10 KiB
Concepts
Shared domain vocabulary for this project — entities, named processes, and status concepts with project-specific meaning. Seeded with core domain vocabulary, then accretes as ce-compound and ce-compound-refresh process learnings; direct edits are fine. Glossary only, not a spec or catch-all.
Bitable
Bitable
A companion service providing multi-dimensional table storage (structured records with typed fields, views, and formula columns) that runs alongside the main AgentKit server. Owns its own database schema, isolated from core tables, and exposes a REST API boundary that all callers — agents and end-users alike — go through. Agents authenticate with an internal service token; end-users authenticate via JWT.
BitableFile
The top-level container that groups related data tables. A multi-dimensional table file (多维表格文件) holds zero or more tables and owns its own metadata (name, icon, description, owner). Tables reference their parent file via file_id. The file is the unit of ownership sharing and the top of the 文件 → 数据表 → 字段/记录 navigation hierarchy, mirroring the Feishu Bitable App and twenty Object container pattern.
Field Ownership
The model controlling which actor may modify a field's definition or values. Every field has an owner of either the agent or the user. Agent-owned fields are written by the ingestion tooling during data import; user-owned fields are edited by humans in the UI. Upsert operations update only agent-owned fields — user-owned fields are never overwritten by agent writes, so human edits survive re-ingestion.
Recalc
The background process that recomputes formula field values after source data changes. When a record's data columns are written, a recalc task is enqueued; a worker claims tasks atomically using row-level locking so concurrent workers partition the queue without coordination, evaluates the formula DAG in topological order, and writes results back. A reaper resets stale in-progress tasks back to pending for crash recovery.
Expert Orchestration
Disclosure Level
The mode controlling how much of a skill's prompt enters the agent context at load time. Full mode injects the complete skill prompt; summary mode injects only name and description. The default is full so that skills predating the field keep their original behavior — only skills that explicitly opt into summary mode are degraded.
State Offloading
The pattern of moving a phase's full output to a shared workspace while keeping only a short summary and a reference key in memory. Prevents unbounded memory growth across long pipelines. Consumers that need the full content resolve it through the reference key, not by reading the in-memory summary.
Pipeline Checkpoint
A crash-recovery snapshot of a pipeline's progress, persisted per phase. On resume, completed phases are skipped, failed phases are preserved as failed, and pending phases execute. Checkpoints are cleared after successful completion; failed runs retain them for retry.
Debate Phase
A phase dynamically inserted into a team plan when divergence is detected between expert outputs. Subject to a per-pipeline limit that persists across resume — the counter is rebuilt from the restored plan, not reset to zero.
Resume
The act of rebuilding a crashed pipeline's runtime state from persisted checkpoints. Restores completed and failed phase statuses, rebuilds runtime counters, and re-persists any dynamically inserted phases so the restored plan matches what was executing at crash time.
Verify Re-injection
The feedback loop triggered when a verification check fails after a final answer is produced. Errors are injected as a new user message into the conversation and the ReAct loop continues for one retry; a second verification failure interrupts execution and returns the error to the user with the verify log. Bounded to one retry to balance auto-correction against token cost, rather than looping until max_steps.
Three-tier Degradation Chain
The agent-level fallback sequence when the primary agent fails: main agent → Recovery tier (reuses ReflexionEngine for Evaluate→Reflect→Retry) → Emergency tier (rule-based fallback returning a structured error with suggestions). Each tier is independently configurable; the Recovery tier avoids new infrastructure by reusing the existing reflection engine, and the Emergency tier replaces the previous static-text fallback with actionable error structure.
Channels & Caching
Per-User Cache Namespace
A security pattern for LLM response caching where the cache key includes user_id so that cached responses are isolated per user. When per_user_namespace=True, anonymous requests (user_id=None) must be rejected from caching — they cannot be namespaced and would pollute other users' cache blocks. should_cache() enforces this by returning False when the namespace is on but user_id is missing.
Webhook Signature Freshness
The timestamp validation layered on top of webhook signature verification that bounds the replay window. Computes abs(now - ts) against a max-age constant (e.g. 300s) and rejects requests outside the window — the absolute value defends against both historical replays and future-dated requests. Without this, a valid signature alone provides zero replay protection.
Webhook Backpressure
The pattern of bounding a webhook handler's in-flight background task set with a cap (typically max_concurrent * 2) and returning HTTP 429 when exceeded. The 2x margin absorbs short spikes; the 429 forces clients to back off rather than snowballing memory and coroutine exhaustion. The task set is also awaited on app shutdown so in-flight replies are not dropped.
Auth
3-State Startup
The auth store's cold-start state machine with values valid / invalid / error (plus the initial pending). On app boot, beginStartup() runs before app.mount(); the router guard awaits waitForStartup() when state is pending so no routing decision is made until the probe completes. invalid redirects to /login; error lets navigation proceed (retryable, refresh token retained); valid proceeds normally. The distinction matters because a network failure (error) must not force a re-login, while a revoked token (invalid) must.
Real-Time Fan-Out
Service Broadcast Callback
The convention for pushing backend state changes to the user's open frontend tabs in real time without coupling domain services to the WebSocket transport. A service accepts an optional async callback at construction; after a successful mutation it best-effort invokes the callback with a typed message envelope. Delivery failure is logged but never rolls back the mutation — the persisted state is the source of truth, the broadcast is a latency optimization. The callback is wired at the composition root (app lifespan) to the portal's per-user fan-out primitive, so the service stays layer-pure. The same callback shape is shared by CRUD services, reminder dispatchers, and sync providers, giving all real-time updates a single exit point.
PLAN_EXEC
Phase State Machine
The four-phase lifecycle (PLANNING → BUILDING → VERIFICATION → DELIVERY) that constrains which tools an agent may call at each step of a PLAN_EXEC run. Each phase has a tool whitelist (e.g., PLANNING allows search/read_file; BUILDING allows write_file); WILDCARD = "*" means all tools allowed (used by DELIVERY by default). Transitions are LLM-driven via AdvancePhaseTool — the LLM decides when to advance, there is no implicit timer. auto_advance_after_steps opts into a step-count trigger as an alternative to explicit advance_phase calls. State lives in ReActEngine.current_phase; PhaseState is an enum with from_string() parser.
PhasePolicy
The dataclass holding the per-phase whitelist (dict[PhaseState, frozenset[str]]) and bash command filter (dict[PhaseState, Callable[[str], bool] | re.Pattern | None]). Constructed via default_policy() (hardcoded KTD5 defaults) or policy_from_config(plan_exec_cfg) (YAML-driven). policy_from_config returns None for empty/disabled config — signaling "opt out, fall back to REACT" (not "use defaults"). is_bash_command_allowed accepts either a Callable (returns True if dangerous; ShellTool._is_dangerous is the default) or a legacy re.Pattern (matches dangerous substrings). The Callable path closes the regex ceiling — regex missed :>file, dd of=file, and unknown binaries.
Phase Violation
The structured event emitted when _check_phase_permission blocks a tool call in PLAN_EXEC mode. Two emission paths: (1) the violation is re-injected into the LLM as a tool result so the loop continues (the LLM can switch tools or call advance_phase); (2) a phase_violation WS event is forwarded to the client for UI feedback (PhaseIndicator component). Violations carry current_phase, tool, violation_kind (tool_not_allowed / bash_command_blocked), message, and command_preview. The engine accumulates violations in _phase_violations and drains them via _drain_phase_violations after each tool dispatch — drains are the caller's responsibility at three sites in execute_stream.
AdvancePhaseTool
The state-transition tool that moves the engine between phases. Calls engine.advance_phase() which transitions current_phase to the next enum value (PLANNING→BUILDING→VERIFICATION→DELIVERY; DELIVERY is terminal). Returns {"previous_phase", "current_phase", "message"}. Bypasses _check_phase_permission (always permitted) — phase advancement is not subject to the whitelist it enforces. Known limitation: the default _loop_threshold=2 fires on the 2nd identical advance_phase({}) call because all transitions produce the same argument hash — needs exemption from loop detection.
_build_phase_engine
The chat.py helper that consolidates PLAN_EXEC engine construction for both WS and REST paths. Returns (engine, tools_with_advance_phase, error_message). Returns (None, None, None) when execution_mode != PLAN_EXEC or plan_exec.enabled=False (both signal "fall back to REACT"). Returns (None, None, error_message) on policy construction failure. The helper ensures WS and REST paths share identical PhasePolicy + AdvancePhaseTool wiring — previously the REST path returned 501 because it had no construction logic.