v1.7.0

Minor — Workflow-UX unification + chat-streaming polish + agent-efficiency. The agent's workflow tool surface is consolidated 8 → 6 with one consistent "workflow" vocabulary across tools, HTTP API and UI, plus a new Saved Workflows library tab. The chat stream gains an animated presence indicator, a live activity ticker, interleaved thinking, and an inline compaction marker. Agent-efficiency work lands orchestrated sub-agent routing and non-lossy compaction (a tool-result recall blob store).

Breaking: the tool renames and the /api/pipelines* → /api/workflows* endpoint move are a hard cut with no alias — external MCP clients / API scripts must update.

Breaking

• Agent tool surface 8 → 6 (#510, #512, #514) — run_pipeline → run_workflow; capture_process + promote_process merged into one save_workflow; step_complete removed.
• HTTP API /api/pipelines* → /api/workflows* (#511) — hard rename, no alias.

Added

• Saved Workflows library (#513, #520) — A library tab listing reusable workflows with run / rename / delete, and expandable per-step details.
• save_workflow tool (#512) — One call to capture and save a reusable workflow; merges the former capture_process + promote_process, with redaction hardening.
• plan_task decoupled (#514) — Planning never executes; it returns a reusable workflow_id that run_workflow runs.
• Chat streaming presence (#494, #497–#499) — Animated lynox-icon presence indicator + live activity ticker coupled to run state.
• Interleaved thinking (#492) — Thinking renders interleaved with text and tool calls in chronological order.
• Compaction marker (#491) — A persistent inline marker shows where the conversation was compacted.
• Downloadable artifacts (#493) — Data-file artifact types are downloadable; HTML artifacts offer a source download.
• Orchestrated sub-agent routing (#507) — Eligible multi-step plans dispatch one sub-agent per step.
• Non-lossy compaction (#508) — Large tool results are evicted to a recall blob store at compaction and fetchable on demand via recall_tool_result.

Fixed

• Main agent no longer auto-downgrades (#517) — The main conversational agent always runs on the configured tier; the per-turn _isSimpleTask heuristic that silently dropped to Haiku is removed.
• Deterministic save_workflow steps (#521) — Step extraction preserves the actual action sequence; the LLM only annotates a fixed step list and can no longer merge or drop steps.
• Per-message usage footer survives an SSE drop (#518) — The done event echoes the run usage so the footer renders even if the turn_end frame is lost.
• Per-message LLM usage persists across thread resume (#486).
• Saved-workflow persistence (#515, #516) — save_workflow and decoupled plan_task persist their pipelines so the library and workflow_id survive a restart.
• Context budget from real API usage (#489) — Replaces the summed estimate that drifted past 100%.
• Truncated max_tokens turns continue (#490) — A turn cut off at the token cap resumes instead of returning empty.
• Voice (#488, #495) — STT glossary no longer rewrites common words to proper nouns; voice messages can't double-submit.
• Speak (#503) — TTS-hostile special characters normalized.
• Web UI (#485) — Mobile nav, chat scroll, artifact deep-links.
• capture_process reads session-wide tool calls (#505).
• Operator-channel hallucination (#502) — The prompt states the real operator channels, stopping Telegram hallucination after the Telegram removal.
• Nav chat-history fills the sidebar (#519) — The thread list grows to fill available height; other nav items pin to the bottom.

Internal

• Agent-efficiency measurement (#506, #509) — Phase-0 measurement protocol + a Tier-1 static cost-regression guard.
• perf(memory) (#487) — A no-query memory_recall is bounded to a ranked, capped subset.
• Shared AgentPresenceIcon (#500) — Extracted; transcription state animated.
• Transcription animation + compaction history block (#501).
• Telegram references removed from README + bench fixtures (#504).
• CI: staging-smoke pnpm setup fixed (#496).