@alexeiled/claude-router 1.7.0 → 1.7.2

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.
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "router",
3
3
  "displayName": "Router",
4
- "version": "1.7.0",
4
+ "version": "1.7.2",
5
5
  "description": "Model and effort routing for the main conversation by task difficulty and activity, advised by a prompt classifier (Jev, Clef, Clef Flash, OpenAI or local Ollama). Preserves subagent model choices.",
6
6
  "author": {
7
7
  "name": "Alexei Ledenev",
package/CHANGELOG.md CHANGED
@@ -1,5 +1,36 @@
1
1
  # Changelog
2
2
 
3
+ ## 1.7.2 (2026-10-10)
4
+
5
+ ### Fixed
6
+
7
+ - A Manual turn ends the watch behind **Escalations** (ACROSS SESSIONS). A turn that moved to a cheaper activity route and finished after routing was turned off mid-turn could count a later escalation against that move.
8
+
9
+ ### Added
10
+
11
+ - More live evidence in `experiments/mod-router/results/activity-live-acceptance.json`: the Sonnet → Haiku → Sonnet → Opus chain with tool calls, Sonnet and Opus thinking blocks after a Haiku detour, and the `shadow` would-route line, with no API errors.
12
+
13
+ ## 1.7.1 (2026-10-10)
14
+
15
+ Readouts the activity routing plan asked for, and a record of the live checks behind 1.7.0.
16
+
17
+ ### Added
18
+
19
+ - The Now tab's **Why** line names the activity cell when activity routing is on and the activity moved the route or the route is an override: `code at low runs on Sonnet 5.5 · high`. Tier moves, holds and refused moves keep their reason.
20
+ - The Usage tab's **ACTIVITY** table adds `est. cost`, the list price of each activity's replies (the same estimate as Routing vs your model, shown only when every reply was priced), and `mostly on`, the route most of its replies used. A narrow pane drops these columns first.
21
+ - **ACROSS SESSIONS** adds **Escalations** (`after a cheaper activity move: 1 in 12 moves`) and **Latency** (`Jev p95 ≤ 350 ms · 20 turns`, the active classifier), each only when there is data. Each session keeps its own record, merged when the pane reads it.
22
+ - The Classifier tab shows each classifier's last checked-in activity probe under its row, such as `activity probe 70/70 · p95 299 ms · Oct 10`, when the configured model is the one that was probed.
23
+ - `experiments/mod-router/results/activity-live-acceptance.json`: four live Claude Code 2.1.296 sessions in agterm across Haiku, Sonnet and Opus with tool calls, `shadow` and `on`, and Opus thinking blocks intact after a Haiku detour, with no API errors.
24
+
25
+ ### Changed
26
+
27
+ - Activity moves refused by an escalation hold or the credits cap now count as refused, in the session's **Switches** line (`· 2 refused`) and in **Activity moves** across sessions. Before, only moves that did not pay back were counted.
28
+ - A probe run without a single usable activity answer no longer replaces the checked-in result.
29
+
30
+ ### Upgrade notes
31
+
32
+ - 1.7.0 keeps reading `activity:stats:v1` after a downgrade. It does not read the new `activity:metrics:v1:*` records, so the Escalations and Latency lines are not shown on 1.7.0.
33
+
3
34
  ## 1.7.0 (2026-10-10)
4
35
 
5
36
  Activity routing is now on by default, with a new default matrix that uses Opus, Sonnet and Haiku. The Usage tab shows what routing cost against running every reply on your own model. The band has one routing on/off control.
@@ -108,14 +108,14 @@ When a turn completes, the Mod labels it from the tool calls the assistant made
108
108
 
109
109
  Two stores hold counts:
110
110
 
111
- - **Session**: `activityStats` in the view, reset when the session ends. It has turns, requests and tokens per activity, switches by tier and by activity, tool agreement, and the shadow readout. The label counted per activity is the one the turn used: the applied activity in `on`, the answer that cleared `activityMass` in `shadow`. Tool agreement always compares the answer that cleared `activityMass`, so it measures the classifier in both modes. The view also keeps an `activities` array aligned with the last 30 replies.
112
- - **Across sessions**: one `$.store` key, `activity:stats:v1`. It holds a confusion matrix of the classifier's raw answer (including `uncertain` and `none`) against the observed bucket, run lengths per activity in five buckets, lateral switches taken and refused, and shadow turns, differences, and the summed next-request estimate of the differing turns with the count it covers. The key set is fixed, so the size is bounded. A 1.6.0 value without the estimate loads with zeros; a stored value with any other shape is read as empty. The view keeps the last value read or written as `activityStore`: read at session start and when the pane opens, replaced after each turn's write and after a reset. The open run is written when the session ends. **Reset stats** clears both stores, and the routing-vs-your-model totals below.
111
+ - **Session**: `activityStats` in the view, reset when the session ends. It has turns, requests, tokens, the list price of priced replies with their count, and routed replies by `model@effort` per activity, switches by tier and by activity, lateral moves taken and refused outside `shadow`, tool agreement, and the shadow readout. The label counted per activity is the one the turn used: the applied activity in `on`, the answer that cleared `activityMass` in `shadow`. Tool agreement always compares the answer that cleared `activityMass`, so it measures the classifier in both modes. The view also keeps an `activities` array aligned with the last 30 replies.
112
+ - **Across sessions**: one `$.store` key and one record per session. `activity:stats:v1` holds a confusion matrix of the classifier's raw answer (including `uncertain` and `none`) against the observed bucket, run lengths per activity in five buckets, lateral switches taken and refused (from the decision's `lateral` field, so a hold or the cash gate counts), and shadow turns, differences, and the summed next-request estimate of the differing turns with the count it covers. Its shape is 1.7.0's: a value with any other key or shape reads as empty, and a 1.6.0 value without the estimate loads with zeros. Each session's `activity:metrics:v1:<session id>` holds a classifier latency histogram per classifier id (22 buckets, at most 8 ids), the cheaper activity moves taken in `on` with the escalations that followed one, and `since`, its first reading; each part reads as empty on its own when its shape is wrong, as after a change of buckets. Keeping them apart from the stats key lets 1.7.0, which never reads them, keep the counts across a downgrade. The metrics follow the savings records below: one writer per key, so sessions that finish turns at the same moment do not overwrite each other; this session's record held in memory and written after each turn that adds a reading; the readout the sum of all records, latency kept to the first 8 classifier ids; and the same reset marker. The stats key has a fixed key set; each metrics record is bounded, and the records grow by one per session that recorded a reading until **Reset stats**, as the savings records do. The view keeps the last values read or written as `activityStore` and `activityMetrics` (the sum): read at session start and when the pane opens, replaced after each turn's write and after a reset. The open run is written when the session ends. **Reset stats** clears the session counts, the stats key and every metrics record, and the routing-vs-your-model totals below.
113
113
 
114
- Nothing is recorded with `activityRouting: off`. Stats run inside a guard: a failed transcript read or store access loses that turn's counts and never fails the turn. A turn from a session that has since changed is dropped.
114
+ Nothing is recorded with `activityRouting: off`. Stats run inside a guard: a failed transcript read or store access loses that turn's counts and never fails the turn. Readings already in the session's metrics record stay in memory after a failed write and go out with the next turn that adds a reading. A turn from a session that has since changed is dropped.
115
115
 
116
116
  ### Routing vs your model
117
117
 
118
- After each main reply, where the response is observed, `savings.mjs` prices the reply on the model that served it and the same tokens on "your model": the session's model and the effort Claude Code sent on the turn's first request. Your model keeps a simulated single cache in the loop as `yours` (`{ total, at, onYours, model, effort }`), which a history reset clears; after `/model` or an effort change the next reply finds it cold, since it was kept for another model or effort. The session's totals are `savings` in the view; at turn completion the turn's totals are added to the session's own `$.store` key `savings:v1:<session id>` (replies, five sums and `since`), under the same reset guard as the activity stats. Each key has one writer, and Claude Code 2.1.296 writes one key under a file lock (observed in its code; the `$.store` types do not promise it), so sessions that finish turns at the same moment do not overwrite each other. The view's `savingsStore` adds up two parts: every other session's key, read at session start and when the pane opens, and this session's record, which the router reads once, holds in memory, adds each finished turn to and then writes, so a pane opened during that write counts the turn once. **Reset stats** first writes the time in `savings:reset:v1`, then deletes every such key. A record whose first reply is older than that time reads as empty, and each session reads the time before it adds a turn, so a session that still holds its record from before a Reset in another session drops it instead of writing it back. The comparison is by milliseconds: a record whose first reply has the reset's millisecond, or one written while the clock had moved back, still counts. A stored value of any other shape reads as empty. A reply with routing off or after an engine fallback is not counted, but it still moves your model's cache. A reply whose cache lifetime cannot be read loses its prices only, and your model's cache still follows it; a usage without rate limits prices writes at five minutes. No failure here fails the turn. [Evaluation](evaluation.md#routing-vs-your-model) gives the method and its limits.
118
+ After each main reply, where the response is observed, `savings.mjs` prices the reply on the model that served it and the same tokens on "your model": the session's model and the effort Claude Code sent on the turn's first request. Your model keeps a simulated single cache in the loop as `yours` (`{ total, at, onYours, model, effort }`), which a history reset clears; after `/model` or an effort change the next reply finds it cold, since it was kept for another model or effort. The session's totals are `savings` in the view; at turn completion the turn's totals are added to the session's own `$.store` key `savings:v1:<session id>` (replies, five sums and `since`), under the same reset guard as the activity stats. Each key has one writer, and Claude Code 2.1.296 writes one key under a file lock (observed in its code; the `$.store` types do not promise it), so sessions that finish turns at the same moment do not overwrite each other. The view's `savingsStore` adds up two parts: every other session's key, read at session start and when the pane opens, and this session's record, which the router reads once, holds in memory, adds each finished turn to and then writes, so a pane opened during that write counts the turn once. **Reset stats** first writes the time in `savings:reset:v1`, then deletes every such key and every activity metrics record. A record whose first reply is older than that time reads as empty, and each session reads the time before it adds a turn, so a session that still holds its record from before a Reset in another session drops it instead of writing it back. The comparison is by milliseconds: a record whose first reply has the reset's millisecond, or one written while the clock had moved back, still counts. A stored value of any other shape reads as empty. A reply with routing off or after an engine fallback is not counted, but it still moves your model's cache. A reply whose cache lifetime cannot be read loses its prices only, and your model's cache still follows it; a usage without rate limits prices writes at five minutes. No failure here fails the turn. [Evaluation](evaluation.md#routing-vs-your-model) gives the method and its limits.
119
119
 
120
120
  Subagent events pass directly to Claude Code. They do not create classifier requests or change main-conversation metrics. This is the selected scope because Mods do not expose whether a subagent model was explicit or inherited.
121
121
 
@@ -56,7 +56,7 @@ The script writes aggregate counts, reason histograms, route transitions, and sh
56
56
 
57
57
  Activity routing has no measured result. The replay above predates it and covers tiers only. It is `on` by default since 1.7, with defaults chosen from Anthropic's published benchmarks and a list-price replay of one developer's sessions; [Activity routing](activity-routing.md#what-the-estimates-mean) gives that replay and its limits. The pane counts turns, requests, switches, and the turns `on` would route differently, with a configured list-price estimate of those turns' next requests, not a measured saving. Two checks look at the classifier's activity label.
58
58
 
59
- The figures plan §11 tracks (tool agreement of at least 80% on `code` against `ops` and `explore`, a negative would-route estimate in `shadow`, the number of activity switches) are read from the Usage tab's **ACROSS SESSIONS** block, which shows the counts kept since the last reset: **Code/ops/explore** for the tool agreement on turns the classifier answered `code`, `ops` or `explore`, **Agreement** and **Mismatch** for all of them, **Shadow** for the would-route count and estimate, **Activity moves** for activity switches inside a tier. `/router` without a UI surface prints both agreements. The underlying counts are the `activity:stats:v1` store described in [Architecture](architecture.md).
59
+ The figures plan §11 tracks (tool agreement of at least 80% on `code` against `ops` and `explore`, Jev p95 latency of at most 450 ms, at most 1 tool-error escalation in 10 cheaper activity moves, a negative would-route estimate in `shadow`, the number of activity switches) are read from the Usage tab's **ACROSS SESSIONS** block, which shows the counts kept since the last reset: **Code/ops/explore** for the tool agreement on turns the classifier answered `code`, `ops` or `explore`, **Agreement** and **Mismatch** for all of them, **Activity moves** for moves inside a tier taken and refused (a refusal by a hold or the credits cap included), **Escalations** for the tool-error escalations that followed a cheaper activity move in `on` while the turns stayed on its route in `on` (turning routing off, a turn with routing off, or a turn in `shadow` or `off` ends the watch, a pin in between does not, and an escalation the credits cap held counts), **Latency** for the active classifier's p95 wait, **Shadow** for the would-route count and estimate. Latency is kept in fixed buckets, so the p95 reads as a bound (`≤ 350 ms`); it is the wait for the classifier's answer, transient retries included, without the routing after it, so it compares with the probe's p95 below, and is not recorded with activity routing `off`. `/router` without a UI surface prints both agreements. The underlying counts are the `activity:stats:v1` store and the per-session `activity:metrics:v1:<session id>` records described in [Architecture](architecture.md).
60
60
 
61
61
  ### Tool agreement
62
62
 
@@ -67,10 +67,10 @@ Each finished turn is labelled from the tools the assistant used, locally and at
67
67
  | `code` | An edit tool (`Edit`, `Write`, `MultiEdit`, `NotebookEdit`) on a path that is not a doc path |
68
68
  | `docs` | Edits only to paths ending `.md`, `.mdx`, `.rst`, or `.txt`, or under a `docs/` directory |
69
69
  | `ops` | A Bash command that is not read-only |
70
- | `read` | Other tool use: reads, searches, web fetches, and Bash commands that only read, such as `ls`, `grep`, `cat`, or `git status` and `git diff` |
70
+ | `read` | Other tool use: reads, searches, web fetches, Task, TodoWrite, MCP tools, and Bash commands that only read, such as `ls`, `grep`, `cat`, or `git status` and `git diff` |
71
71
  | `talk` | No tool calls |
72
72
 
73
- A Bash command counts as read-only only when every command in the pipeline or list is on a short list of readers. Redirects, command substitution, and `find` with `-exec` or `-delete` are not read-only, and neither is any command the list does not name.
73
+ A subagent's edits and commands never reach the main transcript, so a turn that delegates its coding to a Task reads as `read`: agreement on `code` is understated for delegated work. A Bash command counts as read-only only when every command in the pipeline or list is on a short list of readers. Redirects, command substitution, and `find` with `-exec` or `-delete` are not read-only, and neither is any command the list does not name.
74
74
 
75
75
  The classifier's activity agrees with the turn when the bucket is one it allows:
76
76
 
@@ -96,7 +96,7 @@ node scripts/probe-activity.mjs [classifier]
96
96
 
97
97
  The script sends each prompt to one classifier as the router would ask it, with activity routing forced to `shadow`. It reads credentials as `probe-classifier.mjs` does and needs a classifier you can reach. The run is billed on a paid classifier, and it is run by hand, not in CI. It writes accuracy, an expected-by-answered confusion matrix, p50 and p95 latency against the classifier's deadline, and error counts to `experiments/mod-router/results/activity-probe-<classifier>.json`. For Ollama the latency covers both requests. It never prints a key.
98
98
 
99
- Five probe results are checked in, all run on 2026-10-10 with the synthetic set above: Jev against the live typesafe.ai API, OpenAI against the live Decisions API, Clef and Clef Flash against the live Cloudflare API, Ollama with `qwen3.5:9b` on an Apple M2 Pro laptop:
99
+ Five probe results are checked in, run on 2026-10-09 and 2026-10-10 (UTC) with the synthetic set above: Jev against the live typesafe.ai API, OpenAI against the live Decisions API, Clef and Clef Flash against the live Cloudflare API, Ollama with `qwen3.5:9b` on an Apple M2 Pro laptop:
100
100
 
101
101
  | Classifier | Correct | p50 latency | p95 latency | Result file |
102
102
  | ---------- | -------: | ----------: | ----------: | ---------------------------------------------------------------------------------------------------- |
@@ -106,7 +106,7 @@ Five probe results are checked in, all run on 2026-10-10 with the synthetic set
106
106
  | Clef Flash | 69 of 70 | 470 ms | 874 ms | [`activity-probe-clef-flash.json`](../experiments/mod-router/results/activity-probe-clef-flash.json) |
107
107
  | Ollama | 66 of 70 | 2986 ms | 3483 ms | [`activity-probe-ollama.json`](../experiments/mod-router/results/activity-probe-ollama.json) |
108
108
 
109
- Ollama's latency covers both requests and stays inside its 5000 ms deadline; on slower hardware the activity step is skipped when less than 300 ms remain. They measure the label on synthetic prompts, not routing quality or savings. The activity question costs little time: 30 alternating pairs of the same prompts, with and without it, gave Jev p50 259 ms both ways and p95 300 ms without, 306 ms with; OpenAI p50 163 ms and 162 ms, p95 296 ms and 349 ms. OpenAI's first request of a run took about 2.3 s of its 3000 ms deadline. The pane shows no probe accuracy. `policy.activityMass` was not tuned on any classifier; run the probe before you trust it with a new one.
109
+ Ollama's latency covers both requests and stays inside its 5000 ms deadline; on slower hardware the activity step is skipped when less than 300 ms remain. They measure the label on synthetic prompts, not routing quality or savings. The activity question costs little time: 30 alternating pairs of the same prompts, with and without it, gave Jev p50 259 ms both ways and p95 300 ms without, 306 ms with; OpenAI p50 163 ms and 162 ms, p95 296 ms and 349 ms. OpenAI's first request of a run took about 2.3 s of its 3000 ms deadline. The Classifier tab shows each result under its classifier, such as `activity probe 70/70 · p95 299 ms · Oct 10` (UTC date), when the classifier runs the probed model. The script regenerates `lib/probe-results.mjs` from the result files after each run; `npm run probe:results` does it alone, and a test fails when the two disagree. A run without a single activity answer (an activity or `uncertain`) or without a p95 latency, such as one with a revoked key or a model whose answers all lack the activity, exits with an error and writes neither file. `policy.activityMass` was not tuned on any classifier; run the probe before you trust it with a new one.
110
110
 
111
111
  ## Routing vs your model
112
112
 
@@ -39,8 +39,8 @@ The status band is one line above the prompt: a tier meter, the activity, the ro
39
39
  | Route (activity) | With `on`: the route for the tier and activity, marked `(override)` with the base route beside it, or `(base)`. With `shadow`: what `on` would use, and the route in use (`shadow; same route` when they agree). |
40
40
  | Replies | The tier that served each of the last 30 main-conversation replies; a dot where the router did not choose. A second row of letters names each reply's activity: `c` code, `d` debug, `e` explore, `p` plan, `r` review, `o` ops, `w` docs. |
41
41
  | Routing vs your model (Usage) | Routed replies at configured list prices against the same tokens on your model (the session's model and effort), this session and since the last reset (each session against its own model); the difference split into cheaper models, stronger than yours, and cache writes from switches. `too early` below 10 replies. Output length is held the same; answer quality is not measured. |
42
- | Activity stats (Usage) | Counts for this session: turns, requests and share per activity; switches by tier and by activity; tool agreement; in `shadow`, turns `on` would route differently with their summed next-request estimate at configured list prices. |
43
- | Across sessions (Usage) | The counts kept since the last reset: labelled turns, tool agreement on the classifier's raw answer, overall and for `code`, `ops` and `explore`, the two largest mismatches, activity moves inside a tier taken and refused, and the shadow readout. `/router` text shows the agreement. |
42
+ | Activity stats (Usage) | Counts for this session: turns, requests and share per activity, its `est. cost` at configured list prices (`—` unless every reply had a price) and the route it was `mostly on`; switches by tier and by activity, and activity moves refused in `on`; tool agreement; in `shadow`, turns `on` would route differently with their summed next-request estimate at configured list prices. |
43
+ | Across sessions (Usage) | The counts kept since the last reset: labelled turns, tool agreement on the classifier's raw answer, overall and for `code`, `ops` and `explore`, the two largest mismatches, activity moves inside a tier taken and refused, escalations after a cheaper activity move, the active classifier's p95 latency, and the shadow readout. `/router` text shows the agreement. |
44
44
  | Cost | Claude Code's native usage value. A plan's list-price estimates are not subscription cash charges. |
45
45
  | Context bar | Maximum available observed and local estimated context, compared with the routed model's window. The fit rule reserves 20%. |
46
46
  | Observed / estimated input | Claude's local context reading and the router's best current context estimate. `unknown` blocks a move to a smaller window. |
@@ -56,7 +56,7 @@ The router keeps no bill. Routing vs your model is a configured-price estimate,
56
56
 
57
57
  ## Routing and classifier settings
58
58
 
59
- The Routing tab edits each tier's model alias and effort, the baseline tier, the activity overrides and mode, and four policy controls as one draft. The draft does not affect the active turn, survives a tab change or closing the pane, and is dropped at a new session. **Save** validates the whole file with the same loader the Mod starts with, writes `router.json` for future turns, and keeps unrelated keys. A route, activity override, or policy value equal to the built-in default is removed from the file, so later default changes still reach it. The Classifier tab writes the classifier choice and its deadline at once. **Undo** reverts the settings the last pane write changed, from any tab, until the next write or a new session.
59
+ The Routing tab edits each tier's model alias and effort, the baseline tier, the activity overrides and mode, and four policy controls as one draft. The draft does not affect the active turn, survives a tab change or closing the pane, and is dropped at a new session. **Save** validates the whole file with the same loader the Mod starts with, writes `router.json` for future turns, and keeps unrelated keys. A route, activity override, or policy value equal to the built-in default is removed from the file, so later default changes still reach it. The Classifier tab writes the classifier choice and its deadline at once, and shows each classifier's last checked-in activity probe on the model it ran, from `lib/probe-results.mjs`. **Undo** reverts the settings the last pane write changed, from any tab, until the next write or a new session.
60
60
 
61
61
  | Control | Values in the pane | Effect |
62
62
  | ------------------ | ---------------------------------- | ---------------------------------------------------------------------------------------- |
@@ -66,7 +66,7 @@ The Routing tab edits each tier's model alias and effort, the baseline tier, the
66
66
  | Credits cap | $0.50, $1, $2, $5 | Largest estimated cold write on a `credits` model. |
67
67
  | Activity threshold | 50%, 60%, 70%, 80% | `policy.activityMass`: the least probability at which an activity applies. |
68
68
 
69
- Under the routes, **ACTIVITIES** starts with the **Activity routing** selector, `off · shadow · on`, and a hint for the mode. Below it is the effective matrix of the draft: one row per distinct set of cells, `·` where a cell uses the base route, routes abbreviated such as `S·high` for Sonnet · high, and the count of distinct routes, each counted as its own cache. Until the mode is `on` the count carries `once activity routing is on`. **OVERRIDES** reads `in use`, or in yellow `not in use:` with the reason: activity routing is `shadow` or `off`, routing is off, or routing is unavailable. Each override is a row with a model, an effort, and **remove**, and a selector adds one for a free activity and tier. A removed built-in override is saved as the tier's own route, because a file that only omitted it would get the built-in back. Yellow notes flag an override equal to its base route, one stronger than the tier above, and the only cell or tier on its (model, effort) pair, which adds a cache. **Reset routes to defaults** resets the routes only; pending override and mode edits stay in the draft. The Usage tab's **Reset stats** clears the activity counts and routing vs your model, for this session and across sessions.
69
+ Under the routes, **ACTIVITIES** starts with the **Activity routing** selector, `off · shadow · on`, and a hint for the mode. Below it is the effective matrix of the draft: one row per distinct set of cells, `·` where a cell uses the base route, routes abbreviated such as `S·high` for Sonnet · high, and the count of distinct routes, each counted as its own cache. Until the mode is `on` the count carries `once activity routing is on`. **OVERRIDES** reads `in use`, or in yellow `not in use:` with the reason: activity routing is `shadow` or `off`, routing is off, or routing is unavailable. Each override is a row with a model, an effort, and **remove**, and a selector adds one for a free activity and tier. A removed built-in override is saved as the tier's own route, because a file that only omitted it would get the built-in back. Yellow notes flag an override equal to its base route, one stronger than the tier above, and the only cell or tier on its (model, effort) pair, which adds a cache. **Reset routes to defaults** resets the routes only; pending override and mode edits stay in the draft. The Usage tab's **ACTIVITY** table adds each activity's `est. cost`, its replies at list prices, and the route it was `mostly on`; **ACROSS SESSIONS** adds the escalations after a cheaper activity move and the active classifier's p95 latency. The Usage tab's **Reset stats** clears the activity counts and routing vs your model, for this session and across sessions.
70
70
 
71
71
  The classifier section lists every configured classifier as a row with its service host and whether its key and endpoint settings are complete. Selecting a row writes `classifier` to `router.json` at once, or removes it for the default, through the same validation, and resets the new classifier's failure count. A turn already being classified finishes with its own classifier; its readings stay off the pane. **Undo** returns to the previous classifier. The deadline saves at once too, and a built-in classifier's default deadline is written as no override. **Sends** names the host that receives prompt text.
72
72
 
@@ -1,12 +1,12 @@
1
- <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 899 401" role="img" aria-labelledby="title desc">
1
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 899 496" role="img" aria-labelledby="title desc">
2
2
  <title id="title">Router pane, Classifier tab</title>
3
- <desc id="desc">The classifier rows with the active one marked, a missing Jev API key with Set up, the deadline, health, the receiving host and the credentials each classifier needs, and Undo after a switch.</desc>
3
+ <desc id="desc">The classifier rows with the active one marked and each one’s last activity probe, a missing Jev API key with Set up, the deadline, health, the receiving host and the credentials each classifier needs, and Undo after a switch.</desc>
4
4
  <style>
5
5
  text { font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, 'Liberation Mono', monospace; font-size: 13px; }
6
6
  .cap { font-family: system-ui, -apple-system, 'Segoe UI', sans-serif; font-size: 12px; fill: #9a988f; }
7
7
  </style>
8
- <rect width="899" height="401" rx="8" fill="#1c1c1c"/>
9
- <rect x="10" y="10" width="879" height="381" rx="6" fill="none" stroke="#3f5f8f"/>
8
+ <rect width="899" height="496" rx="8" fill="#1c1c1c"/>
9
+ <rect x="10" y="10" width="879" height="476" rx="6" fill="none" stroke="#3f5f8f"/>
10
10
  <text x="24.0" y="34" fill="#5fafd7" font-weight="700" xml:space="preserve">ROUTING</text>
11
11
  <rect x="94.2" y="20" width="62.4" height="18" rx="4" fill="#87afff"/>
12
12
  <text x="94.2" y="34" fill="#0d1016" font-weight="700" xml:space="preserve">[ ◉ on ]</text>
@@ -33,37 +33,42 @@
33
33
  <text x="320.4" y="148" fill="#d7af5f" xml:space="preserve">○ no API key </text>
34
34
  <rect x="429.6" y="134" width="78.0" height="18" rx="4" fill="#2e3340"/>
35
35
  <text x="429.6" y="148" fill="#d0d0d0" xml:space="preserve">[ Set up ]</text>
36
- <rect x="24.0" y="153" width="132.6" height="18" rx="4" fill="#2e3340"/>
37
- <text x="24.0" y="167" fill="#d0d0d0" xml:space="preserve">[ ○ Clef ]</text>
38
- <text x="164.4" y="167" fill="#7c7c7c" xml:space="preserve">api.cloudflare.com</text>
39
- <text x="320.4" y="167" fill="#87d787" xml:space="preserve">● ready</text>
40
- <rect x="24.0" y="172" width="132.6" height="18" rx="4" fill="#87afff"/>
41
- <text x="24.0" y="186" fill="#0d1016" font-weight="700" xml:space="preserve">[ ◉ Clef Flash ]</text>
36
+ <text x="24.0" y="167" fill="#7c7c7c" xml:space="preserve"> activity probe 70/70 · p95 299 ms · Oct 10</text>
37
+ <rect x="24.0" y="172" width="132.6" height="18" rx="4" fill="#2e3340"/>
38
+ <text x="24.0" y="186" fill="#d0d0d0" xml:space="preserve">[ ○ Clef ]</text>
42
39
  <text x="164.4" y="186" fill="#7c7c7c" xml:space="preserve">api.cloudflare.com</text>
43
40
  <text x="320.4" y="186" fill="#87d787" xml:space="preserve">● ready</text>
44
- <rect x="24.0" y="191" width="132.6" height="18" rx="4" fill="#2e3340"/>
45
- <text x="24.0" y="205" fill="#d0d0d0" xml:space="preserve">[ ○ OpenAI ]</text>
46
- <text x="164.4" y="205" fill="#7c7c7c" xml:space="preserve">api.openai.com</text>
47
- <text x="320.4" y="205" fill="#d7af5f" xml:space="preserve">○ no API key </text>
48
- <rect x="429.6" y="191" width="78.0" height="18" rx="4" fill="#2e3340"/>
49
- <text x="429.6" y="205" fill="#d0d0d0" xml:space="preserve">[ Set up ]</text>
50
- <rect x="24.0" y="210" width="132.6" height="18" rx="4" fill="#2e3340"/>
51
- <text x="24.0" y="224" fill="#d0d0d0" xml:space="preserve">[ ○ Ollama ]</text>
52
- <text x="164.4" y="224" fill="#7c7c7c" xml:space="preserve">127.0.0.1:11434</text>
41
+ <text x="24.0" y="205" fill="#7c7c7c" xml:space="preserve"> activity probe 69/70 · p95 883 ms · Oct 9</text>
42
+ <rect x="24.0" y="210" width="132.6" height="18" rx="4" fill="#87afff"/>
43
+ <text x="24.0" y="224" fill="#0d1016" font-weight="700" xml:space="preserve">[ ◉ Clef Flash ]</text>
44
+ <text x="164.4" y="224" fill="#7c7c7c" xml:space="preserve">api.cloudflare.com</text>
53
45
  <text x="320.4" y="224" fill="#87d787" xml:space="preserve">● ready</text>
54
- <text x="24.0" y="243" fill="#d0d0d0" xml:space="preserve">Deadline</text>
55
- <text x="94.2" y="243" fill="#d0d0d0" xml:space="preserve">3000 ms ▾</text>
56
- <text x="187.8" y="243" fill="#7c7c7c" xml:space="preserve">then keep the model</text>
57
- <text x="24.0" y="262" fill="#d0d0d0" xml:space="preserve">Health active · 0 recent failures</text>
58
- <text x="24.0" y="281" fill="#7c7c7c" xml:space="preserve">Sends prompt + 6 recent turns → api.cloudflare.com</text>
59
- <text x="24.0" y="300" fill="#d0d0d0" xml:space="preserve">Credentials </text>
60
- <rect x="125.4" y="286" width="62.4" height="18" rx="4" fill="#2e3340"/>
61
- <text x="125.4" y="300" fill="#d0d0d0" xml:space="preserve">[ Edit ]</text>
62
- <text x="187.8" y="300" fill="#7c7c7c" xml:space="preserve"> opens /plugin configure</text>
63
- <text x="24.0" y="319" fill="#7c7c7c" xml:space="preserve"> Jev: API key · Clef, Clef Flash: API token, account ID · OpenAI: API key · Ollama: no key needed</text>
64
- <text x="24.0" y="357" fill="#87d787" xml:space="preserve">Saved: classifier Clef → Clef Flash. Applies from the next turn. </text>
65
- <rect x="538.8" y="343" width="62.4" height="18" rx="4" fill="#2e3340"/>
66
- <text x="538.8" y="357" fill="#d0d0d0" xml:space="preserve">[ Undo ]</text>
67
- <rect x="24.0" y="362" width="70.2" height="18" rx="4" fill="#2e3340"/>
68
- <text x="24.0" y="376" fill="#d0d0d0" xml:space="preserve">[ Close ]</text>
46
+ <text x="24.0" y="243" fill="#7c7c7c" xml:space="preserve"> activity probe 69/70 · p95 874 ms · Oct 9</text>
47
+ <rect x="24.0" y="248" width="132.6" height="18" rx="4" fill="#2e3340"/>
48
+ <text x="24.0" y="262" fill="#d0d0d0" xml:space="preserve">[ ○ OpenAI ]</text>
49
+ <text x="164.4" y="262" fill="#7c7c7c" xml:space="preserve">api.openai.com</text>
50
+ <text x="320.4" y="262" fill="#d7af5f" xml:space="preserve">○ no API key </text>
51
+ <rect x="429.6" y="248" width="78.0" height="18" rx="4" fill="#2e3340"/>
52
+ <text x="429.6" y="262" fill="#d0d0d0" xml:space="preserve">[ Set up ]</text>
53
+ <text x="24.0" y="281" fill="#7c7c7c" xml:space="preserve"> activity probe 70/70 · p95 286 ms · Oct 10</text>
54
+ <rect x="24.0" y="286" width="132.6" height="18" rx="4" fill="#2e3340"/>
55
+ <text x="24.0" y="300" fill="#d0d0d0" xml:space="preserve">[ ○ Ollama ]</text>
56
+ <text x="164.4" y="300" fill="#7c7c7c" xml:space="preserve">127.0.0.1:11434</text>
57
+ <text x="320.4" y="300" fill="#87d787" xml:space="preserve">● ready</text>
58
+ <text x="24.0" y="319" fill="#7c7c7c" xml:space="preserve"> activity probe 66/70 · p95 3483 ms · Oct 9</text>
59
+ <text x="24.0" y="338" fill="#d0d0d0" xml:space="preserve">Deadline</text>
60
+ <text x="94.2" y="338" fill="#d0d0d0" xml:space="preserve">3000 ms ▾</text>
61
+ <text x="187.8" y="338" fill="#7c7c7c" xml:space="preserve">then keep the model</text>
62
+ <text x="24.0" y="357" fill="#d0d0d0" xml:space="preserve">Health active · 0 recent failures</text>
63
+ <text x="24.0" y="376" fill="#7c7c7c" xml:space="preserve">Sends prompt + 6 recent turns → api.cloudflare.com</text>
64
+ <text x="24.0" y="395" fill="#d0d0d0" xml:space="preserve">Credentials </text>
65
+ <rect x="125.4" y="381" width="62.4" height="18" rx="4" fill="#2e3340"/>
66
+ <text x="125.4" y="395" fill="#d0d0d0" xml:space="preserve">[ Edit ]</text>
67
+ <text x="187.8" y="395" fill="#7c7c7c" xml:space="preserve"> opens /plugin configure</text>
68
+ <text x="24.0" y="414" fill="#7c7c7c" xml:space="preserve"> Jev: API key · Clef, Clef Flash: API token, account ID · OpenAI: API key · Ollama: no key needed</text>
69
+ <text x="24.0" y="452" fill="#87d787" xml:space="preserve">Saved: classifier Clef → Clef Flash. Applies from the next turn. </text>
70
+ <rect x="538.8" y="438" width="62.4" height="18" rx="4" fill="#2e3340"/>
71
+ <text x="538.8" y="452" fill="#d0d0d0" xml:space="preserve">[ Undo ]</text>
72
+ <rect x="24.0" y="457" width="70.2" height="18" rx="4" fill="#2e3340"/>
73
+ <text x="24.0" y="471" fill="#d0d0d0" xml:space="preserve">[ Close ]</text>
69
74
  </svg>
@@ -1,12 +1,12 @@
1
- <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 758 971" role="img" aria-labelledby="title desc">
1
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 657 1104" role="img" aria-labelledby="title desc">
2
2
  <title id="title">Router pane, Usage tab</title>
3
- <desc id="desc">Routing vs your model first: routed replies against the same tokens on your model, Opus 5.5 at xhigh this session and each session’s own model since the last reset, the difference split into cheaper models, stronger models and switch cache writes, the two costs as bars and the replies by tier; then Claude’s own readings, activity counts and estimates.</desc>
3
+ <desc id="desc">Routing vs your model first: routed replies against the same tokens on your model, Opus 5.5 at xhigh this session and each session’s own model since the last reset, the difference split into cheaper models, stronger models and switch cache writes, the two costs as bars and the replies by tier; then Claude’s own readings, activity counts with their estimated cost and route, the counts across sessions, and estimates.</desc>
4
4
  <style>
5
5
  text { font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, 'Liberation Mono', monospace; font-size: 13px; }
6
6
  .cap { font-family: system-ui, -apple-system, 'Segoe UI', sans-serif; font-size: 12px; fill: #9a988f; }
7
7
  </style>
8
- <rect width="758" height="971" rx="8" fill="#1c1c1c"/>
9
- <rect x="10" y="10" width="738" height="951" rx="6" fill="none" stroke="#3f5f8f"/>
8
+ <rect width="657" height="1104" rx="8" fill="#1c1c1c"/>
9
+ <rect x="10" y="10" width="637" height="1084" rx="6" fill="none" stroke="#3f5f8f"/>
10
10
  <text x="24.0" y="34" fill="#5fafd7" font-weight="700" xml:space="preserve">ROUTING</text>
11
11
  <rect x="94.2" y="20" width="62.4" height="18" rx="4" fill="#87afff"/>
12
12
  <text x="94.2" y="34" fill="#0d1016" font-weight="700" xml:space="preserve">[ ◉ on ]</text>
@@ -102,22 +102,33 @@
102
102
  <text x="24.0" y="623" fill="#87d787" xml:space="preserve">seven day ███████░░░░░░░░░ 43% used</text>
103
103
  <text x="24.0" y="661" fill="#5fafd7" font-weight="700" xml:space="preserve">ACTIVITY · this session</text>
104
104
  <text x="203.4" y="661" fill="#7c7c7c" xml:space="preserve"> activity routing on</text>
105
- <text x="24.0" y="680" fill="#7c7c7c" xml:space="preserve"> turns requests share</text>
105
+ <text x="24.0" y="680" fill="#7c7c7c" xml:space="preserve"> turns requests share est. cost mostly on</text>
106
106
  <text x="24.0" y="699" fill="#d0d0d0" xml:space="preserve"> code </text>
107
- <text x="109.8" y="699" fill="#5fafd7" xml:space="preserve">██████████</text>
108
- <text x="187.8" y="699" fill="#d0d0d0" xml:space="preserve"> 14 41 100%</text>
109
- <text x="24.0" y="718" fill="#d0d0d0" xml:space="preserve">Switches 5 · 5 by tier · 0 by activity</text>
110
- <text x="24.0" y="737" fill="#d0d0d0" xml:space="preserve">Agreement classifier vs tools: 12 of 14 turns (86%)</text>
111
- <text x="24.0" y="756" fill="#d0d0d0" xml:space="preserve">Shadow on would route 4 of 14 turns differently · est. −$0.420 … −$0.110 at list prices</text>
112
- <text x="24.0" y="794" fill="#5fafd7" font-weight="700" xml:space="preserve">ACROSS SESSIONS · since the last reset</text>
113
- <text x="24.0" y="813" fill="#7c7c7c" xml:space="preserve"> no turns recorded yet</text>
114
- <rect x="24.0" y="818" width="117.0" height="18" rx="4" fill="#2e3340"/>
115
- <text x="24.0" y="832" fill="#d0d0d0" xml:space="preserve">[ Reset stats ]</text>
116
- <text x="141.0" y="832" fill="#7c7c7c" xml:space="preserve"> clears activity stats and routing vs your model, all sessions</text>
117
- <text x="24.0" y="870" fill="#5fafd7" font-weight="700" xml:space="preserve">ESTIMATES · configured list prices</text>
118
- <text x="24.0" y="889" fill="#d0d0d0" xml:space="preserve">Cache read benefit $1.522</text>
119
- <text x="24.0" y="908" fill="#d0d0d0" xml:space="preserve">Switch tax $0.476</text>
120
- <text x="24.0" y="927" fill="#7c7c7c" xml:space="preserve">Press ? for what these estimates leave out.</text>
121
- <rect x="24.0" y="932" width="70.2" height="18" rx="4" fill="#2e3340"/>
122
- <text x="24.0" y="946" fill="#d0d0d0" xml:space="preserve">[ Close ]</text>
107
+ <text x="109.8" y="699" fill="#5fafd7" xml:space="preserve">██████░░░░</text>
108
+ <text x="187.8" y="699" fill="#d0d0d0" xml:space="preserve"> 14 41 58% $3.12 Sonnet 5.5 · high</text>
109
+ <text x="24.0" y="718" fill="#d0d0d0" xml:space="preserve"> ops </text>
110
+ <text x="109.8" y="718" fill="#5fafd7" xml:space="preserve">███░░░░░░░</text>
111
+ <text x="187.8" y="718" fill="#d0d0d0" xml:space="preserve"> 6 15 25% $0.41 Haiku 5.5 · high</text>
112
+ <text x="24.0" y="737" fill="#d0d0d0" xml:space="preserve"> explore </text>
113
+ <text x="109.8" y="737" fill="#5fafd7" xml:space="preserve">██░░░░░░░░</text>
114
+ <text x="187.8" y="737" fill="#d0d0d0" xml:space="preserve"> 4 8 17% $0.38 Haiku 5.5 · high</text>
115
+ <text x="24.0" y="756" fill="#d0d0d0" xml:space="preserve">Switches 8 · 5 by tier · 3 by activity · 1 refused</text>
116
+ <text x="24.0" y="775" fill="#d0d0d0" xml:space="preserve">Agreement classifier vs tools: 21 of 24 turns (88%)</text>
117
+ <text x="24.0" y="813" fill="#5fafd7" font-weight="700" xml:space="preserve">ACROSS SESSIONS · since the last reset</text>
118
+ <text x="24.0" y="832" fill="#d0d0d0" xml:space="preserve">Labelled 83 turns</text>
119
+ <text x="24.0" y="851" fill="#d0d0d0" xml:space="preserve">Agreement classifier vs tools: 78 of 83 turns (94%)</text>
120
+ <text x="24.0" y="870" fill="#d0d0d0" xml:space="preserve">Code/ops/explore classifier vs tools: 78 of 83 turns (94%)</text>
121
+ <text x="24.0" y="889" fill="#d0d0d0" xml:space="preserve">Mismatch code → read 4 · ops → code 1</text>
122
+ <text x="24.0" y="908" fill="#d0d0d0" xml:space="preserve">Activity moves 9 taken · 4 refused</text>
123
+ <text x="24.0" y="927" fill="#d0d0d0" xml:space="preserve">Escalations after a cheaper activity move: 0 in 7 moves</text>
124
+ <text x="24.0" y="946" fill="#d0d0d0" xml:space="preserve">Latency Jev p95 ≤ 350 ms · 81 turns</text>
125
+ <rect x="24.0" y="951" width="117.0" height="18" rx="4" fill="#2e3340"/>
126
+ <text x="24.0" y="965" fill="#d0d0d0" xml:space="preserve">[ Reset stats ]</text>
127
+ <text x="141.0" y="965" fill="#7c7c7c" xml:space="preserve"> clears activity stats and routing vs your model, all sessions</text>
128
+ <text x="24.0" y="1003" fill="#5fafd7" font-weight="700" xml:space="preserve">ESTIMATES · configured list prices</text>
129
+ <text x="24.0" y="1022" fill="#d0d0d0" xml:space="preserve">Cache read benefit $1.522</text>
130
+ <text x="24.0" y="1041" fill="#d0d0d0" xml:space="preserve">Switch tax $0.476</text>
131
+ <text x="24.0" y="1060" fill="#7c7c7c" xml:space="preserve">Press ? for what these estimates leave out.</text>
132
+ <rect x="24.0" y="1065" width="70.2" height="18" rx="4" fill="#2e3340"/>
133
+ <text x="24.0" y="1079" fill="#d0d0d0" xml:space="preserve">[ Close ]</text>
123
134
  </svg>
@@ -92,7 +92,7 @@ Every write applies from the next turn and can be undone. **Undo** (`u`) puts ba
92
92
 
93
93
  ### Compare routing with your model
94
94
 
95
- ![The Usage tab: routing vs your model, this session and since the last reset, with the difference split into its three parts, bars and replies by tier, then Claude's readings, activity counts and estimates](router-pane-usage.svg)
95
+ ![The Usage tab: routing vs your model, this session and since the last reset, with the difference split into its three parts, bars and replies by tier, then Claude's readings, activity counts with their estimated cost and route, the counts across sessions, and estimates](router-pane-usage.svg)
96
96
 
97
97
  **ROUTING VS YOUR MODEL** opens the Usage tab. Your model is the model and effort the session had before routing: the `--model` or `/model` choice and the session effort, shown in the heading. For every reply routing chose, it prices the tokens Claude reported on the model that served them, and the same tokens on your model, at the list prices in `router.json`.
98
98
 
@@ -164,11 +164,11 @@ The band's activity label appears only when the activity applies, so a weak or `
164
164
 
165
165
  ### Read the activity on the pane
166
166
 
167
- - **Now**: the **Activity** line names the classifier's choice with its share and the two next-best answers, such as `code 84% debug 8% · plan 5%`. The **Route** line shows the route the cell resolves to. **Replies** has a letter row under the strip.
167
+ - **Now**: the **Activity** line names the classifier's choice with its share and the two next-best answers, such as `code 84% debug 8% · plan 5%`. In `on`, **Why** names the cell when the activity moved the route or the route is an override, such as `code at low runs on Sonnet 5.5 · high`, or `low runs on its base route, Haiku 5.5 · high` for a move without an activity; a refused move, a hold, or a tier change keeps its reason. The **Route** line shows the route the cell resolves to. **Replies** has a letter row under the strip.
168
168
  - **Routing**: **ACTIVITIES** starts with the **Activity routing** selector (`off`, `shadow`, `on`) and what the mode does. The matrix under it shows the effective routes per tier, with `·` where a cell uses the base route, and counts the distinct routes, because the router counts each as its own cache. Routes are abbreviated, such as `S·high` for Sonnet · high, and activities with the same cells share a row. **OVERRIDES** says whether the overrides are in use, then lists each override with a model, an effort, and **remove**, plus a selector to add one. **Activity threshold** under **POLICY** is the least probability at which an activity applies. Edits join the routing draft: **Save** and **Discard** apply as for routes. Yellow notes flag an override with no effect (`same as base`), one stronger than the tier above, and the only cell or tier on its (model, effort) pair, which adds a cache.
169
- - **Usage**: **ACTIVITY · this session** lists turns, requests, and share per activity, the switches split by tier and by activity, and **Agreement**, how often the tools a turn used fit the classifier's activity (see [Evaluation](evaluation.md#tool-agreement)). In `shadow` the **Shadow** line counts how many turns `on` would have routed differently, such as `on would route 5 of 20 turns differently · est. −$0.300 … −$0.120 at list prices`. The estimate sums, over those turns, what the next request on the route `on` would use would cost against the route that ran, as a low and a high scenario at the configured list prices in `router.json`. Minus means cheaper. It covers only turns after a measured reply, and adds `for 3` when only three of the turns have one. It is not a measured saving.
169
+ - **Usage**: **ACTIVITY · this session** lists turns, requests, and share per activity, then `est. cost`, the list price of that activity's replies (the same per-reply estimate as Routing vs your model, but only for turns recorded with activity routing `shadow` or `on`, so the rows need not add up to its routed total; `—` when a reply had no price), and `mostly on`, the route most of its replies used, such as `Sonnet 5.5 · high`. A narrow pane drops `mostly on`, then `est. cost`. Below the table are the switches split by tier and by activity, with `· 2 refused` when activity moves were refused in `on` (not worth the cache write, a hold after an escalation, or the credits cap), and **Agreement**, how often the tools a turn used fit the classifier's activity (see [Evaluation](evaluation.md#tool-agreement)). In `shadow` the **Shadow** line counts how many turns `on` would have routed differently, such as `on would route 5 of 20 turns differently · est. −$0.300 … −$0.120 at list prices`. The estimate sums, over those turns, what the next request on the route `on` would use would cost against the route that ran, as a low and a high scenario at the configured list prices in `router.json`. Minus means cheaper. It covers only turns after a measured reply, and adds `for 3` when only three of the turns have one. It is not a measured saving.
170
170
 
171
- **ACROSS SESSIONS · since the last reset** shows the counts kept across sessions: the turns labelled from their tools, **Agreement** over them, **Code/ops/explore**, the same agreement over the turns the classifier answered `code`, `ops` or `explore` (omitted when there are none), the two most frequent mismatches as `predicted → observed count` (`code → read 3`), **Activity moves**, the moves to another route inside a tier, taken and refused, and the **Shadow** line summed over all shadow turns. This agreement compares the classifier's raw answer, so it can differ from the session's. The pane reads these counts when the session starts and when it opens, and after each turn. **Reset stats** clears the activity counts and routing vs your model, for this session and across sessions.
171
+ **ACROSS SESSIONS · since the last reset** shows the counts kept across sessions: the turns labelled from their tools, **Agreement** over them, **Code/ops/explore**, the same agreement over the turns the classifier answered `code`, `ops` or `explore` (omitted when there are none), the two most frequent mismatches as `predicted → observed count` (`code → read 3`), **Activity moves**, the moves to another route inside a tier, taken and refused for any reason, **Escalations**, such as `after a cheaper activity move: 1 in 12 moves`, the tool-error escalations that followed a move to a cheaper route in `on` while the turns stayed on it in `on` (turning routing off, a turn with routing off, or a turn in `shadow` or `off` ends the watch, a pin in between does not, and an escalation the credits cap held counts), **Latency**, the active classifier's p95 wait as a bucket bound such as `Jev p95 ≤ 350 ms · 20 turns` (not recorded with activity routing `off`), and the **Shadow** line summed over all shadow turns. Each of these lines appears only when it has data. This agreement compares the classifier's raw answer, so it can differ from the session's. The pane reads these counts when the session starts and when it opens, and adds this session's turns as they finish; another session's new latency and escalation readings show the next time the pane opens. **Reset stats** clears the activity counts and routing vs your model, for this session and across sessions.
172
172
 
173
173
  `/router` without a UI surface adds an `Activity: ops (81%)` line, the resolved route, and `Agreement across sessions: 19 of 22 turns (86%) · code/ops/explore 9 of 10 (90%)`. It also prints the session against your model, such as `vs your model (Opus 5.5 · xhigh): −$1.84 (−22%) this session, 64 replies, est. at list prices`.
174
174
 
@@ -221,9 +221,9 @@ Router asks one classifier per turn. Five are built in:
221
221
  | OpenAI | OpenAI, `gpt-6-luna` | OpenAI API key | 3,000 ms |
222
222
  | Ollama | Local Ollama, `qwen3.5:9b` | None | 5,000 ms |
223
223
 
224
- ![The Classifier tab: classifier rows with the active one marked, a missing Jev API key with Set up, the deadline, health, the receiving host, the credentials each classifier needs, and Undo after a switch](router-pane-classifier.svg)
224
+ ![The Classifier tab: classifier rows with the active one marked and each one's last activity probe, a missing Jev API key with Set up, the deadline, health, the receiving host, the credentials each classifier needs, and Undo after a switch](router-pane-classifier.svg)
225
225
 
226
- You can save credentials for all of them; only the active one is asked. The **Classifier** tab has one row per classifier: `◉` marks the active one, followed by its service and whether its credentials are complete. A row that lacks one names it, such as `○ no API token`, and has its own **Set up** button.
226
+ You can save credentials for all of them; only the active one is asked. The **Classifier** tab has one row per classifier: `◉` marks the active one, followed by its service and whether its credentials are complete. A row that lacks one names it, such as `○ no API token`, and has its own **Set up** button. Under a row, a dim line gives the classifier's last checked-in [activity probe](evaluation.md#activity-probe-set), such as `activity probe 70/70 · p95 299 ms · Oct 10`; it is omitted when you configured a model other than the probed one.
227
227
 
228
228
  Select a row to switch. The choice is written to `router.json` at once and applies from the next turn. **Undo** in the status bar returns to the previous classifier. A turn that is already being classified finishes with the classifier it started with. The new classifier starts with a clean failure count.
229
229