fix(tool-registry): pass naive UTC datetime to fix PostgreSQL timestamp error on v1.82.0-stable

[joereyna]

Summary

Fixes continuous tool_registry_writer batch_upsert_tools error: ERROR: column "created_at" is of type timestamp without time zone but expression is of type text on v1.82.0-stable with PostgreSQL.

• datetime.now(timezone.utc).isoformat() produces a timezone-aware string ("2026-03-12T13:34:30+00:00") that PostgreSQL rejects when binding to a timestamp without time zone column via execute_raw
• Replace with .replace(tzinfo=None) to produce a timezone-naive UTC datetime that Prisma passes natively — no ::timestamp cast needed, no schema changes required
• Applies to both batch_upsert_tools and update_tool_policy call sites

Impact

Reported as ~185 errors/hour on Cloud Run + Cloud SQL PostgreSQL (v1.82.0-stable). Error fires every ~8 seconds for any deployment with tool calling enabled.

Related

• Contributor PR (closed, not merged): fix: cast ISO timestamp strings to timestamp type in tool_registry_writer SQL #23449 — same root cause, different fix approach (::timestamp cast); closed by author once they confirmed main had a better fix
• Fix on main is a full Prisma ORM refactor ([Feat] Add Tool Policies for AI Gateway  #22732) with schema changes (call_policy → input_policy/output_policy) — not backport-safe to stable
• This patch is the minimal safe fix for the stable branch: 2-line change, no schema changes, no new dependencies

Test plan

• tests/test_litellm/proxy/db/test_tool_registry_writer.py — 11/11 passing locally
• Deploy v1.82.0-stable patch against PostgreSQL and send a chat completion with tool use
• Confirm tool_registry_writer batch_upsert_tools error no longer appears in logs

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Mar 12, 2026 9:04pm

[greptile-apps]

Greptile Summary

This is a minimal, targeted fix for a PostgreSQL type-binding error in litellm/proxy/db/tool_registry_writer.py on the litellm_stable_v1_82_0 branch. The root cause was that datetime.now(timezone.utc).isoformat() produces a timezone-aware ISO 8601 string (e.g. "2026-03-12T13:34:30+00:00"), which PostgreSQL rejects when Prisma binds it to a timestamp without time zone column via execute_raw. Replacing it with .replace(tzinfo=None) produces a naive datetime object that Prisma serialises correctly for that column type.

Key points:

• Fix is applied consistently to both batch_upsert_tools (line 65) and update_tool_policy (line 143).
• No schema changes, no new dependencies — a 2-line surgical patch.
• The fix deliberately avoids datetime.utcnow() (deprecated in Python 3.12+), choosing the modern idiom datetime.now(timezone.utc).replace(tzinfo=None) instead.
• A minor pre-existing observation (not introduced by this PR): now is instantiated inside the for loop in batch_upsert_tools, so each row in a batch receives a marginally different created_at/updated_at timestamp. This could be pulled above the loop for consistency, but it does not affect correctness.
• The PR description references 11/11 unit tests passing, but the test file (tests/test_litellm/proxy/db/test_tool_registry_writer.py) is not part of the diff — only pre-existing or separately-run tests are verified, and a live PostgreSQL smoke test is still marked as TODO.

Confidence Score: 4/5

• Safe to merge for the stable branch — two-line change with a correct fix, though a live PostgreSQL smoke test is still pending.
• The change is semantically correct: datetime.now(timezone.utc).replace(tzinfo=None) produces a naive UTC datetime that Prisma's execute_raw can bind cleanly to a timestamp without time zone column. No regressions are introduced for SQLite deployments since Prisma handles naive datetime serialisation for both backends. Score is 4 rather than 5 only because the PR's own test plan marks the PostgreSQL smoke-test as unchecked, leaving a small gap in end-to-end verification.
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/proxy/db/tool_registry_writer.py	Two call sites changed from .isoformat() (returning a timezone-aware string) to .replace(tzinfo=None) (returning a naive datetime object). Fix is correct and consistent; minor pre-existing issue of now being created inside the loop is unrelated to this PR.

Sequence Diagram

sequenceDiagram
    participant LLM as LLM Response
    participant TRW as tool_registry_writer
    participant Prisma as Prisma Client
    participant PG as PostgreSQL

    LLM->>TRW: discovered tool names
    TRW->>TRW: now = datetime.now(UTC).replace(tzinfo=None)<br/>(naive UTC datetime)
    TRW->>Prisma: execute_raw(INSERT ... $8=now, ...)
    Prisma->>PG: bind now → timestamp without time zone ✅
    PG-->>Prisma: OK / conflict → increment call_count
    Prisma-->>TRW: success
    TRW-->>LLM: upserted

    note over TRW,PG: Previously: .isoformat() produced "2026-03-12T13:34:30+00:00"<br/>→ PostgreSQL rejected text → timestamp cast ❌

Loading

_{Last reviewed commit: a73b1ca}