[kitlangton]

The agent can now tell when a message came from another session or a tool rather than the person, and which one.
A message can say who sent it in its metadata, and the agent reads that as one line before the text. What was typed, what is stored and what the API returns stay exactly as sent.

💡 Why
Sessions message each other: a subagent gets its task from its parent, one session hands work to another, and scripts report back. Today every such message looks to the agent like the person typed it, so senders type "From X:" into the text by hand, and a continued subagent gets no hint at all.

🆕 Now you can

• Send a prompt with metadata["opencode.origin"] set to { kind: "session", sessionID } or { kind: "tool", name }.
• The agent reads <message-origin session="ses_…" title="Fix login"/> (or <message-origin tool="release-bot"/>) before the text.
• The subagent tool and subagent commands now send their parent session as the origin, for new and continued subagents. The existing "You are a subagent…" line stays.

🚦 How it ships
Off by default, behind experimental.message_context. With it off, requests to the model are byte-for-byte what they are today. The origin is still stored, so turning it on later applies to earlier messages too.

  user message -> admission -> history (model reads) -> prompt builder -> model
+ admission: set origin.title from the sender session (trimmed, one line, ≤100 chars), never from the caller
+ history, model reads only: flag off, drop the origin in memory; flag on, keep it as stored
+ prompt builder: "<message-origin …/>\n" + text   (same line in compaction's recent text)

The title the agent reads is the one saved when the message arrived and never changes, so renaming or deleting the sender can't rewrite earlier messages and bust the receiver's prompt cache. A sender with no title yet is named by its session ID alone. Clients show the sender's current title from the session ID.

📊 Proof

Case	What the agent reads
Flag off	the text alone, same request as a message without the key ✅
Sender renamed or deleted after sending	the same line as before ✅
Sender untitled when the message arrived	session="…" with no title, even after it gets one ✅
Caller sends its own title	replaced by the sender's real title ✅
Title or tool name with newlines, padding, 150 chars	one trimmed line, 100 chars ✅
Tool origin	tool="…" ✅
Value that doesn't decode	the text alone ✅
Compaction's recent text	one [User]: line with the origin before the text ✅
Continued subagent	origin on both the first and the follow-up prompt ✅
sessions.context (HTTP, SDK, plugins)	metadata exactly as stored, whatever the flag ✅

14 new tests, plus origin checks added to the subagent tool, subagent command and prompt admission tests.

👀 Read: session/history.ts, loadForModel. The flag is read only where model requests are built (the runner, preview, compaction), inside the session's own location. load and sessions.context stay plain reads of stored metadata. The rest is routine.

_{The generated SDK gains only the optional experimental.message_context boolean; routes, prompt input and message types are unchanged.}

Checks run

• bun run test in packages/core over the session, prompt, subagent, config, compaction, runner, title, instructions, execution and generate tests: 741 pass, 0 fail
• An earlier wider run also covered test/plugin/ and test/schedule.test.ts; its only failures were 6 DigitalOcean OAuth tests that fail the same way on dev without this change
• bun run typecheck in packages/core, packages/schema, packages/client: no errors
• bun run check:generated in packages/client and packages/protocol: clean
• oxlint on the changed files: 0 warnings, 0 errors

---

Added by the opencode bot (AI-generated).

Related: #58367, #52297

Public API: adds the optional experimental.message_context config flag and the opencode.origin prompt metadata key (backwards compatible)

[opencodejr]

Looks sound. I ran the four changed test files in packages/core (75 pass) and tsc --noEmit on core (clean). I also checked that every path that builds a model request (runner, preview, compaction trigger and precompute) reads history with the flag applied, while title generation and sessions.context keep reading metadata as stored.

Public API: adds the optional experimental.message_context boolean to config (schema, OpenAPI spec and generated client types). It also starts a metadata convention, opencode.origin: on admission, a session origin's title is always replaced with the sender's stored title and any extra fields are dropped. The subagent tool and command subagents now store that key on the child's prompt, so sessions.context returns it. Everything is additive and backwards compatible; only callers that already used the opencode.origin key would see it rewritten.

A non-blocking note on the design: the flag works by removing the origin when history is loaded, but SessionHistory.load keeps it and toLLMMessages always writes the line. A future model-request path that reads through load (or store.context) would show the origin even with the flag off. That's fine for now, since every current path is covered.

Related: #58367, #52297