think.
@cloudflare/think github.com/acoyfellow/think-snippets 15 examples · live
15 examples

Small Think agents, each with a live test.

Project Think is Cloudflare's agent runtime. This is a collection of 15 small examples. Each one deploys to a real Worker, runs a test that proves what it claims, and tears itself down afterward.

Bythink.coey.dev·Updated 2026.06
Deploy rpc-chat-memory Run the full suite

Each example makes one claim and ships a test for it. If the test doesn't pass live, it doesn't go in.

About this issue Each contract is isolated: its own Worker, its own Durable Object bindings, its own probe, its own teardown. We do not stage screenshots; we read durable state. Personal Cloudflare deploys only.

think.coey.dev / receipt #26868

15/15 · live
  1. 01 · MemoryTWO REQUESTS → SAME DO → "OCTARINE"
  2. 02 · EvidenceMODEL WROTE BYTES · WE READ BYTES
  3. 03 · RetrievalSEARCH FIRED · READ FIRED · ANSWER CITED
  4. 04 · ProvenanceSQLITE ROW = ASSISTANT ANSWER
  5. 05 · ApprovalDENY → DENY → APPROVE → REPLAY = DENY
  6. 06 · LifecycleBEFORE · STEP · TOOL · AFTER · RESPONSE
  7. 07 · Sub-agentPARENT.STREAMS(CHILD.TURN)
  8. 08 · HandoffTYPED · CHECKSUMMED · TRANSFERRED
  9. 09 · ScheduleALARM → USER MSG → MODEL REPLY
  10. 10 · SandboxGUEST JS WROTE FILE · PARENT READ FILE
  11. 11 · ConcurrencyMIDDLE TURN: SUPERSEDED
  12. 12 · BroadcastTWO SOCKETS · ONE STREAM
  13. 13 · ResumeDISCONNECT → REPLAY → AGREES
  14. 14 · CLI ground truth (HTTP)PROBE READS CF API · AGENT MUST AGREE
  15. 15 · CLI ground truth (sandbox)SANDBOX RAN CF · STDOUT = LIVE ACCOUNT
**** end of transmission ****
01

Memory

Two requests, one Durable Object — fact remembered.

rpc-chat-memory

Read →
02

Evidence

Model writes; we read the bytes back.

workspace-write-read-proof

Read →
03

Retrieval

Search/read fires before the answer.

workspace-search-proof

Read →
04

Provenance

Runtime-only output, captured & served.

server-tool-audit-loop

Read →
05

Approval

Exactly-once, denied otherwise.

tool-approval-headless

Read →
06

Lifecycle

Stable partial order, durably proved.

hook-order-receipt

Read →
07

Sub-agent

Parent streams a child Think turn.

clientless-subagent-rpc

Read →
08

Handoff

Typed checksum envelope between agents.

cross-agent-handoff-envelope

Read →
09

Schedule

Alarm injects a turn, model replies durably.

scheduled-synthetic-turn

Read →
10

Sandbox

Sandbox JS mutates workspace state.

execute-tool-state-edit

Read →
11

Concurrency

Queue keeps all; latest supersedes middle.

concurrency-latest-vs-queue

Read →
12

Broadcast

Two sockets, one stream.

multi-tab-broadcast-protocol

Read →
13

Resume

Disconnect, replay, agree with transcript.

stream-resume-contract

Read →
14

CLI ground truth (HTTP)

Agent runs a real cf CLI over HTTP; the live account is the only source of truth.

cli-http-ground-truth

Read →
15

CLI ground truth (sandbox)

Same real cf CLI, run inside a codemode sandbox; token stays server-side.

cli-sandbox-ground-truth

Read →
Try it

Deploy one, or run the whole set against your own account.

Deploy rpc-chat-memory Run the full suite