diff --git a/src-tauri/src/acp/background_watch.rs b/src-tauri/src/acp/background_watch.rs
index 82a16057f6..a09c35e6ed 100644
--- a/src-tauri/src/acp/background_watch.rs
+++ b/src-tauri/src/acp/background_watch.rs
@@ -39,13 +39,19 @@
 //!
 //! * **Session title** — Claude Code's generated name arrives as a dedicated
 //!   `ai-title` transcript record whenever the background summarizer finishes,
-//!   routinely AFTER the turn that triggered it ended. The ACP adapter only
-//!   pulls the name at turn-end (`maybeUpdateSessionTitle`, claude-agent-acp
-//!   0.69.0), so on a short session there is nothing to read yet and no wire
-//!   event ever follows. These bytes are already being tailed, so the records
-//!   are folded here and handed to [`publish_native_title`] — the same path a
-//!   live ACP title takes. Not activity: it rides alongside the activity event
-//!   rather than inside it (see `run_watch`).
+//!   routinely AFTER the turn that triggered it ended. Through claude-agent-acp
+//!   0.70.0 the adapter only pulled the name at turn-end
+//!   (`maybeUpdateSessionTitle`), so on a short session there was nothing to
+//!   read yet and no wire event ever followed. 0.71.0 (#984) asks the CLI to
+//!   generate one at turn-end instead (`SessionTitles`, a background
+//!   `generate_session_title` control request), so a wire title now usually
+//!   does follow — but generation needs ten characters of collected context
+//!   and can return null, so this fold stays the backstop. These bytes are
+//!   already being tailed, so the records are folded here and handed to
+//!   [`publish_native_title`] — the same skip-cached path a live ACP title
+//!   takes, which also absorbs the overlap when both fire. Not activity: it
+//!   rides alongside the activity event rather than inside it (see
+//!   `run_watch`).
 //!
 //! The watcher is connection-scoped on purpose: background work cannot outlive
 //! the agent CLI process, whose lifetime IS the connection's. Poll ticks are
diff --git a/src-tauri/src/acp/connection.rs b/src-tauri/src/acp/connection.rs
index 581e190642..aace8f6526 100644
--- a/src-tauri/src/acp/connection.rs
+++ b/src-tauri/src/acp/connection.rs
@@ -3737,18 +3737,22 @@ fn build_client_capabilities(
     //
     // codex-acp 1.7.0 added a third, "nativeSubagentSessions" (the draft ACP
     // subagent RFD; the canonical gate is a `clientCapabilities.subagents: {}`
-    // field, with this AIR key as the fallback for SDKs that strip it). It must
+    // field, with this AIR key as the fallback for SDKs that strip it);
+    // claude-agent-acp 0.71.0 speaks it too, plus a fourth, "asyncTasks"
+    // (`async_task_spawned` / `async_task_progress` /
+    // `async_task_state_update`), gated the same way. They must
     // stay out for a harder reason than cost: `agent-client-protocol-schema`
-    // 0.11.7 cannot RECEIVE the result. Its `SessionUpdate` is an
+    // 0.11.7 cannot RECEIVE the results. Its `SessionUpdate` is an
     // internally-tagged enum with no catch-all arm, so the `subagent_spawned` /
     // `subagent_state_update` notifications would fail to deserialize — and
-    // since the adapter switches child messages, thoughts, tools and
+    // since the adapters switch child messages, thoughts, tools and
     // permissions onto a child session id announced only in that first
     // notification, opting in would make subagent work vanish from the timeline
     // rather than render better. Without the advertisement the lifecycle stays
-    // the legacy `subAgentActivity` tool call codeg already renders, whose
-    // shape is unchanged from 1.4.0. Revisit when the schema crate ships both
-    // the capability field and the update variants.
+    // what codeg already renders — codex's legacy `subAgentActivity` tool call
+    // (unchanged from 1.4.0) and claude's `subagent-transcript` streaming
+    // (verified intact through 0.73.0). Revisit when the schema crate ships
+    // both the capability field and the update variants.
     if matches!(agent_type, AgentType::ClaudeCode | AgentType::Codex) {
         meta.insert(
             "jetbrains".to_string(),
@@ -10224,8 +10228,10 @@ fn is_codex_plan_review(
 /// `_meta.permission = {version: 1, title, description?}`. The title is now one
 /// of four fixed strings and the reason lives only in `description`, so a card
 /// built from the tool call alone would read "Edit files" where it used to
-/// explain WHY the edit needs approval. claude-agent-acp does not send this
-/// block; nothing changes for it.
+/// explain WHY the edit needs approval. claude-agent-acp joined with its
+/// 0.71.0 permission rebuild — same `{version: 1, title, description?}`, the
+/// SDK's decisionReason riding in `description` as "Reason: …" — so the hoist
+/// serves both built-ins.
 ///
 /// Hoisting rather than adding an event field is deliberate: the tool call is
 /// already the card's payload end-to-end (`PendingPermissionState.tool_call`,
@@ -13323,14 +13329,17 @@ mod tests {
                 .iter()
                 .any(|v| v.as_str() == Some("sessionFailure")));
             // And nothing else. Adding a capability here is not free — it is
-            // what turns the corresponding behavior on, and neither of the two
+            // what turns the corresponding behavior on, and none of the three
             // that exist is wanted: "agentFileChangeReport"
             // (claude-agent-acp 0.69.0 / codex-acp 1.4.0) buys an extra model
             // round-trip per turn for a clamped, self-reported subset of what
-            // the `workspace_state` watcher already sees, and
-            // "nativeSubagentSessions" (codex-acp 1.7.0) would move subagent
+            // the `workspace_state` watcher already sees;
+            // "nativeSubagentSessions" (codex-acp 1.7.0 / claude-agent-acp
+            // 0.71.0) would move subagent
             // output onto child session ids carried by `SessionUpdate` variants
-            // `agent-client-protocol-schema` 0.11.7 cannot deserialize at all.
+            // `agent-client-protocol-schema` 0.11.7 cannot deserialize at all;
+            // and "asyncTasks" (claude-agent-acp 0.71.0) rides the same
+            // undeliverable `async_task_*` variants.
             // See the reasoning at the advertisement site before relaxing this.
             assert_eq!(
                 capabilities,
diff --git a/src-tauri/src/acp/preflight.rs b/src-tauri/src/acp/preflight.rs
index 3a85bff46f..819e2d17e0 100644
--- a/src-tauri/src/acp/preflight.rs
+++ b/src-tauri/src/acp/preflight.rs
@@ -51,7 +51,7 @@ pub struct CheckItem {
 /// `None` on [`PreflightResult`] for every non-adapter agent.
 #[derive(Debug, Clone, Serialize)]
 pub struct AdapterInfo {
-    /// npm spec codeg installs, e.g. "@agentclientprotocol/claude-agent-acp@0.69.0".
+    /// npm spec codeg installs, e.g. "@agentclientprotocol/claude-agent-acp@0.73.0".
     pub adapter_package: String,
     /// Command the launch gate resolves, e.g. "claude-agent-acp".
     pub adapter_cmd: String,
@@ -724,7 +724,7 @@ mod adapter_tests {
         );
         assert_eq!(
             info.adapter_package,
-            "@agentclientprotocol/claude-agent-acp@0.69.0"
+            "@agentclientprotocol/claude-agent-acp@0.73.0"
         );
         assert_eq!(info.adapter_cmd, "claude-agent-acp");
         assert!(!info.adapter_installed);
diff --git a/src-tauri/src/acp/registry.rs b/src-tauri/src/acp/registry.rs
index bab516976d..9ba3a8793c 100644
--- a/src-tauri/src/acp/registry.rs
+++ b/src-tauri/src/acp/registry.rs
@@ -577,9 +577,112 @@ pub fn get_agent_meta(agent_type: AgentType) -> AcpAgentMeta {
             // can advertise `["sessionFailure", "agentFileChangeReport"]` — an
             // ADDITIVE element in an array codeg only ever membership-tests,
             // so the session-failure gate is unaffected.
+            // 0.70.0 (#1002) is provider plumbing only: the
+            // `providers/list|set|disable` methods and the `providers: {}`
+            // advertisement (both already in 0.69.0) now also cover LOADED
+            // sessions — a switch respawns the SDK behind the live session
+            // with a rebuilt routing env (base-URL/Bedrock/Vertex vars,
+            // local credential checks bypassed via "acp-proxy" sentinels).
+            // All of it sits inside client-called requests codeg never
+            // sends; dependencies, the session_update surface and the Node
+            // floor are unchanged (tarball-diffed — the ambient additions
+            // are a "Claude ACP started" boot line and provider-resolution
+            // logging).
+            // 0.71.0 (nine PRs; claude-agent-sdk 0.3.232→0.3.238, which
+            // vendors runtime 2.1.238; zod floor raised to ^4, #1057; dist
+            // split into modules) adds five session_update KINDS, every one
+            // gated on a client advertisement codeg does not make:
+            // `subagent_spawned`/`subagent_state_update` (#1017) ride only
+            // for a client advertising `clientCapabilities.subagents`
+            // (inexpressible in `agent-client-protocol-schema` 0.11.7) or
+            // AIR `nativeSubagentSessions`, and `async_task_spawned`/
+            // `_progress`/`_state_update` only behind AIR `asyncTasks` —
+            // codeg's AIR array is pinned to exactly ["sessionFailure"]
+            // (tested), which the no-catch-all `SessionUpdate` argument in
+            // codex-acp 1.7.0's NOT-adopted paragraph makes load-bearing:
+            // the same "would fail to deserialize" logic applies verbatim.
+            // Without the advertisements the legacy shape is intact
+            // (tarball-verified): `subagent-transcript` streaming,
+            // `_meta.claudeCode.parentToolUseId`, permissions on the root
+            // session. The advertisement array grows to [sessionFailure,
+            // agentFileChangeReport, nativeSubagentSessions, asyncTasks] and
+            // `sessionCapabilities` gains `subagents: {}` — additive shapes
+            // codeg membership-tests or ignores. What codeg's wire DOES
+            // gain: (a) real session titles (#984) — the CLI's own
+            // auto-titler never arms under the Agent SDK, so the old
+            // turn-end pull mostly republished the raw first prompt; the
+            // adapter now asks for a generated title (a background
+            // `generate_session_title` control request, once per session,
+            // needs ≥10 chars of context, a user /rename always wins) over
+            // the SAME `session_info_update {title, updatedAt}` wire the
+            // SessionInfoUpdate arm already feeds to `publish_native_title`;
+            // the transcript-tail fallback in background_watch.rs stays as
+            // the backstop. (b) The permission layer is rebuilt
+            // (`src/permissions/*`): the option-level `_meta.permission =
+            // {version, changes[]}` contract from 0.64.1/#930 is GONE, and
+            // claude joins codex-acp 1.7.0's REQUEST-level
+            // `_meta.permission = {version: 1, title, description?}`
+            // (description = "Reason: <decisionReason>") — exactly what
+            // `hoist_request_permission_meta` forwards, so the card's reason
+            // line lights up for claude with no code change while
+            // `parsePermissionOptionChanges` simply no longer fires for it.
+            // ExitPlanMode approval gains clear-context lanes (#1004): new
+            // opaque option ids under standard kinds, and accepting one
+            // swaps the underlying SDK session while the SAME ACP turn stays
+            // in flight (`current_mode_update` + `config_option_update`
+            // mid-turn, then the plan continuation streams on) — transparent
+            // to codeg's turn-scoped runtime by construction. (c) Modes
+            // (#1025): the advertised list is now default/acceptEdits/plan/
+            // auto (+ bypassPermissions behind the adapter's ALLOW_BYPASS),
+            // `dontAsk` is still parsed but no longer advertised, and each
+            // mode/option carries `_meta.kind` (standard/plan/auto_review/
+            // full_access) — additive, unread. (d) The prompt response gains
+            // `_meta.quota = {token_count, model_usage[]}` (#1037),
+            // deliberately codex-shaped — unread here, input for a future
+            // usage surface. (e) `session/fork` (registered since 0.69.0)
+            // honors an AIR fork point (`_meta.jetbrains.air.fork.messageId`,
+            // #1046) — client-called; codeg doesn't. (f) #1045: a
+            // `_session/steering` arriving while a permission/elicitation
+            // card is open is delivered at non-interrupting priority instead
+            // of aborting the running cycle into the open prompt; the
+            // `promptRequired` contract is unchanged (tarball-verified), so
+            // the 0.65.0 floor keeps holding, just safer. The Windows
+            // `PowerShell` shell tool is now recognized alongside Bash (same
+            // description-title treatment).
+            // 0.72.0 moves both SDKs (@agentclientprotocol/sdk 1.3.0→1.4.0,
+            // claude-agent-sdk 0.3.238→0.3.252 = runtime 2.1.252) and rides
+            // the new runtime in #1065: a PostModelSwitch hook (fires on CLI
+            // 2.1.251+ only) mirrors switches the adapter didn't drive — a
+            // `/model x` typed as a prompt — back into the picker through
+            // the EXISTING `config_option_update`/`current_mode_update`
+            // surfaces; PreModelSwitch hooks can veto (`setModel` then fails
+            // with "Model switch blocked by a PreModelSwitch hook"); effort
+            // now seeds from the CLI's persisted per-model settings unless
+            // the user pinned it via the picker this session; and results
+            // echo the triggering send's `user_message_uuid` (SDK 0.3.246+),
+            // upgrading the adapter's internal orphan-result heuristics to
+            // an exact join. The ACP SDK bump is API-surface only:
+            // `unstable_createElicitation`/`unstable_completeElicitation`
+            // lose their prefixes while the wire methods stay
+            // `elicitation/create`/`elicitation/complete` (schema method
+            // tables diffed byte-identical) — moot here anyway, codeg
+            // advertises elicitation for codex and deepseek only.
+            // 0.73.0 (#1066) is the runtime and nothing else: dist/ is
+            // byte-identical to 0.72.0's and the only manifest change is
+            // claude-agent-sdk 0.3.252→0.3.257 (runtime 2.1.257). That is
+            // the point of this pin: model availability ships in the
+            // runtime — the API refuses new model ids on old ones ("Claude
+            // Code 2.1.247 does not support this model; version 2.1.251 or
+            // newer is required" for `claude-fable-5-1`, and the old pin's
+            // SDK 0.3.232 vendors 2.1.232) — so from 0.72.0 on the dynamic
+            // model config option lists new Anthropic models with no codeg
+            // change (verified live over stdio: 0.73.0 initializes and
+            // session/new's model option offers Default/Opus 1M/Fable/
+            // Sonnet/Haiku). `engines.node` stays ">=22" across 0.70–0.73,
+            // so `node_required` holds.
             distribution: AgentDistribution::Npx {
-                version: "0.69.0",
-                package: "@agentclientprotocol/claude-agent-acp@0.69.0",
+                version: "0.73.0",
+                package: "@agentclientprotocol/claude-agent-acp@0.73.0",
                 cmd: "claude-agent-acp",
                 args: &[],
                 env: &[],
@@ -734,8 +837,11 @@ pub fn get_agent_meta(agent_type: AgentType) -> AcpAgentMeta {
             // standard ACP fields `parse_permission_tool_call` already reads
             // (`rawInput.command/cwd/url/additionalPermissions`, `locations`,
             // `content`). `handle_permission_request` hoists the request meta
-            // onto the card so the reason survives; claude-agent-acp 0.69.0
-            // still emits `changes[]`, so that parser stays. Codex's option
+            // onto the card so the reason survives; claude-agent-acp joined
+            // this request-level contract in 0.71.0 (its `changes[]` ended at
+            // 0.70.0), so the option-level parser now serves codex's
+            // description-only MCP-approval options and any custom agent
+            // still speaking `changes[]`. Codex's option
             // IDs were also renamed (`allow_for_session`,
             // `accept_execpolicy_amendment`, `apply_network_policy_amendment:N`
             // …) — inert here, codeg only echoes back the selected id — and
@@ -1616,8 +1722,8 @@ mod tests {
     fn registry_pins_current_acp_agent_versions() {
         assert_npx_version(
             AgentType::ClaudeCode,
-            "0.69.0",
-            "@agentclientprotocol/claude-agent-acp@0.69.0",
+            "0.73.0",
+            "@agentclientprotocol/claude-agent-acp@0.73.0",
             Some("22.0.0"),
         );
         assert_npx_version(