@lucascouts/claude-agent-acp-plus 0.12.0 → 0.14.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/acp-agent.js CHANGED
@@ -14,6 +14,7 @@ import { applyAskElicitationResponse, askUserQuestionsToCreateRequest, createEli
14
14
  import { agentName } from "./agent-name.js";
15
15
  import { filterDeprecatedModels } from "./model-deprecation.js";
16
16
  import { SettingsManager } from "./settings.js";
17
+ import { ContextCompactionLifecycle, contextCompactionMetadataFromBoundary, } from "./context-compaction.js";
17
18
  import { createThinkingConfigOption, effectiveThinkingConfig, resolveThinkingSelection, THINKING_CONFIG_ID, } from "./thinking-option.js";
18
19
  import { effortOptionValue, isUltracodeAvailable, isUltracodeValue, ULTRACODE_OPTION_NAME, ULTRACODE_OPTION_VALUE, ultracodeFlagSettings, } from "./ultracode.js";
19
20
  import { handleRewindCommand, parseRewindInvocation } from "./rewind-command.js";
@@ -30,6 +31,8 @@ import { acceptedPlanToolResult, ExitPlanCoordinator, executionDiagnostic, exitP
30
31
  import { DEFAULT_AGENT_ID, EFFORT_CONFIG_ID } from "./session-config-ids.js";
31
32
  import { parseToolResultMeta } from "./tool-result-meta.js";
32
33
  import { AccountUsageTracker } from "./account-usage.js";
34
+ import { ACCOUNT_CONFIG_ID, accountForValue, accountInForce, createAccountConfigOption, envForAccount, readDeclaredAccounts, } from "./accounts.js";
35
+ import { isUsageCommandText, structuredUsageMarkdown } from "./usage-markdown.js";
33
36
  export { DEFAULT_AGENT_ID, EFFORT_CONFIG_ID } from "./session-config-ids.js";
34
37
  import { MODE_CONFIG_ID, SessionModeManager } from "./session-mode.js";
35
38
  export const CLAUDE_CONFIG_DIR = process.env.CLAUDE_CONFIG_DIR ?? path.join(os.homedir(), ".claude");
@@ -112,6 +115,37 @@ function parseSteerRequest(params) {
112
115
  _meta: _meta,
113
116
  };
114
117
  }
118
+ /** How often the account quota windows are refreshed while a session sits idle,
119
+ * in milliseconds; `0` disables polling entirely.
120
+ *
121
+ * Without this the windows move at only three moments: session establishment,
122
+ * the end of every turn, and a pushed `rate_limit_event`. None of them is time
123
+ * — so a five-hour window that resets while nobody is prompting keeps showing
124
+ * the old utilization, and the bar is wrong in the one direction that matters
125
+ * (it under-reports how much is available) until the next turn ends.
126
+ *
127
+ * The floor is not timidity: each tick costs one control request, and neither
128
+ * the five-hour nor the seven-day window can move fast enough for a tighter
129
+ * loop to show anything a 15-second one would miss. */
130
+ const QUOTA_POLL_DEFAULT_MS = 60_000;
131
+ const QUOTA_POLL_FLOOR_MS = 15_000;
132
+ /** Resolve the poll interval from the environment, clamped. An unparseable or
133
+ * negative value falls back to the default rather than disabling the refresh:
134
+ * a typo should not silently return the bars to the stale behaviour this
135
+ * exists to fix. Exactly `0` disables, because that is the only way to ask. */
136
+ export function quotaPollIntervalMs(raw = process.env.CLAUDE_ACP_QUOTA_POLL_MS) {
137
+ if (raw === undefined || raw.trim() === "") {
138
+ return QUOTA_POLL_DEFAULT_MS;
139
+ }
140
+ const parsed = Number(raw);
141
+ if (!Number.isFinite(parsed) || parsed < 0) {
142
+ return QUOTA_POLL_DEFAULT_MS;
143
+ }
144
+ if (parsed === 0) {
145
+ return 0;
146
+ }
147
+ return Math.max(parsed, QUOTA_POLL_FLOOR_MS);
148
+ }
115
149
  /** Result-message origin kinds that mark an AUTONOMOUS cycle — work the
116
150
  * model did on its own (a task-notification followup, a peer/coordinator/
117
151
  * observer message it handled) rather than the user's prompt. Absent,
@@ -481,6 +515,14 @@ export class ClaudeAcpAgent {
481
515
  * return "cancelled". See {@link DEFAULT_FORCE_CANCEL_GRACE_MS}. Mutable so
482
516
  * tests can shrink it. */
483
517
  forceCancelGraceMs = DEFAULT_FORCE_CANCEL_GRACE_MS;
518
+ /** The account selector's last choice, as a selector value (see
519
+ * `accounts.ts`). Held on the AGENT, not on a session, because that is what
520
+ * it governs: the account applies to sessions created after it is set, and
521
+ * the thread that was on screen when the user chose keeps the account it
522
+ * started under (D7 — a session id belongs to one account's local history,
523
+ * so a live thread cannot be carried across a switch). `undefined` until a
524
+ * selection is made, in which case the first declared account stands. */
525
+ selectedAccount;
484
526
  constructor(client, logger) {
485
527
  this.sessions = {};
486
528
  this.client = client;
@@ -1007,6 +1049,12 @@ export class ClaudeAcpAgent {
1007
1049
  session.fileChangeReportRequestIds.add(fileChangeReportRequestId);
1008
1050
  fileChangeAudit = createFileChangeAuditTurnState(fileChangeReportRequestId);
1009
1051
  }
1052
+ // R2.3: exactly `/usage`, and nothing else in the prompt. A second block
1053
+ // (an attached file, a second text part) is a model turn that happens to
1054
+ // mention the command, not the local command itself.
1055
+ const isUsageCommand = params.prompt.length === 1 &&
1056
+ params.prompt[0]?.type === "text" &&
1057
+ isUsageCommandText(params.prompt[0].text);
1010
1058
  session.titles.onPrompt(params.prompt);
1011
1059
  // Each prompt is a Turn whose deferred the persistent consumer settles once
1012
1060
  // the turn's outcome is known. `prompt()` owns no loop: it enqueues the
@@ -1015,6 +1063,9 @@ export class ClaudeAcpAgent {
1015
1063
  const turn = {
1016
1064
  promptUuid,
1017
1065
  isLocalOnlyCommand,
1066
+ ...(isUsageCommand
1067
+ ? { isUsageCommand: true, usageMarkdownAbort: new AbortController() }
1068
+ : {}),
1018
1069
  ...(fileChangeAudit ? { fileChangeAudit } : {}),
1019
1070
  settled: false,
1020
1071
  resolve: () => { },
@@ -1105,7 +1156,14 @@ export class ClaudeAcpAgent {
1105
1156
  // answers to, and it is written out here rather than hidden behind a
1106
1157
  // helper so a rename lands as one named failure in the log below, not as
1107
1158
  // a quiet degradation to no report at all.
1108
- const report = await session.query.usage_EXPERIMENTAL_MAY_CHANGE_DO_NOT_RELY_ON_THIS_API_YET();
1159
+ // `skipBehaviors` drops the local-transcript scan that fills the report's
1160
+ // `behaviors` section. Nothing here reads it — `windowsFrom` touches only
1161
+ // `rate_limits_available` and `rate_limits` — so on every call, not just
1162
+ // the polled ones, that scan was pure cost. `/usage` still asks for the
1163
+ // full report, because it renders those contributions.
1164
+ const report = await session.query.usage_EXPERIMENTAL_MAY_CHANGE_DO_NOT_RELY_ON_THIS_API_YET({
1165
+ skipBehaviors: true,
1166
+ });
1109
1167
  // The session can be torn down or recreated while the request is in
1110
1168
  // flight (provider switch, a changed cwd on session/load). A successor
1111
1169
  // must not inherit its predecessor's windows.
@@ -1142,6 +1200,50 @@ export class ClaudeAcpAgent {
1142
1200
  this.logger.error(`Session ${sessionId}: structured usage report unavailable: ${error}`);
1143
1201
  }
1144
1202
  }
1203
+ /** Start the idle refresh of the account quota windows for this session.
1204
+ *
1205
+ * Idempotent, and deliberately called at the end of every turn rather than
1206
+ * once: the first call arms, the rest return immediately, so no caller has to
1207
+ * know whether this is turn one.
1208
+ *
1209
+ * The tick does nothing while a turn is in flight. That turn's own end
1210
+ * already refreshes the windows, and a second control request would buy the
1211
+ * same answer twice — the same reasoning that excludes autonomous cycles at
1212
+ * the end-of-turn call site. */
1213
+ armAccountUsagePolling(sessionId, session) {
1214
+ if (session.accountUsageTimer) {
1215
+ return;
1216
+ }
1217
+ const intervalMs = quotaPollIntervalMs();
1218
+ if (intervalMs === 0) {
1219
+ return;
1220
+ }
1221
+ // Closure-local rather than a Session field: it guards this timer alone, and
1222
+ // a slow report must not let ticks pile up into a queue of duplicate control
1223
+ // requests against a session that is already struggling to answer one.
1224
+ let inFlight = false;
1225
+ const timer = setInterval(() => {
1226
+ // Re-resolve through the map. A provider update replaces the Session
1227
+ // object under the same id and the replacement arms its own timer, so an
1228
+ // orphaned closure must retire rather than keep polling a husk.
1229
+ if (this.sessions[sessionId] !== session || session.queryClosed) {
1230
+ clearInterval(timer);
1231
+ return;
1232
+ }
1233
+ if (session.turnQueue?.length || session.activeTurn || inFlight) {
1234
+ return;
1235
+ }
1236
+ inFlight = true;
1237
+ // `publishAccountUsage` never rejects — it logs and returns — so this
1238
+ // cannot become an unhandled rejection, and `finally` always re-opens.
1239
+ void this.publishAccountUsage(sessionId, session).finally(() => {
1240
+ inFlight = false;
1241
+ });
1242
+ }, intervalMs);
1243
+ // A quota refresh is never a reason to keep the process alive.
1244
+ timer.unref?.();
1245
+ session.accountUsageTimer = timer;
1246
+ }
1145
1247
  async publishGoalFromPrompt(sessionId, prompt, commandUuid) {
1146
1248
  const goalUpdate = goalUpdateFromPrompt(prompt);
1147
1249
  if (goalUpdate !== undefined) {
@@ -1310,11 +1412,6 @@ export class ClaudeAcpAgent {
1310
1412
  // stop_reason "refusal" and structured stop_details. We capture the
1311
1413
  // human-readable explanation so the terminal `result` can surface it.
1312
1414
  let lastRefusalExplanation = null;
1313
- // Tracks whether we're inside a compaction. The SDK emits the terminal
1314
- // `status` (compact_result success/failed) twice for a single failed
1315
- // compaction, and the two messages are indistinguishable — so we report the
1316
- // outcome only while a compaction is in progress, then clear this.
1317
- let compactionInProgress = false;
1318
1415
  // Anthropic API message id of the assistant message currently being
1319
1416
  // streamed, captured from `message_start` so the streamed chunks that follow
1320
1417
  // (whose delta events don't carry it) can all be tagged with the same,
@@ -1349,6 +1446,12 @@ export class ClaudeAcpAgent {
1349
1446
  * recognizable by the `parentToolUseId` meta that toAcpNotifications
1350
1447
  * stamps from `parent_tool_use_id`, and never reach the top-level feed
1351
1448
  * as the turn's answer. */
1449
+ /** Compaction as one ACP tool lifecycle, replacing the in-progress boolean
1450
+ * this consumer used to infer it from (story 010, R2.1/R2.2). Declared
1451
+ * ahead of `sendUpdate` because that chokepoint consults it; the closure
1452
+ * below only dereferences `sendUpdate` when an update is actually sent,
1453
+ * which is always after both bindings exist. */
1454
+ const compaction = new ContextCompactionLifecycle((notification) => sendUpdate(notification));
1352
1455
  const sendUpdate = async (notification) => {
1353
1456
  const { update } = notification;
1354
1457
  if (isFileChangeAuditReportPhase(session.activeTurn?.fileChangeAudit) &&
@@ -1361,6 +1464,14 @@ export class ClaudeAcpAgent {
1361
1464
  }
1362
1465
  if (update.sessionUpdate === "agent_message_chunk") {
1363
1466
  const claudeMeta = update._meta?.claudeCode;
1467
+ // A failed manual compaction's error also arrives as assistant text;
1468
+ // the tool lifecycle already carried it, so drop that one copy rather
1469
+ // than report the same failure twice (R2.2).
1470
+ if (!claudeMeta?.parentToolUseId &&
1471
+ update.content.type === "text" &&
1472
+ compaction.consumeDuplicateErrorOutput(update.content.text)) {
1473
+ return;
1474
+ }
1364
1475
  if (!claudeMeta?.parentToolUseId) {
1365
1476
  session.emittedAssistantText = true;
1366
1477
  session.titles.onAssistantText(update.content);
@@ -1413,7 +1524,6 @@ export class ClaudeAcpAgent {
1413
1524
  lastAssistantWasUsageLimit = false;
1414
1525
  lastAssistantFailureTitle = undefined;
1415
1526
  lastRefusalExplanation = null;
1416
- compactionInProgress = false;
1417
1527
  // Do NOT reset currentStreamMessageId or streamedBlocks here. Turn
1418
1528
  // activation can fire mid-message (the replayed user echo with
1419
1529
  // --replay-user-messages lands between a message's blocks); clearing the
@@ -1432,6 +1542,22 @@ export class ClaudeAcpAgent {
1432
1542
  if (session.activeTurn)
1433
1543
  session.activeTurn.carriedUsage = undefined;
1434
1544
  };
1545
+ /** Start this turn's structured `/usage` render, once, and hand back the
1546
+ * in-flight promise; undefined when the turn owns no render.
1547
+ *
1548
+ * Started at ACTIVATION rather than when the output arrives, so the
1549
+ * bounded wait overlaps the command instead of following it: by the time
1550
+ * the CLI has printed its answer the report has usually already lost or
1551
+ * won its race, and the user waits for neither. */
1552
+ const ensureUsageMarkdown = (turn) => {
1553
+ if (!turn.isUsageCommand || !turn.usageMarkdownAbort) {
1554
+ return undefined;
1555
+ }
1556
+ // `session.query` is read here rather than captured: a provider switch
1557
+ // can replace it, and the render must ask the query that is live now.
1558
+ turn.usageMarkdown ??= structuredUsageMarkdown(session.query, turn.usageMarkdownAbort.signal, this.logger);
1559
+ return turn.usageMarkdown;
1560
+ };
1435
1561
  /** Promote a queued turn to active: it becomes the one output is attributed
1436
1562
  * to, and its scratch starts fresh. Clears the cancelled flag so a turn
1437
1563
  * enqueued after a prior cancel isn't treated as cancelled. Also clears any
@@ -1443,6 +1569,7 @@ export class ClaudeAcpAgent {
1443
1569
  const activateTurn = (turn) => {
1444
1570
  session.activeTurn = turn;
1445
1571
  session.cancelled = false;
1572
+ ensureUsageMarkdown(turn);
1446
1573
  session.pendingOrphanResults = 0;
1447
1574
  session.orphanCommands?.clear();
1448
1575
  // Two-phase sweep of registry entries the level signal ended (see
@@ -1481,12 +1608,20 @@ export class ClaudeAcpAgent {
1481
1608
  *
1482
1609
  * But an echo-less result can also be an ORPHAN: cancel() settles+removes a
1483
1610
  * queued turn whose user message was already pushed, so the SDK still runs
1484
- * it and emits a result with no uuid to match. Promoting the head for an
1611
+ * it and emits a result with no echo to match. Promoting the head for an
1485
1612
  * orphan would misattribute its stop reason/usage to an unrelated later
1486
1613
  * prompt. `session.pendingOrphanResults` counts exactly how many such
1487
1614
  * orphans are still expected (FIFO, they arrive before any live turn's
1488
- * result), so we skip those and only promote once the count is drained. */
1489
- const ensureActiveTurn = () => {
1615
+ * result), so we skip those and only promote once the count is drained.
1616
+ *
1617
+ * `resultUserMessageUuid` is the result's own join key (SDK 0.3.246+
1618
+ * echoes the triggering send's client uuid on results; absent on older
1619
+ * CLIs, synthetic/meta turns, and session-scoped failures). When present
1620
+ * it upgrades the map lane from positional heuristics to an exact match:
1621
+ * a stamp naming an orphaned command consumes the result outright, and a
1622
+ * stamp naming anything else positively refutes "this is a dead turn's
1623
+ * result", so the dup-over-loss one-skip must not eat it. */
1624
+ const ensureActiveTurn = (resultUserMessageUuid) => {
1490
1625
  if (session.activeTurn) {
1491
1626
  if (!isHeldOpen(session.activeTurn)) {
1492
1627
  return;
@@ -1538,8 +1673,19 @@ export class ClaudeAcpAgent {
1538
1673
  // double-consume it. The unexpected-transition logging in the frame
1539
1674
  // handler is the tripwire for that class of drift.
1540
1675
  if (session.orphanCommands?.size) {
1676
+ // Resolved BEFORE the drain below, which deletes every started/zombie
1677
+ // entry: a stamp naming a "started" orphan would no longer be found
1678
+ // afterwards, and the result would promote the head — misattributing
1679
+ // a dead turn's outcome to a live prompt, the exact failure the map
1680
+ // exists to prevent. The lookup order is load-bearing.
1681
+ const stampedOrphanUuid = resultUserMessageUuid !== undefined && session.orphanCommands.has(resultUserMessageUuid)
1682
+ ? resultUserMessageUuid
1683
+ : undefined;
1541
1684
  let consumedOrphanResult = false;
1542
1685
  let oldestPending;
1686
+ // The started/zombie drain applies regardless of the stamp: commands
1687
+ // folded into the turn that emitted this result share it, and zombies'
1688
+ // late results have already passed (or never existed).
1543
1689
  for (const [uuid, state] of session.orphanCommands) {
1544
1690
  if (state === "started" || state === "zombie") {
1545
1691
  consumedOrphanResult = true;
@@ -1549,17 +1695,33 @@ export class ClaudeAcpAgent {
1549
1695
  oldestPending ??= uuid;
1550
1696
  }
1551
1697
  }
1552
- if (consumedOrphanResult) {
1698
+ if (stampedOrphanUuid !== undefined) {
1699
+ // Exact join: the result names an orphaned command. Delete the
1700
+ // matched entry even when it is still "pending" (its dispatch frame
1701
+ // was lost) and consume the result — no promotion.
1702
+ session.orphanCommands.delete(stampedOrphanUuid);
1553
1703
  return;
1554
1704
  }
1555
- if (oldestPending !== undefined) {
1556
- // No dispatch was seen before this result, so it is very likely a
1557
- // live turn's — but a lost "started" frame would mean it IS the
1558
- // orphan's (dup-over-loss: prefer one wrong skip over
1559
- // misattributing a dead turn's outcome to a live prompt). Grant
1560
- // each pending entry exactly one skip, like the count lane did.
1561
- session.orphanCommands.delete(oldestPending);
1562
- return;
1705
+ if (resultUserMessageUuid !== undefined) {
1706
+ // The stamp names a send that is NOT in the orphan map, so this is
1707
+ // a live turn's result: skip both the consumed-return (its folded
1708
+ // orphans were drained above, but the result itself still needs a
1709
+ // turn) and the dup-over-loss one-skip the stamp refutes, and fall
1710
+ // through to promote the head.
1711
+ }
1712
+ else {
1713
+ if (consumedOrphanResult) {
1714
+ return;
1715
+ }
1716
+ if (oldestPending !== undefined) {
1717
+ // No dispatch was seen before this result, so it is very likely a
1718
+ // live turn's — but a lost "started" frame would mean it IS the
1719
+ // orphan's (dup-over-loss: prefer one wrong skip over
1720
+ // misattributing a dead turn's outcome to a live prompt). Grant
1721
+ // each pending entry exactly one skip, like the count lane did.
1722
+ session.orphanCommands.delete(oldestPending);
1723
+ return;
1724
+ }
1563
1725
  }
1564
1726
  }
1565
1727
  const head = firstUnsettledQueuedTurn();
@@ -1600,6 +1762,49 @@ export class ClaudeAcpAgent {
1600
1762
  * spelling of "a prompt is pending" shared by the head promotion and
1601
1763
  * the autonomous stretch-close guard. */
1602
1764
  const firstUnsettledQueuedTurn = () => (session.turnQueue ?? []).find((t) => !t.settled);
1765
+ /** Claim the structured render for whichever turn is producing this local
1766
+ * command output. Three answers, and the caller must distinguish all
1767
+ * three:
1768
+ *
1769
+ * undefined — not a `/usage` turn, or one of R2.4's three ways out
1770
+ * fired: publish `originalOutput` unchanged
1771
+ * string — the render: publish it INSTEAD of `originalOutput`
1772
+ * null — publish nothing (a duplicate of an already-replaced
1773
+ * frame, or a turn cancelled during the wait)
1774
+ *
1775
+ * No content signature and no text parsing: which turn owns a render was
1776
+ * decided from the prompt, so a command whose output happens to look like
1777
+ * `/usage`'s can never be rewritten. */
1778
+ const takeUsageMarkdown = async (originalOutput) => {
1779
+ const turn = session.activeTurn ?? firstUnsettledQueuedTurn();
1780
+ if (!turn) {
1781
+ return undefined;
1782
+ }
1783
+ const pending = ensureUsageMarkdown(turn);
1784
+ if (!pending) {
1785
+ return undefined;
1786
+ }
1787
+ const markdown = await pending;
1788
+ // That await can span the whole bounded wait, and a cancel landing inside
1789
+ // it ends the turn. A cancelled turn publishes nothing — not the render,
1790
+ // and not the original text either.
1791
+ if (session.cancelled) {
1792
+ return null;
1793
+ }
1794
+ if (markdown === null) {
1795
+ return undefined;
1796
+ }
1797
+ if (turn.usageMarkdownDelivered) {
1798
+ // One local command can reach the client through more than one SDK
1799
+ // message shape. Suppress an exact mirror of the frame already
1800
+ // replaced, but let a later, genuinely different frame (an
1801
+ // interruption diagnostic, say) take the normal path.
1802
+ return turn.usageOriginalOutput === originalOutput ? null : undefined;
1803
+ }
1804
+ turn.usageMarkdownDelivered = true;
1805
+ turn.usageOriginalOutput = originalOutput;
1806
+ return markdown;
1807
+ };
1603
1808
  /** Whether any background subagent this turn spawned is still live —
1604
1809
  * while true, the turn's settlement stays deferred so the subagent's
1605
1810
  * output and permission requests land inside it (see
@@ -1668,6 +1873,9 @@ export class ClaudeAcpAgent {
1668
1873
  // Captured before the settled flip below (isHeldOpen tests !settled).
1669
1874
  const wasHeld = isHeldOpen(turn);
1670
1875
  turn.settled = true;
1876
+ // The turn is over: abandon any structured `/usage` render still in
1877
+ // flight rather than leaving it to run out its own clock (D5).
1878
+ turn.usageMarkdownAbort?.abort();
1671
1879
  disarmForceCancel(session);
1672
1880
  session.turnQueue = (session.turnQueue ?? []).filter((t) => t !== turn);
1673
1881
  session.activeTurn = null;
@@ -1700,6 +1908,7 @@ export class ClaudeAcpAgent {
1700
1908
  }
1701
1909
  this.finishFileChangeAudit(session, turn, "providerError");
1702
1910
  turn.settled = true;
1911
+ turn.usageMarkdownAbort?.abort();
1703
1912
  session.turnQueue = (session.turnQueue ?? []).filter((t) => t !== turn);
1704
1913
  session.activeTurn = null;
1705
1914
  streamedToolInputs.clear();
@@ -1748,6 +1957,7 @@ export class ClaudeAcpAgent {
1748
1957
  this.finishFileChangeAudit(session, turn, "providerError");
1749
1958
  const wasHeld = isHeldOpen(turn);
1750
1959
  turn.settled = true;
1960
+ turn.usageMarkdownAbort?.abort();
1751
1961
  if (wasHeld) {
1752
1962
  // A held turn's answer already streamed and its outcome is
1753
1963
  // recorded — a stream death during the post-answer hold is a
@@ -2073,50 +2283,43 @@ export class ClaudeAcpAgent {
2073
2283
  }
2074
2284
  break;
2075
2285
  case "status": {
2076
- // These banners count as delivered text (via sendUpdate), so
2077
- // an echo-less turn that only ever emits them (e.g. `/compact`,
2078
- // promoted at its own result) doesn't have its result text
2079
- // re-emitted by the issue-#453 fallback.
2286
+ // Compaction is reported as ONE tool lifecycle rather than the
2287
+ // assistant-text banners this branch used to emit (R2.1). The
2288
+ // banners needed a boolean in-progress guard because the SDK
2289
+ // repeats the terminal `status` and the two copies are
2290
+ // indistinguishable here; the lifecycle owns that de-duplication
2291
+ // by phase instead, so the guard is gone rather than inert (R2.2).
2292
+ //
2293
+ // The SDK signals manual `/compact` completion with a status
2294
+ // message carrying `compact_result`, not the `compact_boundary`
2295
+ // message (which only fires when there's content to compact) —
2296
+ // so both frames must be able to close a lifecycle.
2080
2297
  if (message.status === "compacting") {
2081
- compactionInProgress = true;
2082
- await sendUpdate({
2083
- sessionId: message.session_id,
2084
- update: {
2085
- sessionUpdate: "agent_message_chunk",
2086
- content: { type: "text", text: "Compacting..." },
2087
- },
2088
- });
2298
+ await compaction.start(message.session_id, message.uuid);
2089
2299
  }
2090
- else if (message.compact_result === "success" && compactionInProgress) {
2091
- // The SDK signals manual `/compact` completion with a status
2092
- // message carrying `compact_result`, not the `compact_boundary`
2093
- // message (which only fires when there's content to compact).
2094
- compactionInProgress = false;
2095
- await sendUpdate({
2096
- sessionId: message.session_id,
2097
- update: {
2098
- sessionUpdate: "agent_message_chunk",
2099
- content: { type: "text", text: "\n\nCompacting completed." },
2100
- },
2101
- });
2300
+ else if (message.compact_result === "success") {
2301
+ await compaction.finish(message.session_id, message.uuid, "completed");
2102
2302
  }
2103
- else if (message.compact_result === "failed" && compactionInProgress) {
2104
- compactionInProgress = false;
2105
- const reason = message.compact_error ? `: ${message.compact_error}` : ".";
2106
- await sendUpdate({
2107
- sessionId: message.session_id,
2108
- update: {
2109
- sessionUpdate: "agent_message_chunk",
2110
- content: { type: "text", text: `\n\nCompacting failed${reason}` },
2111
- },
2303
+ else if (message.compact_result === "failed") {
2304
+ await compaction.finish(message.session_id, message.uuid, "failed", {
2305
+ ...(message.compact_error ? { error: message.compact_error } : {}),
2112
2306
  });
2113
2307
  }
2114
2308
  break;
2115
2309
  }
2116
2310
  case "compact_boundary": {
2311
+ // This is the only frame carrying the token counts, and it
2312
+ // arrives AFTER the terminal `status` — so it enriches the
2313
+ // lifecycle rather than reporting an outcome again
2314
+ // (`enrichTerminal`): the update omits `status`, leaving exactly
2315
+ // one terminal transition the client can see (R2.2). It also
2316
+ // stands alone when no `status` preceded it (an automatic
2317
+ // compaction, or a replay that dropped the opening frame).
2318
+ const compactMetadata = message.compact_metadata;
2319
+ await compaction.finish(message.session_id, message.uuid, "completed", compactMetadata ? contextCompactionMetadataFromBoundary(compactMetadata) : {}, true);
2117
2320
  // Refresh the displayed usage immediately so the client doesn't
2118
2321
  // keep showing the stale pre-compaction size (e.g. "944k/1m")
2119
- // right after the user sees "Compacting completed", which is
2322
+ // right after the compaction tool call completes, which is
2120
2323
  // confusing and wrong.
2121
2324
  //
2122
2325
  // Prefer the SDK's authoritative post-compaction `used` via
@@ -2131,10 +2334,6 @@ export class ClaudeAcpAgent {
2131
2334
  // `size` keeps coming from session.contextWindowSize —
2132
2335
  // compaction frees occupancy, it doesn't change the model's
2133
2336
  // window.
2134
- //
2135
- // The "Compacting completed." text is emitted from the `status`
2136
- // handler (keyed on `compact_result`), not here, so the failure
2137
- // path gets a message too.
2138
2337
  const usedTokens = await fetchContextUsedTokens(session.query, this.logger);
2139
2338
  lastAssistantUsage = null;
2140
2339
  lastAssistantTotalUsage = usedTokens ?? 0;
@@ -2150,11 +2349,24 @@ export class ClaudeAcpAgent {
2150
2349
  break;
2151
2350
  }
2152
2351
  case "local_command_output": {
2352
+ // A failed `/compact` also prints its error here; the tool
2353
+ // lifecycle already carried it, so consume that one duplicate
2354
+ // (matched by text, once) without hiding other command output.
2355
+ if (compaction.consumeDuplicateErrorOutput(message.content)) {
2356
+ break;
2357
+ }
2358
+ // R2.3/R2.4: a `/usage` turn publishes its structured render
2359
+ // here instead; anything else, and any way out, publishes the
2360
+ // CLI's own text byte-for-byte.
2361
+ const usageMarkdown = await takeUsageMarkdown(message.content);
2362
+ if (usageMarkdown === null) {
2363
+ break;
2364
+ }
2153
2365
  await sendUpdate({
2154
2366
  sessionId: message.session_id,
2155
2367
  update: {
2156
2368
  sessionUpdate: "agent_message_chunk",
2157
- content: { type: "text", text: message.content },
2369
+ content: { type: "text", text: usageMarkdown ?? message.content },
2158
2370
  },
2159
2371
  });
2160
2372
  break;
@@ -2192,6 +2404,10 @@ export class ClaudeAcpAgent {
2192
2404
  // the interrupted turn's tokens entirely (issue #844). Zero
2193
2405
  // when the cancel pre-empted the result (wedge/force-cancel).
2194
2406
  if (session.cancelled && session.activeTurn && !session.activeTurn.settled) {
2407
+ // An interrupt can pre-empt the result entirely, so the
2408
+ // lifecycle's own reset there never ran; close it here or a
2409
+ // half-open compaction would leak into the next turn.
2410
+ compaction.reset();
2195
2411
  settleActive({ stopReason: "cancelled", usage: sessionUsage(session) });
2196
2412
  // An interrupt can pre-empt the turn's result entirely
2197
2413
  // (nothing ran the result-case `finally`), so close the
@@ -2254,6 +2470,9 @@ export class ClaudeAcpAgent {
2254
2470
  else if (!session.cancelled &&
2255
2471
  session.activeTurn &&
2256
2472
  !session.activeTurn.settled) {
2473
+ // Same reason as the cancelled branch: this turn will never
2474
+ // reach the result that would have reset the lifecycle.
2475
+ compaction.reset();
2257
2476
  // Deliberately only the ACTIVE turn: a queued turn that
2258
2477
  // was never echoed is NOT failed here, because an idle
2259
2478
  // can legitimately precede the SDK picking up freshly
@@ -2669,7 +2888,7 @@ export class ClaudeAcpAgent {
2669
2888
  // the map in that case).
2670
2889
  if (!isAutonomousResult) {
2671
2890
  recordResultForOrphanCommands();
2672
- ensureActiveTurn();
2891
+ ensureActiveTurn(message.user_message_uuid);
2673
2892
  // Once the submitted goal command has produced its own result,
2674
2893
  // no older runtime update can still precede it in the ordered
2675
2894
  // SDK stream. Stop suppressing updates even when this runtime
@@ -2700,6 +2919,12 @@ export class ClaudeAcpAgent {
2700
2919
  // through the early break below, which the gated `finally`
2701
2920
  // leaves alone).
2702
2921
  const deliveredAssistantText = session.emittedAssistantText;
2922
+ // The deleted "Compacting…" banners counted as delivered text
2923
+ // simply by going through sendUpdate; tool calls don't, so a turn
2924
+ // that ONLY compacted (e.g. `/compact`, promoted at its own
2925
+ // result) needs this to keep its result text from being re-emitted
2926
+ // by the issue-#453 fallback below.
2927
+ const deliveredCompactionOutput = compaction.hasDeliveredOutput;
2703
2928
  // Every user-turn result terminates a turn (settle, reject, or
2704
2929
  // orphan skip) and the SDK follows it with a trailing
2705
2930
  // `session_state_changed: idle` — record the debt so the idle
@@ -2946,9 +3171,19 @@ export class ClaudeAcpAgent {
2946
3171
  // the fallback there. (Autonomous results never get here —
2947
3172
  // they exit at the early break above — so no background
2948
3173
  // prose can be injected into the feed.)
2949
- if (session.activeTurn?.isLocalOnlyCommand ||
2950
- (!deliveredAssistantText && (message.usage.output_tokens ?? 0) === 0)) {
2951
- for (const notification of toAcpNotifications(message.result, "assistant", params.sessionId, session.toolUseCache, this.client, this.logger)) {
3174
+ const shouldForwardResult = session.activeTurn?.isLocalOnlyCommand ||
3175
+ (!deliveredAssistantText &&
3176
+ !deliveredCompactionOutput &&
3177
+ (message.usage.output_tokens ?? 0) === 0);
3178
+ if (shouldForwardResult) {
3179
+ // A `/usage` turn whose output arrives only on the result.
3180
+ // Claiming it here also stops the raw text following a
3181
+ // render already published from another message shape.
3182
+ const usageMarkdown = await takeUsageMarkdown(message.result);
3183
+ if (usageMarkdown === null) {
3184
+ break;
3185
+ }
3186
+ for (const notification of toAcpNotifications(usageMarkdown ?? message.result, "assistant", params.sessionId, session.toolUseCache, this.client, this.logger)) {
2952
3187
  await sendUpdate(notification);
2953
3188
  }
2954
3189
  }
@@ -3023,6 +3258,11 @@ export class ClaudeAcpAgent {
3023
3258
  finally {
3024
3259
  if (!isAutonomousResult) {
3025
3260
  session.emittedAssistantText = false;
3261
+ // A result closes this turn's compaction lifecycle. Reset here
3262
+ // rather than at idle: an owed idle from this turn can arrive
3263
+ // after the next turn has already started, and would erase that
3264
+ // turn's compaction state instead of its own.
3265
+ compaction.reset();
3026
3266
  // R1.2: a user turn's terminal result is the moment consumption
3027
3267
  // actually changed, so refresh the account quota windows here —
3028
3268
  // in the `finally`, which is the one point EVERY exit from this
@@ -3038,11 +3278,26 @@ export class ClaudeAcpAgent {
3038
3278
  // the trailing idle is, and a stalled control request there is
3039
3279
  // the same exposure `fetchContextUsedTokens` already accepts.
3040
3280
  await this.publishAccountUsage(params.sessionId, session);
3281
+ // A settled turn is the proof that control requests are being
3282
+ // serviced (before the first one they are not — #886/#880), so
3283
+ // this is the earliest honest moment to start the idle refresh.
3284
+ this.armAccountUsagePolling(params.sessionId, session);
3041
3285
  }
3042
3286
  }
3043
3287
  break;
3044
3288
  }
3045
3289
  case "stream_event": {
3290
+ // Compaction streams as its own block type. The deltas carry the
3291
+ // generated summary, which stays internal to the agent — all the
3292
+ // client gets is one keep-alive on the open tool call, so a long
3293
+ // compaction doesn't look stalled.
3294
+ const isCompactionProgress = (message.event.type === "content_block_start" &&
3295
+ message.event.content_block.type === "compaction") ||
3296
+ (message.event.type === "content_block_delta" &&
3297
+ message.event.delta.type === "compaction_delta");
3298
+ if (isCompactionProgress) {
3299
+ await compaction.heartbeat(message.session_id, message.uuid);
3300
+ }
3046
3301
  // `message_start` carries the Anthropic API message id; capture it
3047
3302
  // so the streamed chunks that follow (whose delta events don't carry
3048
3303
  // it) can all be tagged with the same, replay-stable id.
@@ -3265,6 +3520,18 @@ export class ClaudeAcpAgent {
3265
3520
  if (session.cancelled) {
3266
3521
  break;
3267
3522
  }
3523
+ // Synthetic assistant frames carry the CLI's local-command output.
3524
+ // On resume the SDK can replay a stale frame from an earlier compact
3525
+ // attempt after a later compaction completed. Scope the suppression
3526
+ // to the compaction lifecycle and the synthetic frame itself rather
3527
+ // than to the owning turn: one model turn may compact more than
3528
+ // once, and its real assistant response must still be delivered.
3529
+ if (message.type === "assistant" &&
3530
+ message.parent_tool_use_id === null &&
3531
+ message.message.model === "<synthetic>" &&
3532
+ compaction.hasDeliveredOutput) {
3533
+ break;
3534
+ }
3268
3535
  // Snapshot the latest top-level assistant usage and model so the
3269
3536
  // next `result` can emit a usage_update tied to the right context
3270
3537
  // window. Subagent messages are excluded to keep the snapshot
@@ -3303,7 +3570,13 @@ export class ClaudeAcpAgent {
3303
3570
  message.message.content.includes("<local-command-stdout>")) {
3304
3571
  const stripped = stripLocalCommandMetadata(message.message.content);
3305
3572
  if (typeof stripped === "string") {
3306
- for (const notification of toAcpNotifications(stripped, message.message.role, params.sessionId, session.toolUseCache, this.client, this.logger, {
3573
+ // The usual shape for a real `/usage`: the comment above names
3574
+ // it as one of the commands the CLI wraps in these markers.
3575
+ const usageMarkdown = await takeUsageMarkdown(stripped);
3576
+ if (usageMarkdown === null) {
3577
+ break;
3578
+ }
3579
+ for (const notification of toAcpNotifications(usageMarkdown ?? stripped, message.message.role, params.sessionId, session.toolUseCache, this.client, this.logger, {
3307
3580
  clientCapabilities: this.clientCapabilities,
3308
3581
  parentToolUseId: message.parent_tool_use_id,
3309
3582
  cwd: session.cwd,
@@ -3644,6 +3917,11 @@ export class ClaudeAcpAgent {
3644
3917
  await session.queryRecreateInFlight;
3645
3918
  }
3646
3919
  session.cancelled = true;
3920
+ // Every turn's structured `/usage` render is abandoned here — the queue
3921
+ // still holds the active turn at this point, so one sweep covers both.
3922
+ for (const turn of session.turnQueue ?? []) {
3923
+ turn.usageMarkdownAbort?.abort();
3924
+ }
3647
3925
  session.pendingExitPlanModeInterruption = undefined;
3648
3926
  session.pendingExitPlanContextReset = undefined;
3649
3927
  // The stream already ended (see closeQueryStream): every in-flight turn was
@@ -3686,8 +3964,10 @@ export class ClaudeAcpAgent {
3686
3964
  }
3687
3965
  // Each removed queued turn's user message was already pushed to the SDK,
3688
3966
  // which processes input FIFO and will still emit a result for it with no
3689
- // uuid to match. Track those so the consumer skips them (see
3690
- // ensureActiveTurn) rather than misattributing them to the head.
3967
+ // user echo to match (0.3.246+ CLIs do stamp results with the
3968
+ // triggering send's user_message_uuid, which ensureActiveTurn uses as
3969
+ // an exact join when present). Track those so the consumer skips them
3970
+ // (see ensureActiveTurn) rather than misattributing them to the head.
3691
3971
  // msg_lifecycle_v1 CLIs get per-uuid tracking drained by the command's
3692
3972
  // own terminal lifecycle frame — exact under command coalescing, where
3693
3973
  // N queued commands fold into ONE turn emitting one result and a plain
@@ -3874,6 +4154,15 @@ export class ClaudeAcpAgent {
3874
4154
  }
3875
4155
  session.queryClosed = true;
3876
4156
  session.consumer = undefined;
4157
+ // Every teardown path reaches this method, which is why the quota timer is
4158
+ // cleared here and nowhere else: a `clearInterval` repeated at each of the
4159
+ // four call sites is one merge away from missing the fifth. A session whose
4160
+ // stream is closed cannot answer the usage control request anyway, so a
4161
+ // surviving timer would only log a failure a minute, forever.
4162
+ if (session.accountUsageTimer) {
4163
+ clearInterval(session.accountUsageTimer);
4164
+ session.accountUsageTimer = undefined;
4165
+ }
3877
4166
  session.settingsManager.dispose();
3878
4167
  session.input.end();
3879
4168
  session.query.close();
@@ -4619,6 +4908,14 @@ export class ClaudeAcpAgent {
4619
4908
  // same way it does for any unsupported level.
4620
4909
  available: isUltracodeAvailable(newModelInfo?.supportsEffort ? (newModelInfo.supportedEffortLevels ?? []) : [], { disableWorkflows: session.workflowsDisabled }),
4621
4910
  enabled: session.ultracode,
4911
+ }, {
4912
+ disableWorkflows: session.workflowsDisabled,
4913
+ // Threaded for the same reason Thinking is (R1.7): an option not
4914
+ // passed through this rebuild silently drops from the picker on
4915
+ // every model switch. The account is model-independent, so the row
4916
+ // and its selection survive unchanged.
4917
+ accounts: session.declaredAccounts,
4918
+ currentAccount: session.currentAccount,
4622
4919
  });
4623
4920
  // Sync effort with the SDK if it changed after the model switch
4624
4921
  const newEffortOpt = session.configOptions.find((o) => o.id === EFFORT_CONFIG_ID);
@@ -4649,6 +4946,26 @@ export class ClaudeAcpAgent {
4649
4946
  session.currentAgent = value;
4650
4947
  session.configOptions = session.configOptions.map((o) => o.id === configId && typeof o.currentValue === "string" ? { ...o, currentValue: value } : o);
4651
4948
  }
4949
+ else if (configId === ACCOUNT_CONFIG_ID) {
4950
+ // Resolve BEFORE recording anything: an account nobody declared must
4951
+ // leave the session exactly as it was. `setSessionConfigOption`'s shared
4952
+ // validation already refuses a value that is not among the option's
4953
+ // entries; this second check covers the one path that bypasses it (a
4954
+ // client round-tripping `currentValue`) and keeps the refusal here
4955
+ // rather than in a caller (R6.3).
4956
+ const account = accountForValue(session.declaredAccounts ?? [], value);
4957
+ if (!account) {
4958
+ throw new Error(`Invalid value for config option ${configId}: ${value}`);
4959
+ }
4960
+ // No SDK call, and no query recreate: the account is applied through the
4961
+ // per-session `env` at query CREATION (D2/D9 — the SDK already accepts
4962
+ // one, so no Zed patch is involved), and this session's query keeps the
4963
+ // account it was created under (D7).
4964
+ this.selectedAccount = value;
4965
+ session.currentAccount = value;
4966
+ session.accountEnv = envForAccount(account);
4967
+ session.configOptions = session.configOptions.map((o) => o.id === configId && typeof o.currentValue === "string" ? { ...o, currentValue: value } : o);
4968
+ }
4652
4969
  else {
4653
4970
  session.configOptions = session.configOptions.map((o) => o.id === configId && typeof o.currentValue === "string" ? { ...o, currentValue: value } : o);
4654
4971
  if (configId === EFFORT_CONFIG_ID) {
@@ -5247,9 +5564,28 @@ export class ClaudeAcpAgent {
5247
5564
  env: { ...baseSettings?.env, ...providerEnv },
5248
5565
  };
5249
5566
  }
5567
+ // The account this session runs under (R6.3): the selector's last choice
5568
+ // when it names a declared account, else the first declared one. Read from
5569
+ // the resolved settings, because accounts are DECLARED, never discovered
5570
+ // (D4) — nothing here scans a home directory for credential stores. With
5571
+ // none declared there is no overlay and the process's own configuration
5572
+ // home stands, which is exactly the behavior that predates this option.
5573
+ const declaredAccounts = readDeclaredAccounts(settingsManager.getSettings(), this.logger);
5574
+ const accountEntry = accountInForce(declaredAccounts, this.selectedAccount);
5575
+ // The SESSION's account home, never the ADAPTER's (D3): this overlay is
5576
+ // handed to the SDK per session and to nothing else. `process.env` is not
5577
+ // mutated — `CLAUDE_CONFIG_DIR` is read once into a module-level constant
5578
+ // that `settings.ts` uses to find the adapter's OWN settings.json, so a
5579
+ // mutation would be seen by some readers and not others.
5580
+ const accountEnv = accountEntry ? envForAccount(accountEntry.account) : undefined;
5250
5581
  const env = {
5251
5582
  ...process.env,
5252
5583
  ...userProvidedOptions?.env,
5584
+ // After the client-supplied env: the selector is a live user choice,
5585
+ // and an agent entry that pinned CLAUDE_CONFIG_DIR is precisely the
5586
+ // "account as a second agent" shape this option replaces (D1). Before
5587
+ // the provider routing, which sets ANTHROPIC_* and never this key.
5588
+ ...accountEnv,
5253
5589
  // Client-managed LLM routing: `providers/set` config wins, else the
5254
5590
  // legacy gateway auth request. Routing is baked into the query at
5255
5591
  // creation; provider updates recreate loaded queries between turns.
@@ -5533,6 +5869,13 @@ export class ClaudeAcpAgent {
5533
5869
  // `xhigh` on its own is a legitimate selection that is not Ultracode.
5534
5870
  available: isUltracodeAvailable(currentModelInfo?.supportsEffort ? (currentModelInfo.supportedEffortLevels ?? []) : [], { disableWorkflows: workflowsDisabled }),
5535
5871
  enabled: settingsManager.getSettings().ultracode === true,
5872
+ }, {
5873
+ disableWorkflows: workflowsDisabled,
5874
+ // Both halves of the selector, resolved above alongside the env they
5875
+ // produced — so the row the client renders and the configuration
5876
+ // directory the query was actually started with cannot disagree.
5877
+ accounts: declaredAccounts,
5878
+ currentAccount: accountEntry?.value,
5536
5879
  });
5537
5880
  // Apply the initial effort level to the SDK so it matches the UI default
5538
5881
  const initialEffort = configOptions.find((o) => o.id === EFFORT_CONFIG_ID);
@@ -5611,6 +5954,9 @@ export class ClaudeAcpAgent {
5611
5954
  currentAgent,
5612
5955
  ultracode: initialUltracode,
5613
5956
  workflowsDisabled,
5957
+ declaredAccounts,
5958
+ currentAccount: accountEntry?.value,
5959
+ accountEnv,
5614
5960
  fastModeEnabled,
5615
5961
  fastModeDisabledReason,
5616
5962
  abortController,
@@ -5980,8 +6326,12 @@ thinkingEnabled,
5980
6326
  * direct callers/tests, in which case availability is derived from
5981
6327
  * `settings` and the entry renders unselected. */
5982
6328
  ultracode,
5983
- /** Resolved session settings, read only for `disableWorkflows` — the half of
5984
- * the Ultracode gate that is not a model capability. */
6329
+ /** Resolved session settings, read for `disableWorkflows` — the half of the
6330
+ * Ultracode gate that is not a model capability — and, alongside them, the
6331
+ * declared accounts plus the one in force, which decide whether an account
6332
+ * selector is advertised at all (R6.1/R6.2). `accounts` is not a field of
6333
+ * the SDK's `Settings`, so it is carried beside them rather than picked
6334
+ * from them. */
5985
6335
  settings) {
5986
6336
  const options = [
5987
6337
  SessionModeManager.configOption(modes),
@@ -6087,6 +6437,15 @@ settings) {
6087
6437
  ],
6088
6438
  });
6089
6439
  }
6440
+ // The account selector, last because it is the least often touched of the
6441
+ // rows and the only one that governs the NEXT thread rather than this one.
6442
+ // `createAccountConfigOption` returns undefined below two declared accounts,
6443
+ // so the row is genuinely absent there rather than present and empty (D8,
6444
+ // R6.2).
6445
+ const accountOption = createAccountConfigOption(settings?.accounts ?? [], settings?.currentAccount);
6446
+ if (accountOption) {
6447
+ options.push(accountOption);
6448
+ }
6090
6449
  return options;
6091
6450
  }
6092
6451
  // Claude Code CLI persists display strings like "opus[1m]" in settings,