[mattzcarey]

agents/context gets Context (experimental): context blocks bound to a session. The transcript records the system prompt each turn was sent with and every context change delivered after it, and each model request is a projection of that record. Think's context now runs on it, ThinkHarness takes a context option, and Sessions is unchanged.

readonly context = new Context(this.session, {
  storage: this.ctx.storage,
  blocks: [
    { label: "soul", provider: { get: async () => "You are helpful." } },
    { label: "memory", description: "Learned facts", maxTokens: 1100 },
    { label: "environment", provider: environment, whenChanged: "remind" }
  ]
});

// each turn, once the user message is stored
const { system, messages } = await this.context.build();
const tools = await this.context.tools();

Context never rewrites what the model has already been sent:

• the first turn records a snapshot of the blocks on its user message; every later request opens with that exact prompt until a refresh records a new one
• a remind block that changed is delivered after the user message of the turn that noticed it, and recorded there, so later requests render it in the same place instead of moving it to the newest user message every turn
• a user message that already has an answer (a regeneration) is never recorded on; the change rides on that request only
• block values stay in their providers (the existing cf_agents_context_blocks rows), so nothing migrates; a small state row keeps a pending refresh and a pointer to the latest snapshot, so a refresh survives eviction and a bounded history window still opens with the recorded prompt

ContextBlocks keeps only blocks and tools: freezeSystemPrompt, refreshSystemPrompt, reminder, and the prompt store move to Context. ThinkHarness reserves the context metadata key, so client input cannot carry a record.

This replaces the _system_prompt row and the moving reminder (#2772). pi and Claude Code work the same way: context changes land at turn boundaries and stay where they were delivered. Think's prompt-cache test shows the result: a turn after a reminder now extends the previous request instead of losing the last user message from the cache.

Agents workers, Think workers, ThinkHarness, and ai-chat tests ran, plus pnpm run check. Context is covered by integration tests on a real Durable Object and session. Docs: cloudflare/cloudflare-docs#34718.

[changeset-bot]

🦋 Changeset detected

Latest commit: bde8a57

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 3 packages

Name	Type
agents	Minor
@cloudflare/think	Patch
@cloudflare/agent-think	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pkg-pr-new]

Open in StackBlitz

agents

npm i https://pkg.pr.new/agents@2773

@cloudflare/ai-chat

npm i https://pkg.pr.new/@cloudflare/ai-chat@2773

@cloudflare/codemode

npm i https://pkg.pr.new/@cloudflare/codemode@2773

hono-agents

npm i https://pkg.pr.new/hono-agents@2773

@cloudflare/shell

npm i https://pkg.pr.new/@cloudflare/shell@2773

@cloudflare/think

npm i https://pkg.pr.new/@cloudflare/think@2773

@cloudflare/voice

npm i https://pkg.pr.new/@cloudflare/voice@2773

@cloudflare/worker-bundler

npm i https://pkg.pr.new/@cloudflare/worker-bundler@2773

commit: bde8a57

[agent-think]

🟢 agents import sizes: 1 entry point changed, no growth

	Entry point	Exports	Largest gzip change	Size now
🆕	agents/context	4 new	—	95.4 KiB

Changed exports (4)

	Import	Gzip change	Size now
🆕	agents/context#Context	—	95.4 KiB
🆕	agents/context#CONTEXT_METADATA_KEY	—	95 B
🆕	agents/context#SET_CONTEXT_TOOL	—	89 B
🆕	agents/context#SEARCH_CONTEXT_TOOL	—	88 B

How this works

Each runtime export is bundled on its own, minified, and gzipped. Changes smaller than 100 B, or smaller than 1% and 1 KiB, are ignored. Growth over 10% or 5 KiB is marked 🔴. This report is informational and does not fail CI. The workflow artifact contains every measurement.

_{Compared 54ae3e86 → bde8a574 · workflow run · reported by agent-think[bot]}