# The Phase 2 fork: reopened, and why the first answer was wrong

> **CORRECTION, same day.** An earlier version of this file concluded "the new pipeline does not
> execute, keep the subprocess". **That was wrong, and wrong for an embarrassing reason: my probe.**
> The conclusion is reversed below. The mistake is written up rather than deleted, because the shape of
> it is the useful part.

**Decision: the fork is worth taking. Not started, and gated on one open question.**

---

## What actually happened

The serve exposes a newer `/api/session/*` surface, alongside the `/session/*` one every prior document
was written against. It offers, natively, what the parity doc lists as impossible under `stdin: 'ignore'`:
`delivery: "steer" | "queue"` on `POST /prompt`, `/interrupt`, a per-session `text/event-stream` with an
`?after=` cursor, `/compact`, `prompt.files`, and `/permission` + `/question`.

I probed it, saw prompts accepted and never executed, and concluded it was an unfinished pipeline
("`session.next.*` is the tell"). Every one of those probes passed an explicit
`model: {providerID: 'opencode', id: 'claude-sonnet-4-6'}`.

**That model silently does not run on the new pipeline.** No error, no event, no assistant message — the
prompt is admitted, stored, `prompt.admitted` and `prompted` fire, and nothing else ever happens. Drop
the model field and the identical request completes normally.

So I had one broken variable in every experiment and read the result as a property of the system.

## What is actually true, measured on 1.18.16 (both machines upgraded 2026-08-10)

| Claim                                     | Verdict                                                              |
| ----------------------------------------- | -------------------------------------------------------------------- |
| The new pipeline executes turns           | **Yes** — replies normally when no model is forced                   |
| `delivery: "steer"` injects mid-turn      | **Yes, verified** — steered a running turn, output changed to order  |
| `delivery: "queue"` runs after            | **Yes, verified** — two replies, "ONE" then "TWO", zero errors       |
| Model selection works at all              | **Yes** — `POST /api/session/{id}/model` → 204, then runs on it      |
| `claude-sonnet-4-6` works there           | **No** — silent no-op, at create *and* via the model route           |
| `claude-sonnet-4-6` works via `run --dir` | **Yes** — verified end to end the same day                           |

Steer and queue are exactly the two features we hand-built for Claude and that Andre called a game
changer. Having them as primitives, plus a resumable per-session cursor that mirrors officer's durable
replay, is a strong argument for migrating.

## The blocker, now characterised: only free models run

It is not sonnet, and it is not `variant`. Swept four models through `POST /api/session/{id}/model`
followed by a prompt:

| Model                | New pipeline |
| -------------------- | ------------ |
| `longcat-2.0-free`   | **ran**      |
| `ling-3.0-tiny-free` | **ran**      |
| `claude-haiku-4-5`   | never ran    |
| `claude-sonnet-4-6`  | never ran    |
| `gpt-5.1-codex-mini` | never ran    |

**Every `-free` model runs; every paid model silently does not.** Ruled out along the way:

- **Not `variant`.** `claude-sonnet-4-6` advertises `["low","medium","high","max"]` and session create
  echoes back `variant: "default"`, which is not among them — a promising theory that turned out to be
  wrong: setting `variant: "high"` explicitly also never ran. Tested rather than assumed, which is the
  whole lesson of this file.
- **Not missing credentials.** `opencode auth list` shows an OpenCode Zen API key in
  `~/.local/share/opencode/auth.json` plus `ANTHROPIC_API_KEY` in the environment.
- **Not the sidecar's environment.** The *same* sidecar process runs `claude-sonnet-4-6` correctly
  through `opencode run --model`, verified end to end. Same user, same home, same auth file.

So the new pipeline does not resolve paid-model credentials, and fails **silently** rather than
reporting it — while the legacy path and `run` both authenticate fine. Note it *can* surface provider
auth errors when it reaches that far: alpha's default (`nano-gpt`) returned a clean
`401 missing_api_key`. The silence is specific to opencode-zen paid models.

This reads as an upstream bug in an in-progress pipeline, not something configurable on our side.

**Consequence for the fork:** blocked, but precisely. Officer's users pick real models; a harness that
works only on free tiers is not adoptable. Re-run the sweep above after each `opencode upgrade` — the
day a paid model runs there, the migration is unblocked and worth doing immediately, because steer and
queue are already proven.

## Revised recommendation

**Take the fork, targeting the new surface, once the model question is answered.** Not the legacy
`/session/{id}/message` path — that generates fine but has neither steer nor queue, so it buys streaming
at the cost of SSE demux and warm-session lifetime for the two least interesting gaps.

Until then `opencode run --dir` stays, and it is verified working: `session:init` → `assistant:text` →
`result` with cost, on 1.18.16, after the upgrade.

## The lesson, which is the reason this file keeps its history

This project has now hit the same trap three times, each time in a different costume:

- `directory` in the body of `POST /session` — accepted, echoed, ignored. Produced a confident wrong
  answer about per-request directories.
- `location.directory` in the parity doc — a field that did not exist on that surface, which would have
  compiled and silently produced `''`.
- `model` on `POST /api/session` — accepted, echoed back in the response, and fatal to execution.

**OpenCode's API accepts input it does not honour, and says nothing.** So a probe that changes one thing
and sees nothing happen has not learned that the feature is missing; it has learned that *something* is
wrong, and the next step is to remove variables, not to conclude. The corrected probes here each changed
exactly one field.

## Findings worth keeping

- `delivery` defaults to `"steer"` when omitted.
- New-surface responses wrap in `{"data": …}`; the legacy surface returns bare objects, so reading
  `body.id` instead of `body.data.id` silently yields `undefined`.
- The new surface is location-scoped per request: `x-opencode-directory` header, or
  `?location[directory]=` as a `deepObject` query.
- `?after=` genuinely replays a finished session's events.
- **opencode versions: both machines on 1.18.16** as of 2026-08-10 (Mac was 1.18.11, alpha 1.17.9).
  Several comments recorded these backwards; corrected. Policy from Andre: **write against the latest
  version regardless of what alpha happens to run.**