collavre 0.24.0 → 0.25.0

data/app/services/collavre/typo_corrector.rb

@@ -1,188 +0,0 @@
-module Collavre
-  # Server-side typo correction. Sends the text the user is typing to an LLM and
-  # returns a validated *structured edit list* — never a rewritten string. Each
-  # edit is {original, suggestion, reason, confidence}. Validation rejects anything
-  # that is not a small, in-place spelling fix (style rewrites, hallucinated spans,
-  # edits inside code/URLs/mentions), so the UI can trust every returned edit.
-  class TypoCorrector
-    AGENT_EMAIL = "typo-corrector@collavre.local"
-
-    # Reject "corrections" that are really rewrites: an edit is kept only when the
-    # original is short and the suggestion is within this edit distance of it.
-    MAX_ORIGINAL_LENGTH = 40
-    MIN_EDIT_DISTANCE_CAP = 2
-    EDIT_DISTANCE_RATIO = 0.4
-
-    FALLBACK_SYSTEM_PROMPT = <<~PROMPT.freeze
-      You are a typo correction engine. Fix only spelling mistakes and obvious typos,
-      never style, grammar or meaning, in any language. `original` must be an exact
-      substring of the input. Return ONLY JSON:
-      {"edits":[{"original":"x","suggestion":"y","reason":"spelling","confidence":0.0}]}
-    PROMPT
-
-    def initialize(client: nil)
-      @client = client
-    end
-
-    # Returns an Array of Hashes with string keys: "original", "suggestion",
-    # "reason", "confidence". Empty array on blank input or LLM/parse failure.
-    def correct(text)
-      text = text.to_s
-      return [] if text.strip.blank?
-
-      response = client.chat([
-        { role: :user, parts: [ { text: "Text:\n#{text}" } ] }
-      ])
-
-      edits = parse_response(response)
-      validate(edits, text)
-    end
-
-    private
-
-    def client
-      @client ||= build_client
-    end
-
-    def build_client
-      agent = Collavre.user_class.find_by(email: AGENT_EMAIL)
-      AiClient.new(
-        vendor: agent&.llm_vendor.presence || ENV.fetch("COLLAVRE_DEFAULT_LLM_VENDOR", "gemini"),
-        model: agent&.llm_model.presence || ENV.fetch("COLLAVRE_DEFAULT_LLM_MODEL", "gemini-3.1-flash-lite"),
-        system_prompt: agent&.system_prompt.presence || FALLBACK_SYSTEM_PROMPT,
-        llm_api_key: agent&.llm_api_key,
-        gateway_url: agent&.gateway_url,
-        # Vendor adapters (e.g. OpenClaw) resolve the gateway/key from the context
-        # user, not the llm_api_key/gateway_url kwargs. Pass the agent so an
-        # OpenClaw-backed typo agent isn't built with user: nil (which makes the
-        # adapter bail with "Gateway URL not configured" and silently return no edits).
-        context: { user: agent },
-        # Runs on debounced typing of *unsubmitted* text — never persist the draft
-        # to ActivityLog (RubyLlmInteractionLogger writes `messages` there).
-        log_interactions: false
-      )
-    end
-
-    def parse_response(content)
-      return [] if content.blank?
-
-      cleaned = content.gsub(/^```json\s*/, "").gsub(/\s*```$/, "").strip
-      parsed = JSON.parse(cleaned)
-      edits = parsed.is_a?(Hash) ? parsed["edits"] : parsed
-      edits.is_a?(Array) ? edits : []
-    rescue JSON::ParserError => e
-      # Never log the raw response: it can echo the user's unsubmitted draft,
-      # which would leak private text to application logs despite log_interactions: false.
-      Rails.logger.warn("[TypoCorrector] JSON parse error: #{e.message} (#{content.to_s.length} chars)")
-      []
-    end
-
-    def validate(edits, text)
-      skip_ranges = skip_ranges(text)
-      seen = {}
-
-      edits.filter_map do |edit|
-        next unless edit.is_a?(Hash)
-
-        original = edit["original"].to_s
-        suggestion = edit["suggestion"].to_s
-        next if original.empty? || suggestion.empty? || original == suggestion
-        next if original.length > MAX_ORIGINAL_LENGTH
-
-        # `original` must appear verbatim in the text, outside any skip region
-        # (code/URL/mention/markup). We return the offset of that first *valid*
-        # occurrence so the client anchors there — otherwise the client's own
-        # left-to-right search would bind to an earlier occurrence sitting inside
-        # a protected span and edit it (a markdown-canonical/code-safety hole).
-        offset = first_offset_outside_skip_ranges(text, original, skip_ranges)
-        next if offset.nil?
-
-        # Reject big rewrites — keep only minimal in-place fixes.
-        cap = [ MIN_EDIT_DISTANCE_CAP, (original.length * EDIT_DISTANCE_RATIO).ceil ].max
-        next if levenshtein(original, suggestion) > cap
-
-        key = [ original, suggestion ]
-        next if seen[key]
-        seen[key] = true
-
-        {
-          "original" => original,
-          "suggestion" => suggestion,
-          "reason" => edit["reason"].presence || "spelling",
-          "confidence" => clamp_confidence(edit["confidence"]),
-          # Emit the offset in UTF-16 code units, the coordinate space JS string
-          # indexing uses. A Ruby character index would be off by one per astral
-          # character (emoji, etc.) before the typo, so the client's substring
-          # check at `offset` would fail and fall back to indexOf — which can bind
-          # to a protected occurrence the server already excluded.
-          "offset" => utf16_offset(text, offset)
-        }
-      end
-    end
-
-    def clamp_confidence(value)
-      return 0.5 if value.nil?
-
-      Float(value).clamp(0.0, 1.0)
-    rescue ArgumentError, TypeError
-      0.5
-    end
-
-    # Character ranges that must never be edited: fenced code, inline code, URLs,
-    # @mentions, HTML tags, and markdown-canonical <span style> color fragments.
-    def skip_ranges(text)
-      ranges = []
-      patterns = [
-        /```.*?```/m,             # fenced code blocks
-        /`[^`]*`/,                # inline code
-        %r{https?://\S+},         # URLs
-        /@[^\s@]+/,               # @mentions
-        # `[^<>]` (not just `[^>]`) keeps this linear: a stray run of "<<<<" can't
-        # make scan re-walk the tail from every "<" (CodeQL polynomial-ReDoS).
-        /<[^<>]+>/                # HTML / span-style markup
-      ]
-      patterns.each do |pattern|
-        text.scan(pattern) { ranges << Regexp.last_match.offset(0) }
-      end
-      ranges
-    end
-
-    # First character offset where `needle` occurs entirely outside every skip
-    # range, or nil if every occurrence overlaps a protected span.
-    def first_offset_outside_skip_ranges(text, needle, skip_ranges)
-      pos = 0
-      while (idx = text.index(needle, pos))
-        finish = idx + needle.length
-        inside = skip_ranges.any? { |(s, e)| idx < e && finish > s }
-        return idx unless inside
-
-        pos = idx + 1
-      end
-      nil
-    end
-
-    # Convert a Ruby character index into a UTF-16 code-unit offset (JS string
-    # coordinates). Characters outside the BMP count as two code units in JS.
-    def utf16_offset(text, char_index)
-      text[0, char_index].to_s.encode("UTF-16LE").bytesize / 2
-    rescue Encoding::UndefinedConversionError, Encoding::InvalidByteSequenceError
-      char_index
-    end
-
-    # Iterative Levenshtein distance.
-    def levenshtein(a, b)
-      a = a.chars
-      b = b.chars
-      prev = (0..b.length).to_a
-      a.each_with_index do |ac, i|
-        curr = [ i + 1 ]
-        b.each_with_index do |bc, j|
-          cost = ac == bc ? 0 : 1
-          curr << [ curr[j] + 1, prev[j + 1] + 1, prev[j] + cost ].min
-        end
-        prev = curr
-      end
-      prev.last
-    end
-  end
-end

data/app/views/collavre/inbox_items/_item.html.erb

@@ -1,18 +0,0 @@
-<div class="inbox-item" data-id="<%= item.id %>">
-  <div class="message">
-    <% if item.link.present? %>
-      <%= link_to item.localized_message, item.link, class: 'item-link unstyled-link' %>
-    <% else %>
-      <%= h item.localized_message %>
-    <% end %>
-  </div>
-  <div class="time"><%= time_ago_in_words(item.created_at) %></div>
-  <div class="actions">
-    <% if item.state == 'read' %>
-      <button class="mark-unread" data-id="<%= item.id %>"><%= t('collavre.inbox.mark_unread') %></button>
-    <% else %>
-      <button class="mark-read" data-id="<%= item.id %>"><%= t('collavre.inbox.mark_read') %></button>
-    <% end %>
-    <button class="delete-item" data-id="<%= item.id %>"><%= t('collavre.inbox.delete') %></button>
-  </div>
-</div>

data/app/views/collavre/inbox_items/_items.html.erb

@@ -1,3 +0,0 @@
-<% items.each do |item| %>
-  <%= render partial: 'collavre/inbox_items/item', locals: { item: item } %>
-<% end %>

data/app/views/collavre/inbox_items/_list.html.erb

@@ -1,7 +0,0 @@
-<div id="inbox-items" data-next-page="<%= next_page || '' %>">
-  <% if items.any? %>
-    <%= render partial: 'collavre/inbox_items/items', locals: { items: items } %>
-  <% else %>
-    <div class="inbox-empty"><%= t('collavre.inbox.no_messages') %></div>
-  <% end %>
-</div>

data/app/views/collavre/inbox_items/index.html.erb

@@ -1 +0,0 @@
-<%= render partial: 'collavre/inbox_items/list', locals: { items: @inbox_items, next_page: @next_page } %>

data/app/views/collavre/inbox_mailer/daily_summary.html.erb

@@ -1,11 +0,0 @@
-<h1><%= t('collavre.inbox_mailer.daily_summary.greeting') %></h1>
-<ul>
-  <% @items.each do |item| %>
-    <li>
-      <%= item.localized_message %>
-      <% if item.link.present? %>
-        (<a href="<%= item.link %>"><%= t('collavre.inbox_mailer.daily_summary.view') %></a>)
-      <% end %>
-    </li>
-  <% end %>
-</ul>

data/app/views/collavre/inbox_mailer/daily_summary.text.erb

@@ -1,8 +0,0 @@
-<%= t('collavre.inbox_mailer.daily_summary.greeting') %>
-
-<% @items.each do |item| %>
-* <%= item.localized_message %>
-  <% if item.link.present? %>
-    <%= item.link %>
-  <% end %>
-<% end %>

data/app/views/collavre/users/typo_correction.html.erb

@@ -1,50 +0,0 @@
-<h1 class="no-top-margin"><%= t('collavre.users.typo_correction.legend') %></h1>
-
-<%= form_with model: @user, url: collavre.user_path(@user), method: :patch, local: true, html: { class: 'profile-form' } do |f| %>
-  <p class="text-muted"><%= t('collavre.users.typo_correction.description') %></p>
-
-  <div class="checkbox-field">
-    <%= f.check_box :typo_correction_enabled %>
-    <%= f.label :typo_correction_enabled, t('collavre.users.typo_correction.enabled') %>
-  </div>
-
-  <div>
-    <%= f.label :typo_correction_threshold, t('collavre.users.typo_correction.threshold') %>
-    <%= f.number_field :typo_correction_threshold, in: 0..100, step: 1, required: true %>
-    <small class="text-muted"><%= t('collavre.users.typo_correction.threshold_hint') %></small>
-  </div>
-
-  <fieldset>
-    <legend><%= t('collavre.users.typo_correction.device_legend') %></legend>
-    <div class="checkbox-field">
-      <%= f.check_box :typo_correction_on_voice %>
-      <%= f.label :typo_correction_on_voice, t('collavre.users.typo_correction.on_voice') %>
-    </div>
-    <div class="checkbox-field">
-      <%= f.check_box :typo_correction_on_soft_keyboard %>
-      <%= f.label :typo_correction_on_soft_keyboard, t('collavre.users.typo_correction.on_soft_keyboard') %>
-    </div>
-    <div class="checkbox-field">
-      <%= f.check_box :typo_correction_on_physical_keyboard %>
-      <%= f.label :typo_correction_on_physical_keyboard, t('collavre.users.typo_correction.on_physical_keyboard') %>
-    </div>
-  </fieldset>
-
-  <fieldset>
-    <legend><%= t('collavre.users.typo_correction.location_legend') %></legend>
-    <div class="checkbox-field">
-      <%= f.check_box :typo_correction_in_chat %>
-      <%= f.label :typo_correction_in_chat, t('collavre.users.typo_correction.in_chat') %>
-    </div>
-    <div class="checkbox-field">
-      <%= f.check_box :typo_correction_in_editor %>
-      <%= f.label :typo_correction_in_editor, t('collavre.users.typo_correction.in_editor') %>
-    </div>
-  </fieldset>
-
-  <%= f.submit t('collavre.users.update') %>
-<% end %>
-
-<div style="margin-top: 1rem;">
-  <%= link_to t('collavre.users.back'), collavre.user_path(@user) %>
-</div>

data/config/locales/mobile.en.yml

@@ -1,16 +0,0 @@
-en:
-  collavre:
-    mobile:
-      default_task_label: "Task"
-      summary:
-        a_tool: "a tool"
-        approval: "%{label}. Approval requested: %{detail}. Approve it?"
-      reply:
-        approved: "Approved."
-        denied: "Denied."
-        already_decided: "That was already decided."
-        not_authorized: "You're not allowed to decide that one — it may need an administrator."
-        clarify_decision: "Sorry, did you want to approve or deny?"
-        empty_response: "I didn't catch a response."
-        relayed: "Sent your reply to the agent."
-        sent: "Sent."

data/config/locales/mobile.ko.yml

@@ -1,16 +0,0 @@
-ko:
-  collavre:
-    mobile:
-      default_task_label: "작업"
-      summary:
-        a_tool: "도구 실행"
-        approval: "%{label}. 승인 요청: %{detail}. 승인할까요?"
-      reply:
-        approved: "승인했습니다."
-        denied: "거절했습니다."
-        already_decided: "이미 결정된 항목입니다."
-        not_authorized: "그 항목은 결정할 권한이 없습니다. 관리자 승인이 필요할 수 있습니다."
-        clarify_decision: "승인할까요, 거절할까요?"
-        empty_response: "응답을 듣지 못했습니다."
-        relayed: "에이전트에게 답변을 전달했습니다."
-        sent: "전송했습니다."