PMAT-3718: apr run --json reports prompt_tokens/completion_tokens/finish_reason; serve never reports a cut as a stop

crates/apr-cli/src/commands/inference_output.rs

@@ -184,6 +184,8 @@ struct InferenceOutput {
     /// Populated only when `--stream` asked for it (see
     /// [`decode_token_pieces`]) — every other mode renders the whole `text`.
     token_texts: Option<Vec<String>>,
+    /// Prompt and completion counts plus the finish reason (#3718).
+    usage: RunUsage,
 }
 
 /// Execute inference on model
@@ -231,6 +233,7 @@ fn execute_inference(
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: RunUsage::default(),
         })
     }
 }
@@ -299,7 +302,8 @@ fn execute_with_realizar(
     options: &RunOptions,
     _use_mmap: bool,
 ) -> Result<InferenceOutput> {
-    use realizar::{run_inference, InferenceConfig};
+    use realizar::infer::run_inference_report;
+    use realizar::InferenceConfig;
 
     // Get prompt from options or input file
     let prompt = if let Some(ref p) = options.prompt {
@@ -343,7 +347,10 @@ fn execute_with_realizar(
     }
 
     // Run inference via realizar
-    let result = run_inference(&config).map_err(inference_error)?;
+    // #3718: the report carries what only the decode path knows (finish reason,
+    // context window); the prompt count is `input_token_count`, taken after the
+    // chat template, so it is the number the model actually read.
+    let (result, report) = run_inference_report(&config).map_err(inference_error)?;
 
     // Report performance if benchmarking
     if options.benchmark {
@@ -381,6 +388,12 @@ fn execute_with_realizar(
         used_gpu: Some(result.used_gpu),
         generated_tokens,
         token_texts,
+        usage: RunUsage {
+            prompt_tokens: Some(result.input_token_count),
+            completion_tokens: Some(result.generated_token_count),
+            finish_reason: report.finish_reason.map(|r| r.as_str()),
+            context_length: report.context_length,
+        },
     })
 }
 

crates/apr-cli/src/commands/run.rs

@@ -207,6 +207,25 @@ impl Default for RunOptions {
     }
 }
 
+/// What `apr run --json` reports about the prompt and how generation ended (#3718).
+///
+/// RAH (#3716) rebuilt Qwen3.5's tokenizer from the GGUF just to learn how many
+/// prompt tokens the model saw, and could not tell a reply cut at `--max-tokens`
+/// from one that finished. Every field is `None` when the inference path did not
+/// report it, which is `null` in the JSON and never a zero that reads as measured.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub(crate) struct RunUsage {
+    /// Tokens actually fed to the model: after the chat template, including BOS.
+    pub prompt_tokens: Option<usize>,
+    /// Tokens the engine generated. Unlike `RunResult::tokens_generated`, never a
+    /// word-count stand-in.
+    pub completion_tokens: Option<usize>,
+    /// `"stop"` (a stop token ended it) or `"length"` (the budget did).
+    pub finish_reason: Option<&'static str>,
+    /// The model's context window, from its metadata.
+    pub context_length: Option<usize>,
+}
+
 /// Run result
 #[derive(Debug, Clone)]
 pub(crate) struct RunResult {
@@ -230,6 +249,8 @@ pub(crate) struct RunResult {
     /// could be resolved. `--stream` used to emit `"text":""` for every token
     /// because nothing ever decoded the ids one at a time.
     pub token_texts: Option<Vec<String>>,
+    /// Prompt and completion counts plus the finish reason (#3718).
+    pub usage: RunUsage,
 }
 
 /// Resolve a user-supplied model argument into a [`ModelSource`].
@@ -316,6 +337,7 @@ pub(crate) fn run_model(source: &str, options: &RunOptions) -> Result<RunResult>
         used_gpu: output.used_gpu,
         generated_tokens: output.generated_tokens,
         token_texts: output.token_texts,
+        usage: output.usage,
     })
 }
 

crates/apr-cli/src/commands/run_07.rs

@@ -11,4 +11,5 @@ include!("run_tests_chrome_trace.rs");
 include!("run_tests_layer_trace.rs");
 include!("run_tests_stream_output.rs");
 include!("run_tests_accel_reconcile.rs");
+include!("run_tests_usage_3718.rs");
 }

crates/apr-cli/src/commands/run_entry.rs

@@ -527,6 +527,15 @@ fn build_final_json(
         "inference_time_ms": (result.duration_secs * 1000.0 * 100.0).round() / 100.0,
         "used_gpu": result.used_gpu.unwrap_or(false),
         "cached": result.cached,
+        // #3718: what the model read and why it stopped writing. `prompt_tokens` is
+        // the post-chat-template count (BOS included) fed to the model, so a
+        // consumer no longer rebuilds the tokenizer to learn it. `finish_reason`
+        // is "length" when the budget ended the reply, so a cut answer is never
+        // mistaken for a finished one. Each is null when the path did not report it.
+        "prompt_tokens": result.usage.prompt_tokens,
+        "completion_tokens": result.usage.completion_tokens,
+        "finish_reason": result.usage.finish_reason,
+        "context_length": result.usage.context_length,
         // #3602: `used_gpu: false` alone collapses two different outcomes — "no
         // accelerator was asked for" and "one was asked for, attempted, and
         // REFUSED at runtime". A consumer cannot tell a CPU run from a rejected

crates/apr-cli/src/commands/run_tests_accel_reconcile.rs

@@ -62,6 +62,7 @@ fn a_forced_accelerator_that_ran_on_cpu_is_refused() {
         used_gpu: Some(false),
         generated_tokens: Some(vec![9707]),
         token_texts: None,
+        usage: Default::default(),
     };
 
     let err = reconcile_accelerator(true, &result)
@@ -148,6 +149,7 @@ fn gpu_result(used_gpu: Option<bool>) -> RunResult {
         used_gpu,
         generated_tokens: Some(vec![9707]),
         token_texts: None,
+        usage: Default::default(),
     }
 }
 

crates/apr-cli/src/commands/run_tests_chrome_trace.rs

@@ -24,6 +24,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
 
         let json = build_chrome_trace_events(&result, "model.gguf", 16, true);
@@ -54,6 +55,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
 
         let json = build_chrome_trace_events(&result, "empty.gguf", 0, false);
@@ -73,6 +75,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
 
         let json = build_chrome_trace_events(&result, "model.gguf", 10, false);
@@ -92,6 +95,7 @@
             used_gpu: Some(true),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
 
         let json = build_chrome_trace_events(&result, "my-model.gguf", 64, true);
@@ -111,6 +115,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
 
         let json = build_chrome_trace_events(&result, "model.gguf", 10, false);
@@ -140,6 +145,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
 
         let json = build_chrome_trace_events(&result, "model.gguf", 10, false);
@@ -159,6 +165,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         let json = build_chrome_trace_events(&result, "m.gguf", 1, false);
         assert_eq!(json["displayTimeUnit"], "ms");
@@ -178,6 +185,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         // Just ensure no panic; file creation is best-effort
         print_chrome_trace(&result, "test-model.gguf", 32, false, None);
@@ -210,6 +218,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         print_chrome_trace(&result, "test-model.gguf", 32, false, Some(&target));
 
@@ -239,6 +248,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         print_benchmark_results(&result, "model.gguf", "text", 100);
     }
@@ -254,6 +264,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         print_benchmark_results(&result, "model.gguf", "json", 50);
     }
@@ -269,6 +280,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         print_benchmark_results(&result, "model.gguf", "text", 10);
     }

crates/apr-cli/src/commands/run_tests_inference_output.rs

@@ -15,6 +15,7 @@
             used_gpu: Some(false),
             generated_tokens: Some(vec![1, 2, 3, 4, 5]),
             token_texts: None,
+            usage: Default::default(),
         };
         assert_eq!(output.text, "hello");
         assert_eq!(output.tokens_generated, Some(5));
@@ -32,6 +33,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         assert!(output.tokens_generated.is_none());
         assert!(output.inference_ms.is_none());

crates/apr-cli/src/commands/run_tests_layer_trace.rs

@@ -20,6 +20,7 @@
             used_gpu: Some(false),
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         }
     }
 

crates/apr-cli/src/commands/run_tests_model_source.rs

@@ -39,6 +39,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         let _debug = format!("{:?}", result);
         assert_eq!(result.tokens_generated, Some(5));
@@ -191,6 +192,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         let cloned = result.clone();
         assert_eq!(result.text, cloned.text);

crates/apr-cli/src/commands/run_tests_parse_token_ids.rs

@@ -232,6 +232,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         let result_zero = RunResult {
             text: String::new(),
@@ -242,6 +243,7 @@
             used_gpu: None,
             generated_tokens: None,
             token_texts: None,
+            usage: Default::default(),
         };
         assert_ne!(
             result_none.tokens_generated, result_zero.tokens_generated,
@@ -262,6 +264,7 @@
             used_gpu: Some(true),
             generated_tokens: Some(vec![10, 20, 30]),
             token_texts: None,
+            usage: Default::default(),
         };
         assert_eq!(result.text, "output");
         assert!((result.duration_secs - 1.234).abs() < f64::EPSILON);

crates/apr-cli/src/commands/run_tests_stream_output.rs

@@ -16,6 +16,7 @@ fn stream_output_emits_n_plus_one_json_lines() {
             " world".to_string(),
             "!".to_string(),
         ]),
+        usage: Default::default(),
     };
 
     let mut buf: Vec<u8> = Vec::new();
@@ -85,6 +86,7 @@ fn stream_token_events_carry_their_own_decoded_text() {
             " here".to_string(),
             " to help".to_string(),
         ]),
+        usage: Default::default(),
     };
 
     let mut buf: Vec<u8> = Vec::new();
@@ -125,6 +127,7 @@ fn stream_token_events_degrade_to_empty_text_without_a_tokenizer() {
         used_gpu: Some(false),
         generated_tokens: Some(vec![7, 9]),
         token_texts: None,
+        usage: Default::default(),
     };
 
     let mut buf: Vec<u8> = Vec::new();
@@ -150,6 +153,7 @@ fn stream_output_no_tokens_emits_only_final() {
         used_gpu: Some(false),
         generated_tokens: Some(Vec::new()),
         token_texts: None,
+        usage: Default::default(),
     };
 
     let mut buf: Vec<u8> = Vec::new();
@@ -175,6 +179,7 @@ fn stream_output_none_tokens_emits_only_final() {
         used_gpu: None,
         generated_tokens: None,
         token_texts: None,
+        usage: Default::default(),
     };
 
     let mut buf: Vec<u8> = Vec::new();
@@ -199,6 +204,7 @@ fn build_final_json_matches_legacy_json_shape() {
         used_gpu: Some(true),
         generated_tokens: Some(vec![1, 2, 3]),
         token_texts: None,
+        usage: Default::default(),
     };
     let v = build_final_json(&result, "src.apr", 100, false);
     assert_eq!(v["model"], "src.apr");