by Python, not by asking the model to notice.
  • olav doctor names the backends it found — model and endpoint for the LLM, the same for embeddings — so a misconfiguration is visible without a run.
  • Per-script timeouts can be declared in SKILL.md.

Fixes

  • The first-run wizard rejected the numbered choice it had just printed in the embedding menu.
  • API-mode embeddings read a config key that did not exist, so an endpoint you configured was silently ignored.
  • deploy_service reported healthy when a container had already exited — docker compose ps was being read without -a.
  • A single broken audit view failed the whole audit; jobs now degrade individually.

v0.23.1 — Metadata

Changed

  • Project URLs point at https://olavai.com and the documentation site; author and copyright metadata unified.

v0.23.0 — Self-reflection, scheduled jobs, recall benchmark

Also carries the changes internally tagged 0.22.2, which was never published.

Features

  • admin/reflector — a daily self-reflection pass that turns recurring error signatures into constraints the agents recall later.
  • olav cron — enable / disable / list for declared scheduled jobs, opt-in rather than implicit.
  • olav kb bench — measures whether the knowledge base can actually recall what was stored in it, without domain assumptions.
  • olav doctor grew from 3 to 8 zero-LLM checks, adding agents, tools and scaffolding.
  • draw_topology — a one-shot draw.io or Mermaid topology from the writer agent.
  • execute_sql caps per-cell text, so one 70 KB config column can no longer consume the model's whole context.
  • Batfish self-service — the simulator diagnoses a missing Batfish and offers to deploy it instead of failing.

Fixes

  • Scheduled jobs silently did nothing — their log redirect pointed at a directory that did not exist.
  • register_service wrote auth keys the reader never looked for, so a registered credential was ignored.
  • NetBox 4.5 — Bearer nbt_ tokens and the /api prefix are supported.
  • olav doctor --json emitted invalid JSON, because rich wrapped a value mid-line.
  • A slow local model was reported as unreachable rather than slow.

v0.22.1 — Local embedding endpoints

Fixes

  • Ollama, llama.cpp and vLLM embedding endpoints work as configured; the client no longer assumes an OpenAI-shaped response.

v0.22.0 — Zero-ritual onboarding

Features

  • Bare olav in a fresh directory sets itself up — no init ceremony before the first question.
  • olav doctor and the TUI /doctor — a deterministic, zero-LLM health check you can run before blaming the model.
  • A welcome screen that reports real findings from your workspace instead of rotating tips.
  • Conversational self-configuration — the admin agent can edit its own configuration under approval.
  • Undo for workspace writes and cron changes.

v0.21.1 — Per-skill LLM overrides, KB source-tier lifecycle, NetBox CSV sync

Status: in progress (post-rc4, prepping for tag).

Features

  • Per-skill LLM overrides. Each SKILL.md / AGENT.md may now carry an optional llm: block with any subset of model, temperature, max_tokens, base_url, model_provider, num_ctx, num_predict. Missing keys fall through to the global api.json.llm.* default. No profile indirection — single fall-through chain. Used in-tree by:
  • sim, investigate, learner → temperature: 0.0 (deterministic output)
  • core/memory_curator → max_tokens: 4096 (R102 HITL is short)
  • analyzer (no override) — inherits global 32K
  • Global max_tokens raised 16000 → 32768. gemma4:31b nothink can produce long structured tables/reports without truncation. Override per-skill if needed (memory_curator already does the opposite).
  • Source-tier KB lifecycle (dev_docs/79). Every *.guide.yaml carries a source_tier field (vendor / platform / team / user). Priority + weight derived from the tier, not hand-set.
  • R102 conversational commits land at source_tier: user (priority 3, weight 0.83) — operator-spoken rules can't outrank architectural invariants by accident
  • Schema v1 YAML still loads (default = user tier with backward-compat for explicit priority: field)
  • New CLI: olav kb remove <intent> --reason "..." (tombstone YAML + LanceDB delete + kb_audit/<ts>_remove_<intent>.yaml row), olav kb show <intent>, olav kb list-guides --tier <name>
  • kb_audit/ git-trackable audit chain. Every KB-mutating action (R102 commit, kb remove, NetBox push) drops one kb_audit/<iso-ts>_<verb>_<intent>.yaml row with action / intent / actor / timestamp / body_sha256 + outcome counters. Distinct from audit.duckdb — git-tracked so the change record survives independent of the runtime DB.
  • olav user CLI. Friendly aliases for olav admin add-user/.... olav user create <name> requires the Linux user to exist (pwd.getpwnam(name)) by default — OLAV's user model is aligned to the OS, no parallel identity. Token hash + salt stored, plaintext shown once. Override Linux check with --no-verify for containers.
  • Collection-time credential redaction (olav.core.redaction.scrub, ADR-0008). Two-layer pipeline scrubs passwords / SNMP communities / BGP MD5 keys / IPSec PSK / TACACS / RADIUS shared secrets out of collected network config before it lands in audit.duckdb or exported datasets, while keeping IP / hostname / ASN / topology intact so diff and diagnostic logic still work.
  • Collection layer: olav.core.redaction.scrub wraps netconan (Batfish team, ~50 KB pure Python; 55 built-in sensitive-item regex + 5213 reserved words; per-workspace .redaction_salt chmod 600). Wired into netops_init/run.py:_collect_cmd (collection boundary) and AuditEventRecorder.record_message (defense-in-depth for chat/LLM text). Replacement marker: netconanRemovedN.
  • Enterprise gate layer: audit_recorder.redact_sensitive (5-pattern regex producing [REDACTED]) drives Layer 1 of audit_dataset_export.redact_audit_run and runs again at gate-check time as a final escape detector.
  • Configuration: api.json.redaction.{enabled, anon_ip, extra_sensitive_words, extra_reserved_words} + OLAV_REDACTION=0 env-var kill-switch. Install via pip install 'olav[redaction]' to pull netconan>=0.13; without it OLAV still operates but Layer 1 fails open with a WARNING.
  • 16 governance tests in tests/governance/test_redaction.py cover Cisco / Junos / salt management / fail-open semantics / idempotency.
  • NetBox CSV sync chain (dev_docs/71 Ch10). Three-phase workflow decoupling read (netops) from write (services) via a CSV intermediate:
  • olav --agent netops "/export_netbox_csv" — produces NetBox-shaped CSV from netops.devices (column mapping in netops/guides/netbox_csv_export.guide.yaml)
  • olav --agent services "/import_netbox_csv" — dry-run validates shape, writes exports/reports/netbox_import_dry_run.md
  • olav --agent services "/import_netbox_csv --write --endpoint ... --token ..." — actual POST to a running NetBox, 5-step idempotent lookup-or-create chain (site → manufacturer → device-type → device-role → platform → device); writes kb_audit/ row per push

Fixes

  • Audit tables populate from the CLI again. v0.20.3 flipped OLAV_MIDDLEWARE_MODE default to middleware; the CLI invoke path never bound an OlavRunContext to astream_events, so AuditMiddleware._get_context() returned None and silently no-op'd every tool call. Plus record_run_start was called without thread_id so the sessions UPSERT never fired. Both fixed: CLI now mirrors the API server's context-binding pattern, and a thread_id=session_id or run_id is threaded through every turn. Net effect: audit_runs / sessions / audit_events / audit_messages / audit_tool_calls all populate after a single tool-using turn.
  • reasoning_effort: minimal no longer sent. Ollama OpenAI-compat strictly validates this field and accepts only high / medium / low / max / none. Dropped — the sibling chat_template_kwargs.enable_thinking=false already handles the qwen3 / gemma4 nothink path.
  • olav skill install refuses to overwrite platform-owned workspaces. Domain-package installs (e.g. olav-netops) that carried a dev-mirror of audit/ would silently clobber the platform's bundled audit/ on olav skill install. Now guarded by _PLATFORM_OWNED = {core, audit, services} skip list with a --force-overwrite escape.
  • /netops_init ingest gate. Stage 4 now distinguishes "SSH collected" from "DB ingest" and prints ❌ FAILED at the ingest stage with non-zero exit when the staging JSON loaded 0 rows despite successful SSH collection. Previously printed ✅ Network initialization complete even when ingest erred.
  • 5 SQL view schema guides shipped (netops/guides/sql_schema_*.guide.yaml) documenting BGP / OSPF / interface / route / device-and-topology view column names + worked examples. AutoRecall surfaces them on schema-discovery queries; gemma4 nothink lands first-try SQL instead of column-name trial-and-error.
  • olav user subparser added. Previously CLI had only the more verbose olav admin add-user path; olav user create / list / token / revoke are the canonical user-facing form.
  • Platform audit workspace bundled into the wheel. Pre-v0.21.1 pip install olav==0.19.0 && olav init produced only 2 agents (core + services); the audit/ workspace existed in dev but wasn't in the wheel's data path. Fresh install now ships 3 platform agents (audit / core / services).

Migration

  • No flag-day. Schema v1 *.guide.yaml files continue to load, treated as source_tier: user with a deprecation warning at prime time. Migrate at your leisure; the v2 form is schema_version: 2 + an explicit source_tier: field.
  • R102 commits are demoted to user tier. If you were relying on operator-spoken rules outranking platform guides, that no longer happens. Promote via git PR (edit the YAML to source_tier: team and commit through the normal review path).
  • Lab / CAB content moved out. lab/ sub-agent + CAB workflow + TCF schema demo + fault-injection scripts now live in olav-ent. The platform-base runsheet (dev_docs/71) drops former Ch7 + Ch8; the enterprise runsheet picks them up.

v0.21.0 — WebUI agent routing + service lifecycle hardening

Status: in progress (rc4 latest at time of writing).

Fixes

  • Web GUI: agent dropdown now actually switches the running graph. Pre-rc2 the FastAPI server kept a single _agent_instance singleton; clicking "ops" in the dropdown only re-tagged audit events but still executed the core graph. get_agent() is now a per-assistant_id cache; five agents coexist without cross-contamination.
  • POST /reload returns 200 again. rc2 replaced the agent singleton with a cache dict but forgot to update the /reload handler; every reload raised NameError. Handler now iterates the cache, best-effort- closes each agent, and clears state.
  • Audit in middleware mode sees run_id. v0.20.3 flipped OLAV_MIDDLEWARE_MODE default to middleware; the web server never attached an OlavRunContext to astream_events, so AuditMiddleware had no run_id to record against — every web request ran without audit. Fixed by wiring context through in both modes.
  • Thread + agent binding persists across page refresh. WebUI now writes currentThreadId and each thread's agentId to localStorage; F5 restores state fully.
  • core pinned first in /agents. Installing olav-netops adds 4 workspaces that all sort before core alphabetically, so the WebUI dropdown was auto-selecting audit as the default. Server now forces core to index 0.
  • Service lifecycle pins spawned processes to the workspace. Pre-rc4 uvicorn / the agent daemon / the syslog receiver derived their PID + log paths from a module-level PROJECT_ROOT frozen at module-import time. When olav was launched from inside another repo with a pyproject.toml, the spawned processes silently inherited the wrong cwd and read the wrong users.duckdb — every login 401'd. Resolution is now cwd-based, anchored on the .olav/ marker, and re-resolved on every call; Popen(..., cwd=str(workspace_root)) is passed explicitly.
  • olav service start --all works on a fresh wheel install. Not a bug — arg order is olav service <action> --all, which was always correct — but the demo runsheet examples have been re-verified.
  • olav skill install <path> still works via a v0.21.0-rc4 deprecation shim. The singular skill dispatcher was removed in v0.20.3; the shim forwards verbatim to olav agent install and prints a one-line warning. See CLI reference for the migration note.

New in rc4

  • scripts/sync_netops_workspace.py — sync the platform-vendored .olav/workspace/ops/ copy with the authoritative olav-netops/.olav/workspace/ops/ tree. Default direction: root → olav-netops; --reverse pulls the other way; --check is the CI drift guard.

v0.20.3 — deepagents migration complete (P7 final)

  • olav skill install subcommand removed. Use olav agent install. (v0.21.0-rc4 adds a deprecation shim for back-compat; see above.)
  • OLAV_MIDDLEWARE_MODE default flipped to middleware. The legacy callback plugins are still installable via OLAV_MIDDLEWARE_MODE=callback for one release cycle while AuditMiddleware proves itself.
  • parse_subcommand_args sweeping into 7 verbs. Shared helper for quoted/empty/-- argparse edge cases.
  • workspace_discovery.py consolidated. Three duplicate scanners merged into one; all consumers (CLI, web /agents, TUI) go through the same code path.

v0.20.2 — P1 + P2 + P6 cutover

  • Native TUI (via langgraph dev subprocess) is now default, replacing the in-process overlay. OLAV_TUI_MODE=overlay rolls back to the v0.19.x path for emergency cases.
  • Tool-loader auto-discovers .deepagents/agents/<name>/tools/*.py and .deepagents/agents/<name>/skills/<skill>/tools/*.py at startup.
  • v0.20 workspace layout migration available via olav migrate v0_20_layout (idempotent + --dry-run).

v0.20.1 — P3 middleware migration

  • AgentMiddleware subclasses for audit / guardrails / memory-capture / memory-recall / uks-capture, parallel with the callback plugin path. OLAV_MIDDLEWARE_MODE env var selects.

v0.20.0 — P5 graph factory + olav agent install

  • olav agent install <path> introduced as the public install verb. olav skill install still works (same dispatcher under the hood); this release is the start of the rename.
  • graph_factory.build_graph() internal API for constructing the LangGraph per assistant_id — foundation for the per-id cache that landed in v0.21.0-rc2.

v0.19.0 — last pre-deepagents release

See DEMO_RUNSHEET (historical).