Four NEW agent-native CLI harnesses that answer, in order, the time questions every AI agent runs into but usually improvises: what dates does this text actually contain, how long from A to B (with the time model stated, never assumed), what does this machine think time is, and what does this month actually look like. No timezone is ever invented silently.
Read-only date-token recon over one file or stdin: counted classes (ISO date, ISO datetime with offset, RFC-2822, month-name, US slash marked ambiguous and never resolved), invalid-ISO accounting (2026-13-40 stays counted but is reported separately, never mixed into ranges), an aware-only ISO min/max range (naive wall times are never compared against timezone-bearing values), first-three samples per class, and opt-in 10-digit epoch-candidate scanning with explicit --epoch-decode.
Duration math between ISO-8601 timestamps: exact aware-aware arithmetic across differing offsets (13:30+05:30 minus 08:00+00:00 is honestly zero), naive-naive clearly labeled wall-clock, and a MIXED pair or list REFUSED with the reason (exit 2) instead of silently assuming a timezone. List mode sorts one-per-line timestamps (comments skipped) and reports span plus true min/median/max gaps โ an even count averages the middle two and says so by returning a float.
Local system timezone truth and explicit-offset conversion: what the host says right now (offset seconds + display, DST state, tz names, epoch) with the honesty stamp tz_database: not-read (stdlib time only); converts timestamps under two labeled models โ aware inputs are exact, naive inputs are either stamped host-local ASSUMPTION (assumed_local: true) or given an explicit +HH:MM you name; --off on an aware input is refused (it would be a second opinion, not a source of truth); named offsets place naive timestamps (+03:00 12:00 > 09:00 UTC).
Calendar month recon and weekday scheduling math: real month length with honest leap rules (century rule applied โ 1900/2100 are NOT leap years; 2000 is), first/last weekday names, weekday counts, ISO week numbers of the first and last day, the stdlib text table, and --nth/--last weekday lookups ("third Monday of the month") with honest misses โ requesting the 9th Monday in a 4-Monday month is a named exit-1 finding, never clamped.
unzip CLI-Anything_Bundle11_TimeRecon_v1.0.zip
Copy any skill folder into your agent's skills directory (OpenClaw, Claude Code, Codex, or any Agent-Skills host), or run any harness as a plain CLI:
python3 datepeek/agent-harness/cli_anything/datepeek/__main__.py FILE --json python3 durcalc/agent-harness/cli_anything/durcalc/__main__.py A B --json python3 tzpeek/agent-harness/cli_anything/tzpeek/__main__.py now --json python3 calpeek/agent-harness/cli_anything/calpeek/__main__.py 2026-10 --json
Every skill ships its own self-test battery that live-fires the real CLI on the running host (verified 2026-10-19: 10 + 16 + 14 + 14 = 54 checks, 0 failed):
python3 datepeek/agent-harness/cli_anything/datepeek/tests/test_core.py python3 durcalc/agent-harness/cli_anything/durcalc/tests/test_core.py python3 tzpeek/agent-harness/cli_anything/tzpeek/tests/test_core.py python3 calpeek/agent-harness/cli_anything/calpeek/tests/test_core.py
Reads only the file or stdin you point at, computes with the stdlib, writes nothing, opens no sockets, sends no signals. Unreadable paths report honestly (exit 2); mixed time models and impossible months are named refusals, never stack traces.
Apache 2.0 ยท Python 3.8+ standard library only ยท Part of the CLI-Anything family on the SoulTech docs shelf ยท Support: [email protected] โ subject "BUNDLE11 TIME".
Built on SoulTech AI by Kiwann Williams. All Is One. One Is All.