kent/consciousness
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

simplify compaction: Agent owns config, compact() reloads everything

Browse source 
Agent stores AppConfig and prompt_file, so compact() reloads
identity internally — callers no longer pass system_prompt and
personality. restore_from_log() loads entries and calls compact().

Remove soft compaction threshold and pre-compaction nudge (journal
agent handles this). Remove /compact and /context commands (F10
debug screen replaces both). Inline do_compact, emergency_compact,
trim_and_reload into compact(). Rename model_context_window to
context_window, drop unused model parameter.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
This commit is contained in:
Kent Overstreet 2026-04-02 16:08:41 -04:00
parent d419587c1b
commit af3929cc65
2 changed files with 53 additions and 179 deletions
Download patch file Download diff file Expand all files Collapse all files

91
src/agent/runner.rs
View file

@ -69,6 +69,9 @@ pub struct Agent {
pub context: ContextState,
/// Shared live context summary — TUI reads this directly for debug screen.
pub shared_context: SharedContextState,
/// App config — used to reload identity on compaction.
app_config: crate::config::AppConfig,
pub prompt_file: String,
/// Stable session ID for memory-search dedup across turns.
session_id: String,
/// Agent orchestration state (surface-observe, journal, reflect).
@ -90,6 +93,8 @@ impl Agent {
client: ApiClient,
system_prompt: String,
personality: Vec<(String, String)>,
app_config: crate::config::AppConfig,
prompt_file: String,
conversation_log: Option<ConversationLog>,
shared_context: SharedContextState,
) -> Self {
@ -116,6 +121,8 @@ impl Agent {
tokenizer,
context,
shared_context,
app_config,
prompt_file,
session_id,
agent_cycles,
};
@ -181,7 +188,7 @@ impl Agent {
pub fn budget(&self) -> ContextBudget {
let count_str = |s: &str| self.tokenizer.encode_with_special_tokens(s).len();
let count_msg = |m: &Message| crate::thought::context::msg_token_count(&self.tokenizer, m);
let window = crate::thought::context::model_context_window(&self.client.model);
let window = crate::thought::context::context_window();
self.context.budget(&count_str, &count_msg, window)
}
@ -332,7 +339,7 @@ impl Agent {
"[context overflow — compacting and retrying ({}/2)]",
overflow_retries,
)));
self.emergency_compact();
self.compact();
continue;
}
if crate::thought::context::is_stream_error(&err) && empty_retries < 2 {
@ -787,7 +794,7 @@ impl Agent {
// Walk backwards from cutoff, accumulating entries within 5% of context
let count = |s: &str| self.tokenizer.encode_with_special_tokens(s).len();
let context_window = crate::thought::context::model_context_window(&self.client.model);
let context_window = crate::thought::context::context_window();
let journal_budget = context_window * 5 / 100;
dbg_log!("[journal] budget={} tokens ({}*5%)", journal_budget, context_window);
@ -899,21 +906,23 @@ impl Agent {
self.last_prompt_tokens
}
/// Build context window from conversation messages + journal.
/// Used by both compact() (in-memory messages) and restore_from_log()
/// (conversation log). The context window is always:
/// identity + journal summaries + raw recent messages
pub fn compact(&mut self, new_system_prompt: String, new_personality: Vec<(String, String)>) {
self.context.system_prompt = new_system_prompt;
self.context.personality = new_personality;
self.do_compact();
}
/// Dedup memory entries, trim to fit, reload journal for new time range.
fn trim_and_reload(&mut self, entries: &[ConversationEntry]) {
/// Rebuild the context window: reload identity, dedup, trim, reload journal.
pub fn compact(&mut self) {
// Reload identity from config
match crate::config::reload_for_model(&self.app_config, &self.prompt_file) {
Ok((system_prompt, personality)) => {
self.context.system_prompt = system_prompt;
self.context.personality = personality;
}
Err(e) => {
eprintln!("warning: failed to reload identity: {:#}", e);
}
}
// Dedup memory, trim to budget, reload journal
let entries = self.context.entries.clone();
self.context.entries = crate::thought::context::trim_entries(
&self.context,
entries,
&entries,
&self.tokenizer,
);
self.load_startup_journal();
@ -921,58 +930,24 @@ impl Agent {
self.publish_context_state();
}
/// Internal compaction — dedup memory entries and trim to fit.
fn do_compact(&mut self) {
let entries = self.context.entries.clone();
self.trim_and_reload(&entries);
}
/// Emergency compaction using stored config — called on context overflow.
fn emergency_compact(&mut self) {
self.do_compact();
}
/// Restore from the conversation log. Builds the context window
/// the same way compact() does — journal summaries for old messages,
/// raw recent messages. This is the unified startup path.
/// Returns true if the log had content to restore.
pub fn restore_from_log(
&mut self,
system_prompt: String,
personality: Vec<(String, String)>,
) -> bool {
self.context.system_prompt = system_prompt;
self.context.personality = personality;
pub fn restore_from_log(&mut self) -> bool {
let entries = match &self.conversation_log {
Some(log) => match log.read_tail(512 * 1024) {
Ok(entries) if !entries.is_empty() => {
dbglog!("[restore] read {} entries from log tail", entries.len());
entries
}
Ok(_) => {
dbglog!("[restore] log exists but is empty");
return false;
}
Err(e) => {
dbglog!("[restore] failed to read log: {}", e);
return false;
}
Some(log) => match log.read_tail(2 * 1024 * 1024) {
Ok(entries) if !entries.is_empty() => entries,
_ => return false,
},
None => {
dbglog!("[restore] no conversation log configured");
return false;
}
None => return false,
};
// Filter out system messages, dedup memory, trim to fit
let entries: Vec<ConversationEntry> = entries
.into_iter()
// Load extra — compact() will dedup, trim, reload identity + journal
self.context.entries = entries.into_iter()
.filter(|e| e.message().role != Role::System)
.collect();
self.trim_and_reload(&entries);
dbglog!("[restore] {} entries, journal: {} entries",
self.context.entries.len(), self.context.journal.len());
self.compact();
true
}