fix(integrations-claude-code): label 'Current time' as UTC in recall context

[valda]

Problem

The Hindsight recall hook injects Current time - <ts> into the <hindsight_memories> block on every UserPromptSubmit, but the timestamp is rendered as UTC with no timezone label. Client-side LLMs running in non-UTC timezones routinely misread the value as local time.

Concrete repro #1 (Claude Code, JST / UTC+9):

• Local wall clock: 2026-05-11 08:55 (morning)
• Injected line: Current time - 2026-05-10 23:55
• Result: the model reasons that it is late evening and replies things like "sounds like a good place to wrap up for the day" even though it is morning in the user's actual timezone.

Concrete repro #2 — much worse, observed since the original report:

• A long-running Claude Code session had an earlier Current time - <ts> line still sitting in its scrolled-back context (UTC value).
• Later in the same session (local time now ~9 hours ahead in JST), the model re-read that scrolled-back stamp while reasoning about command history, parsed the UTC value as local time, and concluded that a different user had executed the same command roughly 9 hours earlier — i.e. it invented a second actor to explain the apparent time gap.
• It then opened a security investigation into the "unauthorized duplicate execution." Sounds like a joke but it actually happened.
• The 9-hour delta is exactly JST→UTC, which is the smoking gun: same root cause as repro Fix dependencies and docker #1, but the failure mode is paranoia / false security incident rather than just a wrong greeting.

Fix

format_current_time() now returns 2026-05-10 23:55 UTC. The UTC suffix matches the convention already used by the opencode integration:

https://github.com/vectorize-io/hindsight/blob/main/hindsight-integrations/opencode/src/hooks.ts#L117

so this is parity rather than novelty.

Scope

• hindsight-integrations/claude-code/scripts/lib/content.py — the actual change (4 lines)
• hindsight-integrations/claude-code/tests/test_content.py — regression tests (suffix + format-shape)

The same format_current_time() shape exists in the codex and openclaw integrations and is theoretically affected by the same issue. This PR intentionally leaves them out:

• The misreading behaviour has only been observed and confirmed against Claude Code by the reporter.
• hindsight-integrations/codex/scripts/lib/content.py carries pre-existing ruff-format drift in unrelated regions, so editing the same file would either (a) fail verify-generated-files CI or (b) require reformatting ~120 unrelated lines — out of scope for this fix.

Happy to open follow-up PRs for codex / openclaw if you want the parity extended.

Test plan

• pytest hindsight-integrations/claude-code/tests/test_content.py — 57 passed, including the two new tests.
• ruff check --config ./ruff.toml hindsight-integrations/claude-code/scripts/lib/content.py — clean.
• ruff format --config ./ruff.toml --check hindsight-integrations/claude-code/scripts/lib/content.py — already formatted.

Rebase note

Rebased onto 8b10231b Release v0.6.2 to stay current with the v0.6.2 release line.

[nicoloboschi]

nice catch!