platform/docs/opencode-serve-path.md

The serve-based turn path: what it was, and whether to rebuild it

Written 2026-08-10, before deleting event-mapper.ts and the unused half of api/chat/opencode/client.ts. docs/opencode-parity.md item 7 asks for exactly this: read the dead design into a note first, because Phase 2 may rebuild it.

It also answers Phase 2's blocking question, which turned out to be cheap to settle.

---

The question Phase 2 was waiting on

docs/opencode-phase0-review.md narrowed the fork to one thing:

Moving turns onto the serve would reintroduce exactly the single-directory coupling that caused the original pain, unless the serve's API can take a per-request directory. That question — not "why was it replaced" — is now the one to answer first.

It can. Tested against the running serve (opencode 1.17.9 on this machine):

POST /session?directory=/tmp/oc-phase2-probe   →   { directory: "/tmp/oc-phase2-probe" }   ✅ matches
POST /session   body { directory: … }          →   { directory: "<serve cwd>" }            ❌ ignored

directory is a query parameter on every /session* route — create, message, abort, fork, summarize, prompt_async, shell, all of them — alongside a workspace param. It is not a body field, which is why a first probe that sent it in the body appeared to disprove the whole idea.

So the coupling is gone on both paths. opencode run --dir anchors a subprocess turn (verified in the phase-0 review), and ?directory= anchors a serve turn. The single-server constraint that shelved this work does not exist in either architecture any more.

That removes the reason not to move. It does not by itself decide the move — see the trade below.

---

What the dead code actually was

Two files, ~200 of ~390 platform-side lines, all reachable from nothing:

event-mapper.ts (146 lines) — maps OpenCode's SSE event stream to officer's ChatEvents. Handles message.part.updated (text deltas, tool state transitions), message.updated, session.idle and session.error. Its shape assumes a streaming source: partial text arriving as deltas, tool calls transitioning pending → running → completed as separate events.

The SSE half of client.ts — subscribe(sessionId, listener), one shared GET /event stream per base URL demultiplexed to per-session listeners, with reconnect. Plus createSession, postMessage, abort, isServerHealthy.

The live half of client.ts stays: listSessions, getSession, getMessages, deleteSession, renameSession, all used by opencode-sessions.ts for the chat list and transcript reads.

Why this matters for a rebuild

The dead mapper is not a sketch to be dusted off — it is a finished, working shape for a design that was measured against a real event stream. Two things in it are worth keeping if the serve path returns:

1. The delta model. runner.ts emits whole text blocks because opencode run --format json emits whole blocks; the mapper emits deltas because SSE emits deltas. Token streaming (parity Phase 3, item 11) is not new work on the serve path — it is this file.
2. Tool-state transitions. The mapper tracks a tool call across pending/running/completed. The subprocess path only ever sees the finished call.

Both are recoverable from git after deletion (5d077a4 is the last commit where the serve path was live), which is the argument for deleting rather than keeping it compiled-but-unreachable: an unused file rots silently against a moving API, and this one is already pinned to a version two minor releases behind what some machines run.

---

The trade, now that the blocker is gone

Keep the subprocess (opencode run --dir)

• No rewrite; it works today.
• Permanently forfeits: token streaming, mid-turn injection, background tasks, live-session enumeration, reattach-by-id, interrupt-without-teardown. Every one of those is a stdin: 'ignore' consequence.
• One process per turn, no warm state to leak or garbage-collect.

Move turns onto the serve (POST /session/{id}/message?directory=…)

• Unblocks the whole of parity Phase 3 at once — those six capabilities are all downstream of a persistent, addressable session.
• Re-adopts an SSE stream officer must keep alive, demultiplex and reconnect. That machinery already exists in the deleted code, so the cost is smaller than it looks.
• Introduces warm sessions and therefore a lifetime question OpenCode currently does not have: idle GC, orphan adoption after a restart, the same problems the Claude path spent months getting right.
• The serve becomes load-bearing rather than a convenience. Today a serve crash costs session listing; then it would cost every turn in flight.

A third option, not in the parity doc: move only what needs the serve. Keep opencode run for turns and add ?directory=-scoped serve calls for enumeration and reattach. That buys the Live panel and reattach-by-id without warm sessions or an SSE loop. It does not buy streaming or mid-turn injection, which are the two most visible gaps.

Recommendation

Do not start the migration on this pass. The blocker is cleared and that is the deliverable here; choosing between the three is a product call about how much OpenCode should behave like Claude, and it should be made deliberately rather than as a side effect of a cleanup branch.

If it is taken: the third option first. It is incremental, it is the only one with no lifetime questions, and it makes the Live panel — which today silently omits every running OpenCode turn — tell the truth.

---

Verified facts this note rests on

Claim	How
?directory= on POST /session is honoured	Created a session, read directory back: matched
A body directory is ignored	Same call with the field in the body: got the serve's cwd
directory is on every /session* route	Read the serve's own /doc (OpenAPI)
event-mapper.ts has no importers outside client.ts	grep
createSession/postMessage/abort/isServerHealthy have no callers	grep for call sites
listSessions/getSession/getMessages/deleteSession/renameSession are live	all from opencode-sessions.ts

Machine note: this server runs opencode 1.17.9; the phase-0 review's tests ran against 1.18.11 elsewhere. The ?directory= result above is from 1.17.9, so it holds on the older of the two.