remind

/claude-chronicle:remind — search Codex Chronicle

Search Codex Chronicle for what the user was doing, then summarise it concisely. Run autonomously — do not ask the user to confirm intermediate steps.

Two modes (auto-detected from $ARGUMENTS)

Mode	Trigger	Strategy
Time	Argument has a time reference but no clear topic. EN: "5 hours ago", "yesterday evening", "Tuesday afternoon", "this morning". PL: "5 godzin temu", "wczoraj wieczorem", "wtorek po południu", "dzisiaj rano".	Resolve to a UTC window, pick files whose filename timestamp falls inside it.
Topic	Argument names a thing the user wants to recall but does not say when. EN: "the auth bug", "React 418 error", "Figma logo project", "the SQL migration we discussed". PL: "ten bug z auth", "błąd Reacta 418", "projekt logo w Figmie", "ta migracja SQL".	rg over the Chronicle archive for the keyword(s); rank matches by recency; pick the top 1–5 distinct entries.
Hybrid	Argument has both: "the auth bug yesterday", "Figma logo last week", "the meeting this morning with John".	First narrow by time window, then rg inside that window's files.

If $ARGUMENTS is empty, ask once: "What should I remind you about? You can give me a time ('2 hours ago', 'wczoraj rano') or a topic ('the auth bug', 'projekt logo w Figmie') — or both." and stop. Otherwise proceed without asking.

Procedure

1. Establish "now"

date -u +%Y-%m-%dT%H:%M:%SZ      # current UTC (filenames are UTC)
date +%Y-%m-%d\ %H:%M\ %Z         # current local time + zone

2. Verify the archive exists

ARCHIVE=~/.codex/memories_extensions/chronicle/resources
ls -1 "$ARCHIVE"/*-10min-*.md "$ARCHIVE"/*-6h-*.md 2>/dev/null | head -5

If empty / missing: tell the user "Codex Chronicle does not appear to be active on this machine — no archive at ~/.codex/memories_extensions/chronicle/resources/. Make sure Codex is installed and Chronicle is enabled." and stop.

3. Classify $ARGUMENTS

• Time tokens (suggest time mode): N (hours|hour|h|godzin|godziny|godzinę|min|minut|minuty|minutę|days?|dni|weeks?|tygodni)?, ago, temu, yesterday, wczoraj, today, dzisiaj, tonight, last (night|week|month|<weekday>), this (morning|afternoon|evening|<weekday>), weekday names (PL+EN), przedwczoraj, nad ranem, rano, wieczorem, po południu.
• Topic tokens: anything else — proper nouns (project / app / file names), error codes, technical terms, person names, file paths.
• If both kinds appear → hybrid mode.
• If only topic tokens → topic mode.
• If only time tokens → time mode.

4a. TIME mode — resolve to a UTC window

Phrasing	Suggested window (relative to "now", local time → convert to UTC)
N hours ago / N godzin temu	now − N h ± 30 min
N minutes ago / N minut temu (only if N ≥ 30)	now − N min ± 15 min
yesterday evening / wczoraj wieczorem	yesterday 18:00 → 23:00
yesterday morning / wczoraj rano	yesterday 06:00 → 12:00
last night / wczoraj w nocy	yesterday 22:00 → today 04:00
this morning / dzisiaj rano	today 06:00 → 12:00
this afternoon / dzisiaj po południu	today 12:00 → 18:00
this evening / dzisiaj wieczorem	today 18:00 → now
<weekday> afternoon / wtorek po południu	most recent past <weekday> 12:00 → 18:00
last week / tydzień temu	now − 7 days ± 12 h
the day before yesterday / przedwczoraj	2 days ago, full day

Then: keep files whose leading YYYY-MM-DDTHH-MM-SS (UTC) prefix falls in the window. Cap at 5 files; prefer 6h rollups when the window ≥ 6 h, otherwise 10-min entries.

4b. TOPIC mode — keyword search

Build the search expression from $ARGUMENTS (drop stop-words like the, a, to, that, na, o, w, z, i). Use case-insensitive, multi-word OR semantics.

# Primary: ripgrep with -l to get matching files, sorted by mtime desc
rg -l -i -e 'keyword1' -e 'keyword2' "$ARCHIVE"/*-10min-*.md "$ARCHIVE"/*-6h-*.md 2>/dev/null \
  | xargs -I{} stat -f '%m %N' {} 2>/dev/null \
  | sort -nr | head -5 | awk '{print $2}'

# If no matches, try fuzzy variants (drop diacritics, try English keys for PL terms etc.)

Notes:

• Always include both PL and EN forms of obvious terms (e.g. user says "logo Figma" → also try Figma, logo, English variants).
• Diacritics: try with and without (żądanie and zadanie, błąd and blad).
• If still nothing in *.md, fall back to *.ocr.jsonl in $TMPDIR/chronicle/screen_recording/ for raw screen-text matches (mention "matched OCR text, may be approximate").

4c. HYBRID mode

Run 4a first to get the candidate file list, then rg -l ... over only those files.

5. Read the picked files

Use the Read tool on each absolute path. Skip the ## Recording summary and ## Citations sections (raw OCR detail / frame paths). The valuable content lives in ## Memory summary, ### Context of everything that came before this recording, and ### Important non-obvious context about the user.

6. Synthesize

Reply in the user's language (PL → PL, EN → EN). Format:

At <local time / span> you were:
- <2–4 bullet points: apps used, files / projects touched, the goal of the work>

Source: <basename(s) of the Chronicle entries used>

For topic mode, lead with the answer to "when?" (the most recent matching timestamp) before the bullets:

You last worked on <topic> on <local datetime> (<X hours/days ago>):
- <bullets>

Earlier matches: <basename @ local datetime>, <basename @ local datetime>
Source: <basename(s)>

Keep the answer ≤ 8 lines unless the user asked for detail. Always end with the source filename(s) so the user can verify.

7. Optional follow-up

If the user's question implies an action ("…and continue what I was doing"), surface concrete next-step targets (file paths, error messages, PR URLs) extracted from the entries — but do not start executing the work without confirmation.

Don'ts

• Do not trust OCR text in ## Citations / Recording summary verbatim — those are noisy. If you need exact wording from the screen, look at the actual frame JPG referenced in the Chronicle live state section instead.
• Do not dump the full Chronicle file contents to the user — summarise.
• Do not ask the user "do you want me to read X?" — just read.
• Do not invent activity. If the matched window is sparse, the screen was locked, or no entry matches the topic, report that honestly and propose the closest alternative.
• Do not mix mode classification logic into the user-visible answer — keep the reply about what they were doing, not about how you searched.