orizu 0.4.1 → 0.5.0

package/dist/hosted-harness-opencode.js

@@ -0,0 +1,729 @@
+/**
+ * OpenCode driver for the AgentHarness seam (ALI-926 / P3.4, per ADR-005 §4a).
+ *
+ * The `opencode` implementation of `AgentHarness`. Drives a long-running
+ * `opencode serve` process over localhost HTTP/SSE, exactly as OpenInspect's
+ * `bridge.py` does — this module is a DESIGN port of `bridge.py`'s
+ * `_stream_opencode_response_sse` + `_transform_part_to_event` into idiomatic
+ * TypeScript. The WebSocket half of `bridge.py` is NOT ported: the transport is
+ * the RunAPI event bridge (`hosted-run-event-sink.ts`), not a Durable Object.
+ *
+ * DRIVING PROTOCOL (all localhost):
+ *   - POST /session                          → create an OpenCode session
+ *   - GET  /session/{id}                     → validate a resumed session
+ *   - GET  /event                            → SSE stream of message parts
+ *   - POST /session/{id}/prompt_async        → enqueue a prompt (non-blocking)
+ *   - GET  /session/{id}/message             → final-state reconciliation
+ *   - POST /session/{id}/abort               → stop the in-flight prompt
+ *
+ * The stream is opened BEFORE the prompt is POSTed (mirrors bridge.py) so no
+ * early parts are missed. Termination is driven by `session.idle` /
+ * `session.status{idle}` (→ `execution_complete`) or `session.error` (→
+ * `error`). Sub-task / child-session fan-out is intentionally NOT ported —
+ * ADR-005 keeps the RunAPI single-writer-per-run, so `spawn-task` is out of P3.
+ *
+ * TESTABILITY: the driver takes a `baseUrl` (localhost:port) and an injectable
+ * `fetchImpl`, so it runs against a fake in-process OpenCode SSE server with no
+ * real `opencode` binary. `spawnOpenCode` (below) is the P3.5 in-sandbox launch
+ * helper; the driver itself never spawns a process.
+ *
+ * ── Deliberate deltas from bridge.py (kept in ONE place) ─────────────────────
+ * This driver is a faithful port EXCEPT for these consciously-chosen
+ * differences — read them before diffing behavior against `bridge.py`:
+ *   • Sub-task / child-session fan-out is OUT OF SCOPE. bridge.py tracks
+ *     `tracked_child_session_ids`, authorizes child assistant messages, and
+ *     emits `isSubtask` events; ADR-005 keeps the RunAPI single-writer-per-run,
+ *     so `task`-tool child sessions are neither tracked nor forwarded here.
+ *   • Compaction IS NOW PORTED (P2-1). `session.compacted` sets a
+ *     `compactionOccurred` flag; thereafter assistant messages whose parentID
+ *     no longer matches the user message are accepted (except the compaction
+ *     summary itself), in both the live stream and `fetchFinalState` — mirrors
+ *     bridge.py lines ~1300-1307 / 1153-1160 / 1421-1425.
+ *   • Max prompt duration IS NOW PORTED (P3-1). `PROMPT_MAX_DURATION`
+ *     (default 5400s, injectable) runs ALONGSIDE the SSE inactivity deadline;
+ *     on breach the prompt is aborted and a terminal `error` is emitted.
+ *   • SSE framing assumes LF `\n\n` event separators and `data:` lines, exactly
+ *     as OpenCode 1.14.41 emits them (bridge.py `_parse_sse_stream` makes the
+ *     same assumption — it splits on `"\n\n"`, never CRLF). This is pinned to
+ *     `OPENCODE_PINNED_VERSION`; a bump that switches OpenCode to CRLF framing
+ *     would break both ports and must be re-verified.
+ */
+// OpenCode is HARD-PINNED to this version everywhere it is installed/launched.
+// AUDIT RISK #1 (SSE fragility): the entire event model rides OpenCode's
+// undocumented `/event` SSE part schema (`message.part.updated`, `session.idle`,
+// `session.error`, `session.updated`, ...). Versions newer than 1.14.41 broke
+// that stream upstream, so the pin is load-bearing and must be re-verified
+// before any bump. The AgentHarness seam is the mitigation of last resort: swap
+// to a Claude-Agent-SDK loop (ALI-929) rather than chase OpenCode's SSE.
+export const OPENCODE_PINNED_VERSION = '1.14.41';
+const DEFAULT_OPENCODE_PORT = 4096;
+const DEFAULT_SSE_INACTIVITY_MS = 120_000;
+const DEFAULT_REQUEST_TIMEOUT_MS = 30_000;
+// Wall-clock ceiling on a single prompt (port of bridge.py PROMPT_MAX_DURATION,
+// 5400s). Runs alongside the inactivity deadline: inactivity catches a stalled
+// stream, this catches a stream that stays busy forever.
+const DEFAULT_PROMPT_MAX_DURATION_MS = 5_400_000;
+// Placeholder titles OpenCode auto-assigns ("New session - <ISO8601>Z" /
+// "Child session - <ISO8601>Z"); noise, never forwarded (port of bridge.py
+// OPENCODE_DEFAULT_TITLE_RE).
+const OPENCODE_DEFAULT_TITLE_RE = /^(new session|child session) - \d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$/i;
+/** Distinct terminal reason emitted when a prompt breaches PROMPT_MAX_DURATION,
+ *  so the bridge/tests can tell a duration abort from a model error. */
+export const PROMPT_MAX_DURATION_ERROR = 'prompt exceeded max duration';
+class SseInactivityError extends Error {
+    constructor(ms) {
+        super(`OpenCode SSE stream inactive for ${ms}ms (no data received)`);
+        this.name = 'SseInactivityError';
+    }
+}
+class PromptAbortedError extends Error {
+    constructor() {
+        super('prompt aborted');
+        this.name = 'PromptAbortedError';
+    }
+}
+// -- OpenCode ascending id generator (port of OpenCode's id.ts) --------------
+// Monotonic so our user-message id always sorts AFTER any prior assistant
+// message id — this is what prevents OpenCode's prompt loop from early-exiting
+// on `lastUser.id < lastAssistant.id`.
+const BASE62 = '0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz';
+let idLastTimestamp = 0;
+let idCounter = 0;
+function randomBase62(length) {
+    let out = '';
+    for (let i = 0; i < length; i += 1) {
+        out += BASE62[Math.floor(Math.random() * 62)];
+    }
+    return out;
+}
+function ascendingId(prefix) {
+    const nowMs = Date.now();
+    if (nowMs !== idLastTimestamp) {
+        idLastTimestamp = nowMs;
+        idCounter = 0;
+    }
+    idCounter += 1;
+    const encoded = (nowMs * 0x1000 + idCounter) & 0xffffffffffff;
+    const hex = encoded.toString(16).padStart(12, '0');
+    return `${prefix}_${hex}${randomBase62(14)}`;
+}
+// -- Anthropic / OpenAI reasoning-effort mapping (port of bridge.py) ---------
+const ANTHROPIC_THINKING_BUDGETS = { high: 16_000, max: 31_999 };
+const ANTHROPIC_ADAPTIVE_THINKING_MODELS = new Set([
+    'claude-fable-5',
+    'claude-opus-4-6',
+    'claude-opus-4-7',
+    'claude-opus-4-8',
+    'claude-sonnet-4-6',
+]);
+const ANTHROPIC_ADAPTIVE_EFFORTS = new Set(['low', 'medium', 'high', 'xhigh', 'max']);
+export function buildPromptRequestBody(content, model, opencodeMessageId, reasoningEffort) {
+    const body = {
+        parts: [{ type: 'text', text: content }],
+        messageID: opencodeMessageId,
+    };
+    if (!model)
+        return body;
+    const [providerId, modelId] = model.includes('/') ? model.split('/', 2) : ['anthropic', model];
+    const modelSpec = { providerID: providerId, modelID: modelId };
+    if (reasoningEffort) {
+        if (providerId === 'anthropic') {
+            if (ANTHROPIC_ADAPTIVE_THINKING_MODELS.has(modelId)) {
+                const options = { thinking: { type: 'adaptive' } };
+                if (ANTHROPIC_ADAPTIVE_EFFORTS.has(reasoningEffort)) {
+                    options.outputConfig = { effort: reasoningEffort };
+                }
+                modelSpec.options = options;
+            }
+            else {
+                const budget = ANTHROPIC_THINKING_BUDGETS[reasoningEffort];
+                if (budget !== undefined) {
+                    modelSpec.options = { thinking: { type: 'enabled', budgetTokens: budget } };
+                }
+            }
+        }
+        else if (providerId === 'openai') {
+            modelSpec.options = { reasoningEffort, reasoningSummary: 'auto' };
+        }
+    }
+    body.model = modelSpec;
+    return body;
+}
+// -- OpenCode part → HarnessEvent transform (port of _transform_part_to_event) -
+function asRecord(value) {
+    return value && typeof value === 'object' && !Array.isArray(value)
+        ? value
+        : {};
+}
+function extractErrorMessage(error) {
+    const rec = asRecord(error);
+    const data = asRecord(rec.data);
+    if (typeof data.message === 'string')
+        return data.message;
+    if (typeof rec.message === 'string')
+        return rec.message;
+    if (typeof rec.name === 'string')
+        return rec.name;
+    return typeof error === 'string' && error ? error : 'Unknown error';
+}
+// -- SSE parsing over a fetch ReadableStream ---------------------------------
+async function readWithDeadline(reader, signal, inactivityMs) {
+    return new Promise((resolve, reject) => {
+        if (signal.aborted) {
+            reject(new PromptAbortedError());
+            return;
+        }
+        const timer = setTimeout(() => {
+            cleanup();
+            reject(new SseInactivityError(inactivityMs));
+        }, inactivityMs);
+        const onAbort = () => {
+            cleanup();
+            reject(new PromptAbortedError());
+        };
+        const cleanup = () => {
+            clearTimeout(timer);
+            signal.removeEventListener('abort', onAbort);
+        };
+        signal.addEventListener('abort', onAbort, { once: true });
+        reader.read().then(result => {
+            cleanup();
+            resolve(result);
+        }, error => {
+            cleanup();
+            reject(error);
+        });
+    });
+}
+async function* parseSseEvents(body, signal, inactivityMs) {
+    const reader = body.getReader();
+    const decoder = new TextDecoder();
+    let buffer = '';
+    try {
+        for (;;) {
+            const { done, value } = await readWithDeadline(reader, signal, inactivityMs);
+            if (done)
+                break;
+            if (value)
+                buffer += decoder.decode(value, { stream: true });
+            let boundary = buffer.indexOf('\n\n');
+            while (boundary !== -1) {
+                const rawEvent = buffer.slice(0, boundary);
+                buffer = buffer.slice(boundary + 2);
+                const dataLines = [];
+                for (const line of rawEvent.split('\n')) {
+                    if (line.startsWith('data:')) {
+                        const trimmed = line.slice(5).replace(/^ /, '');
+                        if (trimmed)
+                            dataLines.push(trimmed);
+                    }
+                }
+                if (dataLines.length > 0) {
+                    try {
+                        yield asRecord(JSON.parse(dataLines.join('\n')));
+                    }
+                    catch {
+                        // Malformed SSE payload — skip rather than crash the stream.
+                    }
+                }
+                boundary = buffer.indexOf('\n\n');
+            }
+        }
+    }
+    finally {
+        try {
+            reader.releaseLock();
+        }
+        catch {
+            // reader already released
+        }
+        try {
+            await body.cancel();
+        }
+        catch {
+            // stream already closed
+        }
+    }
+}
+// -- The driver --------------------------------------------------------------
+export function createOpenCodeHarness(options) {
+    const fetchImpl = options.fetchImpl ?? globalThis.fetch;
+    const base = options.baseUrl.replace(/\/$/, '');
+    const inactivityMs = options.sseInactivityMs ?? DEFAULT_SSE_INACTIVITY_MS;
+    const requestTimeoutMs = options.requestTimeoutMs ?? DEFAULT_REQUEST_TIMEOUT_MS;
+    const promptMaxDurationMs = options.promptMaxDurationMs ?? DEFAULT_PROMPT_MAX_DURATION_MS;
+    const now = options.now ?? (() => Date.now());
+    let opencodeSessionId = null;
+    let startModel;
+    let startReasoningEffort;
+    let readyEmitted = false;
+    // Set true by the public stop(); makes the next idle terminal resolve as a
+    // cancellation rather than a success (P2-2).
+    let stopRequested = false;
+    // Last title actually forwarded, deduped across the whole harness lifetime so
+    // a repeated identical title is emitted once (P3-6).
+    let lastForwardedTitle = null;
+    /** Drop placeholder/default titles and de-dupe against the last one forwarded;
+     *  returns the title to emit, or null to suppress (port of bridge.py
+     *  _normalize_forwardable_session_title + _session_title_event_once). */
+    function forwardableTitle(rawTitle) {
+        if (typeof rawTitle !== 'string')
+            return null;
+        const trimmed = rawTitle.trim();
+        if (!trimmed || OPENCODE_DEFAULT_TITLE_RE.test(trimmed))
+            return null;
+        if (trimmed === lastForwardedTitle)
+            return null;
+        lastForwardedTitle = trimmed;
+        return trimmed;
+    }
+    async function timedFetch(url, init = {}) {
+        const controller = new AbortController();
+        const timer = setTimeout(() => controller.abort(), requestTimeoutMs);
+        try {
+            return await fetchImpl(url, { ...init, signal: init.signal ?? controller.signal });
+        }
+        finally {
+            clearTimeout(timer);
+        }
+    }
+    async function createSession() {
+        const res = await timedFetch(`${base}/session`, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({}),
+        });
+        if (!res.ok)
+            throw new Error(`OpenCode session create failed (${res.status})`);
+        const data = asRecord(await res.json());
+        const id = typeof data.id === 'string' ? data.id : null;
+        if (!id)
+            throw new Error('OpenCode session create returned no id');
+        return id;
+    }
+    async function validateSession(id) {
+        try {
+            const res = await timedFetch(`${base}/session/${encodeURIComponent(id)}`);
+            return res.ok;
+        }
+        catch {
+            return false;
+        }
+    }
+    async function requestStop() {
+        if (!opencodeSessionId)
+            return;
+        try {
+            await timedFetch(`${base}/session/${encodeURIComponent(opencodeSessionId)}/abort`, {
+                method: 'POST',
+            });
+        }
+        catch {
+            // best-effort: the stream teardown is the real cancellation
+        }
+    }
+    async function* fetchFinalState(sessionId, userMessageIds, allowedAssistantMsgIds, messageText, compactionOccurred) {
+        // After idle, re-read the message list to capture any assistant text the SSE
+        // stream may have missed due to part-ordering (bridge.py does the same). Any
+        // text LONGER than what we already emitted is re-emitted as a token snapshot.
+        let messages;
+        try {
+            const res = await timedFetch(`${base}/session/${encodeURIComponent(sessionId)}/message`);
+            if (!res.ok)
+                return;
+            messages = await res.json();
+        }
+        catch {
+            return;
+        }
+        if (!Array.isArray(messages))
+            return;
+        for (const msg of messages) {
+            const info = asRecord(asRecord(msg).info);
+            if (info.role !== 'assistant')
+                continue;
+            const msgId = typeof info.id === 'string' ? info.id : '';
+            const parentId = typeof info.parentID === 'string' ? info.parentID : '';
+            const isCompactionSummary = info.summary === true;
+            // Accept if parentID matches our user message, OR it was tracked live, OR
+            // compaction re-chained the messages and this isn't the summary itself
+            // (port of bridge.py _fetch_final_message_state, lines ~1420-1426).
+            const accept = userMessageIds.has(parentId) ||
+                (msgId !== '' && allowedAssistantMsgIds.has(msgId)) ||
+                (compactionOccurred && !isCompactionSummary);
+            if (!accept)
+                continue;
+            const parts = Array.isArray(asRecord(msg).parts) ? asRecord(msg).parts : [];
+            const combined = parts
+                .map(p => asRecord(p))
+                .filter(p => p.type === 'text')
+                .map(p => (typeof p.text === 'string' ? p.text : ''))
+                .join('');
+            const previous = messageText.get(msgId) ?? '';
+            if (combined.length > previous.length) {
+                messageText.set(msgId, combined);
+                yield { kind: 'token', messageId: msgId, payload: { text: combined } };
+            }
+        }
+    }
+    return {
+        name: 'opencode',
+        async start(opts) {
+            startModel = opts.model;
+            startReasoningEffort = opts.reasoningEffort;
+            if (opts.resumeAgentSessionId && (await validateSession(opts.resumeAgentSessionId))) {
+                opencodeSessionId = opts.resumeAgentSessionId;
+            }
+            else {
+                opencodeSessionId = await createSession();
+            }
+            return { agentSessionId: opencodeSessionId };
+        },
+        async *runPrompt(prompt, signal) {
+            if (!opencodeSessionId)
+                opencodeSessionId = await createSession();
+            const sessionId = opencodeSessionId;
+            // Per-prompt state: a fresh stop intent and compaction flag each run.
+            stopRequested = false;
+            let compactionOccurred = false;
+            const promptStart = now();
+            if (!readyEmitted) {
+                readyEmitted = true;
+                yield {
+                    kind: 'ready',
+                    messageId: prompt.messageId,
+                    payload: { agentSessionId: sessionId },
+                };
+            }
+            const opencodeMessageId = ascendingId('msg');
+            const requestBody = buildPromptRequestBody(prompt.content, startModel, opencodeMessageId, startReasoningEffort);
+            // Per-part cumulative text, ordered per assistant message so a coalesced
+            // token snapshot carries the whole assistant turn (not a single part).
+            const partText = new Map();
+            const messagePartOrder = new Map();
+            const messageText = (msgId) => (messagePartOrder.get(msgId) ?? []).map(pid => partText.get(pid) ?? '').join('');
+            const allowedAssistantMsgIds = new Set();
+            const userMessageIds = new Set([opencodeMessageId]);
+            const pendingParts = new Map();
+            const emittedToolStates = new Set();
+            const handlePart = (part, delta) => {
+                const partType = typeof part.type === 'string' ? part.type : '';
+                const partId = typeof part.id === 'string' ? part.id : '';
+                const msgId = typeof part.messageID === 'string' ? part.messageID : '';
+                const events = [];
+                if (partType === 'text') {
+                    const full = typeof part.text === 'string' ? part.text : '';
+                    if (typeof delta === 'string' && delta) {
+                        partText.set(partId, (partText.get(partId) ?? '') + delta);
+                    }
+                    else {
+                        partText.set(partId, full);
+                    }
+                    const order = messagePartOrder.get(msgId) ?? [];
+                    if (!order.includes(partId)) {
+                        order.push(partId);
+                        messagePartOrder.set(msgId, order);
+                    }
+                    const text = messageText(msgId);
+                    if (text)
+                        events.push({ kind: 'token', messageId: msgId, payload: { text } });
+                }
+                else if (partType === 'tool') {
+                    const state = asRecord(part.state);
+                    const status = typeof state.status === 'string' ? state.status : '';
+                    const input = asRecord(state.input);
+                    if ((status === 'pending' || status === '') && Object.keys(input).length === 0) {
+                        return events;
+                    }
+                    const callId = typeof part.callID === 'string' ? part.callID : '';
+                    const partSid = typeof part.sessionID === 'string' ? part.sessionID : '';
+                    const toolKey = `tool:${partSid}:${callId}:${status}`;
+                    if (emittedToolStates.has(toolKey))
+                        return events;
+                    emittedToolStates.add(toolKey);
+                    const terminal = status === 'completed' || status === 'error';
+                    const payload = {
+                        tool: typeof part.tool === 'string' ? part.tool : '',
+                        args: input,
+                        callId,
+                        status,
+                    };
+                    if (terminal)
+                        payload.output = typeof state.output === 'string' ? state.output : '';
+                    events.push({ kind: terminal ? 'tool_result' : 'tool_call', messageId: msgId, payload });
+                }
+                else if (partType === 'step-start') {
+                    events.push({ kind: 'step_start', messageId: msgId, payload: {} });
+                }
+                else if (partType === 'step-finish') {
+                    events.push({
+                        kind: 'step_finish',
+                        messageId: msgId,
+                        payload: { cost: part.cost ?? null, tokens: part.tokens ?? null, reason: part.reason ?? null },
+                    });
+                }
+                return events;
+            };
+            let sseResponse;
+            try {
+                sseResponse = await fetchImpl(`${base}/event`, {
+                    headers: { Accept: 'text/event-stream' },
+                    signal,
+                });
+            }
+            catch (error) {
+                yield errorEvent(prompt.messageId, error);
+                return;
+            }
+            if (!sseResponse.ok || !sseResponse.body) {
+                // Cancel the body on the failure path too — a non-2xx response can still
+                // carry an (unconsumed) body that would otherwise leak (P3-7).
+                await cancelBody(sseResponse.body);
+                yield errorEvent(prompt.messageId, new Error(`SSE connection failed (${sseResponse.status})`));
+                return;
+            }
+            const sseBody = sseResponse.body;
+            let sawError = false;
+            let errorText = null;
+            let terminalYielded = false;
+            try {
+                const promptRes = await timedFetch(`${base}/session/${encodeURIComponent(sessionId)}/prompt_async`, {
+                    method: 'POST',
+                    headers: { 'Content-Type': 'application/json' },
+                    body: JSON.stringify(requestBody),
+                });
+                if (promptRes.status !== 200 && promptRes.status !== 204) {
+                    throw new Error(`Async prompt failed (${promptRes.status})`);
+                }
+                for await (const event of parseSseEvents(sseBody, signal, inactivityMs)) {
+                    // Wall-clock max-duration guard (P3-1): a stream that stays busy (or
+                    // heartbeats) forever never trips the inactivity deadline, so bound it
+                    // here — checked on EVERY event, ahead of the heartbeat/continue paths,
+                    // so no event class can starve it (port of bridge.py PROMPT_MAX_DURATION).
+                    if (now() - promptStart > promptMaxDurationMs) {
+                        await requestStop();
+                        const durationTextMap = new Map();
+                        for (const [msgId, order] of messagePartOrder.entries()) {
+                            durationTextMap.set(msgId, order.map(pid => partText.get(pid) ?? '').join(''));
+                        }
+                        for await (const ev of fetchFinalState(sessionId, userMessageIds, allowedAssistantMsgIds, durationTextMap, compactionOccurred)) {
+                            yield ev;
+                        }
+                        terminalYielded = true;
+                        yield {
+                            kind: 'error',
+                            messageId: prompt.messageId,
+                            critical: true,
+                            payload: {
+                                error: `${PROMPT_MAX_DURATION_ERROR} of ${Math.round(promptMaxDurationMs / 1000)}s`,
+                            },
+                        };
+                        return;
+                    }
+                    const eventType = event.type;
+                    if (eventType === 'server.heartbeat' || eventType === 'server.connected')
+                        continue;
+                    const props = asRecord(event.properties);
+                    // session_title (session.updated → session_title). Placeholder/default
+                    // titles are dropped and identical repeats are deduped (P3-6).
+                    if (eventType === 'session.updated') {
+                        const info = asRecord(props.info);
+                        const sid = props.sessionID ?? info.id;
+                        if (sid === sessionId) {
+                            const title = forwardableTitle(info.title);
+                            if (title)
+                                yield { kind: 'session_title', messageId: prompt.messageId, payload: { title } };
+                        }
+                        continue;
+                    }
+                    const eventSessionId = props.sessionID ?? asRecord(props.part).sessionID;
+                    if (eventSessionId && eventSessionId !== sessionId)
+                        continue;
+                    if (eventType === 'message.updated') {
+                        const info = asRecord(props.info);
+                        if (info.sessionID !== sessionId)
+                            continue;
+                        const ocMsgId = typeof info.id === 'string' ? info.id : '';
+                        const parentId = typeof info.parentID === 'string' ? info.parentID : '';
+                        const role = info.role;
+                        const isCompactionSummary = info.summary === true;
+                        if (role === 'user' && ocMsgId)
+                            userMessageIds.add(ocMsgId);
+                        // Accept an assistant message when its parentID matches our user
+                        // message, OR compaction has re-chained the conversation and this is
+                        // not the compaction summary itself (P2-1; port of bridge.py
+                        // lines ~1153-1160).
+                        const acceptAssistant = userMessageIds.has(parentId) || (compactionOccurred && !isCompactionSummary);
+                        if (role === 'assistant' && ocMsgId && acceptAssistant) {
+                            allowedAssistantMsgIds.add(ocMsgId);
+                            const buffered = pendingParts.get(ocMsgId);
+                            if (buffered) {
+                                pendingParts.delete(ocMsgId);
+                                for (const { part, delta } of buffered) {
+                                    for (const ev of handlePart(part, delta))
+                                        yield ev;
+                                }
+                            }
+                        }
+                    }
+                    else if (eventType === 'message.part.updated') {
+                        const part = asRecord(props.part);
+                        const delta = props.delta;
+                        const ocMsgId = typeof part.messageID === 'string' ? part.messageID : '';
+                        if (allowedAssistantMsgIds.has(ocMsgId)) {
+                            for (const ev of handlePart(part, delta))
+                                yield ev;
+                        }
+                        else if (ocMsgId) {
+                            const list = pendingParts.get(ocMsgId) ?? [];
+                            list.push({ part, delta });
+                            pendingParts.set(ocMsgId, list);
+                        }
+                    }
+                    else if (eventType === 'session.compacted') {
+                        // Compaction re-chains messages so subsequent assistant parentIDs no
+                        // longer point at our user message; flip the flag so acceptance falls
+                        // through to the compaction branch above (P2-1).
+                        const compactedSid = props.sessionID ?? asRecord(props.info).id;
+                        if (compactedSid === sessionId)
+                            compactionOccurred = true;
+                        continue;
+                    }
+                    else if (eventType === 'session.idle' ||
+                        (eventType === 'session.status' && asRecord(props.status).type === 'idle')) {
+                        const messageTextMap = new Map();
+                        for (const [msgId, order] of messagePartOrder.entries()) {
+                            messageTextMap.set(msgId, order.map(pid => partText.get(pid) ?? '').join(''));
+                        }
+                        for await (const ev of fetchFinalState(sessionId, userMessageIds, allowedAssistantMsgIds, messageTextMap, compactionOccurred)) {
+                            yield ev;
+                        }
+                        terminalYielded = true;
+                        // A stop()-initiated idle is a cancellation, not a success (P2-2).
+                        yield {
+                            kind: 'execution_complete',
+                            messageId: prompt.messageId,
+                            critical: true,
+                            payload: stopRequested ? { success: false, aborted: true } : { success: true },
+                        };
+                        return;
+                    }
+                    else if (eventType === 'session.error') {
+                        sawError = true;
+                        errorText = extractErrorMessage(props.error);
+                        terminalYielded = true;
+                        yield {
+                            kind: 'error',
+                            messageId: prompt.messageId,
+                            critical: true,
+                            payload: { error: errorText },
+                        };
+                        return;
+                    }
+                }
+                // Stream ended without an explicit idle/error terminal (server closed).
+                if (!terminalYielded) {
+                    yield {
+                        kind: 'execution_complete',
+                        messageId: prompt.messageId,
+                        critical: true,
+                        payload: sawError ? { success: false, error: errorText } : { success: true },
+                    };
+                }
+            }
+            catch (error) {
+                if (error instanceof PromptAbortedError) {
+                    await requestStop();
+                    if (!terminalYielded) {
+                        // Signal-abort resolves as a cancellation (P2-2): `aborted: true` is
+                        // the bridge's cue to finish('cancelled'), never 'failed'.
+                        yield {
+                            kind: 'execution_complete',
+                            messageId: prompt.messageId,
+                            critical: true,
+                            payload: { success: false, aborted: true },
+                        };
+                    }
+                    return;
+                }
+                await requestStop();
+                if (!terminalYielded)
+                    yield errorEvent(prompt.messageId, error);
+            }
+            finally {
+                // Guaranteed body teardown on EVERY exit — including a prompt POST that
+                // throws before parseSseEvents (whose own finally would otherwise be the
+                // only canceller) — so the /event response never leaks (P3-7).
+                await cancelBody(sseBody);
+            }
+        },
+        async stop() {
+            // Mark intent so the resulting idle terminal is reported as cancelled, not
+            // succeeded (P2-2), then best-effort ask OpenCode to abort.
+            stopRequested = true;
+            await requestStop();
+        },
+        async shutdown() {
+            opencodeSessionId = null;
+            readyEmitted = false;
+            stopRequested = false;
+        },
+    };
+}
+/** Best-effort cancel of a fetch response body; never throws. */
+async function cancelBody(body) {
+    if (!body)
+        return;
+    try {
+        await body.cancel();
+    }
+    catch {
+        // already released/closed
+    }
+}
+function errorEvent(messageId, error) {
+    const message = error instanceof Error ? error.message : String(error);
+    return { kind: 'error', messageId, critical: true, payload: { error: message } };
+}
+/**
+ * Build the `OPENCODE_CONFIG_CONTENT` payload OpenCode reads at boot (mirrors
+ * OpenInspect's `start_opencode`). Kept pure so the launch command and the
+ * config are independently testable.
+ */
+export function buildOpenCodeConfigContent(opts) {
+    return JSON.stringify({
+        model: opts.model,
+        permission: opts.permission ?? { '*': { '*': 'allow' } },
+    });
+}
+/**
+ * Launch `opencode serve` locally. Process supervision is intentionally minimal
+ * for this slice — P3.5 runs this INSIDE the Daytona sandbox via the CLI. The
+ * pinned version (`OPENCODE_PINNED_VERSION`) is the one the sandbox image must
+ * install (see the SSE-fragility note on that constant). Returns the localhost
+ * base URL the driver connects to; callers still `start()`/`runPrompt()` the
+ * harness against it.
+ */
+export function spawnOpenCode(opts) {
+    const port = opts.port ?? DEFAULT_OPENCODE_PORT;
+    const env = {
+        OPENCODE_CONFIG_CONTENT: buildOpenCodeConfigContent(opts),
+        ...(opts.env ?? {}),
+    };
+    const cmd = ['opencode', 'serve', '--port', String(port)];
+    const spawnImpl = opts.spawn ??
+        ((command, spawnOpts) => {
+            const bun = globalThis.Bun;
+            if (!bun?.spawn)
+                throw new Error('spawnOpenCode requires Bun.spawn or an injected spawner');
+            return bun.spawn({
+                cmd: command,
+                cwd: spawnOpts.cwd,
+                env: { ...process.env, ...spawnOpts.env },
+                stdout: 'ignore',
+                stderr: 'ignore',
+            });
+        });
+    const child = spawnImpl(cmd, { cwd: opts.cwd, env });
+    return { baseUrl: `http://localhost:${port}`, port, stop: () => child.kill() };
+}