superlocalmemory 4.1.4 → 4.1.6

package/src/superlocalmemory/mcp/session_binding.py

@@ -34,7 +34,13 @@ __all__ = ["resolve_session_id", "SESSION_ENV_VARS"]
 
 #: Checked in order. Hosts set one or the other; SLM's own takes precedence so a
 #: user can override a host that sets its variable to something unhelpful.
-SESSION_ENV_VARS = ("SLM_SESSION_ID", "CLAUDE_SESSION_ID")
+#: ``CLAUDE_CODE_SESSION_ID`` is the actual variable Claude Code's MCP
+#: subprocess environment carries — verified on a live daemon 2026-08-24,
+#: this pipeline settled 6013/6039 outcomes at the neutral 0.5 fallback
+#: because the old two-name list never matched it, and every recall fell
+#: through to the synthetic ``mcp:<agent_id>`` id that step 4 deliberately
+#: excludes from matching.
+SESSION_ENV_VARS = ("SLM_SESSION_ID", "CLAUDE_SESSION_ID", "CLAUDE_CODE_SESSION_ID")
 
 
 def resolve_session_id(

package/src/superlocalmemory/reliability/__init__.py

@@ -0,0 +1,45 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory | https://qualixar.com | https://varunpratap.com
+
+"""Checks that ask whether a mechanism is *effective*, not merely present.
+
+Three questions get confused in a system this size:
+
+* **Implemented** — the code exists. A grep answers this.
+* **Reachable** — something calls it. A call-graph trace answers this.
+* **Effective** — it has actually changed an outcome against real data.
+  **Neither of the above answers this.** Only querying the store does.
+
+A mechanism can pass the first two questions for months and fail the third
+silently: a learner whose reward channel emits a constant still records plays,
+and a conditional path guarded by a missing column still appears in coverage.
+Nothing raises, nothing logs, and every file is present.
+
+The two checks here answer the third question directly.
+
+* :mod:`.prior_distance` — has a Bayesian learner's posterior actually moved
+  away from its prior?
+* :mod:`.join_liveness` — has a schema-guarded code path ever executed against
+  this store, and if not, which requirement is missing?
+
+Both are read-only, both are fail-soft, and neither is on a hot path.
+"""
+
+from superlocalmemory.reliability.join_liveness import (
+    GuardVerdict,
+    check_schema_guards,
+)
+from superlocalmemory.reliability.prior_distance import (
+    DEFAULT_MIN_OBSERVATIONS,
+    LearnerVerdict,
+    check_beta_learners,
+)
+
+__all__ = [
+    "DEFAULT_MIN_OBSERVATIONS",
+    "GuardVerdict",
+    "LearnerVerdict",
+    "check_beta_learners",
+    "check_schema_guards",
+]

package/src/superlocalmemory/reliability/join_liveness.py

@@ -0,0 +1,301 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory | https://qualixar.com | https://varunpratap.com
+
+"""Can a schema-guarded code path run against this store, and has it had the data?
+
+A note on what this proves, because the honest scope is narrower than the
+convenient phrasing. The check reads the schema: it establishes whether the
+guard's requirements are present *now*, and therefore whether the path is
+reachable at all. It does not read an execution history, so a satisfied guard
+means "this would execute", not "this has executed". The inference to "never
+executed" is sound only in the other direction -- if a required column is absent
+from a store that has been in service, the guarded path cannot have run against
+it -- and that is the direction the findings rely on.
+
+Some features are wired behind a guard that asks the schema a question before
+doing any work — "is this table here, does that column exist" — and fall back
+silently when the answer is no. The fallback is correct behaviour: it is what
+keeps an old store openable. But it means a feature can be implemented, called
+on the hot path, and covered by tests, while never once executing against real
+data. Static analysis passes. A call-graph trace passes. Coverage passes. The
+guard returns ``False`` and the feature is arithmetically absent.
+
+Two things make that failure hard to see from inside the code:
+
+1. **The guard is doing its job.** There is no error to raise. Falling back is
+   the designed response to a missing column.
+2. **The fallback is often a neutral value**, which composes into an identity.
+   A decay-rate multiplier that falls back to a trust of 1.0 collapses
+   ``lambda * (1 + kappa * (1 - trust))`` to ``lambda`` — the feature is on, and
+   it computes exactly what having no feature would compute.
+
+So this check does two things a plain schema assertion does not. It records
+whether each named guard **passes right now**, and when a guard fails it looks
+for the required data **elsewhere in the store** — because the common case is
+not that the data is missing, it is that the guard is asking the wrong table.
+
+A guard reported as ``SATISFIED_ELSEWHERE`` is the most actionable outcome
+available: the feature is one re-keyed join away from working, and no backfill
+or migration is required.
+"""
+
+from __future__ import annotations
+
+import logging
+import sqlite3
+from dataclasses import dataclass, field
+from typing import Any
+
+logger = logging.getLogger("superlocalmemory.reliability.join_liveness")
+
+
+@dataclass(frozen=True)
+class Requirement:
+    """One schema object a guard needs: a table, optionally a column on it."""
+
+    table: str
+    column: str | None = None
+
+    def __str__(self) -> str:
+        return f"{self.table}.{self.column}" if self.column else self.table
+
+
+@dataclass(frozen=True)
+class Guard:
+    """A named conditional path and the schema it requires to execute."""
+
+    name: str
+    describes: str
+    requires: tuple[Requirement, ...]
+    #: What the feature computes when the guard fails, in words. Recording this
+    #: is the difference between "a feature is off" and "a feature is off and
+    #: indistinguishable from not having it".
+    fallback_behaviour: str
+
+
+@dataclass(frozen=True)
+class GuardVerdict:
+    name: str
+    describes: str
+    #: LIVE = requirements present, so the path is reachable now. DEAD /     #: SATISFIED_ELSEWHERE = a requirement is absent, so the path cannot have run
+    #: against this store. LIVE is NOT evidence of past execution.
+    verdict: str
+    missing: tuple[str, ...] = field(default=())
+    found_elsewhere: tuple[tuple[str, str, int, float], ...] = field(default=())
+    detail: str = ""
+
+    @property
+    def is_live(self) -> bool:
+        return self.verdict == "LIVE"
+
+
+#: Guards worth reporting on. Each entry names a real conditional in the code,
+#: so that a reader can go from this list to the line that asks the question.
+GUARDS: tuple[Guard, ...] = (
+    Guard(
+        name="trust_weighted_forgetting",
+        describes=(
+            "learning/forgetting_scheduler.py::_has_trust_tables — gates the "
+            "per-fact trust lookup that modulates the decay rate"
+        ),
+        requires=(
+            Requirement("trust_scores"),
+            Requirement("atomic_facts", "created_by"),
+        ),
+        fallback_behaviour=(
+            "every fact takes trust = 1.0, so lambda_eff collapses to "
+            "lambda_base and the decay rate is identical to no trust weighting"
+        ),
+    ),
+)
+
+
+def _table_exists(conn: sqlite3.Connection, table: str) -> bool:
+    return (
+        conn.execute(
+            "SELECT 1 FROM sqlite_master WHERE type='table' AND name=? LIMIT 1",
+            (table,),
+        ).fetchone()
+        is not None
+    )
+
+
+def _columns(conn: sqlite3.Connection, table: str) -> set[str]:
+    try:
+        return {r[1] for r in conn.execute(f'PRAGMA table_info("{table}")')}
+    except sqlite3.Error:
+        return set()
+
+
+def _satisfied(conn: sqlite3.Connection, requirement: Requirement) -> bool:
+    if not _table_exists(conn, requirement.table):
+        return False
+    if requirement.column is None:
+        return True
+    return requirement.column in _columns(conn, requirement.table)
+
+
+def _look_elsewhere(
+    conn: sqlite3.Connection,
+    column: str,
+    *,
+    join_target: str | None = None,
+    join_key: str = "fact_id",
+) -> list[tuple[str, str, int]]:
+    """Find other tables carrying ``column``, with a populated-row count.
+
+    This is the part that turns a failed guard into a fix. A column the guard
+    could not find on its own table is often present on a neighbouring one,
+    already populated.
+
+    **A populated count is not coverage, and reporting it alone overstates the
+    remedy.** A provenance-style table can carry a value on every one of its own
+    rows while describing only part of the set the join needs: rows can be
+    missing for older entities entirely. When ``join_target`` is given, the
+    coverage fraction over that table is measured and returned, because the
+    honest question is not "does this column exist somewhere" but "how much of
+    what the guard needs would the re-keyed join actually resolve".
+    """
+    out: list[tuple[str, str, int]] = []
+    tables = [
+        r[0]
+        for r in conn.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' "
+            "AND name NOT LIKE 'sqlite_%'",
+        )
+    ]
+    for table in tables:
+        if column not in _columns(conn, table):
+            continue
+        try:
+            populated = conn.execute(
+                f'SELECT COUNT(*) FROM "{table}" '
+                f'WHERE "{column}" IS NOT NULL AND TRIM("{column}") <> \'\'',
+            ).fetchone()[0]
+        except sqlite3.Error:
+            populated = 0
+        covered = -1
+        if join_target and join_target != table:
+            try:
+                total = conn.execute(
+                    f'SELECT COUNT(*) FROM "{join_target}"',
+                ).fetchone()[0]
+                if total:
+                    hit = conn.execute(
+                        f'SELECT COUNT(*) FROM "{join_target}" t WHERE EXISTS ('
+                        f'  SELECT 1 FROM "{table}" s WHERE s."{join_key}" = t."{join_key}"'
+                        f'  AND s."{column}" IS NOT NULL AND TRIM(s."{column}") <> \'\')',
+                    ).fetchone()[0]
+                    covered = round(100.0 * hit / total, 1)
+            except sqlite3.Error:
+                covered = -1
+        out.append((table, column, int(populated), covered))
+    return out
+
+
+def _evaluate(conn: sqlite3.Connection, guard: Guard) -> GuardVerdict:
+    missing = [str(r) for r in guard.requires if not _satisfied(conn, r)]
+    if not missing:
+        return GuardVerdict(
+            name=guard.name,
+            describes=guard.describes,
+            verdict="LIVE",
+            detail=(
+                "Every requirement is present, so the guarded path is reachable "
+                "on this store. This is schema evidence, not execution history: "
+                "it does not establish that the path has ever actually run."
+            ),
+        )
+
+    elsewhere: list[tuple[str, str, int, float]] = []
+    for requirement in guard.requires:
+        if requirement.column and not _satisfied(conn, requirement):
+            for found in _look_elsewhere(
+                conn, requirement.column, join_target=requirement.table,
+            ):
+                if found[0] != requirement.table:
+                    elsewhere.append(found)
+
+    populated = [e for e in elsewhere if e[2] > 0]
+    if populated:
+        # Rank by coverage of the guarded table, falling back to row count only
+        # when coverage could not be measured. Picking the largest table instead
+        # would recommend a 4,000-row partial source over a smaller complete one.
+        best = max(populated, key=lambda e: (e[3], e[2]))
+        cov = best[3]
+        if cov < 0:
+            remedy = (
+                f"Re-keying the join onto that table would make the path "
+                f"executable; coverage over the guarded table could not be "
+                f"measured here, so confirm it before relying on the remedy."
+            )
+        elif cov >= 99.5:
+            remedy = (
+                f"That table covers {cov}% of the rows the guard needs, so "
+                f"re-keying the join makes the path live without a backfill."
+            )
+        else:
+            remedy = (
+                f"That table covers only {cov}% of the rows the guard needs. "
+                f"Re-keying the join makes the path executable for those rows "
+                f"and leaves the remainder on the same fallback, so this is a "
+                f"partial remedy and a backfill decision, not a free fix."
+            )
+        return GuardVerdict(
+            name=guard.name,
+            describes=guard.describes,
+            verdict="SATISFIED_ELSEWHERE",
+            missing=tuple(missing),
+            found_elsewhere=tuple(elsewhere),
+            detail=(
+                f"The guard requires {', '.join(missing)}, which is absent, so "
+                f"the path has never executed against this store — "
+                f"{guard.fallback_behaviour}. The data it needs is present on "
+                f"{best[0]}.{best[1]} with {best[2]} populated rows. {remedy}"
+            ),
+        )
+
+    return GuardVerdict(
+        name=guard.name,
+        describes=guard.describes,
+        verdict="DEAD",
+        missing=tuple(missing),
+        detail=(
+            f"The guard requires {', '.join(missing)}, which is absent from "
+            f"this store and not carried by any other table, so the path has "
+            f"never executed — {guard.fallback_behaviour}."
+        ),
+    )
+
+
+def check_schema_guards(
+    memory_db: Any, *, guards: tuple[Guard, ...] = GUARDS,
+) -> list[GuardVerdict]:
+    """Report, per registered guard, whether its path can run on this store.
+
+    ``memory_db`` may be a path or an open connection. Read-only, and fail-soft:
+    an error yields an empty list, because a diagnostic must never be the reason
+    something breaks.
+    """
+    owns_connection = not isinstance(memory_db, sqlite3.Connection)
+    conn: sqlite3.Connection | None = None
+    try:
+        conn = (
+            sqlite3.connect(f"file:{memory_db}?mode=ro", uri=True)
+            if owns_connection
+            else memory_db
+        )
+        return [_evaluate(conn, guard) for guard in guards]
+    except Exception:
+        logger.debug("join-liveness check skipped", exc_info=True)
+        return []
+    finally:
+        if owns_connection and conn is not None:
+            try:
+                conn.close()
+            except Exception:
+                pass
+
+
+__all__ = ["GUARDS", "Guard", "GuardVerdict", "Requirement", "check_schema_guards"]

package/src/superlocalmemory/reliability/prior_distance.py

@@ -0,0 +1,243 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory | https://qualixar.com | https://varunpratap.com
+
+"""Has a Bayesian learner's posterior actually moved off its prior?
+
+A Thompson-sampling selector over Beta posteriors records a play, applies a
+reward, and reports both. None of that tells you whether it learned anything.
+Its own counters cannot: a reward channel that emits one constant value
+increments the posterior on every play, so the play count rises, the timestamps
+advance, and the dashboard looks alive while the distribution never moves.
+
+This is not a hypothetical. A Beta posterior updated as
+``alpha += r; beta += (1 - r)`` is stationary in mean for exactly one reward
+value: ``r = 0.5``. And that value is the usual neutral fallback when a reward
+cannot be attributed to a play. So the failure that produces *no learning at
+all* is also the failure that produces *the most normal-looking counters*.
+
+The signature is exact, not statistical. With a ``Beta(a0, b0)`` prior and *n*
+observations all equal to 0.5::
+
+    alpha - a0  ==  beta - b0  ==  n * 0.5
+
+0.5 is exactly representable in binary floating point, so ``n * 0.5`` is exact
+for any plausible *n*. A learner matching that identity on every unit has
+provably received the neutral value every single time — there is no sampling
+noise to argue about, and one run is enough to establish it.
+
+What this check does NOT claim: that a moving posterior is a *good* one. Motion
+off the prior is necessary for learning, not sufficient. This distinguishes
+"receiving signal" from "receiving nothing", which is the distinction the
+learner's own metrics cannot make.
+"""
+
+from __future__ import annotations
+
+import logging
+import sqlite3
+from dataclasses import dataclass, field
+from typing import Any
+
+logger = logging.getLogger("superlocalmemory.reliability.prior_distance")
+
+#: Below this many observations in total, a posterior sitting at its prior is
+#: expected rather than suspicious, so no verdict is issued.
+DEFAULT_MIN_OBSERVATIONS = 20
+
+#: And below this many observations *per unit*. An aggregate floor alone lets a
+#: store with many units and few observations report STALLED on units that were
+#: never played at all.
+_MIN_OBSERVATIONS_PER_UNIT = 2.0
+
+#: Distance from the prior mean below which a unit counts as unmoved. Kept well
+#: above float noise so that a genuinely tiny update is not reported as none.
+_MEAN_EPSILON = 1e-9
+
+#: Beta learners in this store, as (table, unit column, observation column).
+#: Each is a Beta(1, 1) posterior over a named unit.
+_BETA_LEARNERS: tuple[tuple[str, str, str | None], ...] = (
+    ("bandit_arms", "arm_id", "plays"),
+    ("source_quality", "source_id", None),
+)
+
+_PRIOR_ALPHA = 1.0
+_PRIOR_BETA = 1.0
+
+
+@dataclass(frozen=True)
+class LearnerVerdict:
+    """One Beta learner, and whether its posterior has moved."""
+
+    table: str
+    units: int
+    units_at_prior_mean: int
+    units_matching_neutral_identity: int
+    observations: int
+    verdict: str
+    detail: str
+    sample: tuple[tuple[str, float, float], ...] = field(default=())
+
+    @property
+    def is_stalled(self) -> bool:
+        return self.verdict == "STALLED"
+
+
+def _table_exists(conn: sqlite3.Connection, table: str) -> bool:
+    row = conn.execute(
+        "SELECT 1 FROM sqlite_master WHERE type='table' AND name=? LIMIT 1",
+        (table,),
+    ).fetchone()
+    return row is not None
+
+
+def _columns(conn: sqlite3.Connection, table: str) -> set[str]:
+    return {r[1] for r in conn.execute(f'PRAGMA table_info("{table}")')}
+
+
+def _inspect_learner(
+    conn: sqlite3.Connection,
+    table: str,
+    unit_column: str,
+    observation_column: str | None,
+    *,
+    min_observations: int,
+) -> LearnerVerdict | None:
+    """Inspect one Beta learner. Returns None when the table is absent."""
+    if not _table_exists(conn, table):
+        return None
+
+    available = _columns(conn, table)
+    if not {"alpha", "beta"}.issubset(available):
+        return None
+    unit = unit_column if unit_column in available else "rowid"
+    obs_col = observation_column if (observation_column or "") in available else None
+
+    select = f'SELECT "{unit}", alpha, beta'
+    select += f', "{obs_col}"' if obs_col else ", NULL"
+    rows = conn.execute(f'{select} FROM "{table}"').fetchall()
+    if not rows:
+        return None
+
+    units = len(rows)
+    at_prior_mean = 0
+    neutral_identity = 0
+    observations = 0
+    for _unit, alpha, beta, obs in rows:
+        alpha = float(alpha or 0.0)
+        beta = float(beta or 0.0)
+        total = alpha + beta
+        if total > 0 and abs(alpha / total - 0.5) < _MEAN_EPSILON:
+            at_prior_mean += 1
+        n = int(obs) if obs is not None else None
+        if n is None:
+            # No per-unit observation count. Infer it from the identity itself:
+            # n = (alpha - a0) / 0.5 only holds if every reward was neutral, so
+            # this is checked, never assumed.
+            candidate = (alpha - _PRIOR_ALPHA) * 2.0
+            n = int(round(candidate)) if candidate >= 0 else 0
+        observations += n
+        if n > 0 and (
+            abs((alpha - _PRIOR_ALPHA) - n * 0.5) < _MEAN_EPSILON
+            and abs((beta - _PRIOR_BETA) - n * 0.5) < _MEAN_EPSILON
+        ):
+            neutral_identity += 1
+
+    sample = tuple(
+        (str(r[0]), float(r[1] or 0.0), float(r[2] or 0.0)) for r in rows[:3]
+    )
+
+    # The floor has to bind per unit, not in aggregate. 165 arms sharing 20
+    # observations leaves most of them untouched at exactly the prior, which
+    # satisfies the unmoved test for a reason that carries no information. A
+    # verdict of STALLED must mean "measured inert", never "too sparse to tell".
+    per_unit = observations / units if units else 0.0
+    if observations < min_observations or per_unit < _MIN_OBSERVATIONS_PER_UNIT:
+        verdict = "INSUFFICIENT_DATA"
+        detail = (
+            f"{observations} observations across {units} units "
+            f"({per_unit:.2f} per unit) is too sparse for an unmoved posterior to "
+            f"mean anything; the floors are {min_observations} in total and "
+            f"{_MIN_OBSERVATIONS_PER_UNIT:g} per unit."
+        )
+    elif neutral_identity == units:
+        verdict = "STALLED"
+        detail = (
+            f"All {units} units satisfy (alpha-{_PRIOR_ALPHA:g}) == "
+            f"(beta-{_PRIOR_BETA:g}) == n/2 exactly across {observations} "
+            f"observations, so the rewards sum to exactly n/2 on every unit and "
+            f"each posterior mean is still {0.5}. No unit has acquired any "
+            f"preference. Note the identity constrains the SUM: it is also "
+            f"satisfied by a symmetric non-neutral stream, so confirm against a "
+            f"per-observation record before concluding the reward was constant."
+        )
+    elif at_prior_mean == units:
+        verdict = "STALLED"
+        detail = (
+            f"All {units} units sit at posterior mean 0.5 after {observations} "
+            f"observations, without matching the exact neutral identity. The "
+            f"updates are symmetric but not uniformly 0.5 — inspect the reward "
+            f"source."
+        )
+    else:
+        moved = units - at_prior_mean
+        verdict = "MOVING"
+        detail = (
+            f"{moved} of {units} units have moved off the prior mean across "
+            f"{observations} observations."
+        )
+
+    return LearnerVerdict(
+        table=table,
+        units=units,
+        units_at_prior_mean=at_prior_mean,
+        units_matching_neutral_identity=neutral_identity,
+        observations=observations,
+        verdict=verdict,
+        detail=detail,
+        sample=sample,
+    )
+
+
+def check_beta_learners(
+    learning_db: Any,
+    *,
+    min_observations: int = DEFAULT_MIN_OBSERVATIONS,
+) -> list[LearnerVerdict]:
+    """Report, per Beta learner in ``learning_db``, whether it has learned.
+
+    ``learning_db`` may be a path or an open :class:`sqlite3.Connection`. The
+    database is only read. Fail-soft: any error yields an empty list and a log
+    line, because a diagnostic must never be the reason something breaks.
+    """
+    owns_connection = not isinstance(learning_db, sqlite3.Connection)
+    conn: sqlite3.Connection | None = None
+    try:
+        if owns_connection:
+            conn = sqlite3.connect(f"file:{learning_db}?mode=ro", uri=True)
+        else:
+            conn = learning_db
+        out: list[LearnerVerdict] = []
+        for table, unit_column, observation_column in _BETA_LEARNERS:
+            verdict = _inspect_learner(
+                conn,
+                table,
+                unit_column,
+                observation_column,
+                min_observations=min_observations,
+            )
+            if verdict is not None:
+                out.append(verdict)
+        return out
+    except Exception:
+        logger.debug("prior-distance check skipped", exc_info=True)
+        return []
+    finally:
+        if owns_connection and conn is not None:
+            try:
+                conn.close()
+            except Exception:
+                pass
+
+
+__all__ = ["DEFAULT_MIN_OBSERVATIONS", "LearnerVerdict", "check_beta_learners"]

package/src/superlocalmemory/server/routes/backup.py

@@ -737,7 +737,7 @@ a { color: #00D4AA; }
 <p class="sub">Back up your memories to a private GitHub repository.</p>
 
 <label>Personal Access Token</label>
-<input type="password" id="pat" placeholder="ghp_xxxxxxxxxxxxxxxxxxxx">
+<input type="password" id="pat" placeholder="GitHub personal access token">
 
 <label>Repository Name</label>
 <input type="text" id="repo" value="slm-backup" placeholder="slm-backup">