Merge remote-tracking branch 'origin/master' into paperclip/pr-6222

CHANGELOG.md

@@ -9,6 +9,8 @@
 
 ### Fixed
 
+- **The "Compressing context" card no longer flashes on turns that aren't actually compressing.** The card was driven by a broad status match (anything containing `compressing`), so skip/cooldown/defer notices and post-compression summaries (`Compressed X → Y`) could paint a phantom compression card mid-turn. A dedicated `_is_agent_compression_start_status()` now recognizes only the five genuine start-of-compression markers the Hermes agent emits, and the transcript renderer no longer lets an unrecognized `terminal_status` row paint a compression anchor. Real compression turns still show the card unchanged; verified against every current agent compression emitter. Thanks @carlotestor. (#6184)
+
 - **Transparent Stream no longer shows the start of the final answer twice after a multi-segment turn settles.** In a turn with interim assistant prose interleaved with tool calls, the beginning of the final answer could render twice after the live→settled transition — once as a stale live-token accumulator snapshot, once as the real assistant segment — persisting until reload. The earlier suppression gate (#5758) computed its "after the last tool row" boundary over the combined row list, where settled per-message tool rows re-list tools that ran earlier in the turn, pushing the boundary past the final segment's accumulator so the dedupe never fired. Final-segment eligibility is now judged against the live projection's own chronology (a live-prose row belongs to the final segment only when no *projected* tool row follows it), so the stale prefix is dropped while pre-tool narration and every tool/thinking row are preserved. Thanks @ai-ag2026. (#6189, #5749)
 
 - **The frontend now reads the `msg_limit` ceiling from `/api/session` metadata instead of hand-mirroring the constant.** Following the ceiling clamp shipped in exp-v0.52.98, the backend advertises its `_MAX_MSG_LIMIT` as `_msg_limit_max` in every `/api/session` response, and the frontend reads it dynamically (`_msgLimitMax`, a module-scope value defaulting to the static fallback), so the two can no longer drift and the mirrored constant's drift-guard test is no longer needed. Older backends that omit the field fall back to the built-in default, so mixed-version deployments are unaffected. Thanks @webtecnica. (#6214, #6177)

api/streaming.py

@@ -542,6 +542,48 @@ def _is_fallback_lifecycle_message(kind: str, message: str) -> bool:
     )
 
 
+def _is_agent_compression_start_status(kind: str, message: str) -> bool:
+    """Return True only for real Hermes context-compression start notices.
+
+    WebUI bridges matching lifecycle statuses into an SSE ``compressing`` event
+    and paints the live "Compressing context" worklog divider. The previous
+    matcher used broad substrings such as ``'compressing' in message`` and
+    ``'preflight compression' in message``, which can false-positive on skip /
+    cooldown / unrelated notices and make brand-new low-token turns look like
+    auto-compression.
+
+    Positive markers below match the agent emitters in hermes-agent
+    (``turn_context`` preflight, ``conversation_loop`` pre-API / 413 / too-large,
+    ``conversation_compression`` compaction status). Explicitly reject skip /
+    defer notices so "Skipping preflight compression…" never surfaces as a
+    running compress divider.
+    """
+    k = str(kind or '').strip().lower()
+    m = str(message or '').strip().lower()
+    if k != 'lifecycle' or not m:
+        return False
+    # Skip / cooldown / defer logs must never look like a live compression start.
+    if (
+        'skipping' in m
+        or 'defer' in m
+        or 'cooldown' in m
+        or 'will not start' in m
+    ):
+        return False
+    # Post-compress retry chatter is not a start event.
+    if 'compressed' in m and 'compressing' not in m and 'compression attempt' not in m:
+        return False
+    return (
+        'preflight compression:' in m
+        or 'pre-api compression:' in m
+        or 'compacting context' in m
+        or 'context too large' in m
+        or '— compressing (' in m
+        or '- compressing (' in m
+        or 'compression attempt' in m
+    )
+
+
 def _prewarm_skill_tool_modules():
     """Import tools.skills_tool and tools.skill_manager_tool outside any lock.
 
@@ -7148,16 +7190,7 @@ def _run_agent_streaming(
             and 'http' in _lower
         ):
             _captured_terminal_error[0] = _message
-        _is_compression_start = (
-            _kind == 'lifecycle'
-            and (
-                'preflight compression' in _lower
-                or 'compressing' in _lower
-                or 'compacting context' in _lower
-                or 'context too large' in _lower
-            )
-        )
-        if _is_compression_start:
+        if _is_agent_compression_start_status(_kind, _message):
             put('compressing', {
                 'session_id': session_id,
                 'message': 'Compressing context',

static/messages.js

@@ -3936,8 +3936,38 @@ function attachLiveStream(activeSid, streamId, uploaded=[], options={}){
     if(role==='prose'||kind==='process_prose') return 'token';
     if(role==='thinking'||kind==='reasoning') return 'reasoning';
     if(role==='tool') return row&&row.status==='running'?'tool':'tool_complete';
-    if(role==='terminal'||kind==='terminal_status') return row&&row.status==='running'?'compressing':'done';
-    if(role==='lifecycle'||kind==='lifecycle_status') return 'compressing';
+    // Terminal statuses are done/cancel/error/apperror — never invent a
+    // compression start from a running terminal row (false "Compressing context").
+    if(role==='terminal'||kind==='terminal_status'){
+      const termStatus=String(row&&row.status||'').trim().toLowerCase();
+      if(termStatus==='cancelled'||termStatus==='canceled'||termStatus==='interrupted') return 'cancel';
+      if(termStatus==='error'||termStatus==='failed'||termStatus==='errored') return 'error';
+      if(termStatus==='running') return '';
+      return 'done';
+    }
+    // lifecycle_status is shared by compressing + compressed. Prefer explicit
+    // cues; do not default every lifecycle row to a running compress divider.
+    if(role==='lifecycle'||kind==='lifecycle_status'){
+      const phase=String(row&&(row.phase||row.status)||'').trim().toLowerCase();
+      const text=String(row&&(row.text||row.message||row.label)||'').trim().toLowerCase();
+      if(
+        phase==='done'||phase==='completed'||phase==='compressed'
+        || text.includes('auto-compressed')
+        || text.includes('compression finished')
+        || (text.includes('compressed')&&!text.includes('compressing'))
+      ) return 'compressed';
+      if(
+        phase==='running'||phase==='compressing'
+        || text.includes('compressing context')
+        || text.includes('compacting context')
+        || text.includes('preflight compression')
+        || text.includes('pre-api compression')
+        || text.includes('context too large')
+        || text.includes('compression attempt')
+        || (text.includes('compressing')&&!text.includes('skipping'))
+      ) return 'compressing';
+      return '';
+    }
     return '';
   }
   function _hydrateAnchorRegistryFromActivityScene(scene){

tests/test_auto_compression_card.py

@@ -5,6 +5,7 @@ from api.models import Session
 from api.streaming import (
     _POST_COMPRESSION_TOOL_RESULT_SUMMARY_FLAG,
     _compressed_context_tool_result_summary,
+    _is_agent_compression_start_status,
     _is_fallback_lifecycle_message,
     _merge_display_messages_after_agent_result,
     _message_text,
@@ -447,14 +448,13 @@ def test_agent_status_callback_emits_compressing_and_warning_events():
     assert end != -1, "status callback block end marker not found"
     block = src[start:end]
 
-    # compressing events for compression lifecycle notices
+    # compressing events only via the narrowed helper (no broad substring matcher)
     assert "put('compressing'" in block
     assert "'session_id': session_id" in block
     assert "'message': 'Compressing context'" in block
-    assert "'preflight compression'" in block
-    assert "'compressing'" in block
-    assert "'compacting context'" in block
-    assert "'context too large'" in block
+    assert "_is_agent_compression_start_status(_kind, _message)" in block
+    assert "or 'compressing' in _lower" not in block
+    assert "or 'preflight compression' in _lower" not in block
 
     # warning events with type:fallback for rate-limit/fallback lifecycle notices
     assert "put('warning'" in block
@@ -471,6 +471,77 @@ def test_agent_status_callback_emits_compressing_and_warning_events():
     assert "agent.status_callback = _agent_kwargs.get('status_callback')" in src
 
 
+def test_agent_compression_start_status_matches_real_emitters_only():
+    # Real start notices from hermes-agent emitters
+    assert _is_agent_compression_start_status(
+        "lifecycle",
+        "📦 Preflight compression: ~101,000 tokens >= 96,000 threshold. This may take a moment.",
+    )
+    assert _is_agent_compression_start_status(
+        "lifecycle",
+        "📦 Pre-API compression: ~521,055 tokens near the context/output limit. Compacting before the next model call.",
+    )
+    assert _is_agent_compression_start_status(
+        "lifecycle",
+        "🗜️ Compacting context — summarizing earlier conversation so I can continue...",
+    )
+    assert _is_agent_compression_start_status(
+        "lifecycle",
+        "🗜️ Context too large (~120,000 tokens) — compressing (1/3)...",
+    )
+    assert _is_agent_compression_start_status(
+        "lifecycle",
+        "⚠️  Request payload too large (413) — compression attempt 1/3...",
+    )
+
+    # False positives that previously (or could) light the divider on low-token turns
+    assert not _is_agent_compression_start_status(
+        "lifecycle",
+        "Skipping preflight compression: rough estimate ~20,000 >= 96,000, but last real provider prompt was 18,000 after compression",
+    )
+    assert not _is_agent_compression_start_status(
+        "lifecycle",
+        "Skipping preflight compression: same-session cooldown active (~30 seconds remaining, session abc)",
+    )
+    assert not _is_agent_compression_start_status(
+        "lifecycle",
+        "Skipping Hermes preflight compression for codex app-server (mode=native); Hermes will not start thread compaction here.",
+    )
+    assert not _is_agent_compression_start_status(
+        "lifecycle",
+        "🗜️ Compressed 924 → 143 messages, retrying...",
+    )
+    assert not _is_agent_compression_start_status(
+        "lifecycle",
+        "Rate limited — switching to fallback provider...",
+    )
+    assert not _is_agent_compression_start_status("tool", "compacting context")
+    assert not _is_agent_compression_start_status("lifecycle", "")
+    assert not _is_agent_compression_start_status("lifecycle", "Working on request")
+
+
+def test_snapshot_anchor_hydration_does_not_invent_compressing_rows():
+    src = _read("static/messages.js")
+    start = src.find("function _sourceEventTypeForSnapshotAnchorRow")
+    assert start != -1, "snapshot anchor source helper not found"
+    end = src.find("function _hydrateAnchorRegistryFromActivityScene", start)
+    assert end != -1, "hydrate helper after snapshot source helper not found"
+    block = src[start:end]
+
+    # Old false-positive defaults must be gone
+    assert "return row&&row.status==='running'?'compressing':'done'" not in block
+    assert "if(role==='lifecycle'||kind==='lifecycle_status') return 'compressing';" not in block
+
+    # Terminal running no longer maps to compressing
+    assert "if(termStatus==='running') return '';" in block
+    # Lifecycle requires positive compression cues
+    assert "text.includes('compacting context')" in block
+    assert "return 'compressed';" in block
+    assert "return 'compressing';" in block
+    # Unknown lifecycle rows stay empty (no invented divider)
+    assert "return '';" in block
+
+
 def test_agent_status_callback_wiring():
     src = _read("api/streaming.py")
     assert "_agent_status_callback" in src