zeph_core/
metrics.rs

1// SPDX-FileCopyrightText: 2026 Andrei G <bug-ops>
2// SPDX-License-Identifier: MIT OR Apache-2.0
3
4use std::collections::VecDeque;
5
6use tokio::sync::watch;
7use zeph_common::SecurityEventCategory;
8
9pub use zeph_llm::{ClassifierMetricsSnapshot, TaskMetricsSnapshot};
10pub use zeph_memory::{CategoryScore, ProbeCategory, ProbeVerdict};
11
12/// A single security event record for TUI display.
13#[derive(Debug, Clone)]
14pub struct SecurityEvent {
15    /// Unix timestamp (seconds since epoch).
16    pub timestamp: u64,
17    pub category: SecurityEventCategory,
18    /// Source that triggered the event (e.g., `web_scrape`, `mcp_response`).
19    pub source: String,
20    /// Short description, capped at 128 chars.
21    pub detail: String,
22}
23
24impl SecurityEvent {
25    #[must_use]
26    pub fn new(
27        category: SecurityEventCategory,
28        source: impl Into<String>,
29        detail: impl Into<String>,
30    ) -> Self {
31        // IMP-1: cap source at 64 chars and strip ASCII control chars.
32        let source: String = source
33            .into()
34            .chars()
35            .filter(|c| !c.is_ascii_control())
36            .take(64)
37            .collect();
38        // CR-1: UTF-8 safe truncation using floor_char_boundary (stable since Rust 1.82).
39        let detail = detail.into();
40        let detail = if detail.len() > 128 {
41            let end = detail.floor_char_boundary(127);
42            format!("{}…", &detail[..end])
43        } else {
44            detail
45        };
46        Self {
47            timestamp: std::time::SystemTime::now()
48                .duration_since(std::time::UNIX_EPOCH)
49                .unwrap_or_default()
50                .as_secs(),
51            category,
52            source,
53            detail,
54        }
55    }
56}
57
58/// Ring buffer capacity for security events.
59pub const SECURITY_EVENT_CAP: usize = 100;
60
61/// Lightweight snapshot of a single task row for TUI display.
62///
63/// Captured from the task graph on each metrics tick; kept minimal on purpose.
64#[derive(Debug, Clone)]
65pub struct TaskSnapshotRow {
66    pub id: u32,
67    pub title: String,
68    /// Stringified `TaskStatus` (e.g. `"pending"`, `"running"`, `"completed"`).
69    pub status: String,
70    pub agent: Option<String>,
71    pub duration_ms: u64,
72    /// Truncated error message (first 80 chars) when the task failed.
73    pub error: Option<String>,
74}
75
76/// Lightweight snapshot of a `TaskGraph` for TUI display.
77#[derive(Debug, Clone, Default)]
78pub struct TaskGraphSnapshot {
79    pub graph_id: String,
80    pub goal: String,
81    /// Stringified `GraphStatus` (e.g. `"created"`, `"running"`, `"completed"`).
82    pub status: String,
83    pub tasks: Vec<TaskSnapshotRow>,
84    pub completed_at: Option<std::time::Instant>,
85}
86
87impl TaskGraphSnapshot {
88    /// Returns `true` if this snapshot represents a terminal plan that finished
89    /// more than 30 seconds ago and should no longer be shown in the TUI.
90    #[must_use]
91    pub fn is_stale(&self) -> bool {
92        self.completed_at
93            .is_some_and(|t| t.elapsed().as_secs() > 30)
94    }
95}
96
97/// Counters for the task orchestration subsystem.
98///
99/// Always present in [`MetricsSnapshot`]; zero-valued when orchestration is inactive.
100#[derive(Debug, Clone, Default)]
101pub struct OrchestrationMetrics {
102    pub plans_total: u64,
103    pub tasks_total: u64,
104    pub tasks_completed: u64,
105    pub tasks_failed: u64,
106    pub tasks_skipped: u64,
107}
108
109/// Connection status of a single MCP server for TUI display.
110#[derive(Debug, Clone, PartialEq, Eq)]
111pub enum McpServerConnectionStatus {
112    Connected,
113    Failed,
114}
115
116/// Per-server MCP status snapshot for TUI display.
117#[derive(Debug, Clone)]
118pub struct McpServerStatus {
119    pub id: String,
120    pub status: McpServerConnectionStatus,
121    /// Number of tools provided by this server (0 when failed).
122    pub tool_count: usize,
123    /// Human-readable failure reason. Empty when connected.
124    pub error: String,
125}
126
127/// Bayesian confidence data for a single skill, used by TUI confidence bar.
128#[derive(Debug, Clone, Default)]
129pub struct SkillConfidence {
130    pub name: String,
131    pub posterior: f64,
132    pub total_uses: u32,
133}
134
135/// Snapshot of a single sub-agent's runtime status.
136#[derive(Debug, Clone, Default)]
137pub struct SubAgentMetrics {
138    pub id: String,
139    pub name: String,
140    /// Stringified `TaskState`: "working", "completed", "failed", "canceled", etc.
141    pub state: String,
142    pub turns_used: u32,
143    pub max_turns: u32,
144    pub background: bool,
145    pub elapsed_secs: u64,
146    /// Stringified `PermissionMode`: `"default"`, `"accept_edits"`, `"dont_ask"`,
147    /// `"bypass_permissions"`, `"plan"`. Empty string when mode is `Default`.
148    pub permission_mode: String,
149    /// Path to the directory containing this agent's JSONL transcript file.
150    /// `None` when transcript writing is disabled for this agent.
151    pub transcript_dir: Option<String>,
152}
153
154/// Per-turn latency breakdown for the four agent hot-path phases.
155///
156/// Populated with `Instant`-based measurements at each phase boundary.
157/// All values are in milliseconds.
158#[derive(Debug, Clone, Default)]
159pub struct TurnTimings {
160    pub prepare_context_ms: u64,
161    pub llm_chat_ms: u64,
162    pub tool_exec_ms: u64,
163    pub persist_message_ms: u64,
164}
165
166/// Live snapshot of agent metrics broadcast via a [`tokio::sync::watch`] channel.
167///
168/// Fields are updated at different rates: some once at startup (static), others every turn
169/// (dynamic). For fields that are known at agent startup and do not change during the session,
170/// use [`StaticMetricsInit`] and `AgentBuilder::with_static_metrics` instead of
171/// adding a raw `send_modify` call in the runner.
172#[derive(Debug, Clone, Default)]
173#[allow(clippy::struct_excessive_bools)] // independent boolean flags; bitflags or enum would obscure semantics without reducing complexity
174pub struct MetricsSnapshot {
175    pub prompt_tokens: u64,
176    pub completion_tokens: u64,
177    pub total_tokens: u64,
178    pub context_tokens: u64,
179    pub api_calls: u64,
180    pub active_skills: Vec<String>,
181    pub total_skills: usize,
182    /// Total configured MCP servers (connected + failed).
183    pub mcp_server_count: usize,
184    pub mcp_tool_count: usize,
185    /// Number of successfully connected MCP servers.
186    pub mcp_connected_count: usize,
187    /// Per-server connection status list.
188    pub mcp_servers: Vec<McpServerStatus>,
189    pub active_mcp_tools: Vec<String>,
190    pub sqlite_message_count: u64,
191    pub sqlite_conversation_id: Option<zeph_memory::ConversationId>,
192    pub qdrant_available: bool,
193    pub vector_backend: String,
194    pub embeddings_generated: u64,
195    pub last_llm_latency_ms: u64,
196    pub uptime_seconds: u64,
197    pub provider_name: String,
198    pub model_name: String,
199    pub summaries_count: u64,
200    pub context_compactions: u64,
201    /// Number of times the agent entered the Hard compaction tier, including cooldown-skipped
202    /// turns. Not equal to the actual LLM summarization count — reflects pressure, not action.
203    pub compaction_hard_count: u64,
204    /// User-message turns elapsed after each hard compaction event.
205    /// Entry i = turns between hard compaction i and hard compaction i+1 (or session end).
206    /// Empty when no hard compaction occurred during the session.
207    pub compaction_turns_after_hard: Vec<u64>,
208    pub compression_events: u64,
209    pub compression_tokens_saved: u64,
210    pub tool_output_prunes: u64,
211    /// Compaction probe outcomes (#1609).
212    pub compaction_probe_passes: u64,
213    /// Compaction probe soft failures (summary borderline — compaction proceeded with warning).
214    pub compaction_probe_soft_failures: u64,
215    /// Compaction probe hard failures (compaction blocked due to lossy summary).
216    pub compaction_probe_failures: u64,
217    /// Compaction probe errors (LLM/timeout — non-blocking, compaction proceeded).
218    pub compaction_probe_errors: u64,
219    /// Last compaction probe verdict. `None` before the first probe completes.
220    pub last_probe_verdict: Option<zeph_memory::ProbeVerdict>,
221    /// Last compaction probe score in [0.0, 1.0]. `None` before the first probe
222    /// completes or after an Error verdict (errors produce no score).
223    pub last_probe_score: Option<f32>,
224    /// Per-category scores from the last completed probe.
225    pub last_probe_category_scores: Option<Vec<zeph_memory::CategoryScore>>,
226    /// Configured pass threshold for the compaction probe. Used by TUI for category color-coding.
227    pub compaction_probe_threshold: f32,
228    /// Configured hard-fail threshold for the compaction probe.
229    pub compaction_probe_hard_fail_threshold: f32,
230    pub cache_read_tokens: u64,
231    pub cache_creation_tokens: u64,
232    pub cost_spent_cents: f64,
233    /// Cost per successful task in cents. `None` until at least one task completes.
234    pub cost_cps_cents: Option<f64>,
235    /// Number of successful tasks recorded today.
236    pub cost_successful_tasks: u64,
237    /// Per-provider cost breakdown, sorted by cost descending.
238    pub provider_cost_breakdown: Vec<(String, crate::cost::ProviderUsage)>,
239    pub filter_raw_tokens: u64,
240    pub filter_saved_tokens: u64,
241    pub filter_applications: u64,
242    pub filter_total_commands: u64,
243    pub filter_filtered_commands: u64,
244    pub filter_confidence_full: u64,
245    pub filter_confidence_partial: u64,
246    pub filter_confidence_fallback: u64,
247    pub cancellations: u64,
248    pub server_compaction_events: u64,
249    pub sanitizer_runs: u64,
250    pub sanitizer_injection_flags: u64,
251    /// Injection pattern hits on `ToolResult` (local) sources — likely false positives.
252    ///
253    /// Counts regex hits that fired on content from `shell`, `read_file`, `search_code`, etc.
254    /// These sources are user-owned and not adversarial; a non-zero value indicates a pattern
255    /// that needs tightening or a source reclassification.
256    pub sanitizer_injection_fp_local: u64,
257    pub sanitizer_truncations: u64,
258    pub quarantine_invocations: u64,
259    pub quarantine_failures: u64,
260    /// ML classifier hard-blocked tool outputs (`enforcement_mode=block` only).
261    pub classifier_tool_blocks: u64,
262    /// ML classifier suspicious tool outputs (both enforcement modes).
263    pub classifier_tool_suspicious: u64,
264    /// `TurnCausalAnalyzer` flags: behavioral deviation detected at tool-return boundary.
265    pub causal_ipi_flags: u64,
266    /// VIGIL pre-sanitizer flags: tool outputs matched injection patterns (any action).
267    pub vigil_flags_total: u64,
268    /// VIGIL pre-sanitizer blocks: tool outputs replaced with sentinel (`strict_mode=true`).
269    pub vigil_blocks_total: u64,
270    pub exfiltration_images_blocked: u64,
271    pub exfiltration_tool_urls_flagged: u64,
272    pub exfiltration_memory_guards: u64,
273    pub pii_scrub_count: u64,
274    /// Number of times the PII NER classifier timed out; input fell back to regex-only.
275    pub pii_ner_timeouts: u64,
276    /// Number of times the PII NER circuit breaker tripped (disabled NER for the session).
277    pub pii_ner_circuit_breaker_trips: u64,
278    pub memory_validation_failures: u64,
279    pub rate_limit_trips: u64,
280    pub pre_execution_blocks: u64,
281    pub pre_execution_warnings: u64,
282    /// `true` when a guardrail filter is active for this session.
283    pub guardrail_enabled: bool,
284    /// `true` when guardrail is in warn-only mode (action = warn).
285    pub guardrail_warn_mode: bool,
286    pub sub_agents: Vec<SubAgentMetrics>,
287    pub skill_confidence: Vec<SkillConfidence>,
288    /// Scheduled task summaries: `[name, kind, mode, next_run]`.
289    pub scheduled_tasks: Vec<[String; 4]>,
290    /// Thompson Sampling distribution snapshots: `(provider, alpha, beta)`.
291    pub router_thompson_stats: Vec<(String, f64, f64)>,
292    /// Ring buffer of recent security events (cap 100, FIFO eviction).
293    pub security_events: VecDeque<SecurityEvent>,
294    pub orchestration: OrchestrationMetrics,
295    /// Live snapshot of the currently active task graph. `None` when no plan is active.
296    pub orchestration_graph: Option<TaskGraphSnapshot>,
297    pub graph_community_detection_failures: u64,
298    pub graph_entities_total: u64,
299    pub graph_edges_total: u64,
300    pub graph_communities_total: u64,
301    pub graph_extraction_count: u64,
302    pub graph_extraction_failures: u64,
303    /// `true` when `config.llm.cloud.enable_extended_context = true`.
304    /// Never set for other providers to avoid false positives.
305    pub extended_context: bool,
306    /// Latest compression-guidelines version (0 = no guidelines yet).
307    pub guidelines_version: u32,
308    /// ISO 8601 timestamp of the latest guidelines update (empty if none).
309    pub guidelines_updated_at: String,
310    pub tool_cache_hits: u64,
311    pub tool_cache_misses: u64,
312    pub tool_cache_entries: usize,
313    /// Number of semantic-tier facts in memory (0 when tier promotion disabled).
314    pub semantic_fact_count: u64,
315    /// STT model name (e.g. "whisper-1"). `None` when STT is not configured.
316    pub stt_model: Option<String>,
317    /// Model used for context compaction/summarization. `None` when no summary provider is set.
318    pub compaction_model: Option<String>,
319    /// Temperature of the active provider when using Candle. `None` for API providers.
320    pub provider_temperature: Option<f32>,
321    /// Top-p of the active provider when using Candle. `None` for API providers.
322    pub provider_top_p: Option<f32>,
323    /// Embedding model name (e.g. `"nomic-embed-text"`). Empty when embeddings are disabled.
324    pub embedding_model: String,
325    /// Token budget for context window. `None` when not configured.
326    pub token_budget: Option<u64>,
327    /// Token threshold that triggers soft compaction. `None` when not configured.
328    pub compaction_threshold: Option<u32>,
329    /// Vault backend identifier: "age", "env", or "none".
330    pub vault_backend: String,
331    /// Active I/O channel name: `"cli"`, `"telegram"`, `"tui"`, `"discord"`, `"slack"`.
332    pub active_channel: String,
333    /// Background supervisor: inflight tasks across all classes.
334    pub bg_inflight: u64,
335    /// Background supervisor: total tasks dropped due to concurrency limit (all classes).
336    pub bg_dropped: u64,
337    /// Background supervisor: total tasks completed (all classes).
338    pub bg_completed: u64,
339    /// Background supervisor: inflight enrichment tasks.
340    pub bg_enrichment_inflight: u64,
341    /// Background supervisor: inflight telemetry tasks.
342    pub bg_telemetry_inflight: u64,
343    /// In-flight background shell runs. Empty when none are running or no `ShellExecutor` is wired.
344    pub shell_background_runs: Vec<ShellBackgroundRunRow>,
345    /// Whether self-learning (skill evolution) is enabled.
346    pub self_learning_enabled: bool,
347    /// Whether the semantic response cache is enabled.
348    pub semantic_cache_enabled: bool,
349    /// Whether semantic response caching is enabled (alias for `semantic_cache_enabled`).
350    pub cache_enabled: bool,
351    /// Whether assistant messages are auto-saved to memory.
352    pub autosave_enabled: bool,
353    /// Classifier p50/p95 latency metrics per task (injection, pii, feedback).
354    pub classifier: ClassifierMetricsSnapshot,
355    /// Latency breakdown for the most recently completed agent turn.
356    pub last_turn_timings: TurnTimings,
357    /// Rolling average of per-phase latency over the last 10 turns.
358    pub avg_turn_timings: TurnTimings,
359    /// Maximum per-phase latency observed within the rolling window (tail-latency visibility).
360    ///
361    /// M3: exposes `max_in_window` alongside the rolling average for operational monitoring.
362    pub max_turn_timings: TurnTimings,
363    /// Number of turns included in `avg_turn_timings` and `max_turn_timings` (capped at 10).
364    pub timing_sample_count: u64,
365    /// Total egress (outbound HTTP) requests attempted this session.
366    pub egress_requests_total: u64,
367    /// Egress events dropped due to bounded channel backpressure.
368    pub egress_dropped_total: u64,
369    /// Egress requests blocked by scheme/domain/SSRF policy.
370    pub egress_blocked_total: u64,
371    /// Runtime-resolved context window limit (tokens).
372    ///
373    /// Populated from `resolve_context_budget` after provider pool construction and refreshed on
374    /// every `/provider` switch. `0` means unknown (pre-init or provider has no declared window);
375    /// the TUI gauge renders `"—"` in this case to avoid divide-by-zero.
376    pub context_max_tokens: u64,
377    /// Token count at the time the most recent compaction was triggered. `0` = never compacted.
378    pub compaction_last_before: u64,
379    /// Token count after the most recent compaction completed. `0` = never compacted.
380    pub compaction_last_after: u64,
381    /// Unix epoch milliseconds when the most recent compaction occurred. `0` = never compacted.
382    pub compaction_last_at_ms: u64,
383    /// Active long-horizon goal for TUI display. `None` when no goal is active.
384    pub active_goal: Option<crate::goal::GoalSnapshot>,
385    /// Cocoon sidecar connection state. `None` when Cocoon is not configured.
386    /// `Some(true)` = proxy connected, `Some(false)` = unreachable or disconnected.
387    pub cocoon_connected: Option<bool>,
388    /// Worker count reported by the Cocoon sidecar. `0` when not connected or not configured.
389    pub cocoon_worker_count: u32,
390    /// Number of models available through the Cocoon sidecar.
391    pub cocoon_model_count: usize,
392    /// TON wallet balance in TON units. `None` when unknown or Cocoon not configured.
393    pub cocoon_ton_balance: Option<f64>,
394}
395
396/// Snapshot of a single in-flight background shell run for TUI display.
397///
398/// Populated from [`zeph_tools::BackgroundRunSnapshot`] during the per-turn metrics update in
399/// `reap_background_tasks_and_update_metrics`. The `run_id` is truncated to 8 hex chars for
400/// compact TUI rendering.
401#[derive(Debug, Clone, Default, serde::Serialize)]
402pub struct ShellBackgroundRunRow {
403    /// First 8 hex characters of the run's UUID, sufficient for unique identification in TUI.
404    pub run_id: String,
405    /// Original command, truncated to 80 characters.
406    pub command: String,
407    /// Wall-clock elapsed seconds since spawn.
408    pub elapsed_secs: u64,
409}
410
411/// Configuration-derived fields of [`MetricsSnapshot`] that are known at agent startup and do
412/// not change during the session.
413///
414/// Pass this struct to `AgentBuilder::with_static_metrics` immediately after
415/// `AgentBuilder::with_metrics` to initialize all static fields in one place rather than
416/// through scattered `send_modify` calls in the runner.
417///
418/// # Examples
419///
420/// ```no_run
421/// use zeph_core::metrics::StaticMetricsInit;
422///
423/// let init = StaticMetricsInit {
424///     active_channel: "cli".to_owned(),
425///     ..StaticMetricsInit::default()
426/// };
427/// ```
428#[derive(Debug, Default)]
429pub struct StaticMetricsInit {
430    /// STT model name (e.g. `"whisper-1"`). `None` when STT is not configured.
431    pub stt_model: Option<String>,
432    /// Model used for context compaction/summarization. `None` when no summary provider is set.
433    pub compaction_model: Option<String>,
434    /// Whether the semantic response cache is enabled.
435    ///
436    /// This value is also written to [`MetricsSnapshot::cache_enabled`] which is an alias for the
437    /// same concept.
438    pub semantic_cache_enabled: bool,
439    /// Embedding model name (e.g. `"nomic-embed-text"`). Empty when embeddings are disabled.
440    pub embedding_model: String,
441    /// Whether self-learning (skill evolution) is enabled.
442    pub self_learning_enabled: bool,
443    /// Active I/O channel name: `"cli"`, `"telegram"`, `"tui"`, `"discord"`, `"slack"`.
444    pub active_channel: String,
445    /// Token budget for context window. `None` when not configured.
446    pub token_budget: Option<u64>,
447    /// Token threshold that triggers soft compaction. `None` when not configured.
448    pub compaction_threshold: Option<u32>,
449    /// Vault backend identifier: `"age"`, `"env"`, or `"none"`.
450    pub vault_backend: String,
451    /// Whether assistant messages are auto-saved to memory.
452    pub autosave_enabled: bool,
453    /// Override for the active model name. When `Some`, replaces the model name set by the
454    /// builder from `runtime.model_name` (which may be a placeholder) with the effective model
455    /// resolved from the LLM provider configuration.
456    pub model_name_override: Option<String>,
457}
458
459/// Strip ASCII control characters and ANSI escape sequences from a string for safe TUI display.
460///
461/// Allows tab, LF, and CR; removes everything else in the `0x00–0x1F` range including full
462/// ANSI CSI sequences (`ESC[...`). This prevents escape-sequence injection from LLM planner
463/// output into the TUI.
464fn strip_ctrl(s: &str) -> String {
465    let mut out = String::with_capacity(s.len());
466    let mut chars = s.chars().peekable();
467    while let Some(c) = chars.next() {
468        if c == '\x1b' {
469            // Consume an ANSI CSI sequence: ESC [ <params> <final-byte in 0x40–0x7E>
470            if chars.peek() == Some(&'[') {
471                chars.next(); // consume '['
472                for inner in chars.by_ref() {
473                    if ('\x40'..='\x7e').contains(&inner) {
474                        break;
475                    }
476                }
477            }
478            // Drop ESC and any consumed sequence — write nothing.
479        } else if c.is_control() && c != '\t' && c != '\n' && c != '\r' {
480            // drop other control chars
481        } else {
482            out.push(c);
483        }
484    }
485    out
486}
487
488/// Convert a live `TaskGraph` into a lightweight snapshot for TUI display.
489impl From<&zeph_orchestration::TaskGraph> for TaskGraphSnapshot {
490    fn from(graph: &zeph_orchestration::TaskGraph) -> Self {
491        let tasks = graph
492            .tasks
493            .iter()
494            .map(|t| {
495                let error = t
496                    .result
497                    .as_ref()
498                    .filter(|_| t.status == zeph_orchestration::TaskStatus::Failed)
499                    .and_then(|r| {
500                        if r.output.is_empty() {
501                            None
502                        } else {
503                            // Strip control chars, then truncate at 80 chars (SEC-P6-01).
504                            let s = strip_ctrl(&r.output);
505                            if s.len() > 80 {
506                                let end = s.floor_char_boundary(79);
507                                Some(format!("{}…", &s[..end]))
508                            } else {
509                                Some(s)
510                            }
511                        }
512                    });
513                let duration_ms = t.result.as_ref().map_or(0, |r| r.duration_ms);
514                TaskSnapshotRow {
515                    id: t.id.as_u32(),
516                    title: strip_ctrl(&t.title),
517                    status: t.status.to_string(),
518                    agent: t.assigned_agent.as_deref().map(strip_ctrl),
519                    duration_ms,
520                    error,
521                }
522            })
523            .collect();
524        Self {
525            graph_id: graph.id.to_string(),
526            goal: strip_ctrl(&graph.goal),
527            status: graph.status.to_string(),
528            tasks,
529            completed_at: None,
530        }
531    }
532}
533
534pub struct MetricsCollector {
535    tx: watch::Sender<MetricsSnapshot>,
536}
537
538impl MetricsCollector {
539    #[must_use]
540    pub fn new() -> (Self, watch::Receiver<MetricsSnapshot>) {
541        let (tx, rx) = watch::channel(MetricsSnapshot::default());
542        (Self { tx }, rx)
543    }
544
545    pub fn update(&self, f: impl FnOnce(&mut MetricsSnapshot)) {
546        self.tx.send_modify(f);
547    }
548
549    /// Publish the runtime-resolved context window limit.
550    ///
551    /// Call after the provider pool is constructed (builder) and on every successful
552    /// `/provider` switch so the TUI context gauge reflects the active provider's window.
553    ///
554    /// # Examples
555    ///
556    /// ```rust
557    /// use zeph_core::metrics::MetricsCollector;
558    ///
559    /// let (collector, rx) = MetricsCollector::new();
560    /// collector.set_context_max_tokens(128_000);
561    /// assert_eq!(rx.borrow().context_max_tokens, 128_000);
562    /// ```
563    pub fn set_context_max_tokens(&self, max_tokens: u64) {
564        self.tx.send_modify(|m| m.context_max_tokens = max_tokens);
565    }
566
567    /// Record the outcome of the most recent compaction event.
568    ///
569    /// Sets all three `compaction_last_*` fields atomically. `at_ms` is the Unix epoch in
570    /// milliseconds — obtain via `SystemTime::UNIX_EPOCH.elapsed().unwrap_or_default().as_millis()`.
571    ///
572    /// # Examples
573    ///
574    /// ```rust
575    /// use zeph_core::metrics::MetricsCollector;
576    ///
577    /// let (collector, rx) = MetricsCollector::new();
578    /// collector.record_compaction(50_000, 12_000, 1_700_000_000_000);
579    /// let snap = rx.borrow();
580    /// assert_eq!(snap.compaction_last_before, 50_000);
581    /// assert_eq!(snap.compaction_last_after, 12_000);
582    /// assert_eq!(snap.compaction_last_at_ms, 1_700_000_000_000);
583    /// ```
584    pub fn record_compaction(&self, before: u64, after: u64, at_ms: u64) {
585        self.tx.send_modify(|m| {
586            m.compaction_last_before = before;
587            m.compaction_last_after = after;
588            m.compaction_last_at_ms = at_ms;
589        });
590    }
591
592    /// Returns a clone of the underlying [`watch::Sender`].
593    ///
594    /// Use this to pass the sender to code that requires a raw
595    /// `watch::Sender<MetricsSnapshot>` while the [`MetricsCollector`] is
596    /// also shared (e.g., passed to a `MetricsBridge` layer).
597    #[must_use]
598    pub fn sender(&self) -> watch::Sender<MetricsSnapshot> {
599        self.tx.clone()
600    }
601}
602
603// ---------------------------------------------------------------------------
604// HistogramRecorder
605// ---------------------------------------------------------------------------
606
607/// Per-event histogram recording contract for the agent loop.
608///
609/// Implementors record individual latency observations into Prometheus histograms
610/// (or any other backend). The trait is object-safe: the agent stores an
611/// `Option<Arc<dyn HistogramRecorder>>` and calls these methods at each measurement
612/// point. When `None`, recording is a no-op with zero overhead.
613///
614/// # Contract for implementors
615///
616/// - All methods must be non-blocking; they must not call async code or acquire
617///   mutexes that may block.
618/// - Implementations must be `Send + Sync` — the agent loop runs on the tokio
619///   thread pool and the recorder may be called from multiple tasks.
620///
621/// # Examples
622///
623/// ```rust
624/// use std::sync::Arc;
625/// use std::time::Duration;
626/// use zeph_core::metrics::HistogramRecorder;
627///
628/// struct NoOpRecorder;
629///
630/// impl HistogramRecorder for NoOpRecorder {
631///     fn observe_llm_latency(&self, _: Duration) {}
632///     fn observe_turn_duration(&self, _: Duration) {}
633///     fn observe_tool_execution(&self, _: Duration) {}
634///     fn observe_bg_task(&self, _: &str, _: Duration) {}
635/// }
636///
637/// let recorder: Arc<dyn HistogramRecorder> = Arc::new(NoOpRecorder);
638/// recorder.observe_llm_latency(Duration::from_millis(500));
639/// ```
640pub trait HistogramRecorder: Send + Sync {
641    /// Record a single LLM API call latency observation.
642    fn observe_llm_latency(&self, duration: std::time::Duration);
643
644    /// Record a full agent turn duration observation (context prep + LLM + tools + persist).
645    fn observe_turn_duration(&self, duration: std::time::Duration);
646
647    /// Record a single tool execution latency observation.
648    fn observe_tool_execution(&self, duration: std::time::Duration);
649
650    /// Record a background task completion latency.
651    ///
652    /// `class_label` is `"enrichment"` or `"telemetry"` (from `TaskClass::name()`).
653    fn observe_bg_task(&self, class_label: &str, duration: std::time::Duration);
654}
655
656#[cfg(test)]
657mod tests {
658    #![allow(clippy::field_reassign_with_default)]
659
660    use super::*;
661
662    #[test]
663    fn default_metrics_snapshot() {
664        let m = MetricsSnapshot::default();
665        assert_eq!(m.total_tokens, 0);
666        assert_eq!(m.api_calls, 0);
667        assert!(m.active_skills.is_empty());
668        assert!(m.active_mcp_tools.is_empty());
669        assert_eq!(m.mcp_tool_count, 0);
670        assert_eq!(m.mcp_server_count, 0);
671        assert!(m.provider_name.is_empty());
672        assert_eq!(m.summaries_count, 0);
673        // Phase 2 fields
674        assert!(m.stt_model.is_none());
675        assert!(m.compaction_model.is_none());
676        assert!(m.provider_temperature.is_none());
677        assert!(m.provider_top_p.is_none());
678        assert!(m.active_channel.is_empty());
679        assert!(m.embedding_model.is_empty());
680        assert!(m.token_budget.is_none());
681        assert!(!m.self_learning_enabled);
682        assert!(!m.semantic_cache_enabled);
683    }
684
685    #[test]
686    fn metrics_collector_update_phase2_fields() {
687        let (collector, rx) = MetricsCollector::new();
688        collector.update(|m| {
689            m.stt_model = Some("whisper-1".into());
690            m.compaction_model = Some("haiku".into());
691            m.provider_temperature = Some(0.7);
692            m.provider_top_p = Some(0.95);
693            m.active_channel = "tui".into();
694            m.embedding_model = "nomic-embed-text".into();
695            m.token_budget = Some(200_000);
696            m.self_learning_enabled = true;
697            m.semantic_cache_enabled = true;
698        });
699        let s = rx.borrow();
700        assert_eq!(s.stt_model.as_deref(), Some("whisper-1"));
701        assert_eq!(s.compaction_model.as_deref(), Some("haiku"));
702        assert_eq!(s.provider_temperature, Some(0.7));
703        assert_eq!(s.provider_top_p, Some(0.95));
704        assert_eq!(s.active_channel, "tui");
705        assert_eq!(s.embedding_model, "nomic-embed-text");
706        assert_eq!(s.token_budget, Some(200_000));
707        assert!(s.self_learning_enabled);
708        assert!(s.semantic_cache_enabled);
709    }
710
711    #[test]
712    fn metrics_collector_update() {
713        let (collector, rx) = MetricsCollector::new();
714        collector.update(|m| {
715            m.api_calls = 5;
716            m.total_tokens = 1000;
717        });
718        let snapshot = rx.borrow().clone();
719        assert_eq!(snapshot.api_calls, 5);
720        assert_eq!(snapshot.total_tokens, 1000);
721    }
722
723    #[test]
724    fn metrics_collector_multiple_updates() {
725        let (collector, rx) = MetricsCollector::new();
726        collector.update(|m| m.api_calls = 1);
727        collector.update(|m| m.api_calls += 1);
728        assert_eq!(rx.borrow().api_calls, 2);
729    }
730
731    #[test]
732    fn metrics_snapshot_clone() {
733        let mut m = MetricsSnapshot::default();
734        m.provider_name = "ollama".into();
735        let cloned = m.clone();
736        assert_eq!(cloned.provider_name, "ollama");
737    }
738
739    #[test]
740    fn filter_metrics_tracking() {
741        let (collector, rx) = MetricsCollector::new();
742        collector.update(|m| {
743            m.filter_raw_tokens += 250;
744            m.filter_saved_tokens += 200;
745            m.filter_applications += 1;
746        });
747        collector.update(|m| {
748            m.filter_raw_tokens += 100;
749            m.filter_saved_tokens += 80;
750            m.filter_applications += 1;
751        });
752        let s = rx.borrow();
753        assert_eq!(s.filter_raw_tokens, 350);
754        assert_eq!(s.filter_saved_tokens, 280);
755        assert_eq!(s.filter_applications, 2);
756    }
757
758    #[test]
759    fn filter_confidence_and_command_metrics() {
760        let (collector, rx) = MetricsCollector::new();
761        collector.update(|m| {
762            m.filter_total_commands += 1;
763            m.filter_filtered_commands += 1;
764            m.filter_confidence_full += 1;
765        });
766        collector.update(|m| {
767            m.filter_total_commands += 1;
768            m.filter_confidence_partial += 1;
769        });
770        let s = rx.borrow();
771        assert_eq!(s.filter_total_commands, 2);
772        assert_eq!(s.filter_filtered_commands, 1);
773        assert_eq!(s.filter_confidence_full, 1);
774        assert_eq!(s.filter_confidence_partial, 1);
775        assert_eq!(s.filter_confidence_fallback, 0);
776    }
777
778    #[test]
779    fn summaries_count_tracks_summarizations() {
780        let (collector, rx) = MetricsCollector::new();
781        collector.update(|m| m.summaries_count += 1);
782        collector.update(|m| m.summaries_count += 1);
783        assert_eq!(rx.borrow().summaries_count, 2);
784    }
785
786    #[test]
787    fn cancellations_counter_increments() {
788        let (collector, rx) = MetricsCollector::new();
789        assert_eq!(rx.borrow().cancellations, 0);
790        collector.update(|m| m.cancellations += 1);
791        collector.update(|m| m.cancellations += 1);
792        assert_eq!(rx.borrow().cancellations, 2);
793    }
794
795    #[test]
796    fn security_event_detail_exact_128_not_truncated() {
797        let s = "a".repeat(128);
798        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, "src", s.clone());
799        assert_eq!(ev.detail, s, "128-char string must not be truncated");
800    }
801
802    #[test]
803    fn security_event_detail_129_is_truncated() {
804        let s = "a".repeat(129);
805        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, "src", s);
806        assert!(
807            ev.detail.ends_with('…'),
808            "129-char string must end with ellipsis"
809        );
810        assert!(
811            ev.detail.len() <= 130,
812            "truncated detail must be at most 130 bytes"
813        );
814    }
815
816    #[test]
817    fn security_event_detail_multibyte_utf8_no_panic() {
818        // Each '中' is 3 bytes. 43 chars = 129 bytes — triggers truncation at a multi-byte boundary.
819        let s = "中".repeat(43);
820        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, "src", s);
821        assert!(ev.detail.ends_with('…'));
822    }
823
824    #[test]
825    fn security_event_source_capped_at_64_chars() {
826        let long_source = "x".repeat(200);
827        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, long_source, "detail");
828        assert_eq!(ev.source.len(), 64);
829    }
830
831    #[test]
832    fn security_event_source_strips_control_chars() {
833        let source = "tool\x00name\x1b[31m";
834        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, source, "detail");
835        assert!(!ev.source.contains('\x00'));
836        assert!(!ev.source.contains('\x1b'));
837    }
838
839    #[test]
840    fn security_event_category_as_str() {
841        assert_eq!(SecurityEventCategory::InjectionFlag.as_str(), "injection");
842        assert_eq!(SecurityEventCategory::ExfiltrationBlock.as_str(), "exfil");
843        assert_eq!(SecurityEventCategory::Quarantine.as_str(), "quarantine");
844        assert_eq!(SecurityEventCategory::Truncation.as_str(), "truncation");
845        assert_eq!(
846            SecurityEventCategory::CrossBoundaryMcpToAcp.as_str(),
847            "cross_boundary_mcp_to_acp"
848        );
849    }
850
851    #[test]
852    fn ring_buffer_respects_cap_via_update() {
853        let (collector, rx) = MetricsCollector::new();
854        for i in 0..110u64 {
855            let event = SecurityEvent::new(
856                SecurityEventCategory::InjectionFlag,
857                "src",
858                format!("event {i}"),
859            );
860            collector.update(|m| {
861                if m.security_events.len() >= SECURITY_EVENT_CAP {
862                    m.security_events.pop_front();
863                }
864                m.security_events.push_back(event);
865            });
866        }
867        let snap = rx.borrow();
868        assert_eq!(snap.security_events.len(), SECURITY_EVENT_CAP);
869        // FIFO: earliest events evicted, last one present
870        assert!(snap.security_events.back().unwrap().detail.contains("109"));
871    }
872
873    #[test]
874    fn security_events_empty_by_default() {
875        let m = MetricsSnapshot::default();
876        assert!(m.security_events.is_empty());
877    }
878
879    #[test]
880    fn orchestration_metrics_default_zero() {
881        let m = OrchestrationMetrics::default();
882        assert_eq!(m.plans_total, 0);
883        assert_eq!(m.tasks_total, 0);
884        assert_eq!(m.tasks_completed, 0);
885        assert_eq!(m.tasks_failed, 0);
886        assert_eq!(m.tasks_skipped, 0);
887    }
888
889    #[test]
890    fn metrics_snapshot_includes_orchestration_default_zero() {
891        let m = MetricsSnapshot::default();
892        assert_eq!(m.orchestration.plans_total, 0);
893        assert_eq!(m.orchestration.tasks_total, 0);
894        assert_eq!(m.orchestration.tasks_completed, 0);
895    }
896
897    #[test]
898    fn orchestration_metrics_update_via_collector() {
899        let (collector, rx) = MetricsCollector::new();
900        collector.update(|m| {
901            m.orchestration.plans_total += 1;
902            m.orchestration.tasks_total += 5;
903            m.orchestration.tasks_completed += 3;
904            m.orchestration.tasks_failed += 1;
905            m.orchestration.tasks_skipped += 1;
906        });
907        let s = rx.borrow();
908        assert_eq!(s.orchestration.plans_total, 1);
909        assert_eq!(s.orchestration.tasks_total, 5);
910        assert_eq!(s.orchestration.tasks_completed, 3);
911        assert_eq!(s.orchestration.tasks_failed, 1);
912        assert_eq!(s.orchestration.tasks_skipped, 1);
913    }
914
915    #[test]
916    fn strip_ctrl_removes_escape_sequences() {
917        let input = "hello\x1b[31mworld\x00end";
918        let result = strip_ctrl(input);
919        assert_eq!(result, "helloworldend");
920    }
921
922    #[test]
923    fn strip_ctrl_allows_tab_lf_cr() {
924        let input = "a\tb\nc\rd";
925        let result = strip_ctrl(input);
926        assert_eq!(result, "a\tb\nc\rd");
927    }
928
929    #[test]
930    fn task_graph_snapshot_is_stale_after_30s() {
931        let mut snap = TaskGraphSnapshot::default();
932        // Not stale if no completed_at.
933        assert!(!snap.is_stale());
934        // Not stale if just completed.
935        snap.completed_at = Some(std::time::Instant::now());
936        assert!(!snap.is_stale());
937        // Stale if completed more than 30s ago.
938        snap.completed_at = Some(
939            std::time::Instant::now()
940                .checked_sub(std::time::Duration::from_secs(31))
941                .unwrap(),
942        );
943        assert!(snap.is_stale());
944    }
945
946    // T1: From<&TaskGraph> correctly maps fields including duration_ms and error truncation.
947    #[test]
948    fn task_graph_snapshot_from_task_graph_maps_fields() {
949        use zeph_orchestration::{GraphStatus, TaskGraph, TaskNode, TaskResult, TaskStatus};
950
951        let mut graph = TaskGraph::new("My goal");
952        let mut task = TaskNode::new(0, "Do work", "description");
953        task.status = TaskStatus::Failed;
954        task.assigned_agent = Some("agent-1".into());
955        task.result = Some(TaskResult {
956            output: "error occurred here".into(),
957            artifacts: vec![],
958            duration_ms: 1234,
959            agent_id: None,
960            agent_def: None,
961        });
962        graph.tasks.push(task);
963        graph.status = GraphStatus::Failed;
964
965        let snap = TaskGraphSnapshot::from(&graph);
966        assert_eq!(snap.goal, "My goal");
967        assert_eq!(snap.status, "failed");
968        assert_eq!(snap.tasks.len(), 1);
969        let row = &snap.tasks[0];
970        assert_eq!(row.title, "Do work");
971        assert_eq!(row.status, "failed");
972        assert_eq!(row.agent.as_deref(), Some("agent-1"));
973        assert_eq!(row.duration_ms, 1234);
974        assert!(row.error.as_deref().unwrap().contains("error occurred"));
975    }
976
977    // T2: From impl compiles with orchestration feature active.
978    #[test]
979    fn task_graph_snapshot_from_compiles_with_feature() {
980        use zeph_orchestration::TaskGraph;
981        let graph = TaskGraph::new("feature flag test");
982        let snap = TaskGraphSnapshot::from(&graph);
983        assert_eq!(snap.goal, "feature flag test");
984        assert!(snap.tasks.is_empty());
985        assert!(!snap.is_stale());
986    }
987
988    // T1-extra: long error is truncated with ellipsis.
989    #[test]
990    fn task_graph_snapshot_error_truncated_at_80_chars() {
991        use zeph_orchestration::{TaskGraph, TaskNode, TaskResult, TaskStatus};
992
993        let mut graph = TaskGraph::new("goal");
994        let mut task = TaskNode::new(0, "t", "d");
995        task.status = TaskStatus::Failed;
996        task.result = Some(TaskResult {
997            output: "e".repeat(100),
998            artifacts: vec![],
999            duration_ms: 0,
1000            agent_id: None,
1001            agent_def: None,
1002        });
1003        graph.tasks.push(task);
1004
1005        let snap = TaskGraphSnapshot::from(&graph);
1006        let err = snap.tasks[0].error.as_ref().unwrap();
1007        assert!(err.ends_with('…'), "truncated error must end with ellipsis");
1008        assert!(
1009            err.len() <= 83,
1010            "truncated error must not exceed 80 chars + ellipsis"
1011        );
1012    }
1013
1014    // SEC-P6-01: control chars in task title are stripped.
1015    #[test]
1016    fn task_graph_snapshot_strips_control_chars_from_title() {
1017        use zeph_orchestration::{TaskGraph, TaskNode};
1018
1019        let mut graph = TaskGraph::new("goal\x1b[31m");
1020        let task = TaskNode::new(0, "title\x00injected", "d");
1021        graph.tasks.push(task);
1022
1023        let snap = TaskGraphSnapshot::from(&graph);
1024        assert!(!snap.goal.contains('\x1b'), "goal must not contain escape");
1025        assert!(
1026            !snap.tasks[0].title.contains('\x00'),
1027            "title must not contain null byte"
1028        );
1029    }
1030
1031    #[test]
1032    fn graph_metrics_default_zero() {
1033        let m = MetricsSnapshot::default();
1034        assert_eq!(m.graph_entities_total, 0);
1035        assert_eq!(m.graph_edges_total, 0);
1036        assert_eq!(m.graph_communities_total, 0);
1037        assert_eq!(m.graph_extraction_count, 0);
1038        assert_eq!(m.graph_extraction_failures, 0);
1039    }
1040
1041    #[test]
1042    fn graph_metrics_update_via_collector() {
1043        let (collector, rx) = MetricsCollector::new();
1044        collector.update(|m| {
1045            m.graph_entities_total = 5;
1046            m.graph_edges_total = 10;
1047            m.graph_communities_total = 2;
1048            m.graph_extraction_count = 7;
1049            m.graph_extraction_failures = 1;
1050        });
1051        let snapshot = rx.borrow().clone();
1052        assert_eq!(snapshot.graph_entities_total, 5);
1053        assert_eq!(snapshot.graph_edges_total, 10);
1054        assert_eq!(snapshot.graph_communities_total, 2);
1055        assert_eq!(snapshot.graph_extraction_count, 7);
1056        assert_eq!(snapshot.graph_extraction_failures, 1);
1057    }
1058
1059    #[test]
1060    fn histogram_recorder_trait_is_object_safe() {
1061        use std::sync::Arc;
1062        use std::time::Duration;
1063
1064        struct NoOpRecorder;
1065        impl HistogramRecorder for NoOpRecorder {
1066            fn observe_llm_latency(&self, _: Duration) {}
1067            fn observe_turn_duration(&self, _: Duration) {}
1068            fn observe_tool_execution(&self, _: Duration) {}
1069            fn observe_bg_task(&self, _: &str, _: Duration) {}
1070        }
1071
1072        // Verify the trait can be used as a trait object (object-safe).
1073        let recorder: Arc<dyn HistogramRecorder> = Arc::new(NoOpRecorder);
1074        recorder.observe_llm_latency(Duration::from_millis(500));
1075        recorder.observe_turn_duration(Duration::from_secs(3));
1076        recorder.observe_tool_execution(Duration::from_millis(100));
1077    }
1078}
zeph_core/metrics.rs

zeph_core/
metrics.rs
