Special Topic

Inside DeepSeek Harness

Follow a plugin-based agent foundation from session logs to tool calls, compaction, inboxes, and sandboxed execution. This is an optional source-reading track for people who want to see the abstractions behind the interface.

29 notesOpen reading · no login

Special Topic29notes

THE QUESTION THIS PAGE ANSWERS

ANSWER FIRST

What will the “Inside DeepSeek Harness” AI learning path help you do?

DECISION RULE

Core themes include System Map, Sessions & Loops, Context Engineering, Tool System.

TRY NEXT

Begin with “Everything Is a Plugin: Announcement vs Source,” then choose the next note by the task in front of you.

WATCH FOR

Do not optimize for finishing the list. Explaining one trade-off with your own example matters more than opening more titles.

What this route helps you practice

Open the first note

Each chapter follows a class of real decisions. Follow the sequence, or enter at the problem you are solving today.

29notes

01Everything Is a Plugin: Announcement vs SourceMap the launch copy’s four modes, session log, and plugin ecosystem line by line onto the repo’s YAML presets and package layoutArchitecture10 min↗02Profile / Bundle / Patch: How Far Users Can Reshape the ProductHow the three-layer config lets users swap deep capabilities without editing sourceArchitecture11 min↗03Model-visible ⟺ logged: An Invariant That Crashes in Your FaceStarting from the byte-for-byte check in invariant.ts: why everything the model sees must be reconstructable from the logDeep Dive10 min↗04followup / steer / inject: Dual-Queue InboxBreak down enqueue, claim, and wake semantics for the next-turn and next-step queues, and contrast Claude Code’s interrupt modelInteractive10 min↗05What Happens After Esc: Cancel, Crash Recovery, and Re-entryOne AbortSignal per turn; write the log even on interrupt; crash, restart, and keep goingDeep Dive11 min↗06Goal: Provenance Is PermissionHow long-term goals are stored, who may change them—auth looks at message provenanceDesign Pattern11 min↗07Compaction Dual Paths & replaceGenerationTwo orthogonal paths — pressure trigger and overflow recovery — plus requiring a generation number to prove compaction actually happened before retry is allowedDeep Dive9 min↗08Token Accounting: Replay for Decisions, Projection for DisplayTwo meters, each its own job—compaction decisions never trust the UI numberConcept8 min↗09Spill: When Tool Output Is Too LargeOver-limit output is archived to disk; the model keeps a retrieval credentialHands-on10 min↗10Session Search & Cross-Session CitationsOld sessions are a searchable corpus; citations can carry provenanceDeep Dive9 min↗11Tool Execution Pipeline: Three-Stage Cascade & Monotone GuardA three-stage line from pre-execute to post-execute — Guard can only tighten, never allowArchitecture9 min↗12Tool Output Contract: Value Separated from DisplaySame result — what the model sees and what people see can differDesign Pattern8 min↗13File-Edit Engineering: Read Before WriteThe read / edit / write trio — unread files can’t be editedHands-on9 min↗14DSH-Only Tool Surface: terminal / lsp / jobsWhat the tools others lack each solveCase Study9 min↗15Approvals & Permissions: Two Knobs, One DropdownSandbox mode and approval policy are two independent knobs; presets are just common combosSecurity9 min↗16Sandbox: From seatbelt to the Execution Worldctx.fs and ctx.subprocess share one path namespace, so the whole execution environment can be swappedSecurity10 min↗17Code Mode: One Program Beats Many Tool RoundsAnnouncement calls it PTC, source calls it code mode: worker sandbox isolation, message protocol, and double bookkeepingFrontier11 min↗18Subagent Is a Seam: In-Process to Delegating Claude CodeA sub-Agent is a capability seam — in-process, remote, or another product can plug inArchitecture8 min↗19workflow / schedule / plan / todo: Orchestration PrimitivesWhat each of the four orchestration primitives owns, and why they were never merged into one monolithConcept9 min↗20Skill, Preset, and Self-Modificationcordis_define lets an Agent rewrite its own runtime while runningFrontier10 min↗21MCP & Extensions: Two Paths to External ToolsHow bridging the ecosystem standard and native extensions divide the workDeep Dive10 min↗22LLM Adapter Layer: Single Attempt, Explicit Retry, Dual-Stream PersistenceReasoning and text streams stored separately; retries are explicit eventsDeep Dive9 min↗23Testing a Nondeterministic SystemDeterministic replay, property-based testing, and a fault server built to trick LLM clientsHands-on10 min↗24Persistence Governance: Versions, fork Boundaries, Refuse-to-ParseHow log formats evolve, how fork boundaries are set, and prefer refusing data you can’t readArchitecture10 min↗25Credentials, Settings, Storage, and TelemetryUnsexy and full of traps: credentials resolved every time, secrets never land in configHands-on10 min↗26Multi-Entry & Typert: One Kernel, Five FacesWeb, headless, ACP, SDK, and HTTP share one kernelArchitecture9 min↗27Agent Notes & AGENTS.md: Discipline for Building AI with AIFour-state design notes and coding standards written for AI — your team can copy them immediatelyRecap11 min↗28KV Cache Is an InterfaceTreat prompt prefix stability as a compatibility promiseFrontier10 min↗29Finale: Five Engineering Views — What Should We Copy?Five harnesses’ design-philosophy master table, plus a minimum copy list and a scale-trap listFinale10 min↗