@workweave/router 0.2.10 → 0.2.12

package/README.md

@@ -15,11 +15,13 @@ npx @workweave/router --non-interactive     # reads $WEAVE_ROUTER_KEY, no prompt
 ```
 
 Re-running the installer to pick up changes reuses the key already on disk, so
-you paste it once and never again. `update` is the never-prompting form of that
-(safe for cron; errors instead of asking when no key can be found):
+you paste it once and never again — for every client, not just Claude Code.
+`update` is the never-prompting form of that (safe for cron; errors instead of
+asking when no key can be found):
 
 ```bash
 npx @workweave/router --claude                # reuses the installed key
+npx @workweave/router --codex                 # same for Codex, opencode, and pi
 npx @workweave/router --claude --rotate-key   # ignore it and prompt for a new one
 npx @workweave/router update --claude         # non-interactive refresh in place
 ```
@@ -27,8 +29,11 @@ npx @workweave/router update --claude         # non-interactive refresh in place
 For Claude Code the installed statusline and `/force-model`, `/router-*` slash
 commands also refresh themselves in the background about once a week (never
 overwriting a wrapper you edited). Opt out with `WEAVE_STATUSLINE_UPDATE=0`, or
-just the commands with `WEAVE_COMMANDS_UPDATE=0`. Codex, opencode, and pi have
-no per-turn hook to refresh from — re-run the installer for those.
+just the commands with `WEAVE_COMMANDS_UPDATE=0`. Codex installs native `$` skills
+plus managed `SessionStart`/`Stop` hooks: the latest routed model is reflected in
+the terminal title and a compact `Weave Router · …` status message is shown when
+the router reports a new route. Existing Codex hooks are preserved. OpenCode and pi
+have their own target-specific integrations.
 
 Version-pin for reproducible setups:
 
@@ -48,12 +53,28 @@ npx @workweave/router status --codex    # is Codex on the router or direct?
 Claude Code reads its router setting at launch, so quit and reopen it after an
 on/off. Codex and opencode pick it up on their next run. Inside Claude Code the
 slash commands `/router-off`, `/router-on`, and `/router-status` do the same.
-Codex installs a `$disable-routing` skill that switches its next session back
+Codex installs `$router-status`, `$router-off`, `$router-on`, and
+`$router-models` skills that call the same CLI verbs, plus a
+`$disable-routing` skill that switches its next session back
 to the normal provider; Codex does not support third-party `/disable-routing`
 slash commands. The shell equivalent is `npx @workweave/router disable-routing`.
 Cursor has no config file we own — toggle its base URL override in **Settings →
 Models** instead.
 
+Pick which models the router is allowed to route to:
+
+```bash
+npx @workweave/router models --claude                  # list every model, with its on/off state
+npx @workweave/router models --codex                   # same, for a Codex install
+npx @workweave/router models disable gpt-5.6 --claude  # take one out of rotation
+npx @workweave/router models enable gpt-5.6 --claude   # put it back
+```
+
+Inside Claude Code that's `/router-models` (alias `/models`). Editing needs a
+router that serves the model-selection API; against the Weave-hosted router the
+list still prints and points you at the dashboard, where model selection is an
+organization-wide setting.
+
 Uninstall:
 
 ```bash
@@ -82,18 +103,24 @@ Four install targets:
 - **Codex** (`--codex`) — patches `~/.codex/config.toml` (or
   `<repo>/.codex/config.toml`) with a managed `[model_providers.weave]`
   block plus `model_provider = "weave"`. The provider preserves the existing
-  ChatGPT OAuth login. The public hosted endpoint sends
-  `X-Weave-Router-Strategy: hmm`; `--local` and custom self-hosted URLs keep
-  their router's configured default because its HMM sidecar is optional. HMM or forced
+  ChatGPT OAuth login. No install pins `X-Weave-Router-Strategy`; every
+  endpoint keeps its router's configured default. HMM or forced
   `gpt-5.6-sol`, `gpt-5.6-terra`, and `gpt-5.6-luna` turns use that plan;
   every other selected model uses its WorkWeave deployment or BYOK credential.
   The block lives between begin/end markers
   so re-running the installer rewrites it cleanly and `--uninstall --codex`
   removes it without touching the rest of your config. Codex does not load
-  third-party slash-command files; to send a router directive, type it with
-  one leading space (for example, ` /force-model gpt-5.6-terra`). Its
-  `$disable-routing` skill returns the next Codex session to the default
-  provider without logging out or deleting the router configuration.
+  third-party slash-command files; the installer provides native skills
+  `$force-model` (`$fm`), `$unforce-model` (`$ufm`), and `$router-feedback`
+  (`$rf`), each of which execs a local `scripts/emit.sh` that prints the same
+  leading-space directive Claude Code uses (for example,
+  ` /force-model gpt-5.6-terra`) — you can also type that form directly. It
+  also installs `$router-status`, `$router-off`, `$router-on`, and
+  `$router-models`, which call this installer's own verbs, plus a
+  `$disable-routing` skill that returns the next Codex session to the default
+  provider without logging out or deleting the router configuration. The
+  managed lifecycle hooks also keep the latest routed model in the terminal
+  title and emit a compact status message when the router reports a new route.
 - **opencode** (`--opencode`) — merges a `provider.weave` entry (backed by
   opencode's built-in `@ai-sdk/anthropic` provider) into
   `~/.config/opencode/opencode.json` (or `<repo>/opencode.json` with
@@ -115,8 +142,7 @@ for the full reference.
 
 - Node ≥ 18 (ships with `npx`)
 - `bash` on PATH (macOS / Linux native; Windows needs Git Bash or WSL)
-- `jq` on PATH — used by the Claude Code status line and the opencode/pi JSON
-  merges. Not required for the Codex path.
+- `jq` on PATH — used by the Claude Code status line, the Codex lifecycle helper, and the opencode/pi JSON merges.
 
 ## Why npx
 

package/cc-statusline.sh

@@ -16,9 +16,9 @@
 # Renders:
 #   WEAVE ROUTER — claude-sonnet-4-5 ← claude-opus-4-7 · saved $1.23 · 12.4k in / 3.1k out / 45.2k cached
 #
-# Pricing source of truth: router/eval/pricing.py. Keep these maps in lockstep
-# when prices change. Cache multipliers (1.25× / 0.1×) follow Anthropic's
-# published cache pricing and are stable across the Claude family.
+# Pricing source of truth: internal/router/catalog. Input/output prices and
+# cache-read multipliers are generated by cmd/genprices. Cache creation remains
+# at 1.25× input pending TTL-aware pricing.
 
 set -euo pipefail
 
@@ -159,6 +159,23 @@ weave_command_tracked_by_git() {
   git -C "$(dirname "$1")" ls-files --error-unmatch -- "$1" >/dev/null 2>&1
 }
 
+# weave_installed_command_names lists the wrappers this install may refresh.
+# The statusline ships standalone (no registry.sh beside it), so the installed
+# set — itself written from the registry — is the only source of truth here.
+#
+# Files written before ownership markers existed carry none, so matching on the
+# marker alone would freeze every pre-marker install out of refreshes forever.
+# List every wrapper instead and let the baseline comparison below decide: a
+# file is replaced only when its bytes still match the last canonical copy, so
+# a user-authored command is never touched whether or not it carries a marker.
+weave_installed_command_names() {
+  local dir="$1" file
+  for file in "$dir"/*.md; do
+    [ -f "$file" ] || continue
+    printf '%s\n' "$(basename "$file" .md)"
+  done
+}
+
 # weave_render_command prints $1 with the installer's {{SCOPE}} placeholder
 # replaced by $2, matching how install_slash_commands writes the same file.
 # Trailing newlines are stripped on both sides of every comparison below.
@@ -228,8 +245,7 @@ weave_sync_commands() {
     # Detach stdin (CC pipes JSON to us) so curl can't consume it, and silence
     # everything so no output leaks into the statusline.
     exec </dev/null
-    for name in force-model unforce-model router-feedback fm ufm rf \
-                router-off router-on router-status router-session; do
+    while IFS= read -r name; do
       installed="$cmd_dir/$name.md"
       # Only ever refresh a wrapper that is already installed: a missing one
       # was uninstalled or deliberately deleted, and resurrecting it would be
@@ -258,10 +274,10 @@ weave_sync_commands() {
       if [ -f "$prev" ]; then
         new_body="$(weave_render_command "$raw" "$scope_args")"
         prev_body="$(weave_render_command "$prev" "$scope_args")"
-        installed_body="$(cat "$installed" 2>/dev/null)" || installed_body=""
+        installed_body="$(cat "$installed" 2>/dev/null | sed '/^<!-- weave-router managed command: .* -->$/d')" || installed_body=""
         if [ "$prev_body" = "$installed_body" ] && [ "$new_body" != "$installed_body" ]; then
           tmp="$installed.tmp.$$"
-          if printf '%s\n' "$new_body" >"$tmp" 2>/dev/null; then
+          if printf '%s\n<!-- weave-router managed command: %s -->' "$new_body" "$name" >"$tmp" 2>/dev/null; then
             mv "$tmp" "$installed" 2>/dev/null || rm -f "$tmp"
           else
             rm -f "$tmp"
@@ -269,13 +285,152 @@ weave_sync_commands() {
         fi
       fi
       mv "$raw" "$prev" 2>/dev/null || rm -f "$raw"
-    done
+    done <<EOF
+$(weave_installed_command_names "$cmd_dir")
+EOF
   ) >/dev/null 2>&1 &
   disown 2>/dev/null || true
   return 0
 }
 weave_sync_commands 2>/dev/null || true
 
+# ---------- org "hide terminal surfaces" gate ----------
+#
+# When the org has hidden the router's terminal surfaces, the statusline
+# renders nothing: this prints no output and exits 0, leaving the slot blank.
+# The setting comes from GET /v1/display-settings with the install's own
+# router key, but the foreground path NEVER touches the network: it decides
+# solely from a per-install cache file (TTL WEAVE_DISPLAY_SETTINGS_TTL_SECONDS,
+# default 1h), and a missing or stale cache fails open — the statusline
+# renders normally. When the cache is missing or stale a detached background
+# refresh re-fetches the setting and rewrites the cache atomically, so the
+# next turn picks up the fresh value; a refresh failure simply leaves the
+# cache stale, which keeps failing open rather than pinning the gate closed.
+#
+# Claude Code runs the statusline every turn, so several invocations can see a
+# stale cache at once. Letting each fetch independently is not safe: the
+# responses can land out of order, and the loser's mv would replace a newer
+# setting with an older one AND stamp it fresh, pinning the gate on a stale
+# value for a full TTL. Refreshes are therefore serialized per cache key on a
+# mkdir mutex (atomic everywhere; flock is absent on macOS), held across both
+# the fetch and the write. A refresh that finds the mutex held exits rather
+# than queueing — the in-flight one is at least as fresh as anything it would
+# fetch, so waiting only to overwrite it is the bug. The one path that is not
+# strictly exclusive is reclaiming a lock whose holder died; see below.
+weave_hidden_gate() {
+  command -v curl >/dev/null 2>&1 || return 1
+  command -v jq >/dev/null 2>&1 || return 1
+
+  local cache_dir="${XDG_CACHE_HOME:-$HOME/.cache}/weave-router"
+  mkdir -p "$cache_dir" 2>/dev/null || return 1
+  local self="${BASH_SOURCE[0]:-$0}"
+  local script_slug
+  script_slug="$(printf '%s' "$self" | tr -c 'A-Za-z0-9._-' '_')"
+  local cache="$cache_dir/display-settings${script_slug}"
+  local ttl="${WEAVE_DISPLAY_SETTINGS_TTL_SECONDS:-3600}"
+
+  local now mtime fresh="false"
+  now="$(date +%s 2>/dev/null)" || now=0
+  if [ -f "$cache" ]; then
+    mtime="$(stat -c %Y "$cache" 2>/dev/null || stat -f %m "$cache" 2>/dev/null)" || mtime=0
+    if [ -n "${mtime:-}" ] && [ "$mtime" -gt 0 ] && [ $(( now - mtime )) -lt "$ttl" ]; then
+      fresh="true"
+    fi
+  fi
+
+  # Foreground decision: only a fresh cache hides the surfaces. Anything else
+  # (no cache, stale cache, unreadable cache) renders normally so a slow or
+  # unreachable router can never stall a turn or wedge the statusline blank.
+  if [ "$fresh" = "true" ]; then
+    [ "$(cat "$cache" 2>/dev/null)" = "1" ]
+    return
+  fi
+
+  # Background refresh for the next invocation. Resolve the router base URL
+  # and key inside the subshell from the Claude Code settings the installer
+  # wrote: project/--dir installs put both under <base>/.claude alongside
+  # this script (key in settings.local.json), while a user-scope install
+  # lives under ~/.weave and reads ~/.claude. Resolve relative to the
+  # script's own location, falling back to user scope, so a project install
+  # never reads (or leaks) the user-scope key. ANTHROPIC_BASE_URL and
+  # WEAVE_ROUTER_BASE_URL may also be set in the environment. A file:// base
+  # URL is the offline/test seam: curl reads it as the response body
+  # directly, so the endpoint path is meaningless for it; real router URLs
+  # (https) get /v1/display-settings appended.
+  (
+    exec </dev/null
+
+    # Take the per-cache-key mutex, or bail. mkdir is the portable atomic
+    # test-and-set. A crashed holder would otherwise block refreshes forever,
+    # so a lock older than the fetch timeout is treated as abandoned and
+    # reclaimed. Releasing from a trap covers every exit path below.
+    lock="$cache.lock"
+    if ! mkdir "$lock" 2>/dev/null; then
+      lock_mtime="$(stat -c %Y "$lock" 2>/dev/null || stat -f %m "$lock" 2>/dev/null)" || lock_mtime=0
+      lock_now="$(date +%s 2>/dev/null)" || lock_now=0
+      if [ "${lock_mtime:-0}" -le 0 ] || [ $(( lock_now - lock_mtime )) -le 30 ]; then
+        exit 0
+      fi
+      # Reclaiming an abandoned lock must not be delete-then-recreate. With
+      # `rm -rf` + `mkdir`, refreshers that all see the same stale lock each
+      # delete the next one's freshly created directory, so many end up holding
+      # it at once (measured at 50 claimants: 4-11 concurrent holders) and their
+      # writes can land out of order. Renaming is atomic, so only one racer can
+      # move a given directory aside and the losers exit instead of clobbering
+      # the winner (same measurement: 1-2). It is not a perfect mutex — a
+      # straggler can still reclaim the new lock a winner just created, since
+      # nothing distinguishes it from the stale one — but real refreshes arrive
+      # one per turn rather than 50 at once, and the staleness threshold below
+      # is what bounds the rest.
+      dead="$lock.dead.$$"
+      mv "$lock" "$dead" 2>/dev/null || exit 0
+      rm -rf "$dead" 2>/dev/null
+      mkdir "$lock" 2>/dev/null || exit 0
+    fi
+    trap 'rmdir "$lock" 2>/dev/null' EXIT
+
+    self_dir="$(cd "$(dirname "$self")" 2>/dev/null && pwd)"
+    settings_base="$HOME"
+    case "$self_dir" in
+      */.claude) settings_base="${self_dir%/.claude}" ;;
+    esac
+    settings="$settings_base/.claude/settings.json"
+    local_settings="$settings_base/.claude/settings.local.json"
+    base_url="${WEAVE_ROUTER_BASE_URL:-${ANTHROPIC_BASE_URL:-}}"
+    key="${WEAVE_ROUTER_KEY:-}"
+    if [ -z "$key" ] && [ -f "$settings" ]; then
+      key="$(jq -r '.env.ANTHROPIC_CUSTOM_HEADERS // "" | split("\n")[] | select(startswith("X-Weave-Router-Key:")) | sub("^X-Weave-Router-Key:[[:space:]]*";"")' "$settings" 2>/dev/null | head -n1)"
+    fi
+    if [ -z "$key" ] && [ -f "$local_settings" ]; then
+      key="$(jq -r '.env.ANTHROPIC_CUSTOM_HEADERS // "" | split("\n")[] | select(startswith("X-Weave-Router-Key:")) | sub("^X-Weave-Router-Key:[[:space:]]*";"")' "$local_settings" 2>/dev/null | head -n1)"
+    fi
+    if [ -z "$base_url" ] && [ -f "$settings" ]; then
+      base_url="$(jq -r '.env.ANTHROPIC_BASE_URL // empty' "$settings" 2>/dev/null)"
+    fi
+    [ -n "$base_url" ] && [ -n "$key" ] || exit 0
+    url="${base_url%/}"
+    case "$url" in
+      file://*) ;;
+      *) url="$url/v1/display-settings" ;;
+    esac
+    body="$(curl -fsS --max-time 5 -H "X-Weave-Router-Key: $key" "$url" 2>/dev/null)" || exit 0
+    hidden="$(printf '%s' "$body" | jq -r '.hide_terminal_surfaces // false' 2>/dev/null)"
+    tmp="$cache.tmp.$$"
+    if [ "$hidden" = "true" ]; then
+      printf '1' >"$tmp" 2>/dev/null && mv "$tmp" "$cache" 2>/dev/null
+    else
+      printf '0' >"$tmp" 2>/dev/null && mv "$tmp" "$cache" 2>/dev/null
+    fi
+    rm -f "$tmp" 2>/dev/null
+  ) >/dev/null 2>&1 &
+  disown 2>/dev/null || true
+  return 1
+}
+
+if weave_hidden_gate </dev/null; then
+  exit 0
+fi
+
 input="$(cat)"
 transcript_path="$(printf '%s' "$input" | jq -r '.transcript_path // empty')"
 # Prefer model.id over display_name: pricing keys + the routed model id in
@@ -316,6 +471,7 @@ prices='{
     "claude-sonnet-5":                  0.003,
     "deepseek/deepseek-v4-flash":       0.0001134,
     "deepseek/deepseek-v4-pro":         0.00174,
+    "deepseek/deepseek-v4-pro-0813":    0.00174,
     "gemini-2.0-flash":                 0.0001,
     "gemini-2.0-flash-lite":            0.000075,
     "gemini-2.5-flash":                 0.0003,
@@ -328,6 +484,7 @@ prices='{
     "gemini-3.5-flash":                 0.0015,
     "gemini-3.5-flash-lite":            0.0003,
     "gemini-3.6-flash":                 0.0015,
+    "gemini-3.7-flash":                 0.0015,
     "google/gemma-4-26b-a4b-it":        0.00015,
     "gpt-4.1":                          0.002,
     "gpt-4.1-mini":                     0.0004,
@@ -347,7 +504,9 @@ prices='{
     "gpt-5.5-nano":                     0.00015,
     "gpt-5.5-pro":                      0.03,
     "gpt-5.6-luna":                     0.001,
+    "gpt-5.6-luna-pro":                 0.001,
     "gpt-5.6-sol":                      0.005,
+    "gpt-5.6-sol-pro":                  0.005,
     "gpt-5.6-terra":                    0.0025,
     "grok-4.5":                         0.002,
     "grok-4.6":                         0.002,
@@ -370,7 +529,9 @@ prices='{
     "xiaomi/mimo-v2.5-pro":             0.001,
     "z-ai/glm-5":                       0.001,
     "z-ai/glm-5.1":                     0.0014,
-    "z-ai/glm-5.2":                     0.0014
+    "z-ai/glm-5.2":                     0.0014,
+    "z-ai/glm-5.3":                     0.0014,
+    "z-ai/glm-5.3-flash":               0.00015
   },
   "output": {
     "claude-fable-5":                   0.05,
@@ -387,6 +548,7 @@ prices='{
     "claude-sonnet-5":                  0.015,
     "deepseek/deepseek-v4-flash":       0.0002791,
     "deepseek/deepseek-v4-pro":         0.00348,
+    "deepseek/deepseek-v4-pro-0813":    0.00348,
     "gemini-2.0-flash":                 0.0004,
     "gemini-2.0-flash-lite":            0.0003,
     "gemini-2.5-flash":                 0.0012,
@@ -399,6 +561,7 @@ prices='{
     "gemini-3.5-flash":                 0.009,
     "gemini-3.5-flash-lite":            0.0025,
     "gemini-3.6-flash":                 0.0075,
+    "gemini-3.7-flash":                 0.0075,
     "google/gemma-4-26b-a4b-it":        0.0006,
     "gpt-4.1":                          0.008,
     "gpt-4.1-mini":                     0.0016,
@@ -418,7 +581,9 @@ prices='{
     "gpt-5.5-nano":                     0.0006,
     "gpt-5.5-pro":                      0.18,
     "gpt-5.6-luna":                     0.006,
+    "gpt-5.6-luna-pro":                 0.006,
     "gpt-5.6-sol":                      0.03,
+    "gpt-5.6-sol-pro":                  0.03,
     "gpt-5.6-terra":                    0.015,
     "grok-4.5":                         0.006,
     "grok-4.6":                         0.006,
@@ -441,7 +606,86 @@ prices='{
     "xiaomi/mimo-v2.5-pro":             0.003,
     "z-ai/glm-5":                       0.0032,
     "z-ai/glm-5.1":                     0.0044,
-    "z-ai/glm-5.2":                     0.0044
+    "z-ai/glm-5.2":                     0.0044,
+    "z-ai/glm-5.3":                     0.0044,
+    "z-ai/glm-5.3-flash":               0.0005
+  },
+  "cache_read": {
+    "claude-fable-5":                   0.1,
+    "claude-haiku-4-5":                 0.1,
+    "claude-opus-4-0":                  0.1,
+    "claude-opus-4-1":                  0.1,
+    "claude-opus-4-5":                  0.1,
+    "claude-opus-4-6":                  0.1,
+    "claude-opus-4-7":                  0.1,
+    "claude-opus-4-8":                  0.1,
+    "claude-opus-5":                    0.1,
+    "claude-sonnet-4-5":                0.1,
+    "claude-sonnet-4-6":                0.1,
+    "claude-sonnet-5":                  0.1,
+    "deepseek/deepseek-v4-flash":       0.2,
+    "deepseek/deepseek-v4-pro":         0.11494252873563218,
+    "deepseek/deepseek-v4-pro-0813":    0.11494252873563218,
+    "gemini-2.0-flash":                 0.25,
+    "gemini-2.0-flash-lite":            0.25,
+    "gemini-2.5-flash":                 0.1,
+    "gemini-2.5-flash-lite":            0.1,
+    "gemini-2.5-pro":                   0.1,
+    "gemini-3-flash-preview":           0.1,
+    "gemini-3-pro-preview":             0.1,
+    "gemini-3.1-flash-lite-preview":    0.1,
+    "gemini-3.1-pro-preview":           0.1,
+    "gemini-3.5-flash":                 0.1,
+    "gemini-3.5-flash-lite":            0.1,
+    "gemini-3.6-flash":                 0.1,
+    "gemini-3.7-flash":                 0.1,
+    "google/gemma-4-26b-a4b-it":        0.1,
+    "gpt-4.1":                          0.25,
+    "gpt-4.1-mini":                     0.25,
+    "gpt-4.1-nano":                     0.25,
+    "gpt-4o":                           0.5,
+    "gpt-4o-mini":                      0.5,
+    "gpt-5":                            0.1,
+    "gpt-5-chat":                       0.1,
+    "gpt-5-mini":                       0.1,
+    "gpt-5-nano":                       0.1,
+    "gpt-5.4":                          0.1,
+    "gpt-5.4-mini":                     0.1,
+    "gpt-5.4-nano":                     0.1,
+    "gpt-5.4-pro":                      1,
+    "gpt-5.5":                          0.1,
+    "gpt-5.5-mini":                     0.1,
+    "gpt-5.5-nano":                     0.1,
+    "gpt-5.5-pro":                      1,
+    "gpt-5.6-luna":                     0.1,
+    "gpt-5.6-luna-pro":                 0.1,
+    "gpt-5.6-sol":                      0.1,
+    "gpt-5.6-sol-pro":                  0.1,
+    "gpt-5.6-terra":                    0.1,
+    "grok-4.5":                         0.25,
+    "grok-4.6":                         0.25,
+    "minimax/minimax-m2.7":             0.2,
+    "minimax/minimax-m3":               0.2,
+    "mistralai/mistral-small-2603":     0.1,
+    "moonshotai/kimi-k2.5":             0.5,
+    "moonshotai/kimi-k2.6":             0.1684,
+    "moonshotai/kimi-k2.7":             0.2,
+    "moonshotai/kimi-k3":               0.1,
+    "qwen/qwen3-235b-a22b-2507":        0.5,
+    "qwen/qwen3-30b-a3b-instruct-2507": 0.1684,
+    "qwen/qwen3-coder":                 0.1684,
+    "qwen/qwen3-coder-next":            0.5,
+    "qwen/qwen3-next-80b-a3b-instruct": 0.5,
+    "qwen/qwen3.5-flash-02-23":         0.1,
+    "qwen/qwen3.6-35b-a3b":             0.1,
+    "qwen/qwen3.7-plus":                0.2,
+    "qwen/qwen3.8-max":                 0.125,
+    "xiaomi/mimo-v2.5-pro":             0.1,
+    "z-ai/glm-5":                       0.2,
+    "z-ai/glm-5.1":                     0.18571428571428572,
+    "z-ai/glm-5.2":                     0.18571428571428572,
+    "z-ai/glm-5.3":                     0.18571428571428572,
+    "z-ai/glm-5.3-flash":               0.2
   }
 }'
 # END_GENERATED_PRICES
@@ -496,6 +740,12 @@ if [[ -n "$transcript_path" && -f "$transcript_path" ]]; then
   #   * unrecognized model        → "… isn't a recognized model · keeping
   #                                  automatic routing" — a NO-OP: the prior
   #                                  pin, if any, is left untouched
+  #   * listing (historical)      → "… pick a model by id …" / "no models are
+  #                                  available to pin …" — also NO-OPs. The
+  #                                  router no longer emits these, but a
+  #                                  transcript written while it did still
+  #                                  carries them, and they must not read as
+  #                                  a clear or a live pin loses its [forced]
   # These persist on disk (the ingress stripper only scrubs them from upstream
   # requests). Classify each weave-router turn newest-first, skip the no-op
   # "rejected" acks, and let the latest real state change decide: an "applied"
@@ -509,6 +759,7 @@ if [[ -n "$transcript_path" && -f "$transcript_path" ]]; then
         | ([.message.content[]? | select(.type? == "text") | .text] | join(" ") | gsub("[\n\r]"; " ")) as $t
         | if ($t | test("force-model applied:")) then "APPLIED " + ($t | capture("force-model applied: (?<m>[^ ]+)").m)
           elif ($t | test("isn.t a recognized model")) then "REJECTED"
+          elif ($t | test("pick a model by id|no models are available to pin")) then "REJECTED"
           else "CLEARED" end' 2>/dev/null \
     | grep -m1 -v '^REJECTED$' || true)"
   if [[ "$force_state" == APPLIED\ * ]]; then
@@ -519,12 +770,9 @@ if [[ -n "$transcript_path" && -f "$transcript_path" ]]; then
   # Compute a session running total: savings across every assistant turn
   # whose marker reports a requested ≠ routed swap, plus cumulative token
   # counts across every assistant turn (rerouted or not — total work the
-  # session has done). cache_creation is priced at 1.25× input, cache_read
-  # at 0.1× — both ratios are stable across the Claude family and a no-op
-  # when the provider doesn't return those fields. Cache reads ARE included
-  # in the savings comparison: both costs apply the same 0.1× weight to
-  # cache_read_input_tokens, so the delta reflects the model-price
-  # difference on the cached portion as well.
+  # session has done). cache_creation is priced at 1.25× input; cache_read
+  # uses each model's generated catalog multiplier. Both are no-ops when the
+  # provider does not return those fields.
   #
   # The marker regex tolerates the optional "(<provider>)" segment and a
   # `[1m]` / `-YYYYMMDD` suffix on either model name so transcripts written
@@ -557,14 +805,17 @@ if [[ -n "$transcript_path" && -f "$transcript_path" ]]; then
       } as $t |
       (if $requested == "" or $requested == $rm then 0
        else
-         ($p.input[$rm] // null)        as $rin  | ($p.output[$rm] // null)        as $rout |
-         ($p.input[$requested] // null) as $sin  | ($p.output[$requested] // null) as $sout |
-         if ($rin == null or $rout == null or $sin == null or $sout == null) then 0
+         ($p.input[$rm] // null)             as $rin  | ($p.output[$rm] // null)             as $rout |
+         ($p.cache_read[$rm] // null)        as $rcr  |
+         ($p.input[$requested] // null)      as $sin  | ($p.output[$requested] // null)      as $sout |
+         ($p.cache_read[$requested] // null) as $scr  |
+         if ($rin == null or $rout == null or $rcr == null or $sin == null or $sout == null or $scr == null) then 0
          else
-           (($t.in + 1.25 * $t.cwrt + 0.1 * $t.crd) / 1000) as $input_units |
-           ($t.out / 1000)                                  as $output_units |
-           ($input_units * $rin + $output_units * $rout)    as $routed_cost |
-           ($input_units * $sin + $output_units * $sout)    as $requested_cost |
+           (($t.in + 1.25 * $t.cwrt + $rcr * $t.crd) / 1000) as $routed_input_units |
+           (($t.in + 1.25 * $t.cwrt + $scr * $t.crd) / 1000) as $requested_input_units |
+           ($t.out / 1000)                                    as $output_units |
+           ($routed_input_units * $rin + $output_units * $rout)       as $routed_cost |
+           ($requested_input_units * $sin + $output_units * $sout)    as $requested_cost |
            ($requested_cost - $routed_cost)
          end
        end) as $savings |

package/codex-skills/fm/SKILL.md

@@ -0,0 +1,15 @@
+---
+name: fm
+description: "Alias for force-model — pin this Codex session to a specific model through the Weave Router."
+---
+
+<!-- weave-router managed fm skill -->
+
+When the user invokes `$fm <model-id>`, run this skill's `scripts/emit.sh` using the
+absolute path shown in this skill's `<path>` (its directory + `/scripts/emit.sh`).
+Pass the model id as arguments, quoted as a single argument list.
+Do not send a user message and do not type a Codex slash command.
+
+The script prints a leading-space `/force-model` line. The Weave Router reads that
+exec output, pins the session, and continues this turn — so run the script once and
+do not repeat its output as your answer. Then report the router's response.

package/codex-skills/fm/scripts/emit.sh

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+# Alias of force-model/scripts/emit.sh
+set -euo pipefail
+if [ "$#" -eq 0 ] || [ -z "${1:-}" ]; then
+  echo "usage: emit.sh <model-id>" >&2
+  exit 1
+fi
+printf ' /force-model %s\n' "$*"

package/codex-skills/force-model/SKILL.md

@@ -0,0 +1,15 @@
+---
+name: force-model
+description: "Pin this Codex session to a specific model through the Weave Router."
+---
+
+<!-- weave-router managed force-model skill -->
+
+When the user invokes `$force-model <model-id>` (or `$fm`), run this skill's `scripts/emit.sh` using the
+absolute path shown in this skill's `<path>` (its directory + `/scripts/emit.sh`).
+Pass the model id as arguments, quoted as a single argument list.
+Do not send a user message and do not type a Codex slash command.
+
+The script prints a leading-space `/force-model` line. The Weave Router reads that
+exec output, pins the session, and continues this turn — so run the script once and
+do not repeat its output as your answer. Then report the router's response.

package/codex-skills/force-model/scripts/emit.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+# Print a leading-space /force-model directive. Codex runs this via exec;
+# the Weave Router intercepts the tool output and pins the session.
+set -euo pipefail
+if [ "$#" -eq 0 ] || [ -z "${1:-}" ]; then
+  echo "usage: emit.sh <model-id>" >&2
+  exit 1
+fi
+printf ' /force-model %s\n' "$*"

package/codex-skills/rf/SKILL.md

@@ -0,0 +1,15 @@
+---
+name: rf
+description: "Alias for router-feedback — submit feedback about a Weave Router decision or model performance."
+---
+
+<!-- weave-router managed rf skill -->
+
+When the user invokes `$rf <feedback>`, run this skill's `scripts/emit.sh` using the
+absolute path shown in this skill's `<path>` (its directory + `/scripts/emit.sh`).
+Pass the feedback text as arguments, quoted as a single argument list.
+Do not send a user message and do not type a Codex slash command.
+
+The script prints a leading-space `/router-feedback` line. The Weave Router reads that
+exec output, records the feedback, and continues this turn — so run the script once and
+do not repeat its output as your answer. Then report the router's response.

package/codex-skills/rf/scripts/emit.sh

@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+set -euo pipefail
+if [ "$#" -eq 0 ] || [ -z "${1:-}" ]; then
+  echo "usage: emit.sh <feedback>" >&2
+  exit 1
+fi
+printf ' /router-feedback %s\n' "$*"

package/codex-skills/router-feedback/SKILL.md

@@ -0,0 +1,15 @@
+---
+name: router-feedback
+description: "Submit feedback about a Weave Router decision or model performance."
+---
+
+<!-- weave-router managed router-feedback skill -->
+
+When the user invokes `$router-feedback <feedback>` (or `$rf`), run this skill's `scripts/emit.sh` using the
+absolute path shown in this skill's `<path>` (its directory + `/scripts/emit.sh`).
+Pass the feedback text as arguments, quoted as a single argument list.
+Do not send a user message and do not type a Codex slash command.
+
+The script prints a leading-space `/router-feedback` line. The Weave Router reads that
+exec output, records the feedback, and continues this turn — so run the script once and
+do not repeat its output as your answer. Then report the router's response.

package/codex-skills/router-feedback/scripts/emit.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+# Print a leading-space /router-feedback directive. Codex runs this via exec;
+# the Weave Router intercepts the tool output and records the feedback.
+set -euo pipefail
+if [ "$#" -eq 0 ] || [ -z "${1:-}" ]; then
+  echo "usage: emit.sh <feedback>" >&2
+  exit 1
+fi
+printf ' /router-feedback %s\n' "$*"