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    /// Reasoning tokens from the last turn (`OpenAI` o-series only).
179    ///
180    /// This is a **subset** of `completion_tokens` and must not be added to cost separately.
181    pub reasoning_tokens: u64,
182    pub context_tokens: u64,
183    pub api_calls: u64,
184    pub active_skills: Vec<String>,
185    pub total_skills: usize,
186    /// Total configured MCP servers (connected + failed).
187    pub mcp_server_count: usize,
188    pub mcp_tool_count: usize,
189    /// Number of successfully connected MCP servers.
190    pub mcp_connected_count: usize,
191    /// Per-server connection status list.
192    pub mcp_servers: Vec<McpServerStatus>,
193    pub active_mcp_tools: Vec<String>,
194    pub sqlite_message_count: u64,
195    pub sqlite_conversation_id: Option<zeph_memory::ConversationId>,
196    pub qdrant_available: bool,
197    pub vector_backend: String,
198    pub embeddings_generated: u64,
199    pub last_llm_latency_ms: u64,
200    pub uptime_seconds: u64,
201    pub provider_name: String,
202    pub model_name: String,
203    pub summaries_count: u64,
204    pub context_compactions: u64,
205    /// Number of times the agent entered the Hard compaction tier, including cooldown-skipped
206    /// turns. Not equal to the actual LLM summarization count — reflects pressure, not action.
207    pub compaction_hard_count: u64,
208    /// User-message turns elapsed after each hard compaction event.
209    /// Entry i = turns between hard compaction i and hard compaction i+1 (or session end).
210    /// Empty when no hard compaction occurred during the session.
211    pub compaction_turns_after_hard: Vec<u64>,
212    pub compression_events: u64,
213    pub compression_tokens_saved: u64,
214    pub tool_output_prunes: u64,
215    /// Compaction probe outcomes (#1609).
216    pub compaction_probe_passes: u64,
217    /// Compaction probe soft failures (summary borderline — compaction proceeded with warning).
218    pub compaction_probe_soft_failures: u64,
219    /// Compaction probe hard failures (compaction blocked due to lossy summary).
220    pub compaction_probe_failures: u64,
221    /// Compaction probe errors (LLM/timeout — non-blocking, compaction proceeded).
222    pub compaction_probe_errors: u64,
223    /// Last compaction probe verdict. `None` before the first probe completes.
224    pub last_probe_verdict: Option<zeph_memory::ProbeVerdict>,
225    /// Last compaction probe score in [0.0, 1.0]. `None` before the first probe
226    /// completes or after an Error verdict (errors produce no score).
227    pub last_probe_score: Option<f32>,
228    /// Per-category scores from the last completed probe.
229    pub last_probe_category_scores: Option<Vec<zeph_memory::CategoryScore>>,
230    /// Configured pass threshold for the compaction probe. Used by TUI for category color-coding.
231    pub compaction_probe_threshold: f32,
232    /// Configured hard-fail threshold for the compaction probe.
233    pub compaction_probe_hard_fail_threshold: f32,
234    pub cache_read_tokens: u64,
235    pub cache_creation_tokens: u64,
236    pub cost_spent_cents: f64,
237    /// Cost per successful task in cents. `None` until at least one task completes.
238    pub cost_cps_cents: Option<f64>,
239    /// Number of successful tasks recorded today.
240    pub cost_successful_tasks: u64,
241    /// Per-provider cost breakdown, sorted by cost descending.
242    pub provider_cost_breakdown: Vec<(String, crate::cost::ProviderUsage)>,
243    pub filter_raw_tokens: u64,
244    pub filter_saved_tokens: u64,
245    pub filter_applications: u64,
246    pub filter_total_commands: u64,
247    pub filter_filtered_commands: u64,
248    pub filter_confidence_full: u64,
249    pub filter_confidence_partial: u64,
250    pub filter_confidence_fallback: u64,
251    pub cancellations: u64,
252    pub server_compaction_events: u64,
253    pub sanitizer_runs: u64,
254    pub sanitizer_injection_flags: u64,
255    /// Injection pattern hits on `ToolResult` (local) sources — likely false positives.
256    ///
257    /// Counts regex hits that fired on content from `shell`, `read_file`, `search_code`, etc.
258    /// These sources are user-owned and not adversarial; a non-zero value indicates a pattern
259    /// that needs tightening or a source reclassification.
260    pub sanitizer_injection_fp_local: u64,
261    pub sanitizer_truncations: u64,
262    pub quarantine_invocations: u64,
263    pub quarantine_failures: u64,
264    /// ML classifier hard-blocked tool outputs (`enforcement_mode=block` only).
265    pub classifier_tool_blocks: u64,
266    /// ML classifier suspicious tool outputs (both enforcement modes).
267    pub classifier_tool_suspicious: u64,
268    /// `TurnCausalAnalyzer` flags: behavioral deviation detected at tool-return boundary.
269    pub causal_ipi_flags: u64,
270    /// VIGIL pre-sanitizer flags: tool outputs matched injection patterns (any action).
271    pub vigil_flags_total: u64,
272    /// VIGIL pre-sanitizer blocks: tool outputs replaced with sentinel (`strict_mode=true`).
273    pub vigil_blocks_total: u64,
274    pub exfiltration_images_blocked: u64,
275    pub exfiltration_tool_urls_flagged: u64,
276    pub exfiltration_memory_guards: u64,
277    pub pii_scrub_count: u64,
278    /// Number of times the PII NER classifier timed out; input fell back to regex-only.
279    pub pii_ner_timeouts: u64,
280    /// Number of times the PII NER circuit breaker tripped (disabled NER for the session).
281    pub pii_ner_circuit_breaker_trips: u64,
282    pub memory_validation_failures: u64,
283    pub rate_limit_trips: u64,
284    pub pre_execution_blocks: u64,
285    pub pre_execution_warnings: u64,
286    /// `true` when a guardrail filter is active for this session.
287    pub guardrail_enabled: bool,
288    /// `true` when guardrail is in warn-only mode (action = warn).
289    pub guardrail_warn_mode: bool,
290    pub sub_agents: Vec<SubAgentMetrics>,
291    pub skill_confidence: Vec<SkillConfidence>,
292    /// Scheduled task summaries: `[name, kind, mode, next_run]`.
293    pub scheduled_tasks: Vec<[String; 4]>,
294    /// Thompson Sampling distribution snapshots: `(provider, alpha, beta)`.
295    pub router_thompson_stats: Vec<(String, f64, f64)>,
296    /// Ring buffer of recent security events (cap 100, FIFO eviction).
297    pub security_events: VecDeque<SecurityEvent>,
298    pub orchestration: OrchestrationMetrics,
299    /// Live snapshot of the currently active task graph. `None` when no plan is active.
300    pub orchestration_graph: Option<TaskGraphSnapshot>,
301    pub graph_community_detection_failures: u64,
302    pub graph_entities_total: u64,
303    pub graph_edges_total: u64,
304    pub graph_communities_total: u64,
305    pub graph_extraction_count: u64,
306    pub graph_extraction_failures: u64,
307    /// `true` when `config.llm.cloud.enable_extended_context = true`.
308    /// Never set for other providers to avoid false positives.
309    pub extended_context: bool,
310    /// Latest compression-guidelines version (0 = no guidelines yet).
311    pub guidelines_version: u32,
312    /// ISO 8601 timestamp of the latest guidelines update (empty if none).
313    pub guidelines_updated_at: String,
314    pub tool_cache_hits: u64,
315    pub tool_cache_misses: u64,
316    pub tool_cache_entries: usize,
317    /// Number of semantic-tier facts in memory (0 when tier promotion disabled).
318    pub semantic_fact_count: u64,
319    /// STT model name (e.g. "whisper-1"). `None` when STT is not configured.
320    pub stt_model: Option<String>,
321    /// Model used for context compaction/summarization. `None` when no summary provider is set.
322    pub compaction_model: Option<String>,
323    /// Temperature of the active provider when using Candle. `None` for API providers.
324    pub provider_temperature: Option<f32>,
325    /// Top-p of the active provider when using Candle. `None` for API providers.
326    pub provider_top_p: Option<f32>,
327    /// Embedding model name (e.g. `"nomic-embed-text"`). Empty when embeddings are disabled.
328    pub embedding_model: String,
329    /// Token budget for context window. `None` when not configured.
330    pub token_budget: Option<u64>,
331    /// Token threshold that triggers soft compaction. `None` when not configured.
332    pub compaction_threshold: Option<u32>,
333    /// Vault backend identifier: "age", "env", or "none".
334    pub vault_backend: String,
335    /// Active I/O channel name: `"cli"`, `"telegram"`, `"tui"`, `"discord"`, `"slack"`.
336    pub active_channel: String,
337    /// Background supervisor: inflight tasks across all classes.
338    pub bg_inflight: u64,
339    /// Background supervisor: total tasks dropped due to concurrency limit (all classes).
340    pub bg_dropped: u64,
341    /// Background supervisor: total tasks completed (all classes).
342    pub bg_completed: u64,
343    /// Background supervisor: inflight enrichment tasks.
344    pub bg_enrichment_inflight: u64,
345    /// Background supervisor: inflight telemetry tasks.
346    pub bg_telemetry_inflight: u64,
347    /// In-flight background shell runs. Empty when none are running or no `ShellExecutor` is wired.
348    pub shell_background_runs: Vec<ShellBackgroundRunRow>,
349    /// Whether self-learning (skill evolution) is enabled.
350    pub self_learning_enabled: bool,
351    /// Whether the semantic response cache is enabled.
352    pub semantic_cache_enabled: bool,
353    /// Whether semantic response caching is enabled (alias for `semantic_cache_enabled`).
354    pub cache_enabled: bool,
355    /// Whether assistant messages are auto-saved to memory.
356    pub autosave_enabled: bool,
357    /// Classifier p50/p95 latency metrics per task (injection, pii, feedback).
358    pub classifier: ClassifierMetricsSnapshot,
359    /// Latency breakdown for the most recently completed agent turn.
360    pub last_turn_timings: TurnTimings,
361    /// Rolling average of per-phase latency over the last 10 turns.
362    pub avg_turn_timings: TurnTimings,
363    /// Maximum per-phase latency observed within the rolling window (tail-latency visibility).
364    ///
365    /// M3: exposes `max_in_window` alongside the rolling average for operational monitoring.
366    pub max_turn_timings: TurnTimings,
367    /// Number of turns included in `avg_turn_timings` and `max_turn_timings` (capped at 10).
368    pub timing_sample_count: u64,
369    /// Total egress (outbound HTTP) requests attempted this session.
370    pub egress_requests_total: u64,
371    /// Egress events dropped due to bounded channel backpressure.
372    pub egress_dropped_total: u64,
373    /// Egress requests blocked by scheme/domain/SSRF policy.
374    pub egress_blocked_total: u64,
375    /// Runtime-resolved context window limit (tokens).
376    ///
377    /// Populated from `resolve_context_budget` after provider pool construction and refreshed on
378    /// every `/provider` switch. `0` means unknown (pre-init or provider has no declared window);
379    /// the TUI gauge renders `"—"` in this case to avoid divide-by-zero.
380    pub context_max_tokens: u64,
381    /// Token count at the time the most recent compaction was triggered. `0` = never compacted.
382    pub compaction_last_before: u64,
383    /// Token count after the most recent compaction completed. `0` = never compacted.
384    pub compaction_last_after: u64,
385    /// Unix epoch milliseconds when the most recent compaction occurred. `0` = never compacted.
386    pub compaction_last_at_ms: u64,
387    /// Active long-horizon goal for TUI display. `None` when no goal is active.
388    pub active_goal: Option<crate::goal::GoalSnapshot>,
389    /// Cocoon sidecar connection state. `None` when Cocoon is not configured.
390    /// `Some(true)` = proxy connected, `Some(false)` = unreachable or disconnected.
391    pub cocoon_connected: Option<bool>,
392    /// Worker count reported by the Cocoon sidecar. `0` when not connected or not configured.
393    pub cocoon_worker_count: u32,
394    /// Number of models available through the Cocoon sidecar.
395    pub cocoon_model_count: usize,
396    /// TON wallet balance in TON units. `None` when unknown or Cocoon not configured.
397    pub cocoon_ton_balance: Option<f64>,
398}
399
400/// Snapshot of a single in-flight background shell run for TUI display.
401///
402/// Populated from [`zeph_tools::BackgroundRunSnapshot`] during the per-turn metrics update in
403/// `reap_background_tasks_and_update_metrics`. The `run_id` is truncated to 8 hex chars for
404/// compact TUI rendering.
405#[derive(Debug, Clone, Default, serde::Serialize)]
406pub struct ShellBackgroundRunRow {
407    /// First 8 hex characters of the run's UUID, sufficient for unique identification in TUI.
408    pub run_id: String,
409    /// Original command, truncated to 80 characters.
410    pub command: String,
411    /// Wall-clock elapsed seconds since spawn.
412    pub elapsed_secs: u64,
413}
414
415/// Configuration-derived fields of [`MetricsSnapshot`] that are known at agent startup and do
416/// not change during the session.
417///
418/// Pass this struct to `AgentBuilder::with_static_metrics` immediately after
419/// `AgentBuilder::with_metrics` to initialize all static fields in one place rather than
420/// through scattered `send_modify` calls in the runner.
421///
422/// # Examples
423///
424/// ```no_run
425/// use zeph_core::metrics::StaticMetricsInit;
426///
427/// let init = StaticMetricsInit {
428///     active_channel: "cli".to_owned(),
429///     ..StaticMetricsInit::default()
430/// };
431/// ```
432#[derive(Debug, Default)]
433pub struct StaticMetricsInit {
434    /// STT model name (e.g. `"whisper-1"`). `None` when STT is not configured.
435    pub stt_model: Option<String>,
436    /// Model used for context compaction/summarization. `None` when no summary provider is set.
437    pub compaction_model: Option<String>,
438    /// Whether the semantic response cache is enabled.
439    ///
440    /// This value is also written to [`MetricsSnapshot::cache_enabled`] which is an alias for the
441    /// same concept.
442    pub semantic_cache_enabled: bool,
443    /// Embedding model name (e.g. `"nomic-embed-text"`). Empty when embeddings are disabled.
444    pub embedding_model: String,
445    /// Whether self-learning (skill evolution) is enabled.
446    pub self_learning_enabled: bool,
447    /// Active I/O channel name: `"cli"`, `"telegram"`, `"tui"`, `"discord"`, `"slack"`.
448    pub active_channel: String,
449    /// Token budget for context window. `None` when not configured.
450    pub token_budget: Option<u64>,
451    /// Token threshold that triggers soft compaction. `None` when not configured.
452    pub compaction_threshold: Option<u32>,
453    /// Vault backend identifier: `"age"`, `"env"`, or `"none"`.
454    pub vault_backend: String,
455    /// Whether assistant messages are auto-saved to memory.
456    pub autosave_enabled: bool,
457    /// Override for the active model name. When `Some`, replaces the model name set by the
458    /// builder from `runtime.model_name` (which may be a placeholder) with the effective model
459    /// resolved from the LLM provider configuration.
460    pub model_name_override: Option<String>,
461}
462
463/// Strip ASCII control characters and ANSI escape sequences from a string for safe TUI display.
464///
465/// Allows tab, LF, and CR; removes everything else in the `0x00–0x1F` range including full
466/// ANSI CSI sequences (`ESC[...`). This prevents escape-sequence injection from LLM planner
467/// output into the TUI.
468fn strip_ctrl(s: &str) -> String {
469    let mut out = String::with_capacity(s.len());
470    let mut chars = s.chars().peekable();
471    while let Some(c) = chars.next() {
472        if c == '\x1b' {
473            // Consume an ANSI CSI sequence: ESC [ <params> <final-byte in 0x40–0x7E>
474            if chars.peek() == Some(&'[') {
475                chars.next(); // consume '['
476                for inner in chars.by_ref() {
477                    if ('\x40'..='\x7e').contains(&inner) {
478                        break;
479                    }
480                }
481            }
482            // Drop ESC and any consumed sequence — write nothing.
483        } else if c.is_control() && c != '\t' && c != '\n' && c != '\r' {
484            // drop other control chars
485        } else {
486            out.push(c);
487        }
488    }
489    out
490}
491
492/// Convert a live `TaskGraph` into a lightweight snapshot for TUI display.
493impl From<&zeph_orchestration::TaskGraph> for TaskGraphSnapshot {
494    fn from(graph: &zeph_orchestration::TaskGraph) -> Self {
495        let tasks = graph
496            .tasks
497            .iter()
498            .map(|t| {
499                let error = t
500                    .result
501                    .as_ref()
502                    .filter(|_| t.status == zeph_orchestration::TaskStatus::Failed)
503                    .and_then(|r| {
504                        if r.output.is_empty() {
505                            None
506                        } else {
507                            // Strip control chars, then truncate at 80 chars (SEC-P6-01).
508                            let s = strip_ctrl(&r.output);
509                            if s.len() > 80 {
510                                let end = s.floor_char_boundary(79);
511                                Some(format!("{}…", &s[..end]))
512                            } else {
513                                Some(s)
514                            }
515                        }
516                    });
517                let duration_ms = t.result.as_ref().map_or(0, |r| r.duration_ms);
518                TaskSnapshotRow {
519                    id: t.id.as_u32(),
520                    title: strip_ctrl(&t.title),
521                    status: t.status.to_string(),
522                    agent: t.assigned_agent.as_deref().map(strip_ctrl),
523                    duration_ms,
524                    error,
525                }
526            })
527            .collect();
528        Self {
529            graph_id: graph.id.to_string(),
530            goal: strip_ctrl(&graph.goal),
531            status: graph.status.to_string(),
532            tasks,
533            completed_at: None,
534        }
535    }
536}
537
538pub struct MetricsCollector {
539    tx: watch::Sender<MetricsSnapshot>,
540}
541
542impl MetricsCollector {
543    #[must_use]
544    pub fn new() -> (Self, watch::Receiver<MetricsSnapshot>) {
545        let (tx, rx) = watch::channel(MetricsSnapshot::default());
546        (Self { tx }, rx)
547    }
548
549    pub fn update(&self, f: impl FnOnce(&mut MetricsSnapshot)) {
550        self.tx.send_modify(f);
551    }
552
553    /// Publish the runtime-resolved context window limit.
554    ///
555    /// Call after the provider pool is constructed (builder) and on every successful
556    /// `/provider` switch so the TUI context gauge reflects the active provider's window.
557    ///
558    /// # Examples
559    ///
560    /// ```rust
561    /// use zeph_core::metrics::MetricsCollector;
562    ///
563    /// let (collector, rx) = MetricsCollector::new();
564    /// collector.set_context_max_tokens(128_000);
565    /// assert_eq!(rx.borrow().context_max_tokens, 128_000);
566    /// ```
567    pub fn set_context_max_tokens(&self, max_tokens: u64) {
568        self.tx.send_modify(|m| m.context_max_tokens = max_tokens);
569    }
570
571    /// Record the outcome of the most recent compaction event.
572    ///
573    /// Sets all three `compaction_last_*` fields atomically. `at_ms` is the Unix epoch in
574    /// milliseconds — obtain via `SystemTime::UNIX_EPOCH.elapsed().unwrap_or_default().as_millis()`.
575    ///
576    /// # Examples
577    ///
578    /// ```rust
579    /// use zeph_core::metrics::MetricsCollector;
580    ///
581    /// let (collector, rx) = MetricsCollector::new();
582    /// collector.record_compaction(50_000, 12_000, 1_700_000_000_000);
583    /// let snap = rx.borrow();
584    /// assert_eq!(snap.compaction_last_before, 50_000);
585    /// assert_eq!(snap.compaction_last_after, 12_000);
586    /// assert_eq!(snap.compaction_last_at_ms, 1_700_000_000_000);
587    /// ```
588    pub fn record_compaction(&self, before: u64, after: u64, at_ms: u64) {
589        self.tx.send_modify(|m| {
590            m.compaction_last_before = before;
591            m.compaction_last_after = after;
592            m.compaction_last_at_ms = at_ms;
593        });
594    }
595
596    /// Returns a clone of the underlying [`watch::Sender`].
597    ///
598    /// Use this to pass the sender to code that requires a raw
599    /// `watch::Sender<MetricsSnapshot>` while the [`MetricsCollector`] is
600    /// also shared (e.g., passed to a `MetricsBridge` layer).
601    #[must_use]
602    pub fn sender(&self) -> watch::Sender<MetricsSnapshot> {
603        self.tx.clone()
604    }
605}
606
607// ---------------------------------------------------------------------------
608// HistogramRecorder
609// ---------------------------------------------------------------------------
610
611/// Per-event histogram recording contract for the agent loop.
612///
613/// Implementors record individual latency observations into Prometheus histograms
614/// (or any other backend). The trait is object-safe: the agent stores an
615/// `Option<Arc<dyn HistogramRecorder>>` and calls these methods at each measurement
616/// point. When `None`, recording is a no-op with zero overhead.
617///
618/// # Contract for implementors
619///
620/// - All methods must be non-blocking; they must not call async code or acquire
621///   mutexes that may block.
622/// - Implementations must be `Send + Sync` — the agent loop runs on the tokio
623///   thread pool and the recorder may be called from multiple tasks.
624///
625/// # Examples
626///
627/// ```rust
628/// use std::sync::Arc;
629/// use std::time::Duration;
630/// use zeph_core::metrics::HistogramRecorder;
631///
632/// struct NoOpRecorder;
633///
634/// impl HistogramRecorder for NoOpRecorder {
635///     fn observe_llm_latency(&self, _: Duration) {}
636///     fn observe_turn_duration(&self, _: Duration) {}
637///     fn observe_tool_execution(&self, _: Duration) {}
638///     fn observe_bg_task(&self, _: &str, _: Duration) {}
639/// }
640///
641/// let recorder: Arc<dyn HistogramRecorder> = Arc::new(NoOpRecorder);
642/// recorder.observe_llm_latency(Duration::from_millis(500));
643/// ```
644pub trait HistogramRecorder: Send + Sync {
645    /// Record a single LLM API call latency observation.
646    fn observe_llm_latency(&self, duration: std::time::Duration);
647
648    /// Record a full agent turn duration observation (context prep + LLM + tools + persist).
649    fn observe_turn_duration(&self, duration: std::time::Duration);
650
651    /// Record a single tool execution latency observation.
652    fn observe_tool_execution(&self, duration: std::time::Duration);
653
654    /// Record a background task completion latency.
655    ///
656    /// `class_label` is `"enrichment"` or `"telemetry"` (from `TaskClass::name()`).
657    fn observe_bg_task(&self, class_label: &str, duration: std::time::Duration);
658}
659
660#[cfg(test)]
661mod tests {
662    #![allow(clippy::field_reassign_with_default)]
663
664    use super::*;
665
666    #[test]
667    fn default_metrics_snapshot() {
668        let m = MetricsSnapshot::default();
669        assert_eq!(m.total_tokens, 0);
670        assert_eq!(m.api_calls, 0);
671        assert!(m.active_skills.is_empty());
672        assert!(m.active_mcp_tools.is_empty());
673        assert_eq!(m.mcp_tool_count, 0);
674        assert_eq!(m.mcp_server_count, 0);
675        assert!(m.provider_name.is_empty());
676        assert_eq!(m.summaries_count, 0);
677        // Phase 2 fields
678        assert!(m.stt_model.is_none());
679        assert!(m.compaction_model.is_none());
680        assert!(m.provider_temperature.is_none());
681        assert!(m.provider_top_p.is_none());
682        assert!(m.active_channel.is_empty());
683        assert!(m.embedding_model.is_empty());
684        assert!(m.token_budget.is_none());
685        assert!(!m.self_learning_enabled);
686        assert!(!m.semantic_cache_enabled);
687    }
688
689    #[test]
690    fn metrics_collector_update_phase2_fields() {
691        let (collector, rx) = MetricsCollector::new();
692        collector.update(|m| {
693            m.stt_model = Some("whisper-1".into());
694            m.compaction_model = Some("haiku".into());
695            m.provider_temperature = Some(0.7);
696            m.provider_top_p = Some(0.95);
697            m.active_channel = "tui".into();
698            m.embedding_model = "nomic-embed-text".into();
699            m.token_budget = Some(200_000);
700            m.self_learning_enabled = true;
701            m.semantic_cache_enabled = true;
702        });
703        let s = rx.borrow();
704        assert_eq!(s.stt_model.as_deref(), Some("whisper-1"));
705        assert_eq!(s.compaction_model.as_deref(), Some("haiku"));
706        assert_eq!(s.provider_temperature, Some(0.7));
707        assert_eq!(s.provider_top_p, Some(0.95));
708        assert_eq!(s.active_channel, "tui");
709        assert_eq!(s.embedding_model, "nomic-embed-text");
710        assert_eq!(s.token_budget, Some(200_000));
711        assert!(s.self_learning_enabled);
712        assert!(s.semantic_cache_enabled);
713    }
714
715    #[test]
716    fn metrics_collector_update() {
717        let (collector, rx) = MetricsCollector::new();
718        collector.update(|m| {
719            m.api_calls = 5;
720            m.total_tokens = 1000;
721        });
722        let snapshot = rx.borrow().clone();
723        assert_eq!(snapshot.api_calls, 5);
724        assert_eq!(snapshot.total_tokens, 1000);
725    }
726
727    #[test]
728    fn metrics_collector_multiple_updates() {
729        let (collector, rx) = MetricsCollector::new();
730        collector.update(|m| m.api_calls = 1);
731        collector.update(|m| m.api_calls += 1);
732        assert_eq!(rx.borrow().api_calls, 2);
733    }
734
735    #[test]
736    fn metrics_snapshot_clone() {
737        let mut m = MetricsSnapshot::default();
738        m.provider_name = "ollama".into();
739        let cloned = m.clone();
740        assert_eq!(cloned.provider_name, "ollama");
741    }
742
743    #[test]
744    fn filter_metrics_tracking() {
745        let (collector, rx) = MetricsCollector::new();
746        collector.update(|m| {
747            m.filter_raw_tokens += 250;
748            m.filter_saved_tokens += 200;
749            m.filter_applications += 1;
750        });
751        collector.update(|m| {
752            m.filter_raw_tokens += 100;
753            m.filter_saved_tokens += 80;
754            m.filter_applications += 1;
755        });
756        let s = rx.borrow();
757        assert_eq!(s.filter_raw_tokens, 350);
758        assert_eq!(s.filter_saved_tokens, 280);
759        assert_eq!(s.filter_applications, 2);
760    }
761
762    #[test]
763    fn filter_confidence_and_command_metrics() {
764        let (collector, rx) = MetricsCollector::new();
765        collector.update(|m| {
766            m.filter_total_commands += 1;
767            m.filter_filtered_commands += 1;
768            m.filter_confidence_full += 1;
769        });
770        collector.update(|m| {
771            m.filter_total_commands += 1;
772            m.filter_confidence_partial += 1;
773        });
774        let s = rx.borrow();
775        assert_eq!(s.filter_total_commands, 2);
776        assert_eq!(s.filter_filtered_commands, 1);
777        assert_eq!(s.filter_confidence_full, 1);
778        assert_eq!(s.filter_confidence_partial, 1);
779        assert_eq!(s.filter_confidence_fallback, 0);
780    }
781
782    #[test]
783    fn summaries_count_tracks_summarizations() {
784        let (collector, rx) = MetricsCollector::new();
785        collector.update(|m| m.summaries_count += 1);
786        collector.update(|m| m.summaries_count += 1);
787        assert_eq!(rx.borrow().summaries_count, 2);
788    }
789
790    #[test]
791    fn cancellations_counter_increments() {
792        let (collector, rx) = MetricsCollector::new();
793        assert_eq!(rx.borrow().cancellations, 0);
794        collector.update(|m| m.cancellations += 1);
795        collector.update(|m| m.cancellations += 1);
796        assert_eq!(rx.borrow().cancellations, 2);
797    }
798
799    #[test]
800    fn security_event_detail_exact_128_not_truncated() {
801        let s = "a".repeat(128);
802        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, "src", s.clone());
803        assert_eq!(ev.detail, s, "128-char string must not be truncated");
804    }
805
806    #[test]
807    fn security_event_detail_129_is_truncated() {
808        let s = "a".repeat(129);
809        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, "src", s);
810        assert!(
811            ev.detail.ends_with('…'),
812            "129-char string must end with ellipsis"
813        );
814        assert!(
815            ev.detail.len() <= 130,
816            "truncated detail must be at most 130 bytes"
817        );
818    }
819
820    #[test]
821    fn security_event_detail_multibyte_utf8_no_panic() {
822        // Each '中' is 3 bytes. 43 chars = 129 bytes — triggers truncation at a multi-byte boundary.
823        let s = "中".repeat(43);
824        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, "src", s);
825        assert!(ev.detail.ends_with('…'));
826    }
827
828    #[test]
829    fn security_event_source_capped_at_64_chars() {
830        let long_source = "x".repeat(200);
831        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, long_source, "detail");
832        assert_eq!(ev.source.len(), 64);
833    }
834
835    #[test]
836    fn security_event_source_strips_control_chars() {
837        let source = "tool\x00name\x1b[31m";
838        let ev = SecurityEvent::new(SecurityEventCategory::InjectionFlag, source, "detail");
839        assert!(!ev.source.contains('\x00'));
840        assert!(!ev.source.contains('\x1b'));
841    }
842
843    #[test]
844    fn security_event_category_as_str() {
845        assert_eq!(SecurityEventCategory::InjectionFlag.as_str(), "injection");
846        assert_eq!(SecurityEventCategory::ExfiltrationBlock.as_str(), "exfil");
847        assert_eq!(SecurityEventCategory::Quarantine.as_str(), "quarantine");
848        assert_eq!(SecurityEventCategory::Truncation.as_str(), "truncation");
849        assert_eq!(
850            SecurityEventCategory::CrossBoundaryMcpToAcp.as_str(),
851            "cross_boundary_mcp_to_acp"
852        );
853    }
854
855    #[test]
856    fn ring_buffer_respects_cap_via_update() {
857        let (collector, rx) = MetricsCollector::new();
858        for i in 0..110u64 {
859            let event = SecurityEvent::new(
860                SecurityEventCategory::InjectionFlag,
861                "src",
862                format!("event {i}"),
863            );
864            collector.update(|m| {
865                if m.security_events.len() >= SECURITY_EVENT_CAP {
866                    m.security_events.pop_front();
867                }
868                m.security_events.push_back(event);
869            });
870        }
871        let snap = rx.borrow();
872        assert_eq!(snap.security_events.len(), SECURITY_EVENT_CAP);
873        // FIFO: earliest events evicted, last one present
874        assert!(snap.security_events.back().unwrap().detail.contains("109"));
875    }
876
877    #[test]
878    fn security_events_empty_by_default() {
879        let m = MetricsSnapshot::default();
880        assert!(m.security_events.is_empty());
881    }
882
883    #[test]
884    fn orchestration_metrics_default_zero() {
885        let m = OrchestrationMetrics::default();
886        assert_eq!(m.plans_total, 0);
887        assert_eq!(m.tasks_total, 0);
888        assert_eq!(m.tasks_completed, 0);
889        assert_eq!(m.tasks_failed, 0);
890        assert_eq!(m.tasks_skipped, 0);
891    }
892
893    #[test]
894    fn metrics_snapshot_includes_orchestration_default_zero() {
895        let m = MetricsSnapshot::default();
896        assert_eq!(m.orchestration.plans_total, 0);
897        assert_eq!(m.orchestration.tasks_total, 0);
898        assert_eq!(m.orchestration.tasks_completed, 0);
899    }
900
901    #[test]
902    fn orchestration_metrics_update_via_collector() {
903        let (collector, rx) = MetricsCollector::new();
904        collector.update(|m| {
905            m.orchestration.plans_total += 1;
906            m.orchestration.tasks_total += 5;
907            m.orchestration.tasks_completed += 3;
908            m.orchestration.tasks_failed += 1;
909            m.orchestration.tasks_skipped += 1;
910        });
911        let s = rx.borrow();
912        assert_eq!(s.orchestration.plans_total, 1);
913        assert_eq!(s.orchestration.tasks_total, 5);
914        assert_eq!(s.orchestration.tasks_completed, 3);
915        assert_eq!(s.orchestration.tasks_failed, 1);
916        assert_eq!(s.orchestration.tasks_skipped, 1);
917    }
918
919    #[test]
920    fn strip_ctrl_removes_escape_sequences() {
921        let input = "hello\x1b[31mworld\x00end";
922        let result = strip_ctrl(input);
923        assert_eq!(result, "helloworldend");
924    }
925
926    #[test]
927    fn strip_ctrl_allows_tab_lf_cr() {
928        let input = "a\tb\nc\rd";
929        let result = strip_ctrl(input);
930        assert_eq!(result, "a\tb\nc\rd");
931    }
932
933    #[test]
934    fn task_graph_snapshot_is_stale_after_30s() {
935        let mut snap = TaskGraphSnapshot::default();
936        // Not stale if no completed_at.
937        assert!(!snap.is_stale());
938        // Not stale if just completed.
939        snap.completed_at = Some(std::time::Instant::now());
940        assert!(!snap.is_stale());
941        // Stale if completed more than 30s ago.
942        snap.completed_at = Some(
943            std::time::Instant::now()
944                .checked_sub(std::time::Duration::from_secs(31))
945                .unwrap(),
946        );
947        assert!(snap.is_stale());
948    }
949
950    // T1: From<&TaskGraph> correctly maps fields including duration_ms and error truncation.
951    #[test]
952    fn task_graph_snapshot_from_task_graph_maps_fields() {
953        use zeph_orchestration::{GraphStatus, TaskGraph, TaskNode, TaskResult, TaskStatus};
954
955        let mut graph = TaskGraph::new("My goal");
956        let mut task = TaskNode::new(0, "Do work", "description");
957        task.status = TaskStatus::Failed;
958        task.assigned_agent = Some("agent-1".into());
959        task.result = Some(TaskResult {
960            output: "error occurred here".into(),
961            artifacts: vec![],
962            duration_ms: 1234,
963            agent_id: None,
964            agent_def: None,
965        });
966        graph.tasks.push(task);
967        graph.status = GraphStatus::Failed;
968
969        let snap = TaskGraphSnapshot::from(&graph);
970        assert_eq!(snap.goal, "My goal");
971        assert_eq!(snap.status, "failed");
972        assert_eq!(snap.tasks.len(), 1);
973        let row = &snap.tasks[0];
974        assert_eq!(row.title, "Do work");
975        assert_eq!(row.status, "failed");
976        assert_eq!(row.agent.as_deref(), Some("agent-1"));
977        assert_eq!(row.duration_ms, 1234);
978        assert!(row.error.as_deref().unwrap().contains("error occurred"));
979    }
980
981    // T2: From impl compiles with orchestration feature active.
982    #[test]
983    fn task_graph_snapshot_from_compiles_with_feature() {
984        use zeph_orchestration::TaskGraph;
985        let graph = TaskGraph::new("feature flag test");
986        let snap = TaskGraphSnapshot::from(&graph);
987        assert_eq!(snap.goal, "feature flag test");
988        assert!(snap.tasks.is_empty());
989        assert!(!snap.is_stale());
990    }
991
992    // T1-extra: long error is truncated with ellipsis.
993    #[test]
994    fn task_graph_snapshot_error_truncated_at_80_chars() {
995        use zeph_orchestration::{TaskGraph, TaskNode, TaskResult, TaskStatus};
996
997        let mut graph = TaskGraph::new("goal");
998        let mut task = TaskNode::new(0, "t", "d");
999        task.status = TaskStatus::Failed;
1000        task.result = Some(TaskResult {
1001            output: "e".repeat(100),
1002            artifacts: vec![],
1003            duration_ms: 0,
1004            agent_id: None,
1005            agent_def: None,
1006        });
1007        graph.tasks.push(task);
1008
1009        let snap = TaskGraphSnapshot::from(&graph);
1010        let err = snap.tasks[0].error.as_ref().unwrap();
1011        assert!(err.ends_with('…'), "truncated error must end with ellipsis");
1012        assert!(
1013            err.len() <= 83,
1014            "truncated error must not exceed 80 chars + ellipsis"
1015        );
1016    }
1017
1018    // SEC-P6-01: control chars in task title are stripped.
1019    #[test]
1020    fn task_graph_snapshot_strips_control_chars_from_title() {
1021        use zeph_orchestration::{TaskGraph, TaskNode};
1022
1023        let mut graph = TaskGraph::new("goal\x1b[31m");
1024        let task = TaskNode::new(0, "title\x00injected", "d");
1025        graph.tasks.push(task);
1026
1027        let snap = TaskGraphSnapshot::from(&graph);
1028        assert!(!snap.goal.contains('\x1b'), "goal must not contain escape");
1029        assert!(
1030            !snap.tasks[0].title.contains('\x00'),
1031            "title must not contain null byte"
1032        );
1033    }
1034
1035    #[test]
1036    fn graph_metrics_default_zero() {
1037        let m = MetricsSnapshot::default();
1038        assert_eq!(m.graph_entities_total, 0);
1039        assert_eq!(m.graph_edges_total, 0);
1040        assert_eq!(m.graph_communities_total, 0);
1041        assert_eq!(m.graph_extraction_count, 0);
1042        assert_eq!(m.graph_extraction_failures, 0);
1043    }
1044
1045    #[test]
1046    fn graph_metrics_update_via_collector() {
1047        let (collector, rx) = MetricsCollector::new();
1048        collector.update(|m| {
1049            m.graph_entities_total = 5;
1050            m.graph_edges_total = 10;
1051            m.graph_communities_total = 2;
1052            m.graph_extraction_count = 7;
1053            m.graph_extraction_failures = 1;
1054        });
1055        let snapshot = rx.borrow().clone();
1056        assert_eq!(snapshot.graph_entities_total, 5);
1057        assert_eq!(snapshot.graph_edges_total, 10);
1058        assert_eq!(snapshot.graph_communities_total, 2);
1059        assert_eq!(snapshot.graph_extraction_count, 7);
1060        assert_eq!(snapshot.graph_extraction_failures, 1);
1061    }
1062
1063    #[test]
1064    fn histogram_recorder_trait_is_object_safe() {
1065        use std::sync::Arc;
1066        use std::time::Duration;
1067
1068        struct NoOpRecorder;
1069        impl HistogramRecorder for NoOpRecorder {
1070            fn observe_llm_latency(&self, _: Duration) {}
1071            fn observe_turn_duration(&self, _: Duration) {}
1072            fn observe_tool_execution(&self, _: Duration) {}
1073            fn observe_bg_task(&self, _: &str, _: Duration) {}
1074        }
1075
1076        // Verify the trait can be used as a trait object (object-safe).
1077        let recorder: Arc<dyn HistogramRecorder> = Arc::new(NoOpRecorder);
1078        recorder.observe_llm_latency(Duration::from_millis(500));
1079        recorder.observe_turn_duration(Duration::from_secs(3));
1080        recorder.observe_tool_execution(Duration::from_millis(100));
1081    }
1082}
zeph_core/metrics.rs

zeph_core/
metrics.rs
