← reports

#2292 · prompt history: empty-array rows persisted then skipped at read (zod too_small; 11 rows observed)

Bug Priority: Low Effort: S (suggested in issue) threads open on GitHub 2026-08-24 · base 494f66526

Verdict: REPRODUCED (live on a dev instance and with a unit test at the exact write path) · Root-cause confidence: high

1. TL;DR

bb remembers every prompt a user sends so the composer can recall it later ("prompt history", table prompt_history_entries). A thread can legitimately be started without a prompt: bb thread fork <id> with no --prompt, the SDK's threads.fork without input, or the side-chat panel's launcher fork, all clone the source provider session and leave the new thread idle. For those starts the server still writes a prompt-history row whose stored input is the JSON string []. When history is read back, the row is parsed with a schema that requires at least one item, so it fails validation, is dropped, and a WARN Skipping malformed prompt history row is logged on every read for as long as the row exists. Nothing crashes and the user sees nothing wrong (the app also drops empty drafts client-side); the defect is a wasted row plus perpetual log noise. The cause is that the single write funnel recordAcceptedPromptHistoryEntry persists args.input unconditionally, and a no-prompt fork is user-initiated and root-level (forks are not hierarchy children), so it resolves to the project scope and is written. PR #2295 adds an input.length === 0 guard at that funnel; I verified it fixes the write path in unit tests and live.

2. Claims vs findings

3. Environment

• bb source at 494f66526913557ab076e048218236f0a6610927 (main, 2026-08-24), own worktree, detached HEAD.
• macOS 26.5.2 (Darwin 25.5.0), Apple Silicon; Node v22.23.1; pnpm 9.15.0; codex-cli 0.149.1 (provider codex, model gpt-5.6-sol, permission mode accept-edits).
• Dev instance from scripts/bb-dev-app current: App http://localhost:15422, Server http://localhost:23422, Host daemon http://127.0.0.1:31422, data dir ~/.bb-dev/bb-machines-HOST.getbb.app-checkouts-bb-.claude-worktrees-wf_846839f8-f8a-63-d4073e0642a3 (deleted after the investigation). Host id host_6ntxdnapd7, project proj_ba66tywb3e on scratch repo /tmp/bb-2292-qa.

4. Minimal reproduction

4a. Live (CLI + sqlite3), ~1 minute, one tiny codex turn

1. Start an isolated dev instance from a checkout of 494f66526 and create a project on it.

   pnpm install --frozen-lockfile --prefer-offline && pnpm exec turbo run build
   scripts/bb-dev-app current            # prints App/Server/Host daemon URLs and the data dir
   mkdir -p /tmp/bb-2292-qa && cd /tmp/bb-2292-qa && git init -q -b main . && echo "# qa" > README.md && git add . && git commit -qm init
   curl -s http://localhost:23422/api/v1/hosts          # copy the host id
   curl -s -X POST http://localhost:23422/api/v1/projects -H 'content-type: application/json' \
     -d '{"name":"qa","source":{"type":"local_path","path":"/tmp/bb-2292-qa","hostId":"host_6ntxdnapd7"}}'
   # → {"id":"proj_ba66tywb3e", ...}

2. Spawn a source thread with a provider that supports fork (codex, claude-code, pi) and wait until it is idle.

   BB_SERVER_URL=http://localhost:23422 BB_HOST_DAEMON_PORT=31422 BB_PROJECT_ID=proj_ba66tywb3e \
     pnpm bb:dev thread spawn --project proj_ba66tywb3e --provider codex --permission-mode accept-edits \
     --title "2292 source" --prompt "Reply only with ok." --json
   # → "id": "thr_kqdsa8qm7t"
   curl -s http://localhost:23422/api/v1/threads/thr_kqdsa8qm7t | python3 -c 'import sys,json; print(json.load(sys.stdin)["status"])'
   # → idle

3. Look at the table before forking: one valid row for the source prompt.

   DB=<data dir>/bb.db
   sqlite3 -header "$DB" "select id, scope, thread_id, request_sequence, input from prompt_history_entries order by created_at"
   id|scope|thread_id|request_sequence|input
   phist_jxd9x97ahk|project|thr_kqdsa8qm7t|1|[{"type":"text","text":"Reply only with ok.","mentions":[]}]

4. Fork the thread without a prompt (an "idle fork", exactly what the CLI help offers: --prompt "Optional first prompt; omit for an idle fork").

   BB_SERVER_URL=http://localhost:23422 BB_HOST_DAEMON_PORT=31422 BB_PROJECT_ID=proj_ba66tywb3e \
     node packages/scripts/dist/commands/run-cli.js thread fork thr_kqdsa8qm7t --workspace reuse --title "2292 idle fork" --json
   # → "id": "thr_hz6xuhkvky", "parentThreadId": null, "sourceThreadId": "thr_kqdsa8qm7t", "originKind": "fork"

5. Look at the table again.

   sqlite3 -header "$DB" "select id, scope, thread_id, request_sequence, input from prompt_history_entries order by created_at"
   
   expected: only the one valid row from step 3 (there was no prompt to remember)
   actual:
   id|scope|thread_id|request_sequence|input
   phist_jxd9x97ahk|project|thr_kqdsa8qm7t|1|[{"type":"text","text":"Reply only with ok.","mentions":[]}]
   phist_deg4hty4uq|project|thr_hz6xuhkvky|6|[]

6. Read project prompt history through the API, then look at the server log.

   curl -s http://localhost:23422/api/v1/projects/proj_ba66tywb3e/prompt-history
   [{"id":"phist_jxd9x97ahk","createdAt":1787597928194,"input":[{"type":"text","text":"Reply only with ok.","mentions":[]}]}]
   
   grep -n "Skipping malformed prompt history row" ~/.bb-dev/launchers/<instance>/dev.log
   111:@bb/server:dev: [11:59:23] WARN: [server] Skipping malformed prompt history row {"entryId":"phist_deg4hty4uq","requestSequence":6,"threadId":"thr_hz6xuhkvky","errorMessage":"[\n  {\n    \"origin\": \"array\",\n    \"code\": \"too_small\",\n    \"minimum\": 1,\n    \"inclusive\": true,\n    \"path\": [],\n    \"message\": \"Too small: expected array to have >=1 items\"\n  }\n]","errorName":"ZodError"}

   The API silently omits the [] row and the server warns. Every subsequent read of this project's history (the composer fetches it on open) repeats the warn.

The fork thread's own client/turn/requested event (seq 6, the requestSequence stored on the bad row) carries "initiator": "user", "input": [], "target": {"kind": "thread-start"} — see 2292/repro/fork-event-seq6.json.

4b. Unit test at the exact write path (fails on base)

Save as apps/server/test/services/prompt-history-2292.test.ts and run cd apps/server && pnpm exec vitest run test/services/prompt-history-2292.test.ts. On 494f66526 the assertion at line 86 fails because the stored inputs contain "[]"; with PR #2295 applied it passes.

// Repro for get-bb/bb#2292: a user-initiated thread-start with empty input
// (an idle fork / side-chat tip fork) persists a prompt_history_entries row
// whose input is "[]". The read path rejects that row with zod
// (`z.array(promptInputSchema).min(1)`) and logs a warn on every read.
//
// On base 494f66526 the FIRST assertion fails: a row with input "[]" exists.
import { describe, expect, it, vi } from "vitest";
import {
  createConnection,
  createProject,
  createThread,
  migrate,
  noopNotifier,
  promptHistoryEntries,
  upsertHost,
} from "@bb/db";
import {
  listProjectPromptHistory,
  recordAcceptedPromptHistoryEntry,
} from "../../src/services/prompt-history.js";
import { textInput } from "../helpers/prompt-input.js";

describe("#2292 empty-input prompt history rows", () => {
  it("does not persist a row for a user-initiated empty-input thread start", () => {
    const db = createConnection(":memory:");
    migrate(db);
    const host = upsertHost(db, noopNotifier, {
      name: "test-host",
      type: "persistent",
    });
    const project = createProject(db, noopNotifier, {
      name: "Project A",
      source: { type: "local_path", hostId: host.id, path: "/tmp/project-a" },
    }).project;
    const logger = {
      debug: vi.fn(),
      error: vi.fn(),
      info: vi.fn(),
      warn: vi.fn(),
    };

    // A forked thread is NOT a hierarchy child (parentThreadId stays null),
    // so the scope resolver picks "project" for a user-initiated start.
    const fork = createThread(db, noopNotifier, {
      projectId: project.id,
      providerId: "codex",
      originKind: "fork",
    });
    expect(fork.parentThreadId).toBeNull();

    // A valid entry that should still be recalled afterwards.
    const normal = createThread(db, noopNotifier, {
      projectId: project.id,
      providerId: "codex",
    });
    recordAcceptedPromptHistoryEntry(
      { db },
      {
        thread: normal,
        input: textInput("Investigate auth flow"),
        initiator: "user",
        target: { kind: "thread-start" },
        requestSequence: 1,
      },
    );

    // Exactly what requestThreadProvision passes for `bb thread fork <id>`
    // without --prompt, or a side-chat launcher fork with no reply anchor.
    recordAcceptedPromptHistoryEntry(
      { db },
      {
        thread: fork,
        input: [],
        initiator: "user",
        target: { kind: "thread-start" },
        requestSequence: 1,
      },
    );

    const storedInputs = db
      .select({ input: promptHistoryEntries.input })
      .from(promptHistoryEntries)
      .all()
      .map((row) => row.input);
    // FAILS on 494f66526: storedInputs contains "[]".
    expect(storedInputs).not.toContain("[]");

    // Read path: the "[]" row is dropped with a warn on every read.
    const history = listProjectPromptHistory(
      { db, logger },
      { projectId: project.id, limit: 50 },
    );
    expect(history.map((entry) => entry.input)).toEqual([
      textInput("Investigate auth flow"),
    ]);
    expect(logger.warn).not.toHaveBeenCalled();
  });
});

$ cd apps/server && pnpm exec vitest run test/services/prompt-history-2292.test.ts     # at 494f66526
 FAIL  @bb/server  test/services/prompt-history-2292.test.ts > #2292 empty-input prompt history rows > does not persist a row for a user-initiated empty-input thread start
AssertionError: expected [ …(2) ] to not include '[]'
 ❯ test/services/prompt-history-2292.test.ts:86:30
 Test Files  1 failed (1)
      Tests  1 failed (1)

Repro files: 2292/repro/ — prompt-history-2292.test.ts, unit-test-base-494f66526.clean.log, db-after-fork.txt, api-project-prompt-history.json, server-log-warn.txt, fork-event-seq6.json, cli-fork-output.json, pr-2295.diff, unit-test-with-pr-2295.log, cli-fork-with-pr-2295.txt, typecheck-with-pr-2295.log. No screenshots: the bug has no visual surface (the app hides empty drafts; the only observable artefacts are the DB row and the log line).

5. Root cause

Mechanism, end to end:

1. A fork request may carry no prompt. forkThreadRequestSchema.input is .min(1).optional() (threads.ts#L177); createThreadForkFromRequest defaults it to [] (thread-fork.ts#L109-L115). It only marks the start as agent-initiated (startedOnBehalfOf) when there is an agentContextSeed and no visible input (isSeedOnlyIdleFork, #L144-L146). A fork with neither stays initiator: "user". createThreadRequestSchema accepts the empty array because originKind !== null (#L146-L152).
2. Forks are not hierarchy children. createThreadFromRequest sets hierarchyParentThreadId = originKind === null ? requestInput.parentThreadId : undefined (thread-create.ts#L612-L617; comment at #L644: "Forks and side chats are not hierarchy children"). The new thread therefore has parentThreadId: null — confirmed in the live output.
3. requestThreadProvision appends the client/turn/requested event and then unconditionally calls the history funnel with the same input (thread-provisioning.ts#L254-L260).
4. recordAcceptedPromptHistoryEntry decides the scope purely from initiator, target.kind and parentThreadId (prompt-history.ts#L159-L175): user + thread-start + root → "project". It never looks at the input, and persists it as-is (#L249-L270); createPromptHistoryEntry JSON-stringifies [] (db/prompt-history.ts#L51-L75).
5. Reads parse every row with z.array(promptInputSchema).min(1) (#L24) inside buildPromptHistoryEntries, whose catch block drops the row and logs the warn (#L134-L157). The row is never cleaned up, so the warn repeats on every read.

Why this is the real cause and not a read-side bug: the write and read sides disagree on the invariant. The read side encodes "a remembered prompt has at least one item" (which is the right invariant: there is nothing to recall from an empty prompt, and the runtime never even runs a turn for it — thread-lifecycle.ts#L731-L736). The write side treats "a user-initiated thread-start request was accepted" as sufficient, which became false when empty-input forks were introduced. Per AGENTS.md ("validate at the boundary, then pass typed values"), the row should never be written.

Which real surfaces trigger it: bb thread fork <id> without --prompt/--file/--image/--agent-context-seed (fork.ts#L72); SDK threads.fork({sourceThreadId}); the side-chat launcher fork ("no anchor, no seed", side-chat/app.tsx#L345) and any side chat whose anchor equals the source's last message (resolveReplySeedText returns null → no seed, side-chat/server.ts#L201-L213). Side chats that do get a seed start as initiator: "agent" and never write a row. The web app's "fork from message" flow navigates to the composer and submits a real prompt, so it is not affected.

Deeper observation (not this bug): the same funnel records the full input, so a fork that carries both an agentContextSeed and a visible prompt stores the agent-only seed item in project prompt history. promptInputToDraft does not appear to filter on visibility (only threadQueuedMessages.ts does). I did not drive that path; it is a candidate for a separate issue if recall of agent-only text is undesirable.

6. Proposed fix (first principles)

Reject empty input at the single write funnel — exactly what PR #2295 does — so every call site (provisioning, reprovisioning, send, queued flush) is covered and the stored invariant matches the read schema:

// apps/server/src/services/prompt-history.ts, recordAcceptedPromptHistoryEntry
if (args.input.length === 0) {
  return false; // no prompt was sent, so there is nothing to recall
}

Keep the read-side skip as defense in depth for rows that already exist. Optionally add a one-off Drizzle migration DELETE FROM prompt_history_entries WHERE input = '[]' so users who already have bad rows (the issue's 11) stop getting a warn on every history read; the rows carry no information, so deleting them cannot lose anything. What could go wrong: nothing observable — the return value is ignored by all callers (grep confirms no caller reads it), new-turn targets already require min(1) input at the API boundary, and the guard does not alter scope resolution. No wire shape changes, so HOST_DAEMON_PROTOCOL_VERSION does not need a bump.

7. PR review

PR #2295 — "fix(server): never persist empty-input prompt history rows" (bradhallett, cross-repo fork, head 32ce011bf)

What it changes: +11 lines in apps/server/src/services/prompt-history.ts adding the args.input.length === 0 → return false guard before scope resolution, and +36 lines in apps/server/test/services/prompt-history.test.ts ("does not persist empty-input turns as prompt history rows") asserting the funnel returns false, the table stays empty, listProjectPromptHistory returns [], and no warn is logged. No other files. Diff saved at 2292/repro/pr-2295.diff.

Root cause or symptom? Root cause. The guard sits at the only production writer, so no path can persist an empty array, and the stored invariant now matches the read schema. Layer is correct (server owns prompt-history policy; nothing crosses the daemon boundary). No casts, no unknown, no accepted-but-ignored fields, no wire change.

What I ran (PR diff applied onto 494f66526 with git apply; applies cleanly):

• pnpm exec vitest run test/services/prompt-history.test.ts test/services/prompt-history-2292.test.ts → 2 files, 9 tests passed, including the PR's new test and my repro (which fails on base). Log: unit-test-with-pr-2295.log.
• pnpm exec turbo run typecheck --filter=@bb/server → green. Log: typecheck-with-pr-2295.log.
• Live: restarted my dev instance from the patched tree and repeated step 4 of the repro (bb thread fork thr_kqdsa8qm7t, no prompt). New thread thr_4cbfetdhjw was created, landed idle, and no prompt_history_entries row was written (table still holds exactly the original valid row plus the pre-fix [] row). Output: cli-fork-with-pr-2295.txt.
• Mutation check: with the PR's test file applied but the source guard reverted to base, the PR's test fails at prompt-history.test.ts:512 (AssertionError: expected true to be false — the funnel returned true and wrote the row); my repro fails at its not.toContain("[]") assertion. Both pass with the guard restored.

Findings:

Verdict: MERGE (after CI is approved and green). The change is minimal, at the right layer, addresses the root cause, and is verified by a test that fails without it. The only substantive gap is the missing one-off cleanup of already-persisted rows, which can be a follow-up.

8. Related issues

• #2295 — the fix PR reviewed above.
• #2190 (open) — SQLite db silently reverted to a 2-week-old state; the issue explicitly rules it out as a duplicate. Agreed: unrelated.
• #1321 (merged PR) — host-daemon undeliverable-event fix; unrelated.
• #2052 (open) — CLI human output of thread history/project history is raw JSON; touches the same history surfaces but a different defect.
• Candidate follow-up (not filed): agent-only agentContextSeed items are stored in prompt history for forks that also carry a visible prompt (section 5, "deeper observation").

9. Appendix

Commands run (in order)

gh issue view 2292 --json number,title,body,labels,state,createdAt,author,comments
gh pr view 2295 && gh pr diff 2295
pnpm install --frozen-lockfile --prefer-offline; pnpm exec turbo run build        # 18/18 tasks, cached
git checkout 494f66526; git fetch origin main
git log --oneline 494f66526..origin/main -- apps/server/src/services/prompt-history.ts apps/server/src/routes apps/server/src/services/threads packages/db/src packages/domain/src   # empty
grep -rn recordAcceptedPromptHistoryEntry apps/server/src                          # 4 call sites
scripts/bb-dev-app current; scripts/bb-dev-app env
curl -s http://localhost:23422/api/v1/hosts
curl -s -X POST http://localhost:23422/api/v1/projects ... (qa project on /tmp/bb-2292-qa)
pnpm bb:dev thread spawn --project proj_ba66tywb3e --provider codex --permission-mode accept-edits --title "2292 source" --prompt "Reply only with ok." --json
cd apps/server && pnpm exec vitest run test/services/prompt-history-2292.test.ts   # FAILS on base (expected)
sqlite3 -header $DB "select ... from prompt_history_entries"                      # before fork: 1 row
node packages/scripts/dist/commands/run-cli.js thread fork thr_kqdsa8qm7t --workspace reuse --title "2292 idle fork" --json
sqlite3 -header $DB "select ... from prompt_history_entries"                      # after fork: 2 rows, second is []
curl -s http://localhost:23422/api/v1/projects/proj_ba66tywb3e/prompt-history     # 1 entry returned
curl -s http://localhost:23422/api/v1/threads/thr_hz6xuhkvky/events?limit=50      # seq 6: initiator user, input []
grep -n "Skipping malformed prompt history row" ~/.bb-dev/launchers/<instance>/dev.log
gh pr diff 2295 > pr-2295.diff; git apply --check pr-2295.diff; git apply pr-2295.diff
cd apps/server && pnpm exec vitest run test/services/prompt-history.test.ts test/services/prompt-history-2292.test.ts   # 9 passed
pnpm exec turbo run typecheck --filter=@bb/server                                 # green
scripts/bb-dev-app current                                                        # restart on patched tree
node packages/scripts/dist/commands/run-cli.js thread fork thr_kqdsa8qm7t --workspace reuse --title "2292 idle fork with PR" --json
sqlite3 -header $DB "select ... from prompt_history_entries"                      # still 2 rows: no new [] row
git checkout -- apps/server/src/services/prompt-history.ts apps/server/test/services/prompt-history.test.ts; rm apps/server/test/services/prompt-history-2292.test.ts
pnpm dev:stop; pkill -f wf_846839f8-f8a-63; rm -rf ~/.bb-dev/...-wf_846839f8-f8a-63-d4073e0642a3 /tmp/bb-2292-qa; lsof -nP -iTCP -sTCP:LISTEN | grep -E ":15422|:23422|:31422"   # ports free

Thread rows on the dev instance after both forks

id|parent_thread_id|source_thread_id|origin_kind|status
thr_kqdsa8qm7t||||idle
thr_hz6xuhkvky||thr_kqdsa8qm7t|fork|idle          # created at base: wrote the [] row
thr_4cbfetdhjw||thr_kqdsa8qm7t|fork|starting→idle # created with PR #2295 applied: no row

Fork thread event types (base run)

1 client/turn/requested   (inherited from source)
2 turn/started
3 turn/input/accepted
4 item/completed
5 turn/completed
6 client/turn/requested   initiator=user, input=[], target=thread-start   ← requestSequence of the bad row
7 client/thread/start
8 thread/identity
9 thread/contextWindowUsage/updated
10 thread/started
11 thread/identity
12 thread/name/updated
13 system/thread-provisioning

Scope resolver (base)

function resolveAcceptedPromptHistoryScope(args) {
  if (args.initiator !== "user") return null;
  if (args.target.kind !== "thread-start") return "thread";
  if (args.thread.parentThreadId !== null) return null;   // forks have parentThreadId === null → falls through
  return "project";
}

Other logs: dev instance start, restart on patched tree, turbo build.