@workweave/router 0.2.3 → 0.2.5

package/cc-statusline.sh

@@ -130,14 +130,18 @@ normalize_model() {
 prices='{
   "input": {
     "claude-fable-5":                   0.01,
-    "claude-haiku-4-5":                 0.0008,
+    "claude-haiku-4-5":                 0.001,
+    "claude-opus-4-0":                  0.015,
+    "claude-opus-4-1":                  0.015,
+    "claude-opus-4-5":                  0.005,
     "claude-opus-4-6":                  0.005,
     "claude-opus-4-7":                  0.005,
     "claude-opus-4-8":                  0.005,
     "claude-sonnet-4-5":                0.003,
     "claude-sonnet-4-6":                0.003,
-    "deepseek/deepseek-v4-flash":       0.00014,
-    "deepseek/deepseek-v4-pro":         0.00174,
+    "claude-sonnet-5":                  0.003,
+    "deepseek/deepseek-v4-flash":       0.0001134,
+    "deepseek/deepseek-v4-pro":         0.001318,
     "gemini-2.0-flash":                 0.0001,
     "gemini-2.0-flash-lite":            0.000075,
     "gemini-2.5-flash":                 0.0003,
@@ -148,6 +152,7 @@ prices='{
     "gemini-3.1-flash-lite-preview":    0.0001,
     "gemini-3.1-pro-preview":           0.002,
     "gemini-3.5-flash":                 0.0015,
+    "google/gemma-4-26b-a4b-it":        0.00015,
     "gpt-4.1":                          0.002,
     "gpt-4.1-mini":                     0.0004,
     "gpt-4.1-nano":                     0.0001,
@@ -157,10 +162,10 @@ prices='{
     "gpt-5-chat":                       0.0025,
     "gpt-5-mini":                       0.0005,
     "gpt-5-nano":                       0.0001,
-    "gpt-5.4":                          0.003,
-    "gpt-5.4-mini":                     0.0004,
-    "gpt-5.4-nano":                     0.0001,
-    "gpt-5.4-pro":                      0.02,
+    "gpt-5.4":                          0.0025,
+    "gpt-5.4-mini":                     0.00075,
+    "gpt-5.4-nano":                     0.0002,
+    "gpt-5.4-pro":                      0.03,
     "gpt-5.5":                          0.005,
     "gpt-5.5-mini":                     0.0005,
     "gpt-5.5-nano":                     0.00015,
@@ -177,22 +182,27 @@ prices='{
     "qwen/qwen3-coder-next":            0.0005,
     "qwen/qwen3-next-80b-a3b-instruct": 0.00015,
     "qwen/qwen3.5-flash-02-23":         0.00005,
-    "qwen/qwen3.6-35b-a3b":             0.00015,
+    "qwen/qwen3.6-35b-a3b":             0.000172,
     "qwen/qwen3.7-plus":                0.0004,
     "xiaomi/mimo-v2.5-pro":             0.001,
-    "z-ai/glm-5":                       0.0006,
-    "z-ai/glm-5.1":                     0.00105
+    "z-ai/glm-5":                       0.001,
+    "z-ai/glm-5.1":                     0.0014,
+    "z-ai/glm-5.2":                     0.0014
   },
   "output": {
     "claude-fable-5":                   0.05,
-    "claude-haiku-4-5":                 0.004,
+    "claude-haiku-4-5":                 0.005,
+    "claude-opus-4-0":                  0.075,
+    "claude-opus-4-1":                  0.075,
+    "claude-opus-4-5":                  0.025,
     "claude-opus-4-6":                  0.025,
     "claude-opus-4-7":                  0.025,
     "claude-opus-4-8":                  0.025,
     "claude-sonnet-4-5":                0.015,
     "claude-sonnet-4-6":                0.015,
-    "deepseek/deepseek-v4-flash":       0.00028,
-    "deepseek/deepseek-v4-pro":         0.00348,
+    "claude-sonnet-5":                  0.015,
+    "deepseek/deepseek-v4-flash":       0.0002791,
+    "deepseek/deepseek-v4-pro":         0.0026361,
     "gemini-2.0-flash":                 0.0004,
     "gemini-2.0-flash-lite":            0.0003,
     "gemini-2.5-flash":                 0.0012,
@@ -203,6 +213,7 @@ prices='{
     "gemini-3.1-flash-lite-preview":    0.0004,
     "gemini-3.1-pro-preview":           0.008,
     "gemini-3.5-flash":                 0.009,
+    "google/gemma-4-26b-a4b-it":        0.0006,
     "gpt-4.1":                          0.008,
     "gpt-4.1-mini":                     0.0016,
     "gpt-4.1-nano":                     0.0004,
@@ -212,14 +223,14 @@ prices='{
     "gpt-5-chat":                       0.01,
     "gpt-5-mini":                       0.002,
     "gpt-5-nano":                       0.0004,
-    "gpt-5.4":                          0.012,
-    "gpt-5.4-mini":                     0.0016,
-    "gpt-5.4-nano":                     0.0004,
-    "gpt-5.4-pro":                      0.08,
-    "gpt-5.5":                          0.04,
+    "gpt-5.4":                          0.015,
+    "gpt-5.4-mini":                     0.0045,
+    "gpt-5.4-nano":                     0.00125,
+    "gpt-5.4-pro":                      0.18,
+    "gpt-5.5":                          0.03,
     "gpt-5.5-mini":                     0.0025,
     "gpt-5.5-nano":                     0.0006,
-    "gpt-5.5-pro":                      0.12,
+    "gpt-5.5-pro":                      0.18,
     "minimax/minimax-m2.7":             0.0012,
     "minimax/minimax-m3":               0.0012,
     "mistralai/mistral-small-2603":     0.0006,
@@ -232,11 +243,12 @@ prices='{
     "qwen/qwen3-coder-next":            0.0012,
     "qwen/qwen3-next-80b-a3b-instruct": 0.0012,
     "qwen/qwen3.5-flash-02-23":         0.00015,
-    "qwen/qwen3.6-35b-a3b":             0.00095,
+    "qwen/qwen3.6-35b-a3b":             0.0012002,
     "qwen/qwen3.7-plus":                0.0016,
     "xiaomi/mimo-v2.5-pro":             0.003,
-    "z-ai/glm-5":                       0.00208,
-    "z-ai/glm-5.1":                     0.0035
+    "z-ai/glm-5":                       0.0032,
+    "z-ai/glm-5.1":                     0.0044,
+    "z-ai/glm-5.2":                     0.0044
   }
 }'
 # END_GENERATED_PRICES
@@ -425,20 +437,23 @@ elif [[ "$routed" == "failure" ]]; then
   # swap or compute savings against a non-model.
   printf '%s — %s%s' "$brand" "$routed" "$tokens_clause"
 elif [[ -n "$routed" ]]; then
-  # Show the savings clause only when the session is genuinely net-saving.
+  # Always show the savings clause, flooring a non-positive total at $0.00.
   # session_savings is "0.0000" on fresh sessions or sessions where every
-  # turn routed back to the selected model; it can also go negative when
-  # sticky routing forces a haiku-tagged side-call up to a cached
-  # sonnet/opus decision. In both cases the word "saved" would mislead,
-  # so drop the savings clause but keep the token totals.
-  has_savings="false"
-  if [[ -n "$session_savings" ]] \
-     && awk -v v="$session_savings" 'BEGIN{exit !(v+0 > 0.005)}'; then
-    has_savings="true"
+  # turn routed back to the selected model, and can go negative when routing
+  # upgrades a turn to a pricier model (e.g. a sticky/hard-pinned side-call,
+  # or a hard turn escalated to opus). "saved -$X" would mislead, so clamp
+  # the display to $0.00 rather than dropping the clause — a $0.00 readout
+  # tells the user the router ran and simply didn't beat their selection.
+  # When the CC selection is unknown ("?" or empty) there's nothing to
+  # compare against, so drop the "← selection" arrow and just show routed.
+  display_savings="$session_savings"
+  if [[ -z "$display_savings" ]] \
+     || awk -v v="$display_savings" 'BEGIN{exit !(v+0 < 0)}'; then
+    display_savings="0"
   fi
-  if [[ "$has_savings" == "true" ]]; then
+  if [[ -n "$selected_display" && "$selected_display" != "?" ]]; then
     printf '%s — %s ← %s · saved %s%s' \
-      "$brand" "$routed" "$selected_display" "$(fmt_money "$session_savings")" "$tokens_clause"
+      "$brand" "$routed" "$selected_display" "$(fmt_money "$display_savings")" "$tokens_clause"
   else
     printf '%s — %s%s' "$brand" "$routed" "$tokens_clause"
   fi

package/install.sh

@@ -533,20 +533,26 @@ TOML
   chmod 600 "$config_file"
 }
 
-# write_opencode_config merges TWO managed providers into opencode's
-# opencode.json plus the bundled Codex-subscription plugin:
-#   - provider.weave        : Anthropic-shaped (@ai-sdk/anthropic) — the router
-#                             speaks the Anthropic Messages API natively, so the
-#                             bundled provider works unmodified. The default.
-#   - provider.weave-codex  : OpenAI-shaped (@ai-sdk/openai, Responses wire) —
-#                             lets a caller's own ChatGPT (Codex) subscription
-#                             pay for their opencode turns. Dormant until the
-#                             user runs `opencode auth login` → ChatGPT and the
-#                             weave-codex plugin's loader injects the bearer.
+# write_opencode_config merges the managed Weave provider(s) into opencode's
+# opencode.json plus the bundled subscription plugin:
+#   - provider.weave        : OpenAI/Responses-shaped (@ai-sdk/openai → /v1/responses).
+#                             The single request provider. The router routes every
+#                             turn across all models the caller's subscriptions +
+#                             Weave key can pay for, and bills the plan matching the
+#                             model it served. The default.
+#   - provider.weave-claude : login-only storage for a Claude (Pro/Max) subscription
+#                             (no models, never serves requests). Written only when
+#                             the bundled plugin is present, since the Claude login
+#                             method lives in the plugin. The `weave` loader reads
+#                             this slot and attaches the Claude sub.
 # The plugin (src bundled at $script_dir/opencode-weave/src/index.ts) is dropped
 # into $opencode_dir/.weave/ and registered via opencode.json's `plugin` array
-# by absolute path — scope-independent (no reliance on an auto-load dir).
-# Re-running rewrites both blocks in-place via jq; uninstall strips them.
+# by absolute path — scope-independent (no reliance on an auto-load dir). It owns
+# both the ChatGPT (on `weave`) and Claude (on `weave-claude`) logins and attaches
+# whichever subscriptions are connected to every request via the router's
+# dedicated X-Weave-*-Subscription headers. Re-running rewrites the blocks
+# in-place via jq (and strips the legacy `weave-codex` provider); uninstall
+# strips them.
 #
 # Usage: write_opencode_config <config_file_path> <base_url> <api_key> [user_email] [user_name]
 write_opencode_config() {
@@ -573,61 +579,53 @@ write_opencode_config() {
 
   # Headline models we surface in opencode's picker. The router re-routes
   # each request anyway, so this list is mostly UX — what shows up when the
-  # user runs /models inside opencode. Keep it short and Anthropic-shaped
-  # so the bundled @ai-sdk/anthropic provider can request them.
+  # user runs /models inside opencode. We list a mix of GPT + Claude so the
+  # picker reflects that the router spans families; whichever model serves the
+  # turn, its matching subscription (if connected) pays.
   #
-  # apiKey is set to the router key as well as planted in headers. opencode's
-  # @ai-sdk/anthropic provider treats apiKey as required at config-parse time
-  # and otherwise falls back to ANTHROPIC_API_KEY from the environment;
-  # without this, a user who's never had an Anthropic key in their shell hits
-  # a startup error before the router ever sees a request. The router itself
-  # ignores the value (auth runs off X-Weave-Router-Key); apiKey here is just
-  # a placeholder that satisfies the SDK's "is auth configured" check.
-  # baseURL KEEPS its /v1 here — do NOT align it with the pi block, which is
-  # root. opencode's @ai-sdk/anthropic (Vercel) appends only /messages to
-  # baseURL (its default is api.anthropic.com/v1), so /v1 yields the correct
-  # /v1/messages. Dropping it would hit /messages and 404. (pi uses the official
-  # @anthropic-ai/sdk, which appends /v1/messages to a root baseURL — opposite
-  # convention.) Verified by install/pi-router/test/opencode_smoke.sh.
+  # npm is @ai-sdk/openai and baseURL KEEPS its /v1 here: opencode's
+  # @ai-sdk/openai provider appends /responses, yielding the router's
+  # /v1/responses surface (the canonical inbound — the router translates to
+  # Anthropic/OSS as it routes, and ships verbatim to the Codex backend for GPT
+  # turns). apiKey is the router key as a parse-time placeholder; the router
+  # authenticates off X-Weave-Router-Key (planted in headers) and the plugin's
+  # loader attaches subscriptions via the dedicated X-Weave-*-Subscription
+  # headers, so the apiKey value is never used upstream.
   local block
   block="$(jq -n \
     --arg url "$block_url/v1" \
     --arg key "$block_key" \
     --argjson headers "$headers_json" '
     {
-      npm: "@ai-sdk/anthropic",
+      npm: "@ai-sdk/openai",
       name: "Weave Router",
       options: { apiKey: $key, baseURL: $url, headers: $headers },
       models: {
-        "claude-fable-5":    { name: "Claude Fable 5 (via Weave Router)" },
-        "claude-opus-4-8":   { name: "Claude Opus 4.8 (via Weave Router)" },
-        "claude-opus-4-7":   { name: "Claude Opus 4.7 (via Weave Router)" },
+        "claude-sonnet-5":   { name: "Claude Sonnet 5 (via Weave Router)" },
         "claude-sonnet-4-6": { name: "Claude Sonnet 4.6 (via Weave Router)" },
-        "claude-haiku-4-5":  { name: "Claude Haiku 4.5 (via Weave Router)" }
+        "claude-opus-4-8":   { name: "Claude Opus 4.8 (via Weave Router)" },
+        "claude-haiku-4-5":  { name: "Claude Haiku 4.5 (via Weave Router)" },
+        "gpt-5.5":           { name: "GPT-5.5 (via Weave Router)" },
+        "gpt-5.4":           { name: "GPT-5.4 (via Weave Router)" }
       }
     }
   ')"
 
-  # The OpenAI-shaped Codex-subscription provider. Reuses the same X-Weave-*
-  # header triplet (router key + identity) — the plugin's loader adds only the
-  # dynamic subscription Authorization + ChatGPT-Account-Id on top. baseURL
-  # KEEPS /v1: opencode's @ai-sdk/openai provider appends /responses, yielding
-  # the router's /v1/responses Codex passthrough. apiKey is the router key as a
-  # parse-time placeholder (the loader overrides it once a subscription is
-  # linked), mirroring the weave block above.
-  local codex_block
-  codex_block="$(jq -n \
+  # Login-only storage provider for a Claude (Pro/Max) subscription. It serves
+  # no requests (no models), so its npm/baseURL are inert — opencode just needs
+  # it registered so the plugin's Claude login method has a home and `opencode
+  # auth login` lists it. The `weave` loader reads this slot off disk and
+  # attaches the Claude sub. Reuses the same router key + identity headers.
+  local claude_block
+  claude_block="$(jq -n \
     --arg url "$block_url/v1" \
     --arg key "$block_key" \
     --argjson headers "$headers_json" '
     {
       npm: "@ai-sdk/openai",
-      name: "Weave Router (Codex subscription)",
+      name: "Weave Router — Claude plan",
       options: { apiKey: $key, baseURL: $url, headers: $headers },
-      models: {
-        "gpt-5.5": { name: "GPT-5.5 — Codex subscription (via Weave Router)" },
-        "gpt-5.4": { name: "GPT-5.4 — Codex subscription (via Weave Router)" }
-      }
+      models: {}
     }
   ')"
 
@@ -653,7 +651,7 @@ write_opencode_config() {
     chmod 644 "$plugin_spec"
     plugin_arg="$plugin_spec"
   else
-    warn "opencode Codex plugin source not found at $plugin_src — skipping the weave-codex plugin registration. (Use a packaged 'npx @workweave/router' install.)"
+    warn "opencode subscription plugin source not found at $plugin_src — skipping the Claude login + subscription routing. (Use a packaged 'npx @workweave/router' install.)"
   fi
 
   # Merge into any existing opencode.json. We always overwrite provider.weave
@@ -661,21 +659,23 @@ write_opencode_config() {
   # file (other providers, mcp, agent settings) untouched. Top-level `model` is
   # only set when the user hasn't already picked one.
   #
-  # The weave-codex provider AND its plugin entry are written together only when
-  # the bundled plugin was present and copied ($plugin non-empty). Codex
-  # subscription auth depends on that plugin, so registering the provider
-  # without it (e.g. the `curl | sh` path, which carries no plugin source)
-  # would leave a non-working provider — instead we omit it (and strip any
-  # stale one from a prior install).
+  # The weave-claude login provider AND the plugin entry are written together
+  # only when the bundled plugin was present and copied ($plugin non-empty): the
+  # Claude login method lives in the plugin, so registering the provider without
+  # it (e.g. the `curl | sh` path, which carries no plugin source) would leave a
+  # non-working login — instead we omit it (and strip any stale one). The legacy
+  # `weave-codex` provider is always stripped: the single Responses `weave`
+  # provider supersedes it.
   local merged
   if [ -f "$config_file" ]; then
     merged="$(jq \
       --argjson block "$block" \
-      --argjson codex "$codex_block" \
+      --argjson claude "$claude_block" \
       --arg plugin "$plugin_arg" \
       --arg pluginspec "$plugin_spec" '
       .provider = ((.provider // {}) | .weave = $block)
-      | (if $plugin != "" then .provider["weave-codex"] = $codex else .provider |= del(."weave-codex") end)
+      | (.provider |= del(."weave-codex"))
+      | (if $plugin != "" then .provider["weave-claude"] = $claude else .provider |= del(."weave-claude") end)
       # Register the managed plugin path when we installed it; otherwise strip a
       # stale entry left by a prior install (the provider was just removed, so a
       # lingering plugin reference would be dead weight).
@@ -686,23 +686,24 @@ write_opencode_config() {
                    else . end)
          end)
       | (if (.model // "") == "" then .model = "weave/claude-sonnet-4-6" else . end)
-      # If we just stripped weave-codex (plugin-less re-install) but the default
-      # model still points at it, reset to the Anthropic weave model so opencode
-      # does not boot with a model whose provider no longer exists.
-      | (if $plugin == "" and (.model // "" | tostring | startswith("weave-codex/")) then .model = "weave/claude-sonnet-4-6" else . end)
+      # Reset the default model if it points at a provider we just removed: the
+      # retired weave-codex, or weave-claude on a plugin-less re-install (it has
+      # no models anyway). Otherwise opencode boots with a dangling model.
+      | (if (.model // "" | tostring | startswith("weave-codex/")) then .model = "weave/claude-sonnet-4-6" else . end)
+      | (if $plugin == "" and (.model // "" | tostring | startswith("weave-claude/")) then .model = "weave/claude-sonnet-4-6" else . end)
       | (.["$schema"] //= "https://opencode.ai/config.json")
     ' "$config_file")"
   else
     merged="$(jq -n \
       --argjson block "$block" \
-      --argjson codex "$codex_block" \
+      --argjson claude "$claude_block" \
       --arg plugin "$plugin_arg" '
       {
         "$schema": "https://opencode.ai/config.json",
         model: "weave/claude-sonnet-4-6",
         provider: { weave: $block }
       }
-      | (if $plugin != "" then .provider["weave-codex"] = $codex | .plugin = [$plugin] else . end)
+      | (if $plugin != "" then .provider["weave-claude"] = $claude | .plugin = [$plugin] else . end)
     ')"
   fi
   printf '%s\n' "$merged" >"$config_file"
@@ -1402,9 +1403,9 @@ toggle_claude() {
       else
         # Park just the router-owned env keys, then strip them so Claude Code
         # falls back to its own Anthropic login.
-        jq '{env: ((.env // {}) | {ANTHROPIC_BASE_URL, ANTHROPIC_CUSTOM_HEADERS} | with_entries(select(.value != null)))}' "$settings_file" >"$parked"
+        jq '{env: ((.env // {}) | {ANTHROPIC_BASE_URL, ANTHROPIC_CUSTOM_HEADERS, ENABLE_TOOL_SEARCH} | with_entries(select(.value != null)))}' "$settings_file" >"$parked"
         chmod 600 "$parked"
-        merged="$(jq '(.env // {}) |= del(.ANTHROPIC_BASE_URL, .ANTHROPIC_CUSTOM_HEADERS)
+        merged="$(jq '(.env // {}) |= del(.ANTHROPIC_BASE_URL, .ANTHROPIC_CUSTOM_HEADERS, .ENABLE_TOOL_SEARCH)
                       | (if (.env // {} | length) == 0 then del(.env) else . end)' "$settings_file")"
         printf '%s\n' "$merged" >"$settings_file"
       fi
@@ -1543,8 +1544,8 @@ toggle_opencode() {
     model="$(jq -r '.model // empty' "$f" 2>/dev/null || true)"
     if [ "$(jq -r '((.provider // {}) | has("weave"))' "$f" 2>/dev/null || true)" = "true" ]; then has_weave="true"; fi
   fi
-  # Either managed provider counts as router-on: the Anthropic-shaped `weave/…`
-  # default or a caller-selected `weave-codex/…` Codex-subscription model.
+  # A managed model counts as router-on: the Responses-shaped `weave/…` default,
+  # or a legacy `weave-codex/…` model parked by a pre-upgrade install.
   case "$model" in weave/* | weave-codex/*) on="true" ;; esac
 
   case "$mode" in
@@ -1588,10 +1589,10 @@ toggle_opencode() {
           | "weave/" + .
         ' "$f" 2>/dev/null || echo "weave/claude-sonnet-4-6")"
       fi
-      # Never restore a weave-codex model whose provider is no longer
-      # registered (e.g. a plugin-less re-install stripped it after `off`
-      # parked a weave-codex/… model). Fall back to the Anthropic weave default
-      # so `on` can't leave opencode pointing at a missing provider.
+      # Never restore a legacy weave-codex model whose provider is no longer
+      # registered (every current install strips weave-codex). Fall back to the
+      # `weave` default so `on` can't leave opencode pointing at a missing
+      # provider.
       case "$restore_model" in
         weave-codex/*)
           if [ "$(jq -r '((.provider // {}) | has("weave-codex"))' "$f" 2>/dev/null || true)" != "true" ]; then
@@ -1849,13 +1850,14 @@ if [ "$target" = "opencode" ]; then
   printf "\n"
   printf "%s✓%s %s%sWeave Router installed for opencode.%s\n" \
     "$C_GREEN" "$C_RESET" "$C_BOLD" "$C_BRAND" "$C_RESET"
-  # Surface the optional Codex-subscription path only when this run actually
-  # registered the weave-codex provider. Gate on the provider's presence in the
-  # written config (authoritative) rather than a plugin file on disk — a
-  # leftover plugin from a prior install can outlive a plugin-less re-install
-  # that stripped the provider, which would make these instructions misleading.
-  if jq -e '(.provider // {}) | has("weave-codex")' "$opencode_config_file" >/dev/null 2>&1; then
-    info "To pay for opencode turns with your own ChatGPT (Codex) subscription: run ${C_BOLD}opencode auth login${C_RESET} → ${C_BOLD}ChatGPT Pro/Plus${C_RESET}, then pick a ${C_BOLD}weave-codex${C_RESET} model."
+  # Surface the optional subscription-routing path only when this run actually
+  # registered the weave-claude login provider (which is written together with
+  # the plugin). Gate on its presence in the written config (authoritative)
+  # rather than a plugin file on disk — a leftover plugin from a prior install
+  # can outlive a plugin-less re-install that stripped the provider, which would
+  # make these instructions misleading.
+  if jq -e '(.provider // {}) | has("weave-claude")' "$opencode_config_file" >/dev/null 2>&1; then
+    info "Optional: connect your AI plans so they pay for the matching turns. Run ${C_BOLD}opencode auth login${C_RESET} → ${C_BOLD}Weave Router${C_RESET} for ${C_BOLD}ChatGPT Pro/Plus${C_RESET} (GPT/Codex turns) and/or ${C_BOLD}Weave Router — Claude plan${C_RESET} for ${C_BOLD}Claude Pro/Max${C_RESET} (Claude turns). The router still routes every turn; your Weave key pays for the rest."
   fi
   if [ -n "$install_dir" ]; then
     # --dir installs land outside opencode's discovery roots, so the caller
@@ -2069,14 +2071,18 @@ normalize_model() {
 prices='{
   "input": {
     "claude-fable-5":                   0.01,
-    "claude-haiku-4-5":                 0.0008,
+    "claude-haiku-4-5":                 0.001,
+    "claude-opus-4-0":                  0.015,
+    "claude-opus-4-1":                  0.015,
+    "claude-opus-4-5":                  0.005,
     "claude-opus-4-6":                  0.005,
     "claude-opus-4-7":                  0.005,
     "claude-opus-4-8":                  0.005,
     "claude-sonnet-4-5":                0.003,
     "claude-sonnet-4-6":                0.003,
-    "deepseek/deepseek-v4-flash":       0.00014,
-    "deepseek/deepseek-v4-pro":         0.00174,
+    "claude-sonnet-5":                  0.003,
+    "deepseek/deepseek-v4-flash":       0.0001134,
+    "deepseek/deepseek-v4-pro":         0.001318,
     "gemini-2.0-flash":                 0.0001,
     "gemini-2.0-flash-lite":            0.000075,
     "gemini-2.5-flash":                 0.0003,
@@ -2087,6 +2093,7 @@ prices='{
     "gemini-3.1-flash-lite-preview":    0.0001,
     "gemini-3.1-pro-preview":           0.002,
     "gemini-3.5-flash":                 0.0015,
+    "google/gemma-4-26b-a4b-it":        0.00015,
     "gpt-4.1":                          0.002,
     "gpt-4.1-mini":                     0.0004,
     "gpt-4.1-nano":                     0.0001,
@@ -2096,10 +2103,10 @@ prices='{
     "gpt-5-chat":                       0.0025,
     "gpt-5-mini":                       0.0005,
     "gpt-5-nano":                       0.0001,
-    "gpt-5.4":                          0.003,
-    "gpt-5.4-mini":                     0.0004,
-    "gpt-5.4-nano":                     0.0001,
-    "gpt-5.4-pro":                      0.02,
+    "gpt-5.4":                          0.0025,
+    "gpt-5.4-mini":                     0.00075,
+    "gpt-5.4-nano":                     0.0002,
+    "gpt-5.4-pro":                      0.03,
     "gpt-5.5":                          0.005,
     "gpt-5.5-mini":                     0.0005,
     "gpt-5.5-nano":                     0.00015,
@@ -2116,22 +2123,27 @@ prices='{
     "qwen/qwen3-coder-next":            0.0005,
     "qwen/qwen3-next-80b-a3b-instruct": 0.00015,
     "qwen/qwen3.5-flash-02-23":         0.00005,
-    "qwen/qwen3.6-35b-a3b":             0.00015,
+    "qwen/qwen3.6-35b-a3b":             0.000172,
     "qwen/qwen3.7-plus":                0.0004,
     "xiaomi/mimo-v2.5-pro":             0.001,
-    "z-ai/glm-5":                       0.0006,
-    "z-ai/glm-5.1":                     0.00105
+    "z-ai/glm-5":                       0.001,
+    "z-ai/glm-5.1":                     0.0014,
+    "z-ai/glm-5.2":                     0.0014
   },
   "output": {
     "claude-fable-5":                   0.05,
-    "claude-haiku-4-5":                 0.004,
+    "claude-haiku-4-5":                 0.005,
+    "claude-opus-4-0":                  0.075,
+    "claude-opus-4-1":                  0.075,
+    "claude-opus-4-5":                  0.025,
     "claude-opus-4-6":                  0.025,
     "claude-opus-4-7":                  0.025,
     "claude-opus-4-8":                  0.025,
     "claude-sonnet-4-5":                0.015,
     "claude-sonnet-4-6":                0.015,
-    "deepseek/deepseek-v4-flash":       0.00028,
-    "deepseek/deepseek-v4-pro":         0.00348,
+    "claude-sonnet-5":                  0.015,
+    "deepseek/deepseek-v4-flash":       0.0002791,
+    "deepseek/deepseek-v4-pro":         0.0026361,
     "gemini-2.0-flash":                 0.0004,
     "gemini-2.0-flash-lite":            0.0003,
     "gemini-2.5-flash":                 0.0012,
@@ -2142,6 +2154,7 @@ prices='{
     "gemini-3.1-flash-lite-preview":    0.0004,
     "gemini-3.1-pro-preview":           0.008,
     "gemini-3.5-flash":                 0.009,
+    "google/gemma-4-26b-a4b-it":        0.0006,
     "gpt-4.1":                          0.008,
     "gpt-4.1-mini":                     0.0016,
     "gpt-4.1-nano":                     0.0004,
@@ -2151,14 +2164,14 @@ prices='{
     "gpt-5-chat":                       0.01,
     "gpt-5-mini":                       0.002,
     "gpt-5-nano":                       0.0004,
-    "gpt-5.4":                          0.012,
-    "gpt-5.4-mini":                     0.0016,
-    "gpt-5.4-nano":                     0.0004,
-    "gpt-5.4-pro":                      0.08,
-    "gpt-5.5":                          0.04,
+    "gpt-5.4":                          0.015,
+    "gpt-5.4-mini":                     0.0045,
+    "gpt-5.4-nano":                     0.00125,
+    "gpt-5.4-pro":                      0.18,
+    "gpt-5.5":                          0.03,
     "gpt-5.5-mini":                     0.0025,
     "gpt-5.5-nano":                     0.0006,
-    "gpt-5.5-pro":                      0.12,
+    "gpt-5.5-pro":                      0.18,
     "minimax/minimax-m2.7":             0.0012,
     "minimax/minimax-m3":               0.0012,
     "mistralai/mistral-small-2603":     0.0006,
@@ -2171,11 +2184,12 @@ prices='{
     "qwen/qwen3-coder-next":            0.0012,
     "qwen/qwen3-next-80b-a3b-instruct": 0.0012,
     "qwen/qwen3.5-flash-02-23":         0.00015,
-    "qwen/qwen3.6-35b-a3b":             0.00095,
+    "qwen/qwen3.6-35b-a3b":             0.0012002,
     "qwen/qwen3.7-plus":                0.0016,
     "xiaomi/mimo-v2.5-pro":             0.003,
-    "z-ai/glm-5":                       0.00208,
-    "z-ai/glm-5.1":                     0.0035
+    "z-ai/glm-5":                       0.0032,
+    "z-ai/glm-5.1":                     0.0044,
+    "z-ai/glm-5.2":                     0.0044
   }
 }'
 # END_GENERATED_PRICES
@@ -2364,20 +2378,23 @@ elif [[ "$routed" == "failure" ]]; then
   # swap or compute savings against a non-model.
   printf '%s — %s%s' "$brand" "$routed" "$tokens_clause"
 elif [[ -n "$routed" ]]; then
-  # Show the savings clause only when the session is genuinely net-saving.
+  # Always show the savings clause, flooring a non-positive total at $0.00.
   # session_savings is "0.0000" on fresh sessions or sessions where every
-  # turn routed back to the selected model; it can also go negative when
-  # sticky routing forces a haiku-tagged side-call up to a cached
-  # sonnet/opus decision. In both cases the word "saved" would mislead,
-  # so drop the savings clause but keep the token totals.
-  has_savings="false"
-  if [[ -n "$session_savings" ]] \
-     && awk -v v="$session_savings" 'BEGIN{exit !(v+0 > 0.005)}'; then
-    has_savings="true"
-  fi
-  if [[ "$has_savings" == "true" ]]; then
+  # turn routed back to the selected model, and can go negative when routing
+  # upgrades a turn to a pricier model (e.g. a sticky/hard-pinned side-call,
+  # or a hard turn escalated to opus). "saved -$X" would mislead, so clamp
+  # the display to $0.00 rather than dropping the clause — a $0.00 readout
+  # tells the user the router ran and simply didn't beat their selection.
+  # When the CC selection is unknown ("?" or empty) there's nothing to
+  # compare against, so drop the "← selection" arrow and just show routed.
+  display_savings="$session_savings"
+  if [[ -z "$display_savings" ]] \
+     || awk -v v="$display_savings" 'BEGIN{exit !(v+0 < 0)}'; then
+    display_savings="0"
+  fi
+  if [[ -n "$selected_display" && "$selected_display" != "?" ]]; then
     printf '%s — %s ← %s · saved %s%s' \
-      "$brand" "$routed" "$selected_display" "$(fmt_money "$session_savings")" "$tokens_clause"
+      "$brand" "$routed" "$selected_display" "$(fmt_money "$display_savings")" "$tokens_clause"
   else
     printf '%s — %s%s' "$brand" "$routed" "$tokens_clause"
   fi
@@ -2414,14 +2431,19 @@ if [ -n "$user_name" ]; then
 fi
 custom_headers="$custom_headers"$'\n'"X-App: claude-code"
 
+# Setting ANTHROPIC_BASE_URL makes Claude Code treat us as non-first-party and
+# disable MCP tool-search deferral, inlining every tool schema into every request
+# — a large uncompactable prefix that can push a session into an autocompact
+# thrash loop. ENABLE_TOOL_SEARCH=auto restores on-demand loading (Claude Code's
+# own first-party default).
 if [ "$scope" = "project" ] && [ -z "$install_dir" ]; then
   jq -n --arg url "$base_url" --arg sl "$statusline_path_for_settings" '{
-    env: { ANTHROPIC_BASE_URL: $url },
+    env: { ANTHROPIC_BASE_URL: $url, ENABLE_TOOL_SEARCH: "auto" },
     statusLine: { type: "command", command: $sl }
   }' >"$tmp_patch"
 else
   jq -n --arg url "$base_url" --arg header "$custom_headers" --arg sl "$statusline_path_for_settings" '{
-    env: { ANTHROPIC_BASE_URL: $url, ANTHROPIC_CUSTOM_HEADERS: $header },
+    env: { ANTHROPIC_BASE_URL: $url, ANTHROPIC_CUSTOM_HEADERS: $header, ENABLE_TOOL_SEARCH: "auto" },
     statusLine: { type: "command", command: $sl }
   }' >"$tmp_patch"
 fi