auto_recall hook queries non-existent 'heat' column instead of 'heat_base'

[PSGSupport]

Summary

mcp_server/hooks/auto_recall.py queries memories.heat, but the actual schema column is heat_base. As a result, every UserPromptSubmit hook fire fails with:

[cortex-auto-recall] FTS query failed: column "heat" does not exist
LINE 2:             SELECT id, content, heat, domain, agent_context,...
                                        ^

The hook then exits 0 with no injection — silent failure on every user prompt. Per-prompt memory injection is broken.

Reproducer

On a fresh Cortex install (3.14.5 + 3.15.0):

echo '{"prompt": "any text"}' | python -m mcp_server.hooks.auto_recall
# stderr: [cortex-auto-recall] FTS query failed: column "heat" does not exist
# stdout: (empty)

Affected lines

mcp_server/hooks/auto_recall.py, lines 144-170:

SELECT id, content, heat, domain, agent_context, is_protected,        # heat → heat_base
       ts_rank_cd(content_tsv, q) AS rank
FROM memories,
     plainto_tsquery('english', %s) q
WHERE content_tsv @@ q
  AND heat >= %s                                                       # heat → heat_base
  AND NOT is_benchmark
ORDER BY is_protected DESC, rank DESC, heat DESC                       # heat → heat_base
LIMIT %s

The heat_base column comes from pg_schema.py / migrations; the heat field returned by recall_memories() is a derived/effective heat (computed via effective_heat() PL/pgSQL function), not a stored column.

Fix

Three options:

1. Use heat_base directly (simplest, what I patched locally):

   SELECT id, content, heat_base, domain, agent_context, is_protected, ...
   AND heat_base >= %s
   ORDER BY is_protected DESC, rank DESC, heat_base DESC

   Caveat: doesn't apply lazy decay. The hook's purpose is fast filter, so raw heat_base is acceptable here.

2. Use the effective_heat() PL/pgSQL function to mirror the production recall path:

   SELECT id, content, effective_heat(heat_base, heat_base_set_at, plasticity, no_decay) AS heat, ...

3. Use the recall_memories() server-side procedure instead of inline SQL, matching the main retrieval path. Heaviest of the three but stays consistent with production.

I'd lean toward #2 — keeps the lazy-decay semantics that the docstring at line 28 advertises ("heat_base filter") while preserving the time-decayed scoring intent.

Environment

• Cortex 3.14.5 (also confirmed in 3.15.0 via inspection)
• Python 3.12 / Postgres 16 + pgvector
• Discovered while wiring auto_recall as a UserPromptSubmit hook in ~/.claude/settings.json

[cdeust]

Fixed in b102c51.

Root cause confirmed: memories.heat does not exist as a stored column. The actual storage column is heat_base; heat is only a derived field returned by effective_heat() and recall_memories(). The hook's broad except Exception swallowed the PG error so injection broke silently on every fire.

Fix: switched five SQL sites across three hooks (auto_recall.py, session_start.py, agent_briefing.py) to effective_heat(m, NOW()), matching production recall_memories's lazy A3 decay semantics. Note the reporter's described effective_heat(heat_base, ...) per-column signature is incorrect — the actual function takes a memories row composite (pg_schema.py:592).

Regression guardrails:

• tests_py/hooks/test_auto_recall.py — end-to-end smoke against real PG schema
• tests_py/infrastructure/test_schema_integrity.py — runs EXPLAIN on every memories-touching SQL blob in hooks/handlers; verified to catch the original bug when the fix is reverted.

Live smoke on Mac PG: 3 memories injected, exit 0, no errors.

[cdeust]

This one rewarded a deeper audit than the report itself asked for — your suggested option-2 (use effective_heat() for lazy-decay semantics) was the right call, and the implementation also surfaced 4 sister bugs at the same class-of-bug shape: session_start.py:107,147,150 (anchors + decisions fetch) and agent_briefing.py:270-276,298-304 (agent + team-decisions FTS passes). All 5 sites were silently failing on every UserPromptSubmit / SessionStart / agent-briefing fire across every platform. New tests_py/infrastructure/test_schema_integrity.py now parses every static SQL blob in hooks/handlers and asserts column existence — the abstraction barrier that prevents this class of bug from coming back. Five high-quality reports in one day from your team is a strong contributor signal; thank you. v3.15.1 shipped.