Live Stream: add compression recovery action

api/compression_recovery.py

@@ -0,0 +1,111 @@
+"""Compression-exhausted recovery metadata and intent helpers."""
+
+from __future__ import annotations
+
+import re
+import time
+from typing import Any
+
+
+COMPRESSION_RECOVERY_TERMINAL_STATE = "compression_exhausted"
+COMPRESSION_RECOVERY_ACTION_START_FOCUSED = "start_focused_continuation"
+
+
+_GENERIC_CONTINUATION_INTENTS = frozenset(
+    {
+        "continue",
+        "continue please",
+        "go on",
+        "keep going",
+        "resume",
+        "proceed",
+        "carry on",
+        "继续",
+        "继续吧",
+        "接着",
+        "接着做",
+        "继续做",
+        "继续执行",
+    }
+)
+
+
+def _positive_int(value: Any) -> int:
+    try:
+        parsed = int(value or 0)
+    except (TypeError, ValueError):
+        return 0
+    return parsed if parsed > 0 else 0
+
+
+def _normalize_intent_text(text: str) -> str:
+    raw = str(text or "").strip().lower()
+    if not raw:
+        return ""
+    # Keep CJK letters while dropping punctuation/emoji/noise around short intents.
+    return re.sub(r"[\W_]+", " ", raw, flags=re.UNICODE).strip()
+
+
+def is_generic_continuation_intent(text: str) -> bool:
+    """Return True only for short, content-free continuation requests."""
+
+    normalized = _normalize_intent_text(text)
+    if normalized in _GENERIC_CONTINUATION_INTENTS:
+        return True
+    # A single repeated word such as "continue continue" is still generic, but
+    # longer prompts like "continue by summarizing file X" are substantive.
+    parts = normalized.split()
+    return bool(parts) and len(parts) <= 2 and all(part in _GENERIC_CONTINUATION_INTENTS for part in parts)
+
+
+def build_compression_recovery_payload(session, *, message: str = "", details: str = "") -> dict:
+    """Build the durable UI/route payload for a compression-exhausted turn."""
+
+    source_sid = str(getattr(session, "session_id", "") or "")
+    last_prompt_tokens = _positive_int(getattr(session, "last_prompt_tokens", None))
+    threshold_tokens = _positive_int(getattr(session, "threshold_tokens", None))
+    context_length = _positive_int(getattr(session, "context_length", None))
+    payload = {
+        "terminal_state": COMPRESSION_RECOVERY_TERMINAL_STATE,
+        "recommended_action": COMPRESSION_RECOVERY_ACTION_START_FOCUSED,
+        "source_session_id": source_sid,
+        "created_at": time.time(),
+        "title": "Context compression exhausted",
+        "summary": (
+            "This run could not safely shrink the conversation enough to continue in place. "
+            "Start a focused continuation, then describe the next narrow task."
+        ),
+        "action_label": "Start focused continuation",
+        "message": str(message or "").strip(),
+        "details": str(details or "").strip()[:1200],
+        "last_prompt_tokens": last_prompt_tokens,
+        "threshold_tokens": threshold_tokens,
+        "context_length": context_length,
+    }
+    return payload
+
+
+def stamp_compression_exhausted_recovery(session, *, message: str = "", details: str = "") -> dict:
+    """Persist recovery metadata on a session and return the message payload."""
+
+    payload = build_compression_recovery_payload(session, message=message, details=details)
+    session.recommended_recovery_action = payload["recommended_action"]
+    session.compression_recovery = payload
+    return payload
+
+
+def compression_recovery_payload_for_session(session) -> dict | None:
+    payload = getattr(session, "compression_recovery", None)
+    if not isinstance(payload, dict):
+        return None
+    if payload.get("terminal_state") != COMPRESSION_RECOVERY_TERMINAL_STATE:
+        return None
+    action = str(payload.get("recommended_action") or getattr(session, "recommended_recovery_action", "") or "")
+    if action != COMPRESSION_RECOVERY_ACTION_START_FOCUSED:
+        return None
+    return payload
+
+
+def clear_compression_recovery(session) -> None:
+    session.recommended_recovery_action = None
+    session.compression_recovery = {}

api/models.py

@@ -1051,6 +1051,8 @@ class Session:
                  context_engine_state=None,
                  context_length=None, threshold_tokens=None,
                  last_prompt_tokens=None,
+                 compression_recovery=None,
+                 recommended_recovery_action=None,
                  truncation_watermark=None,
                  truncation_boundary=None,
                  gateway_routing=None, gateway_routing_history=None,
@@ -1102,6 +1104,8 @@ class Session:
         self.context_length = context_length
         self.threshold_tokens = threshold_tokens
         self.last_prompt_tokens = last_prompt_tokens
+        self.compression_recovery = compression_recovery if isinstance(compression_recovery, dict) else {}
+        self.recommended_recovery_action = recommended_recovery_action
         self.truncation_watermark = truncation_watermark
         self.truncation_boundary = truncation_boundary
         self.gateway_routing = gateway_routing if isinstance(gateway_routing, dict) else None
@@ -1172,6 +1176,7 @@ class Session:
             'context_engine', 'compression_anchor_engine', 'compression_anchor_mode',
             'compression_anchor_details', 'context_engine_state',
             'context_length', 'threshold_tokens', 'last_prompt_tokens',
+            'compression_recovery', 'recommended_recovery_action',
             'truncation_watermark',
             'truncation_boundary',
             'gateway_routing', 'gateway_routing_history', 'llm_title_generated', 'manual_title',
@@ -1419,6 +1424,8 @@ class Session:
             'context_length': self.context_length,
             'threshold_tokens': self.threshold_tokens,
             'last_prompt_tokens': self.last_prompt_tokens,
+            'compression_recovery': self.compression_recovery,
+            'recommended_recovery_action': self.recommended_recovery_action,
             'gateway_routing': self.gateway_routing,
             'gateway_routing_history': self.gateway_routing_history,
             'manual_title': self.manual_title,

api/routes.py

@@ -42,6 +42,12 @@ from api.agent_sessions import (
     read_session_lineage_report,
 )
 from api.compression_anchor import visible_messages_for_anchor
+from api.compression_recovery import (
+    COMPRESSION_RECOVERY_ACTION_START_FOCUSED,
+    clear_compression_recovery,
+    compression_recovery_payload_for_session,
+    is_generic_continuation_intent,
+)
 from api.session_events import (
     add_session_list_changed_listener,
     publish_session_list_changed,
@@ -13099,6 +13105,9 @@ def handle_post(handler, parsed) -> bool:
             )
         return j(handler, {"session": s.compact() | {"messages": s.messages}})
 
+    if parsed.path == "/api/session/compression-recovery/start":
+        return _handle_session_compression_recovery_start(handler, body)
+
     if parsed.path == "/api/session/duplicate":
         try:
             sid = body.get("session_id")
@@ -19724,6 +19733,89 @@ def _handle_bg_task_complete_ack(handler, body):
     )
 
 
+def _handle_session_compression_recovery_start(handler, body):
+    try:
+        require(body, "session_id")
+    except ValueError as e:
+        return bad(handler, str(e))
+    sid = str(body.get("session_id") or "").strip()
+    if not sid:
+        return bad(handler, "session_id is required")
+    if _session_is_subagent_view_only(sid):
+        return bad(handler, "Subagent sessions are view-only and cannot start compression recovery from WebUI", 400)
+    try:
+        source = get_session(sid)
+    except KeyError:
+        return bad(handler, "Session not found", 404)
+    if not _session_visible_to_active_profile(getattr(source, "profile", None), handler):
+        return bad(handler, "Session not found", 404)
+    recovery = compression_recovery_payload_for_session(source)
+    if not recovery:
+        return bad(handler, "Session does not have a compression recovery action.", 409)
+    action = str(recovery.get("recommended_action") or "")
+    if action != COMPRESSION_RECOVERY_ACTION_START_FOCUSED:
+        return bad(handler, "Unsupported compression recovery action.", 409)
+
+    title = str(getattr(source, "title", None) or "Untitled").strip() or "Untitled"
+    if not title.endswith(" (focused continuation)"):
+        title = f"{title} (focused continuation)"
+    copied_session = Session(
+        session_id=uuid.uuid4().hex[:12],
+        title=title,
+        workspace=getattr(source, "workspace", get_last_workspace()),
+        model=getattr(source, "model", None),
+        model_provider=getattr(source, "model_provider", None),
+        messages=[],
+        tool_calls=[],
+        pinned=False,
+        archived=False,
+        project_id=getattr(source, "project_id", None),
+        profile=getattr(source, "profile", None),
+        personality=getattr(source, "personality", None),
+        enabled_toolsets=copy.deepcopy(getattr(source, "enabled_toolsets", None)),
+        context_length=getattr(source, "context_length", None),
+        threshold_tokens=getattr(source, "threshold_tokens", None),
+        gateway_routing=copy.deepcopy(getattr(source, "gateway_routing", None)),
+        gateway_routing_history=copy.deepcopy(getattr(source, "gateway_routing_history", None) or []),
+        parent_session_id=getattr(source, "session_id", sid),
+        worktree_path=getattr(source, "worktree_path", None),
+        worktree_branch=getattr(source, "worktree_branch", None),
+        worktree_repo_root=getattr(source, "worktree_repo_root", None),
+        worktree_created_at=getattr(source, "worktree_created_at", None),
+    )
+    # Preserve the workspace/model/profile lane, but intentionally start with an
+    # empty model-facing transcript so a focused follow-up does not replay the
+    # exhausted state.db/context tail.
+    copied_session.context_messages = []
+    copied_session.composer_draft = {"text": "", "files": []}
+    try:
+        copied_session.save()
+    except Exception as e:
+        logger.exception("failed to persist compression recovery session for %s", sid)
+        return bad(handler, f"Failed to start compression recovery: {_sanitize_error(e)}", 500)
+
+    with LOCK:
+        SESSIONS[copied_session.session_id] = copied_session
+        SESSIONS.move_to_end(copied_session.session_id)
+        _evict_sessions_over_cap()
+    publish_session_list_changed(
+        "session_compression_recovery",
+        profile=getattr(copied_session, "profile", None),
+        session_id=getattr(copied_session, "session_id", None),
+    )
+    session_payload = redact_session_data(copied_session.compact() | {"messages": copied_session.messages})
+    return j(
+        handler,
+        {
+            "ok": True,
+            "session": session_payload,
+            "source_session_id": sid,
+            "recommended_recovery_action": action,
+            "message": "Started a focused continuation. Describe the next narrow task to continue.",
+        },
+    )
+
+
 def _handle_goal_command(handler, body):
     """Handle WebUI /goal command controls and optional kickoff stream."""
     try:
@@ -19979,6 +20071,19 @@ def _handle_chat_start(handler, body, diag=None):
             return bad(handler, "message is required")
         diag.stage("normalize_attachments") if diag else None
         attachments = _normalize_chat_attachments(body.get("attachments") or [])[:20]
+        recovery = compression_recovery_payload_for_session(s)
+        if recovery and not attachments and is_generic_continuation_intent(msg):
+            return j(
+                handler,
+                {
+                    "error": "This session exhausted context compression. Start a focused continuation, then describe the next narrow task.",
+                    "type": "compression_recovery_required",
+                    "recommended_recovery_action": recovery.get("recommended_action"),
+                    "compression_recovery": recovery,
+                    "session_id": getattr(s, "session_id", body["session_id"]),
+                },
+                status=409,
+            )
         diag.stage("resolve_workspace") if diag else None
         try:
             workspace = _resolve_chat_workspace_with_recovery(s, body.get("workspace"))
@@ -20038,16 +20143,43 @@ def _handle_chat_start(handler, body, diag=None):
         }
         if not gateway_chat_enabled and moa_config is not None:
             start_run_kwargs["moa_config"] = moa_config
-        response = _start_run(
-            s,
-            **start_run_kwargs,
-        )
+        recovery_cleared_for_start = None
+        def _restore_cleared_recovery():
+            if recovery_cleared_for_start is None:
+                return None
+            s.compression_recovery = recovery_cleared_for_start
+            s.recommended_recovery_action = recovery_cleared_for_start.get("recommended_action")
+            try:
+                s.save()
+            except Exception as restore_err:
+                logger.exception("failed to restore compression recovery after chat start rejection for %s", getattr(s, "session_id", None))
+                return restore_err
+            return None
+
+        if recovery:
+            recovery_cleared_for_start = copy.deepcopy(recovery)
+            clear_compression_recovery(s)
+        try:
+            response = _start_run(
+                s,
+                **start_run_kwargs,
+            )
+        except Exception:
+            _restore_cleared_recovery()
+            raise
         # Map adapter-selection NotImplementedError (501) onto the legacy
         # bad-request response shape that this route exposed historically
         # before the helper extraction.
         if response.get("_status") == 501 and "error" in response:
+            restore_err = _restore_cleared_recovery()
+            if restore_err is not None:
+                return bad(handler, f"failed to restore compression recovery: {_sanitize_error(restore_err)}", 500)
             return j(handler, {"error": response["error"]}, status=501)
         status = int(response.pop("_status", 200) or 200)
+        if status >= 400 and recovery_cleared_for_start is not None:
+            restore_err = _restore_cleared_recovery()
+            if restore_err is not None:
+                return bad(handler, f"failed to restore compression recovery: {_sanitize_error(restore_err)}", 500)
         diag.stage("response_write") if diag else None
         return j(handler, response, status=status)
     finally:

api/session_recovery.py

@@ -484,6 +484,8 @@ def _state_db_row_to_sidecar(row: dict) -> dict:
         'context_length': None,
         'threshold_tokens': None,
         'last_prompt_tokens': None,
+        'compression_recovery': {},
+        'recommended_recovery_action': None,
         'gateway_routing': None,
         'gateway_routing_history': [],
         'llm_title_generated': False,

api/streaming.py

@@ -44,6 +44,7 @@ from api.config import (
 )
 from api.helpers import redact_session_data, _redact_text
 from api.compression_anchor import is_context_compression_marker, visible_messages_for_anchor
+from api.compression_recovery import stamp_compression_exhausted_recovery
 from api.metering import meter
 from api.run_journal import RunJournalWriter
 from api.todo_state import attach_todo_state, emit_todo_state
@@ -8569,6 +8570,15 @@ def _run_agent_streaming(
                             'timestamp': int(time.time()),
                             '_error': True,
                         }
+                        if _err_type == 'compression_exhausted':
+                            _recovery = stamp_compression_exhausted_recovery(
+                                s,
+                                message=_error_payload.get('message') or _err_label,
+                                details=_error_payload.get('details') or '',
+                            )
+                            _error_message['_compressionRecovery'] = _recovery
+                            _error_payload['compression_recovery'] = _recovery
+                            _error_payload['recommended_recovery_action'] = _recovery.get('recommended_action')
                         if _error_payload.get('details'):
                             _error_message['provider_details'] = _error_payload['details']
                         if _err_type == 'cancelled':
@@ -9461,6 +9471,7 @@ def _run_agent_streaming(
         _exc_is_not_found = _classification['type'] == 'model_not_found'  # detects '404', 'not found', 'does not exist', and 'invalid model'.
         _exc_is_cancelled = _classification['type'] == 'cancelled'
         _exc_is_interrupted = _classification['type'] == 'interrupted'
+        _exc_is_compression_exhausted = _classification['type'] == 'compression_exhausted'
 
         # The user hint still points to Settings / `hermes model` from _classify_provider_error().
         if _exc_is_quota:
@@ -9581,6 +9592,10 @@ def _run_agent_streaming(
             _exc_label, _exc_type, _exc_hint = (
                 _classification['label'], _classification['type'], _classification['hint'],
             )
+        elif _exc_is_compression_exhausted:
+            _exc_label, _exc_type, _exc_hint = (
+                _classification['label'], _classification['type'], _classification['hint'],
+            )
         else:
             _exc_label, _exc_type, _exc_hint = 'Error', 'error', ''
 
@@ -9620,6 +9635,15 @@ def _run_agent_streaming(
                     'timestamp': int(time.time()),
                     '_error': True,
                 }
+                if _exc_type == 'compression_exhausted':
+                    _recovery = stamp_compression_exhausted_recovery(
+                        s,
+                        message=_error_payload.get('message') or err_str,
+                        details=_error_payload.get('details') or '',
+                    )
+                    _error_message['_compressionRecovery'] = _recovery
+                    _error_payload['compression_recovery'] = _recovery
+                    _error_payload['recommended_recovery_action'] = _recovery.get('recommended_action')
                 if _error_payload.get('details'):
                     _error_message['provider_details'] = _error_payload['details']
                 if _exc_type == 'cancelled':

api/webui_session_db.py

@@ -51,6 +51,8 @@ _METADATA_FIELDS = frozenset(
         "context_length",
         "threshold_tokens",
         "last_prompt_tokens",
+        "compression_recovery",
+        "recommended_recovery_action",
         "truncation_watermark",
         "truncation_boundary",
         "gateway_routing",

docs/rfcs/session-sse-contract-v1.md

@@ -38,8 +38,8 @@ against current source before any route is added.
 - This RFC does **not** implement `GET /api/sessions/{session_id}/events`. No
   route, handler, or related code is added in this PR.
 - This RFC does **not** modify `GET /api/sessions/events` (the existing global
-  session-list invalidation stream routed at `api/routes.py:12345-12346` and
-  implemented by `_handle_session_events_stream()` at `api/routes.py:16177`).
+  session-list invalidation stream routed at `api/routes.py:12351-12352` and
+  implemented by `_handle_session_events_stream()` at `api/routes.py:16186`).
 - This RFC does **not** replace or modify existing streams: `/api/chat/stream`,
   `/api/approval/stream`, or `/api/clarify/stream`.
 - This RFC does **not** introduce Android, iOS, or PWA client code.
@@ -53,8 +53,8 @@ against current source before any route is added.
 ### Existing global session-list stream
 
 `GET /api/sessions/events` is a **different endpoint** from the one this RFC
-proposes. It is routed at `api/routes.py:12345-12346` and implemented by
-`_handle_session_events_stream()` at `api/routes.py:16177`. It emits bare
+proposes. It is routed at `api/routes.py:12351-12352` and implemented by
+`_handle_session_events_stream()` at `api/routes.py:16186`. It emits bare
 `sessions_changed` events and keepalives for any change to the session list. It
 is a global invalidation signal, not a per-session lifecycle stream. The proposed
 `GET /api/sessions/{session_id}/events` is per-session and path-distinct.
@@ -73,19 +73,19 @@ Line ranges in this inventory were verified against WebUI `master` when this
 RFC was written. Function, constant, and endpoint names are the stable anchors
 if source layout moves later.
 
-- `_parse_run_journal_event_id()` (`api/routes.py:15673-15686`) and
-  `_parse_run_journal_after_seq()` (`api/routes.py:15688-15701`) parse the replay
+- `_parse_run_journal_event_id()` (`api/routes.py:15682-15695`) and
+  `_parse_run_journal_after_seq()` (`api/routes.py:15697-15710`) parse the replay
   cursor from the `after_event_id` / `after_seq` **query params** (not the
   `Last-Event-ID` header — that header is the *proposed* new-endpoint contract
   below, §Reconnect).
-- `_runner_event_id()` at `api/routes.py:15765-15772` constructs the event `id`
+- `_runner_event_id()` at `api/routes.py:15774-15781` constructs the event `id`
   field as `stream_id:seq`.
 - SSE frames carry their `id:` via the `_sse_with_id()` helper, emitted on the
-  live `/api/chat/stream` path at `api/routes.py:15918`, on the runner-observe
-  path at `api/routes.py:15811`, and during journal replay at
-  `api/routes.py:15721` / `15734`.
+  live `/api/chat/stream` path at `api/routes.py:15927`, on the runner-observe
+  path at `api/routes.py:15820`, and during journal replay at
+  `api/routes.py:15730` / `15743`.
 - `_replay_run_journal()` reads events by `(session_id, stream_id)` at
-  `api/routes.py:15703-15735`.
+  `api/routes.py:15712-15744`.
 - `api/streaming.py:6265-6285` writes current live agent streams to
   `STREAMS[stream_id]`.
 - `api/streaming.py:6620-6634` appends SSE events to the run journal and carries
@@ -165,7 +165,7 @@ position.
 
 **`event_id` is opaque to clients.** Its current source-compatible form is
 `stream_id:seq`, as constructed by `_runner_event_id()` at
-`api/routes.py:15765-15772`. Clients must treat it as an opaque string and must
+`api/routes.py:15774-15781`. Clients must treat it as an opaque string and must
 not parse or construct cursor values.
 
 **`seq` is monotonic within a stream/run.** It is not a session-global counter
@@ -183,7 +183,7 @@ events. The live `STREAMS[stream_id]` queue (`api/streaming.py:6265-6285`) is
 not a reliable replay source because it holds only recent in-memory state.
 
 A future implementation must replay from the run journal via the existing
-`_replay_run_journal()` path (`api/routes.py:15703-15735`) and fall back to the
+`_replay_run_journal()` path (`api/routes.py:15712-15744`) and fall back to the
 snapshot mechanism when journal entries are unavailable for a given cursor.
 
 ## Snapshot fallback

docs/troubleshooting.md

@@ -121,6 +121,44 @@ The on-disk locations below assume the default `~/.hermes/webui` state directory
 
 ---
 
+## "Context compression exhausted" after a long-running turn
+
+**Symptom.** A long-running session, often with many tool calls or a small
+context-window model, ends with a `Context compression exhausted` error instead
+of a final answer. The message includes a recovery action labeled `Start focused
+continuation`.
+
+**Why.** Automatic compression could not shrink the current conversation enough
+to continue safely in the same model-facing context. The exhausted session is
+terminal: sending a bare "continue", "go on", or "继续" would usually replay the
+same oversized state and fail again, so the WebUI points the user to a focused
+linked continuation instead.
+
+**Diagnostic.**
+
+1. Open the session JSON under your WebUI state directory, for example:
+   ```bash
+   jq '.recommended_recovery_action, .compression_recovery' \
+     ~/.hermes/webui/sessions/<session_id>.json
+   ```
+2. A recoverable exhausted turn should report:
+   - `recommended_recovery_action: "start_focused_continuation"`
+   - `compression_recovery.terminal_state: "compression_exhausted"`
+   - the final assistant error message carrying `_compressionRecovery`
+
+**Fix.** Use the `Start focused continuation` action in the exhausted message.
+The new linked session preserves the workspace, model, profile, project, and
+toolset lane, but intentionally starts with an empty model-facing transcript so
+the oversized exhausted tail is not replayed. After the new session opens,
+describe the next narrow task explicitly instead of sending a bare continuation.
+
+**When to file a bug.** File a bug if the exhausted message has no recovery
+action, the action creates a session with the old oversized context/messages
+replayed into the model-facing transcript, or a bare "continue" starts another
+turn in the exhausted session instead of being blocked with recovery guidance.
+
+---
+
 ## Other troubleshooting
 
 This document grows over time. If a recurring failure mode isn't covered here yet, add it via PR. The format for each entry: **Symptom → Why → Diagnostic commands → Fix → When to file a bug**.

static/messages.js

@@ -1326,6 +1326,11 @@ async function send(){
   _flushSelectionBlocksToComposer();
   text=$('msg').value.trim();
   if(!text&&!S.pendingFiles.length){_sendInProgress=false;_sendInProgressSid=null;return;}
+  if(typeof shouldInterceptCompressionRecoveryContinuation==='function'&&shouldInterceptCompressionRecoveryContinuation(text,S.pendingFiles)){
+    if(typeof showCompressionRecoveryContinuationHint==='function') showCompressionRecoveryContinuationHint();
+    _sendInProgress=false;_sendInProgressSid=null;
+    return;
+  }
 
   // #5472: snapshot the ORIGINAL user-typed composer state now — before slash
   // rewrites (/moa, bundles) mutate `text` and before uploadPendingFiles()
@@ -5680,7 +5685,8 @@ function attachLiveStream(activeSid, streamId, uploaded=[], options={}){
               if(typeof _setActiveSessionUrl==='function') _setActiveSessionUrl(S.session.session_id);
             }
           } else {
-            S.messages.push({role:'assistant',content:`**${label}:** ${d.message}${hint}`,provider_details:details,provider_details_label:detailsLabel});
+            const recovery=(d.compression_recovery&&typeof d.compression_recovery==='object')?d.compression_recovery:null;
+            S.messages.push({role:'assistant',content:`**${label}:** ${d.message}${hint}`,provider_details:details,provider_details_label:detailsLabel,_compressionRecovery:recovery||undefined});
             _attachProjectedAnchorSceneToLastAssistant(S.messages);
           }
         }catch(_){

static/style.css

@@ -5726,6 +5726,14 @@ main.main > #mainPlugin{display:none;}
 .status-card-label{color:var(--muted);font-weight:550;background:rgba(0,0,0,.015);}
 .status-card-value{color:var(--text);word-break:break-word;font-family:var(--font-mono);}
 @media (max-width:700px){.status-card{margin-left:0;}.status-card-head{flex-direction:column;}.status-card-session-copy{max-width:100%;}.status-card-grid{grid-template-columns:1fr;}.status-card-label{padding-bottom:2px;border-top:1px solid var(--border-subtle);}.status-card-value{padding-top:2px;border-top:0;}}
+.compression-recovery-card{display:flex;align-items:center;justify-content:space-between;gap:14px;margin:12px 0 4px;padding:14px 16px;border:1px solid var(--border-subtle);border-radius:var(--radius-card);background:var(--surface-subtle);box-shadow:0 8px 20px rgba(0,0,0,.04);}
+.compression-recovery-copy{display:flex;flex-direction:column;gap:3px;min-width:0;}
+.compression-recovery-title{color:var(--text);font-size:14px;font-weight:650;line-height:1.35;}
+.compression-recovery-summary{color:var(--muted);font-size:12px;line-height:1.45;}
+.compression-recovery-action{display:inline-flex;align-items:center;justify-content:center;gap:7px;min-height:32px;padding:6px 11px;border:1px solid var(--accent-bg-strong);border-radius:7px;background:var(--accent-bg);color:var(--accent-text);font:inherit;font-size:12px;font-weight:650;white-space:nowrap;cursor:pointer;}
+.compression-recovery-action:hover{background:var(--accent-bg-strong);color:var(--accent-text);}
+.compression-recovery-action:disabled{opacity:.65;cursor:progress;}
+@media (max-width:700px){.compression-recovery-card{flex-direction:column;align-items:stretch;margin-left:0;}.compression-recovery-action{width:100%;}}
 .assistant-segment-anchor { display: none; }
 
 /* ── Classic conversation layout: user right, half-width; assistant left ── */

static/ui.js

@@ -417,6 +417,89 @@ function _statusCardHtml(card){
   </div>`;
 }
 
+function _compressionRecoveryHtml(recovery, sessionId){
+  if(!recovery||typeof recovery!=='object') return '';
+  if(String(recovery.terminal_state||'')!=='compression_exhausted') return '';
+  const action=String(recovery.recommended_action||'');
+  if(action!=='start_focused_continuation') return '';
+  const sid=String(sessionId||recovery.source_session_id||'');
+  const title=String(recovery.title||'Context compression exhausted');
+  const summary=String(recovery.summary||'Start a focused continuation, then describe the next narrow task.');
+  const actionLabel=String(recovery.action_label||'Start focused continuation');
+  const icon=(typeof li==='function')?li('git-branch',14):'';
+  return `<div class="compression-recovery-card" data-compression-recovery-card="1">
+    <div class="compression-recovery-copy">
+      <div class="compression-recovery-title">${esc(title)}</div>
+      <div class="compression-recovery-summary">${esc(summary)}</div>
+    </div>
+    <button class="compression-recovery-action" type="button" data-recovery-session-id="${esc(sid)}" onclick="startCompressionRecovery(this);event.stopPropagation()">${icon}<span>${esc(actionLabel)}</span></button>
+  </div>`;
+}
+
+function _activeCompressionRecoveryPayload(){
+  if(!S||!S.session) return null;
+  const recovery=S.session.compression_recovery;
+  if(recovery&&typeof recovery==='object'&&String(recovery.terminal_state||'')==='compression_exhausted') return recovery;
+  const messages=Array.isArray(S.messages)?S.messages:[];
+  for(let i=messages.length-1;i>=0;i--){
+    const msg=messages[i];
+    const msgRecovery=msg&&msg._compressionRecovery;
+    if(msgRecovery&&typeof msgRecovery==='object'&&String(msgRecovery.terminal_state||'')==='compression_exhausted') return msgRecovery;
+  }
+  return null;
+}
+
+function isGenericCompressionContinuationIntent(text){
+  const raw=String(text||'').trim().toLowerCase();
+  if(!raw) return false;
+  const normalized=raw.replace(/[^\p{L}\p{N}]+/gu,' ').trim();
+  const generic=new Set(['continue','continue please','go on','keep going','resume','proceed','carry on','继续','继续吧','接着','接着做','继续做','继续执行']);
+  if(generic.has(normalized)) return true;
+  const parts=normalized.split(/\s+/).filter(Boolean);
+  return !!parts.length&&parts.length<=2&&parts.every(part=>generic.has(part));
+}
+
+function shouldInterceptCompressionRecoveryContinuation(text, files){
+  const hasFiles=Array.isArray(files)&&files.length>0;
+  if(hasFiles||!isGenericCompressionContinuationIntent(text)) return false;
+  const recovery=_activeCompressionRecoveryPayload();
+  return !!(recovery&&String(recovery.recommended_action||'')==='start_focused_continuation');
+}
+
+function showCompressionRecoveryContinuationHint(){
+  const card=document.querySelector('[data-compression-recovery-card="1"]');
+  if(card&&typeof card.scrollIntoView==='function'){
+    try{card.scrollIntoView({block:'center',behavior:'smooth'});}catch(_){card.scrollIntoView();}
+    const btn=card.querySelector('.compression-recovery-action');
+    if(btn&&typeof btn.focus==='function') setTimeout(()=>btn.focus(),120);
+  }
+  if(typeof showToast==='function') showToast('This session exhausted context compression. Start a focused continuation, then describe the next narrow task.',4500,'warning');
+}
+
+async function startCompressionRecovery(btn){
+  const sourceSid=String((btn&&btn.dataset&&btn.dataset.recoverySessionId)||(S.session&&S.session.session_id)||'').trim();
+  if(!sourceSid) return;
+  if(btn){btn.disabled=true;btn.classList.add('loading');}
+  try{
+    const data=await api('/api/session/compression-recovery/start',{method:'POST',body:JSON.stringify({session_id:sourceSid})});
+    const sid=data&&data.session&&data.session.session_id;
+    if(sid){
+      try{localStorage.setItem('hermes-webui-session',sid);}catch(_){}
+      if(typeof loadSession==='function') await loadSession(sid,{preserveActiveInput:false});
+      else if(data.session){S.session=data.session;S.messages=data.session.messages||[];syncTopbar();renderMessages();}
+      if(typeof renderSessionList==='function') await renderSessionList();
+      if(typeof _setActiveSessionUrl==='function') _setActiveSessionUrl(sid);
+      if(typeof showToast==='function') showToast((data&&data.message)||'Started focused continuation.',3000,'success');
+      const composer=$('msg');
+      if(composer&&typeof composer.focus==='function') composer.focus();
+    }
+  }catch(e){
+    if(typeof showToast==='function') showToast('Compression recovery failed: '+(e&&e.message||e),5000,'error');
+  }finally{
+    if(btn){btn.disabled=false;btn.classList.remove('loading');}
+  }
+}
+
 const MESSAGE_RENDER_WINDOW_DEFAULT=50;
 const MESSAGE_VIRTUAL_THRESHOLD_ROWS=80;
 const MESSAGE_VIRTUAL_BUFFER_PX=900;
@@ -13390,6 +13473,8 @@ function renderMessages(options){
       const summary=m.provider_details_label||'Provider details';
       bodyHtml += `<details class="provider-error-details"><summary>${esc(String(summary))}</summary><pre><code>${esc(String(m.provider_details))}</code></pre></details>`;
     }
+    const recoveryHtml=(!isUser&&m._compressionRecovery) ? _compressionRecoveryHtml(m._compressionRecovery, (S.session&&S.session.session_id)||'') : '';
+    if(recoveryHtml) bodyHtml += recoveryHtml;
     const statusHtml = (!isUser&&m._statusCard) ? _statusCardHtml(m._statusCard) : '';
     const isEditableUser=isUser&&rawIdx===lastUserRawIdx;
     const editBtn  = isEditableUser ? `<button class="msg-action-btn" title="${t('edit_message')}" onclick="editMessage(this)">${li('pencil',13)}</button>` : '';
@@ -13584,7 +13669,7 @@ function renderMessages(options){
       if((isCompactWorklogMode()||isTransparentStream())&&_assistantThinkingBelongsInWorklog(m, rawIdx, toolCallAssistantIdxs)) assistantThinking.set(rawIdx, thinkingText);
       else if(window._showThinking!==false) seg.insertAdjacentHTML('beforeend', _thinkingCardHtml(thinkingText));
     }
-    const hasVisibleBody=!!(String(content||'').trim()||filesHtml||statusHtml);
+    const hasVisibleBody=!!(String(content||'').trim()||filesHtml||statusHtml||recoveryHtml);
     if(statusHtml){
       seg.insertAdjacentHTML('beforeend', statusHtml);
     }else if(hasVisibleBody){

tests/test_auto_compression_terminal_failure.py

@@ -135,6 +135,9 @@ def test_compression_exhausted_after_session_rotation_preserves_snapshot_and_err
     assert payload["session"]["session_id"] == new_sid
     assert payload["old_session_id"] == old_sid
     assert payload["new_session_id"] == new_sid
+    assert payload["recommended_recovery_action"] == "start_focused_continuation"
+    assert payload["compression_recovery"]["terminal_state"] == "compression_exhausted"
+    assert payload["compression_recovery"]["source_session_id"] == new_sid
 
     old_payload = json.loads((session_dir / f"{old_sid}.json").read_text(encoding="utf-8"))
     new_payload = json.loads((session_dir / f"{new_sid}.json").read_text(encoding="utf-8"))
@@ -144,7 +147,10 @@ def test_compression_exhausted_after_session_rotation_preserves_snapshot_and_err
     assert new_payload["session_id"] == new_sid
     assert new_payload["parent_session_id"] == old_sid
     assert new_payload["pre_compression_snapshot"] is False
+    assert new_payload["recommended_recovery_action"] == "start_focused_continuation"
+    assert new_payload["compression_recovery"]["recommended_action"] == "start_focused_continuation"
     assert new_payload["messages"][-1]["_error"] is True
+    assert new_payload["messages"][-1]["_compressionRecovery"]["recommended_action"] == "start_focused_continuation"
     assert "Context compression exhausted" in new_payload["messages"][-1]["content"]
     assert old_sid not in streaming.SESSIONS
     assert streaming.SESSIONS[new_sid].session_id == new_sid
@@ -428,6 +434,8 @@ def test_apperror_payload_enriched_before_enqueue(tmp_path, monkeypatch):
     assert payload_after["session_id"] == new_sid
     assert payload_after["old_session_id"] == old_sid
     assert payload_after["new_session_id"] == new_sid
+    assert payload_after["recommended_recovery_action"] == "start_focused_continuation"
+    assert payload_after["compression_recovery"]["recommended_action"] == "start_focused_continuation"
 
 
 def test_exception_apperror_payload_includes_session_id_before_enqueue():

tests/test_compression_recovery_action.py

@@ -0,0 +1,248 @@
+import io
+import json
+from pathlib import Path
+
+from api import models, routes
+from api.compression_recovery import (
+    compression_recovery_payload_for_session,
+    is_generic_continuation_intent,
+    stamp_compression_exhausted_recovery,
+)
+from api.models import Session
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+class _JSONHandler:
+    headers = {}
+
+    def __init__(self):
+        self.status = None
+        self.wfile = io.BytesIO()
+        self.headers_sent = {}
+
+    def send_response(self, status):
+        self.status = status
+
+    def send_header(self, key, value):
+        self.headers_sent[key] = value
+
+    def end_headers(self):
+        pass
+
+
+def _payload(handler):
+    raw = handler.wfile.getvalue().decode("utf-8")
+    return json.loads(raw) if raw else {}
+
+
+def _isolate_sessions(monkeypatch, tmp_path):
+    session_dir = tmp_path / "sessions"
+    session_dir.mkdir()
+    monkeypatch.setattr(models, "SESSION_DIR", session_dir)
+    monkeypatch.setattr(models, "SESSION_INDEX_FILE", session_dir / "_index.json")
+    models.SESSIONS.clear()
+    routes.SESSIONS.clear()
+    return session_dir
+
+
+def test_generic_continuation_intent_is_scoped_to_empty_continue_words():
+    assert is_generic_continuation_intent("continue")
+    assert is_generic_continuation_intent("继续吧。")
+    assert is_generic_continuation_intent("go on")
+    assert not is_generic_continuation_intent("continue by summarizing the workspace changes")
+    assert not is_generic_continuation_intent("继续修复 4685 的恢复卡")
+
+
+def test_chat_start_blocks_generic_continue_after_compression_exhausted(monkeypatch, tmp_path):
+    _isolate_sessions(monkeypatch, tmp_path)
+    sid = "recoverychat1"
+    session = Session(
+        session_id=sid,
+        title="Recovery",
+        workspace=str(tmp_path),
+        model="gpt-4o",
+        messages=[{"role": "user", "content": "long task"}],
+    )
+    stamp_compression_exhausted_recovery(session, message="Context length exceeded.")
+    models.SESSIONS[sid] = session
+    routes.SESSIONS[sid] = session
+
+    handler = _JSONHandler()
+    routes._handle_chat_start(handler, {"session_id": sid, "message": "继续"})
+    payload = _payload(handler)
+
+    assert handler.status == 409
+    assert payload["type"] == "compression_recovery_required"
+    assert payload["recommended_recovery_action"] == "start_focused_continuation"
+    assert payload["compression_recovery"]["terminal_state"] == "compression_exhausted"
+
+
+def test_chat_start_keeps_recovery_when_substantive_prompt_fails_validation(monkeypatch, tmp_path):
+    session_dir = _isolate_sessions(monkeypatch, tmp_path)
+    sid = "recoverychat2"
+    session = Session(
+        session_id=sid,
+        title="Recovery",
+        workspace=str(tmp_path),
+        model="gpt-4o",
+        messages=[{"role": "user", "content": "long task"}],
+    )
+    stamp_compression_exhausted_recovery(session, message="Context length exceeded.")
+    session.save()
+    models.SESSIONS[sid] = session
+    routes.SESSIONS[sid] = session
+    monkeypatch.setattr(
+        routes,
+        "_resolve_chat_workspace_with_recovery",
+        lambda *_args, **_kwargs: (_ for _ in ()).throw(ValueError("bad workspace")),
+    )
+
+    handler = _JSONHandler()
+    routes._handle_chat_start(handler, {"session_id": sid, "message": "continue by checking the repo"})
+    saved = json.loads((session_dir / f"{sid}.json").read_text(encoding="utf-8"))
+
+    assert handler.status == 400
+    assert saved["recommended_recovery_action"] == "start_focused_continuation"
+    assert saved["compression_recovery"]["terminal_state"] == "compression_exhausted"
+
+
+def test_chat_start_clears_recovery_when_substantive_prompt_starts(monkeypatch, tmp_path):
+    session_dir = _isolate_sessions(monkeypatch, tmp_path)
+    sid = "recoverychat3"
+    session = Session(
+        session_id=sid,
+        title="Recovery",
+        workspace=str(tmp_path),
+        model="gpt-4o",
+        messages=[{"role": "user", "content": "long task"}],
+    )
+    stamp_compression_exhausted_recovery(session, message="Context length exceeded.")
+    session.save()
+    models.SESSIONS[sid] = session
+    routes.SESSIONS[sid] = session
+    monkeypatch.setattr(routes, "_resolve_chat_workspace_with_recovery", lambda *_args, **_kwargs: str(tmp_path))
+    monkeypatch.setattr(routes, "_read_profile_model_config", lambda *_args, **_kwargs: (None, None, {}))
+    monkeypatch.setattr(
+        routes,
+        "_resolve_compatible_session_model_state",
+        lambda requested_model, requested_provider, **_kwargs: (requested_model or "gpt-4o", requested_provider or "openai", False),
+    )
+    monkeypatch.setattr(routes, "webui_gateway_chat_enabled", lambda _config: False)
+
+    def _fake_start_run(run_session, **_kwargs):
+        assert compression_recovery_payload_for_session(run_session) is None
+        run_session.save()
+        return {"session_id": sid, "stream_id": "stream1", "_status": 200}
+
+    monkeypatch.setattr(routes, "_start_run", _fake_start_run)
+
+    handler = _JSONHandler()
+    routes._handle_chat_start(handler, {"session_id": sid, "message": "continue by checking the repo"})
+    payload = _payload(handler)
+    saved = json.loads((session_dir / f"{sid}.json").read_text(encoding="utf-8"))
+
+    assert handler.status == 200
+    assert payload["stream_id"] == "stream1"
+    assert saved["compression_recovery"] == {}
+    assert saved["recommended_recovery_action"] is None
+
+
+def test_chat_start_restores_recovery_when_substantive_prompt_start_is_rejected(monkeypatch, tmp_path):
+    session_dir = _isolate_sessions(monkeypatch, tmp_path)
+    sid = "recoverychat4"
+    session = Session(
+        session_id=sid,
+        title="Recovery",
+        workspace=str(tmp_path),
+        model="gpt-4o",
+        messages=[{"role": "user", "content": "long task"}],
+    )
+    stamp_compression_exhausted_recovery(session, message="Context length exceeded.")
+    session.save()
+    models.SESSIONS[sid] = session
+    routes.SESSIONS[sid] = session
+    monkeypatch.setattr(routes, "_resolve_chat_workspace_with_recovery", lambda *_args, **_kwargs: str(tmp_path))
+    monkeypatch.setattr(routes, "_read_profile_model_config", lambda *_args, **_kwargs: (None, None, {}))
+    monkeypatch.setattr(
+        routes,
+        "_resolve_compatible_session_model_state",
+        lambda requested_model, requested_provider, **_kwargs: (requested_model or "gpt-4o", requested_provider or "openai", False),
+    )
+    monkeypatch.setattr(routes, "webui_gateway_chat_enabled", lambda _config: False)
+
+    def _fake_start_run(run_session, **_kwargs):
+        assert compression_recovery_payload_for_session(run_session) is None
+        run_session.save()
+        return {"error": "session already has an active stream", "_status": 409}
+
+    monkeypatch.setattr(routes, "_start_run", _fake_start_run)
+
+    handler = _JSONHandler()
+    routes._handle_chat_start(handler, {"session_id": sid, "message": "continue by checking the repo"})
+    saved = json.loads((session_dir / f"{sid}.json").read_text(encoding="utf-8"))
+
+    assert handler.status == 409
+    assert saved["recommended_recovery_action"] == "start_focused_continuation"
+    assert saved["compression_recovery"]["terminal_state"] == "compression_exhausted"
+
+
+def test_recovery_start_creates_focused_linked_session(monkeypatch, tmp_path):
+    session_dir = _isolate_sessions(monkeypatch, tmp_path)
+    sid = "recoverysrc1"
+    session = Session(
+        session_id=sid,
+        title="Long task",
+        workspace=str(tmp_path),
+        model="gpt-4o",
+        model_provider="openai",
+        profile="default",
+        project_id="proj_1",
+        messages=[{"role": "user", "content": "long task"}],
+        context_messages=[{"role": "user", "content": "large context"}],
+    )
+    stamp_compression_exhausted_recovery(session, message="Context length exceeded.")
+    session.save()
+    models.SESSIONS[sid] = session
+    routes.SESSIONS[sid] = session
+
+    handler = _JSONHandler()
+    routes._handle_session_compression_recovery_start(handler, {"session_id": sid})
+    payload = _payload(handler)
+
+    assert handler.status == 200
+    new_session = payload["session"]
+    assert new_session["session_id"] != sid
+    assert new_session["parent_session_id"] == sid
+    assert new_session["workspace"] == str(tmp_path)
+    assert new_session["model"] == "gpt-4o"
+    assert new_session["model_provider"] == "openai"
+    assert new_session["messages"] == []
+
+    saved = json.loads((session_dir / f"{new_session['session_id']}.json").read_text(encoding="utf-8"))
+    assert saved["parent_session_id"] == sid
+    assert saved["context_messages"] == []
+    assert compression_recovery_payload_for_session(session)["recommended_action"] == "start_focused_continuation"
+
+
+def test_recovery_metadata_is_persisted_and_exposed_in_compact_session():
+    session = Session(session_id="recovermeta1", title="x", messages=[])
+    recovery = stamp_compression_exhausted_recovery(session, message="Context length exceeded.")
+    compact = session.compact()
+
+    assert recovery["terminal_state"] == "compression_exhausted"
+    assert compact["recommended_recovery_action"] == "start_focused_continuation"
+    assert compact["compression_recovery"]["recommended_action"] == "start_focused_continuation"
+
+
+def test_compression_recovery_ui_wires_card_action_and_send_intercept():
+    ui = (ROOT / "static/ui.js").read_text(encoding="utf-8")
+    messages = (ROOT / "static/messages.js").read_text(encoding="utf-8")
+
+    assert "function _compressionRecoveryHtml" in ui
+    assert "data-compression-recovery-card=\"1\"" in ui
+    assert "api('/api/session/compression-recovery/start'" in ui
+    assert "function shouldInterceptCompressionRecoveryContinuation" in ui
+    assert "shouldInterceptCompressionRecoveryContinuation(text,S.pendingFiles)" in messages
+    assert "_compressionRecovery:recovery||undefined" in messages