15 KiB
DeepSeek Harness Architecture
The DeepSeek Harness SDK builds on Cordis: everything is a plugin, including the shipped loop.
Overview
A harness is one Cordis context. Packages add services (ctx.llm, ctx.tools, ctx.sessions), typed events (agent/request, tools/pre-execute, session/event), and disposable prompt, tool, provider, adapter, and listener registrations.
packages/core/ groups the default agent flow; surrounding capabilities are equally first-class Cordis plugins.
Default Services
| ctx key | Package | Role |
|---|---|---|
| — | dsh-scope |
scoped-context registration primitive (library) |
ctx.sessions |
dsh-session |
in-memory event-sourced sessions |
ctx.systemPrompt |
dsh-system-prompt |
ordered prompt sections, tool schemas, and prompt variables |
ctx.tools |
dsh-tools |
tool registry and execution pipeline |
ctx.agents |
dsh-agent |
live agents, delegated creation, agent/* events, and process-local initiator scope |
ctx.agentLoop |
dsh-agent-loop |
concrete Agent driver |
Capability Services
| ctx key | Package family | Role |
|---|---|---|
ctx.llm |
llm/ |
adapter registry and streaming model calls |
ctx.tokenMeter |
llm/token-meter |
singleton replay-aware request/surface pressure |
ctx.bash |
bash/ |
foreground/background command execution |
ctx.sandbox |
sandbox/ |
same-world process confinement (argv wrapping, per-call policy) |
ctx.codeRuntime |
code-runtime/ |
model-written program execution |
ctx.fs |
fs/ |
filesystem provider primitives and policy events |
ctx.skills |
skill/ |
skill provider registry and progressive disclosure |
ctx.web |
web/ |
search/fetch provider registries |
ctx.compact |
compact/ |
session-log compaction |
ctx.subagents |
subagent/ |
named delegation providers |
ctx.tasks |
tasks/ |
background task registry + generic task_* control tools |
ctx.workflows |
workflow/ |
script-driven multi-agent orchestration |
ctx.goals |
goal/ |
persisted same-session goals |
ctx.sessionPersistence |
session-persistence/ |
durable storage for session logs |
ctx.sessionQuery |
session-query/ |
live-preferred logical-corpus exact reads and relationship traces |
Event
Events form the service extension API; see the exhaustive events catalog and producer/consumer map.
Event Domains
- Session events are durable, replayable facts: boundaries, messages, tool activity, steering, compaction, and tool-owned records append to the log and flow through
session/event. - Agent events carry the live
Agenthandle for status, diagnostics, prompt admission, request shaping, result validation, and continuation policy. - Capability events belong to their owning seam;
tools/*,llm/*,system-prompt/*,fs/*, andsubagent/*attach policy and adapters without importing the loop.
Interception Semantics
Waterfall events behave like around-middleware: a listener delegates by calling next(); returning without it vetoes or takes over. Full rule: Cordis waterfall semantics.
Default Loop Lifecycle
The shipped loop drains prompts through checkpoints; every pause is exposed to plugins through services or events.
A turn drains queued input in an append-only session until the model requests no more tools or plugin continuation. A step is one model request plus its tool executions. In the flow below (sequence companion), quoted names are durable session events and event names are extension points.
Startup resolves identity. No id mints <config-id>-session-<uuid>; sessionId resumes or creates; resumeSessionId requires history. Active failures emit agent-loop/config-start-failed(sessionId, error), so front doors reject work; teardown stays silent.
Turn Flow
choose declarative identity and fresh/resume path
-> prepare private session + agent.ctx -> await unpublished setup
-> enter session + agent -> session/created -> agent/created
-> enable driving -> agent/session-start(source) -> start driver
forever:
wait for queued messages
emit agent/status(running)
TURN:
'turn/start'
each queued message -> agent/prompt-submit
allowed prompt -> 'user/message' plus injected context
every prompt blocked -> 'turn/end'(rejected)
STEP loop:
drain steering
assemble system prompt and tool schemas
agent/session-prefix (first step)
agent/pre-step
snapshot the derived messages (the reconstruction boundary)
'step/start'
agent/request (config only) -> log request/header -> llm/stream (frozen)
on final adapter-path or terminal in-band failure:
'step/end'
agent/request-error(original error, consecutive retry attempt, signal)
retry in the next numbered step or preserve the original error
otherwise:
'assistant/chunk'
agent/step-result
'assistant/message' (transformed content or empty success anchor after step-result rejection)
schedule tool calls by ctx.tools.executionMode:
exclusive -> one-call barrier
parallel -> rolling pool, <= maxParallelToolCalls in flight; reclassify before start
each start -> 'tool/call' -> ordered tools/pre-execute -> concurrent tools/execute
each model-order result -> ordered tools/post-execute -> 'tool/result'
append accepted tool-batch context after all recorded results, then steering
agent/post-step
'step/end'
agent/turn-continuation
agent/turn-stop (terminal policy)
stop unless tools or continuation policy ask for another step
'turn/end'
checkpoint persistence and notify idle/running status
Each step assembles ordered prompt sections, tool schemas, and {{name}} variables; unknown or valueless references fail the turn. dsh-system-prompt owns the harness identity and default persona, which an agent scope may shadow. The loop supplies model and cwd (prompt-ownership RFC).
Tool-time context—including async agent.inject() notices and post-tool additionalContexts—settles after recorded results. Steering drains before agent/post-step, which observes durable output, results, context, and steering before signal closure. Leftovers become queued input. Terminal agent/turn-stop runs after continuation and steering folding, remains authoritative through turn close and flush, and discards later steering while preserving queued prompts.
dsh-compact-basic handles pressure and canonical overflow at these checkpoints; retry requires a balanced surface replacement (RFC).
Failure Boundaries
The turn is the containment boundary. Final adapter-path and terminal in-band failures close the step before agent/request-error; retry opens a numbered step; otherwise, the provider error survives. Attempts reset on success.
Other failures use agent/error. Cancellation beats recovery; undispatched calls get synthetic ABORTED results. Effective cancel() emits agent/cancel-requested before queue clearing or abort; observers cannot veto it, and idle calls emit nothing. Disposal awaits quiescence.
Every session event is turn-enclosed. Reloading preserves an interrupted tail and closes it with a synthetic interrupted turn end. Failures after durable turn close report only through agent/error because no safe in-turn position remains. Each turn has one TurnEndReason; TurnEndReasonMap owns the variants.
Agent Handles
ctx.agents owns live agents and returns AgentHandle { agent, dispose() }. Plugins drive Agent through send(), steer(), inject(), cancel(), and whenIdle(). The caller fiber and factory provider structurally co-own programmatic lifecycles; the consumer handle is the only other teardown capability. All owners await one disposer.
Agent Scope
Every live agent owns a scoped agent.ctx. Its registrations shadow globals, receive only that agent's dispatches, and unwind with it; async effects such as background-task cleanup are awaited. CreateAgentOptions.setup(agentCtx) composes the scope before publication. Typed resolvers derive carrier checks from merged Events signatures and scopeTarget (semantic-gates RFC). See the agent-scope RFC and subagent composition controls. AgentLoop runs drivers inside ctx.agents.withInitiator(); private orchestration derives agent.session; other identities stay explicit (RFC).
State
Session Log
The session log is the source of truth. deriveMessages() projects session events into the Message[] sent to the model; raw assistant/chunk events stay in the log for replay and UI fidelity. Replay, fork, resume, transcript rendering, telemetry, and persistence all derive from the same event stream.
Model-visible ⟺ logged: the log reconstructs every request — messages at step/start fronted by the header's session prefix, headers by folding request/header — and dev invariants assert this (reconstructability RFC).
Durability is a plugin concern. Persistence backends buffer synchronous session/event notifications and the loop awaits a turn-end checkpoint before moving on. The SessionPersistence seam stores SessionEvent directly, with metadata in SessionHeader; JSONL and SQLite share one contract suite.
Model Content
Messages contain typed blocks (text, reasoning, tool-call, tool-result) derived from merge-extensible ContentBlockMap; the same pattern types MessageSource, FinishReason, TurnTrigger, and TurnEndReason. New block types coordinate adapters, UI bridges, compaction pricing, token metering, and persistence as one repo-wide contract; replay measurement types live in token-meter.md.
Streaming uses raw chunks (block-start through finish) and BlockAssembler. The loop logs and assembles chunks, storing provider/model provenance plus replay state. An LlmAdapter implements stream(), registers provider routes, and may expose selector metadata; it resolves and validates model ids. Replay state reaches targets only when both routes map to one adapter instance, which owns validation and conversion. The contract lives in llm-streaming.md.
Extension And Composition
Capability Pattern
A swappable capability usually splits into interface / implementation / consumer: service/events, a backend, and model-facing tools/prompts. Bash is the reference; the capability graph maps each family.
Exceptions combine layers: LLM interface/consumer; filesystem policy; web registries; named skill/subagent providers. Subagents spawn fresh, fork a completed-turn prefix, or use ACP children (subagent.md).
dsh-workspace-context composes agent/session-prefix baselines and ctx.fs changes; its RFC records isolation. dsh-paths owns shared paths.
Bundles And Apps
dsh-agent-spine-demo bundles the default spine (README). Apps add front doors; TUI/ACP mount commands (ui/). dsh-jsonrpc-agent boots external cordis.yml, including the Python SDK default (Python SDK). Deployments stay thin with swappable backends/tools (examples/, runnable wirings, graph atlas).
Where New Behavior Goes
New behavior attaches to a documented extension point; a loop change updates this map.
| Goal | Mechanism |
|---|---|
| Add a model provider | register an adapter on ctx.llm |
| Add a model-facing capability | register on ctx.tools; schemas enter prompt assembly |
| Add shell execution | implement and register a ctx.bash backend |
| Add a human command | register on ctx.commands; adapters discover and dispatch it without a model turn |
| Add background work | register on ctx.tasks; generic task_* tools collect or stop it |
| Add filesystem access or policy | implement a ctx.fs provider or listen on fs/* policy events |
| Confine spawned processes | a ctx.sandbox backend; consumers wrap their argv before spawning |
| Intercept a request, tool, or turn | use its agent/* or tools/* event; agent/turn-stop is the serial terminal stop |
| Add a session-stable prefix outside history | compose agent/session-prefix; the request header logs it |
| Add UI or editor integration | drive ctx.agents and render from session/event |
| Add durable session state | add a SessionEventMap member and render/replay from the log |
| Manage a same-session objective | call ctx.goals; drive continuation through Agent and agent/* seams |
| Fork a live session | use ctx.sessions.fork(source, boundary?, childSessionId?) |
| Scope a registration to one agent | use that agent's agent.ctx (see Agent Scope) |
The extension cookbook carries plugin skeletons and the feature-to-seam map; step-by-step guides cover packages, tools, LLM adapters, and vendored packages.
Quick Reference
- Domain terms in the glossary
- Type definitions in core-data-structures/
- Exact event and service signatures in events
- services catalogs
- package contracts in the package map
- RFCs