@workweave/router 0.2.7 → 0.2.8

package/README.md

@@ -32,6 +32,9 @@ 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
+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.
 
@@ -62,10 +65,19 @@ Four install targets:
   api.anthropic.com.
 - **Codex** (`--codex`) — patches `~/.codex/config.toml` (or
   `<repo>/.codex/config.toml`) with a managed `[model_providers.weave]`
-  block plus `model_provider = "weave"`. OpenAI plan credentials flow
-  through to api.openai.com. 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.
+  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
+  `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.
 - **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

package/cc-statusline.sh

@@ -164,7 +164,7 @@ 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,
@@ -199,6 +199,7 @@ 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,
@@ -212,8 +213,9 @@ 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,
@@ -233,7 +235,7 @@ 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,
@@ -268,6 +270,7 @@ 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,
@@ -281,8 +284,9 @@ 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,

package/codex-skills/disable-routing/SKILL.md

@@ -0,0 +1,22 @@
+---
+name: disable-routing
+description: "Switch Codex back to its normal provider for future sessions."
+---
+
+<!-- weave-router managed disable-routing skill -->
+
+# Disable Weave routing
+
+When the user invokes `$disable-routing`, switch the current Codex installation
+off the Weave Router without logging out or deleting its router configuration.
+
+1. Explain that the change takes effect on the next `codex` launch and that it
+   can later be reversed with `npx --package @workweave/router -y -- weave-router on --codex{{SCOPE}}`.
+2. Run exactly:
+
+   ```bash
+   npx --package @workweave/router -y -- weave-router off --codex{{SCOPE}}
+   ```
+
+3. Report the command result. Do not uninstall the router or alter any other
+   Codex settings.

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
@@ -2130,7 +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,
@@ -2165,6 +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,
@@ -2178,8 +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,
@@ -2199,7 +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,
@@ -2234,6 +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,
@@ -2247,8 +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,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workweave/router",
-  "version": "0.2.7",
+  "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

@@ -31,8 +31,12 @@ from npm on next start and loads this extension via its `pi.extensions` field.
   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
@@ -65,7 +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.
@@ -83,5 +87,5 @@ 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.

package/pi-router/src/compaction.ts

@@ -1,26 +1,150 @@
 /**
- * EXPERIMENTAL, off by default (enable with WEAVE_CHEAP_COMPACTION=1).
+ * Compaction safeguards for Pi's routed provider.
  *
- * Intent: run context compaction through the cheapest routing knobs. The
- * catch is that compaction bypasses provider hooks and reuses the session's
- * static provider headers (the main loop's quality knobs), and
- * `session_before_compact` is the only per-turn lever. Producing a fully
- * router-routed cheap CompactionResult here is not yet validated, so this
- * handler currently defers to pi's built-in compaction (returns no override)
- * and only reserves the flag + surfaces that it's active. Promote to a real
- * cheap path once validated end-to-end.
+ * Pi 0.74 checks automatic compaction only after an entire agent loop. A long
+ * tool loop can therefore cross the virtual model's context window between
+ * turns. Pi then clamps the next request to max_tokens=1, which looks exactly
+ * like an SDK quota probe to the router and produces a one-token response.
+ *
+ * Preserve real probes, but restore a usable budget for a clamped request that
+ * contains an actual tool result. Once the agent finishes, compact using the
+ * highest context reading seen during the run; routed providers tokenize and
+ * report caches differently, so the final response alone is not authoritative.
+ */
+
+import type {
+	AgentEndEvent,
+	BeforeProviderRequestEvent,
+	ExtensionAPI,
+	ExtensionContext,
+	SessionCompactEvent,
+	TurnEndEvent,
+} from "@mariozechner/pi-coding-agent";
+import type { AssistantMessage } from "@mariozechner/pi-ai";
+
+const PROBE_MAX_TOKENS = 4;
+const CONTINUATION_MAX_TOKENS = 16_384;
+const COMPACTION_RESERVE_TOKENS = 16_384;
+const STATUS_KEY = "weave-compaction";
+
+interface ProviderPayload {
+	max_tokens?: unknown;
+	messages?: unknown;
+	tools?: unknown;
+}
+
+interface ProviderMessage {
+	content?: unknown;
+}
+
+type Schedule = (callback: () => void) => unknown;
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+	return typeof value === "object" && value !== null;
+}
+
+function containsToolResult(messages: unknown): boolean {
+	if (!Array.isArray(messages)) return false;
+	return messages.some((message: unknown) => {
+		if (!isRecord(message)) return false;
+		const content = (message as ProviderMessage).content;
+		return Array.isArray(content) && content.some((block: unknown) => isRecord(block) && block.type === "tool_result");
+	});
+}
+
+/**
+ * Repair only the Pi context-exhaustion shape. Genuine SDK probes have no tool
+ * result transcript and remain at max_tokens=1..4 for the router to hard-pin.
  */
+export function repairClampedToolContinuation(payload: unknown): boolean {
+	if (!isRecord(payload)) return false;
+	const body = payload as ProviderPayload;
+	if (typeof body.max_tokens !== "number" || body.max_tokens < 1 || body.max_tokens > PROBE_MAX_TOKENS) return false;
+	if (!Array.isArray(body.tools) || body.tools.length === 0) return false;
+	if (!containsToolResult(body.messages)) return false;
+	body.max_tokens = CONTINUATION_MAX_TOKENS;
+	return true;
+}
+
+function contextTokens(message: AssistantMessage): number {
+	const usage = message.usage;
+	if (usage.totalTokens > 0) return usage.totalTokens;
+	return usage.input + usage.output + usage.cacheRead + usage.cacheWrite;
+}
+
+function latestCompactionId(ctx: ExtensionContext): string | undefined {
+	const branch = ctx.sessionManager.getBranch();
+	for (let i = branch.length - 1; i >= 0; i--) {
+		const entry = branch[i];
+		if (entry?.type === "compaction") return entry.id;
+	}
+	return undefined;
+}
+
+export function registerCompaction(pi: ExtensionAPI, schedule: Schedule = (callback) => setTimeout(callback, 0)): void {
+	let highWaterTokens = 0;
+	let lastTurnTokens = 0;
+	let repairedContinuation = false;
+	let compactionScheduled = false;
+
+	const resetRun = () => {
+		highWaterTokens = 0;
+		lastTurnTokens = 0;
+		repairedContinuation = false;
+	};
+
+	const finishCompaction = (ctx: ExtensionContext) => {
+		compactionScheduled = false;
+		resetRun();
+		if (ctx.hasUI) ctx.ui.setStatus(STATUS_KEY, undefined);
+	};
+
+	pi.on("session_start", resetRun);
+	pi.on("agent_start", resetRun);
+	pi.on("session_compact", (_event: SessionCompactEvent, ctx: ExtensionContext) => finishCompaction(ctx));
+
+	pi.on("before_provider_request", (event: BeforeProviderRequestEvent) => {
+		if (repairClampedToolContinuation(event.payload)) repairedContinuation = true;
+	});
+
+	pi.on("turn_end", (event: TurnEndEvent) => {
+		if (event.message.role !== "assistant") return;
+		lastTurnTokens = contextTokens(event.message as AssistantMessage);
+		highWaterTokens = Math.max(highWaterTokens, lastTurnTokens);
+	});
+
+	pi.on("agent_end", (_event: AgentEndEvent, ctx: ExtensionContext) => {
+		if (process.env.WEAVE_PI_AUTO_COMPACTION === "0" || compactionScheduled) return;
+		const contextWindow = ctx.model?.contextWindow ?? ctx.getContextUsage()?.contextWindow ?? 0;
+		if (contextWindow <= COMPACTION_RESERVE_TOKENS) return;
+		const threshold = contextWindow - COMPACTION_RESERVE_TOKENS;
+		// Pi's built-in check runs immediately after this event and owns the
+		// ordinary final-turn threshold case. Starting another compaction while
+		// that async summary is in flight would race it.
+		if (lastTurnTokens > threshold) return;
+		if (!repairedContinuation && highWaterTokens <= threshold) return;
+
+		compactionScheduled = true;
+		const previousCompactionId = latestCompactionId(ctx);
 
-import type { ExtensionAPI, ExtensionContext, SessionBeforeCompactEvent } from "@mariozechner/pi-coding-agent";
-
-export function registerCheapCompaction(pi: ExtensionAPI): void {
-	let announced = false;
-	pi.on("session_before_compact", (_event: SessionBeforeCompactEvent, ctx: ExtensionContext) => {
-		if (ctx.hasUI && !announced) {
-			announced = true;
-			ctx.ui.setStatus("weave-compaction", "cheap compaction: experimental (using built-in)");
-		}
-		// Defer to built-in compaction until the router-routed cheap path is validated.
-		return undefined;
+		// Pi is still settling agent_end while extension handlers run. Schedule
+		// manual compaction for the next event-loop turn to avoid abort/wait
+		// recursion, and let Pi's own auto-compaction win if it already ran.
+		schedule(() => {
+			if (!compactionScheduled) return;
+			const currentCompactionId = latestCompactionId(ctx);
+			if (currentCompactionId !== previousCompactionId) {
+				finishCompaction(ctx);
+				return;
+			}
+			if (ctx.hasUI) ctx.ui.setStatus(STATUS_KEY, "compacting routed context...");
+			ctx.compact({
+				onComplete: () => finishCompaction(ctx),
+				onError: (error) => {
+					compactionScheduled = false;
+					if (ctx.hasUI) ctx.ui.setStatus(STATUS_KEY, `compaction failed: ${error.message}`);
+				},
+			});
+		});
 	});
 }

package/pi-router/src/config.ts

@@ -225,9 +225,10 @@ export const WEAVE_MODELS: ProviderModelConfig[] = [
 	model("claude-opus-4-7", "Claude Opus 4.7 (via Weave Router)", 64000),
 	model("claude-sonnet-4-6", "Claude Sonnet 4.6 (via Weave Router)", 64000),
 	model("claude-haiku-4-5", "Claude Haiku 4.5 (via Weave Router)", 32000),
+	model("grok-4.6", "Grok 4.6 (via Weave Router)", 131072, 500000),
 ];
 
-function model(id: string, name: string, maxTokens: number): ProviderModelConfig {
+function model(id: string, name: string, maxTokens: number, contextWindow: number = 200000): ProviderModelConfig {
 	return {
 		id,
 		name,
@@ -235,7 +236,7 @@ function model(id: string, name: string, maxTokens: number): ProviderModelConfig
 		input: ["text", "image"],
 		// Real cost is decided by the router per request and is unknown client-side.
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-		contextWindow: 200000,
+		contextWindow,
 		maxTokens,
 	};
 }

package/pi-router/src/index.ts

@@ -9,7 +9,7 @@
  *                   detection.
  *   - Loom UI:      branded header, Wooly animation, actual route, and saved $.
  *   - safety:       block catastrophic bash (unless WEAVE_NO_SAFETY=1).
- *   - compaction:   experimental cheap path (only when WEAVE_CHEAP_COMPACTION=1).
+ *   - compaction:   protect long tool loops, then compact routed context.
  *   - dispatch:     parallel, context-isolated subagents — top-level process
  *                   only (no grandchildren).
  *
@@ -20,7 +20,7 @@
 import { fileURLToPath } from "node:url";
 import type { ExtensionAPI } from "@mariozechner/pi-coding-agent";
 import { isSubagent } from "./config.js";
-import { registerCheapCompaction } from "./compaction.js";
+import { registerCompaction } from "./compaction.js";
 import { registerDispatch } from "./dispatch.js";
 import { registerForceModelCommands } from "./force-model.js";
 import { registerMetadata } from "./metadata.js";
@@ -40,9 +40,9 @@ export default function (pi: ExtensionAPI): void {
 	registerMetadata(pi);
 	registerForceModelCommands(pi);
 	registerRoutedModel(pi);
+	registerCompaction(pi);
 
 	if (process.env.WEAVE_NO_SAFETY !== "1") registerSafety(pi);
-	if (process.env.WEAVE_CHEAP_COMPACTION === "1") registerCheapCompaction(pi);
 
 	// Only the top-level process fans out. Children (WEAVE_PI_SUBAGENT=1) load
 	// this same extension but get no dispatch tool, so subagents can't spawn

package/pi-router/src/pricing.generated.ts

@@ -6,7 +6,7 @@ export interface ModelPricing {
 	outputUsdPerMillion: number;
 }
 
-export const PRICING_VERSION = "catalog-sha256:a13e743aa9694647";
+export const PRICING_VERSION = "catalog-sha256:f2841a252235d113";
 
 export const MODEL_PRICING: Readonly<Record<string, ModelPricing>> = Object.freeze({
 	"claude-fable-5": { inputUsdPerMillion: 10, outputUsdPerMillion: 50 },
@@ -22,7 +22,7 @@ export const MODEL_PRICING: Readonly<Record<string, ModelPricing>> = Object.free
 	"claude-sonnet-4-6": { inputUsdPerMillion: 3, outputUsdPerMillion: 15 },
 	"claude-sonnet-5": { inputUsdPerMillion: 3, outputUsdPerMillion: 15 },
 	"deepseek/deepseek-v4-flash": { inputUsdPerMillion: 0.1134, outputUsdPerMillion: 0.2791 },
-	"deepseek/deepseek-v4-pro": { inputUsdPerMillion: 1.318, outputUsdPerMillion: 2.6361 },
+	"deepseek/deepseek-v4-pro": { inputUsdPerMillion: 1.74, outputUsdPerMillion: 3.48 },
 	"gemini-2.0-flash": { inputUsdPerMillion: 0.1, outputUsdPerMillion: 0.4 },
 	"gemini-2.0-flash-lite": { inputUsdPerMillion: 0.075, outputUsdPerMillion: 0.3 },
 	"gemini-2.5-flash": { inputUsdPerMillion: 0.3, outputUsdPerMillion: 1.2 },
@@ -57,6 +57,7 @@ export const MODEL_PRICING: Readonly<Record<string, ModelPricing>> = Object.free
 	"gpt-5.6-sol": { inputUsdPerMillion: 5, outputUsdPerMillion: 30 },
 	"gpt-5.6-terra": { inputUsdPerMillion: 2.5, outputUsdPerMillion: 15 },
 	"grok-4.5": { inputUsdPerMillion: 2, outputUsdPerMillion: 6 },
+	"grok-4.6": { inputUsdPerMillion: 2, outputUsdPerMillion: 6 },
 	"minimax/minimax-m2.7": { inputUsdPerMillion: 0.3, outputUsdPerMillion: 1.2 },
 	"minimax/minimax-m3": { inputUsdPerMillion: 0.3, outputUsdPerMillion: 1.2 },
 	"mistralai/mistral-small-2603": { inputUsdPerMillion: 0.2, outputUsdPerMillion: 0.6 },
@@ -70,8 +71,9 @@ export const MODEL_PRICING: Readonly<Record<string, ModelPricing>> = Object.free
 	"qwen/qwen3-coder-next": { inputUsdPerMillion: 0.5, outputUsdPerMillion: 1.2 },
 	"qwen/qwen3-next-80b-a3b-instruct": { inputUsdPerMillion: 0.15, outputUsdPerMillion: 1.2 },
 	"qwen/qwen3.5-flash-02-23": { inputUsdPerMillion: 0.05, outputUsdPerMillion: 0.15 },
-	"qwen/qwen3.6-35b-a3b": { inputUsdPerMillion: 0.172, outputUsdPerMillion: 1.2002 },
+	"qwen/qwen3.6-35b-a3b": { inputUsdPerMillion: 0.15, outputUsdPerMillion: 1 },
 	"qwen/qwen3.7-plus": { inputUsdPerMillion: 0.4, outputUsdPerMillion: 1.6 },
+	"qwen/qwen3.8-max": { inputUsdPerMillion: 2, outputUsdPerMillion: 6 },
 	"xiaomi/mimo-v2.5-pro": { inputUsdPerMillion: 1, outputUsdPerMillion: 3 },
 	"z-ai/glm-5": { inputUsdPerMillion: 1, outputUsdPerMillion: 3.2 },
 	"z-ai/glm-5.1": { inputUsdPerMillion: 1.4, outputUsdPerMillion: 4.4 },

package/pi-router/src/ui.ts

@@ -26,6 +26,16 @@ export interface RouterUiSnapshot {
 	savings: SavingsAggregate;
 }
 
+type LegacyCompatibleExtensionContext = ExtensionContext & { mode?: string };
+
+function isTuiContext(ctx: ExtensionContext): boolean {
+	const mode = (ctx as LegacyCompatibleExtensionContext).mode;
+	// Pi 0.74.2 supports headers and widgets but predates ctx.mode. In that
+	// release hasUI is true only for the interactive TUI, so it is the safe
+	// compatibility signal. Newer Pi releases expose mode and take precedence.
+	return mode === "tui" || (mode === undefined && ctx.hasUI);
+}
+
 function brand(text: string): string {
 	return `${BRAND_OPEN}${text}${BRAND_CLOSE}`;
 }
@@ -45,7 +55,7 @@ function headerLines(theme: Theme, width: number): string[] {
 }
 
 export function installLoomUi(ctx: ExtensionContext): void {
-	if (ctx.mode !== "tui") return;
+	if (!isTuiContext(ctx)) return;
 	ctx.ui.setTitle("Loom · Weave Router");
 	ctx.ui.setHeader((_tui: TUI, theme: Theme) => ({
 		invalidate() {},
@@ -59,7 +69,7 @@ export function installLoomUi(ctx: ExtensionContext): void {
 export function clearLoomUi(ctx: ExtensionContext): void {
 	if (!ctx.hasUI) return;
 	ctx.ui.setStatus(STATUS_KEY, undefined);
-	if (ctx.mode === "tui") {
+	if (isTuiContext(ctx)) {
 		ctx.ui.setWidget(WOOLY_WIDGET_KEY, undefined);
 		ctx.ui.setHeader(undefined);
 	}
@@ -74,7 +84,8 @@ export function updateRouterStatus(ctx: ExtensionContext, snapshot: RouterUiSnap
 	else if (requestedModel) route = `${requestedModel} · awaiting route`;
 	else route = "automatic routing";
 
-	const label = ctx.mode === "tui" ? ctx.ui.theme.bold(brand("WEAVE ROUTER")) : "WEAVE ROUTER";
+	const tui = isTuiContext(ctx);
+	const label = tui ? ctx.ui.theme.bold(brand("WEAVE ROUTER")) : "WEAVE ROUTER";
 	const detail = forcedModel ? `${forcedModel} [forced]` : `${route} · ${formatSavings(savings)}`;
-	ctx.ui.setStatus(STATUS_KEY, `${label} — ${ctx.mode === "tui" ? ctx.ui.theme.fg("dim", detail) : detail}`);
+	ctx.ui.setStatus(STATUS_KEY, `${label} — ${tui ? ctx.ui.theme.fg("dim", detail) : detail}`);
 }

package/uninstall.sh

@@ -30,6 +30,7 @@ install_dir=""
 target="claude"
 
 err()  { printf "\033[31merror:\033[0m %s\n" "$*" >&2; }
+warn() { printf "\033[33mwarning:\033[0m %s\n" "$*" >&2; }
 info() { printf "\033[36m==>\033[0m %s\n" "$*"; }
 ok()   { printf "\033[32m✓\033[0m %s\n" "$*"; }
 
@@ -91,6 +92,7 @@ fi
 # Markers must stay in sync with install.sh. Keep verbatim.
 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 -->"
 
 # strip_codex_block rewrites config.toml without the managed block and any
 # top-level `model_provider = "weave"` that lived outside the markers (which
@@ -436,6 +438,26 @@ if [ "$target" = "codex" ]; then
     rmdir "$codex_prompts_dir" 2>/dev/null || true
   fi
 
+  # The installer-owned skill switches Codex back to its normal provider. It
+  # is separate from config.toml, so remove it when removing the router too.
+  # A same-named skill without our marker belongs to the user and is preserved.
+  codex_skill_dir="$codex_dir/skills/disable-routing"
+  codex_skill_file="$codex_skill_dir/SKILL.md"
+  if [ -d "$codex_skill_dir" ]; then
+    refuse_if_symlink "$codex_skill_dir"
+    if [ -f "$codex_skill_file" ]; then
+      refuse_if_symlink "$codex_skill_file"
+      if grep -Fq "$WEAVE_CODEX_SKILL_MARKER" "$codex_skill_file"; then
+        rm -f "$codex_skill_file"
+        ok "Removed $codex_skill_file"
+      else
+        warn "Leaving user-owned Codex skill at $codex_skill_file untouched."
+      fi
+    fi
+    rmdir "$codex_skill_dir" 2>/dev/null || true
+    rmdir "$codex_dir/skills" 2>/dev/null || true
+  fi
+
   if [ -n "$install_dir" ]; then
     ok "Weave Router uninstalled from $install_dir (Codex)."
   else