shockbot/agents/opencode.ts

/**
 * OpenCode agent — secure harness around OpenCode CLI.
 *
 * transparently wraps OpenCode with a security layer:
 * - bash: "deny" via OPENCODE_CONFIG_CONTENT (agent cannot shell out)
 * - OPENCODE_PERMISSION: filesystem sandbox — deny all external paths except /tmp
 * - MCP ShellTool provides restricted shell (filtered env, no secrets)
 * - MCP server injected alongside project config (not replacing)
 * - ASKPASS handles git auth separately (token never in subprocess env)
 *
 * the agent process itself gets full env (needs LLM API keys, PATH, etc.).
 * security is enforced at the tool layer, not the process layer.
 */
import { execFileSync } from "node:child_process";
import { mkdirSync, writeFileSync } from "node:fs";
import { join } from "node:path";
import { performance } from "node:perf_hooks";
import { pullfrogMcpName } from "../external.ts";
import { BEDROCK_MODEL_ID_ENV, modelAliases } from "../models.ts";
import type { ToolState } from "../toolState.ts";
import { getIdleMs, markActivity } from "../utils/activity.ts";
import { type AgentDiagnostic, formatAgentHangBody } from "../utils/agentHangReport.ts";
import { formatJsonValue, log } from "../utils/cli.ts";
import { installFromNpmTarball } from "../utils/install.ts";
import { findProviderErrorMatch } from "../utils/providerErrors.ts";
import { addSkill, installBundledSkills } from "../utils/skills.ts";
import {
  DEFAULT_MAX_RETAINED_BYTES,
  SPAWN_ACTIVITY_TIMEOUT_CODE,
  SpawnTimeoutError,
  spawn,
  TailBuffer,
} from "../utils/subprocess.ts";
import { ThinkingTimer } from "../utils/timer.ts";
import type { TodoTracker } from "../utils/todoTracking.ts";
import { getDevDependencyVersion } from "../utils/version.ts";
import {
  PULLFROG_BUS_EVENT_TYPE,
  PULLFROG_OPENCODE_PLUGIN_FILENAME,
  PULLFROG_OPENCODE_PLUGIN_SOURCE,
} from "./opencodePlugin.ts";
import {
  buildLearningsReflectionPrompt,
  runPostRunRetryLoop,
  shouldRunReflection,
} from "./postRun.ts";
import { REVIEWER_AGENT_NAME, REVIEWER_SYSTEM_PROMPT } from "./reviewer.ts";
import { formatWithLabel, ORCHESTRATOR_LABEL, SessionLabeler } from "./sessionLabeler.ts";
import {
  type AgentResult,
  type AgentRunContext,
  type AgentUsage,
  agent,
  logTokenTable,
  MAX_STDERR_LINES,
} from "./shared.ts";
import { deriveSubagentModels } from "./subagentModels.ts";

async function installOpencodeCli(): Promise<string> {
  return await installFromNpmTarball({
    packageName: "opencode-ai",
    version: getDevDependencyVersion("opencode-ai"),
    executablePath: "bin/opencode",
    installDependencies: true,
  });
}

// ── config ─────────────────────────────────────────────────────────────────────

type OpenCodeConfig = {
  mcp?: Record<string, unknown>;
  permission?: Record<string, unknown>;
  provider?: Record<string, unknown>;
  agent?: Record<string, unknown>;
  experimental?: Record<string, unknown>;
  model?: string;
  enabled_providers?: string[];
  [key: string]: unknown;
};

// NOTE: OpenCode's per-call `max_tokens` defaults to 32_000. We previously
// overrode this via `OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX = 5000` in #616
// to lower OpenRouter's per-call upfront budget reservation — back when the
// `ROUTER_PER_RUN_LIMIT_USD = 25` per-run key cap meant that reservation was
// a hard gate that could lock low-balance accounts out of starting a run.
//
// That gate is gone (see `app/api/proxy-token/route.ts` ~line 422 — "Per-run
// key budget … is decoupled from wallet balance"); the router now mints
// keys with `keyLimitCents = balance + buffer` ($50 / $5 / $0). The override
// no longer materially helps, and as a hard per-call output truncation it
// actively hurt: a single `create_pull_request_review` tool_use with many
// inline comments would truncate mid-stream past 5K output tokens, the JSON
// was unparseable, and the tool never invoked. We hit this on PR #710's
// verify-downshift PR. Removed in #710 — using OpenCode's 32K default.
//
// If you need to re-cap output for some reason, set
// `OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX` in the action env. OpenCode's
// top-level `limit.output` config field has no read site (silently dropped
// on merge in session/llm.ts), so the env var is the only working knob.

/**
 * Build the `provider.google.models[id].options` map that pins every direct-Google
 * Gemini alias to `thinkingLevel: "high"`. Sourced from the model registry so
 * adding/renaming a Google alias in `action/models.ts` flows through automatically.
 */
export function geminiHighThinkingOverrides(): Record<string, { options: object }> {
  return Object.fromEntries(
    modelAliases
      .filter((a) => a.provider === "google")
      .map((a) => [
        a.resolve.replace(/^google\//, ""),
        { options: { thinkingConfig: { thinkingLevel: "high" } } },
      ])
  );
}

function buildSecurityConfig(ctx: AgentRunContext, model: string | undefined): string {
  const config: OpenCodeConfig = {
    permission: {
      bash: "deny",
      edit: "allow",
      read: "allow",
      webfetch: "allow",
      external_directory: "allow",
      skill: "allow",
    },
    mcp: {
      [pullfrogMcpName]: { type: "remote", url: ctx.mcpServerUrl },
    },
    agent: (() => {
      const cfg = buildReviewerAgentConfig(model);
      const reviewerModel = (cfg[REVIEWER_AGENT_NAME] as { model?: string })?.model ?? "(inherit)";
      log.info(`» subagent models: reviewfrog=${reviewerModel}`);
      return cfg;
    })(),
    // opt into opencode's experimental `batch` tool (added in
    // anomalyco/opencode PR #2983, opt-in via `experimental.batch_tool`). it
    // exposes a single `batch` tool that runs 1-25 independent tool calls
    // (read/grep/glob/bash/etc.) concurrently in one assistant turn, which
    // collapses the dominant grep→20×read pattern into a single round trip.
    // edits are explicitly disallowed inside the batch upstream. paired with
    // the "Parallel tool execution" guidance in utils/instructions.ts so the
    // model actually reaches for it. see wiki/prompt.md.
    experimental: { batch_tool: true },
    // gemini-3 thinking pinned to high for review depth; gpt and anthropic
    // effort set elsewhere (gpt: upstream default, anthropic: --effort flag in claude.ts).
    provider: { google: { models: geminiHighThinkingOverrides() } },
  };

  if (model) {
    config.model = model;

    const slashIndex = model.indexOf("/");
    if (slashIndex > 0) {
      config.enabled_providers = [model.slice(0, slashIndex).toLowerCase()];
    }
  }

  return JSON.stringify(config);
}

/**
 * Read-only `reviewfrog` subagent for lens-based review.
 *
 * Non-mutative + non-recursive — enforced by the prose system prompt in
 * reviewer.ts.
 *
 * Per-subagent `model:` override is driven by the registry in
 * `action/models.ts` via each alias's `subagentModel` field — see
 * `deriveSubagentModels` for the reverse-lookup. Currently wired:
 * Anthropic opus → sonnet, OpenAI gpt-pro → gpt and gpt → gpt-5.4,
 * Google gemini-pro → gemini-flash. Other providers (xai, deepseek,
 * moonshot) and already-cheap tiers inherit (no override) — either the
 * absolute savings are too small to justify or there's no clean
 * cheaper-but-capable sibling.
 */
function buildReviewerAgentConfig(orchestratorModel: string | undefined): Record<string, unknown> {
  const overrides = deriveSubagentModels(orchestratorModel);
  return {
    [REVIEWER_AGENT_NAME]: {
      description:
        "Read-only review subagent for lens-based code review (correctness, security, billing-subsystem, etc.). " +
        "Reads only — no writes, no state-changing shell or MCP calls, no nested subagent dispatch.",
      mode: "subagent",
      prompt: REVIEWER_SYSTEM_PROMPT,
      ...(overrides.reviewer !== undefined ? { model: overrides.reviewer } : {}),
    },
  };
}

// ── model auto-select fallback ──────────────────────────────────────────────────
//
// steps 1–2 of model resolution (PULLFROG_MODEL env, slug resolution) are handled
// by resolveModel() in utils/agent.ts before the agent runs. this fallback only
// handles step 3: auto-select via `opencode models`.

function getOpenCodeModels(cliPath: string): string[] {
  try {
    const output = execFileSync(cliPath, ["models"], {
      encoding: "utf-8",
      timeout: 30_000,
      env: process.env,
    });
    return output
      .split("\n")
      .map((line) => line.trim())
      .filter(Boolean);
  } catch (error) {
    log.debug(
      `» failed to run \`opencode models\`: ${error instanceof Error ? error.message : String(error)}`
    );
    return [];
  }
}

const AUTO_SELECT_WARNING =
  "select a model explicitly in the Pullfrog console (https://pullfrog.com/console) to avoid this.";

function autoSelectModel(cliPath: string): string | undefined {
  const availableModels = getOpenCodeModels(cliPath);
  const availableSet = new Set(availableModels);
  if (availableSet.size > 0) {
    log.debug(`» opencode models (${availableSet.size}): ${availableModels.join(", ")}`);
    // skip hidden aliases (internal subagent-tier targets like opencode/gpt-5.4) —
    // they should never surface as a user-facing orchestrator pick. mirrors the
    // selectable-list filter in components/ModelSelector.tsx and action/commands/init.ts.
    const match =
      modelAliases.find((a) => !a.hidden && a.preferred && availableSet.has(a.resolve)) ??
      modelAliases.find((a) => !a.hidden && availableSet.has(a.resolve));
    if (match) {
      log.info(
        `» model: ${match.resolve} (auto-selected${match.preferred ? " — preferred" : ""} curated match)`
      );
      log.warning(`» model auto-selected. ${AUTO_SELECT_WARNING}`);
      return match.resolve;
    }
    log.info(
      `» opencode has ${availableSet.size} models but none match curated aliases — letting OpenCode auto-select`
    );
  }

  log.warning(`» no model resolved. letting OpenCode auto-select. ${AUTO_SELECT_WARNING}`);
  return undefined;
}

// ── NDJSON event types ─────────────────────────────────────────────────────────

interface OpenCodeInitEvent {
  type: "init";
  timestamp?: string;
  session_id?: string;
  model?: string;
  [key: string]: unknown;
}

interface OpenCodeMessageEvent {
  type: "message";
  timestamp?: string;
  role?: "user" | "assistant";
  content?: string;
  delta?: boolean;
  [key: string]: unknown;
}

interface OpenCodeTextEvent {
  type: "text";
  timestamp?: string;
  sessionID?: string;
  part?: { id?: string; type?: string; text?: string; [key: string]: unknown };
  [key: string]: unknown;
}

interface OpenCodeStepStartEvent {
  type: "step_start";
  timestamp?: string;
  sessionID?: string;
  part?: { id?: string; type?: string; [key: string]: unknown };
  [key: string]: unknown;
}

interface OpenCodeStepFinishEvent {
  type: "step_finish";
  timestamp?: string;
  sessionID?: string;
  part?: {
    id?: string;
    type?: string;
    reason?: string;
    cost?: number;
    tokens?: {
      input?: number;
      output?: number;
      reasoning?: number;
      cache?: { read?: number; write?: number };
    };
    [key: string]: unknown;
  };
  [key: string]: unknown;
}

/**
 * tool-part state, mirroring opencode's `ToolState` (anomalyco/opencode
 * `session/message-v2.ts`). error parts carry the reason on `error`,
 * completed parts on `output` — reading the wrong field is what caused
 * the silent `(no error message)` log in #662.
 *
 * Named `ToolPartState` locally (not `ToolState`) so it doesn't shadow the
 * action-wide `ToolState` imported above.
 */
type ToolPartState =
  | { status: "pending" | "running"; input?: unknown }
  | { status: "completed"; input?: unknown; output: string }
  | { status: "error"; input?: unknown; error: string };

interface OpenCodeToolUseEvent {
  type: "tool_use";
  timestamp?: number;
  sessionID?: string;
  part?: {
    id?: string;
    callID?: string;
    tool?: string;
    state?: ToolPartState;
  };
  [key: string]: unknown;
}

interface OpenCodeToolResultEvent {
  type: "tool_result";
  timestamp?: number;
  sessionID?: string;
  part?: { callID?: string; state?: ToolPartState };
  tool_id?: string;
  status?: "success" | "error";
  output?: string;
  [key: string]: unknown;
}

interface OpenCodeResultEvent {
  type: "result";
  timestamp?: string;
  status?: "success" | "error";
  stats?: {
    total_tokens?: number;
    input_tokens?: number;
    output_tokens?: number;
    duration_ms?: number;
    tool_calls?: number;
  };
  [key: string]: unknown;
}

interface OpenCodeErrorEvent {
  type: "error";
  timestamp?: string;
  sessionID?: string;
  // opencode emits the error message under `error.data.message`, not at the
  // top level. see anomalyco/opencode packages/opencode/src/cli/cmd/run.ts.
  error?: {
    name?: string;
    data?: { message?: string; [key: string]: unknown };
    [key: string]: unknown;
  };
  [key: string]: unknown;
}

/**
 * Envelope event emitted by our `.opencode/plugin/pullfrog-events.ts` (the
 * source lives in `opencodePlugin.ts`). The plugin subscribes to opencode's
 * bus via `bus.subscribeAll()` and re-emits non-orchestrator
 * `message.part.updated` events on stdout so subagent activity surfaces here.
 *
 * `bus_event.properties.part` matches the same `Part` shape that opencode's
 * `cli/cmd/run.ts` uses to drive its own emit() calls, so we can route the
 * inner part through the existing `tool_use` / `step_start` / `step_finish`
 * / `text` handlers by synthesizing the equivalent OpenCode-style event.
 */
interface OpenCodeBusEnvelopeEvent {
  type: "pullfrog_bus_event";
  bus_event?: {
    type?: string;
    properties?: {
      part?: {
        sessionID?: string;
        type?: string;
        time?: { end?: number | string };
        state?: { status?: string };
        [key: string]: unknown;
      };
      [key: string]: unknown;
    };
    [key: string]: unknown;
  };
  [key: string]: unknown;
}

type OpenCodeEvent =
  | OpenCodeInitEvent
  | OpenCodeMessageEvent
  | OpenCodeTextEvent
  | OpenCodeStepStartEvent
  | OpenCodeStepFinishEvent
  | OpenCodeToolUseEvent
  | OpenCodeToolResultEvent
  | OpenCodeResultEvent
  | OpenCodeErrorEvent
  | OpenCodeBusEnvelopeEvent;

// ── runner ──────────────────────────────────────────────────────────────────────

type RunParams = {
  label: string;
  cliPath: string;
  args: string[];
  cwd: string;
  env: Record<string, string | undefined>;
  toolState: ToolState;
  todoTracker?: TodoTracker | undefined;
  onActivityTimeout?: (() => void) | undefined;
  onToolUse?: ((event: { toolName: string; input: unknown }) => void) | undefined;
};

async function runOpenCode(params: RunParams): Promise<AgentResult> {
  const startTime = performance.now();
  let eventCount = 0;

  let finalOutput = "";
  let accumulatedTokens = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
  // per-step `part.cost` sums across the whole session. sourced from models.dev
  // inside opencode — present for every supported provider (Anthropic, OpenAI,
  // Google, xAI, DeepSeek, Moonshot, OpenRouter sub-providers, etc.).
  let accumulatedCostUsd = 0;
  let tokensLogged = false;
  const toolCallTimings = new Map<string, number>();
  let currentStepId: string | null = null;
  let currentStepType: string | null = null;
  let stepHistory: Array<{ stepId: string; stepType: string; toolCalls: string[] }> = [];

  // per-session labeler so parallel subagent log lines can be differentiated.
  // the orchestrator's task tool_use events seed the labeler; the next
  // previously-unseen sessionID consumes the head of the pending-label queue.
  // upstream opencode's `cli/cmd/run.ts` filters subagent events out of its
  // NDJSON stream (`part.sessionID !== sessionID`), so we ship a per-run
  // plugin (`action/agents/opencodePlugin.ts`, written into the tmpdir at
  // setup) that re-emits non-orchestrator `message.part.updated` events. those
  // arrive here as `pullfrog_bus_event` envelopes and feed the labeler with
  // real data per subagent session.
  const labeler = new SessionLabeler();
  function eventLabel(event: Record<string, unknown>): string {
    const sid = event.sessionID ?? event.session_id;
    return labeler.labelFor(typeof sid === "string" ? sid : null);
  }
  function withLabel(label: string, message: string): string {
    return label === ORCHESTRATOR_LABEL ? message : formatWithLabel(label, message);
  }

  // one ThinkingTimer per session — sharing a single timer across sessions
  // conflated cross-session interleaving (parent thinks → child tool_call,
  // or child returns → parent dispatches next) as parent thinking time. each
  // timer formats its log lines through the session label so the "thought
  // for X" attribution is visible in the merged stream.
  const thinkingTimers = new Map<string, ThinkingTimer>();
  function timerFor(label: string): ThinkingTimer {
    let t = thinkingTimers.get(label);
    if (!t) {
      const formatLine = (line: string) =>
        label === ORCHESTRATOR_LABEL ? line : formatWithLabel(label, line);
      t = new ThinkingTimer(formatLine);
      thinkingTimers.set(label, t);
    }
    return t;
  }

  // tracks per-task dispatch metadata so the matching tool_result can log a
  // labeled "» subagent finished: lens=X duration=Ys" line. this is the most
  // useful per-lens observability available given that subagent-internal
  // events aren't streamed.
  //
  // matching strategy is hybrid because opencode does NOT reliably emit a
  // tool_result with a callID equal to the originating tool_use.callID for
  // the `task` tool (verified empirically in T3 — 5 task dispatches recorded
  // here, 0 finish lines fired, yet aggregation succeeded so results did
  // arrive on the stream). we keep an exact-match Map for the fast path, and
  // also a FIFO queue for the fallback path where the callID mismatches.
  // the queue + map share entries by reference so popping one removes both.
  interface TaskDispatch {
    label: string;
    startedAt: number;
    toolUseCallID: string;
  }
  const taskDispatchByCallID = new Map<string, TaskDispatch>();
  const pendingTaskDispatches: TaskDispatch[] = [];
  // every non-task tool_use callID we've observed. lets us tell, on a
  // tool_result, whether its callID belongs to a known non-task tool (in
  // which case we never fall back to FIFO) or is unrecognised (in which case
  // a long-output result is a strong "this is probably a task result with a
  // mismatched callID" signal).
  const knownNonTaskCallIDs = new Set<string>();

  function emitSubagentFinished(
    dispatch: TaskDispatch,
    status: string,
    output: unknown,
    matchKind: "exact" | "fifo"
  ) {
    const subagentDuration = performance.now() - dispatch.startedAt;
    const outputStr = typeof output === "string" ? output : "";
    const outputPreview = outputStr.length > 120 ? `${outputStr.slice(0, 120)}…` : outputStr;
    const matchSuffix = matchKind === "fifo" ? " [fifo-matched]" : "";
    log.info(
      `» subagent finished: ${dispatch.label} (${(subagentDuration / 1000).toFixed(1)}s, status=${status})${matchSuffix}` +
        (outputPreview ? ` — ${outputPreview.replace(/\n/g, " ")}` : "")
    );
    taskDispatchByCallID.delete(dispatch.toolUseCallID);
    const idx = pendingTaskDispatches.indexOf(dispatch);
    if (idx >= 0) pendingTaskDispatches.splice(idx, 1);
  }

  function buildUsage(): AgentUsage | undefined {
    const totalInput =
      accumulatedTokens.input + accumulatedTokens.cacheRead + accumulatedTokens.cacheWrite;
    return totalInput > 0 || accumulatedTokens.output > 0
      ? {
          agent: "pullfrog",
          inputTokens: totalInput,
          outputTokens: accumulatedTokens.output,
          cacheReadTokens: accumulatedTokens.cacheRead || undefined,
          cacheWriteTokens: accumulatedTokens.cacheWrite || undefined,
          costUsd: accumulatedCostUsd > 0 ? accumulatedCostUsd : undefined,
        }
      : undefined;
  }

  const handlers = {
    init: (event: OpenCodeInitEvent) => {
      // bind this sessionID to a label so subsequent events (tool_use,
      // tool_result, text, message) route to the right prefix. for the
      // first session this is "orchestrator"; for subagents it pops from
      // the pending-dispatch queue.
      const label = labeler.labelFor(event.session_id ?? null);
      log.debug(
        withLabel(
          label,
          `» ${params.label} init: session_id=${event.session_id || "unknown"}, model=${event.model || "unknown"}`
        )
      );
      log.debug(withLabel(label, `» ${params.label} init event (full): ${JSON.stringify(event)}`));
      // only reset run-wide state on the orchestrator's init — child sessions
      // emit their own init events and we don't want them to clobber the
      // parent's accumulated counters.
      if (label === ORCHESTRATOR_LABEL) {
        finalOutput = "";
        accumulatedTokens = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
        accumulatedCostUsd = 0;
        tokensLogged = false;
      } else {
        log.info(`» ${params.label} subagent init: ${label} (session ${event.session_id || "?"})`);
      }
    },
    message: (event: OpenCodeMessageEvent) => {
      const label = eventLabel(event);
      if (event.role === "assistant" && event.content?.trim()) {
        const message = event.content.trim();
        if (event.delta) {
          log.debug(
            withLabel(
              label,
              `» ${params.label} thinking: ${message.substring(0, 300)}${message.length > 300 ? "..." : ""}`
            )
          );
        } else {
          log.debug(
            withLabel(
              label,
              `» ${params.label} message (${event.role}): ${message.substring(0, 100)}${message.length > 100 ? "..." : ""}`
            )
          );
          // same reasoning as `text` handler — only orchestrator's non-delta
          // assistant message is the run output; subagent reports stay scoped
          // to the box / debug log.
          if (label === ORCHESTRATOR_LABEL) {
            finalOutput = message;
          }
        }
      } else if (event.role === "user") {
        log.debug(
          withLabel(
            label,
            `» ${params.label} message (${event.role}): ${event.content?.substring(0, 100) || ""}${event.content && event.content.length > 100 ? "..." : ""}`
          )
        );
      }
    },
    text: (event: OpenCodeTextEvent) => {
      if (event.part?.text?.trim()) {
        const message = event.part.text.trim();
        const label = eventLabel(event);
        const boxTitle = label === ORCHESTRATOR_LABEL ? params.label : `${params.label} [${label}]`;
        log.box(message, { title: boxTitle });
        // only the orchestrator's final text is the run's "output" — children
        // emit their own text on report-back, which would clobber the parent's
        // final answer if we accepted any text into finalOutput.
        if (label === ORCHESTRATOR_LABEL) {
          finalOutput = message;
        }
      }
    },
    step_start: (event: OpenCodeStepStartEvent) => {
      const stepType = event.part?.type || "unknown";
      const stepId = event.part?.id || "unknown";
      currentStepId = stepId;
      currentStepType = stepType;
      stepHistory.push({ stepId, stepType, toolCalls: [] });
    },
    step_finish: async (event: OpenCodeStepFinishEvent) => {
      const stepId = event.part?.id || "unknown";
      const eventTokens = event.part?.tokens;
      if (eventTokens) {
        accumulatedTokens.input += eventTokens.input || 0;
        accumulatedTokens.output += eventTokens.output || 0;
        accumulatedTokens.cacheRead += eventTokens.cache?.read || 0;
        accumulatedTokens.cacheWrite += eventTokens.cache?.write || 0;
      }
      // step_finish.part.cost is a per-step delta (not a running total) —
      // OpenCode emits varying per-event values that sum to the session cost.
      // verified empirically across Anthropic, OpenAI, Gemini, xAI, DeepSeek,
      // Moonshot, and OpenRouter (see pullfrog-baseline/opencode-*.log).
      // guard against NaN/Infinity — a single poison value would make the
      // running total un-recoverable for the rest of the session.
      if (typeof event.part?.cost === "number" && Number.isFinite(event.part.cost)) {
        accumulatedCostUsd += event.part.cost;
      }
      if (currentStepId === stepId) {
        currentStepId = null;
        currentStepType = null;
      }
    },
    tool_use: (event: OpenCodeToolUseEvent) => {
      const toolName = event.part?.tool;
      const toolId = event.part?.callID;
      if (!toolName || !toolId) {
        log.info(
          `» tool_use event missing toolName or toolId: ${JSON.stringify(event).substring(0, 500)}`
        );
        return;
      }

      // when the orchestrator dispatches a subagent via the `task` tool, push
      // a label for the upcoming child session so its events are attributable.
      // record BEFORE label lookup: this event's session is the parent (whose
      // label is already bound); the dispatch label is for the next new
      // sessionID that appears.
      if (toolName === "task") {
        // may have been pre-registered via the plugin's early task-dispatch
        // announcement (`pullfrog_bus_event` handler). dedupe on callID so
        // we don't record the same dispatch twice (which would corrupt the
        // FIFO label queue).
        if (!taskDispatchByCallID.has(toolId)) {
          const taskInput = (event.part?.state?.input ?? {}) as {
            description?: string;
            subagent_type?: string;
            prompt?: string;
          };
          const dispatchedLabel = labeler.recordTaskDispatch(taskInput);
          // dual-index by callID (fast path) AND in a FIFO queue (fallback path
          // for when opencode's task tool_result carries a different callID).
          const dispatch: TaskDispatch = {
            label: dispatchedLabel,
            startedAt: performance.now(),
            toolUseCallID: toolId,
          };
          taskDispatchByCallID.set(toolId, dispatch);
          pendingTaskDispatches.push(dispatch);
          log.info(
            `» dispatching subagent: ${dispatchedLabel}` +
              (taskInput.subagent_type ? ` (subagent_type=${taskInput.subagent_type})` : "")
          );
        }
      } else {
        // remember non-task callIDs so a later tool_result with that callID
        // is correctly identified as not-a-task (and we don't FIFO-pop a
        // pending task by mistake).
        knownNonTaskCallIDs.add(toolId);
      }

      const label = eventLabel(event);

      if (stepHistory.length > 0) {
        stepHistory[stepHistory.length - 1]!.toolCalls.push(toolName);
      }

      if (params.onToolUse) {
        params.onToolUse({
          toolName,
          input: event.part?.state?.input,
        });
      }

      timerFor(label).markToolCall();
      const inputFormatted = formatJsonValue(event.part?.state?.input || {});
      const toolCallLine =
        inputFormatted !== "{}" ? `» ${toolName}(${inputFormatted})` : `» ${toolName}()`;
      log.info(withLabel(label, toolCallLine));

      if (event.part?.state?.status === "completed" && event.part.state.output) {
        log.debug(withLabel(label, `  output: ${event.part.state.output}`));
      }
      // surface tool errors at info level. opencode emits tool parts at
      // status="error" through the same `tool_use` event the CLI's run-loop
      // (and our injected plugin for subagent parts) emits — without this
      // branch the only signal in the user's logs is `» <tool>(...)` with
      // no indication the call failed.
      if (event.part?.state?.status === "error") {
        log.info(withLabel(label, `» tool call failed: ${event.part.state.error}`));
      }

      // agent's explicit MCP report_progress takes priority over todo tracking
      if (toolName.includes("report_progress") && params.todoTracker) {
        log.debug("» report_progress detected, disabling todo tracking");
        params.todoTracker.cancel();
      }

      // parse todowrite events for live progress tracking
      if (toolName === "todowrite" && params.todoTracker?.enabled) {
        params.todoTracker.update(event.part?.state?.input);
      }
    },
    tool_result: (event: OpenCodeToolResultEvent) => {
      const toolId = event.part?.callID || event.tool_id;
      const state = event.part?.state;
      const status = state?.status ?? event.status ?? "unknown";
      const payload =
        state?.status === "completed"
          ? state.output
          : state?.status === "error"
            ? state.error
            : event.output;
      const label = eventLabel(event);

      timerFor(label).markToolResult();

      // surface subagent completion at info level — opencode otherwise hides
      // per-task timing in debug-only logs, so a parallel multi-lens fan-out
      // looks like N dispatches followed by a long quiet gap then a single
      // assistant turn. with this line you can see each lens finishing.
      //
      // matching is hybrid: exact callID first; FIFO fallback when the
      // tool_result's callID is unrecognised. opencode does not consistently
      // surface matching callIDs for the `task` tool, so the FIFO path is the
      // one that fires in practice. we only fall through to FIFO when the
      // callID is brand-new (not in `knownNonTaskCallIDs`) so genuinely
      // non-task tool_results never accidentally pop a pending task.
      if (taskDispatchByCallID.size > 0 || pendingTaskDispatches.length > 0) {
        if (toolId && taskDispatchByCallID.has(toolId)) {
          const dispatch = taskDispatchByCallID.get(toolId);
          if (dispatch) emitSubagentFinished(dispatch, status, payload, "exact");
        } else {
          const callIDIsKnownNonTask = toolId ? knownNonTaskCallIDs.has(toolId) : false;
          if (!callIDIsKnownNonTask && pendingTaskDispatches.length > 0) {
            const dispatch = pendingTaskDispatches[0]!;
            emitSubagentFinished(dispatch, status, payload, "fifo");
          }
        }
      }

      if (toolId) {
        const toolStartTime = toolCallTimings.get(toolId);
        if (toolStartTime) {
          const toolDuration = performance.now() - toolStartTime;
          toolCallTimings.delete(toolId);
          const stepContext = currentStepId ? ` (step=${currentStepType || "unknown"})` : "";
          log.debug(
            withLabel(
              label,
              `» ${params.label} tool_result${stepContext}: id=${toolId}, status=${status}, duration=${Math.round(toolDuration)}ms`
            )
          );
          if (payload) {
            log.debug(withLabel(label, `  output: ${payload}`));
          }
          if (toolDuration > 5000) {
            log.info(
              withLabel(
                label,
                `» tool call took ${(toolDuration / 1000).toFixed(1)}s - may indicate network latency`
              )
            );
          }
        }
      }
      if (status === "error") {
        log.info(withLabel(label, `» tool call failed: ${payload ?? "(no error message)"}`));
      } else if (payload) {
        log.debug(withLabel(label, `tool output: ${payload}`));
      }
    },
    error: (event: OpenCodeErrorEvent) => {
      // opencode emits a `type=error` event when a provider call fails (e.g.
      // 401 Invalid authentication credentials). the underlying CLI still
      // exits 0 because the error was returned cleanly by the LLM SDK, so
      // unless we capture this event the run is reported as success.
      agentErrorEvent = event;
      const errorName = event.error?.name || "unknown";
      const errorMessage = event.error?.data?.message || event.error?.name || JSON.stringify(event);
      log.info(`» ${params.label} error event: ${errorName}: ${errorMessage}`);
    },
    result: async (event: OpenCodeResultEvent) => {
      const status = event.status || "unknown";
      const duration = event.stats?.duration_ms || 0;
      const toolCalls = event.stats?.tool_calls || 0;
      log.info(
        `» ${params.label} result: status=${status}, duration=${duration}ms, tool_calls=${toolCalls}`
      );

      if (event.status === "error") {
        log.info(`» ${params.label} failed: ${JSON.stringify(event)}`);
      } else {
        // the final `result` event only carries input_tokens/output_tokens and
        // no cache breakdown — accumulatedTokens (summed across step_finish
        // events) is strictly more accurate, so we prefer it unconditionally.
        log.info(`» run complete: tool_calls=${toolCalls}, duration=${duration}ms`);

        if (
          (accumulatedTokens.input > 0 ||
            accumulatedTokens.output > 0 ||
            accumulatedTokens.cacheRead > 0 ||
            accumulatedTokens.cacheWrite > 0) &&
          !tokensLogged
        ) {
          logTokenTable({ ...accumulatedTokens, costUsd: accumulatedCostUsd });
          tokensLogged = true;
        }
      }
    },
    [PULLFROG_BUS_EVENT_TYPE]: async (event: OpenCodeBusEnvelopeEvent) => {
      // surface subagent activity that opencode's CLI run-loop discards (it
      // filters `part.sessionID !== sessionID`). our injected plugin
      // (action/agents/opencodePlugin.ts) re-emits non-orchestrator
      // `message.part.updated` bus events; here we synthesize the equivalent
      // CLI-style event for each known part type and dispatch through the
      // existing handlers so labeling, attribution, and logging all reuse the
      // same code path as the orchestrator's events. mirrors the dispatch
      // logic in opencode-ai's `cli/cmd/run.ts` `loop()` function.
      const busEvent = event.bus_event;
      if (!busEvent || busEvent.type !== "message.part.updated") return;
      const part = busEvent.properties?.part;
      if (!part || typeof part.sessionID !== "string") return;
      const sessionID = part.sessionID;
      const partType = part.type;

      // early task dispatch: the orchestrator's task tool fires bus events at
      // status=running BEFORE the subagent's first message.part.updated, but
      // the CLI's run-loop only emits the matching tool_use NDJSON event at
      // status=completed (after the subagent finishes). without
      // pre-registering the dispatch label here, the labeler binds the
      // subagent's sessionID to a generic `subagent#N` fallback before the
      // CLI's tool_use ever fires recordTaskDispatch. dedupe against
      // taskDispatchByCallID so the late tool_use handler doesn't double-add.
      if (partType === "tool") {
        const status = part.state?.status;
        const partWithToolFields = part as {
          tool?: string;
          callID?: string;
          state?: { status?: string; input?: unknown };
        };
        // only running (not pending) — at pending state.input is still {}.
        // by running, the LLM has filled in description/subagent_type/prompt.
        // mirrors the same check in the plugin source.
        const isOrchestratorTaskDispatch =
          partWithToolFields.tool === "task" && status === "running";
        if (isOrchestratorTaskDispatch) {
          const callID = partWithToolFields.callID;
          if (typeof callID === "string" && !taskDispatchByCallID.has(callID)) {
            const taskInput = (partWithToolFields.state?.input ?? {}) as {
              description?: string;
              subagent_type?: string;
              prompt?: string;
            };
            const dispatchedLabel = labeler.recordTaskDispatch(taskInput);
            const dispatch: TaskDispatch = {
              label: dispatchedLabel,
              startedAt: performance.now(),
              toolUseCallID: callID,
            };
            taskDispatchByCallID.set(callID, dispatch);
            pendingTaskDispatches.push(dispatch);
            log.info(
              `» dispatching subagent: ${dispatchedLabel}` +
                (taskInput.subagent_type ? ` (subagent_type=${taskInput.subagent_type})` : "")
            );
          }
          return;
        }
        if (status !== "completed" && status !== "error") return;
        await handlers.tool_use({
          type: "tool_use",
          sessionID,
          part,
        } as OpenCodeToolUseEvent);
        return;
      }
      // intentionally NOT routing subagent step_start / step_finish through
      // the orchestrator's handlers:
      //   - step_finish carries `tokens` and `cost` and the handler folds
      //     them into the run-wide accumulators. surfacing subagent steps
      //     here would inflate the orchestrator's usage telemetry — and
      //     either double-count (if opencode also bills child tokens back
      //     up to the parent session) or just over-report. the existing
      //     init/message/text handlers all gate on ORCHESTRATOR_LABEL for
      //     the same reason.
      //   - step_start mutates `currentStepId` / `currentStepType` /
      //     `stepHistory`, which are orchestrator-scoped — using them to
      //     attribute subagent activity in the orchestrator's tool-use
      //     timing log would be wrong.
      // the subagent's tool calls and text still surface (handled below)
      // — that's the user-visible activity.
      if (partType === "step-start" || partType === "step-finish") return;
      if (partType === "text" && part.time?.end !== undefined) {
        await handlers.text({
          type: "text",
          sessionID,
          part,
        } as OpenCodeTextEvent);
        return;
      }
    },
  };

  const recentStderr: string[] = [];

  let lastProviderError: string | null = null;
  let agentErrorEvent: OpenCodeErrorEvent | null = null;

  // shared with main.ts via toolState. updated in place as events stream and
  // stderr accumulates so the outer activity-timeout catch sees the same
  // context the harness's own catch path uses to format `result.error`.
  // recentStderr is shared by reference; the scalar fields are mirrored on
  // each update below.
  const diagnostic: AgentDiagnostic = {
    label: params.label,
    recentStderr,
    lastProviderError: undefined,
    eventCount: 0,
  };
  params.toolState.agentDiagnostic = diagnostic;

  // capped accumulator for the agent's narration. used as a post-run fallback
  // when `finalOutput` (the orchestrator's final assistant message) is empty.
  // unbounded `output += text` previously grew to ~1 GiB on multi-lens Reviews
  // and contributed to the wrapper-level RangeError. retain:"none" on spawn
  // skips the duplicate buffer there; this TailBuffer caps the agent layer.
  const output = new TailBuffer(DEFAULT_MAX_RETAINED_BYTES);
  let stdoutBuffer = "";

  try {
    const result = await spawn({
      cmd: params.cliPath,
      args: params.args,
      cwd: params.cwd,
      env: params.env,
      activityTimeout: 300_000,
      onActivityTimeout: params.onActivityTimeout,
      stdio: ["ignore", "pipe", "pipe"],
      // node_modules/opencode-ai/bin/opencode is a Node shim that spawnSyncs
      // the native opencode-<plat>-<arch> binary with stdio:"inherit". without
      // a process-group kill, SIGKILL hits only the shim, the native binary
      // is reparented to PID 1, holds our stdout pipe open, and `child.close`
      // never fires — producing zombie runs. detached + killGroup nukes the
      // whole tree.
      killGroup: true,
      // we already drain every chunk via onStdout/onStderr (NDJSON parsing
      // + recentStderr ring buffer). retaining a second copy in the spawn
      // wrapper would grow unbounded for multi-lens Reviews and previously
      // crashed the wrapper with RangeError at ~1 GiB. see issue #680.
      retain: "none",
      // NB: we used to pass `isPausedExternally: isSubagentInFlight` to suspend
      // the activity timer during subagent dispatches. unnecessary now that
      // our injected plugin (action/agents/opencodePlugin.ts) re-emits
      // subagent `message.part.updated` events on opencode's stdout — those
      // arrive at child.stdout here, fire updateActivity(), and reset
      // lastActivityTime naturally. verified empirically in PR #634
      // (~3.3 plugin events/sec during a typical subagent run).
      onStdout: async (chunk) => {
        const text = chunk.toString();
        output.append(text);
        markActivity();

        stdoutBuffer += text;
        const lines = stdoutBuffer.split("\n");
        stdoutBuffer = lines.pop() || "";

        for (const line of lines) {
          const trimmed = line.trim();
          if (!trimmed) continue;

          let event: OpenCodeEvent;
          try {
            event = JSON.parse(trimmed) as OpenCodeEvent;
          } catch {
            log.debug(`» non-JSON stdout line: ${trimmed.substring(0, 200)}`);
            continue;
          }

          eventCount++;
          diagnostic.eventCount = eventCount;
          log.debug(JSON.stringify(event, null, 2));

          const timeSinceLastActivity = getIdleMs();
          if (timeSinceLastActivity > 10000) {
            const activeToolCalls = toolCallTimings.size;
            const toolCallInfo =
              activeToolCalls > 0
                ? ` (waiting for ${activeToolCalls} tool call${activeToolCalls > 1 ? "s" : ""})`
                : ` (${params.label} may be processing internally - LLM calls, planning, etc.)`;
            log.info(
              `» no activity for ${(timeSinceLastActivity / 1000).toFixed(1)}s${toolCallInfo} (${eventCount} events processed so far)`
            );
          }
          markActivity();

          const handler = handlers[event.type as keyof typeof handlers];
          if (!handler) {
            log.info(
              `» ${params.label} event (unhandled): type=${event.type}, data=${JSON.stringify(event).substring(0, 500)}`
            );
            continue;
          }
          try {
            await handler(event as never);
          } catch (err) {
            log.info(
              `» ${params.label} handler for type=${event.type} threw: ${err instanceof Error ? err.message : String(err)}`
            );
          }
        }
      },
      onStderr: (chunk) => {
        const trimmed = chunk.trim();
        if (!trimmed) return;

        recentStderr.push(trimmed);
        if (recentStderr.length > MAX_STDERR_LINES) recentStderr.shift();

        const match = findProviderErrorMatch(trimmed);
        if (match) {
          lastProviderError = match.label;
          diagnostic.lastProviderError = match.label;
          log.info(`» provider error detected (${match.label}): ${match.excerpt}`);
        } else {
          log.debug(trimmed);
        }
      },
    });

    if (result.exitCode === 0) {
      await params.todoTracker?.flush();
    } else {
      params.todoTracker?.cancel();
    }

    // any pending task dispatches that never got a matching tool_result are
    // surfaced here so the gap is visible rather than silently swallowed.
    // this happens when opencode delivers the subagent's reply through a
    // path other than tool_result (e.g. inlined into the next assistant
    // message). flushing here is best-effort attribution — the durations
    // reported are upper bounds (the subagent could have finished any time
    // between dispatch and run-end), but the labels and ordering are exact.
    //
    // NB: the `result` event handler is dead in opencode (opencode never
    // emits a `result`-typed event), which is why this flush lives here in
    // the post-subprocess block instead.
    if (pendingTaskDispatches.length > 0) {
      for (const dispatch of [...pendingTaskDispatches]) {
        const elapsed = performance.now() - dispatch.startedAt;
        log.info(
          `» subagent finished (inferred at run-end): ${dispatch.label} (≤${(elapsed / 1000).toFixed(1)}s) — no matching tool_result observed; subagent reply likely arrived via assistant message`
        );
      }
      pendingTaskDispatches.length = 0;
      taskDispatchByCallID.clear();
    }

    const duration = performance.now() - startTime;
    log.info(
      `» ${params.label} completed in ${Math.round(duration)}ms with exit code ${result.exitCode}`
    );

    if (eventCount === 0) {
      const stderrContext = recentStderr.join("\n");
      const diagnosis = lastProviderError
        ? `provider error: ${lastProviderError}`
        : "unknown cause (no stdout events received)";
      log.info(`» ${params.label} produced 0 events (${diagnosis})`);
      if (stderrContext) log.info(`» last stderr output:\n${stderrContext}`);
    }

    if (
      !tokensLogged &&
      (accumulatedTokens.input > 0 ||
        accumulatedTokens.output > 0 ||
        accumulatedTokens.cacheRead > 0 ||
        accumulatedTokens.cacheWrite > 0)
    ) {
      logTokenTable({ ...accumulatedTokens, costUsd: accumulatedCostUsd });
      tokensLogged = true;
    }

    const usage = buildUsage();

    if (result.exitCode !== 0) {
      const errorContext = lastProviderError ? ` (${lastProviderError})` : "";
      // result.stdout / result.stderr are empty because we pass retain:"none"
      // to spawn (see issue #680); use the agent's bounded mirrors instead.
      const stdoutSnapshot = output.toString();
      const stderrSnapshot = recentStderr.join("\n");
      const errorMessage =
        stderrSnapshot ||
        stdoutSnapshot ||
        `unknown error - no output from OpenCode CLI${errorContext}`;
      log.error(
        `${params.label} exited with code ${result.exitCode}${errorContext}: ${errorMessage}`
      );
      log.debug(`stdout: ${stdoutSnapshot.substring(0, 500)}`);
      log.debug(`stderr: ${stderrSnapshot.substring(0, 500)}`);
      return {
        success: false,
        output: finalOutput || stdoutSnapshot,
        error: errorMessage,
        usage,
      };
    }

    if (eventCount === 0 && lastProviderError) {
      return {
        success: false,
        output: finalOutput || output.toString(),
        error: `provider error: ${lastProviderError}`,
        usage,
      };
    }

    if (agentErrorEvent) {
      const errorEvent: OpenCodeErrorEvent = agentErrorEvent;
      const errorName = errorEvent.error?.name || "agent error";
      const errorMessage =
        errorEvent.error?.data?.message || errorEvent.error?.name || JSON.stringify(errorEvent);
      return {
        success: false,
        output: finalOutput || output.toString(),
        error: `${errorName}: ${errorMessage}`,
        usage,
      };
    }

    return { success: true, output: finalOutput || output.toString(), usage };
  } catch (error) {
    params.todoTracker?.cancel();
    const duration = performance.now() - startTime;
    const errorMessage = error instanceof Error ? error.message : String(error);
    const isActivityTimeout =
      error instanceof SpawnTimeoutError && error.code === SPAWN_ACTIVITY_TIMEOUT_CODE;

    const stderrContext = recentStderr.slice(-10).join("\n");
    const diagnosis = lastProviderError
      ? `likely cause: ${lastProviderError}`
      : eventCount === 0
        ? "OpenCode produced 0 stdout events - check if the model provider is reachable"
        : `${eventCount} events were processed before the hang`;

    log.info(
      `» ${params.label} ${isActivityTimeout ? "hung" : "failed"} after ${(duration / 1000).toFixed(1)}s: ${errorMessage}`
    );
    log.info(`» diagnosis: ${diagnosis}`);
    if (stderrContext)
      log.info(
        `» recent stderr (last ${Math.min(recentStderr.length, 10)} lines):\n${stderrContext}`
      );

    const body = formatAgentHangBody({ diagnostic, isHang: isActivityTimeout, errorMessage });
    return {
      success: false,
      output: finalOutput || output.toString(),
      error: body ?? `${errorMessage} [${diagnosis}]`,
      usage: buildUsage(),
    };
  }
}

// ── agent ───────────────────────────────────────────────────────────────────────

export const opencode = agent({
  name: "opencode",
  install: installOpencodeCli,
  run: async (ctx) => {
    const cliPath = await installOpencodeCli();

    const rawModel = ctx.payload.proxyModel ?? ctx.resolvedModel ?? autoSelectModel(cliPath);

    // bedrock route: opencode's `amazon-bedrock` provider expects the model
    // string in `amazon-bedrock/<bedrock-id>` form. the bare AWS model ID
    // (what the user puts in `BEDROCK_MODEL_ID`) needs the prefix added.
    // detect via env-var sentinel — same pattern as claude.ts.
    //
    // we deliberately do NOT gate on `!isBedrockAnthropicId(rawModel)` here:
    // Anthropic-on-Bedrock normally routes to claude-code (per `resolveAgent`),
    // but `PULLFROG_AGENT=opencode` is the documented escape hatch for forcing
    // opencode regardless. when that override fires, opencode still needs the
    // `amazon-bedrock/` prefix or the provider lookup fails with
    // "Model not found: <modelId>/.". the Anthropic-vs-other discriminant
    // only belongs in `resolveAgent`.
    const bedrockModelId = process.env[BEDROCK_MODEL_ID_ENV]?.trim();
    const isBedrockRoute =
      rawModel !== undefined && bedrockModelId !== undefined && bedrockModelId === rawModel;
    const model = isBedrockRoute ? `amazon-bedrock/${rawModel}` : rawModel;

    const homeEnv = {
      HOME: ctx.tmpdir,
      XDG_CONFIG_HOME: join(ctx.tmpdir, ".config"),
    };

    mkdirSync(join(homeEnv.XDG_CONFIG_HOME, "opencode"), { recursive: true });

    // drop our bus-event surfacing plugin into opencode's global config dir
    // (which we've redirected to the per-run tmpdir via XDG_CONFIG_HOME).
    // opencode auto-discovers plugins from `<Global.Path.config>/{plugin,plugins}/*.{ts,js}`
    // (see `packages/opencode/src/config/config.ts:633` calling
    // `ConfigPlugin.load(dir)`), so this lands in the loader without any
    // config wiring. critically: this MUST be inside the tmpdir, never the
    // user's repo working tree — see AGENTS.md.
    const opencodePluginDir = join(homeEnv.XDG_CONFIG_HOME, "opencode", "plugin");
    mkdirSync(opencodePluginDir, { recursive: true });
    writeFileSync(
      join(opencodePluginDir, PULLFROG_OPENCODE_PLUGIN_FILENAME),
      PULLFROG_OPENCODE_PLUGIN_SOURCE
    );

    const agentBrowserVersion = getDevDependencyVersion("agent-browser");
    addSkill({
      ref: `vercel-labs/agent-browser@v${agentBrowserVersion}`,
      skill: "agent-browser",
      env: homeEnv,
      agent: "opencode",
    });

    installBundledSkills({ home: homeEnv.HOME });

    // base args shared between initial run and continue runs
    const baseArgs = ["run", "--format", "json", "--print-logs"];

    // OPENCODE_PERMISSION has absolute highest precedence (merged after managed/MDM configs).
    // external_directory gates ALL native filesystem tools (Read, Write, Edit, Glob, Grep, etc.)
    // for paths outside the project root. last-match-wins: deny everything, then allow /tmp.
    const permissionOverride = JSON.stringify({
      external_directory: { "*": "deny", "/tmp/*": "allow" },
    });

    const env: Record<string, string | undefined> = {
      ...process.env,
      ...homeEnv,
      OPENCODE_CONFIG_CONTENT: buildSecurityConfig(ctx, model),
      OPENCODE_PERMISSION: permissionOverride,
      GOOGLE_GENERATIVE_AI_API_KEY:
        process.env.GOOGLE_GENERATIVE_AI_API_KEY || process.env.GEMINI_API_KEY,
    };

    const repoDir = process.cwd();

    log.debug(`» starting Pullfrog (OpenCode): ${cliPath} ${baseArgs.join(" ")}`);
    log.debug(`» working directory: ${repoDir}`);

    const runParams = {
      label: "Pullfrog",
      cliPath,
      cwd: repoDir,
      env,
      toolState: ctx.toolState,
      todoTracker: ctx.todoTracker,
      onActivityTimeout: ctx.onActivityTimeout,
      onToolUse: ctx.onToolUse,
    };

    const result = await runOpenCode({
      ...runParams,
      args: [...baseArgs, ctx.instructions.full],
    });

    // post-run retry loop aggregates usage across the initial run + every
    // resume, so the caller sees the whole session — not just the final
    // slice. opencode always accepts `--continue`, so no canResume guard.
    // the reflection prompt fires once after gates go clean, as a dedicated
    // turn that nudges the agent to persist learnings.
    return runPostRunRetryLoop({
      ctx,
      initialResult: result,
      initialUsage: result.usage,
      reflectionPrompt:
        ctx.toolState.learningsFilePath && shouldRunReflection(ctx.toolState.selectedMode)
          ? buildLearningsReflectionPrompt(ctx.toolState.learningsFilePath)
          : undefined,
      resume: async (c) =>
        runOpenCode({
          ...runParams,
          args: [...baseArgs, "--continue", c.prompt],
        }),
    });
  },
});