switchroom 0.19.27 → 0.19.28

package/vendor/hindsight-memory/settings.json

@@ -2,7 +2,7 @@
   "hindsightApiUrl": "",
   "bankId": "claude_code",
   "bankMission": "You are a Claude Code AI assistant. Focus on technical discussions, decisions, and context relevant to the user's projects.",
-  "retainMission": "Extract durable facts that will still be true and useful weeks from now: user preferences and standing rules, ongoing projects and recurring commitments, technical and architectural decisions with their rationale, and people/tool relationships. A preference revealed by a request is durable — record the preference (what the user likes, wants, or always does), not the request itself.\n\nA TOOL RESULT IS NOT A FACT. Before extracting, ask: is the subject of this\ncandidate a file path, a command/process/agent/session id, a temp directory, or\nthe location where some output was written? If yes, drop it — it is transcript\nexhaust, not memory.\n\nNEVER extract:\n- Tool results verbatim or paraphrased. Concretely, never produce a fact whose\n  text resembles any of these: \"File created successfully at /path/to/file\",\n  \"A background command with ID bctz4yskm is running, and its output will be\n  written to /tmp/...\", \"Async agent a745598ba84e71df1 was launched successfully\n  and is running in the background\", \"User executed a Bash command to sleep for\n  200 seconds\", \"The assistant used grep to locate 'truncateSync' in src/foo.ts\".\n- Anything mentioning a path under /tmp, a scratchpad directory, or a .tmp file.\n- Agent tool-use traces or narration of what the assistant did (e.g. \"the\n  assistant used X to query Y\", \"ran a search\", \"sent the message\").\n- In-flight workflow/process narration (a sub-task started, paused, or is still\n  running) — retain the outcome only once the task completes or a decision is made.\n- Operation, request, batch, agent, command or session IDs, UUIDs, hashes, or error codes.\n- Slash commands the user typed and their effects (e.g. \"User issued /clear to\n  reset assistant state\").\n- Hindsight's own errors, retries, backlogs, or internal state — the memory\n  system's self-reports are not memories.\n- Restatements of the user's current request or the task in progress.\n- Transient state (unread counts, build status, what is running right now) unless\n  the fact is explicitly dated, in which case record it as a dated observation.\n- Greetings, acknowledgements, and routine operational chatter.\n\nIf a candidate fact matches an exclusion, drop it rather than rewording it. If\nnothing durable remains, return an empty facts list.",
+  "retainMission": "Extract durable facts that will still be true and useful weeks from now: user preferences and standing rules, ongoing projects and recurring commitments, technical and architectural decisions with their rationale, and people/tool relationships. A preference revealed by a request is durable — record the preference (what the user likes, wants, or always does), not the request itself.\n\nA TOOL RESULT IS NOT A FACT. Before extracting, ask: is the subject of this\ncandidate a file path, a command/process/agent/session id, a temp directory, or\nthe location where some output was written? If yes, drop it — it is transcript\nexhaust, not memory.\n\nNEVER extract:\n- Tool results verbatim or paraphrased. Concretely, never produce a fact whose\n  text resembles any of these: \"File created successfully at /path/to/file\",\n  \"A background command with ID bctz4yskm is running, and its output will be\n  written to /tmp/...\", \"Async agent a745598ba84e71df1 was launched successfully\n  and is running in the background\", \"User executed a Bash command to sleep for\n  200 seconds\", \"The assistant used grep to locate 'truncateSync' in src/foo.ts\".\n- Anything mentioning a path under /tmp, a scratchpad directory, or a .tmp file.\n- Agent tool-use traces or narration of what the assistant did (e.g. \"the\n  assistant used X to query Y\", \"ran a search\", \"sent the message\").\n- In-flight workflow/process narration (a sub-task started, paused, or is still\n  running) — retain the outcome only once the task completes or a decision is made.\n- Operation, request, batch, agent, command or session IDs, UUIDs, hashes, or error codes.\n- Slash commands the user typed and their effects (e.g. \"User issued /clear to\n  reset assistant state\").\n- Hindsight's own errors, retries, backlogs, or internal state — the memory\n  system's self-reports are not memories.\n- Restatements of the user's current request or the task in progress.\n- Volatile state written as a timeless assertion. A version, count, size,\n  backlog, status, or any \"X is running Y\" / \"X is at Y\" / \"X is currently Y\"\n  claim is true only at the instant it was said. Concretely, never produce a\n  fact whose text resembles any of these: \"Switchroom fleet is running image\n  version v0.18.19\", \"The switchroom repo is at /path/to/fleet, version\n  v0.19.5\", \"Bank overlord has 43155 pending consolidations\", \"The build is\n  currently green\". If the claim is worth keeping, put the date INSIDE the\n  fact text (\"As of 2026-07-19 the fleet was running v0.18.19\"); if you\n  cannot date it, drop it. An undated one is recalled forever as though it\n  were still true, which is worse than not remembering it at all.\n- Transient state (unread counts, build status, what is running right now) unless\n  the fact is explicitly dated, in which case record it as a dated observation.\n- Greetings, acknowledgements, and routine operational chatter.\n\nIf a candidate fact matches an exclusion, drop it rather than rewording it. If\nnothing durable remains, return an empty facts list.",
   "autoRecall": true,
   "autoRetain": true,
   "retainMode": "full-session",

package/vendor/hindsight-memory/tests/test_hooks.py

@@ -104,9 +104,18 @@ class TestRecallHook:
             raise OSError("connection refused")
 
         hook_input = make_hook_input(prompt="What is my project about?")
-        # Should not raise — graceful degradation
+        # Should not raise — graceful degradation. Since switchroom #3619 that
+        # degradation is no longer SILENT: an unreachable OWN bank emits the
+        # degraded-recall disclosure, so the agent says "I could not check"
+        # rather than asserting there is no prior context. This case asserted
+        # the pre-#3619 silence and had been RED on main ever since. Nothing
+        # noticed because no CI job ran this file; #3688 wired the suite into
+        # ci-tests-python.yml, which is what surfaced it. The behaviour itself
+        # is pinned CI-side by scripts/tests/test_recall_degraded_notice.py.
         output = _run_hook("recall", hook_input, monkeypatch, tmp_path, urlopen_side_effect=raise_error)
-        assert output.strip() == ""
+        data = json.loads(output)
+        assert data["hookSpecificOutput"]["hookEventName"] == "UserPromptSubmit"
+        assert "DEGRADED" in data["hookSpecificOutput"]["additionalContext"]
 
     def test_output_format_matches_claude_code_spec(self, monkeypatch, tmp_path):
         memory = make_memory("User prefers Python")