Glossary
To go from zero to understanding DSH, skim these terms first. Each entry gives a "one-liner" and a source link.
A–Z
ACP (Agent Client Protocol)
An automation transport (not a human-interaction layer) that exposes a harness agent to programmatic clients; the cross-process sub Agent provider acp uses it. → Sub Agents
Agent-default-model
The default model selection (ctx.agentDefaultModel) shared by every Agent entry point, requiring { provider, model }; used by web / headless / API entry points and can be overridden by in-session Models settings. → Multi-model
Attachment
The persistent binary attachment seam behind ctx.attachments: validates and atomically commits immutable image bytes, while the session log stores only the sha256: reference and validated metadata. → Attachments
Bundle
A plugin bundle carrying its own patch, added to a profile as one layer. Declared by dsh.bundle.patch. → Plugin Anatomy
Checkpoint (semantic checkpoint)
session-checkpoint-policy forces a flush before model requests, before tool external side effects, and at each step boundary; on refusal it fails closed. → Session System
Compaction
The compaction capability family: ctx.compaction seam plus the compaction-basic backend, which folds the oldest surface into a summary under token pressure, reusing the KV cache. → Context and Tokens
Cordis
DSH's composition framework (vendored, published only as @deepseek-ai/cordis), providing Context, dependency injection, plugin mounting, and events. → Architecture
Credentials (credential references)
The ctx.credentials seam: config stores only references, not secrets, resolved per operation; priority is process environment > .credentials.yaml > working directory .env > $DSH_HOME/.env. → Credential Management
Dual-face (client dual-face)
A package that has both a node half and a browser half (dsh.client). → Web UI
E2B
An experimental cloud sandbox POC: moves the filesystem + subprocess execution world into an E2B Linux sandbox; opt-in, not mounted by default. → E2B Cloud Sandbox
Feedback
Two contracts: /feedback appends a log-only feedback/record event; message feedback is an editable rating sidecar bound to a single assistant message; neither enters the model. → Feedback
Guard (behavior guard)
packages/guard, two self-contained behavior plugins (repeat-tool-reminder / timeout-policy), watching for no-output patterns and per-call budgets; a non-replaceable capability. → Tool Execution
Patch
Stacking entries in cordis.patch.yml (insert / override by id) that decide "which layer overrides which". → Plugin Anatomy
Persona
Composable persona rows, shadowing deployment personas. → Agent Presets
Plan-mode (planning mode)
Per-agent planning collaboration state plus a policy prompt section (soft guidance); entered via /plan, exited with exit_plan_mode approval; does not switch models. → Multi-model
Preset (permission preset)
An operating-mode profile (workspace-write / danger-full-access) plus approval. → Permissions
Preset (agent-presets)
The per-session fixed composition (agent.cordis.yml). → Agent Presets
Profile
The $DSH_HOME/profiles/<name> directory, holding bundle layers plus user patch layers. → Configuration
Projection (session projection)
ctx.sessionProjections, whole-log derived views for clients (current surface / title / token ledger); fold shortcuts, non-authoritative. → Session System
Scope
Context layering (global / agent) that determines visibility and masking. → Plugin Anatomy
Seam (capability seam)
Abstract capability interfaces (ctx.llm / ctx.tools / ctx.sandbox...), with swappable implementations. → Architecture
Session / Turn / Step
The three-level lifecycle of session / turn / step. → First Session
Session query
The ctx.sessionQuery session-retrieval seam: authorized reads of live/durable logs, relationship traces, SQLite FTS5 search; the session_search tool is not mounted by default. → Session Query
Spill
The ctx.spillStore tool-output overflow seam: overly long plaintext is written to a session-scoped file, while the model side gets a bounded preview plus a locator. → Spill Storage
Surface
The derived message layer on top of the session log that the model sees. → Sessions
Token-meter
The replayable metering singleton ctx.tokenMeter, estimating with a fixed heuristic (4 chars/token plus structural overhead); the basis for compaction decisions. → Context and Tokens
Tool
A model-callable capability (schema + output + execute). → Tool Execution
The Easily Confused Pairs
| Pair | Explanation |
|---|---|
tools/* vs tool/* | Cordis in-process notifications (plural) vs session log events (singular). → Event System |
| patch layer vs settings layer | static deployer configuration vs user runtime settings. → Configuration |
| permission vs sandbox | operating-mode profiles vs command execution boundaries. → Permissions |
| plan-mode vs multi-model | /plan does not change the model; multi-model relies on multiple providers. → Models |
| surface vs projection | the derived message layer the model sees vs log-derived views for clients. → Sessions |
| shell vs terminal | one-shot command wrapper vs persistent owner-scoped interactive terminal session. → Subprocess and Terminal |
Read Deeper as Needed
Each term is expanded in detail on its corresponding page; to follow a knowledge-level path, start from the Learning Path.