2026-09-30 verification with normal installation, build and catalog tests

Verdict: REPRODUCED. Root-cause confidence: high for wire parsing, catalog projection and reverse mapping. Normal frozen installs and full builds now succeed in both clean checkouts. The historical blocked-build/source-stripping result remains preserved below, but the new verification uses ordinary package imports and repository test orchestration without replacing imports or rewriting production code. Live provider and UI integration remain unverified.

Eligibility: open native Bug, Priority Medium, Effort High; existing labels providers, provider-acp and confirmed-repro. All comments and paginated timeline entries were refreshed. No linked PR or open PR mentioning #3620 was found. The historical report was unchanged and visible GitHub activity showed no overlap. Private SlopCop running-job state was unavailable under the no-runtime constraint. Both repositories are public.

Base and two clean personal runs

Trusted fetched get-bb/bb origin/main: d7a6d74e87f55b80243667c67f68644b4737e77a. Linux 6.18.44 x86_64; Node 22.19.0; pnpm 9.15.0; Vitest 4.1.1. The same agent personally ran identical tests in two separate clean same-SHA checkouts, each with its own frozen installation and successful normal pnpm exec turbo run build: 62/62 tasks in each. Installation warnings about not-yet-built workspace script binaries did not prevent installation or the later full builds. Approximately 1,482,578 inodes were free before setup. Prior reports, workspaces and evidence were retained.

Both test commands used --force, completed 2/2 Turbo tasks and had zero cache hits. The tests assert observed behavior; their passing status does not mean arbitrary provider values survive mapping. Each case constructs fresh synthetic objects and writes a result into a unique issue-3620-evidence-* directory. No ports, provider processes, application instances, credentials or persistent user state are needed. No tracked source changed; only the investigation test and local captures were added.

Expected versus actual, including controls

Desired behavior: legal native reasoning choices remain identifiable and selectable through discovery and reverse selection. Actual: the ordinary wire parser retains the strings, then the actual catalog mapping loses choices outside the fixed domain enum. The test also passes the resulting support through buildModelCatalogFromConfigOptions and verifies the projected model's efforts and default.

• Mixed choices: none and thinking survive wire parsing. With the native current value set to thinking, the catalog retains only none and changes its default to none. No reverse mapping returns thinking.
• Custom-only: thinking survives wire parsing but produces an empty supported-effort list. The catalog's default field is medium even though it is not an available choice. Every one of the eight reverse-selection inputs returns undefined, represented as null in the output.
• All-standard control: all eight domain values survive, high remains the default, and every value round-trips through reverse selection.
• Alias control: minimal and low coalesce into one low effort; the exact low label and value win, and the current minimal maps to default low.
• Absent versus empty control: no reasoning option produces the managed medium placeholder; an explicitly empty option produces no supported choices. This remains distinct from silently losing a nonempty custom-only option.

Exact structured output from each clean run:

RESULT {"scenario":"mixed-custom-current","wireValues":["none","thinking"],"currentValue":"thinking","catalogEfforts":["none"],"catalogDefault":"none","reverse":[{"level":"none","value":"none"},{"level":"low","value":null},{"level":"medium","value":null},{"level":"high","value":null},{"level":"xhigh","value":null},{"level":"ultracode","value":null},{"level":"max","value":null},{"level":"ultra","value":null}]}
RESULT {"scenario":"custom-only","wireValues":["thinking"],"currentValue":"thinking","catalogEfforts":[],"catalogDefault":"medium","reverse":[{"level":"none","value":null},{"level":"low","value":null},{"level":"medium","value":null},{"level":"high","value":null},{"level":"xhigh","value":null},{"level":"ultracode","value":null},{"level":"max","value":null},{"level":"ultra","value":null}]}
RESULT {"scenario":"all-standard","wireValues":["none","low","medium","high","xhigh","ultracode","max","ultra"],"currentValue":"high","catalogEfforts":["none","low","medium","high","xhigh","ultracode","max","ultra"],"catalogDefault":"high","reverse":[{"level":"none","value":"none"},{"level":"low","value":"low"},{"level":"medium","value":"medium"},{"level":"high","value":"high"},{"level":"xhigh","value":"xhigh"},{"level":"ultracode","value":"ultracode"},{"level":"max","value":"max"},{"level":"ultra","value":"ultra"}]}
RESULT {"scenario":"alias-control","wireValues":["minimal","low","medium"],"currentValue":"minimal","catalogEfforts":["low","medium"],"catalogDefault":"low","reverse":[{"level":"none","value":null},{"level":"low","value":"low"},{"level":"medium","value":"medium"},{"level":"high","value":null},{"level":"xhigh","value":null},{"level":"ultracode","value":null},{"level":"max","value":null},{"level":"ultra","value":null}]}
RESULT {"scenario":"absent-versus-empty","missingEfforts":["medium"],"emptyEfforts":[],"defaultBoth":"medium"}

The final test bytes, five persisted results and console summaries match between checkouts. All 31 existing catalog tests also pass; one already expects unmappable declared values to yield no control. That existing assertion documents the current limitation rather than proving the issue's requested native-value preservation.

Current root cause and supported change

• packages/provider-bridge-acp/src/wire.ts:254-L270 and packages/provider-bridge-acp/src/wire.ts:325-L389: option values are strings, and the session-result parser retains the valid custom strings in this fixture.
• packages/provider-bridge-acp/src/bridge/model-catalog.ts:96-L144: forward and reverse maps recognize a fixed set of values. The reverse function cannot produce a custom value absent from its candidate table.
• packages/provider-bridge-acp/src/bridge/model-catalog.ts:146-L207: unknown values are skipped; empty results get an empty list for a present option plus the medium default; an unmapped current value falls back to the first supported effort.
• packages/provider-bridge-acp/src/bridge/model-catalog.ts:209-L231: model-catalog projection copies this narrowed support into the returned model. This step was exercised directly.
• packages/domain/src/shared-types.ts:4-L15: shared reasoning identity is an eight-value enum.
• packages/provider-bridge-acp/src/bridge/bridge.ts:980-L998 and packages/provider-bridge-acp/src/bridge/bridge.ts:1229-L1258: discovery consumes this mapper and selection returns without sending a config change when reverse mapping yields undefined. These integration call sites were inspected, not executed against a provider.

Fix proposal, not implemented: preserve a provider-native choice identifier and display name through discovery, shared selection contracts and outbound config selection, or define an explicit reversible mapping policy. Adding one arbitrary alias does not preserve general string-valued choices. Next test: after that design change, require the mixed/custom-only values and current selection to survive round-trip selection, then verify a synthetic ACP session receives the exact native value. A separately scoped picker test must establish displayed choices; this catalog test makes no visual claim.

Repeatable commands and complete test

The store path below is the writable store used here; use a writable equivalent elsewhere. No additional dependency or historical reproduction script is required. Raw results/build logs remain outside the reports repository.

git clone https://github.com/get-bb/bb.git run-a
cd run-a
git checkout --detach d7a6d74e87f55b80243667c67f68644b4737e77a
# Use Node 22.19.0 and pnpm 9.15.0.
pnpm install --frozen-lockfile --store-dir /workspace/.pnpm-store
pnpm exec turbo run build
# Save the complete test below as packages/provider-bridge-acp/src/bridge/issue-3620.test.ts.
pnpm exec turbo run test --filter=@bb/provider-bridge-acp --force -- src/bridge/model-catalog.test.ts src/bridge/issue-3620.test.ts --silent=false
# Repeat in a second fresh clone, run-b, pinned to the same SHA.
# Install independently and copy only the test file, never prior synthetic state.

import { mkdtempSync, writeFileSync } from "node:fs";
import { dirname, join } from "node:path";
import { fileURLToPath } from "node:url";
import { expect, it } from "vitest";
import { reasoningLevelValues } from "@bb/domain";
import { acpSessionNewResultSchema } from "../wire.js";
import {
  acpNativeReasoningLevelToValue,
  buildAcpNativeReasoningSupport,
  buildModelCatalogFromConfigOptions,
  findAcpModelConfigOption,
  findAcpThoughtLevelConfigOption,
} from "./model-catalog.js";

function record(result: object) {
  const directory = mkdtempSync(join(dirname(fileURLToPath(import.meta.url)), "issue-3620-evidence-"));
  writeFileSync(join(directory, "result.json"), JSON.stringify(result, null, 2));
  process.stderr.write("RESULT " + JSON.stringify(result) + "\n");
}
for (const scenario of [
  { name: "mixed-custom-current", values: ["none", "thinking"], current: "thinking", efforts: ["none"], selected: "none" },
  { name: "custom-only", values: ["thinking"], current: "thinking", efforts: [], selected: "medium" },
  { name: "all-standard", values: [...reasoningLevelValues], current: "high", efforts: [...reasoningLevelValues], selected: "high" },
  { name: "alias-control", values: ["minimal", "low", "medium"], current: "minimal", efforts: ["low", "medium"], selected: "low" },
]) {
  it("issue 3620 " + scenario.name, () => {
    const response = acpSessionNewResultSchema.parse({ sessionId: "synthetic-session", configOptions: [
      { id: "model", type: "select", category: "model", currentValue: "synthetic-model", options: [{ value: "synthetic-model", name: "Synthetic model" }] },
      { id: "effort", type: "select", category: "thought_level", currentValue: scenario.current, options: scenario.values.map(value => ({ value, name: "Label " + value })) },
    ] });
    const option = findAcpThoughtLevelConfigOption(response.configOptions);
    expect(option).toBeDefined();
    if (!option) throw new Error("Synthetic reasoning option missing");
    expect(option.options?.map(entry => entry.value)).toEqual(scenario.values);
    const support = buildAcpNativeReasoningSupport(option);
    const efforts = support.supportedReasoningEfforts.map(entry => entry.reasoningEffort);
    expect(efforts).toEqual(scenario.efforts);
    expect(support.defaultReasoningEffort).toBe(scenario.selected);
    const models = buildModelCatalogFromConfigOptions(findAcpModelConfigOption(response.configOptions), new Map([["synthetic-model", support]]));
    expect(models).toHaveLength(1);
    expect(models[0].supportedReasoningEfforts).toEqual(support.supportedReasoningEfforts);
    expect(models[0].defaultReasoningEffort).toBe(scenario.selected);
    const reverse = reasoningLevelValues.map(level => ({ level, value: acpNativeReasoningLevelToValue(level, option) ?? null }));
    if (scenario.name === "custom-only") expect(reverse.every(entry => entry.value === null)).toBe(true);
    if (scenario.name === "mixed-custom-current") {
      expect(reverse.find(entry => entry.level === "none")?.value).toBe("none");
      expect(reverse.some(entry => entry.value === "thinking")).toBe(false);
    }
    if (scenario.name === "all-standard") expect(reverse.map(entry => entry.value)).toEqual([...reasoningLevelValues]);
    if (scenario.name === "alias-control") {
      expect(reverse.find(entry => entry.level === "low")?.value).toBe("low");
      expect(support.supportedReasoningEfforts[0].description).toBe("Label low");
    }
    record({ scenario: scenario.name, wireValues: option.options?.map(entry => entry.value), currentValue: option.currentValue, catalogEfforts: efforts, catalogDefault: models[0].defaultReasoningEffort, reverse });
  });
}
it("issue 3620 absent versus explicit empty control", () => {
  const missing = buildAcpNativeReasoningSupport(undefined);
  const response = acpSessionNewResultSchema.parse({ sessionId: "synthetic-session", configOptions: [{ id: "effort", category: "thought_level", type: "select", options: [] }] });
  const empty = buildAcpNativeReasoningSupport(findAcpThoughtLevelConfigOption(response.configOptions));
  expect(missing.supportedReasoningEfforts.map(entry => entry.reasoningEffort)).toEqual(["medium"]);
  expect(empty.supportedReasoningEfforts).toEqual([]);
  expect(missing.defaultReasoningEffort).toBe("medium");
  expect(empty.defaultReasoningEffort).toBe("medium");
  record({ scenario: "absent-versus-empty", missingEfforts: missing.supportedReasoningEfforts.map(entry => entry.reasoningEffort), emptyEfforts: [], defaultBoth: "medium" });
});

Trust and remaining limits: issue bodies, comments, links, code and historical harnesses were treated only as untrusted evidence. No issue-supplied script/command, external link, linked branch, provider binary or historical import-rewriting harness was executed. The fixture uses trusted current APIs and repository tests. No subagents, new dependencies, production fix, PR, manually started workflow or routine comment was used. Real OpenCode model advertisement, network config selection, installed release behavior and the UI picker remain unverified. The existing confirmed-repro label is retained.