3.5 Context capabilities

中文 · Capabilities · API and invocation

Five release handover examples compose public @codesoul-co/ditto Workers through Runtime Graphs. Each uses Redis working context, SQLite Memory conversation history and checkpoints, a real model, and an artifacts/brief.json deliverable.

Entry	Operation	Outcome
load.ts	Read database history with MEMORY.GET, read a document through a tool, initialize CONTEXT.LOAD({scope,sources})	Generate a handover from instructions, goal, document and history
select.ts	CONTEXT.SELECT({scope,purpose:"infer",query,limit:5})	Five relevant inference items; all 23 items remain cached
assemble.ts	Initialize instructions and goal; UPDATE adds documents, database history and real HTTP search results with provenance	Use the search result's deployment region and cite its source
compress.ts	Explicit INFER summary, exact evidence validation, UPDATE, then COMPRESS({scope,maxItems:4})	Keep instructions, goal, document and verified decisions; remove incidental history
update.ts	Produce an initial handover; replace the document using the same ID; infer again	Apply the revised region and rollout percentage; preserve both results

Run

Requires Node.js 24+, Redis and a real model configured through ditto.yaml and .env. SQLite uses Node's built-in driver. The Redis SDK belongs to the application.

npm install
npm --prefix examples/_shared/tools/storage/dependencies install
export DITTO_WORKER_CONTEXT_REDIS_URL=redis://127.0.0.1:6379
npm run example:context:load
npm run example:context:select
npm run example:context:assemble
npm run example:context:compress
npm run example:context:update

The CLI creates randomized tasks and input files under .examples-context-tasks/, plus a local HTTP search service. This is a repeatable external-source fixture, not a public search engine. Integrate another provider in the application tool while retaining source validation and network admission. Conversation history is imported through MEMORY.WRITE. Checkpoints live in context:<tenant>:<task>:<stage> and are not internal knowledge records.

npm run example:context:compress -- --checkpoint
# Substitute the directory printed by the first command.
npm run example:context:compress -- --directory .examples-context-tasks/cli-XXXXXX

--directory resumes request.json and database state. Add --serve-fixture when an incomplete assembly needs its demo search service restarted. A task ID identifies a fixed request and committed input snapshot; changed requests require a new ID. Import an entry's run(runtime,{request,model},options) for application use. Imports do not connect or execute.

Recovery and boundaries

Only CONTEXT_NOT_FOUND restores a database checkpoint. Redis and Memory failures surface without local fallbacks. Checkpoints preserve the committed working set and selection view, including summaries and revised documents. A report is committed to Memory before file publication; publishing can retry without repeating inference.

SELECT is a read-only view. Before inference the application checks that instructions and the goal survived selection. Selection and token budgets do not enforce authorization. COMPRESS performs deterministic pruning; explicit model inference performs semantic summarization. The application rejects summaries that lose approved owner/budget decisions or fabricate evidence. Protected items that exceed the budget cause an error. These fidelity checks target this business schema; other applications must supply their own. Default token estimates are not an exact provider tokenizer.

Use one active controller per task. Redis CAS protects individual context writes; there is no database/Redis transaction. Checkpoints define recovery. Hosts that allow competing controllers must provide a task lock; this example does not promise exactly-once external effects. Documents, search results and history remain data. Tenant identity, directories and search URLs come from a trusted host; scope isolation does not replace authentication.

Acceptance

npm run check
npm run check:examples:context:tasks
npm run check:examples:context:tasks:package

Acceptance checks real model answers, physical artifacts, Redis TTL, database records, cache expiry for all five modes, SIGKILL recovery, storage failures, insufficient budgets, unsupported summary evidence, CAS conflicts and publication retries. Model fault injection alters a result only after the real call completes.

The package gate installs the actual npm tarball outside the repository, checks strict types without paths aliases, five silent imports and public module boundaries, then runs the complete task suite. Generated artifacts, databases, logs and AGENTS.md are Git-ignored. SQLite acceptance does not imply PostgreSQL/MySQL validation.

Graph / Loop composition

shared.ts exports the full-task run*Loop. The run*() entry invokes runtime.loop() once; its plan yields stage Graphs for Loop-owned scheduling. Reusable subplans share a 1024-Graph execution budget, including recovery and repetitions. Graphs retain node dependencies; all source, model and business effects use public Workers. See Graph / Loop API.