@workweave/router 0.2.6 → 0.2.8

package/install.sh

@@ -52,6 +52,7 @@
 #   npx @workweave/router off --claude                     # route directly to Anthropic again (Claude Code)
 #   npx @workweave/router on --codex                       # route through the Weave Router again (Codex)
 #   npx @workweave/router status --opencode                # report whether opencode is on the router or direct
+#   npx @workweave/router disable-routing                   # Codex shortcut: use Codex's normal provider again
 # Claude Code reads env at launch, so an off/on takes effect on the next
 # `claude` start; Codex and opencode re-read config every invocation.
 # Cursor's base URL lives in its own settings UI (no file we own), so there's
@@ -61,8 +62,9 @@ set -euo pipefail
 
 # ---------- defaults ----------
 
-# The hosted Weave Router URL. Override with --base-url for self-hosted.
-DEFAULT_BASE_URL="${WEAVE_ROUTER_URL:-https://router.workweave.ai}"
+# The public hosted Weave Router URL. Override with --base-url for self-hosted.
+HOSTED_BASE_URL="https://router.workweave.ai"
+DEFAULT_BASE_URL="${WEAVE_ROUTER_URL:-$HOSTED_BASE_URL}"
 
 
 scope="user"
@@ -88,6 +90,7 @@ target_explicit="false"
 # "status" toggle or report an existing install without touching the router
 # key/identity — see the toggle_* helpers and the dispatch block below.
 mode="install"
+disable_routing_alias="false"
 
 # ---------- helpers ----------
 
@@ -297,6 +300,7 @@ refuse_if_symlink() {
 # --codex) can find and replace the block instead of duplicating it.
 WEAVE_CODEX_BEGIN_MARKER="# >>> weave-router managed (do not edit between markers) >>>"
 WEAVE_CODEX_END_MARKER="# <<< weave-router managed <<<"
+WEAVE_CODEX_SKILL_MARKER="<!-- weave-router managed disable-routing skill -->"
 
 # ---------- identity helpers ----------
 #
@@ -415,8 +419,12 @@ resolve_user_email() {
 
 # write_codex_config writes a managed [model_providers.weave] block to the
 # Codex CLI's config.toml. Sets `model_provider = "weave"` at the top level so
-# Codex picks the routed provider by default. Both lines live inside the
-# managed-block markers so uninstall removes them cleanly. We strip any
+# Codex picks the routed provider by default. The provider requires OpenAI
+# authentication, preserving the user's ChatGPT plan credential while the
+# router key is sent independently. The router applies OAuth only to its
+# native gpt-5.6 Sol/Terra/Luna family; all other routed models use WorkWeave
+# deployment or BYOK credentials. Both settings live inside the managed-block
+# markers so uninstall removes them cleanly. We strip any
 # top-level `model_provider = ...` declaration OUTSIDE the markers before
 # appending so the file doesn't end up with a duplicate key (TOML rejects
 # that). Inline `model_provider` keys inside `[profiles.*]` sections stay
@@ -462,6 +470,14 @@ write_codex_config() {
   # that share the same router key. The router otherwise has to guess from
   # User-Agent.
   headers_parts="${headers_parts}, \"X-App\" = \"codex\""
+  # Match the public hosted Claude Code installation: Codex needs this
+  # explicit policy selection because that endpoint otherwise uses its
+  # cluster default. Self-hosted endpoints may not run the optional HMM
+  # sidecar, so every custom URL deliberately keeps the router's own default.
+  # Keep this inside the managed block so a re-install can change it safely.
+  if [ "$block_url" = "$HOSTED_BASE_URL" ]; then
+    headers_parts="${headers_parts}, \"X-Weave-Router-Strategy\" = \"hmm\""
+  fi
   local headers_line="http_headers = { ${headers_parts} }"
 
   local block
@@ -476,6 +492,7 @@ model_provider = "weave"
 name = "Weave Router"
 base_url = "${esc_url}/v1"
 wire_api = "responses"
+requires_openai_auth = true
 ${headers_line}
 ${WEAVE_CODEX_END_MARKER}
 TOML
@@ -774,7 +791,8 @@ write_pi_models_config() {
         { id: "claude-sonnet-4-6", name: "Claude Sonnet 4.6 (via Weave Router)", reasoning: true, input: ["text","image"], contextWindow: 200000, maxTokens: 64000 },
         { id: "claude-haiku-4-5",  name: "Claude Haiku 4.5 (via Weave Router)",  reasoning: true, input: ["text","image"], contextWindow: 200000, maxTokens: 32000 },
         { id: "gpt-5.6-sol",       name: "GPT-5.6 Sol (via Weave Router)",       reasoning: true, input: ["text","image"], contextWindow: 1050000, maxTokens: 128000 },
-        { id: "grok-4.5",          name: "Grok 4.5 (via Weave Router)",          reasoning: true, input: ["text","image"], contextWindow: 500000, maxTokens: 131072 }
+        { id: "grok-4.5",          name: "Grok 4.5 (via Weave Router)",          reasoning: true, input: ["text","image"], contextWindow: 500000, maxTokens: 131072 },
+        { id: "grok-4.6",          name: "Grok 4.6 (via Weave Router)",          reasoning: true, input: ["text","image"], contextWindow: 500000, maxTokens: 131072 }
       ]
     }
   ')"
@@ -944,6 +962,12 @@ while [ $# -gt 0 ]; do
       # npm wrapper forwards argv verbatim so either form reaches us.
       mode="${1#--}"; shift
       ;;
+    disable-routing|--disable-routing)
+      # Convenience alias for the Codex-specific off toggle. Unlike generic
+      # `off`, it deliberately resolves the target after all flags are parsed,
+      # so `disable-routing --claude` cannot silently change Claude settings.
+      mode="off"; disable_routing_alias="true"; shift
+      ;;
     -h|--help)
       usage 0
       ;;
@@ -953,6 +977,15 @@ while [ $# -gt 0 ]; do
   esac
 done
 
+if [ "$disable_routing_alias" = "true" ]; then
+  if [ "$target_explicit" = "true" ] && [ "$target" != "codex" ]; then
+    err "disable-routing only applies to Codex; omit the client flag or use --codex."
+    exit 2
+  fi
+  target="codex"
+  target_explicit="true"
+fi
+
 # Toggle verbs only flip config install.sh already wrote: no key, no identity,
 # no prompts. Require an explicit client so we never guess which config to
 # touch, and suppress every interactive prompt downstream.
@@ -1681,20 +1714,22 @@ else
   info "No identity set — router traffic will be attributed by account UUID only."
 fi
 
-# ---------- slash command wrappers (shared by both targets) ----------
+# ---------- slash command wrappers (Claude Code and opencode) ----------
 #
-# Claude Code intercepts any prompt starting with "/" as a local slash command,
-# so a typed /force-model would resolve to "Unknown command" and never reach
-# the router. Codex CLI does the same (its built-in / menu has its own set).
-# Drop wrapper markdown files into the per-target commands directory so the
-# slash invocation expands locally into a literal "/force-model …" prompt that
-# the router's first-line parser picks up.
+# Claude Code and opencode expand command Markdown locally rather than sending
+# the literal invocation to the model, so they need wrappers that expand to a
+# router directive.
 #
 # Layout:
 #   Claude:  <settings_dir>/commands/{force-model,unforce-model}.md  → /force-model
-#   Codex:   <codex_dir>/prompts/{force-model,unforce-model}.md      → /prompts:force-model
 #   opencode: <commands_dir>/{force-model,unforce-model}.md          → /force-model
 #
+# Codex intentionally is not listed: its slash menu only exposes built-in
+# commands and does not load these Markdown wrappers. A user can send a router
+# directive by starting it with one literal space (for example,
+# ` /force-model gpt-5.6-terra`); Codex submits the trimmed prompt to the
+# router without interpreting it as a local slash command.
+#
 # Files come from install/commands/ in the repo (or the colocated commands/
 # directory the npm package ships alongside install.sh).
 install_slash_commands() {
@@ -1716,11 +1751,11 @@ install_slash_commands() {
   fi
   mkdir -p "$dst_dir"
 
-  # force-model/unforce-model are router-intercepted prompt expansions and
-  # apply to every target. The router-off/on/status wrappers shell out to this
-  # installer to flip the *local* config, so they're Claude Code-only and need
-  # the install scope baked into the command (the .md can't discover it at
-  # invocation time). {{SCOPE}} is substituted accordingly.
+  # force-model/unforce-model are router-intercepted prompt expansions for the
+  # command-capable clients. The router-off/on/status wrappers shell out to
+  # this installer to flip the *local* config, so they're Claude Code-only and
+  # need the install scope baked into the command (the .md can't discover it
+  # at invocation time). {{SCOPE}} is substituted accordingly.
   installed="force-model, unforce-model, router-feedback, fm, ufm, rf"
   cmds="force-model unforce-model router-feedback fm ufm rf"
   if [ "$target" = "claude" ]; then
@@ -1755,12 +1790,103 @@ install_slash_commands() {
   ok "Slash commands written to $dst_dir ($installed)"
 }
 
+# Codex builds before custom prompt discovery existed were given wrapper files
+# under ~/.codex/prompts. They were never invocable as the advertised aliases.
+# Remove only byte-for-byte copies of our old wrappers; retain any prompt the
+# user has changed or created independently.
+remove_obsolete_codex_prompt_wrappers() {
+  local dst_dir="$1"
+  local commands_src_dir=""
+  local candidate cmd src dst
+
+  [ -d "$dst_dir" ] || return 0
+  for candidate in \
+    "$script_dir/commands" \
+    "$script_dir/../commands"
+  do
+    if [ -d "$candidate" ]; then
+      commands_src_dir="$candidate"
+      break
+    fi
+  done
+  [ -n "$commands_src_dir" ] || return 0
+
+  if [ "$scope" = "project" ] || [ -n "$install_dir" ]; then
+    refuse_if_symlink "$dst_dir"
+  fi
+
+  for cmd in force-model unforce-model router-feedback fm ufm rf; do
+    src="$commands_src_dir/$cmd.md"
+    dst="$dst_dir/$cmd.md"
+    [ -f "$src" ] && [ -f "$dst" ] || continue
+    # A symlink is user-controlled at user scope. Leave it untouched; project
+    # scope rejects it through the same guard used by normal installs.
+    [ -L "$dst" ] && continue
+    if [ "$scope" = "project" ] || [ -n "$install_dir" ]; then
+      refuse_if_symlink "$dst"
+    fi
+    cmp -s "$src" "$dst" && rm -f "$dst"
+  done
+  rmdir "$dst_dir" 2>/dev/null || true
+}
+
+# Codex skills are the supported local extension surface for an action that
+# must edit Codex's configuration. A literal third-party slash command cannot
+# do this because Codex reserves `/…` for built-ins. The skill is invoked as
+# `$disable-routing`, asks Codex to run the existing safe toggle, and leaves
+# the managed router block in place for a later re-enable.
+install_codex_disable_routing_skill() {
+  local skill_src=""
+  local candidate dst_dir dst_file scope_args body
+  for candidate in \
+    "$script_dir/codex-skills/disable-routing/SKILL.md" \
+    "$script_dir/../codex-skills/disable-routing/SKILL.md"
+  do
+    if [ -f "$candidate" ]; then
+      skill_src="$candidate"
+      break
+    fi
+  done
+  [ -n "$skill_src" ] || { warn "Codex disable-routing skill template is missing; router configuration is still installed."; return 0; }
+  grep -Fq "$WEAVE_CODEX_SKILL_MARKER" "$skill_src" \
+    || { warn "Codex disable-routing skill template has no ownership marker; leaving skills unchanged."; return 0; }
+
+  dst_dir="$codex_dir/skills/disable-routing"
+  dst_file="$dst_dir/SKILL.md"
+  if [ "$scope" = "project" ] || [ -n "$install_dir" ]; then
+    refuse_if_symlink "$codex_dir/skills"
+    refuse_if_symlink "$dst_dir"
+    refuse_if_symlink "$dst_file"
+  elif [ -L "$codex_dir/skills" ] || [ -L "$dst_dir" ] || [ -L "$dst_file" ]; then
+    warn "Codex disable-routing skill path contains a symlink; leaving it untouched."
+    return 0
+  fi
+  if [ -e "$dst_file" ] && { [ ! -f "$dst_file" ] || ! grep -Fq "$WEAVE_CODEX_SKILL_MARKER" "$dst_file"; }; then
+    warn "A user-owned Codex disable-routing skill already exists at $dst_file; leaving it untouched."
+    return 0
+  fi
+  mkdir -p "$dst_dir"
+
+  scope_args=""
+  if [ -n "$install_dir" ]; then
+    scope_args=" --dir $(printf '%q' "$install_dir")"
+  elif [ "$scope" = "project" ]; then
+    scope_args=" --scope project"
+  fi
+  body="$(<"$skill_src")"
+  body="${body//\{\{SCOPE\}\}/$scope_args}"
+  printf '%s\n' "$body" >"$dst_file"
+  ok "Codex skill installed: \$disable-routing"
+}
+
 # ---------- codex install path (dispatch + exit before the Claude-only writes) ----------
 
 if [ "$target" = "codex" ]; then
   write_codex_config "$codex_config_file" "$base_url" "$api_key" "$user_email" "$user_name"
   ok "Codex config written to $codex_config_file"
-  install_slash_commands "$codex_dir/prompts"
+  remove_obsolete_codex_prompt_wrappers "$codex_dir/prompts"
+  install_codex_disable_routing_skill
+  info "Codex router directives: begin the message with one space, e.g. ' /force-model gpt-5.6-terra'."
 
   # Project scope: ensure the per-teammate config (which holds the router key)
   # is gitignored. The base URL is the same for every teammate, so a
@@ -1999,10 +2125,21 @@ set -euo pipefail
 # lives in their per-user cache dir, and on no-content-change days we skip
 # the mv entirely so the repo working tree stays clean. When upstream does
 # change, the first teammate's commit propagates the new version to the rest.
+# A pricing-table miss also triggers a refresh off-schedule — see
+# weave_refresh_on_price_miss below.
 #
 # Opt out entirely with `export WEAVE_STATUSLINE_UPDATE=0`. Override the
 # source with `WEAVE_STATUSLINE_URL=...`, e.g. for self-hosters who fork.
+#
+# $1 is an optional stamp suffix, so a caller refreshing for a reason other than
+# "the interval elapsed" rate-limits on its own clock rather than sharing the
+# periodic check's budget. $2=1 drops the stamp again if the download fails, for
+# callers recovering from a known-bad state where waiting out the full interval
+# on a transient network error is worse than retrying next turn.
 weave_self_refresh() {
+  local stamp_suffix="${1:-}"
+  local retry_on_fail="${2:-0}"
+
   [ "${WEAVE_STATUSLINE_UPDATE:-1}" = "0" ] && return 0
   command -v curl >/dev/null 2>&1 || return 0
 
@@ -2019,7 +2156,7 @@ weave_self_refresh() {
   mkdir -p "$cache_dir" 2>/dev/null || return 0
   local script_slug
   script_slug="$(printf '%s' "$self" | tr -c 'A-Za-z0-9._-' '_')"
-  local stamp="$cache_dir/checked-at${script_slug}"
+  local stamp="$cache_dir/checked-at${script_slug}${stamp_suffix}"
 
   local now stamp_mtime
   now="$(date +%s 2>/dev/null)" || return 0
@@ -2043,7 +2180,13 @@ weave_self_refresh() {
   : > "$stamp" 2>/dev/null || return 0
 
   local url="${WEAVE_STATUSLINE_URL:-https://raw.githubusercontent.com/workweave/router/main/install/cc-statusline.sh}"
-  local tmp="${self}.tmp.$$"
+  # $$ alone is not unique: two calls can run in one invocation (the periodic
+  # check and a pricing-miss retry both fire on a cold cache) and would then
+  # curl -o into the same path and mv over each other, installing a truncated
+  # script that can never self-heal. The stamp suffix is the right key — it is
+  # what makes two callers mutually exclusive in the first place, so callers
+  # that could overlap necessarily have different suffixes.
+  local tmp="${self}.tmp.$$${stamp_suffix}"
   (
     # Detach stdin (CC pipes JSON to us) so curl can't accidentally consume
     # it, and silence all output so nothing leaks into the statusline.
@@ -2062,6 +2205,11 @@ weave_self_refresh() {
       fi
     else
       rm -f "$tmp"
+      # A download that never landed shouldn't spend the caller's whole
+      # interval; drop the stamp so the next turn can try again.
+      if [ "$retry_on_fail" = "1" ]; then
+        rm -f "$stamp"
+      fi
     fi
   ) >/dev/null 2>&1 &
   disown 2>/dev/null || true
@@ -2108,7 +2256,7 @@ prices='{
     "claude-sonnet-4-6":                0.003,
     "claude-sonnet-5":                  0.003,
     "deepseek/deepseek-v4-flash":       0.0001134,
-    "deepseek/deepseek-v4-pro":         0.001318,
+    "deepseek/deepseek-v4-pro":         0.00174,
     "gemini-2.0-flash":                 0.0001,
     "gemini-2.0-flash-lite":            0.000075,
     "gemini-2.5-flash":                 0.0003,
@@ -2143,6 +2291,7 @@ prices='{
     "gpt-5.6-sol":                      0.005,
     "gpt-5.6-terra":                    0.0025,
     "grok-4.5":                         0.002,
+    "grok-4.6":                         0.002,
     "minimax/minimax-m2.7":             0.0003,
     "minimax/minimax-m3":               0.0003,
     "mistralai/mistral-small-2603":     0.0002,
@@ -2156,8 +2305,9 @@ 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":             0.000172,
+    "qwen/qwen3.6-35b-a3b":             0.00015,
     "qwen/qwen3.7-plus":                0.0004,
+    "qwen/qwen3.8-max":                 0.002,
     "xiaomi/mimo-v2.5-pro":             0.001,
     "z-ai/glm-5":                       0.001,
     "z-ai/glm-5.1":                     0.0014,
@@ -2177,7 +2327,7 @@ prices='{
     "claude-sonnet-4-6":                0.015,
     "claude-sonnet-5":                  0.015,
     "deepseek/deepseek-v4-flash":       0.0002791,
-    "deepseek/deepseek-v4-pro":         0.0026361,
+    "deepseek/deepseek-v4-pro":         0.00348,
     "gemini-2.0-flash":                 0.0004,
     "gemini-2.0-flash-lite":            0.0003,
     "gemini-2.5-flash":                 0.0012,
@@ -2212,6 +2362,7 @@ prices='{
     "gpt-5.6-sol":                      0.03,
     "gpt-5.6-terra":                    0.015,
     "grok-4.5":                         0.006,
+    "grok-4.6":                         0.006,
     "minimax/minimax-m2.7":             0.0012,
     "minimax/minimax-m3":               0.0012,
     "mistralai/mistral-small-2603":     0.0006,
@@ -2225,8 +2376,9 @@ 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":             0.0012002,
+    "qwen/qwen3.6-35b-a3b":             0.001,
     "qwen/qwen3.7-plus":                0.0016,
+    "qwen/qwen3.8-max":                 0.006,
     "xiaomi/mimo-v2.5-pro":             0.003,
     "z-ai/glm-5":                       0.0032,
     "z-ai/glm-5.1":                     0.0044,
@@ -2365,6 +2517,42 @@ if [[ -n "$transcript_path" && -f "$transcript_path" ]]; then
   ) || true
 fi
 
+# ---------- refresh when a model isn't in the pricing table ----------
+#
+# A model that shipped after this copy of the script has no price entry, so the
+# jq guard above zeroes savings on every turn and the line reads "saved $0.00"
+# — indistinguishable from "the router ran and didn't beat your selection". The
+# periodic check heals that eventually, but a week late, and a model launch is
+# exactly when the number gets looked at. Refresh off-schedule instead.
+#
+# Keyed on its own stamp per unpriced id: a model we never price (self-hosted,
+# unrecognized) then costs one download per interval, not one per turn, and
+# can't starve the periodic check.
+weave_refresh_on_price_miss() {
+  local candidates="" m
+  for m in "$@"; do
+    case "$m" in
+      "" | "?" | failure | weave-router | "<synthetic>") continue ;;
+    esac
+    candidates="${candidates}${m}"$'\n'
+  done
+  [ -n "$candidates" ] || return 0
+
+  # A malformed $prices emits nothing here, which fails closed (no refresh)
+  # rather than re-downloading every turn.
+  local missing
+  missing="$(printf '%s' "$candidates" \
+    | jq -rR --argjson p "$prices" \
+        'select($p.input[.] == null or $p.output[.] == null)' 2>/dev/null \
+    | head -n 1)" || return 0
+  [ -n "$missing" ] || return 0
+
+  local model_slug
+  model_slug="$(printf '%s' "$missing" | tr -c 'A-Za-z0-9._-' '_')"
+  weave_self_refresh ".miss.${model_slug}" 1
+}
+weave_refresh_on_price_miss "$requested_norm" "$routed" 2>/dev/null || true
+
 # Brand color (#FF6C47) on terminals that grok 24-bit truecolor — that's
 # every modern one (iTerm2, Apple Terminal, vscode, ghostty, alacritty,
 # wezterm, kitty). Falls back gracefully on any escape-stripping terminal.
@@ -2480,12 +2668,20 @@ custom_headers="$custom_headers"$'\n'"X-App: claude-code"
 if [ "$scope" = "project" ] && [ -z "$install_dir" ]; then
   jq -n --arg url "$base_url" --arg sl "$statusline_path_for_settings" '{
     env: { ANTHROPIC_BASE_URL: $url, ENABLE_TOOL_SEARCH: "auto" },
-    statusLine: { type: "command", command: $sl }
+    statusLine: { type: "command", command: $sl },
+    attribution: {
+      commit: "Co-Authored-By: Weave Router <noreply@workweave.ai>",
+      pr: "🤖 Generated with [Weave Router](https://router.workweave.ai)"
+    }
   }' >"$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, ENABLE_TOOL_SEARCH: "auto" },
-    statusLine: { type: "command", command: $sl }
+    statusLine: { type: "command", command: $sl },
+    attribution: {
+      commit: "Co-Authored-By: Weave Router <noreply@workweave.ai>",
+      pr: "🤖 Generated with [Weave Router](https://router.workweave.ai)"
+    }
   }' >"$tmp_patch"
 fi
 
@@ -2501,6 +2697,7 @@ if [ -f "$settings_file" ]; then
     | (if (.env | length) == 0 then del(.env) else . end)
     | del(.apiKeyHelper)
     | (if $b.statusLine then .statusLine = $b.statusLine else . end)
+    | (if $b.attribution then .attribution = $b.attribution else . end)
   ' "$settings_file" "$tmp_patch")"
   printf '%s\n' "$merged" >"$settings_file"
 else

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workweave/router",
-  "version": "0.2.6",
+  "version": "0.2.8",
   "description": "One-command installer that points Claude Code, Codex, opencode, or pi at the Weave Router. For pi it also ships the routing extension, loaded via pi.extensions.",
   "bin": {
     "weave-router": "bin.js",
@@ -17,6 +17,7 @@
     "uninstall.sh",
     "cc-statusline.sh",
     "commands/",
+    "codex-skills/",
     "pi-router/",
     "opencode-weave/",
     "README.md",

package/pi-router/README.md

@@ -20,21 +20,37 @@ from npm on next start and loads this extension via its `pi.extensions` field.
 
 ## What it does
 
+- **Loom experience on stock pi.** Replaces pi's startup header through the
+  public extension API, adds Wooly's responsive orange terminal animation, and
+  keeps pi's own runtime/footer intact. Wooly is visual only: there is no
+  dialogue box, narration, coaching request, or separate Loom runtime.
 - **Automatic model selection.** All pi traffic flows through the router, which
   selects the model per request. You don't pick a model — the router does.
+- **Force-model commands.** `/fm <model>` and `/force-model <model>` pin the
+  current router session; `/ufm` and `/unforce-model` resume automatic routing.
+  The persistent status changes to `WEAVE ROUTER — <model> [forced]` after the
+  router validates and canonicalizes the requested model.
 - **Per-process routing bias.** Static `x-weave-routing-*` knob headers bias the
-  router: quality on the main loop, speed + cheap on subagents, cheapest on
-  compaction.
+  router: quality on the main loop and speed + cheap on subagents.
+- **Long tool-loop compaction.** Pi 0.74 can cross its context threshold inside
+  an uninterrupted tool loop before its normal post-run compaction check. The
+  extension preserves a usable output budget for the real continuation and
+  compacts once the loop settles, while leaving ordinary threshold compaction
+  to Pi.
 - **Sticky sessions.** `metadata.user_id = "pi:<sessionId>"` pins the main loop
   to one model for the session; subagents get their own pins.
 - **`dispatch` tool — parallel, context-isolated subagents.** pi has none
   natively. `dispatch` spawns child `pi` processes (read-only by default), runs
   them concurrently, and returns only each subagent's final answer — intermediate
   tool output stays in the child, so the main context stays small.
-- **Routed-model display.** Shows the model the router actually chose
-  (`x-router-model`) in the status bar, and opts the request out of the router's
-  in-band routing badge (`X-Weave-Routing-Marker: off`) — pi can't render that
-  separate marker text block inline, and the status bar already conveys the model.
+- **Persistent route + savings display.** Shows
+  `WEAVE ROUTER — <routed> ← <selected> · saved $X.XX` below pi's native footer
+  data. Savings compare the selected and routed catalog prices against the same
+  input/output/cache usage, accumulate across the reachable session branch,
+  and survive resume. Unknown catalog prices are labeled `unpriced` instead of
+  silently contributing zero; costlier routing is labeled `extra`, not savings.
+- **No duplicate in-band badge.** Sets `X-Weave-Routing-Marker: off` because the
+  persistent status already conveys the actual model.
 - **Safety backstop.** Blocks a few catastrophic shell commands (`rm -rf /`,
   `mkfs`, `dd of=/dev/…`, fork bombs, force-push to main). Disable with
   `WEAVE_NO_SAFETY=1`.
@@ -53,7 +69,7 @@ from npm on next start and loads this extension via its `pi.extensions` field.
 | `WEAVE_PI_ALLOW_SUBAGENT_TOOLS` | unset | `1` lets `dispatch` grant subagents write/exec tools (bash, write, edit); default strips them |
 | `WEAVE_ROUTING_ALPHA` / `…_SPEED_WEIGHT` / `…_OUTPUT_COST_RATIO` / `…_EXPECTED_OUTPUT_TOKENS` | role preset | Override individual routing knobs (main process only — children always use their role preset) |
 | `WEAVE_NO_SAFETY` | unset | `1` disables the catastrophic-bash gate |
-| `WEAVE_CHEAP_COMPACTION` | unset | `1` enables the (experimental) cheap-compaction path |
+| `WEAVE_PI_AUTO_COMPACTION` | unset | `0` disables the routed tool-loop compaction safeguard |
 
 Internal: `WEAVE_PI_SUBAGENT=1` and `WEAVE_PI_SUBAGENT_ID` are set by `dispatch`
 on child processes; don't set them yourself.
@@ -64,7 +80,12 @@ Routing through the router switches pi from Claude **subscription OAuth** to
 **per-token** billing on the router deployment's key (or your BYOK key). BYOK
 skips cross-provider failover; deployment-key billing is the default.
 
+The displayed savings are a client-side estimate from the router's generated
+model-price catalog. Cache writes use 1.25× input price and cache reads use 0.1×
+input price, matching the Claude Code statusline. The ledger stores its catalog
+version with each response so resumed totals remain auditable.
+
 ## Notes
 
-- Cheap compaction is currently a reserved flag — the handler defers to pi's
-  built-in compaction until the router-routed path is validated end-to-end.
+- Actual SDK/router probes keep their small output budget. Only a probe-sized
+  request carrying a real tool-result continuation is repaired.