mirror of
https://github.com/pchuan98/codex.git
synced 2026-07-01 00:31:56 +08:00
51fc4b0559
## Why Memory startup used hardcoded OpenAI model slugs for extraction and consolidation. That works for the default OpenAI-compatible path, but provider-specific backends can require different model identifiers. In particular, Amazon Bedrock should use its Bedrock model ID for these background memory requests instead of the OpenAI `gpt-5.4-mini` / `gpt-5.4` slugs. ## What Changed - Added provider-owned preferred memory model methods alongside `approval_review_preferred_model`. - Updated memory extraction and consolidation to resolve their default model through the active `ModelProvider`. - Added Amazon Bedrock overrides so both memory stages use `openai.gpt-5.4` through Bedrock’s provider-specific model ID. - Kept explicit `memories.extract_model` and `memories.consolidation_model` config overrides taking precedence. - Added startup coverage for default OpenAI and Bedrock memory model selection. #closes #26288
578 lines
20 KiB
Rust
578 lines
20 KiB
Rust
use crate::build_consolidation_prompt;
|
|
use crate::memory_root;
|
|
use crate::metrics::MEMORY_PHASE_TWO_E2E_MS;
|
|
use crate::metrics::MEMORY_PHASE_TWO_INPUT;
|
|
use crate::metrics::MEMORY_PHASE_TWO_JOBS;
|
|
use crate::metrics::MEMORY_PHASE_TWO_TOKEN_USAGE;
|
|
use crate::prune_old_extension_resources;
|
|
use crate::rebuild_raw_memories_file_from_memories;
|
|
use crate::runtime::MemoryStartupContext;
|
|
use crate::runtime::SpawnedConsolidationAgent;
|
|
use crate::sync_rollout_summaries_from_memories;
|
|
use crate::workspace::memory_workspace_diff;
|
|
use crate::workspace::prepare_memory_workspace;
|
|
use crate::workspace::reset_memory_workspace_baseline;
|
|
use crate::workspace::write_workspace_diff;
|
|
use codex_config::Constrained;
|
|
use codex_core::config::Config;
|
|
use codex_features::Feature;
|
|
use codex_model_provider::ModelProvider;
|
|
use codex_protocol::ThreadId;
|
|
use codex_protocol::protocol::AgentStatus;
|
|
use codex_protocol::protocol::AskForApproval;
|
|
use codex_protocol::protocol::SandboxPolicy;
|
|
use codex_protocol::protocol::TokenUsage;
|
|
use codex_protocol::user_input::UserInput;
|
|
use codex_state::Stage1Output;
|
|
use codex_state::StateRuntime;
|
|
use std::collections::HashMap;
|
|
use std::path::Path;
|
|
use std::sync::Arc;
|
|
use std::time::Duration;
|
|
|
|
#[derive(Debug, Clone, Default)]
|
|
struct Claim {
|
|
token: String,
|
|
watermark: i64,
|
|
}
|
|
|
|
#[derive(Debug, Clone, Default)]
|
|
struct Counters {
|
|
input: i64,
|
|
}
|
|
|
|
/// Runs memory phase 2 (aka consolidation) in strict order. The method represents the linear
|
|
/// flow of the consolidation phase.
|
|
pub async fn run(context: Arc<MemoryStartupContext>, config: Arc<Config>) {
|
|
let phase_two_e2e_timer = context.start_timer(MEMORY_PHASE_TWO_E2E_MS);
|
|
|
|
let Some(db) = context.state_db() else {
|
|
// This should not happen.
|
|
return;
|
|
};
|
|
let root = memory_root(&config.codex_home);
|
|
let max_raw_memories = config.memories.max_raw_memories_for_consolidation;
|
|
let max_unused_days = config.memories.max_unused_days;
|
|
|
|
// 1. Claim the global Phase 2 lock before touching the memory workspace.
|
|
let claim = match job::claim(context.as_ref(), db.as_ref()).await {
|
|
Ok(claim) => claim,
|
|
Err(e) => {
|
|
context.counter(MEMORY_PHASE_TWO_JOBS, /*inc*/ 1, &[("status", e)]);
|
|
return;
|
|
}
|
|
};
|
|
|
|
// 2. Ensure the memories root has a git baseline repository.
|
|
if let Err(err) = prepare_memory_workspace(&root).await {
|
|
tracing::error!("failed preparing memory workspace: {err}");
|
|
job::failed(
|
|
context.as_ref(),
|
|
db.as_ref(),
|
|
&claim,
|
|
"failed_prepare_workspace",
|
|
)
|
|
.await;
|
|
return;
|
|
}
|
|
|
|
// 3. Build the locked-down config used by the consolidation agent.
|
|
let Some(agent_config) = agent::get_config(config.as_ref(), context.provider()) else {
|
|
// If we can't get the config, we can't consolidate.
|
|
tracing::error!("failed to get agent config");
|
|
job::failed(
|
|
context.as_ref(),
|
|
db.as_ref(),
|
|
&claim,
|
|
"failed_sandbox_policy",
|
|
)
|
|
.await;
|
|
return;
|
|
};
|
|
|
|
// 4. Load current DB-backed Phase 2 inputs.
|
|
let raw_memories = match db
|
|
.memories()
|
|
.get_phase2_input_selection(max_raw_memories, max_unused_days)
|
|
.await
|
|
{
|
|
Ok(raw_memories) => raw_memories,
|
|
Err(err) => {
|
|
tracing::error!("failed to list stage1 outputs from global: {err}");
|
|
job::failed(
|
|
context.as_ref(),
|
|
db.as_ref(),
|
|
&claim,
|
|
"failed_load_stage1_outputs",
|
|
)
|
|
.await;
|
|
return;
|
|
}
|
|
};
|
|
let raw_memory_count = raw_memories.len();
|
|
let new_watermark = get_watermark(claim.watermark, &raw_memories);
|
|
|
|
// 5. Sync the current inputs into the memory workspace.
|
|
if let Err(err) = sync_phase2_workspace_inputs(&root, &raw_memories).await {
|
|
tracing::error!("failed syncing phase2 workspace inputs: {err}");
|
|
job::failed(
|
|
context.as_ref(),
|
|
db.as_ref(),
|
|
&claim,
|
|
"failed_sync_workspace_inputs",
|
|
)
|
|
.await;
|
|
return;
|
|
}
|
|
|
|
// 6. Use git to decide whether the synced workspace actually changed.
|
|
let workspace_diff = match memory_workspace_diff(&root).await {
|
|
Ok(diff) => diff,
|
|
Err(err) => {
|
|
tracing::error!("failed checking memory workspace changes: {err}");
|
|
job::failed(
|
|
context.as_ref(),
|
|
db.as_ref(),
|
|
&claim,
|
|
"failed_workspace_status",
|
|
)
|
|
.await;
|
|
return;
|
|
}
|
|
};
|
|
if !workspace_diff.has_changes() {
|
|
tracing::error!("Phase 2 no changes");
|
|
// We check only after sync of the file system.
|
|
job::succeed(
|
|
context.as_ref(),
|
|
db.as_ref(),
|
|
&claim,
|
|
new_watermark,
|
|
&raw_memories,
|
|
"succeeded_no_workspace_changes",
|
|
)
|
|
.await;
|
|
return;
|
|
}
|
|
|
|
// 7. Persist the diff for the consolidation agent to inspect.
|
|
if let Err(err) = write_workspace_diff(&root, &workspace_diff).await {
|
|
tracing::error!("failed writing memory workspace diff file: {err}");
|
|
job::failed(
|
|
context.as_ref(),
|
|
db.as_ref(),
|
|
&claim,
|
|
"failed_workspace_diff_file",
|
|
)
|
|
.await;
|
|
return;
|
|
}
|
|
|
|
// 8. Spawn the consolidation agent.
|
|
let prompt = agent::get_prompt(&root);
|
|
let agent = match context
|
|
.spawn_consolidation_agent(agent_config, prompt)
|
|
.await
|
|
{
|
|
Ok(agent) => agent,
|
|
Err(err) => {
|
|
tracing::error!("failed to spawn global memory consolidation agent: {err}");
|
|
job::failed(context.as_ref(), db.as_ref(), &claim, "failed_spawn_agent").await;
|
|
return;
|
|
}
|
|
};
|
|
|
|
// 9. Hand off completion handling, heartbeats, and baseline reset.
|
|
agent::handle(
|
|
Arc::clone(&context),
|
|
claim,
|
|
new_watermark,
|
|
raw_memories.clone(),
|
|
root,
|
|
agent,
|
|
phase_two_e2e_timer,
|
|
);
|
|
|
|
// 10. Emit dispatch metrics.
|
|
let counters = Counters {
|
|
input: raw_memory_count as i64,
|
|
};
|
|
emit_metrics(context.as_ref(), counters);
|
|
}
|
|
|
|
async fn sync_phase2_workspace_inputs(
|
|
root: &Path,
|
|
raw_memories: &[Stage1Output],
|
|
) -> std::io::Result<()> {
|
|
let raw_memory_count = raw_memories.len();
|
|
sync_rollout_summaries_from_memories(root, raw_memories, raw_memory_count).await?;
|
|
rebuild_raw_memories_file_from_memories(root, raw_memories, raw_memory_count).await?;
|
|
prune_old_extension_resources(root).await;
|
|
Ok(())
|
|
}
|
|
|
|
mod job {
|
|
use super::*;
|
|
|
|
pub(super) async fn claim(
|
|
context: &MemoryStartupContext,
|
|
db: &StateRuntime,
|
|
) -> Result<Claim, &'static str> {
|
|
let claim = db
|
|
.memories()
|
|
.try_claim_global_phase2_job(context.thread_id(), crate::stage_two::JOB_LEASE_SECONDS)
|
|
.await
|
|
.map_err(|e| {
|
|
tracing::error!("failed to claim job: {e}");
|
|
"failed_claim"
|
|
})?;
|
|
let (token, watermark) = match claim {
|
|
codex_state::Phase2JobClaimOutcome::Claimed {
|
|
ownership_token,
|
|
input_watermark,
|
|
} => {
|
|
context.counter(
|
|
MEMORY_PHASE_TWO_JOBS,
|
|
/*inc*/ 1,
|
|
&[("status", "claimed")],
|
|
);
|
|
(ownership_token, input_watermark)
|
|
}
|
|
codex_state::Phase2JobClaimOutcome::SkippedRetryUnavailable => {
|
|
return Err("skipped_retry_unavailable");
|
|
}
|
|
codex_state::Phase2JobClaimOutcome::SkippedCooldown => {
|
|
return Err("skipped_cooldown");
|
|
}
|
|
codex_state::Phase2JobClaimOutcome::SkippedRunning => return Err("skipped_running"),
|
|
};
|
|
|
|
Ok(Claim { token, watermark })
|
|
}
|
|
|
|
pub(super) async fn failed(
|
|
context: &MemoryStartupContext,
|
|
db: &StateRuntime,
|
|
claim: &Claim,
|
|
reason: &'static str,
|
|
) {
|
|
context.counter(MEMORY_PHASE_TWO_JOBS, /*inc*/ 1, &[("status", reason)]);
|
|
if matches!(
|
|
db.memories()
|
|
.mark_global_phase2_job_failed(
|
|
&claim.token,
|
|
reason,
|
|
crate::stage_two::JOB_RETRY_DELAY_SECONDS,
|
|
)
|
|
.await,
|
|
Ok(false)
|
|
) {
|
|
let _ = db
|
|
.memories()
|
|
.mark_global_phase2_job_failed_if_unowned(
|
|
&claim.token,
|
|
reason,
|
|
crate::stage_two::JOB_RETRY_DELAY_SECONDS,
|
|
)
|
|
.await;
|
|
}
|
|
}
|
|
|
|
pub(super) async fn succeed(
|
|
context: &MemoryStartupContext,
|
|
db: &StateRuntime,
|
|
claim: &Claim,
|
|
completion_watermark: i64,
|
|
selected_outputs: &[codex_state::Stage1Output],
|
|
reason: &'static str,
|
|
) -> bool {
|
|
context.counter(MEMORY_PHASE_TWO_JOBS, /*inc*/ 1, &[("status", reason)]);
|
|
db.memories()
|
|
.mark_global_phase2_job_succeeded(&claim.token, completion_watermark, selected_outputs)
|
|
.await
|
|
.unwrap_or(false)
|
|
}
|
|
}
|
|
|
|
mod agent {
|
|
use super::*;
|
|
use tracing::warn;
|
|
|
|
pub(super) fn get_config(config: &Config, provider: &dyn ModelProvider) -> Option<Config> {
|
|
let root = memory_root(&config.codex_home);
|
|
let mut agent_config = config.clone();
|
|
|
|
agent_config.cwd = root.clone();
|
|
// Consolidation threads must never feed back into phase-1 memory generation.
|
|
agent_config.ephemeral = true;
|
|
agent_config.memories.generate_memories = false;
|
|
agent_config.memories.use_memories = false;
|
|
agent_config.include_apps_instructions = false;
|
|
agent_config.mcp_servers = Constrained::allow_only(HashMap::new());
|
|
// Approval policy
|
|
agent_config.permissions.approval_policy = Constrained::allow_only(AskForApproval::Never);
|
|
// Consolidation runs as an internal worker and must not recursively delegate.
|
|
let _ = agent_config.features.disable(Feature::SpawnCsv);
|
|
let _ = agent_config.features.disable(Feature::Collab);
|
|
let _ = agent_config.features.disable(Feature::MemoryTool);
|
|
let _ = agent_config.features.disable(Feature::Apps);
|
|
let _ = agent_config.features.disable(Feature::Plugins);
|
|
let _ = agent_config
|
|
.features
|
|
.disable(Feature::SkillMcpDependencyInstall);
|
|
|
|
// Sandbox policy
|
|
let writable_roots = vec![root];
|
|
// The consolidation agent only needs local memory-root write access and no network.
|
|
let consolidation_sandbox_policy = SandboxPolicy::WorkspaceWrite {
|
|
writable_roots,
|
|
network_access: false,
|
|
exclude_tmpdir_env_var: true,
|
|
exclude_slash_tmp: true,
|
|
};
|
|
agent_config
|
|
.permissions
|
|
.set_legacy_sandbox_policy(consolidation_sandbox_policy, agent_config.cwd.as_path())
|
|
.ok()?;
|
|
|
|
agent_config.model = Some(
|
|
config
|
|
.memories
|
|
.consolidation_model
|
|
.clone()
|
|
.unwrap_or_else(|| provider.memory_consolidation_preferred_model().to_string()),
|
|
);
|
|
agent_config.model_reasoning_effort = Some(crate::stage_two::REASONING_EFFORT);
|
|
|
|
Some(agent_config)
|
|
}
|
|
|
|
pub(super) fn get_prompt(root: &Path) -> Vec<UserInput> {
|
|
let prompt = build_consolidation_prompt(root);
|
|
vec![UserInput::Text {
|
|
text: prompt,
|
|
text_elements: vec![],
|
|
}]
|
|
}
|
|
|
|
/// Handle the agent while it is running.
|
|
#[allow(clippy::too_many_arguments)]
|
|
pub(super) fn handle(
|
|
context: Arc<MemoryStartupContext>,
|
|
claim: Claim,
|
|
new_watermark: i64,
|
|
selected_outputs: Vec<codex_state::Stage1Output>,
|
|
memory_root: codex_utils_absolute_path::AbsolutePathBuf,
|
|
agent: SpawnedConsolidationAgent,
|
|
phase_two_e2e_timer: Option<codex_otel::Timer>,
|
|
) {
|
|
let Some(db) = context.state_db() else {
|
|
return;
|
|
};
|
|
|
|
tokio::spawn(async move {
|
|
let _phase_two_e2e_timer = phase_two_e2e_timer;
|
|
let SpawnedConsolidationAgent { thread_id, thread } = agent;
|
|
|
|
// Loop the agent until we have the final status.
|
|
let final_status =
|
|
loop_agent(db.clone(), claim.token.clone(), thread_id, &thread).await;
|
|
|
|
if matches!(final_status, AgentStatus::Completed(_)) {
|
|
if let Some(token_usage) = thread
|
|
.token_usage_info()
|
|
.await
|
|
.map(|info| info.total_token_usage)
|
|
{
|
|
emit_token_usage_metrics(context.as_ref(), &token_usage);
|
|
}
|
|
// Do not reset the workspace baseline if we lost the lock.
|
|
let still_owns_lock = match db
|
|
.memories()
|
|
.heartbeat_global_phase2_job(
|
|
&claim.token,
|
|
crate::stage_two::JOB_LEASE_SECONDS,
|
|
)
|
|
.await
|
|
.inspect_err(|err| {
|
|
tracing::error!(
|
|
"failed confirming global memory consolidation ownership before resetting workspace baseline: {err}"
|
|
);
|
|
}) {
|
|
Ok(true) => true,
|
|
Ok(false) => {
|
|
tracing::error!(
|
|
"lost global memory consolidation ownership before resetting workspace baseline"
|
|
);
|
|
false
|
|
}
|
|
Err(_) => {
|
|
job::failed(context.as_ref(), &db, &claim, "failed_confirm_ownership")
|
|
.await;
|
|
false
|
|
}
|
|
};
|
|
if still_owns_lock {
|
|
if let Err(err) = reset_memory_workspace_baseline(&memory_root).await {
|
|
tracing::error!("failed resetting memory workspace baseline: {err}");
|
|
job::failed(context.as_ref(), &db, &claim, "failed_workspace_commit").await;
|
|
} else if !job::succeed(
|
|
context.as_ref(),
|
|
&db,
|
|
&claim,
|
|
new_watermark,
|
|
&selected_outputs,
|
|
"succeeded",
|
|
)
|
|
.await
|
|
{
|
|
tracing::error!(
|
|
"failed marking global memory consolidation job succeeded after resetting workspace baseline"
|
|
);
|
|
}
|
|
}
|
|
} else {
|
|
job::failed(context.as_ref(), &db, &claim, "failed_agent").await;
|
|
}
|
|
|
|
let cleanup_context = Arc::clone(&context);
|
|
tokio::spawn(async move {
|
|
if let Err(err) = cleanup_context
|
|
.shutdown_consolidation_agent(SpawnedConsolidationAgent { thread_id, thread })
|
|
.await
|
|
{
|
|
warn!(
|
|
"failed to auto-close global memory consolidation agent {thread_id}: {err}"
|
|
);
|
|
}
|
|
});
|
|
});
|
|
}
|
|
|
|
async fn loop_agent(
|
|
db: Arc<StateRuntime>,
|
|
token: String,
|
|
thread_id: ThreadId,
|
|
thread: &codex_core::CodexThread,
|
|
) -> AgentStatus {
|
|
let mut heartbeat_interval =
|
|
tokio::time::interval(Duration::from_secs(crate::stage_two::JOB_HEARTBEAT_SECONDS));
|
|
heartbeat_interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
|
|
let mut status_poll_interval = tokio::time::interval(Duration::from_secs(1));
|
|
status_poll_interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
|
|
let session_termination = thread.wait_until_terminated();
|
|
tokio::pin!(session_termination);
|
|
|
|
loop {
|
|
let status = thread.agent_status().await;
|
|
if is_final_agent_status(&status) {
|
|
break status;
|
|
}
|
|
|
|
tokio::select! {
|
|
_ = &mut session_termination => {
|
|
let status = thread.agent_status().await;
|
|
if is_final_agent_status(&status) {
|
|
break status;
|
|
}
|
|
tracing::warn!(
|
|
"memory consolidation agent {thread_id} exited before final status; last status was {status:?}"
|
|
);
|
|
break AgentStatus::Errored(format!(
|
|
"memory consolidation agent exited before final status: {status:?}"
|
|
));
|
|
}
|
|
_ = status_poll_interval.tick() => {
|
|
}
|
|
_ = heartbeat_interval.tick() => {
|
|
match db
|
|
.memories()
|
|
.heartbeat_global_phase2_job(
|
|
&token,
|
|
crate::stage_two::JOB_LEASE_SECONDS,
|
|
)
|
|
.await
|
|
{
|
|
Ok(true) => {}
|
|
Ok(false) => {
|
|
tracing::warn!(
|
|
"lost global phase-2 ownership during heartbeat for memory consolidation agent {thread_id}"
|
|
);
|
|
break AgentStatus::Errored(
|
|
"lost global phase-2 ownership during heartbeat".to_string(),
|
|
);
|
|
}
|
|
Err(err) => {
|
|
tracing::warn!(
|
|
"phase-2 heartbeat update failed for memory consolidation agent {thread_id}: {err}"
|
|
);
|
|
break AgentStatus::Errored(format!(
|
|
"phase-2 heartbeat update failed: {err}"
|
|
));
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
pub(super) fn get_watermark(
|
|
claimed_watermark: i64,
|
|
latest_memories: &[codex_state::Stage1Output],
|
|
) -> i64 {
|
|
latest_memories
|
|
.iter()
|
|
.map(|memory| memory.source_updated_at.timestamp())
|
|
.max()
|
|
.unwrap_or(claimed_watermark)
|
|
.max(claimed_watermark)
|
|
}
|
|
|
|
fn is_final_agent_status(status: &AgentStatus) -> bool {
|
|
!matches!(
|
|
status,
|
|
AgentStatus::PendingInit | AgentStatus::Running | AgentStatus::Interrupted
|
|
)
|
|
}
|
|
|
|
fn emit_metrics(context: &MemoryStartupContext, counters: Counters) {
|
|
if counters.input > 0 {
|
|
context.counter(MEMORY_PHASE_TWO_INPUT, counters.input, &[]);
|
|
}
|
|
|
|
context.counter(
|
|
MEMORY_PHASE_TWO_JOBS,
|
|
/*inc*/ 1,
|
|
&[("status", "agent_spawned")],
|
|
);
|
|
}
|
|
|
|
fn emit_token_usage_metrics(context: &MemoryStartupContext, token_usage: &TokenUsage) {
|
|
context.histogram(
|
|
MEMORY_PHASE_TWO_TOKEN_USAGE,
|
|
token_usage.total_tokens.max(0),
|
|
&[("token_type", "total")],
|
|
);
|
|
context.histogram(
|
|
MEMORY_PHASE_TWO_TOKEN_USAGE,
|
|
token_usage.input_tokens.max(0),
|
|
&[("token_type", "input")],
|
|
);
|
|
context.histogram(
|
|
MEMORY_PHASE_TWO_TOKEN_USAGE,
|
|
token_usage.cached_input(),
|
|
&[("token_type", "cached_input")],
|
|
);
|
|
context.histogram(
|
|
MEMORY_PHASE_TWO_TOKEN_USAGE,
|
|
token_usage.output_tokens.max(0),
|
|
&[("token_type", "output")],
|
|
);
|
|
context.histogram(
|
|
MEMORY_PHASE_TWO_TOKEN_USAGE,
|
|
token_usage.reasoning_output_tokens.max(0),
|
|
&[("token_type", "reasoning_output")],
|
|
);
|
|
}
|