@alexeiled/pi-model-router 0.6.5 → 0.7.0

package/CHANGELOG.md

@@ -1,5 +1,32 @@
 # Changelog
 
+## [0.7.0] - 2026-09-22
+
+### Behavior change
+
+- A Choice below `confidenceThreshold` is not discarded. The router selects the lowest tier whose cumulative probability reaches `probabilityThreshold` (default 0.8). Abstention mass counts for the baseline tier.
+- Each Jev tier is a structured Choice option with `covers`, `notFor` and `examples`. The instructions are a structured object that names the state fields.
+- The router retries one transient Jev status (`408`, `429`, `5xx`) inside the existing total budget. The backoff honors `Retry-After`. Permanent statuses and cancellation are not retried.
+- The response validator accepts omitted zero-mass options and two-decimal rounding.
+- The request token estimate adds 400 tokens of headroom instead of 200.
+
+### Removed
+
+- The `low-confidence` outcome. Debug output shows `selected`, `basis`, `route-p` and `route-threshold` instead.
+- The `/router` verbs `status`, `profile`, `fix`, `disable`, `debug`, `?` and per-tier `thinking`. A removed verb prints its replacement and does nothing.
+
+### Added
+
+- `jev.probabilityThreshold`, `jev.retry.maxAttempts`, `jev.retry.backoffMs` and `classifierModel.timeoutMs`. Defaults are unchanged.
+- The `/router` verbs `off`, `log [on|off|clear]` and argument-free `widget`. Top-level completion lists verbs and profile names.
+- A skipped advisor records `bypassReason` and the footer shows it: `advice skipped: pinned high`, `over budget`, `only high eligible`, `tool turn`.
+- An `invalid-response` names the failing local check. Remote text is not retained.
+
+### Documentation
+
+- `docs/README.md` is the index. `docs/jev-advisor.md` holds the Jev guide. `docs/research/` holds dated experiment reports. `docs/archive/` holds superseded reports. File names are lowercase.
+- Parallel Noul questions were tested and not added. See `docs/research/jev-routing-policy.md`.
+
 ## [0.6.5] - 2026-09-22
 
 - The router now sends structured Jev state for the current request, recent dialogue, and optional tool evidence.

package/README.md

@@ -95,12 +95,12 @@ pi -e ./extensions/index.ts
 
 - Generation and classification use Pi's provider registry, including native/custom providers and credential-specific URLs. Only the optional Jev advisor uses separate HTTPS transport.
 - Fallbacks run only before content is emitted; cancellation does not retry. Every target must support the requested input and exact thinking level; explicit unsupported effort is not silently reduced. Omitted thinking defaults to `off` for non-reasoning targets, including fallbacks.
-- Jev gets a configurable total advisory budget via `jev.timeoutMs`: 1500 ms by default, with no additional routing cap or retry. The separate classifier-only compatibility path retains its 10-second bound and 256-token output limit. Failure or uncertainty means eligible baseline; caller cancellation stops generation.
+- Jev has one total advisory budget (`jev.timeoutMs`, 1500 ms by default) that includes its single transient retry. The classifier-only compatibility path has its own budget (`classifierModel.timeoutMs`, 10 s by default) and never retries. Failure or uncertainty means the eligible baseline; caller cancellation stops generation.
 - Valid same-turn tool continuations reuse the actual prior route before either advisor. Pins, budget policy and a single eligible primary candidate also bypass advisors. Invalid continuations choose a compatible local route without advice; incompatible Google thought-signature replay fails plainly.
 - Pi owns tool execution permissions and per-request authentication. The router checks configured provider/profile identity, not which backend login is currently behind a provider. No private authentication storage is read.
 - Context trimming preserves system instructions and whole active tool turns. It is a text estimate, not a guarantee that images or a large active turn fit.
 
-See [architecture](https://github.com/alexei-led/pi-model-router/blob/main/docs/ARCHITECTURE.md) and [release procedure](https://github.com/alexei-led/pi-model-router/blob/main/docs/RELEASING.md).
+See [docs/architecture.md](docs/architecture.md) and [docs/releasing.md](docs/releasing.md).
 
 ## Configuration
 
@@ -136,8 +136,8 @@ The extension stores the last selected profile in `~/.pi/agent/model-router-stat
 
 | Field                   | Description                                                                       |
 | ----------------------- | --------------------------------------------------------------------------------- |
-| `classifierModel`       | (Optional) Pi model used for four-tier semantic advice only when Jev is not active (disabled, not opted in or missing a key). Supports model aliases. Failure means baseline. |
-| `jev`                   | (Optional, user config only) External advisor settings; requires global enablement, a key and an explicit `profiles.<name>.jev.enabled` opt-in. Disabled by default. |
+| `classifierModel`       | (Optional) Pi model used for four-tier semantic advice only when Jev is not active (disabled, not opted in or missing a key). A string, or `{ "model", "thinking", "timeoutMs" }` (`timeoutMs` defaults to 10000). Supports model aliases. Failure means baseline. |
+| `jev`                   | (Optional, user config only) External advisor settings; requires global enablement, a key and an explicit `profiles.<name>.jev.enabled` opt-in. Disabled by default. Reference: [docs/jev-advisor.md](docs/jev-advisor.md). |
 | `ui.statusLine`        | `compact` (default) or `detailed`. Display only; project config may override it. Widget/debug always include full diagnostics. |
 | `maxSessionBudget`      | (Optional) Soft generation-cost threshold in USD. Unpinned requests prefer eligible medium-or-lower tiers and skip advisors. Not a spending cap; classifier and Jev costs are excluded. |
 | `phaseBias`, `rules`    | Deprecated and ignored, with a fixed value-free warning. Remove these fields; there is no legacy keyword mode. |
@@ -174,264 +174,69 @@ remains, and model tier never grants or restricts tool permissions.
 
 ### Optional Jev advisor: user config only
 
-Jev makes one bounded TypeSafe System One Choice request per eligible new user
-turn, without retries. It chooses only among the active profile's eligible
-primary tier/model/thinking pairs. Fallback models are not extra Jev choices.
-Pins, budget policy, a single eligible primary candidate and tool continuations
-skip Jev and the classifier.
-
-Malformed responses, `uncertain`, low confidence, timeout and HTTP errors go
-directly to the eligible baseline, without a classifier cascade. When Jev is not
-active (including a missing key), the optional Pi classifier is a separate
-compatibility path; without it, the router uses baseline directly. Jev cannot select another profile or an arbitrary model,
-provider account or thinking level. Explicit generation fallback chains may
-still cross providers, as configured by you.
-
-Configure Jev **only** in `~/.pi/agent/model-router.json` (or the agent directory
-selected by Pi). Both global enablement and an explicit user-level profile opt-in
-are required. Work profiles remain disabled unless you explicitly approve sending
-their bounded recent conversation text externally. All project-level `jev` settings, including
-profile opt-ins, are ignored with a warning, before merging user credentials.
+Jev is TypeSafe's System One model. When enabled, the router asks it once per
+new user turn to pick one of the active profile's eligible tier/model/thinking
+pairs from bounded recent conversation text. It cannot choose another profile,
+an unconfigured model or thinking level, or grant tool permissions.
 
-```json
-{
-  "jev": {
-    "enabled": true,
-    "apiKey": "<rendered by chezmoi/1Password>",
-    "endpoint": "https://api.typesafe.ai/v1/systemone",
-    "model": "jev-1.13.0",
-    "timeoutMs": 1500,
-    "confidenceThreshold": 0.65,
-    "maxStateTokens": 3000,
-    "context": {
-      "previousTurns": 2,
-      "maxHistoryTokens": 500,
-      "toolResults": "last-error",
-      "maxToolTokens": 250
-    },
-    "mode": "advisory"
-  },
-  "profiles": {
-    "personal": {
-      "jev": { "enabled": true },
-      "high": { "model": "openai/gpt-5.4-pro", "thinking": "high" },
-      "medium": { "model": "google/gemini-flash-latest", "thinking": "medium" },
-      "low": { "model": "openai/gpt-5.4-nano", "thinking": "off" },
-      "micro": { "model": "openai/gpt-5.4-nano", "thinking": "off" }
-    }
-  }
-}
-```
-
-The endpoint, model, timeout, confidence threshold, state limit and mode shown
-above are defaults. Only HTTPS endpoints without embedded credentials, query
-parameters or fragments are accepted. `timeoutMs` defaults to 1500 ms and must
-be a positive finite number within Node's timer range (at most 2147483647 ms).
-There is no product-level cap: 4000 or 5000 ms are valid if you prefer waiting
-longer before falling back. It sets the total Jev advisory budget, including
-request and response-body time; there is no separate 750 ms cap. Confidence must
-be 0–1. `maxStateTokens` is an estimated preflight budget from 1–24000. The separate classifier-only
-path keeps a 10-second bound. Neither path retries or starts generation after
-caller cancellation.
-
-If Jev frequently falls back because requests time out, try `"timeoutMs": 3000`
-in your user config. Existing explicit values such as 750 remain unchanged;
-remove the field or set it to 1500 to use the new default. Increasing the timeout
-does not lower the confidence threshold or guarantee a different route. After
-upgrading, start a new Pi session; use `/router thinking auto` to clear any
-unwanted effort override in an existing session.
-
-**External data:** Jev receives bounded text in three named JSON fields:
-`currentRequest`, `recentDialogue` and `recentToolEvidence`, plus candidate
-tier/model/thinking identifiers. The default includes up to two prior user turns
-with their last text replies, and at most the last tool result of the immediately
-previous turn **if Pi marks that result as an error**. Selection and head/tail
-truncation are deterministic, with no keyword scoring or summarizer call. System
-prompts, raw config, credentials from config, thinking blocks, tool-call arguments
-and image/binary blocks are not extracted.
-This is not a redaction service: text itself may contain secrets or private data,
-including tool output. Approve this external-data handling before enabling a
-profile, especially work. Short replies, other languages and imperfect sentences
-are advisor input, not local intent branches. Semantic classification and confidence
-are probabilistic, not a security sandbox; Pi owns tool permissions.
-
-Jev classifies the **latest user request**, using earlier messages only as
-context. Criteria describe the reasoning each tier supports, not just its name.
-The objective is **quality-first**: prefer frontier reasoning when it can materially
-improve correctness, completeness or reduce rework, even if a smaller model could
-probably complete the task. Direct retrieval and mechanical work still favor
-micro/low. This is semantic advice, not a local keyword or complexity heuristic.
-Do not increase the context limit or lower the threshold just to raise confidence.
-Confidence measures decisiveness across choices, **not** the chance that the
-selected generation model will succeed. It is distinct from the selected option's
-probability. See [Jev Choice](https://docs.typesafe.ai/primitives/choice).
-
-Concurrent calls for the same turn share one Jev request and its original deadline.
-A repeated same-turn call reuses the validated decision rather than reverting to
-baseline. Cancelling one waiter does not cancel another; the transport is aborted
-when no waiters remain. Each new user turn can choose a different backend and
-thinking level. Tool continuations keep their validated route. The logical
-`router/<profile>` stays selected throughout; this is not conversation-wide pinning.
-
-### Context selection and tuning
-
-Configure `jev.context` only in user config. Project Jev settings remain ignored;
-profile privacy opt-in is still required. Partial context settings inherit defaults.
-Invalid values or unknown context keys reject the Jev config with a value-free warning.
-
-| Setting | Default | Meaning |
-| --- | --- | --- |
-| `maxStateTokens` | `3000` | Estimated selected-state token budget, including excerpt markers. Range 1–24000. |
-| `context.previousTurns` | `2` | Previous user turns, each with its last non-empty assistant text reply. Integer 0–20; not transport-message count. |
-| `context.maxHistoryTokens` | `500` | Shared estimated-token ceiling for prior dialogue, integer 0–24000. |
-| `context.toolResults` | `"last-error"` | `"none"`, `"last"` or `"last-error"`. The latter includes the last result only when its native `isError` flag is true. |
-| `context.maxToolTokens` | `250` | Estimated-token ceiling for that one tool result, integer 0–24000. |
-
-Priority is current request → recent dialogue → tool evidence. Individual ceilings
-never expand the total estimated-token budget. The full current request wins when
-it fits; otherwise its beginning and end are kept. Prior turns also use head/tail
-excerpts when needed, with `truncated: true`. Unused space need not be filled. The
-20-turn cap also bounds JSON metadata overhead.
-
-TypeSafe publishes Jev's post-response `usage.input_tokens`, but no tokenizer or
-preflight count API. OpenAI tokenizers are not compatible substitutes: in a small
-EN/RU/code/emoji calibration, `cl100k`/`o200k` underestimated actual Jev requests by
-26–49%. The router therefore uses a documented conservative estimate: ASCII/4,
-non-ASCII UTF-8 bytes/2, then a 10% margin. The serialized request adds 200 tokens
-of measured envelope headroom. A 28000 estimated-request safety gate leaves room
-below Jev's stricter 32k `state + longest question` limit; its 64k whole-request
-limit is not binding for this single Choice question. See TypeSafe's
-[model limits](https://docs.typesafe.ai/models) and
-[long-context guidance](https://docs.typesafe.ai/model-jaggedness/jev-1.13).
-
-Only the last tool result of the immediately previous user turn is eligible, even
-when more dialogue turns are selected. `last-error` does not parse stdout for words
-such as `ERROR`, search backwards for an old failure, or resurrect a failure after
-a later successful result. A tool can report a meaningful failure as ordinary text
-with `isError: false`; choose `last` when that distinction matters. Tool arguments
-are always excluded. Empty/thinking/tool-call-only assistant messages cannot consume
-dialogue slots. Older intermediate assistant narration is not selected.
-
-Suggested overrides (merge into `jev.context`):
-
-- **Independent tasks:** `{"previousTurns": 0, "toolResults": "none"}`.
-- **Dialogue only:** `{"previousTurns": 2, "toolResults": "none"}`.
-- **Tool-heavy diagnosis:** `{"toolResults": "last", "maxToolTokens": 500}`.
-- **Longer follow-ups:** `{"previousTurns": 4, "maxHistoryTokens": 1000}`.
-
-The default is a conservative data-selection compromise from live experiments,
-not a guarantee of higher confidence. Larger windows did not consistently help;
-the most useful clear improvement was retaining a request at the end of long text.
-Some short/ambiguous follow-ups still use the configured baseline. See
-[context experiments](docs/JEV-CONTEXT-VALIDATION.md) for results and limitations.
-
-`/router status` shows effective settings. Widget/debug show estimated current/
-history/tool tokens, included turns/results, truncated blocks, the estimated full
-request and Jev's actual post-response input usage. Character counts may remain in
-persisted diagnostics for compatibility, but are not configuration budgets. No
-selected text is added to these metrics. This affects **Jev only**: generation still receives
-Pi's normal context. The separate Pi-classifier compatibility path is unchanged.
-
-### Quality-first fallback
-
-If avoiding underpowered answers matters more than extra cost/latency, set
-`"baselineTier": "high"` in an existing profile with a configured high tier:
+Enable it only in the user config, with a global `jev.enabled`, an API key and an
+explicit per-profile opt-in. Project-level `jev` settings are ignored with a
+warning. Enabling a profile is approval to send its bounded conversation text,
+including tool output, to TypeSafe; the router does not redact it.
 
 ```json
 {
-  "profiles": {
-    "personal": {
-      "baselineTier": "high"
-    }
-  }
+  "jev": { "enabled": true, "apiKey": "<rendered by chezmoi/1Password>" },
+  "profiles": { "personal": { "jev": { "enabled": true } } }
 }
 ```
 
-Merge this into the existing profile; it is not a complete standalone profile.
-Uncertain, low-confidence, failed or timed-out advice then prefers the eligible high
-route. Confident micro/low advice still wins. Pins, live capabilities, explicit
-fallback order and the soft budget still apply; high is not a forced minimum.
-Profiles without this setting retain their existing baseline policy. The extension
-never edits user configuration or privacy opt-ins automatically.
-
-### Routing diagnostics and display
+Everything else has defaults: `jev-1.13.0`, a 1500 ms total budget with one
+transient retry, a 0.65 confidence threshold, a 0.8 probability threshold and a
+3000-token context budget of the current request, two prior turns and the last
+native-error tool result. A Choice below the confidence threshold is resolved
+from its full probability distribution rather than discarded; an `uncertain`
+answer, a timeout or an error uses the local baseline directly. See
+[docs/jev-advisor.md](docs/jev-advisor.md) for the full configuration reference,
+context tuning, the acceptance policy, footer/widget diagnostics and
+troubleshooting.
+
+### Routing diagnostics
+
+The footer shows profile, tier, model/thinking and the advisor outcome, for
+example `🧭 Jev → high c91% · 807ms`. `ui.statusLine: "detailed"` adds the
+top option's probability and the request start time. When a configured
+advisor is not asked, the footer says why: `advice skipped: pinned high`,
+`over budget`, `only high eligible`, `tool turn`. `/router` and
+`/router widget` show full metrics; `/router log on` keeps the last 50
+decisions in branch-safe session state and `/router log` summarizes them.
+Field meanings, outcome codes and fixes are in
+[docs/jev-advisor.md](docs/jev-advisor.md#diagnostics).
 
-```json
-{
-  "ui": { "statusLine": "compact" }
-}
-```
+## Commands
 
-- **`compact` (default):** profile, tier, model/thinking, advisor outcome, confidence
-  and latency. Omits the repeated provider prefix to fit split panes.
-  Examples: `🧭 Jev → high c91% · 807ms`,
-  `🧭 Jev high c35% <65% → baseline · 764ms`,
-  `🧭 Jev: no tier chosen → baseline · 860ms`,
-  `🧭 Jev: timeout → baseline · 5.0s`.
-- **`detailed`:** adds selected probability and local request-start time. Example:
-  `🧭 Jev high c35% <65% → baseline · 764ms · p48% @18:34:49`.
-  Use this on wide terminals; long model/profile names can truncate a footer.
-- **Widget / status:** `/router widget on` or `/router status` shows full metrics,
-  including the Jev model label, HTTP status, candidate count, estimated context/request tokens and actual server input usage.
-- **History:** `/router debug on`, then `/router debug show`. The last 50 decisions
-  are saved in branch-safe `router-state` session entries and restored on resume.
-  Debug off stops collecting history; the latest decision still persists.
-- **Statistics:** `/router debug stats` reports unique HTTP requests, advised tiers,
-  outcome counts/rates and median latency. Statistics cover only the retained
-  decision window, **not session lifetime**. Locally generated request IDs deduplicate
-  shared requests, cached routes and tool continuations, including after resume.
-  Older decisions without IDs are excluded. `/router debug clear` clears the window.
-
-`c` is confidence; `p` is the selected option's probability. `<65%` explains a
-confidence rejection. `ms`/`s` is local request-to-validated-result time, not pure model inference
-time. `@` is the original request's local start time. `reuse` / `tool route` means
-no new Jev request: the displayed metrics belong to the original routing attempt.
-`baseline` (or `base` in older traces) means deterministic local baseline, not necessarily the medium tier.
-`no tier chosen` means Jev could not judge the required capability from the supplied
-context. It does not prove that context was missing. Compact mode omits abstention
-scores; widget/debug label them `abstention-confidence` and `abstention-p`, not
-confidence in the generation model. The acceptance threshold is not applied to abstention.
-`local baseline` / `advice bypassed` distinguishes no advisor from a rejected answer.
-
-Failures are distinguished as `low-confidence`, `uncertain`, `invalid-response`,
-`http-error`, `network-error`, `deadline`, `cancelled` or `unavailable`. A quick
-low-confidence rejection is **not a timeout**; increasing timeout will not fix it.
-Only validated choices, numeric diagnostics, recognized version labels and locally
-generated request IDs are retained. State/debug never
-retain the Jev key, endpoint, request text, raw response or remote explanations.
-Older explanations are discarded as non-rendered `legacy` metadata; Pi's own
-conversation transcript is separate from router state.
-
-For chezmoi, use a **private template**, for example
-`private_model-router.json.tmpl` under your agent-directory source path. Render
-only the `apiKey` value using a reference such as
-`{{ onepasswordRead "op://Personal/TypeSafe/apiKey" | toJson }}` (unquoted in the
-JSON template). Adapt the vault/item reference locally. Keep the rendered file
-out of Git and restrict permissions to `0600` (`chmod 600` on Unix); verify the
-mode without printing the file. Never commit rendered credentials or 1Password
-output. No environment variable is required, and the extension never executes a
-secret-lookup command. The repository example contains only a placeholder and
-keeps Jev disabled.
+One verb per concern. A verb without an argument shows the state it controls.
 
-## Commands
+| Command | Description |
+| --- | --- |
+| `/router` | Status: profile, pin, thinking override, cost, Jev settings, last decision. |
+| `/router <profile>` | Switch profile and enable the router. Same as `/model router/<profile>`. |
+| `/router off` | Leave the router and restore the previous non-router model. |
+| `/router pin <tier\|auto>` | Pin the active profile to `high`, `medium`, `low` or `micro`; `auto` clears. |
+| `/router thinking <level\|auto>` | Override thinking for every tier; `auto` clears. Pi's own thinking selector applies the same override. Per-tier levels belong in the profile config. An override that leaves no eligible route is rejected. |
+| `/router log [on\|off\|clear]` | Recent decisions and Jev statistics for the retained window; `on`/`off` control collection (last 50, saved in session state), `clear` forgets them. |
+| `/router widget` | Toggle the status widget. |
+| `/router reload` | Reload `model-router.json`. |
+| `/router help` | Usage. |
 
-| Command                     | Description                                                                     |
-| --------------------------- | ------------------------------------------------------------------------------- |
-| `/router`                   | Show detailed status, current profile, spend, and settings.                     |
-| `/router status`            | Alias for `/router` (show current status).                                      |
-| `/router profile [name]`    | Switch to a profile or list available ones (enables router if off).             |
-| `/router pin <t\|a>`        | Pin a tier (high/medium/low/micro/auto) for the active profile.                      |
-| `/router fix <tier>`        | Correct the _last_ decision and pin that tier for the current profile.          |
-| `/router thinking <level>`  | Override thinking level for all tiers (e.g. `/router thinking max`). Unsupported tiers are skipped; an override that leaves no eligible route is rejected without changing any tier. |
-| `/router thinking <tier> <level>` | Override thinking level for a specific tier (e.g. `/router thinking low off`). |
-| `/router disable`           | Disable the router and switch back to the last non-router model.                |
-| `/router widget <on\|off>`  | Toggle the persistent state widget (supports `toggle`).                         |
-| `/router debug <on\|off>`   | Toggle router debug state; use `show` or `clear` for local decision history. |
-| `/router debug stats`       | Deduplicated Jev counts, advised tiers, fallback rates and median latency for retained history. |
-| `/router reload`            | Hot-reload the configuration JSON.                                              |
-| `/router help`              | Show usage help for all subcommands.                                            |
+Removed verbs (`status`, `profile`, `fix`, `disable`, `debug`, `?`, per-tier
+`thinking`) answer with their replacement and do nothing.
 
 ## Documentation
 
-- [Architecture Guide](docs/ARCHITECTURE.md): Deep dive into the routing logic and modular design.
-- [Sample Configuration](model-router.example.json): Profile examples (`auto`, `cheap`, `deep`, `anthropic`).
+- [docs/README.md](docs/README.md): index of all project documents.
+- [docs/architecture.md](docs/architecture.md): routing flow, module boundaries, state and persistence.
+- [docs/jev-advisor.md](docs/jev-advisor.md): Jev configuration, context tuning, acceptance policy, diagnostics and troubleshooting.
+- [docs/releasing.md](docs/releasing.md): release procedure.
+- [docs/research/](docs/research/): experiments behind the current defaults.
+- [model-router.example.json](model-router.example.json): sample profiles.

package/extensions/classifier.ts

@@ -5,10 +5,10 @@ import type {
 } from '@earendil-works/pi-ai';
 import type { ExtensionContext } from '@earendil-works/pi-coding-agent';
 import { isRouterTier, parseCanonicalModelRef } from './config';
+import { DEFAULT_CLASSIFIER_TIMEOUT_MS } from './constants';
 import { extractTextFromContent, getBoundedRecentContext } from './context';
 import type { ClassifierTier, RouterPhase } from './types';
 
-const CLASSIFIER_TIMEOUT_MS = 10_000;
 const CLASSIFIER_MAX_TOKENS = 256;
 
 export const runClassifier = async (
@@ -18,7 +18,7 @@ export const runClassifier = async (
   currentPhase?: RouterPhase,
   thinking?: ThinkingLevel,
   signal?: AbortSignal,
-  routingDeadline = performance.now() + CLASSIFIER_TIMEOUT_MS,
+  routingDeadline = performance.now() + DEFAULT_CLASSIFIER_TIMEOUT_MS,
 ): Promise<{ tier: ClassifierTier } | undefined> => {
   try {
     const remaining = routingDeadline - performance.now();
@@ -51,9 +51,8 @@ export const runClassifier = async (
         },
       ],
     };
-    const timeout = AbortSignal.timeout(
-      Math.max(1, Math.ceil(Math.min(CLASSIFIER_TIMEOUT_MS, remaining))),
-    );
+    // The caller's deadline already reflects the configured classifier budget.
+    const timeout = AbortSignal.timeout(Math.max(1, Math.ceil(remaining)));
     const classifierSignal = signal
       ? AbortSignal.any([signal, timeout])
       : timeout;