[Blind-Striker]

M1 — walking skeleton, complete

This branch closes M1: the SDK goes from a compiled model closure (#15) to a callable client demonstrated live against a real server. It also executes the v2 retarget sealed in ADR-0005 — the generation target moves to upstream's packages/protocol/openapi.json on the v2 branch, pinned at commit a6a712a, and the legacy 1.x surface is never built.

What landed

• v2 retarget — pin swap to v2-branch a6a712a (104 paths / 120 operations / 324 schemas), the single-element validation-only allOf admit rule in the ingestion wall, and the regenerated model closure.
• Callable surface, fully generated (ADR-0008) — SpecIR plus curation bind into client/operation/envelope/route/error plans; generic emitters render OpenCodeClient.GetHealthAsync, Sessions.GetSessionClient(...), SessionClient.GetMessageAsync, guarded envelopes, OpenCodeRoutes (nested per-group containers, template + escaped-builder pairs), and response adapters as committed source. Emitters stay name-blind; names ride the plan.
• Hand-written behavior core — one Pipeline owns endpoint authority, Basic-auth/User-Agent decoration, response buffering, throw-versus-NoThrow, and transport-failure mapping over the ADR-0007 exception spine (OpenCodeException → OpenCodeApiException; transport/protocol failures as OpenCodeTransportException).
• Mechanical naming policies — envelope payload names derive mechanically from operation subject tokens with a fail-closed spine-collision check (envelopePayloadNames is an override ledger for collisions only); the User-Agent version token strips + build metadata and omits itself entirely rather than falling back silently.
• Public API locked — a reviewed PublicApiGenerator baseline shared across TFMs; the generation writer refuses unmanifested overwrites and headerless manifest entries.
• Tests — 526 across the tooling and SDK legs (net8.0/net9.0/net10.0 everywhere; the net472 leg runs on Windows), including a contract matrix over the callable surface.
• Sandbox — tests/OpenCode.Sdk.Sandbox, a committed debug playground for driving the SDK against a real opencode2 serve under a debugger.

Packing still fails deliberately on the .generation-incomplete marker — the generation profile is two operations wide; packaging unblocks when the surface completes (M5).

Live demonstration (2026-08-13)

Server: opencode2 v0.0.0-next-17403 (@opencode-ai/cli@next); spec pin: v2-branch a6a712a.

$ opencode2 serve --hostname 127.0.0.1 --port 41733
server listening on http://127.0.0.1:41733
server password <generated>

$ curl -s -o /dev/null -w "%{http_code}\n" http://127.0.0.1:41733/api/health
401

# IDs obtained through the pinned surface (v2.session.list, v2.message.list):
$ curl -su "opencode:<pw>" "http://127.0.0.1:41733/api/session?limit=1" | jq -r '.data[0].id'
ses_00e0d1436ffe86YL3QKbrn3Ldy
$ curl -su "opencode:<pw>" ".../api/session/ses_00e0d1436ffe86YL3QKbrn3Ldy/message" | jq -r '.data[0].id'
msg_ffb0e86fa001n28S9q0u7F7JxM

# The two generated methods, from a scratch console with a ProjectReference to OpenCode.Sdk:
$ dotnet run -- http://127.0.0.1:41733 <pw> ses_00e0d1436ffe86YL3QKbrn3Ldy msg_ffb0e86fa001n28S9q0u7F7JxM
health: status=200 healthy=True version=0.0.0-next-17403 pid=3910076
message: status=200 type=SessionMessageAssistant id=msg_ffb0e86fa001n28S9q0u7F7JxM

Content detail (second run):

message: status=200 type=SessionMessageAssistant
  id:      msg_ffb0e86fa001n28S9q0u7F7JxM
  agent:   build
  model:   ModelRef { Id = gpt-5.6-sol, ProviderId = openai, Variant = max }
  finish:    cost: 0  tokens: TokenUsageInfo { Input = 0, Output = 0, Reasoning = 0, Cache = TokenUsageInfoCache { Read = 0, Write = 0 } }
  content: 1 part(s)
    [reasoning] **Summarizing findings and planning next steps**

Version skew note

The demo server is opencode2 v0.0.0-next-17403 while the spec pin is v2-branch commit a6a712a — deliberately not the same snapshot. Pinned-snapshot policy (ADR-0005) is the mechanism that absorbs upstream's continuous pre-GA churn: refreshes are deliberate, at milestone boundaries (M2 is the next sanctioned moment), never HEAD-tracking.

Review follow-ups

Multi-agent review findings triaged into issues #17–#25 (milestone-anchored; see docs/ROADMAP.md). The blocker set and the performance-test infrastructure landed on this branch.

Fixes #17. Fixes #18.