spqrz/consciousness
1
0
Fork 0
forked from kent/consciousness
Code Pull requests Activity

refactor: extract context window building into context.rs

Browse source 
Move context window construction (build_context_window, plan_context,
render_journal_text, assemble_context), token counting, error
classification, and related helpers from agent.rs into context.rs.

All extracted functions are pure — they take inputs and return values
with no mutable state access. State mutation stays in agent.rs
(compact, restore_from_log, load_startup_journal).

agent.rs: 1504 → 987 lines (-517)
context.rs: 365 lines (new)
Net: -152 lines (duplicate comments removed)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Kent Overstreet 2026-03-21 15:55:30 -04:00
parent d04d41e993
commit db48d57917
3 changed files with 206 additions and 488 deletions
Download patch file Download diff file Expand all files Collapse all files

253
poc-agent/src/context.rs
View file

@ -1,14 +1,52 @@
// context.rs — Context window building and management
//
// Pure functions for building the agent's context window from journal
// entries and conversation messages. No mutable state.
// entries and conversation messages. No mutable state — all functions
// take inputs and return new values. State mutation happens in agent.rs.
use crate::journal;
use crate::types::{ContextPlan, ContextState, Message};
use crate::types::*;
use chrono::{DateTime, Utc};
use tiktoken_rs::CoreBPE;
/// Look up a model's context window size in tokens.
pub fn model_context_window(model: &str) -> usize {
let m = model.to_lowercase();
if m.contains("opus") || m.contains("sonnet") {
200_000
} else if m.contains("qwen") {
131_072
} else {
128_000
}
}
/// Context budget in tokens: 60% of the model's context window.
fn context_budget_tokens(model: &str) -> usize {
model_context_window(model) * 60 / 100
}
/// Allocation plan for the context window.
pub struct ContextPlan {
header_start: usize,
full_start: usize,
entry_count: usize,
conv_trim: usize,
_conv_count: usize,
_full_tokens: usize,
_header_tokens: usize,
_conv_tokens: usize,
_available: usize,
}
/// Build a context window from conversation messages + journal entries.
///
/// Allocation strategy: identity and memory are fixed costs. The
/// remaining budget (minus 25% reserve for model output) is split
/// between journal and conversation. Conversation gets priority —
/// it's what's happening now. Journal fills the rest, newest first.
///
/// Returns (messages, journal_text) — caller stores journal_text in ContextState.
pub fn build_context_window(
context: &ContextState,
conversation: &[Message],
@ -17,44 +55,50 @@ pub fn build_context_window(
) -> (Vec<Message>, String) {
let journal_path = journal::default_journal_path();
let all_entries = journal::parse_journal(&journal_path);
crate::dbglog!("[ctx] {} journal entries from {}", all_entries.len(), journal_path.display());
dbglog!("[ctx] {} journal entries from {}", all_entries.len(), journal_path.display());
let count = |s: &str| tokenizer.encode_with_special_tokens(s).len();
let system_prompt = context.system_prompt.clone();
let context_message = context.render_context_message();
// Cap memory to 50% of the context budget so conversation always
// gets space. Truncate at the last complete section boundary.
let max_tokens = context_budget_tokens(model);
let memory_cap = max_tokens / 2;
let memory_tokens = count(&context_message);
let context_message = if memory_tokens > memory_cap {
crate::dbglog!("[ctx] memory too large: {} tokens > {} cap, truncating", memory_tokens, memory_cap);
dbglog!("[ctx] memory too large: {} tokens > {} cap, truncating", memory_tokens, memory_cap);
truncate_at_section(&context_message, memory_cap, &count)
} else {
context_message
};
let recent_start = find_journal_cutoff(conversation, all_entries.last());
dbglog!("[ctx] journal cutoff: {} of {} conversation messages are 'recent'",
conversation.len() - recent_start, conversation.len());
let recent = &conversation[recent_start..];
let plan = plan_context(&system_prompt, &context_message, recent, &all_entries, model, &count);
let plan = plan_context(
&system_prompt,
&context_message,
recent,
&all_entries,
model,
&count,
);
let journal_text = render_journal_text(&all_entries, &plan);
let messages = assemble_context(system_prompt, context_message, &journal_text, recent, &plan);
dbglog!("[ctx] plan: header_start={} full_start={} entry_count={} conv_trim={} journal_text={} chars",
plan.header_start, plan.full_start, plan.entry_count, plan.conv_trim, journal_text.len());
let messages = assemble_context(
system_prompt, context_message, &journal_text,
recent, &plan,
);
(messages, journal_text)
}
pub fn model_context_window(model: &str) -> usize {
let m = model.to_lowercase();
if m.contains("opus") || m.contains("sonnet") { 200_000 }
else if m.contains("qwen") { 131_072 }
else { 128_000 }
}
fn context_budget_tokens(model: &str) -> usize {
model_context_window(model) * 60 / 100
}
fn plan_context(
pub fn plan_context(
system_prompt: &str,
context_message: &str,
recent: &[Message],
@ -63,40 +107,56 @@ fn plan_context(
count: &dyn Fn(&str) -> usize,
) -> ContextPlan {
let max_tokens = context_budget_tokens(model);
let identity_cost = count(system_prompt);
let memory_cost = count(context_message);
let reserve = max_tokens / 4;
let available = max_tokens.saturating_sub(identity_cost).saturating_sub(memory_cost).saturating_sub(reserve);
let available = max_tokens
.saturating_sub(identity_cost)
.saturating_sub(memory_cost)
.saturating_sub(reserve);
let conv_costs: Vec<usize> = recent.iter().map(|m| msg_token_count_fn(m, count)).collect();
let total_conv: usize = conv_costs.iter().sum();
let journal_min = available * 15 / 100;
let journal_budget = available.saturating_sub(total_conv).max(journal_min);
let full_budget = journal_budget * 70 / 100;
let header_budget = journal_budget.saturating_sub(full_budget);
// Phase 1: Full entries (newest first)
let mut full_used = 0;
let mut n_full = 0;
for entry in entries.iter().rev() {
let cost = count(&entry.content) + 10;
if full_used + cost > full_budget { break; }
if full_used + cost > full_budget {
break;
}
full_used += cost;
n_full += 1;
}
let full_start = entries.len().saturating_sub(n_full);
// Phase 2: Header-only entries (continuing backward)
let mut header_used = 0;
let mut n_headers = 0;
for entry in entries[..full_start].iter().rev() {
let first_line = entry.content.lines().find(|l| !l.trim().is_empty()).unwrap_or("(empty)");
let first_line = entry
.content
.lines()
.find(|l| !l.trim().is_empty())
.unwrap_or("(empty)");
let cost = count(first_line) + 10;
if header_used + cost > header_budget { break; }
if header_used + cost > header_budget {
break;
}
header_used += cost;
n_headers += 1;
}
let header_start = full_start.saturating_sub(n_headers);
// Trim oldest conversation if it exceeds budget
let journal_used = full_used + header_used;
let mut conv_trim = 0;
let mut trimmed_conv = total_conv;
@ -104,34 +164,69 @@ fn plan_context(
trimmed_conv -= conv_costs[conv_trim];
conv_trim += 1;
}
while conv_trim < recent.len() && recent[conv_trim].role != crate::types::Role::User {
// Walk forward to user message boundary
while conv_trim < recent.len() && recent[conv_trim].role != Role::User {
conv_trim += 1;
}
dbglog!("[plan] model={} max_tokens={} available={} (identity={} memory={} reserve={})",
model, max_tokens, available, identity_cost, memory_cost, reserve);
dbglog!("[plan] conv: {} msgs, {} tokens total, trimming {} msgs → {} tokens",
recent.len(), total_conv, conv_trim, trimmed_conv);
dbglog!("[plan] journal: {} full entries ({}t) + {} headers ({}t)",
n_full, full_used, n_headers, header_used);
ContextPlan {
header_start, full_start, entry_count: entries.len(), conv_trim,
_conv_count: recent.len(), _full_tokens: full_used, _header_tokens: header_used,
_conv_tokens: trimmed_conv, _available: available,
header_start,
full_start,
entry_count: entries.len(),
conv_trim,
_conv_count: recent.len(),
_full_tokens: full_used,
_header_tokens: header_used,
_conv_tokens: trimmed_conv,
_available: available,
}
}
fn render_journal_text(entries: &[journal::JournalEntry], plan: &ContextPlan) -> String {
if plan.header_start >= plan.entry_count { return String::new(); }
pub fn render_journal_text(
entries: &[journal::JournalEntry],
plan: &ContextPlan,
) -> String {
let has_journal = plan.header_start < plan.entry_count;
if !has_journal {
return String::new();
}
let mut text = String::from("[Earlier in this conversation — from your journal]\n\n");
for entry in &entries[plan.header_start..plan.full_start] {
let first_line = entry.content.lines().find(|l| !l.trim().is_empty()).unwrap_or("(empty)");
text.push_str(&format!("## {} — {}\n", entry.timestamp.format("%Y-%m-%dT%H:%M"), first_line));
let first_line = entry
.content
.lines()
.find(|l| !l.trim().is_empty())
.unwrap_or("(empty)");
text.push_str(&format!(
"## {} — {}\n",
entry.timestamp.format("%Y-%m-%dT%H:%M"),
first_line,
));
}
let n_headers = plan.full_start - plan.header_start;
let n_full = plan.entry_count - plan.full_start;
if n_headers > 0 && n_full > 0 { text.push_str("\n---\n\n"); }
if n_headers > 0 && n_full > 0 {
text.push_str("\n---\n\n");
}
for entry in &entries[plan.full_start..] {
text.push_str(&format!("## {}\n\n{}\n\n", entry.timestamp.format("%Y-%m-%dT%H:%M"), entry.content));
text.push_str(&format!(
"## {}\n\n{}\n\n",
entry.timestamp.format("%Y-%m-%dT%H:%M"),
entry.content
));
}
text
}
@ -143,10 +238,12 @@ fn assemble_context(
plan: &ContextPlan,
) -> Vec<Message> {
let mut messages = vec![Message::system(system_prompt)];
if !context_message.is_empty() { messages.push(Message::user(context_message)); }
if !context_message.is_empty() {
messages.push(Message::user(context_message));
}
let final_recent = &recent[plan.conv_trim..];
if !journal_text.is_empty() {
messages.push(Message::user(journal_text.to_string()));
} else if !final_recent.is_empty() {
@ -157,6 +254,7 @@ fn assemble_context(
.to_string(),
));
}
messages.extend(final_recent.iter().cloned());
messages
}
@ -174,26 +272,42 @@ fn truncate_at_section(text: &str, max_tokens: usize, count: &dyn Fn(&str) -> us
let mut best = 0;
for &end in &boundaries[1..] {
let slice = &text[..end];
if count(slice) <= max_tokens { best = end; }
else { break; }
if count(slice) <= max_tokens {
best = end;
} else {
break;
}
}
if best == 0 {
best = text.len().min(max_tokens * 3);
}
if best == 0 { best = text.len().min(max_tokens * 3); }
let truncated = &text[..best];
crate::dbglog!("[ctx] truncated memory from {} to {} chars ({} tokens)", text.len(), truncated.len(), count(truncated));
dbglog!("[ctx] truncated memory from {} to {} chars ({} tokens)",
text.len(), truncated.len(), count(truncated));
truncated.to_string()
}
fn find_journal_cutoff(conversation: &[Message], newest_entry: Option<&journal::JournalEntry>) -> usize {
let cutoff = match newest_entry { Some(entry) => entry.timestamp, None => return 0 };
fn find_journal_cutoff(
conversation: &[Message],
newest_entry: Option<&journal::JournalEntry>,
) -> usize {
let cutoff = match newest_entry {
Some(entry) => entry.timestamp,
None => return 0,
};
let mut split = conversation.len();
for (i, msg) in conversation.iter().enumerate() {
if let Some(ts) = parse_msg_timestamp(msg) {
if ts > cutoff { split = i; break; }
if ts > cutoff {
split = i;
break;
}
}
}
while split > 0 && split < conversation.len() && conversation[split].role != crate::types::Role::User {
while split > 0 && split < conversation.len() && conversation[split].role != Role::User {
split -= 1;
}
split
@ -201,32 +315,51 @@ fn find_journal_cutoff(conversation: &[Message], newest_entry: Option<&journal::
fn msg_token_count_fn(msg: &Message, count: &dyn Fn(&str) -> usize) -> usize {
let content = msg.content.as_ref().map_or(0, |c| match c {
crate::types::MessageContent::Text(s) => count(s),
crate::types::MessageContent::Parts(parts) => parts.iter().map(|p| match p {
crate::types::ContentPart::Text { text } => count(text),
crate::types::ContentPart::ImageUrl { .. } => 85,
}).sum(),
MessageContent::Text(s) => count(s),
MessageContent::Parts(parts) => parts
.iter()
.map(|p| match p {
ContentPart::Text { text } => count(text),
ContentPart::ImageUrl { .. } => 85,
})
.sum(),
});
let tools = msg.tool_calls.as_ref().map_or(0, |calls| {
calls
.iter()
.map(|c| count(&c.function.arguments) + count(&c.function.name))
.sum()
});
let tools = msg.tool_calls.as_ref().map_or(0, |calls| calls.iter().map(|c| count(&c.function.arguments) + count(&c.function.name)).sum());
content + tools
}
fn parse_msg_timestamp(msg: &Message) -> Option<DateTime<Utc>> {
msg.timestamp.as_ref().and_then(|ts| DateTime::parse_from_rfc3339(ts).ok()).map(|dt| dt.with_timezone(&Utc))
/// Count the token footprint of a message using BPE tokenization.
pub fn msg_token_count(tokenizer: &CoreBPE, msg: &Message) -> usize {
msg_token_count_fn(msg, &|s| tokenizer.encode_with_special_tokens(s).len())
}
/// Detect context window overflow errors from the API.
pub fn is_context_overflow(err: &anyhow::Error) -> bool {
let msg = err.to_string().to_lowercase();
msg.contains("context length") || msg.contains("token limit") || msg.contains("too many tokens")
|| msg.contains("maximum context") || msg.contains("prompt is too long") || msg.contains("request too large")
|| msg.contains("input validation error") || msg.contains("content length limit")
msg.contains("context length")
|| msg.contains("token limit")
|| msg.contains("too many tokens")
|| msg.contains("maximum context")
|| msg.contains("prompt is too long")
|| msg.contains("request too large")
|| msg.contains("input validation error")
|| msg.contains("content length limit")
|| (msg.contains("400") && msg.contains("tokens"))
}
/// Detect model/provider errors delivered inside the SSE stream.
pub fn is_stream_error(err: &anyhow::Error) -> bool {
err.to_string().contains("model stream error")
}
pub fn msg_token_count(tokenizer: &CoreBPE, msg: &Message) -> usize {
msg_token_count_fn(msg, &|s| tokenizer.encode_with_special_tokens(s).len())
fn parse_msg_timestamp(msg: &Message) -> Option<DateTime<Utc>> {
msg.timestamp
.as_ref()
.and_then(|ts| DateTime::parse_from_rfc3339(ts).ok())
.map(|dt| dt.with_timezone(&Utc))
}