fix: slim down token usage

This commit is contained in:
eric
2026-04-04 12:37:50 +02:00
parent 97f329c825
commit 1240ab946b
55 changed files with 6799 additions and 2333 deletions

View File

@@ -6,7 +6,7 @@ use anyhow::Result;
use crate::app::{AppEvent, ControlCommand};
use crate::controller::{executor, goal_checker, planner, verifier};
use crate::model::{
ControllerPhase, GoalStatus, SessionEntry, SessionSource, SessionStream, TaskConfig,
ControllerPhase, GoalStatus, SessionEntry, SessionSource, SessionStream, StepStatus, TaskConfig,
};
use crate::repo;
use crate::storage::toon;
@@ -30,32 +30,44 @@ pub fn runtime_loop(
loop {
let mut plan = toon::read_plan(&config.plan_file)?;
let mut state = toon::read_state(&config.state_file)?;
if recover_stale_execution_state(&config, &mut plan, &mut state, &event_tx)? {
continue;
}
let goal_md = toon::read_markdown(&config.goal_file)?;
let standards_md = toon::read_markdown(&config.standards_file)?;
let _ = event_tx.send(AppEvent::Snapshot {
goal_md,
standards_md,
plan: plan.clone(),
state: state.clone(),
});
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
match control_rx.try_recv() {
Ok(ControlCommand::Quit) => break,
Ok(ControlCommand::Pause) => {
crate::controller::state::pause(&mut state);
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
continue;
}
Ok(ControlCommand::Resume) => {
crate::controller::state::resume(&mut state);
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
continue;
}
Ok(ControlCommand::Stop) => {
state.phase = ControllerPhase::Blocked;
state.goal_status = GoalStatus::Blocked;
state.notes.push("Stopped by user".to_string());
state.set_stop_reason("Stopped by user.");
let reason = state
.phase_notice()
.unwrap_or_else(|| "Stopped by user.".to_string());
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
let _ = event_tx.send(AppEvent::Session(SessionEntry {
source: SessionSource::Warning,
stream: SessionStream::Status,
title: "Notice".to_string(),
tag: Some(config.controller_id()),
body: reason,
run_id: repo::next_run_id(),
}));
continue;
}
Ok(ControlCommand::Submit(text)) => {
@@ -102,8 +114,10 @@ pub fn runtime_loop(
if goal_checker::is_done(&plan, &state)? {
state.phase = ControllerPhase::Done;
state.clear_stop_reason();
state.goal_status = GoalStatus::Done;
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
let _ = event_tx.send(AppEvent::Session(SessionEntry {
source: SessionSource::Controller,
stream: SessionStream::Status,
@@ -115,7 +129,8 @@ pub fn runtime_loop(
continue;
}
if state.replan_required || plan.has_no_actionable_steps() {
let resumable_step = resumable_step(&plan, &state);
if resumable_step.is_none() && (state.replan_required || plan.has_no_actionable_steps()) {
let _ = event_tx.send(AppEvent::Session(SessionEntry {
source: SessionSource::Planner,
stream: SessionStream::Status,
@@ -129,16 +144,29 @@ pub fn runtime_loop(
state.replan_required = false;
toon::write_plan(&config.plan_file, &plan)?;
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
continue;
}
let Some(step) = planner::next_step(&plan, &state) else {
let Some(step) = resumable_step.or_else(|| planner::next_step(&plan, &state)) else {
state.phase = ControllerPhase::Blocked;
state.goal_status = GoalStatus::Blocked;
state.notes.push(
"No actionable step remained and autonomous replan produced nothing.".to_string(),
state.set_stop_reason(
"No actionable step remained and autonomous replan produced nothing.",
);
let reason = state
.phase_notice()
.unwrap_or_else(|| "Controller is blocked.".to_string());
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
let _ = event_tx.send(AppEvent::Session(SessionEntry {
source: SessionSource::Warning,
stream: SessionStream::Status,
title: "Notice".to_string(),
tag: Some(config.controller_id()),
body: reason,
run_id: repo::next_run_id(),
}));
continue;
};
@@ -150,49 +178,68 @@ pub fn runtime_loop(
body: format!("Executing {}", step.title),
run_id: repo::next_run_id(),
}));
state.clear_stop_reason();
state.replan_required = false;
state
.blocked_steps
.retain(|blocked_step| blocked_step != &step.id);
plan.mark_active(&step.id);
state.current_step_id = Some(step.id.clone());
state.iteration += 1;
toon::write_plan(&config.plan_file, &plan)?;
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
let exec = executor::implement(&repo_root, &config, &plan, &step, &event_tx)?;
if goal_checker::needs_goal_clarification(&exec) {
state.phase = ControllerPhase::Planning;
state.notes.push(format!(
"Execution requested goal clarification while processing {}",
state.set_stop_reason(format!(
"Execution requested goal clarification while processing {}.",
step.id
));
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
continue;
}
let verification = verifier::verify_step(&repo_root, &exec, &event_tx)?;
if !verification.passed {
plan.mark_blocked(&step.id);
plan.append_step_note(&step.id, verification.summary.as_str());
state.last_verification = Some(verification);
state.blocked_steps.push(step.id.clone());
state.replan_required = true;
state.set_stop_reason(format!("Verification failed for {}.", step.id));
toon::write_plan(&config.plan_file, &plan)?;
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
continue;
}
let cleanup = verifier::verify_cleanup(&config, &step, &exec)?;
if !cleanup.passed {
plan.mark_todo(&step.id);
plan.append_step_note(&step.id, cleanup.summary.as_str());
state.last_cleanup_summary = Some(cleanup);
state.set_stop_reason(format!(
"Cleanup requirements were not satisfied for {}.",
step.id
));
toon::write_plan(&config.plan_file, &plan)?;
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
continue;
}
let tests = verifier::run_tests(&repo_root, &exec, &event_tx)?;
if !tests.passed {
plan.mark_todo(&step.id);
plan.append_step_note(&step.id, tests.summary.as_str());
state.last_full_test_summary = Some(tests);
state.set_stop_reason(format!("Tests failed for {}.", step.id));
toon::write_plan(&config.plan_file, &plan)?;
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
continue;
}
@@ -200,7 +247,215 @@ pub fn runtime_loop(
state.complete_step(&step, verification, cleanup, tests);
toon::write_plan(&config.plan_file, &plan)?;
toon::write_state(&config.state_file, &state)?;
emit_snapshot(&event_tx, &goal_md, &standards_md, &plan, &state);
}
Ok(())
}
fn emit_snapshot(
event_tx: &Sender<AppEvent>,
goal_md: &str,
standards_md: &str,
plan: &crate::model::Plan,
state: &crate::model::ControllerState,
) {
let _ = event_tx.send(AppEvent::Snapshot {
goal_md: goal_md.to_string(),
standards_md: standards_md.to_string(),
plan: plan.clone(),
state: state.clone(),
});
}
fn resumable_step(
plan: &crate::model::Plan,
state: &crate::model::ControllerState,
) -> Option<crate::model::PlanStep> {
let current_step_id = state.current_step_id.as_deref()?;
plan.steps
.iter()
.find(|step| {
step.id == current_step_id
&& matches!(
step.status,
StepStatus::Todo | StepStatus::Active | StepStatus::Blocked
)
})
.cloned()
}
fn recover_stale_execution_state(
config: &TaskConfig,
plan: &mut crate::model::Plan,
state: &mut crate::model::ControllerState,
event_tx: &Sender<AppEvent>,
) -> Result<bool> {
if state.current_step_id.is_some() {
return Ok(false);
}
let active_steps = plan.active_step_ids();
if active_steps.is_empty() {
return Ok(false);
}
for step_id in &active_steps {
plan.mark_todo(step_id);
plan.append_step_note(
step_id,
"Controller recovered this step from stale active state and returned it to todo.",
);
}
state.phase = ControllerPhase::Executing;
state.goal_status = GoalStatus::InProgress;
state.clear_stop_reason();
state.replan_required = false;
let reason = format!(
"Recovered stale active step state for {}. Reset {} to todo.",
config.controller_id(),
active_steps.join(", ")
);
state.notes.push(reason.clone());
toon::write_plan(&config.plan_file, plan)?;
toon::write_state(&config.state_file, state)?;
let _ = event_tx.send(AppEvent::Session(SessionEntry {
source: SessionSource::Controller,
stream: SessionStream::Status,
title: "Notice".to_string(),
tag: Some(config.controller_id()),
body: reason,
run_id: repo::next_run_id(),
}));
Ok(true)
}
#[cfg(test)]
mod tests {
use std::sync::mpsc;
use tempfile::tempdir;
use super::*;
use crate::model::{ControllerState, Plan, PlanStep, StepStatus};
use crate::storage::toon;
#[test]
fn recovers_stale_active_step_without_current_step() {
let temp = tempdir().expect("tempdir");
let mut config = TaskConfig::default_for("stale-active");
let root = temp.path().join(".agent/controllers/stale-active");
config.goal_file = root.join("goal.md");
config.plan_file = root.join("plan.toon");
config.state_file = root.join("state.toon");
config.standards_file = root.join("standards.md");
let mut plan = Plan {
version: 1,
goal_summary: "goal".to_string(),
steps: vec![PlanStep {
id: "s1".to_string(),
title: "Scope".to_string(),
status: StepStatus::Active,
attempts: 1,
..PlanStep::default()
}],
};
let mut state = ControllerState {
phase: ControllerPhase::Blocked,
goal_status: GoalStatus::Blocked,
..ControllerState::default()
};
state
.set_stop_reason("No actionable step remained and autonomous replan produced nothing.");
toon::ensure_controller_files(&config).expect("ensure files");
let (event_tx, event_rx) = mpsc::channel();
let recovered = recover_stale_execution_state(&config, &mut plan, &mut state, &event_tx)
.expect("recover");
assert!(recovered);
assert!(matches!(plan.steps[0].status, StepStatus::Todo));
assert!(plan.steps[0].notes.contains("stale active state"));
assert!(matches!(state.phase, ControllerPhase::Executing));
assert!(matches!(state.goal_status, GoalStatus::InProgress));
assert!(state.stop_reason.is_none());
let event = event_rx.recv().expect("notice event");
match event {
AppEvent::Session(entry) => assert!(entry.body.contains("Reset s1 to todo")),
other => panic!("unexpected event: {other:?}"),
}
}
#[test]
fn resumable_step_prefers_current_blocked_or_active_step() {
let plan = Plan {
version: 1,
goal_summary: "goal".to_string(),
steps: vec![
PlanStep {
id: "s1".to_string(),
title: "Scope".to_string(),
status: StepStatus::Blocked,
..PlanStep::default()
},
PlanStep {
id: "s2".to_string(),
title: "Other".to_string(),
status: StepStatus::Done,
..PlanStep::default()
},
],
};
let state = ControllerState {
phase: ControllerPhase::Blocked,
goal_status: GoalStatus::Blocked,
current_step_id: Some("s1".to_string()),
replan_required: true,
..ControllerState::default()
};
let resumed = resumable_step(&plan, &state).expect("expected resumable step");
assert_eq!(resumed.id, "s1");
}
#[test]
fn emit_snapshot_clones_current_plan_and_state() {
let (event_tx, event_rx) = mpsc::channel();
let plan = Plan {
version: 1,
goal_summary: "goal".to_string(),
steps: vec![PlanStep {
id: "s1".to_string(),
title: "Scope".to_string(),
status: StepStatus::Active,
..PlanStep::default()
}],
};
let state = ControllerState {
phase: ControllerPhase::Executing,
current_step_id: Some("s1".to_string()),
..ControllerState::default()
};
emit_snapshot(&event_tx, "goal body", "standards body", &plan, &state);
let event = event_rx.recv().expect("snapshot event");
match event {
AppEvent::Snapshot {
goal_md,
standards_md,
plan,
state,
} => {
assert_eq!(goal_md, "goal body");
assert_eq!(standards_md, "standards body");
assert!(matches!(plan.steps[0].status, StepStatus::Active));
assert_eq!(state.current_step_id.as_deref(), Some("s1"));
}
other => panic!("unexpected event: {other:?}"),
}
}
}