fix(tz): compute log filename date from configured timezone

[josemoreno801-netizen]

Summary

The daily log filename (memory-YYYY-MM-DD.log) is being stamped with UTC instead of the user's configured timezone, producing next-day filenames after ~20:00 local in zones west of UTC.

Root cause: scripts/log.sh:43 computes MEMORY_LOG_DATE with TZ="$REMEMBER_TZ" date +%Y-%m-%d, but REMEMBER_TZ isn't defined until scripts/log.sh:132 — 89 lines later. With an empty TZ= prefix, macOS/BSD date silently falls back to UTC. The log() function's internal timestamps were correct because they're computed lazily, after REMEMBER_TZ was populated. Filenames were wrong; content was right.

How I reproduced it

At 23:12 EDT on 2026-04-22:

• TZ="" date → 2026-04-23 03:12 UTC ✗
• TZ="America/New_York" date → 2026-04-22 23:12 EDT ✓

silent-stone/.remember/logs/ had both memory-2026-04-22.log and memory-2026-04-23.log side-by-side, with EDT-correct timestamps (23:12:06) written inside the 04-23-named file.

Fix

• scripts/log.sh — move REMEMBER_CONFIG / config() / REMEMBER_TZ definitions above MEMORY_LOG_DATE. Introduce _remember_date() so an empty REMEMBER_TZ uses system local time (omit TZ= prefix), not UTC. Change the fallback default from "Europe/Paris" to empty string. export REMEMBER_TZ so Python subprocesses inherit it.
• pipeline/_tz.py (new) — shared now() / today_str() / time_str() reading REMEMBER_TZ via zoneinfo. Invalid TZ names fall back silently to system local.
• pipeline/log.py — filename + timestamp now route through _tz.
• pipeline/shell.py — cmd_consolidate's "today" filter uses _tz.today_str(). This was a silent correctness bug beyond filenames: the filter excludes today's staging files from consolidation by substring-matching today against filenames. If Python disagreed with shell about "today," staging files could be consolidated prematurely.
• config.example.json — remove the "timezone": "UTC" landmine so users who copy the example verbatim don't inherit the same bug.
• scripts/post-tool-hook.sh — autonomous save-log filename now uses _remember_date (runs after log.sh is sourced).
• README.md — clarify that timezone is optional and defaults to system local.

Regression tests (all new)

• tests/test_tz.py — 7 tests covering REMEMBER_TZ resolution, empty/unset/invalid fallbacks, and the specific day-boundary frozen at 2026-04-23 03:12 UTC → 2026-04-22 23:12 EDT.
• tests/test_log_sh.py — 3 shell-level regressions running log.sh in subprocesses with a forced TZ=UTC system env and timezone: America/Los_Angeles config. Verifies filenames track the configured zone, and that an empty config correctly falls back to system local instead of UTC.
• tests/test_log.py — 2 new Python regressions.
• tests/test_shell.py — 1 new regression for cmd_consolidate's today-filter (with a mock_con.assert_not_called() guard rail so a future bug here can't accidentally hit the real Haiku API, as my first RED run unfortunately did).

Test plan

• pytest tests/ — 199 passed (baseline was 186), coverage 99.12% (≥80% enforced).
• Manual end-to-end: log.sh with timezone: America/New_York + system TZ=UTC → filename matches EDT date.
• Manual end-to-end: empty config + system TZ=America/Los_Angeles → filename matches LA date, not UTC.
• Round-trip: timezone: UTC + system TZ=America/New_York → filename matches UTC (verifies config actually drives the date, not system clock).
• Python: log() with frozen clock at 03:12 UTC + REMEMBER_TZ=America/New_York → memory-2026-04-22.log with 23:12 timestamp inside.
• Python: cmd_consolidate's today-filter correctly skips a staging file named for EDT "today" (04-22) even when UTC date is 04-23.

🤖 Generated with Claude Code

[fdaviddpt]

Review

Nice work — the bug diagnosis is solid and the fix is correct. The load-ordering issue in log.sh is a real problem and your reproduction is convincing.

Required before merge: rebase onto current main

PR #35 landed after you branched. It changed REMEMBER_CONFIG to use PIPELINE_DIR:

# Current main:
REMEMBER_CONFIG="${PIPELINE_DIR:-${PROJECT_DIR:-.}/.claude/remember}/config.json"

# Your PR still has:
REMEMBER_CONFIG="${PROJECT_DIR:-.}/.claude/remember/config.json"

This will conflict on merge. The fix is straightforward — just rebase onto main and use PIPELINE_DIR in the moved-up REMEMBER_CONFIG line.

Minor suggestions (non-blocking)

1. Frozen datetime helper is duplicated in test_log.py, test_shell.py, and test_tz.py. Consider extracting to tests/conftest.py as a shared fixture.

2. export REMEMBER_TZ — good catch, we didn't have that. Python subprocesses inheriting the configured timezone is important for correctness.

What's good

• _tz.py is clean and minimal. ZoneInfoNotFoundError → None (system local) is the right fallback.
• _remember_date() correctly avoids TZ="" — that's the actual fix for the bug.
• The cmd_consolidate today-filter fix catches a real secondary bug.
• Shell-level tests in test_log_sh.py are the most valuable — they reproduce the exact load-ordering bug.
• mock_con.assert_not_called() guard prevents accidental Haiku API calls during test failures. Smart.

Once rebased, this is ready to merge. 👍

[fdaviddpt]

Merged via local merge to main. Thanks for the solid bug fix and thorough test plan, @josemoreno801-netizen!