switchroom 0.19.32 → 0.19.34

package/telegram-plugin/tests/status-pin-store.test.ts

@@ -116,12 +116,15 @@ describe("status-pin-store", () => {
     expect(idOnly(loadStatusPins(PATH, fs))).toEqual([]);
   });
 
-  it("fail-open: wrong envelope version / shape loads as []", () => {
-    // v5 is an unknown FUTURE version (v1–v4 are the supported set).
-    const { fs: f1 } = memFs({ [PATH]: JSON.stringify({ v: 5, pins: [pin()] }) });
+  it("fail-open: a shape that is not an envelope at all loads as []", () => {
+    const { fs: f1 } = memFs({ [PATH]: JSON.stringify({ v: 1, pins: "nope" }) });
     expect(loadStatusPins(PATH, f1)).toEqual([]);
-    const { fs: f2 } = memFs({ [PATH]: JSON.stringify({ v: 1, pins: "nope" }) });
+    const { fs: f2 } = memFs({ [PATH]: JSON.stringify({ pins: [pin()] }) });
     expect(loadStatusPins(PATH, f2)).toEqual([]);
+    const { fs: f3 } = memFs({ [PATH]: JSON.stringify({ v: "4", pins: [pin()] }) });
+    expect(loadStatusPins(PATH, f3)).toEqual([]);
+    const { fs: f4 } = memFs({ [PATH]: JSON.stringify({ v: 0, pins: [pin()] }) });
+    expect(loadStatusPins(PATH, f4)).toEqual([]);
   });
 
   it("drops malformed rows but keeps valid ones", () => {
@@ -604,8 +607,61 @@ describe("status-pin-store — envelope version compat", () => {
     expect(loadStatusPins(PATH, fs)).toHaveLength(1);
   });
 
-  it("rejects an unknown future envelope version (fail-open [])", () => {
-    const { fs } = memFs({ [PATH]: JSON.stringify({ v: 5, pins: [pin()] }) });
+  /**
+   * #3957 — a DOWNGRADE must degrade, not discard.
+   *
+   * The pre-#3953 reader has already shipped, so the v3→v4 case can never be
+   * retroactively fixed; this pins the forward-looking half so the NEXT bump
+   * behaves. A build that does not know the envelope version still reads every
+   * row it can structurally validate — because for THIS file "discard" is not
+   * "do nothing", it is "forget which messages we pinned", which turns every
+   * pin the newer build took into a permanent orphan.
+   */
+  it("reads an unknown FUTURE envelope version, keeping every row it can validate", () => {
+    const { fs } = memFs({
+      [PATH]: JSON.stringify({
+        v: 99,
+        pins: [
+          // A future row: known fields valid, plus a field this build has never
+          // heard of. Kept — unknown fields are additive by the bump rule.
+          { ...pin({ messageId: 715 }), someFutureField: "v99" },
+          { pinKey: "k", chatId: "c", messageId: "not-a-number" }, // still dropped
+        ],
+      }),
+    });
+    const loaded = loadStatusPins(PATH, fs);
+    expect(loaded).toHaveLength(1);
+    expect(loaded[0].messageId).toBe(715);
+  });
+
+  it("a rolled-back build still UNPINS the orphans a newer build recorded (#3957)", async () => {
+    // The outcome that matters: not "the rows parse" but "the boot cleanup can
+    // still reach them". Against the fail-open reader this unpins nothing.
+    const { fs } = memFs({
+      [PATH]: JSON.stringify({
+        v: 99,
+        pins: [
+          { pinKey: "fg:c:3", chatId: "-100123", messageId: 715, unknownV99: 1 },
+          { pinKey: "wk:agent-x", chatId: "-100999", messageId: 42 },
+        ],
+      }),
+    });
+
+    const unpinned: Array<[string, number]> = [];
+    const res = await runStatusPinBootCleanup({
+      path: PATH,
+      fs,
+      unpin: async (chatId, messageId) => {
+        unpinned.push([chatId, messageId]);
+      },
+      log: () => {},
+    });
+
+    expect(unpinned).toEqual([
+      ["-100123", 715],
+      ["-100999", 42],
+    ]);
+    expect(res.cleared).toBe(2);
     expect(idOnly(loadStatusPins(PATH, fs))).toEqual([]);
   });
 

package/vendor/hindsight-memory/scripts/subagent_retain.py

@@ -90,14 +90,37 @@ SIDECHAIN_MAX_READ_BYTES = 8 * 1024 * 1024
 SIDECHAIN_SCAN_FRESH_WINDOW_S = 300
 
 # Extraction-framing header prepended to the retained content. The retain API
-# has no per-call mission (mission is bank-level), so we bias the consolidation
-# engine toward PROCESS facts with a short in-content note. Deterministic text —
-# it does not change the document_id (that is computed from the raw slice before
-# this is prepended).
+# has no per-call mission (mission is bank-level), so the only lever available
+# here is a short in-content note. Deterministic text — it does not change the
+# document_id (that is computed from the raw slice before this is prepended).
+#
+# 2026-07-29 rewrite. The original text asked the extractor for "PROCESS facts:
+# files/paths touched, commands that worked, ..." — i.e. it explicitly SOLICITED
+# the exact classes the bank-level mission (`DEFAULT_RETAIN_MISSION` in
+# src/memory/hindsight.ts) enumerates as NEVER-extract: file paths, tool-result
+# exhaust, narration of what the assistant did, volatile state. Because this
+# header sits inside the content the extractor reads, it read as a local
+# override and countermanded the bank mission on the sidechain path. The header
+# now REINFORCES the bank mission instead of fighting it: it keeps the genuinely
+# durable half of the old wording (decisions with their rationale, non-obvious
+# technique, structural dead ends) and names the ephemera to drop.
+# Governing test: would this still be worth knowing in a month?
 SIDECHAIN_MISSION_HEADER = (
-    "[sidechain sub-agent work log — extract PROCESS facts: files/paths touched, "
-    "commands that worked, decisions made, dead ends hit and why. Ignore the "
-    "restated mission prose and routine tool chatter.]"
+    "[sidechain sub-agent work log. Apply the bank's retain mission unchanged; "
+    "this is raw work exhaust and most of it is NOT memorable. Extract only "
+    "durable knowledge — something that would still be worth knowing in a month: "
+    "decisions made and the reasoning behind them; a non-obvious technique or "
+    "constraint that would save the next agent real time; a STRUCTURAL dead end "
+    "(\"X cannot work because Y\") as opposed to an incidental one (\"the command "
+    "failed so I retried\"). "
+    "Do NOT extract: the status of in-flight or unfinished work; PR, issue, "
+    "branch, review or CI state; counts, metrics, versions, sizes or any other "
+    "value that changes; container, process or runtime snapshots; narration of "
+    "what the agent did, including lists of the files, paths or commands it "
+    "touched; paths under /tmp or a scratchpad directory; session, agent, "
+    "request or operation IDs; the restated mission prose; routine tool chatter. "
+    "Anything whose truth expires when this session does is not a memory. "
+    "If nothing durable remains, extract nothing.]"
 )
 
 

package/vendor/hindsight-memory/scripts/tests/test_subagent_retain.py

@@ -331,8 +331,10 @@ class RunSubagentRetain(unittest.TestCase):
         self.assertIn("agent_type:worker", captured["tags"])
         # Deterministic id in a DISTINCT namespace from parent retains.
         self.assertTrue(captured["document_id"].startswith("parentsess-sub-af5-r"))
-        # Process-fact framing header is prepended.
-        self.assertIn("extract PROCESS facts", captured["content"])
+        # Durable-knowledge framing header is prepended, verbatim and first.
+        self.assertTrue(
+            captured["content"].startswith(subagent_retain.SIDECHAIN_MISSION_HEADER)
+        )
         # Distinct provenance context.
         self.assertEqual(captured["context"], "claude-code-sidechain")
         self.assertEqual(captured["metadata"]["parent_session_id"], "parentsess")
@@ -459,6 +461,102 @@ class RunSubagentRetain(unittest.TestCase):
             self.assertIn(k, result["payload"])
 
 
+class SidechainMissionHeader(unittest.TestCase):
+    """The header must REINFORCE the bank-level retain mission, not countermand it.
+
+    The header is prepended into the retained content, where the extractor reads
+    it as a local instruction — so anything it solicits is written to memory even
+    when `DEFAULT_RETAIN_MISSION` (src/memory/hindsight.ts) lists that same class
+    under NEVER extract. The pre-2026-07-29 text asked for "PROCESS facts:
+    files/paths touched, commands that worked, ...", which is that exact
+    contradiction. These tests pin the fix: the solicit half asks only for
+    durable knowledge, and the ephemeral classes appear ONLY as exclusions.
+    """
+
+    SPLIT = "Do NOT extract"
+
+    def _halves(self):
+        header = subagent_retain.SIDECHAIN_MISSION_HEADER
+        self.assertIn(
+            self.SPLIT,
+            header,
+            "header must carry an explicit exclusion clause",
+        )
+        solicit, _, exclude = header.partition(self.SPLIT)
+        return header, solicit, exclude
+
+    def test_header_does_not_solicit_ephemeral_classes(self):
+        """Ephemera may be named as exclusions, never in the 'extract X' half."""
+        header = subagent_retain.SIDECHAIN_MISSION_HEADER
+        # The pre-fix solicitations, banned anywhere in the header — there is no
+        # phrasing of these that is a request for durable knowledge.
+        for gone in ("PROCESS facts", "files/paths touched", "commands that worked"):
+            self.assertNotIn(
+                gone,
+                header,
+                f"header still asks for {gone!r}, which the bank mission forbids",
+            )
+        _, solicit, _ = self._halves()
+        lowered = solicit.lower()
+        for banned in (
+            "process fact",
+            "files/paths",
+            "paths touched",
+            "commands that worked",
+            "/tmp",
+            "status",
+            "session id",
+            "container",
+            "metric",
+        ):
+            self.assertNotIn(
+                banned,
+                lowered,
+                f"header solicits ephemeral class {banned!r}: {solicit!r}",
+            )
+
+    def test_header_excludes_every_expiring_class(self):
+        _, _, exclude = self._halves()
+        lowered = exclude.lower()
+        for required in (
+            "in-flight",
+            "pr,",  # PR / issue / branch / CI state
+            "ci state",
+            "counts",
+            "versions",
+            "container",
+            "narration",
+            "/tmp",
+            "scratchpad",
+            "session, agent, request or operation ids",
+        ):
+            self.assertIn(
+                required,
+                lowered,
+                f"header fails to exclude {required!r}",
+            )
+
+    def test_header_still_solicits_decisions_with_rationale(self):
+        """The single most valuable sub-agent output must survive the rewrite."""
+        _, solicit, _ = self._halves()
+        lowered = solicit.lower()
+        self.assertIn("decisions made", lowered)
+        self.assertTrue(
+            any(w in lowered for w in ("reasoning", "rationale", "why")),
+            f"decisions are solicited without their rationale: {solicit!r}",
+        )
+        # And the other two durable classes the old wording earned.
+        self.assertIn("dead end", lowered)
+        self.assertIn("structural", lowered)
+
+    def test_header_states_the_durability_bar(self):
+        header, solicit, _ = self._halves()
+        self.assertIn("durable", solicit.lower())
+        self.assertIn("worth knowing in a month", solicit.lower())
+        # Bounded: this is prepended to every sidechain retain payload.
+        self.assertLess(len(header), 1500)
+
+
 class HookRegistration(unittest.TestCase):
     def test_subagentstop_registered_in_plugin_hooks_json(self):
         hooks_path = os.path.abspath(

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- 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.",
+  "retainMission": "Extract durable facts that will still be true and useful weeks from now, once this session is forgotten.\n\nDURABILITY GATE. Emit a candidate ONLY if it is one of these five classes:\n- PREFERENCE — what the user likes, wants, or always does; a standing rule or correction.\n- DECISION — a settled choice that changes how future work is done, including a choice NOT to do something. A decision about the mechanics of the CURRENT task (which worker to dispatch, which branch to rebase, which PR to merge now, what to do next) is process narration, not a durable decision — drop it unless it establishes a standing rule or permanently changes a system.\n- FINDING — a root cause, a measurement, or verified behaviour of a system. Include the number.\n- OUTCOME — a completed result that changed the world: what shipped, what a thing turned out to be. Not the act of shipping it.\n- RELATIONSHIP — who a person is, what a project or tool is, and how they connect.\nIf a candidate fits none of the five, it is not a memory. Drop it; do not reword it into one.\n\nA 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- The act of delegating, dispatching, spawning, launching, steering or merging\n  work — including when it succeeded. \"X was dispatched and completed\" is the\n  session describing itself. Record only what the work LEARNED or CHANGED.\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\nWrite each fact so it stands alone: name the thing, the number, and the date. A\nsentence that only makes sense while reading this transcript is not durable.\n\nIf a candidate fact matches an exclusion, drop it rather than rewording it. If\nnothing durable remains, return an empty facts list.\n",
   "autoRecall": true,
   "autoRetain": true,
   "retainMode": "full-session",