1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
/// Agent state management
use crate::types::{StopReason, ToolResult};
use oxi_ai::{ContentBlock, Message, TextContent};
use parking_lot::RwLock;
use serde::{Deserialize, Serialize};
use std::sync::Arc;
/// Agent execution state
///
/// Tracks the full lifecycle of an agent conversation including messages,
/// token usage, tool results, and iteration progress.
///
/// Derives `Serialize`/`Deserialize` for session persistence and
/// cross-process state transfer (e.g. oxios supervisor serialization).
#[derive(Debug, Clone, Serialize, Deserialize, Default)]
pub struct AgentState {
/// Conversation message history (user, assistant, and tool-result messages).
pub messages: Vec<Message>,
/// Current agent loop iteration (incremented after each assistant turn).
pub iteration: usize,
/// The reason the last turn stopped, if any.
pub stop_reason: Option<StopReason>,
/// Accumulated results from tool executions in the current conversation.
pub tool_results: Vec<ToolResult>,
/// Cumulative token count (input + output) across all turns.
pub total_tokens: usize,
/// Cumulative prompt / input tokens across all turns.
pub input_tokens: usize,
/// Cumulative completion / output tokens across all turns.
pub output_tokens: usize,
/// **Most-recent** reported input-token count from a single LLM response.
///
/// Unlike [`Self::input_tokens`] (which is cumulative across all turns),
/// this is overwritten on every `ProviderEvent::Done` with the count
/// that turn actually sent to the model. It is the **ground-truth** signal
/// used by compaction when available — see [`Self::current_token_source`].
///
/// `None` until the first `Done` event has been observed.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub last_input_tokens: Option<usize>,
/// Heuristic token estimate (`bytes/4`) that was current when
/// `last_input_tokens` was last set. Used to detect the bytes/4 drift
/// described in #28 and surface it as a warning.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub last_estimate_at_report: Option<usize>,
/// Divergence factor (reported / estimate) at the time of the last
/// `Done`. Surfaced in logs so the operator can see how badly the
/// `bytes/4` heuristic is undercounting on token-dense workloads.
/// `None` until first divergence observation.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub last_estimate_divergence: Option<f64>,
}
/// Source of the value driving compaction / context-size decisions.
///
/// `Real` is the provider-reported input-token count from the most recent
/// `Done` event — ground truth. `Heuristic` is the legacy
/// `serialized_json.len() / 4` estimate. `None` means no messages have
/// been sent yet (the loop has not observed any size at all).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum TokenSource {
/// No size observation available yet (cold start).
None,
/// `bytes/4` estimate; reliable only for token-sparse prose.
Heuristic(usize),
/// Provider-reported input token count from the most recent `Done`.
Real(usize),
}
impl AgentState {
/// Create a new, default-initialized agent state.
pub fn new() -> Self {
Self::default()
}
/// Add a user message
pub fn add_user_message(&mut self, content: String) {
self.messages
.push(Message::User(oxi_ai::UserMessage::new(content)));
}
/// Add an assistant message
pub fn add_assistant_message(&mut self, content: String) {
let mut assistant =
oxi_ai::AssistantMessage::new(oxi_ai::Api::AnthropicMessages, "agent", "agent-model");
assistant.content = vec![ContentBlock::Text(TextContent::new(content))];
self.messages.push(Message::Assistant(assistant));
}
/// Add a tool result message to both the message history and the tool results list.
pub fn add_tool_result(&mut self, tool_call_id: String, content: String) {
let content_for_result = content.clone();
let tool_result_msg = oxi_ai::ToolResultMessage::new(
tool_call_id.clone(),
"tool",
vec![ContentBlock::Text(TextContent::new(content))],
);
self.messages
.push(oxi_ai::Message::ToolResult(tool_result_msg));
self.tool_results
.push(ToolResult::success(tool_call_id, content_for_result));
}
/// Increment the iteration counter after an assistant turn completes.
pub fn increment_iteration(&mut self) {
self.iteration += 1;
}
/// Record the reason the last turn stopped.
pub fn set_stop_reason(&mut self, reason: StopReason) {
self.stop_reason = Some(reason);
}
/// Accumulate token usage from a completed LLM call.
///
/// `input` is **the input-token count for the just-completed turn** —
/// NOT a per-turn delta. The provider reports a fresh per-turn
/// `usage.input_tokens` on every `Done`; we accumulate it into
/// [`Self::input_tokens`] for lifetime accounting and **also** cache
/// it as the most-recent observation via [`Self::record_provider_turn`].
pub fn record_usage(&mut self, input: usize, output: usize) {
self.input_tokens += input;
self.output_tokens += output;
self.total_tokens += input + output;
}
/// Record the most recent provider-reported input-token count and the
/// heuristic estimate that was current at the time of the report, so
/// the loop can use the real count for compaction decisions and
/// surface drift (issue #28 gap 2).
///
/// `input_tokens` is the value from `ProviderEvent::Done.message.usage.input`.
/// `estimate_at_report` is the bytes/4 estimate taken at the same moment
/// (i.e. the value the legacy path *would* have used for compaction).
pub fn record_provider_turn(&mut self, input_tokens: usize, estimate_at_report: usize) {
self.last_input_tokens = Some(input_tokens);
self.last_estimate_at_report = Some(estimate_at_report);
// Divergence = reported / estimate. A value > 1.0 means the
// heuristic under-counted; the documented failure in #28 saw
// ~3.5×. Guard against the zero-estimate case to avoid Inf/NaN.
self.last_estimate_divergence = if estimate_at_report > 0 {
Some(input_tokens as f64 / estimate_at_report as f64)
} else if input_tokens > 0 {
// An estimate of 0 against a non-zero report is the worst-case
// divergence: the heuristic was essentially blind to the
// context. Surface this as a high multiplier.
Some(f64::INFINITY)
} else {
Some(1.0)
};
}
/// Current best estimate of context size, tagged with its source.
///
/// - `Real(n)` if the last completed turn reported `usage.input_tokens`.
/// - `Heuristic(n)` only before the first `Done` is observed (cold start).
/// - `None` if there are no messages yet.
///
/// Callers (notably `maybe_compact`) should **prefer** `Real` and **only**
/// fall back to `Heuristic` on cold start. See issue #28.
pub fn current_token_source(&self) -> TokenSource {
if let Some(real) = self.last_input_tokens {
TokenSource::Real(real)
} else if !self.messages.is_empty() {
TokenSource::Heuristic(self.estimate_tokens())
} else {
TokenSource::None
}
}
/// Clear all state, resetting for a new conversation.
pub fn clear(&mut self) {
self.messages.clear();
self.iteration = 0;
self.stop_reason = None;
self.tool_results.clear();
self.total_tokens = 0;
self.input_tokens = 0;
self.output_tokens = 0;
self.last_input_tokens = None;
self.last_estimate_at_report = None;
self.last_estimate_divergence = None;
}
/// Replace the entire message history (used after context compaction).
pub fn replace_messages(&mut self, messages: Vec<Message>) {
self.messages = messages;
}
/// Rough token-count estimate based on the serialized message JSON length.
pub fn estimate_tokens(&self) -> usize {
let json = serde_json::to_string(&self.messages).unwrap_or_default();
json.len() / 4 // Rough approximation
}
/// Returns `true` if the agent has signaled a stop reason.
pub fn is_complete(&self) -> bool {
self.stop_reason.is_some()
}
}
/// Thread-safe agent state wrapper.
#[derive(Default, Clone)]
pub struct SharedState {
state: Arc<RwLock<AgentState>>,
}
impl SharedState {
/// Create a new SharedState with default (empty) agent state.
pub fn new() -> Self {
Self::default()
}
/// Obtain a snapshot of the current agent state.
pub fn get_state(&self) -> AgentState {
self.state.read().clone()
}
/// Mutably update the agent state under a write lock.
pub fn update<F>(&self, f: F)
where
F: FnOnce(&mut AgentState),
{
let mut state = self.state.write();
f(&mut state);
}
/// Reset the state for a new conversation (delegates to [`AgentState::clear`]).
pub fn reset(&self) {
let mut state = self.state.write();
state.clear();
}
}