Current version: 1.7.0 (August 2026)
v1.7.0 — Starter Process Library
- Starter process library: 14 ready-to-use process definitions seeded into the account as ordinary, editable markdown files.
- New
processStarterSeedtool — copies the library server-side; never overwrites existing files. - Patch: collision check matches both filename and process id.
v1.6.1 — Second Opinion hardening
- Handles models that reject the
temperatureparameter (strip and retry once). - Typed error when a reasoning model burns its whole output budget on thinking blocks.
v1.6 — Second Opinion (cross-model review)
- New
secondOpiniontool: adversarial review of an analysis by a different AI model, using a separate user-supplied API key that never enters the model's context. - Unconditional API-key scrubber on every tool result, applied before truncation.
- Chat input intercepts pasted API keys with a warning.
v1.5.3 — Privacy Mode: receipts and staleness
- Privacy receipts logged with every dispatch (active manifest + session stats).
- 15-minute redaction-dictionary TTL to catch changes made outside Sonar.
- Address book addressee names added to the dictionary.
- Synthetic canary customer makes the self-test independent of real data.
v1.5.2 — Privacy Mode: STRICT level
- Redact before clip — truncation can no longer split a protected value.
- Privacy toggle cycles Off → ON → STRICT; STRICT also blocks free-text tools.
- Schema-driven classification of custom fields (email/phone/credential fields auto-protected).
v1.5.1 — Chokepoint exemption
- Exemption for tools whose results contain only rule identifiers, never NetSuite data.
v1.5.0 — Privacy spec flow
- Describe privacy requirements in plain English; they compile into an enforced, tighten-only spec.
- Spec terms survive every dictionary rebuild.
v1.4.1 – v1.4.3 — Privacy Mode fixes
- Word-boundary anchoring on redaction scans.
- Entity mutations invalidate the redaction dictionary.
- Fail-closed serialization chokepoint; eager dictionary build at boot.
v1.4 — Privacy Mode
- Dispatch-layer entity pseudonymization: customer/vendor/employee identities withheld from the AI model via column-manifest minimization plus a redaction dictionary backstop.
- Toggled only from the header UI — the model cannot flip it.
v1.3 — Parallel subagents and Process runs
- Run 2–3 read-only research subagents concurrently under one approval.
- Process definitions: supervised multi-step pipelines with a live checklist card; approval gates are never relaxed.
v1.2 / v1.2.1 — Subagents
- Opt-in read-only researcher subagent: a nested agent loop with its own private context and live activity display.
- Fixes: accurate token accounting; collapsible running cards.
v1.1 — Guardrails
- Agent permission mask (Standard / No-deletes / Read-only), enforced at the tool-dispatch layer.
- Production deletes require explicit user confirmation, recorded in the audit trail.
- Governance-unit visibility; dry-run previews before every mutation.
- Report runs serialized per session with period verification.
v1.0 — Foundation
- Single stateless Suitelet: SPA chat UI, LLM relay, and tool dispatcher for records, files, and SuiteQL.
- Audit trail: every mutating tool call logged as
SONAR_AUDITentries in the script execution log. - Conversation state and API key live only in browser
localStorage— nothing stored server-side.