1컴퓨터에서 일이 실행된다는 뜻terminal, shell, process, event를 먼저 잡는다.Dalkkak의 가장 밑에는 화면이 아니라 로컬 process가 있다. 화면은 그 process와 대화하는 UI다.At the bottom, Dalkkak runs local processes. The UI is just the control surface around them.TerminalShellProcessEventCode focus: spawn a process, stream output, show close/error state
2왜 web-only가 안 되는지browser sandbox와 native desktop의 차이를 이해한다.브라우저는 내 로컬 Claude/Codex process를 직접 띄울 수 없다. 그래서 Tauri와 Rust backend가 필요하다.A browser-only app cannot safely spawn local Claude or Codex processes, so Dalkkak needs a native shell.TauriWebViewRust backendPTYCode focus: React invoke -> Rust command -> OS process
3터미널 pane 하나가 어떻게 뜨는지renderer, command, PTY, event stream, xterm을 연결한다.UI가 Rust에게 PTY를 만들라고 요청하고, Rust는 child process 출력을 event로 보내고, xterm이 그린다.The renderer asks Rust to spawn a PTY, Rust streams bytes back, and xterm renders the terminal.React rendererTauri commandPTYxterm.jsWindow emitCode focus: invoke('pty_spawn'), listen('pty-output'), term.write(data)
4왜 환경변수 때문에 앱이 깨지는지PATH, shell, GUI app env, Homebrew path를 이해한다.iTerm에서는 되는 명령도 Tauri 앱에서 PATH가 다르면 command not found가 난다.A GUI app may have a different PATH than iTerm, so spawned CLIs need explicit environment hygiene.Env hygienePATHShellEOF markerCode focus: construct child env, prepend Homebrew paths, print visible failure
5Dalkkak이 왜 AI를 직접 부르지 않는지BYO AI와 control plane의 차이를 말한다.Dalkkak은 모델이 아니라 작업공간과 통제층이다. 사용자의 Claude/Codex가 worker다.Dalkkak is not the model. It is the operating layer around the user's own AI workers.BYO AIAgentReAct loopTranscriptCode focus: spawn claude/codex as a local CLI process, observe transcript
6일이 chat으로 사라지지 않게 만드는 법WorkItem, receipt, evidence, follow-up을 잡는다.AI가 한 일은 WorkItem timeline, receipt, artifact, follow-up으로 남아야 다시 볼 수 있다.Work must become durable records: WorkItems, receipts, artifacts, and follow-ups.WorkItemReceiptEvidenceFollow-upWorkItem spineCode focus: WorkItem -> action -> artifact -> receipt -> follow-up
7기억을 저장한다는 뜻GraphStore, JSONL, append-only, node/edge/provenance를 이해한다.Dalkkak의 기억은 그냥 요약 문서가 아니라 검증된 노드와 엣지가 쌓이는 append-only ledger다.Dalkkak memory is an append-only graph ledger with provenance, not just chat history.GraphStoreJSONLAppend-onlyGraphNodeGraphEdgeProvenanceCode focus: validate node, append JSON line, query by startup/type
8AI에게 내 기억을 주는 법RAG, memory bundle, MCP를 연결한다.AI가 기억하는 게 아니라, Dalkkak이 관련 기억을 찾아 context로 넣어줘야 한다.The model does not magically remember; Dalkkak retrieves relevant memory and passes it as context.RAGMemory bundleMCPLocal MCP serverTool gatewayCode focus: MCP handler -> graph query -> cited context bundle
9실제 외부 action을 왜 바로 하면 안 되는지dry-run, approval, guardrail, external mutation boundary를 이해한다.이메일/환불/삭제 같은 외부 부작용은 dry-run, approval, receipt 없이 실행하면 안 된다.External mutations must go through dry-run, approval, receipts, and deterministic guardrails.Action SDKDry-runexternalMutationBlockedApprovalGuardrailCode focus: if approval pending, return blocked receipt instead of mutating
10여러 AI에게 일을 나눠주는 법ParallelRun, AgentTask, worktree, file ownership, merge queue를 이해한다.병렬 AI 작업은 터미널 여러 개가 아니라 소유권, 증거, 검증, merge 순서가 있는 운영 시스템이다.Parallel AI work needs ownership, evidence, verification, and a merge queue.ParallelRunAgentTaskGit worktreeFile ownershipMerge queueCode focus: task packet with allowed files, branch, tests, result artifact
11완료 판정을 AI에게 맡기면 왜 망하는지intent, acceptance, verification, final gate를 이해한다.worker는 done candidate를 만들 뿐이고, 완료는 증거와 gate가 판단한다.The worker can propose done, but the system decides completion through evidence and gates.Intent CompilerAcceptance criteriaVerificationRunFinal GateScope integrityCode focus: done_candidate -> verification -> review -> human acceptance
12제품이 되려면 무엇이 아직 부족한지eval, injection, reliability, observability, release trust를 냉정히 구분한다.Dalkkak은 substrate가 강하지만 trust/safety는 아직 더 만들어야 한다. 이걸 알아야 제대로 방향을 잡는다.The substrate is real, but the trust layer still needs evals, injection defense, reliability, and stronger gates.Eval harnessPrompt injectionReliability layerObservabilityRelease readiness gateCode focus: fixture tests, bounded retries, trace IDs, release evidence
1Founder workspaceThe human sees startups, zones, panes, WorkItems, approvals, and receipts.사용자가 실제로 보는 조종석이다. 여러 사업과 작업을 한 앱에서 보고, AI 세션을 지휘하고, 결과 증거를 확인한다.PortfolioHomeWorkspaceZonePage
2React rendererThe UI layer running inside the Tauri WebView.브라우저처럼 생긴 화면 레이어다. 버튼, 카드, 터미널 화면, 그래프 화면을 그린다.ReactViteTailwind
3Tauri shellThe native desktop wrapper connecting UI to local OS capabilities.웹 UI가 macOS의 파일, 프로세스, 알림, 업데이트 같은 기능을 안전하게 쓰게 해주는 껍데기다.Tauri commandsWindow eventsplugins
4Rust local backendOwns PTY, tmux, graph capture, memory, vault, mission scanning, and local commands.무거운 로컬 일을 하는 엔진이다. 터미널을 띄우고, 로그를 모으고, 기억을 저장하고, 위험한 경계를 지킨다.pty.rstmuxcapture.rs
5Terminal substrateThe pane/process layer where Claude Code, Codex, shells, and commands actually run.AI가 실제 명령을 치고 파일을 바꾸는 작업장이다. Dalkkak은 이 작업장을 한 화면에 여러 개 만든다.portable-ptytmuxpane registry
6Connective layerTurns scattered work into a durable graph of facts, actions, evidence, and edges.일이 끝나고 사라지지 않게 만드는 기억층이다. 커밋, 에이전트 요약, 액션 결과를 노드와 엣지로 남긴다.GraphStoreJSONLGraphNode
7Action spineRepresents work as WorkItem -> action -> artifact -> edge -> receipt -> follow-up.버튼을 눌렀을 때 무엇을 했고, 어떤 산출물이 생겼고, 다음 일이 무엇인지 남기는 업무 척추다.WorkItemAction SDKreceipt
8Local MCP gatewayExposes Dalkkak memory/workflow tools to BYO AI clients without giving broad mutation power.Claude/Codex가 Dalkkak의 기억을 조회할 수 있게 하는 표준 도구 서버다. 우선 읽기 중심이어야 안전하다.apps/local-mcp-serverhandlersschemas
9Cloud/web layerHandles workspace sync, RBAC, web views, tickets, presence, and team-facing surfaces.로컬 앱만으로 안 되는 동기화와 팀/웹 기능을 맡는다. Supabase RLS가 접근권한의 핵심이다.apps/webapps/serverSupabase
10Trust gatesPrevents AI from silently overclaiming, leaking secrets, or performing risky actions.AI를 믿는 게 아니라 증거, 승인, 검증, 방어 규칙으로 통제하는 층이다. 제품의 신뢰 핵심이다.approvalguardrailseval
PartialTerminal/workspace substrate existsTauri desktop, React renderer, Rust PTY/session pieces, panes, zones, and many UI surfaces exist.The UI is broad and technical; product coherence depends on WorkItem/memory/mission surfaces becoming simpler.Next: Explain every visible surface through one model: workspace -> WorkItem -> action -> receipt -> memory.
HasBYO Claude/Codex direction is clearThe product does not primarily call AI itself; it wraps and augments user-owned AI sessions.People may mistake it for a managed chatbot/codegen product.Next: Keep saying: Dalkkak is the control plane, not the model.
PartialOperating memory is real but incompleteGraphStore/JSONL, WorkItem memory spine, graph nodes/edges, and local MCP read surfaces exist.Semantic retrieval, embeddings, deletion/governance, and full citation discipline are incomplete.Next: Harden retrieval, citation, provenance UI, and memory lifecycle.
PartialAction spine exists mostly as safe dry-runAction contracts, receipts, approval metadata, and mock/dry-run boundaries exist.Real Gmail/Stripe/Instagram style mutation connectors are not fully wired.Next: Ship one real connector vertical with injection defense, approval, receipt, retry, and rollback.
PartialMission Control model is strongDocs, scanner/model/UI sections, merge queue, verification evidence, and many tests exist.Full dispatch/monitor/reissue/integration workflow is not fully productized.Next: Turn manual parallel-agent chaos into a guided run from split planner to final gate.
PartialCloud/web layer has a foundationSupabase schema, forced RLS, web workspace, tickets, widget sync, and presence pieces exist.Desktop/cloud/web product boundaries and API contracts are still evolving.Next: Clarify what lives local, what syncs, and what web can safely do.
MissingAI eval is the biggest trust gapTraditional tests exist, but fixed AI quality evals are not yet strong.You cannot know whether AI-generated drafts/actions are good at scale.Next: Build eval fixtures for support, billing, release, and coding assistant outputs.
MissingPrompt injection defense is mandatory before real connectorsThe docs identify it as critical, but full defense is not built.Untrusted email/web/customer text can hijack agent behavior.Next: Add untrusted-content boundaries, policy checks, and adversarial fixtures before mutation connectors.
MissingReliability layer is not completeSome idempotent memory design exists, but action retry/timeout/fallback/replay is not complete.Real external actions could double-run, hang, or fail invisibly.Next: Define action state machine with timeout, retry cap, idempotency key, and recovery receipt.
PartialRelease trust work existsRelease docs, gates, signing/updater concepts, and claim cleanup are present.Release claims can drift from actual verified behavior.Next: Keep release readiness gate tied to command evidence and smoke checks.
Canonical product blueprintddalkkak/docs/BLUEPRINT.mdMission, layers, reality checkpoint, product boundaries
AI component glossaryddalkkak/docs/product/ai-engineering-components-glossary.mdAI concepts, current status, trust gaps
WorkItem memory spineddalkkak/docs/product/workitem-memory-spine-integration-and-blueprint-status-2026-06-17.mdWorkItem -> action -> artifact -> receipt -> follow-up
Mission Control specddalkkak/docs/product/parallel-agent-mission-control-product-spec-2026-06-17.mdParallelRun, AgentTask, result inbox, merge queue
Shared agent rulesddalkkak/docs/product/shared-agent-operating-rules-2026-06-17.mdNo fake done, evidence-first parallel-agent protocol
Rust local backendddalkkak/apps/desktop/src-tauri/srcPTY, graph capture, memory, vault, mission scanner