# Loop
A complete record of one agent-loop execution, stored as `LoopRecord`. Loops are the iterations within a Session. Each Loop contains Turns (steps), tracks its model/provider configuration, accumulates usage, and links to parent/child loops for tree navigation.
Loops are created by `agent_loop` (origin loops) or `agent_loop_continue` (continuation loops). The `SessionRecorder` materializes `LoopRecord` structs from the `AgentStart` / `AgentEnd` event pairs.
## Concept Overview
```
Loop [EXISTS — LoopRecord]
├── HEADER
│ ├── loop_id [EXISTS] — "{session_id}.{config_segment}.{N}"
│ ├── status [EXISTS] — Pending/Running/Completed/Rejected/Aborted
│ ├── continuation_kind [EXISTS] — Initial/Default/Rerun/Branch/Compaction
│ ├── parent_loop_id [EXISTS]
│ ├── timing [EXISTS] — started_at, ended_at
│ ├── Model [EXISTS] — falls back: Loop → Agent default
│ ├── config [EXISTS] — LoopConfigSnapshot
│ ├── usage, compaction_block [EXISTS]
│ └── Callbacks: before_loop / after_loop / on_error [EXISTS]
├── LINE ITEMS: Turns [EXISTS as events and struct]
├── LINE ITEMS: Same-session children, Sub-agent spawns [EXISTS]
├── LINE ITEMS: Parallel group [EXISTS]
└── LINE ITEMS: Events [EXISTS]
```
---
## HEADER
| `loop_id` | `String` | `[EXISTS]` | Unique identifier. Format: `"{session_id}.{config_segment}.{N}"`. The config_segment encodes which model/provider produced this loop. N is a monotonic counter per (session, config). |
| `session_id` | `String` | `[EXISTS]` | Session this loop belongs to. |
| `agent_id` | `String` | `[EXISTS]` | Agent that ran this loop. |
| `status` | `LoopStatus` | `[EXISTS]` | Lifecycle state: Pending, Running, Completed, Rejected, Aborted. See Status section below. |
| `continuation_kind` | `ContinuationKind` | `[EXISTS]` | How this loop relates to its parent. `Initial` for origin loops (`agent_loop`). `Default` for regular continuations. `Rerun` for retries. `Branch` for branch explorations. `Compaction` for standalone compaction passes. |
| `parent_loop_id` | `Option<String>` | `[EXISTS]` | The loop that directly preceded this one. `None` for origin loops. For sub-agent loops, points to the tool-call loop in a different session. |
| `started_at` | `DateTime<Utc>` | `[EXISTS]` | Timestamp from `AgentStart`. |
| `ended_at` | `Option<DateTime<Utc>>` | `[EXISTS]` | Timestamp from `AgentEnd`. `None` while running or pending. |
| `rejection` | `Option<String>` | `[EXISTS]` | Set when `AgentEnd.rejection` is `Some` (input filter blocked the run). |
| `metadata` | `Option<serde_json::Value>` | `[EXISTS]` | Opaque caller-supplied metadata from `AgentStart` (e.g., request id, trace ID). |
### Model for this Loop `[EXISTS]`
The model/provider identity is captured as a lightweight snapshot, not the full config (which contains secrets and non-serializable closures).
| `config` | `Option<LoopConfigSnapshot>` | `[EXISTS]` | Populated from `AgentStart.config_snapshot` or the first `Message::Assistant` seen. `None` if loop ended before any assistant message and no snapshot was provided. |
| `config.model` | `String` | `[EXISTS]` | Model id string (e.g., `"claude-opus-4-6"`, `"gpt-4o"`). |
| `config.provider` | `String` | `[EXISTS]` | Provider name (e.g., `"anthropic"`, `"openai"`). |
| `config.config_id` | `Option<String>` | `[EXISTS]` | Stable config identity from `AgentLoopConfig.config_id`. Matches the `config_segment` in `loop_id`. |
| `config.name` | `Option<String>` | `[EXISTS]` | Model display name. |
| `config.api` | `Option<ApiProtocol>` | `[EXISTS]` | Which API protocol was used (e.g., `AnthropicMessages`, `OpenAiCompletions`). |
| `config.base_url` | `Option<String>` | `[EXISTS]` | Provider base URL. |
| `config.reasoning` | `Option<bool>` | `[EXISTS]` | Whether this model supports reasoning/thinking. |
| `config.context_window` | `Option<u32>` | `[EXISTS]` | Context window size in tokens. |
| `config.max_tokens` | `Option<u32>` | `[EXISTS]` | Max output tokens per response. |
| `config.thinking_level` | `Option<ThinkingLevel>` | `[EXISTS]` | Reasoning depth level for this loop. Formerly a Session-level attribute; now per-loop. |
| `config.temperature` | `Option<f32>` | `[EXISTS]` | Sampling temperature. Formerly a Session-level attribute; now per-loop. |
**Model fallback hierarchy**: Loop (`AgentLoopConfig.model_config`) -> Agent default (`BasicAgent.model_config`).
### Usage `[EXISTS]`
| `usage` | `Usage` | `[EXISTS]` | Token usage from `AgentEnd.usage`. Accumulated across all turns in this loop. Fields: input, output, reasoning, cache_read, cache_write, total_tokens. |
### Compaction `[EXISTS]`
| `compaction_block` | `Option<CompactionBlock>` | `[EXISTS]` | Non-destructive compaction overlay. When `Some`, the context loader uses this block instead of raw messages. Original messages remain untouched. |
### Status `[EXISTS]`
Lifecycle state of a `LoopRecord`. Enum `LoopStatus`:
```
Pending -> Running -> Completed
-> Rejected
-> Aborted
```
| `Pending` | `[EXISTS]` | Loop id appeared in `ParallelLoopStart` but `AgentStart` has not yet arrived. Only for parallel-evaluation branches. |
| `Running` | `[EXISTS]` | `AgentStart` was received; the loop is executing. |
| `Completed` | `[EXISTS]` | `AgentEnd` was received with no rejection. |
| `Rejected` | `[EXISTS]` | `AgentEnd` was received with `rejection: Some(_)`. Input filter blocked the run. |
| `Aborted` | `[EXISTS]` | `SessionRecorder::flush` was called before `AgentEnd` arrived (e.g., process shutdown). |
### Callbacks `[EXISTS]`
| `before_loop` | `[EXISTS]` | Fires before `AgentStart` is emitted. Defined as `BeforeLoopFn` on `AgentLoopConfig`. Blank by default. |
| `after_loop` | `[EXISTS]` | Fires after `AgentEnd` is emitted. Defined as `AfterLoopFn`. Receives messages and usage. Blank by default. |
| `on_error` | `[EXISTS]` | Fires when `StopReason::Error` is encountered. Defined as `OnErrorFn`. Blank by default. |
---
## LINE ITEMS: Turns (Steps) `[EXISTS]` as events and struct
Turns exist as `TurnStart` / `TurnEnd` event pairs in the loop's event stream, and as materialized `Turn` structs on `LoopRecord.turns`. See [turn.md](turn.md).
| `turns` | `Vec<Turn>` | `[EXISTS]` | Materialized turn records. Built by `SessionRecorder` from event pairs. Empty for old sessions (backward compat via `#[serde(default)]`). |
| (event-pair) | — | `[EXISTS]` | Each turn is also bounded by `TurnStart` and `TurnEnd` events in `self.events`. |
---
## LINE ITEMS: Same-session Children `[EXISTS]`
| `children_loop_ids` | `Vec<String>` | `[EXISTS]` | Loop IDs of same-session child loops (continuations, reruns, branches). Parent->children direction. Does not include cross-session sub-agent children. |
---
## LINE ITEMS: Sub-agent Spawns (Cross-session) `[EXISTS]`
| `child_loop_refs` | `Vec<ChildLoopRef>` | `[EXISTS]` | Cross-session links to sub-agent loops spawned by tool calls. Each entry has: `tool_call_id`, `tool_name`, `child_loop_id`, `child_session_id`. |
`ChildLoopRef` fields:
| `tool_call_id` | `String` | `[EXISTS]` | The `ToolCall.id` that triggered sub-agent execution. |
| `tool_name` | `String` | `[EXISTS]` | The tool name that performed the spawn. |
| `child_loop_id` | `String` | `[EXISTS]` | The sub-agent's `AgentStart.loop_id`. |
| `child_session_id` | `String` | `[EXISTS]` | The sub-agent's session. Extracted from `child_loop_id` prefix. |
---
## LINE ITEMS: Parallel Group `[EXISTS]`
Set when this loop was part of an evaluational-parallelism group (`agent_loop_parallel`).
| `parallel_group` | `Option<ParallelGroupRecord>` | `[EXISTS]` | `None` for non-parallel loops. |
| `all_loop_ids` | `Vec<String>` | `[EXISTS]` | All branch loop IDs in config order. |
| `selected_loop_id` | `String` | `[EXISTS]` | The winning branch's loop ID. |
| `selected_config_index` | `usize` | `[EXISTS]` | 0-based index of the winner in the original configs. |
| `evaluation_usage` | `Usage` | `[EXISTS]` | Token usage from the judge LLM (zero for non-judge strategies). |
| `is_selected` | `bool` | `[EXISTS]` | `true` if this `LoopRecord` is the evaluation winner. |
---
## LINE ITEMS: Events `[EXISTS]`
| `events` | `Vec<LoopEvent>` | `[EXISTS]` | Ordered event stream for this loop. |
Each `LoopEvent` has:
| `sequence` | `u64` | `[EXISTS]` | Monotonic counter (0-based). Gaps indicate filtered events (e.g., streaming deltas when `include_streaming_events` is false). |
| `event` | `AgentEvent` | `[EXISTS]` | The original event. `event.loop_id()` matches this `LoopRecord.loop_id`. |
---
## Messages `[EXISTS]`
| `messages` | `Vec<AgentMessage>` | `[EXISTS]` | All new messages produced by this loop, from `AgentEnd.messages`. Authoritative for replay and branching. |
---
## Loop Origin Classification
| `None` | `Initial` | Fresh origin loop (`agent_loop`) |
| `Some(p)`, same session | `Default` | Regular continuation |
| `Some(p)`, same session | `Rerun` | Retry / error recovery |
| `Some(p)`, same session | `Branch` | Branch exploration |
| `Some(p)`, different session | `Initial` | Sub-agent loop (spawned by a tool) |
---
## Code Reference
| `src/session/model.rs` | `LoopRecord` struct, `LoopStatus` enum, `LoopConfigSnapshot` struct, `ChildLoopRef` struct, `ParallelGroupRecord` struct, `LoopEvent` struct, `OpenLoop` struct. |
| `src/agent_loop/run.rs` | `run_loop` function — the core loop engine. Implements the outer loop (follow-ups) and inner loop (tool calls + steering). Accumulates `Usage`, fires turn events and hooks. |
---
## Conceptual Notes
- **Model fallback** is Loop -> Agent default. Session no longer carries model/thinking/temperature fields; these are tracked per-loop in `LoopConfigSnapshot`.
- **Turns as a struct** are materialized on `LoopRecord.turns` as `Vec<Turn>`. Built by `SessionRecorder` from `TurnStart`/`TurnEnd` event pairs. The flat `messages` field is kept independently for compaction and context building. Old sessions without `turns` deserialize with an empty vec.
- **LoopConfigSnapshot** intentionally does not store the full `AgentLoopConfig` because it contains API keys and non-serializable hook closures. The snapshot captures model identity plus key parameters (thinking_level, temperature, context_window, max_tokens, etc.) for cost attribution, replay identification, parallel branch differentiation, and per-loop config tracking.