Order Samurai 2.2.0

Order Samurai 2.2.0

Your dashboard history survives rebuilds, several metrics read more accurately, and the code behind the dashboard was reorganised with no change in behaviour. If you're coming from 2.1.1, this release also brings everything from 2.1.2 to 2.1.4, summarised further down.

To upgrade: re-run the installer, or download order-samurai-core.zip below and check it against order-samurai-core.zip.sha256. Then run samurai doctor.

What you'll notice

• History stops getting wiped. Rebuilding the dashboard snapshot used to erase the series it can't recompute (scout counts, run-log and security signals) and leave a single point. They now keep their earlier values, and your newest live reading survives the rebuild. If telemetry can't be read, history is left untouched instead of being rewritten empty. If the write fails or is interrupted (full disk, Ctrl-C), the previous file is restored. To get the old behaviour back, set BACKFILL_CARRY_FORWARD=false.
• Sparklines and trend arrows compare like with like. They now compare the current 30-day value only with earlier 30-day readings, not with one-week totals. Right after a rebuild you may see only two points per metric, and anomaly and trend warnings come back as new readings build up.
• Project scores may shift. A telemetry tag now counts toward one project only, so sessions shared by similarly named projects are no longer counted twice.
• Fewer refresh failures and fake numbers. A transcript deleted mid-scan is skipped instead of failing the whole refresh. A missing or unreadable vibe score shows "no data" instead of a made-up number.
• Idle weeks are no longer failures. Mechanism Liveness reports "idle by design" instead of FAIL when the reflex heartbeat shows the engine was deliberately idle and no run happened. A corrupt or unreadable execution log still fails.
• Cache Hit Rate is stricter. It counts only usage records whose token counts are fully known. The number is expected to stay the same on a standard install; setups that only have raw telemetry now show "no data" instead of a guess.
• Skill Routing Adherence is stricter. It credits a skill only when it was invoked in the same session after the detection, and ignores the hiring-review nudge. Values can change if you have routing logs.
• Daily digest link. The "Review:" link now reads http://localhost:4322/ (same local dashboard as before).

Known limits

• Three new metrics show SIMULATED on a standard install: Stale Branches, Tickets Stuck at the Lane, and Approved, Never Executed. They only work in the AgenticaOS development layout and are never flagged, so you can ignore them. Their fix hints mention /git-hygiene and /captain, which Order Samurai doesn't ship.
• The dashboard card for Mechanism Liveness shows no value and no explanation for the new "idle by design" state yet.

For anyone calling the Python API

• call_routed_llm() (also agentica_core.model_router.call_llm) gained two optional arguments, num_ctx and force_json, placed before project. If your own code passes project positionally, pass it as project= instead. Order Samurai's own code already does.
• New opt-in LLMGateway.generate_strict(): one isolated Claude CLI call with no tools, no MCP and bounded input and output. Nothing in Order Samurai calls it yet.
• agentica_core/aggregate.py is now a facade over six aggregate_* modules. Every name you imported before still imports, and output is byte-identical on the recorded test payloads.

Since 2.1.1 (2.1.2 to 2.1.4, which shipped without notes)

• Protection for more coding agents. samurai install now installs the prompt-injection guard for Codex, Cursor and Gemini CLI as well as Claude Code, and detects which ones you have. It names agents it can't protect yet (Windsurf, Goose, Cline, OpenCode, Aider, Continue) and exits with an error instead of claiming success when it finds no supported agent. Each agent has coverage limits, listed in the CHANGELOG; for Gemini CLI, the guard only runs in folders you've marked as trusted.
• The secret scrubber works on a standard install. In 2.1.1 it depended on a file a standard install doesn't include, so it scrubbed nothing. samurai doctor now runs each hook for real instead of only checking that it's registered.
• Install works on Homebrew Python and other "externally managed" Pythons, using a private environment in ~/.samurai/venv.
• More reliable installs. The web installer retries when the zip and its checksum don't match right after a release. Install no longer registers the hooks twice when a hook registry already runs them, and uninstall no longer reports success when it couldn't remove a hook.
• Quieter, safer guard log. It no longer records tool calls that passed, and it redacts secrets from the entries it keeps.
• Clearer failures. The daily digest prints a warning before falling back to Mail.app. Doctor says why a hook's criticality is unknown.
• Dashboard. Previous and next controls for categorised reflex stacks, and the selected card stays put when other alerts clear.
• One support address: support@agentica-llc.biz, now printed at the end of samurai doctor and the installer.

Full details are in CHANGELOG.md.