@weave-os/router 0.2.14

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (69) hide show
  1. package/LICENSE +90 -0
  2. package/README.md +172 -0
  3. package/bin.js +68 -0
  4. package/cc-statusline.sh +951 -0
  5. package/codex-skills/disable-routing/SKILL.md +22 -0
  6. package/codex-skills/fm/SKILL.md +15 -0
  7. package/codex-skills/fm/scripts/emit.sh +8 -0
  8. package/codex-skills/force-model/SKILL.md +15 -0
  9. package/codex-skills/force-model/scripts/emit.sh +9 -0
  10. package/codex-skills/rf/SKILL.md +15 -0
  11. package/codex-skills/rf/scripts/emit.sh +7 -0
  12. package/codex-skills/router-feedback/SKILL.md +15 -0
  13. package/codex-skills/router-feedback/scripts/emit.sh +9 -0
  14. package/codex-skills/router-models/SKILL.md +51 -0
  15. package/codex-skills/router-off/SKILL.md +22 -0
  16. package/codex-skills/router-on/SKILL.md +22 -0
  17. package/codex-skills/router-status/SKILL.md +19 -0
  18. package/codex-skills/ufm/SKILL.md +14 -0
  19. package/codex-skills/ufm/scripts/emit.sh +3 -0
  20. package/codex-skills/unforce-model/SKILL.md +14 -0
  21. package/codex-skills/unforce-model/scripts/emit.sh +4 -0
  22. package/codex-status.sh +378 -0
  23. package/commands/beta.md +5 -0
  24. package/commands/fm.md +6 -0
  25. package/commands/force-model.md +6 -0
  26. package/commands/models.md +46 -0
  27. package/commands/rf.md +6 -0
  28. package/commands/router-feedback.md +6 -0
  29. package/commands/router-models.md +46 -0
  30. package/commands/router-off.md +10 -0
  31. package/commands/router-on.md +10 -0
  32. package/commands/router-session.md +15 -0
  33. package/commands/router-status.md +10 -0
  34. package/commands/ufm.md +5 -0
  35. package/commands/unforce-model.md +5 -0
  36. package/directives.tsv +13 -0
  37. package/install.sh +5092 -0
  38. package/opencode-weave/README.md +78 -0
  39. package/opencode-weave/package.json +18 -0
  40. package/opencode-weave/src/index.ts +655 -0
  41. package/package.json +79 -0
  42. package/pi-router/README.md +123 -0
  43. package/pi-router/package.json +25 -0
  44. package/pi-router/skills/install-lsps/SKILL.md +75 -0
  45. package/pi-router/skills/lsp-guide/SKILL.md +63 -0
  46. package/pi-router/src/beta.ts +21 -0
  47. package/pi-router/src/compaction.ts +188 -0
  48. package/pi-router/src/config.ts +293 -0
  49. package/pi-router/src/context-window.ts +12 -0
  50. package/pi-router/src/dispatch.ts +377 -0
  51. package/pi-router/src/force-model.ts +126 -0
  52. package/pi-router/src/index.ts +62 -0
  53. package/pi-router/src/lsp-broker.ts +255 -0
  54. package/pi-router/src/lsp-client.ts +435 -0
  55. package/pi-router/src/lsp-format.ts +230 -0
  56. package/pi-router/src/lsp-install.ts +215 -0
  57. package/pi-router/src/lsp-protocol.ts +128 -0
  58. package/pi-router/src/lsp-servers.ts +361 -0
  59. package/pi-router/src/lsp.ts +529 -0
  60. package/pi-router/src/metadata.ts +32 -0
  61. package/pi-router/src/pricing.generated.ts +91 -0
  62. package/pi-router/src/provider.ts +64 -0
  63. package/pi-router/src/routed-model.ts +182 -0
  64. package/pi-router/src/safety.ts +66 -0
  65. package/pi-router/src/savings.ts +191 -0
  66. package/pi-router/src/ui.ts +91 -0
  67. package/pi-router/src/wooly.ts +408 -0
  68. package/registry.sh +102 -0
  69. package/uninstall.sh +788 -0
@@ -0,0 +1,951 @@
1
+ #!/usr/bin/env bash
2
+ #
3
+ # Claude Code statusline for the Weave Router. CC pipes a JSON blob on stdin
4
+ # whose `transcript_path` points at the JSONL log of the current session and
5
+ # whose `model.display_name` is the user's CC-side model selection. The
6
+ # router rewrites each request's `model` field before forwarding, so
7
+ # Anthropic/OpenAI/Google return `message.model = <routed>` in the SSE
8
+ # stream and CC stores that in the transcript verbatim. Per-turn savings
9
+ # come from comparing each turn's routed cost against what the user's
10
+ # selection would have cost on the same tokens. Works identically for
11
+ # local docker and the managed cloud router — no sidecar, no DB, no auth.
12
+ #
13
+ # Wire up by adding to ~/.claude/settings.json:
14
+ # { "statusLine": { "type": "command", "command": "/abs/path/to/cc-statusline.sh" } }
15
+ #
16
+ # Renders:
17
+ # WEAVE ROUTER — claude-sonnet-4-5 ← claude-opus-4-7 · saved $1.23 · 12.4k in / 3.1k out / 45.2k cached
18
+ #
19
+ # Pricing source of truth: internal/router/catalog. Input/output prices and
20
+ # cache-read multipliers are generated by cmd/genprices. Cache creation remains
21
+ # at 1.25× input pending TTL-aware pricing.
22
+
23
+ set -euo pipefail
24
+
25
+ # ---------- background self-refresh ----------
26
+ #
27
+ # Once every WEAVE_STATUSLINE_UPDATE_INTERVAL_DAYS (default 7), check
28
+ # raw.githubusercontent.com for a newer copy of this script and swap it in
29
+ # atomically. Runs in a forked subshell so the current Claude turn never
30
+ # blocks; the next turn picks up the new version. Applies to both user-scope
31
+ # (~/.weave/cc-statusline.sh) and project-scope (<repo>/.claude/cc-statusline.sh)
32
+ # installs — project teammates rate-limit independently because the stamp
33
+ # lives in their per-user cache dir, and on no-content-change days we skip
34
+ # the mv entirely so the repo working tree stays clean. When upstream does
35
+ # change, the first teammate's commit propagates the new version to the rest.
36
+ # A pricing-table miss also triggers a refresh off-schedule — see
37
+ # weave_refresh_on_price_miss below.
38
+ #
39
+ # Opt out entirely with `export WEAVE_STATUSLINE_UPDATE=0`. Override the
40
+ # source with `WEAVE_STATUSLINE_URL=...`, e.g. for self-hosters who fork.
41
+ #
42
+ # $1 is an optional stamp suffix, so a caller refreshing for a reason other than
43
+ # "the interval elapsed" rate-limits on its own clock rather than sharing the
44
+ # periodic check's budget. $2=1 drops the stamp again if the download fails, for
45
+ # callers recovering from a known-bad state where waiting out the full interval
46
+ # on a transient network error is worse than retrying next turn.
47
+ weave_self_refresh() {
48
+ local stamp_suffix="${1:-}"
49
+ local retry_on_fail="${2:-0}"
50
+
51
+ [ "${WEAVE_STATUSLINE_UPDATE:-1}" = "0" ] && return 0
52
+ command -v curl >/dev/null 2>&1 || return 0
53
+
54
+ local self="${BASH_SOURCE[0]:-$0}"
55
+ [ -f "$self" ] && [ -w "$self" ] || return 0
56
+
57
+ local interval_days="${WEAVE_STATUSLINE_UPDATE_INTERVAL_DAYS:-7}"
58
+ local interval_seconds=$(( interval_days * 86400 ))
59
+
60
+ # Stamp lives in the per-user cache dir, keyed by absolute script path so
61
+ # multiple repos (and the user-scope copy) rate-limit independently and no
62
+ # stray file ever lands inside a repo working tree.
63
+ local cache_dir="${XDG_CACHE_HOME:-$HOME/.cache}/weave-router"
64
+ mkdir -p "$cache_dir" 2>/dev/null || return 0
65
+ local script_slug
66
+ script_slug="$(printf '%s' "$self" | tr -c 'A-Za-z0-9._-' '_')"
67
+ local stamp="$cache_dir/checked-at${script_slug}${stamp_suffix}"
68
+
69
+ local now stamp_mtime
70
+ now="$(date +%s 2>/dev/null)" || return 0
71
+ if [ -f "$stamp" ]; then
72
+ # Try GNU `stat -c %Y` first; on macOS (BSD stat) -c isn't recognized
73
+ # and exits non-zero, so we fall through to `stat -f %m`. The reverse
74
+ # order is broken: GNU `stat -f` is `--file-system`, which silently
75
+ # succeeds with multi-line filesystem info instead of failing, leaving
76
+ # $stamp_mtime as garbage and disabling the rate-limit check entirely.
77
+ stamp_mtime="$(stat -c %Y "$stamp" 2>/dev/null || stat -f %m "$stamp" 2>/dev/null)" || stamp_mtime=0
78
+ else
79
+ stamp_mtime=0
80
+ fi
81
+ if [ -n "${stamp_mtime:-}" ] && [ "$stamp_mtime" -gt 0 ] \
82
+ && [ $(( now - stamp_mtime )) -lt "$interval_seconds" ]; then
83
+ return 0
84
+ fi
85
+
86
+ # Touch the stamp BEFORE forking so concurrent statusline invocations
87
+ # (Claude calls us on every turn) don't all kick off downloads.
88
+ : > "$stamp" 2>/dev/null || return 0
89
+
90
+ local url="${WEAVE_STATUSLINE_URL:-https://raw.githubusercontent.com/weave-os/router/main/install/cc-statusline.sh}"
91
+ # $$ alone is not unique: two calls can run in one invocation (the periodic
92
+ # check and a pricing-miss retry both fire on a cold cache) and would then
93
+ # curl -o into the same path and mv over each other, installing a truncated
94
+ # script that can never self-heal. The stamp suffix is the right key — it is
95
+ # what makes two callers mutually exclusive in the first place, so callers
96
+ # that could overlap necessarily have different suffixes.
97
+ local tmp="${self}.tmp.$$${stamp_suffix}"
98
+ (
99
+ # Detach stdin (CC pipes JSON to us) so curl can't accidentally consume
100
+ # it, and silence all output so nothing leaks into the statusline.
101
+ exec </dev/null
102
+ if curl -fsSL --max-time 15 "$url" -o "$tmp" 2>/dev/null \
103
+ && [ -s "$tmp" ] \
104
+ && head -n 1 "$tmp" | grep -q '^#!.*bash' \
105
+ && [ "$(wc -c < "$tmp")" -ge 1024 ]; then
106
+ # No-op when the download matches what's already on disk — keeps git
107
+ # status clean for project-scope teammates during a routine refresh.
108
+ if cmp -s "$tmp" "$self"; then
109
+ rm -f "$tmp"
110
+ else
111
+ chmod +x "$tmp" 2>/dev/null || true
112
+ mv "$tmp" "$self" 2>/dev/null || rm -f "$tmp"
113
+ fi
114
+ else
115
+ rm -f "$tmp"
116
+ # A download that never landed shouldn't spend the caller's whole
117
+ # interval; drop the stamp so the next turn can try again.
118
+ if [ "$retry_on_fail" = "1" ]; then
119
+ rm -f "$stamp"
120
+ fi
121
+ fi
122
+ ) >/dev/null 2>&1 &
123
+ disown 2>/dev/null || true
124
+ return 0
125
+ }
126
+ weave_self_refresh 2>/dev/null || true
127
+
128
+ # ---------- background slash-command refresh ----------
129
+ #
130
+ # The /force-model, /router-* … wrappers under <install>/.claude/commands/ are
131
+ # written once at install time and then never touched again, so a user who
132
+ # doesn't re-run the installer keeps whatever set shipped that day. This script
133
+ # is the only thing Claude Code invokes on every turn, so it doubles as the
134
+ # refresh point for them: same rate limit, same detached fork, same
135
+ # content-diff no-op as the self-refresh above.
136
+ #
137
+ # Never clobber a wrapper the user edited. The only way to tell an edit from a
138
+ # stale copy is to remember what we last downloaded, so each canonical file is
139
+ # cached (unrendered) per install under the user cache dir and a wrapper is
140
+ # replaced only when its bytes still match that baseline. With no baseline yet
141
+ # we can't prove anything, so the first run only seeds the cache and a swap can
142
+ # happen from the next one on. install.sh seeds it too, so fresh installs skip
143
+ # that warm-up round.
144
+ #
145
+ # Never touch a wrapper git tracks either. A project-scope install writes the
146
+ # wrappers into the repo's own .claude/commands/, and unlike cc-statusline.sh
147
+ # the installer does not gitignore them — so an unattended weekly rewrite would
148
+ # surface as unexplained dirty files (and could ride along in someone's commit).
149
+ # Only the installer changes tracked files, and only when a human runs it.
150
+ #
151
+ # Opt out with WEAVE_COMMANDS_UPDATE=0 (WEAVE_STATUSLINE_UPDATE=0 disables this
152
+ # too, along with every other network path here). Override the source with
153
+ # WEAVE_COMMANDS_URL_BASE=..., e.g. for self-hosters who fork.
154
+
155
+ # weave_command_tracked_by_git returns 0 when $1 is a file git tracks. Used to
156
+ # leave repo-committed wrappers alone; no git (or no repo) means untracked.
157
+ weave_command_tracked_by_git() {
158
+ command -v git >/dev/null 2>&1 || return 1
159
+ git -C "$(dirname "$1")" ls-files --error-unmatch -- "$1" >/dev/null 2>&1
160
+ }
161
+
162
+ # weave_installed_command_names lists the wrappers this install may refresh.
163
+ # The statusline ships standalone (no registry.sh beside it), so the installed
164
+ # set — itself written from the registry — is the only source of truth here.
165
+ #
166
+ # Files written before ownership markers existed carry none, so matching on the
167
+ # marker alone would freeze every pre-marker install out of refreshes forever.
168
+ # List every wrapper instead and let the baseline comparison below decide: a
169
+ # file is replaced only when its bytes still match the last canonical copy, so
170
+ # a user-authored command is never touched whether or not it carries a marker.
171
+ weave_installed_command_names() {
172
+ local dir="$1" file
173
+ for file in "$dir"/*.md; do
174
+ [ -f "$file" ] || continue
175
+ printf '%s\n' "$(basename "$file" .md)"
176
+ done
177
+ }
178
+
179
+ # weave_render_command prints $1 with the installer's {{SCOPE}} placeholder
180
+ # replaced by $2, matching how install_slash_commands writes the same file.
181
+ # Trailing newlines are stripped on both sides of every comparison below.
182
+ weave_render_command() {
183
+ local body
184
+ body="$(cat "$1" 2>/dev/null)" || return 1
185
+ printf '%s' "${body//\{\{SCOPE\}\}/$2}"
186
+ }
187
+
188
+ weave_sync_commands() {
189
+ [ "${WEAVE_STATUSLINE_UPDATE:-1}" = "0" ] && return 0
190
+ [ "${WEAVE_COMMANDS_UPDATE:-1}" = "0" ] && return 0
191
+ command -v curl >/dev/null 2>&1 || return 0
192
+
193
+ local self="${BASH_SOURCE[0]:-$0}" self_dir cmd_dir
194
+ self_dir="$(cd "$(dirname "$self")" 2>/dev/null && pwd -P)" || return 0
195
+ # User scope installs this script at <base>/.weave/, project and --dir at
196
+ # <base>/.claude/. Claude Code reads commands from <base>/.claude/commands in
197
+ # both layouts.
198
+ case "${self_dir##*/}" in
199
+ .weave) cmd_dir="${self_dir%/*}/.claude/commands" ;;
200
+ .claude) cmd_dir="$self_dir/commands" ;;
201
+ *) return 0 ;;
202
+ esac
203
+ [ -d "$cmd_dir" ] && [ -w "$cmd_dir" ] || return 0
204
+
205
+ local cache_dir="${XDG_CACHE_HOME:-$HOME/.cache}/weave-router"
206
+ local dir_slug
207
+ dir_slug="$(printf '%s' "$cmd_dir" | tr -c 'A-Za-z0-9._-' '_')"
208
+ # Baseline is keyed by install, not shared: two installs refresh on their own
209
+ # clocks, and a shared baseline updated by one would make the other's still-
210
+ # canonical wrappers look user-edited and freeze them forever.
211
+ local baseline_dir="$cache_dir/commands${dir_slug}"
212
+ local stamp="$cache_dir/checked-at${dir_slug}.commands"
213
+ mkdir -p "$baseline_dir" 2>/dev/null || return 0
214
+
215
+ local interval_days="${WEAVE_STATUSLINE_UPDATE_INTERVAL_DAYS:-7}"
216
+ local now stamp_mtime
217
+ now="$(date +%s 2>/dev/null)" || return 0
218
+ if [ -f "$stamp" ]; then
219
+ stamp_mtime="$(stat -c %Y "$stamp" 2>/dev/null || stat -f %m "$stamp" 2>/dev/null)" || stamp_mtime=0
220
+ else
221
+ stamp_mtime=0
222
+ fi
223
+ if [ -n "${stamp_mtime:-}" ] && [ "$stamp_mtime" -gt 0 ] \
224
+ && [ $(( now - stamp_mtime )) -lt $(( interval_days * 86400 )) ]; then
225
+ return 0
226
+ fi
227
+ # Stamp before forking, same as the self-refresh: Claude Code calls us on
228
+ # every turn and concurrent invocations must not all start downloading.
229
+ : > "$stamp" 2>/dev/null || return 0
230
+
231
+ # router-off/on/status bake this install's scope selector into their npx
232
+ # line (upstream carries {{SCOPE}} there). Recover it from the installed copy
233
+ # so a project or --dir install keeps toggling its own config; when it can't
234
+ # be recovered those three are skipped rather than rewritten to point at the
235
+ # user-scope install.
236
+ local scope_args="" scope_known="false" off="$cmd_dir/router-off.md"
237
+ if [ -f "$off" ] && grep -Eq '^`npx @(workweave/router|weave-os/router) off --claude.*`$' "$off" 2>/dev/null; then
238
+ scope_known="true"
239
+ scope_args="$(sed -En 's#^`npx @(workweave/router|weave-os/router) off --claude(.*)`$#\2#p' "$off" | head -n 1)"
240
+ fi
241
+
242
+ local url_base="${WEAVE_COMMANDS_URL_BASE:-https://raw.githubusercontent.com/weave-os/router/main/install/commands}"
243
+ local name installed raw prev tmp new_body prev_body installed_body
244
+ (
245
+ # Detach stdin (CC pipes JSON to us) so curl can't consume it, and silence
246
+ # everything so no output leaks into the statusline.
247
+ exec </dev/null
248
+ while IFS= read -r name; do
249
+ installed="$cmd_dir/$name.md"
250
+ # Only ever refresh a wrapper that is already installed: a missing one
251
+ # was uninstalled or deliberately deleted, and resurrecting it would be
252
+ # a surprise. A symlink is user-owned; leave it alone. A git-tracked
253
+ # wrapper belongs to the repo — rewriting it would dirty a working tree
254
+ # nobody asked us to touch.
255
+ [ -f "$installed" ] || continue
256
+ [ -L "$installed" ] && continue
257
+ weave_command_tracked_by_git "$installed" && continue
258
+
259
+ raw="$baseline_dir/$name.md.tmp.$$"
260
+ curl -fsSL --max-time 15 "$url_base/$name.md" -o "$raw" 2>/dev/null || { rm -f "$raw"; continue; }
261
+ # Shape check: every wrapper opens with YAML front matter, so a 404 page
262
+ # or a truncated body can never be installed as a command.
263
+ if [ ! -s "$raw" ] || [ "$(head -n 1 "$raw")" != "---" ]; then
264
+ rm -f "$raw"
265
+ continue
266
+ fi
267
+
268
+ prev="$baseline_dir/$name.md"
269
+ if grep -q '{{SCOPE}}' "$raw" && [ "$scope_known" != "true" ]; then
270
+ mv "$raw" "$prev" 2>/dev/null || rm -f "$raw"
271
+ continue
272
+ fi
273
+
274
+ if [ -f "$prev" ]; then
275
+ new_body="$(weave_render_command "$raw" "$scope_args")"
276
+ prev_body="$(weave_render_command "$prev" "$scope_args")"
277
+ installed_body="$(cat "$installed" 2>/dev/null | sed '/^<!-- weave-router managed command: .* -->$/d')" || installed_body=""
278
+ if [ "$prev_body" = "$installed_body" ] && [ "$new_body" != "$installed_body" ]; then
279
+ tmp="$installed.tmp.$$"
280
+ if printf '%s\n<!-- weave-router managed command: %s -->' "$new_body" "$name" >"$tmp" 2>/dev/null; then
281
+ mv "$tmp" "$installed" 2>/dev/null || rm -f "$tmp"
282
+ else
283
+ rm -f "$tmp"
284
+ fi
285
+ fi
286
+ fi
287
+ mv "$raw" "$prev" 2>/dev/null || rm -f "$raw"
288
+ done <<EOF
289
+ $(weave_installed_command_names "$cmd_dir")
290
+ EOF
291
+ ) >/dev/null 2>&1 &
292
+ disown 2>/dev/null || true
293
+ return 0
294
+ }
295
+ weave_sync_commands 2>/dev/null || true
296
+
297
+ # ---------- org "hide terminal surfaces" gate ----------
298
+ #
299
+ # When the org has hidden the router's terminal surfaces, the statusline
300
+ # renders nothing: this prints no output and exits 0, leaving the slot blank.
301
+ # The setting comes from GET /v1/display-settings with the install's own
302
+ # router key, but the foreground path NEVER touches the network: it decides
303
+ # solely from a per-install cache file (TTL WEAVE_DISPLAY_SETTINGS_TTL_SECONDS,
304
+ # default 1h), and a missing or stale cache fails open — the statusline
305
+ # renders normally. When the cache is missing or stale a detached background
306
+ # refresh re-fetches the setting and rewrites the cache atomically, so the
307
+ # next turn picks up the fresh value; a refresh failure simply leaves the
308
+ # cache stale, which keeps failing open rather than pinning the gate closed.
309
+ #
310
+ # Claude Code runs the statusline every turn, so several invocations can see a
311
+ # stale cache at once. Letting each fetch independently is not safe: the
312
+ # responses can land out of order, and the loser's mv would replace a newer
313
+ # setting with an older one AND stamp it fresh, pinning the gate on a stale
314
+ # value for a full TTL. Refreshes are therefore serialized per cache key on a
315
+ # mkdir mutex (atomic everywhere; flock is absent on macOS), held across both
316
+ # the fetch and the write. A refresh that finds the mutex held exits rather
317
+ # than queueing — the in-flight one is at least as fresh as anything it would
318
+ # fetch, so waiting only to overwrite it is the bug. The one path that is not
319
+ # strictly exclusive is reclaiming a lock whose holder died; see below.
320
+ weave_hidden_gate() {
321
+ command -v curl >/dev/null 2>&1 || return 1
322
+ command -v jq >/dev/null 2>&1 || return 1
323
+
324
+ local cache_dir="${XDG_CACHE_HOME:-$HOME/.cache}/weave-router"
325
+ mkdir -p "$cache_dir" 2>/dev/null || return 1
326
+ local self="${BASH_SOURCE[0]:-$0}"
327
+ local script_slug
328
+ script_slug="$(printf '%s' "$self" | tr -c 'A-Za-z0-9._-' '_')"
329
+ local cache="$cache_dir/display-settings${script_slug}"
330
+ local ttl="${WEAVE_DISPLAY_SETTINGS_TTL_SECONDS:-3600}"
331
+
332
+ local now mtime fresh="false"
333
+ now="$(date +%s 2>/dev/null)" || now=0
334
+ if [ -f "$cache" ]; then
335
+ mtime="$(stat -c %Y "$cache" 2>/dev/null || stat -f %m "$cache" 2>/dev/null)" || mtime=0
336
+ if [ -n "${mtime:-}" ] && [ "$mtime" -gt 0 ] && [ $(( now - mtime )) -lt "$ttl" ]; then
337
+ fresh="true"
338
+ fi
339
+ fi
340
+
341
+ # Foreground decision: only a fresh cache hides the surfaces. Anything else
342
+ # (no cache, stale cache, unreadable cache) renders normally so a slow or
343
+ # unreachable router can never stall a turn or wedge the statusline blank.
344
+ if [ "$fresh" = "true" ]; then
345
+ [ "$(cat "$cache" 2>/dev/null)" = "1" ]
346
+ return
347
+ fi
348
+
349
+ # Background refresh for the next invocation. Resolve the router base URL
350
+ # and key inside the subshell from the Claude Code settings the installer
351
+ # wrote: project/--dir installs put both under <base>/.claude alongside
352
+ # this script (key in settings.local.json), while a user-scope install
353
+ # lives under ~/.weave and reads ~/.claude. Resolve relative to the
354
+ # script's own location, falling back to user scope, so a project install
355
+ # never reads (or leaks) the user-scope key. ANTHROPIC_BASE_URL and
356
+ # WEAVE_ROUTER_BASE_URL may also be set in the environment. A file:// base
357
+ # URL is the offline/test seam: curl reads it as the response body
358
+ # directly, so the endpoint path is meaningless for it; real router URLs
359
+ # (https) get /v1/display-settings appended.
360
+ (
361
+ exec </dev/null
362
+
363
+ # Take the per-cache-key mutex, or bail. mkdir is the portable atomic
364
+ # test-and-set. A crashed holder would otherwise block refreshes forever,
365
+ # so a lock older than the fetch timeout is treated as abandoned and
366
+ # reclaimed. Releasing from a trap covers every exit path below.
367
+ lock="$cache.lock"
368
+ if ! mkdir "$lock" 2>/dev/null; then
369
+ lock_mtime="$(stat -c %Y "$lock" 2>/dev/null || stat -f %m "$lock" 2>/dev/null)" || lock_mtime=0
370
+ lock_now="$(date +%s 2>/dev/null)" || lock_now=0
371
+ if [ "${lock_mtime:-0}" -le 0 ] || [ $(( lock_now - lock_mtime )) -le 30 ]; then
372
+ exit 0
373
+ fi
374
+ # Reclaiming an abandoned lock must not be delete-then-recreate. With
375
+ # `rm -rf` + `mkdir`, refreshers that all see the same stale lock each
376
+ # delete the next one's freshly created directory, so many end up holding
377
+ # it at once (measured at 50 claimants: 4-11 concurrent holders) and their
378
+ # writes can land out of order. Renaming is atomic, so only one racer can
379
+ # move a given directory aside and the losers exit instead of clobbering
380
+ # the winner (same measurement: 1-2). It is not a perfect mutex — a
381
+ # straggler can still reclaim the new lock a winner just created, since
382
+ # nothing distinguishes it from the stale one — but real refreshes arrive
383
+ # one per turn rather than 50 at once, and the staleness threshold below
384
+ # is what bounds the rest.
385
+ dead="$lock.dead.$$"
386
+ mv "$lock" "$dead" 2>/dev/null || exit 0
387
+ rm -rf "$dead" 2>/dev/null
388
+ mkdir "$lock" 2>/dev/null || exit 0
389
+ fi
390
+ trap 'rmdir "$lock" 2>/dev/null' EXIT
391
+
392
+ self_dir="$(cd "$(dirname "$self")" 2>/dev/null && pwd)"
393
+ settings_base="$HOME"
394
+ case "$self_dir" in
395
+ */.claude) settings_base="${self_dir%/.claude}" ;;
396
+ esac
397
+ settings="$settings_base/.claude/settings.json"
398
+ local_settings="$settings_base/.claude/settings.local.json"
399
+ base_url="${WEAVE_ROUTER_BASE_URL:-${ANTHROPIC_BASE_URL:-}}"
400
+ key="${WEAVE_ROUTER_KEY:-}"
401
+ if [ -z "$key" ] && [ -f "$settings" ]; then
402
+ key="$(jq -r '.env.ANTHROPIC_CUSTOM_HEADERS // "" | split("\n")[] | select(startswith("X-Weave-Router-Key:")) | sub("^X-Weave-Router-Key:[[:space:]]*";"")' "$settings" 2>/dev/null | head -n1)"
403
+ fi
404
+ if [ -z "$key" ] && [ -f "$local_settings" ]; then
405
+ key="$(jq -r '.env.ANTHROPIC_CUSTOM_HEADERS // "" | split("\n")[] | select(startswith("X-Weave-Router-Key:")) | sub("^X-Weave-Router-Key:[[:space:]]*";"")' "$local_settings" 2>/dev/null | head -n1)"
406
+ fi
407
+ if [ -z "$base_url" ] && [ -f "$settings" ]; then
408
+ base_url="$(jq -r '.env.ANTHROPIC_BASE_URL // empty' "$settings" 2>/dev/null)"
409
+ fi
410
+ [ -n "$base_url" ] && [ -n "$key" ] || exit 0
411
+ url="${base_url%/}"
412
+ case "$url" in
413
+ file://*) ;;
414
+ *) url="$url/v1/display-settings" ;;
415
+ esac
416
+ body="$(curl -fsS --max-time 5 -H "X-Weave-Router-Key: $key" "$url" 2>/dev/null)" || exit 0
417
+ hidden="$(printf '%s' "$body" | jq -r '.hide_terminal_surfaces // false' 2>/dev/null)"
418
+ tmp="$cache.tmp.$$"
419
+ if [ "$hidden" = "true" ]; then
420
+ printf '1' >"$tmp" 2>/dev/null && mv "$tmp" "$cache" 2>/dev/null
421
+ else
422
+ printf '0' >"$tmp" 2>/dev/null && mv "$tmp" "$cache" 2>/dev/null
423
+ fi
424
+ rm -f "$tmp" 2>/dev/null
425
+ ) >/dev/null 2>&1 &
426
+ disown 2>/dev/null || true
427
+ return 1
428
+ }
429
+
430
+ if weave_hidden_gate </dev/null; then
431
+ exit 0
432
+ fi
433
+
434
+ input="$(cat)"
435
+ transcript_path="$(printf '%s' "$input" | jq -r '.transcript_path // empty')"
436
+ # Prefer model.id over display_name: pricing keys + the routed model id in
437
+ # the transcript are canonical ids (e.g. claude-opus-4-7), while display_name
438
+ # is a human label ("Opus 4.7 (1M context)") that won't hit the pricing table,
439
+ # zeroing out savings. id passes through normalize_model cleanly.
440
+ selected_display="$(printf '%s' "$input" | jq -r '.model.id // .model.display_name // "?"')"
441
+
442
+ # Normalize a model id to a pricing-table key. CC + the decisions log carry
443
+ # two flavors of annotation we don't want in the lookup:
444
+ # * date suffix: claude-opus-4-7-20260101 → claude-opus-4-7
445
+ # * variant tag: claude-opus-4-7[1m] → claude-opus-4-7
446
+ # The 1M-context variant prices ~2× base for prompts >200k tokens, but for
447
+ # the "saved $X vs your selection" UX the base rate is the right comparison
448
+ # — we're measuring the model swap, not the context tier. Used below on the
449
+ # routed and requested model ids from the decisions log / transcript.
450
+ normalize_model() {
451
+ printf '%s' "$1" | sed -E 's/\[[^]]*\]$//; s/-[0-9]{8}$//'
452
+ }
453
+
454
+ # USD per 1k tokens. Generated from internal/observability/otel/pricing.go
455
+ # (USD/1M there, ÷1000 here) by cmd/genprices. Do not hand-edit — run
456
+ # `make generate` after updating pricing.go.
457
+ # BEGIN_GENERATED_PRICES
458
+ prices='{
459
+ "input": {
460
+ "claude-fable-5": 0.01,
461
+ "claude-fable-5-1": 0.01,
462
+ "claude-haiku-4-5": 0.001,
463
+ "claude-opus-4-0": 0.015,
464
+ "claude-opus-4-1": 0.015,
465
+ "claude-opus-4-5": 0.005,
466
+ "claude-opus-4-6": 0.005,
467
+ "claude-opus-4-7": 0.005,
468
+ "claude-opus-4-8": 0.005,
469
+ "claude-opus-5": 0.005,
470
+ "claude-sonnet-4-5": 0.003,
471
+ "claude-sonnet-4-6": 0.003,
472
+ "claude-sonnet-5": 0.003,
473
+ "deepseek/deepseek-v4-flash": 0.0001134,
474
+ "deepseek/deepseek-v4-pro": 0.00174,
475
+ "deepseek/deepseek-v4-pro-0813": 0.00174,
476
+ "gemini-2.0-flash": 0.0001,
477
+ "gemini-2.0-flash-lite": 0.000075,
478
+ "gemini-2.5-flash": 0.0003,
479
+ "gemini-2.5-flash-lite": 0.0001,
480
+ "gemini-2.5-pro": 0.00125,
481
+ "gemini-3-flash-preview": 0.0005,
482
+ "gemini-3-pro-preview": 0.002,
483
+ "gemini-3.1-flash-lite-preview": 0.0001,
484
+ "gemini-3.1-pro-preview": 0.002,
485
+ "gemini-3.5-flash": 0.0015,
486
+ "gemini-3.5-flash-lite": 0.0003,
487
+ "gemini-3.6-flash": 0.0015,
488
+ "gemini-3.7-flash": 0.0015,
489
+ "gemini-3.8-flash": 0.0015,
490
+ "google/gemma-4-26b-a4b-it": 0.00015,
491
+ "gpt-4.1": 0.002,
492
+ "gpt-4.1-mini": 0.0004,
493
+ "gpt-4.1-nano": 0.0001,
494
+ "gpt-4o": 0.0025,
495
+ "gpt-4o-mini": 0.00015,
496
+ "gpt-5": 0.0025,
497
+ "gpt-5-chat": 0.0025,
498
+ "gpt-5-mini": 0.0005,
499
+ "gpt-5-nano": 0.0001,
500
+ "gpt-5.4": 0.0025,
501
+ "gpt-5.4-mini": 0.00075,
502
+ "gpt-5.4-nano": 0.0002,
503
+ "gpt-5.4-pro": 0.03,
504
+ "gpt-5.5": 0.005,
505
+ "gpt-5.5-mini": 0.0005,
506
+ "gpt-5.5-nano": 0.00015,
507
+ "gpt-5.5-pro": 0.03,
508
+ "gpt-5.6-luna": 0.001,
509
+ "gpt-5.6-luna-pro": 0.001,
510
+ "gpt-5.6-sol": 0.005,
511
+ "gpt-5.6-sol-pro": 0.005,
512
+ "gpt-5.6-terra": 0.0025,
513
+ "grok-4.5": 0.002,
514
+ "grok-4.6": 0.002,
515
+ "minimax/minimax-m2.7": 0.0003,
516
+ "minimax/minimax-m3": 0.0003,
517
+ "mistralai/mistral-small-2603": 0.0002,
518
+ "moonshotai/kimi-k2.5": 0.0006,
519
+ "moonshotai/kimi-k2.6": 0.00095,
520
+ "moonshotai/kimi-k2.7": 0.00095,
521
+ "moonshotai/kimi-k3": 0.003,
522
+ "muse-spark-1.3": 0.00125,
523
+ "qwen/qwen3-235b-a22b-2507": 0.0002266,
524
+ "qwen/qwen3-30b-a3b-instruct-2507": 0.00015,
525
+ "qwen/qwen3-coder": 0.0009,
526
+ "qwen/qwen3-coder-next": 0.0005,
527
+ "qwen/qwen3-next-80b-a3b-instruct": 0.00015,
528
+ "qwen/qwen3.5-flash-02-23": 0.00005,
529
+ "qwen/qwen3.6-35b-a3b": 0.00015,
530
+ "qwen/qwen3.7-plus": 0.0004,
531
+ "qwen/qwen3.8-max": 0.002,
532
+ "xiaomi/mimo-v2.5-pro": 0.001,
533
+ "z-ai/glm-5": 0.001,
534
+ "z-ai/glm-5.1": 0.0014,
535
+ "z-ai/glm-5.2": 0.0014,
536
+ "z-ai/glm-5.3": 0.0014,
537
+ "z-ai/glm-5.3-flash": 0.00015
538
+ },
539
+ "output": {
540
+ "claude-fable-5": 0.05,
541
+ "claude-fable-5-1": 0.05,
542
+ "claude-haiku-4-5": 0.005,
543
+ "claude-opus-4-0": 0.075,
544
+ "claude-opus-4-1": 0.075,
545
+ "claude-opus-4-5": 0.025,
546
+ "claude-opus-4-6": 0.025,
547
+ "claude-opus-4-7": 0.025,
548
+ "claude-opus-4-8": 0.025,
549
+ "claude-opus-5": 0.025,
550
+ "claude-sonnet-4-5": 0.015,
551
+ "claude-sonnet-4-6": 0.015,
552
+ "claude-sonnet-5": 0.015,
553
+ "deepseek/deepseek-v4-flash": 0.0002791,
554
+ "deepseek/deepseek-v4-pro": 0.00348,
555
+ "deepseek/deepseek-v4-pro-0813": 0.00348,
556
+ "gemini-2.0-flash": 0.0004,
557
+ "gemini-2.0-flash-lite": 0.0003,
558
+ "gemini-2.5-flash": 0.0012,
559
+ "gemini-2.5-flash-lite": 0.0004,
560
+ "gemini-2.5-pro": 0.005,
561
+ "gemini-3-flash-preview": 0.002,
562
+ "gemini-3-pro-preview": 0.008,
563
+ "gemini-3.1-flash-lite-preview": 0.0004,
564
+ "gemini-3.1-pro-preview": 0.008,
565
+ "gemini-3.5-flash": 0.009,
566
+ "gemini-3.5-flash-lite": 0.0025,
567
+ "gemini-3.6-flash": 0.0075,
568
+ "gemini-3.7-flash": 0.0075,
569
+ "gemini-3.8-flash": 0.0075,
570
+ "google/gemma-4-26b-a4b-it": 0.0006,
571
+ "gpt-4.1": 0.008,
572
+ "gpt-4.1-mini": 0.0016,
573
+ "gpt-4.1-nano": 0.0004,
574
+ "gpt-4o": 0.01,
575
+ "gpt-4o-mini": 0.0006,
576
+ "gpt-5": 0.01,
577
+ "gpt-5-chat": 0.01,
578
+ "gpt-5-mini": 0.002,
579
+ "gpt-5-nano": 0.0004,
580
+ "gpt-5.4": 0.015,
581
+ "gpt-5.4-mini": 0.0045,
582
+ "gpt-5.4-nano": 0.00125,
583
+ "gpt-5.4-pro": 0.18,
584
+ "gpt-5.5": 0.03,
585
+ "gpt-5.5-mini": 0.0025,
586
+ "gpt-5.5-nano": 0.0006,
587
+ "gpt-5.5-pro": 0.18,
588
+ "gpt-5.6-luna": 0.006,
589
+ "gpt-5.6-luna-pro": 0.006,
590
+ "gpt-5.6-sol": 0.03,
591
+ "gpt-5.6-sol-pro": 0.03,
592
+ "gpt-5.6-terra": 0.015,
593
+ "grok-4.5": 0.006,
594
+ "grok-4.6": 0.006,
595
+ "minimax/minimax-m2.7": 0.0012,
596
+ "minimax/minimax-m3": 0.0012,
597
+ "mistralai/mistral-small-2603": 0.0006,
598
+ "moonshotai/kimi-k2.5": 0.003,
599
+ "moonshotai/kimi-k2.6": 0.004,
600
+ "moonshotai/kimi-k2.7": 0.004,
601
+ "moonshotai/kimi-k3": 0.015,
602
+ "muse-spark-1.3": 0.00425,
603
+ "qwen/qwen3-235b-a22b-2507": 0.0009064,
604
+ "qwen/qwen3-30b-a3b-instruct-2507": 0.0006,
605
+ "qwen/qwen3-coder": 0.0027,
606
+ "qwen/qwen3-coder-next": 0.0012,
607
+ "qwen/qwen3-next-80b-a3b-instruct": 0.0012,
608
+ "qwen/qwen3.5-flash-02-23": 0.00015,
609
+ "qwen/qwen3.6-35b-a3b": 0.001,
610
+ "qwen/qwen3.7-plus": 0.0016,
611
+ "qwen/qwen3.8-max": 0.006,
612
+ "xiaomi/mimo-v2.5-pro": 0.003,
613
+ "z-ai/glm-5": 0.0032,
614
+ "z-ai/glm-5.1": 0.0044,
615
+ "z-ai/glm-5.2": 0.0044,
616
+ "z-ai/glm-5.3": 0.0044,
617
+ "z-ai/glm-5.3-flash": 0.0005
618
+ },
619
+ "cache_read": {
620
+ "claude-fable-5": 0.1,
621
+ "claude-fable-5-1": 0.025,
622
+ "claude-haiku-4-5": 0.1,
623
+ "claude-opus-4-0": 0.1,
624
+ "claude-opus-4-1": 0.1,
625
+ "claude-opus-4-5": 0.1,
626
+ "claude-opus-4-6": 0.1,
627
+ "claude-opus-4-7": 0.1,
628
+ "claude-opus-4-8": 0.1,
629
+ "claude-opus-5": 0.1,
630
+ "claude-sonnet-4-5": 0.1,
631
+ "claude-sonnet-4-6": 0.1,
632
+ "claude-sonnet-5": 0.1,
633
+ "deepseek/deepseek-v4-flash": 0.2,
634
+ "deepseek/deepseek-v4-pro": 0.11494252873563218,
635
+ "deepseek/deepseek-v4-pro-0813": 0.11494252873563218,
636
+ "gemini-2.0-flash": 0.25,
637
+ "gemini-2.0-flash-lite": 0.25,
638
+ "gemini-2.5-flash": 0.1,
639
+ "gemini-2.5-flash-lite": 0.1,
640
+ "gemini-2.5-pro": 0.1,
641
+ "gemini-3-flash-preview": 0.1,
642
+ "gemini-3-pro-preview": 0.1,
643
+ "gemini-3.1-flash-lite-preview": 0.1,
644
+ "gemini-3.1-pro-preview": 0.1,
645
+ "gemini-3.5-flash": 0.1,
646
+ "gemini-3.5-flash-lite": 0.1,
647
+ "gemini-3.6-flash": 0.1,
648
+ "gemini-3.7-flash": 0.1,
649
+ "gemini-3.8-flash": 0.1,
650
+ "google/gemma-4-26b-a4b-it": 0.1,
651
+ "gpt-4.1": 0.25,
652
+ "gpt-4.1-mini": 0.25,
653
+ "gpt-4.1-nano": 0.25,
654
+ "gpt-4o": 0.5,
655
+ "gpt-4o-mini": 0.5,
656
+ "gpt-5": 0.1,
657
+ "gpt-5-chat": 0.1,
658
+ "gpt-5-mini": 0.1,
659
+ "gpt-5-nano": 0.1,
660
+ "gpt-5.4": 0.1,
661
+ "gpt-5.4-mini": 0.1,
662
+ "gpt-5.4-nano": 0.1,
663
+ "gpt-5.4-pro": 1,
664
+ "gpt-5.5": 0.1,
665
+ "gpt-5.5-mini": 0.1,
666
+ "gpt-5.5-nano": 0.1,
667
+ "gpt-5.5-pro": 1,
668
+ "gpt-5.6-luna": 0.1,
669
+ "gpt-5.6-luna-pro": 0.1,
670
+ "gpt-5.6-sol": 0.1,
671
+ "gpt-5.6-sol-pro": 0.1,
672
+ "gpt-5.6-terra": 0.1,
673
+ "grok-4.5": 0.25,
674
+ "grok-4.6": 0.25,
675
+ "minimax/minimax-m2.7": 0.2,
676
+ "minimax/minimax-m3": 0.2,
677
+ "mistralai/mistral-small-2603": 0.1,
678
+ "moonshotai/kimi-k2.5": 0.5,
679
+ "moonshotai/kimi-k2.6": 0.1684,
680
+ "moonshotai/kimi-k2.7": 0.2,
681
+ "moonshotai/kimi-k3": 0.1,
682
+ "muse-spark-1.3": 0.12,
683
+ "qwen/qwen3-235b-a22b-2507": 0.5,
684
+ "qwen/qwen3-30b-a3b-instruct-2507": 0.1684,
685
+ "qwen/qwen3-coder": 0.1684,
686
+ "qwen/qwen3-coder-next": 0.5,
687
+ "qwen/qwen3-next-80b-a3b-instruct": 0.5,
688
+ "qwen/qwen3.5-flash-02-23": 0.1,
689
+ "qwen/qwen3.6-35b-a3b": 0.1,
690
+ "qwen/qwen3.7-plus": 0.2,
691
+ "qwen/qwen3.8-max": 0.125,
692
+ "xiaomi/mimo-v2.5-pro": 0.1,
693
+ "z-ai/glm-5": 0.2,
694
+ "z-ai/glm-5.1": 0.18571428571428572,
695
+ "z-ai/glm-5.2": 0.18571428571428572,
696
+ "z-ai/glm-5.3": 0.18571428571428572,
697
+ "z-ai/glm-5.3-flash": 0.2
698
+ }
699
+ }'
700
+ # END_GENERATED_PRICES
701
+
702
+ routed=""
703
+ forced="false"
704
+ forced_model=""
705
+ session_savings=""
706
+ tot_in=0
707
+ tot_out=0
708
+ tot_cache_read=0
709
+ tot_cache_write=0
710
+
711
+ # Per-turn savings compare each turn's routed cost (priced from
712
+ # message.model in the transcript) against what the CC-side model selection
713
+ # (selected_display) would have cost on the same tokens. The selection
714
+ # isn't strictly the per-turn "requested" model — CC tags some background
715
+ # side-calls (compaction probes, title-gen) with a different model id —
716
+ # but for those the planner short-circuits to a hard pin and the savings
717
+ # math zeroes out anyway. Turns where routed == selection or where either
718
+ # model isn't in the pricing table emit 0 savings; the tokens clause
719
+ # always renders.
720
+
721
+ # Normalize the CC-side selection once for use in the jq math below.
722
+ requested_norm="$(normalize_model "$selected_display")"
723
+
724
+ if [[ -n "$transcript_path" && -f "$transcript_path" ]]; then
725
+ # macOS ships `tail -r`, GNU coreutils ships `tac`. Either works to walk the
726
+ # JSONL in reverse so we can grab the latest assistant turn.
727
+ if command -v tac >/dev/null 2>&1; then reverse=(tac); else reverse=(tail -r); fi
728
+
729
+ # CC stamps message.model = "<synthetic>" on assistant turns it generated
730
+ # locally (errored requests, cancellations, tool-only stubs) instead of a
731
+ # real model id. Show that as "failure" rather than leaking the internal
732
+ # sentinel into the statusline.
733
+ routed="$("${reverse[@]}" "$transcript_path" 2>/dev/null \
734
+ | jq -r 'select(.type=="assistant") | .message.model // empty' \
735
+ | head -n 1 || true)"
736
+ if [[ "$routed" == "<synthetic>" ]]; then
737
+ routed="failure"
738
+ else
739
+ routed="$(normalize_model "$routed")"
740
+ fi
741
+
742
+ # Detect an active /force-model pin from the router's synthetic ack turns —
743
+ # the only turns stamped message.model == "weave-router". The router emits
744
+ # one whenever a pin changes state:
745
+ # * /force-model → "force-model applied: <model> (<provider>) …"
746
+ # * /unforce-model → "force-model cleared …"
747
+ # * loop / no-progress break → "… clearing the session pin …" (expires the
748
+ # pin, including a user-forced one)
749
+ # * unrecognized model → "… isn't a recognized model · keeping
750
+ # automatic routing" — a NO-OP: the prior
751
+ # pin, if any, is left untouched
752
+ # * listing (historical) → "… pick a model by id …" / "no models are
753
+ # available to pin …" — also NO-OPs. The
754
+ # router no longer emits these, but a
755
+ # transcript written while it did still
756
+ # carries them, and they must not read as
757
+ # a clear or a live pin loses its [forced]
758
+ # These persist on disk (the ingress stripper only scrubs them from upstream
759
+ # requests). Classify each weave-router turn newest-first, skip the no-op
760
+ # "rejected" acks, and let the latest real state change decide: an "applied"
761
+ # marker means the session is pinned (and names the model); anything else
762
+ # (cleared / loop-break / no-progress) means automatic routing has resumed.
763
+ # Restricting to weave-router turns keeps a normal reply that merely quotes
764
+ # these phrases from flipping the tag. (A silent server-side TTL expiry emits
765
+ # no turn and so can't be reflected here — the pin TTL outlives a session.)
766
+ force_state="$("${reverse[@]}" "$transcript_path" 2>/dev/null \
767
+ | jq -r 'select(.type=="assistant" and .message.model=="weave-router")
768
+ | ([.message.content[]? | select(.type? == "text") | .text] | join(" ") | gsub("[\n\r]"; " ")) as $t
769
+ | if ($t | test("force-model applied:")) then "APPLIED " + ($t | capture("force-model applied: (?<m>[^ ]+)").m)
770
+ elif ($t | test("isn.t a recognized model")) then "REJECTED"
771
+ elif ($t | test("pick a model by id|no models are available to pin")) then "REJECTED"
772
+ else "CLEARED" end' 2>/dev/null \
773
+ | grep -m1 -v '^REJECTED$' || true)"
774
+ if [[ "$force_state" == APPLIED\ * ]]; then
775
+ forced="true"
776
+ forced_model="${force_state#APPLIED }"
777
+ fi
778
+
779
+ # Compute a session running total: savings across every assistant turn
780
+ # whose marker reports a requested ≠ routed swap, plus cumulative token
781
+ # counts across every assistant turn (rerouted or not — total work the
782
+ # session has done). cache_creation is priced at 1.25× input; cache_read
783
+ # uses each model's generated catalog multiplier. Both are no-ops when the
784
+ # provider does not return those fields.
785
+ #
786
+ # The marker regex tolerates the optional "(<provider>)" segment and a
787
+ # `[1m]` / `-YYYYMMDD` suffix on either model name so transcripts written
788
+ # against context-tiered or dated model ids still parse cleanly.
789
+ #
790
+ # Dedup note: CC writes one JSONL entry per *content block* in an
791
+ # assistant turn (text, text, tool_use → 3 entries), and every entry
792
+ # carries the same `message.usage`. Summing per-entry triple-counts the
793
+ # turn. We dedupe on (message.id, message.usage) before summing:
794
+ # * For native Anthropic upstreams message.id is unique per turn, so
795
+ # this collapses the content-block fan-out cleanly.
796
+ # * For non-Anthropic upstreams that round-trip through the router's
797
+ # translator, message.id can be a constant placeholder
798
+ # ("msg_translated"); usage still differs per turn (input_tokens
799
+ # grows), so the composite key keeps turns distinct. Two turns with
800
+ # byte-identical id AND usage would still collapse, but that's a
801
+ # genuine retry/duplicate we want to drop.
802
+ read -r session_savings tot_in tot_out tot_cache_read tot_cache_write < <(
803
+ jq -rs --argjson p "$prices" --arg requested "$requested_norm" '
804
+ [.[] | select(.type=="assistant")] |
805
+ unique_by([.message.id, .message.usage]) |
806
+ .[] |
807
+ .message as $m |
808
+ ($m.model // "" | sub("\\[[^]]*\\]$"; "") | sub("-[0-9]{8}$"; "")) as $rm |
809
+ {
810
+ in: ($m.usage.input_tokens // 0),
811
+ out: ($m.usage.output_tokens // 0),
812
+ cwrt: ($m.usage.cache_creation_input_tokens // 0),
813
+ crd: ($m.usage.cache_read_input_tokens // 0)
814
+ } as $t |
815
+ (if $requested == "" or $requested == $rm then 0
816
+ else
817
+ ($p.input[$rm] // null) as $rin | ($p.output[$rm] // null) as $rout |
818
+ ($p.cache_read[$rm] // null) as $rcr |
819
+ ($p.input[$requested] // null) as $sin | ($p.output[$requested] // null) as $sout |
820
+ ($p.cache_read[$requested] // null) as $scr |
821
+ if ($rin == null or $rout == null or $rcr == null or $sin == null or $sout == null or $scr == null) then 0
822
+ else
823
+ (($t.in + 1.25 * $t.cwrt + $rcr * $t.crd) / 1000) as $routed_input_units |
824
+ (($t.in + 1.25 * $t.cwrt + $scr * $t.crd) / 1000) as $requested_input_units |
825
+ ($t.out / 1000) as $output_units |
826
+ ($routed_input_units * $rin + $output_units * $rout) as $routed_cost |
827
+ ($requested_input_units * $sin + $output_units * $sout) as $requested_cost |
828
+ ($requested_cost - $routed_cost)
829
+ end
830
+ end) as $savings |
831
+ "\($savings) \($t.in) \($t.out) \($t.crd) \($t.cwrt)"
832
+ ' "$transcript_path" 2>/dev/null \
833
+ | awk 'BEGIN{s=0; i=0; o=0; r=0; w=0}
834
+ {s+=$1; i+=$2; o+=$3; r+=$4; w+=$5}
835
+ END{printf "%.4f %d %d %d %d\n", s, i, o, r, w}'
836
+ ) || true
837
+ fi
838
+
839
+ # ---------- refresh when a model isn't in the pricing table ----------
840
+ #
841
+ # A model that shipped after this copy of the script has no price entry, so the
842
+ # jq guard above zeroes savings on every turn and the line reads "saved $0.00"
843
+ # — indistinguishable from "the router ran and didn't beat your selection". The
844
+ # periodic check heals that eventually, but a week late, and a model launch is
845
+ # exactly when the number gets looked at. Refresh off-schedule instead.
846
+ #
847
+ # Keyed on its own stamp per unpriced id: a model we never price (self-hosted,
848
+ # unrecognized) then costs one download per interval, not one per turn, and
849
+ # can't starve the periodic check.
850
+ weave_refresh_on_price_miss() {
851
+ local candidates="" m
852
+ for m in "$@"; do
853
+ case "$m" in
854
+ "" | "?" | failure | weave-router | "<synthetic>") continue ;;
855
+ esac
856
+ candidates="${candidates}${m}"$'\n'
857
+ done
858
+ [ -n "$candidates" ] || return 0
859
+
860
+ # A malformed $prices emits nothing here, which fails closed (no refresh)
861
+ # rather than re-downloading every turn.
862
+ local missing
863
+ missing="$(printf '%s' "$candidates" \
864
+ | jq -rR --argjson p "$prices" \
865
+ 'select($p.input[.] == null or $p.output[.] == null)' 2>/dev/null \
866
+ | head -n 1)" || return 0
867
+ [ -n "$missing" ] || return 0
868
+
869
+ local model_slug
870
+ model_slug="$(printf '%s' "$missing" | tr -c 'A-Za-z0-9._-' '_')"
871
+ weave_self_refresh ".miss.${model_slug}" 1
872
+ }
873
+ weave_refresh_on_price_miss "$requested_norm" "$routed" 2>/dev/null || true
874
+
875
+ # Brand color (#FF6C47) on terminals that grok 24-bit truecolor — that's
876
+ # every modern one (iTerm2, Apple Terminal, vscode, ghostty, alacritty,
877
+ # wezterm, kitty). Falls back gracefully on any escape-stripping terminal.
878
+ brand=$'\033[38;2;255;108;71mWEAVE ROUTER\033[0m'
879
+
880
+ # Format helpers.
881
+ fmt_money() {
882
+ awk -v v="$1" 'BEGIN{
883
+ if (v == "" || v+0 == 0) { printf "$0.00"; exit }
884
+ if (v+0 < 0.005 && v+0 > -0.005){ printf "<$0.01"; exit }
885
+ if (v+0 < 0) { printf "-$%.2f", -v+0; exit }
886
+ printf "$%.2f", v
887
+ }'
888
+ }
889
+
890
+ fmt_tok() {
891
+ awk -v v="$1" 'BEGIN{
892
+ v = v+0
893
+ if (v >= 1000000) { printf "%.1fM", v/1000000; exit }
894
+ if (v >= 1000) { printf "%.1fk", v/1000; exit }
895
+ printf "%d", v
896
+ }'
897
+ }
898
+
899
+ # cache_read tokens are the cached portion of every prompt that the
900
+ # provider serves at 0.1× input price; cache_write tokens are the bytes
901
+ # that get newly cached on this turn at 1.25× input price. They behave
902
+ # completely differently both in cost and in what they tell the user
903
+ # about session-level efficiency, so we surface them separately rather
904
+ # than summing into a single "cached" number that conflates the two.
905
+ # Each clause is shown only when nonzero, so quiet sessions stay quiet.
906
+ tokens_clause=""
907
+ if [[ "$tot_in" -gt 0 || "$tot_out" -gt 0 || "$tot_cache_read" -gt 0 || "$tot_cache_write" -gt 0 ]]; then
908
+ tokens_clause=" · $(fmt_tok "$tot_in") in / $(fmt_tok "$tot_out") out"
909
+ if [[ "$tot_cache_read" -gt 0 ]]; then
910
+ tokens_clause+=" / $(fmt_tok "$tot_cache_read") cache read"
911
+ fi
912
+ if [[ "$tot_cache_write" -gt 0 ]]; then
913
+ tokens_clause+=" / $(fmt_tok "$tot_cache_write") cache write"
914
+ fi
915
+ fi
916
+
917
+ if [[ "$forced" == "true" ]]; then
918
+ # Session is pinned via /force-model. The "← selection · saved $X" clause
919
+ # describes automatic routing and would be misleading on a manual pin, so
920
+ # show the pinned model with a [forced] tag instead. forced_model comes from
921
+ # the marker; fall back to the routed/selected id if parsing came up empty.
922
+ forced_display="${forced_model:-${routed:-$selected_display}}"
923
+ printf '%s — %s [forced]%s' "$brand" "$forced_display" "$tokens_clause"
924
+ elif [[ "$routed" == "failure" ]]; then
925
+ # Latest turn was a CC-synthesized error stub — don't claim a routing
926
+ # swap or compute savings against a non-model.
927
+ printf '%s — %s%s' "$brand" "$routed" "$tokens_clause"
928
+ elif [[ -n "$routed" ]]; then
929
+ # Always show the savings clause, flooring a non-positive total at $0.00.
930
+ # session_savings is "0.0000" on fresh sessions or sessions where every
931
+ # turn routed back to the selected model, and can go negative when routing
932
+ # upgrades a turn to a pricier model (e.g. a sticky/hard-pinned side-call,
933
+ # or a hard turn escalated to opus). "saved -$X" would mislead, so clamp
934
+ # the display to $0.00 rather than dropping the clause — a $0.00 readout
935
+ # tells the user the router ran and simply didn't beat their selection.
936
+ # When the CC selection is unknown ("?" or empty) there's nothing to
937
+ # compare against, so drop the "← selection" arrow and just show routed.
938
+ display_savings="$session_savings"
939
+ if [[ -z "$display_savings" ]] \
940
+ || awk -v v="$display_savings" 'BEGIN{exit !(v+0 < 0)}'; then
941
+ display_savings="0"
942
+ fi
943
+ if [[ -n "$selected_display" && "$selected_display" != "?" ]]; then
944
+ printf '%s — %s ← %s · saved %s%s' \
945
+ "$brand" "$routed" "$selected_display" "$(fmt_money "$display_savings")" "$tokens_clause"
946
+ else
947
+ printf '%s — %s%s' "$brand" "$routed" "$tokens_clause"
948
+ fi
949
+ else
950
+ printf '%s — %s%s' "$brand" "$selected_display" "$tokens_clause"
951
+ fi