zeph-core
Core agent loop, configuration, context builder, metrics, vault, and sub-agent orchestration for Zeph.
Overview
Core orchestration crate for the Zeph agent. Manages the main agent loop, bootstraps the application from TOML configuration with environment variable overrides, and assembles the LLM context from conversation history, skills, and memory. Includes sub-agent orchestration with zero-trust permission grants, background execution, filtered tool/skill access, persistent memory scopes, lifecycle hooks, persistent JSONL transcript storage with resume-by-ID, A2A-based in-process communication channels, and /agent CLI commands for runtime management. All other workspace crates are coordinated through zeph-core.
Key modules
| Module | Description |
|---|---|
agent |
Agent<C> — main loop driving inference and tool execution; ToolExecutor erased via Box<dyn ErasedToolExecutor>; supports external cancellation via with_cancel_signal(); EnvironmentContext cached at bootstrap and partially refreshed (git branch, model name) on skill reload only |
agent::context_manager |
ContextManager — owns token budget, compaction threshold, and safety margin; should_compact() is O(1) — reads cached_prompt_tokens set by the LLM response rather than scanning the message list |
agent::tool_orchestrator |
ToolOrchestrator — owns max iteration limit, doom-loop detection (rolling hash window with in-place hashing, no intermediate String allocation), summarization flag, and overflow config |
agent::learning_engine |
LearningEngine — owns LearningConfig, tracks per-turn reflection state; delegates self-learning decisions to is_enabled() / mark_reflection_used() |
agent::feedback_detector |
FeedbackDetector (regex) and JudgeDetector (LLM-backed) — implicit correction detection from user messages; JudgeDetector runs in background via tokio::spawn with sliding-window rate limiter (5 calls / 60 s) and XML-escaped adversarial-defense prompt; adaptive threshold gates judge invocation to the regex uncertainty zone |
agent::tool_execution |
Tool call handling, redaction, result processing; both the fenced-block path (handle_tool_result) and the structured tool-call path unconditionally emit LoopbackEvent::ToolStart (UUID generated per call) before execution and LoopbackEvent::ToolOutput (matching UUID, is_error flag) after; call_llm_with_retry() / call_chat_with_tools_retry() — auto-detect ContextLengthExceeded, compact context, and retry (max 2 attempts); prune_stale_tool_outputs invokes count_tokens once per ToolResult part |
agent::message_queue |
Message queue management |
agent::builder |
Agent builder API |
agent::commands |
Chat command dispatch (skills, feedback, skill management via /skill install, /skill remove, /skill reject <name> <reason>, sub-agent management via /agent, etc.) |
agent::utils |
Shared agent utilities |
bootstrap |
AppBuilder — fluent builder for application startup; split into submodules: config (config resolution, vault arg parsing), health (health check, provider warmup), mcp (MCP manager and registry), provider (provider factory functions), skills (skill matcher, embedding model helpers) |
channel |
Channel trait defining I/O adapters; LoopbackChannel / LoopbackHandle for headless daemon I/O (LoopbackHandle exposes cancel_signal: Arc<Notify> for session cancellation); LoopbackEvent::ToolStart / LoopbackEvent::ToolOutput carry per-tool UUIDs and is_error flag for ACP lifecycle notifications; Attachment / AttachmentKind for multimodal inputs |
config |
TOML config with ZEPH_* env overrides; typed ConfigError (Io, Parse, Validation, Vault) |
context |
LLM context assembly from history, skills, memory; resilient compaction with reactive context-overflow retry (max 2 attempts), middle-out progressive tool response removal (10/20/50/100% tiers), 9-section structured compaction prompt, LLM-free metadata fallback via build_metadata_summary() with safe UTF-8 truncation; parallel chunked summarization; tool-pair summarization via maybe_summarize_tool_pair() — when visible pairs exceed tool_call_cutoff, oldest pair is LLM-summarized with XML-delimited prompt and originals hidden via agent_visible=false; visibility-aware history loading (agent-only vs user-visible messages); durable compaction via replace_conversation(); active context compression via CompressionStrategy (reactive/proactive) compresses before capacity limits are hit; uses shared Arc<TokenCounter> for accurate tiktoken-based budget tracking |
cost |
Token cost tracking and budgeting |
daemon |
Background daemon mode with PID file lifecycle (optional feature) |
metrics |
Runtime metrics collection |
project |
Project-level context detection |
redact |
Regex-based secret redaction (AWS, OpenAI, Anthropic, Google, GitLab, HuggingFace, npm, Docker) |
vault |
Secret storage and resolution via vault providers (age-encrypted read/write); secrets stored as BTreeMap for deterministic JSON serialization on every vault.save() call; scans ZEPH_SECRET_* keys to build the custom-secrets map used by skill env injection; all secret values are held as Zeroizing<String> (zeroize-on-drop) and are not Clone |
instructions |
load_instructions() — auto-detects and loads provider-specific instruction files (CLAUDE.md, AGENTS.md, GEMINI.md, zeph.md) from the working directory; injects content into the volatile system prompt section with symlink boundary check, null byte guard, and 256 KiB per-file size cap. InstructionWatcher subscribes to filesystem events via notify-debouncer-mini (500 ms debounce) and reloads instruction_blocks in-place on any .md change — no agent restart required |
skill_loader |
SkillLoaderExecutor — ToolExecutor that exposes the load_skill tool to the LLM; accepts a skill name, looks it up in the shared Arc<RwLock<SkillRegistry>>, and returns the full SKILL.md body (truncated to MAX_TOOL_OUTPUT_CHARS); skill name is capped at 128 characters; unknown names return a human-readable error message rather than a hard error |
scheduler_executor |
SchedulerExecutor — ToolExecutor that exposes three LLM-callable tools: schedule_periodic (add a recurring cron task), schedule_deferred (add a one-shot task at a specific ISO 8601 UTC time), and cancel_task (remove a task by name); communicates with the scheduler via mpsc::Sender<SchedulerMessage> and validates input lengths and cron expressions before forwarding; only present when the scheduler feature is enabled |
hash |
content_hash — BLAKE3 hex digest utility |
pipeline |
Composable, type-safe step chains for multi-stage workflows |
subagent |
Sub-agent orchestration: SubAgentManager lifecycle with background execution, SubAgentDef YAML definitions with 4-level resolution priority (CLI > project > user > config) and scope labels, PermissionGrants zero-trust delegation, FilteredToolExecutor scoped tool access (with tools.except additional denylist), PermissionMode enum (Default, AcceptEdits, DontAsk, BypassPermissions, Plan), max_turns turn cap, A2A in-process channels, SubAgentState lifecycle enum (Submitted, Working, Completed, Failed, Canceled), real-time status tracking, persistent JSONL transcript storage with resume-by-ID (TranscriptWriter/TranscriptReader, TranscriptMeta sidecar, prefix-based ID lookup, automatic old transcript sweep); CRUD helpers: serialize_to_markdown() (round-trip Markdown serialization), save_atomic() (write-rename with parent-dir creation and name validation), delete_file(), default_template() (scaffold for new definitions); AgentsCommand enum drives the zeph agents CLI subcommands |
subagent::hooks |
Lifecycle hooks for sub-agents: HookDef (shell command with timeout and fail-open/closed policy), HookMatcher (pipe-separated tool-name patterns), SubagentHooks (per-agent PreToolUse/PostToolUse from YAML frontmatter); config-level SubagentStart/SubagentStop events; fire_hooks() executes sequentially with env-cleared sandbox and child kill on timeout |
subagent::memory |
Persistent memory scopes for sub-agents: MemoryScope enum (User, Project, Local), resolve_memory_dir() / ensure_memory_dir() for directory lifecycle, load_memory_content() reads MEMORY.md (first 200 lines, 256 KiB cap, symlink boundary check, null byte guard), escape_memory_content() prevents prompt injection via <agent-memory> tag escaping. Memory is auto-injected into the sub-agent system prompt and Read/Write/Edit tools are auto-enabled |
Re-exports: Agent, content_hash, DiffData
Configuration
Key AgentConfig fields (TOML section [agent]):
| Field | Type | Default | Env override | Description |
|---|---|---|---|---|
name |
string | "zeph" |
— | Agent display name |
max_tool_iterations |
usize | 10 |
— | Max tool calls per turn |
summary_model |
string? | null |
— | Model used for context summarization |
auto_update_check |
bool | true |
ZEPH_AUTO_UPDATE_CHECK |
Check GitHub releases for a newer version on startup / via scheduler |
Key InstructionConfig fields (TOML section [agent.instructions]):
| Field | Type | Default | Description |
|---|---|---|---|
auto_detect |
bool | true |
Auto-detect provider-specific files (CLAUDE.md, AGENTS.md, GEMINI.md) |
extra_files |
Vec<PathBuf> |
[] |
Additional instruction files (absolute or relative to cwd) |
max_size_bytes |
u64 | 262144 |
Per-file size cap (256 KiB); files exceeding this are skipped |
[!NOTE]
zeph.mdand.zeph/zeph.mdare always loaded regardless ofauto_detect. Use--instruction-file <path>at the CLI to supply extra files at startup without modifying the config file.
[!TIP] Instruction files support hot reload — edit any watched
.mdfile while the agent is running and the updated content is applied within 500 ms on the next inference turn. The watcher starts automatically when at least one instruction path is resolved.
Key DocumentConfig fields (TOML section [memory.documents]):
| Field | Type | Default | Description |
|---|---|---|---|
collection |
string | "zeph_documents" |
Qdrant collection for document chunks |
chunk_size |
usize | 512 |
Target tokens per chunk |
chunk_overlap |
usize | 64 |
Overlap between chunks |
top_k |
usize | 3 |
Max chunks injected per context-build turn |
rag_enabled |
bool | false |
Enable automatic RAG context injection from zeph_documents |
Key MemoryConfig fields (TOML section [memory]):
| Field | Type | Default | Description |
|---|---|---|---|
vector_backend |
"qdrant" / "sqlite" |
"qdrant" |
Vector search backend |
token_safety_margin |
f32 | 1.0 |
Safety multiplier for tiktoken-based token budget (validated: must be >= 1.0) |
redact_credentials |
bool | true |
Scrub secrets and paths before LLM context injection |
autosave_assistant |
bool | false |
Persist assistant responses to semantic memory automatically |
autosave_min_length |
usize | 20 |
Minimum response length (chars) to trigger autosave |
tool_call_cutoff |
usize | 6 |
Max visible tool call/response pairs before oldest is summarized via LLM |
sqlite_pool_size |
u32 | 5 |
SQLite connection pool size for memory storage |
response_cache_cleanup_interval_secs |
u64 | 3600 |
Interval for expiring stale response cache entries |
[]
= true # set to false to disable update notifications
Set ZEPH_AUTO_UPDATE_CHECK=false to disable without changing the config file.
Skill commands
| Command | Description |
|---|---|
/skill list |
List loaded skills with trust level and match count |
/skill install <url> |
Install a skill from a remote URL |
/skill remove <name> |
Remove an installed skill |
/skill reject <name> <reason> |
Record a typed rejection and trigger immediate skill improvement |
[!TIP]
/skill rejectprovides the strongest feedback signal. The rejection is persisted with aFailureKinddiscriminant to theoutcome_detailcolumn and immediately updates the Wilson score posterior for Bayesian re-ranking.
Self-learning configuration
Key AgentConfig.learning fields (TOML section [agent.learning]):
| Field | Type | Default | Description |
|---|---|---|---|
correction_detection |
bool | true |
Enable FeedbackDetector implicit correction capture |
correction_confidence_threshold |
f64 | 0.7 |
Minimum detector confidence to persist a UserCorrection |
correction_recall_limit |
usize | 5 |
Max corrections retrieved per context-build turn |
correction_min_similarity |
f64 | 0.75 |
Minimum embedding similarity for correction recall |
detector_mode |
"regex" / "judge" |
"regex" |
Detection strategy: regex-only or LLM-backed judge with adaptive regex fallback |
judge_model |
string | "" |
Model for the judge detector (e.g. "claude-sonnet-4-6"); empty = use primary provider |
judge_adaptive_low |
f32 | 0.5 |
Regex confidence below this value skips judge invocation (treated as "not a correction") |
judge_adaptive_high |
f32 | 0.8 |
Regex confidence above this value skips judge invocation (high-confidence regex match accepted) |
Key LlmConfig fields for EMA routing (TOML section [llm]):
| Field | Type | Default | Description |
|---|---|---|---|
router_ema_enabled |
bool | false |
Enable per-provider EMA latency tracking and reordering |
router_ema_alpha |
f64 | 0.1 |
EMA smoothing factor (lower = slower adaptation) |
router_reorder_interval |
u64 | 60 |
Seconds between provider list reordering |
Sub-agent Commands
In-session commands for managing sub-agents:
| Command | Description |
|---|---|
/agent list |
List available sub-agent definitions |
/agent spawn <name> <prompt> |
Spawn a sub-agent with a task prompt |
/agent bg <name> <prompt> |
Spawn a background sub-agent |
/agent status |
Show active sub-agents with state, turns, and elapsed time |
/agent cancel <id> |
Cancel a running sub-agent by ID prefix |
/agent resume <id> <prompt> |
Resume a completed sub-agent session with a new prompt (restores JSONL transcript history) |
/agent approve <id> |
Approve a pending secret request |
/agent deny <id> |
Deny a pending secret request |
@agent_name <prompt> |
Mention shorthand for /agent spawn (disambiguated from file references) |
Sub-agents run as independent tokio tasks with their own LLM provider and filtered tool executor. Each sub-agent receives only explicitly granted tools, skills, and secrets via PermissionGrants. Conversation history is persisted as JSONL transcripts with .meta.json sidecars, enabling session resumption via /agent resume <id> <prompt> — the resumed agent inherits the original definition, tools, and full message history.
Lifecycle hooks can be attached at two levels: config-level SubagentStart/SubagentStop hooks (in [agents.hooks]) fire on spawn and completion, while per-agent PreToolUse/PostToolUse hooks (defined in the agent YAML frontmatter) fire around each tool call, matched by pipe-separated tool-name patterns. All hooks run as shell commands in an env-cleared sandbox with configurable timeout and fail-open/closed policy.
Agents management CLI
zeph agents provides CRUD management of sub-agent definition files outside of a running session:
| Command | Description |
|---|---|
zeph agents list |
Print all discovered definitions with name, scope, description, and model |
zeph agents show <name> |
Print full detail of a single definition |
zeph agents create <name> --description <desc> [--dir <path>] [--model <id>] |
Scaffold a new .md definition via default_template + save_atomic |
zeph agents edit <name> |
Open the definition file in $VISUAL / $EDITOR (validates parse on exit) |
zeph agents delete <name> [--yes] |
Delete a definition file with interactive confirmation |
[!TIP] The same CRUD operations are available interactively in the TUI agents panel — press
ain the TUI to open the panel, thenc(create),e(edit),d(delete), Enter (detail view).
Installation
License
MIT