0.7.0

Added

• /config → Lazy Tools: choose which tools are worth prefilling. A new tab with the feature
  switch, the retriever, how many matches a search returns, and a core / auxiliary tier per toolset
  and per configured MCP server. Core tools are in the model's prompt from the first token and are paid
  for on every query; auxiliary tools are not in the prompt at all - the agent finds them with
  search_tools and then calls them normally, so they cost nothing until they are needed, at the price
  of one extra round the first time. The tab carries a short explanation of that trade, and every row
  below the switch is visibly locked (and says why) until the feature is on.

  Off by default, and stored in settings.json under toolPolicy (toolSearch,
  auxiliaryMiddleware, auxiliaryTools, coreMCPServers, toolSearchModel, toolSearchLimit), so
  an existing project is unaffected until you open the tab. The MCP tier lives there rather than in
  mcp.json, which may be a shared .mcp.json that other tools read.

• Retriever choice, with its download state. Space cycles lexical (no model) → ColBERT 350M (8-bit) → ColBERT 350M (bf16), each showing ready or not downloaded, 410 MB. The ColBERT
  encoders score every query token against every tool token (late interaction), which reads intent
  considerably better than term overlap; lexical needs no model at all. Both models appear in /model →
  Local like any other, so they can be pulled ahead of time (ripple model pull mlx-community/LFM2.5-ColBERT-350M-8bit) and deleted. They can never be selected as a planner.

• The launch banner reports lazy tools with a tool search row next to main agent / vision,
  naming the retriever and how many tools are held back (ColBERT 350M (8-bit) · 18 tools on demand).
  The count comes from the live agent, so a toolset tiered auxiliary but also disabled is not counted as
  available. No row when the feature is off.

• /tools and /mcp show each tool's tier. A filled ● for a core tool, a hollow ○ plus an
  [auxiliary] tag for one the agent has to search for, and a note on a toolset that is entirely
  auxiliary; /mcp shows tier: core|auxiliary beside each server's approval mode. Without this an
  auxiliary tool looked identical to a core one, which made a configured tier indistinguishable from a
  setting that had been lost.

Changed

• Every /config tab says what it is for. Each one now opens with a blue ⓘ box naming what that
  tab governs and the trade-off you are making there - what a capability costs in prompt tokens, why
  a container is slower than the local shell, why the compaction threshold rather than the model's
  window is what keeps a session inside your memory, that the prefill cache is pure disk-for-speed.
  Only Lazy Tools had one before, as plain grey text; the rows' own summaries are unchanged.

• /model → Local and Remote are the same view now. Local was a flat twenty-one-row list with the
  family repeated on every row, the repo id printed for all of them, no search, and nothing about the
  two numbers that decide whether a model fits. It has the Remote tab's two levels instead: model
  families first - split into LLM and Embedding, because a retrieval encoder has no LM head
  and picking one as a planner only fails at load - then that family's models when you press enter,
  under a heading per role (Text, Vision, Text + Vision for a unified VLM like Ornith
  that plans and sees images). A family is the model line, so LFM2.5's vision conversions sit
  inside it under the Vision heading rather than in a family of their own.

  Each model row is one line of aligned columns: ✓ on disk / ○ not yet, the weight format, the
  download size, the context window, and the tokens one turn may generate. The Hugging Face
  id, what the model is for, and the "default" note move to a subtitle under the highlighted row,
  which halves the list's height. Above it, the same context line the Remote tab carries.

• Remote model rows carry what Local's do. The same ✓/○ marker, context window and output budget,
  in the same cells, grouped by the same role headings - so comparing a free remote model with one
  you could download is reading across rather than translating. Where a local row prices itself in
  gigabytes a remote one is simply free; the output budget comes from the catalog's
  max_completion_tokens, newly decoded.

• Every list overlay says what it is. /model's three tabs, /tools and /mcp open with the
  same blue ⓘ box the /config tabs carry - what the list is, and what enter does to a row (which on
  Local and Remote changes as you drill in).

• The Sandbox tab says it needs Apple Containers. A second amber ⚠ box under the explanation,
  titled needs apple containers: the sandbox runs on Apple's container tool, which is not part of
  macOS, and nothing on the tab hinted that its one switch depends on something you have to install.
  It names the install and container system start, and spells out what happens without it -
  failover falls back to the local shell, so commands still run, just unsandboxed. The README and
  docs carry the same prerequisite.

• Lazy tools is flagged experimental. Its /config tab opens with an amber ⚠ box titled
  lazy tools - EXPERIMENTAL! rather than the blue ⓘ the settled tabs carry, and the text leads with
  why: a tool the model cannot see is a tool it may not think to look for, so a tiering that suits
  one project can quietly change how the agent behaves in another. The docs and README say so too.

• The Local tab is searchable, like the Remote one. It carries the same bordered input: type to
  narrow the list, backspace to edit, ctrl-u to clear. The query matches a model's name, family, id,
  weight format and role, so thinking, gemma, 4-bit, vision and embedding all work, and the
  highlighted model stays highlighted as the query is refined. esc clears the query before it
  closes the overlay. Because printable keys are now the search, removing a model is ctrl-x, not
  x.

• The banner wraps instead of clipping. In a narrow terminal the left pane cut its values off at
  an ellipsis - tool search ColBERT 350M (8-bit) · 35 tools o…, available mcps: deepwiki, … -
  which is exactly the text the banner exists to show. Values now wrap onto continuation lines
  indented under the label, and when the label leaves under ~18 columns beside it (a 58-column
  terminal) the value takes the whole pane on its own lines rather than being broken mid-word.

• The banner and status line name the model family-first. LFM2.5 · 1.2B Instruct rather than a
  bare 1.2B Instruct, which said nothing about which 1.2B once several families were downloaded.
  Same in the vision row and the model-switch notes.

• The Select tab's model picker and ripple model list carry the same facts. Picking a main agent
  or vision model shows each local option's purpose, weight format, size and context window; ripple model list is grouped by family and lists the context window and output budget alongside the size.

• Lazy tools says what it costs on a small planner, and which toolset to move first. The /config
  tab and the docs now warn that the feature wants a capable planner - the smallest models search less
  reliably and one was seen answering from a tool's description instead of calling it, inventing a note
  it had never read (fixed upstream in DeepAgents, which now tells the model after every search that it
  received definitions rather than results). Both also name filesystem as the toolset worth making
  auxiliary first: a small planner surface-matches the request against the schema in front of it, so
  "list my apple notes" reaches for ls and "read my clipboard" for read_file, and taking those out of
  the prompt removes the wrong answer instead of arguing with it - 3/8 to 4/5 on a 2.6B, after three
  rewrites of the prompt had barely moved it. See STEPS/ISSUES/tool-search-small-planners.md.

• --model no longer accepts a retrieval encoder. Planner ids are validated against the language
  catalog rather than the whole model catalog, so a ColBERT id is rejected up front instead of booting
  into a model that cannot load.

Fixed

• The context meter read far below the truth. The status line's percentage was a running tally of
  what the UI watched stream past - the user's prompt, plus one per generated token - so it counted
  neither the tool schemas (paid on every request, and never streamed) nor any tool result (a file
  read, shell output, a search), which is most of a real context. An empty session showed 0% when the
  prompt already cost thousands of tokens, a turn that read a large file barely moved the needle, and
  automatic compaction fired at what looked like a fraction of the window. /fresh didn't reset it
  either. The meter is now measured from the agent - the new ReactAgent.contextTokens(threadId:),
  the same number the compaction trigger tests itself against - after every turn, compaction, model
  switch, /config rebuild and /fresh, and seeded at launch (so a resumed session opens with its
  real size). Per-token nudges still move it mid-turn; the measurement settles it at the end. This is
  the behaviour the compaction docs already described.