# People, recent correspondents first, and the four-run day (#1943, stage 3) Since #1942, investigating a person no longer digests all of their mail: our code gathers it, writes it to evidence files with an index, and one turn searches those files for what the page needs (#1850). This stage decides **who** is investigated next, **over which window**, and **when** the daily round does it. Before it, the busiest person came first and was read over the full 150 days every time, so the round's top pages never fit a day's calls (#1723). ## Commands ```sh co rem investigate people --list # the order and the cost; nothing read or spent co rem investigate people # the next 5, last 14 days first co rem investigate people --limit 1 # just the most recent correspondent co rem investigate people --recent-days 7 --limit 0 # everyone written to this week ``` `co rem investigate people` is the command it always was; its order and windows changed. It states the cost first — one model call a person, the mail the map counted for the full investigations, how many are updates, and what one full investigation measured on this machine — then investigates one person after another. It stops starting people at the weekly Codex budget, at this run's `--budget`, or at the floor kept for your own work, and ends by saying how many people are left. `projects`, `orgs`, `skills` and `all` keep their order (most mail or sessions first). A single page and `me` are unchanged. ## Order People the owner wrote to at least once come first, then people who wrote more than once and were never answered, then one-mail contacts (#1974). Within each, the last `--recent-days` (default 14) first, then by the map's mail count decayed by the weeks since the last mail. Left out: the owner (`investigate me`, listed first while never investigated), addresses that may be the owner's, automated senders, and vendors whose domain also sends notices. The overview `co rem investigate` and `people --list` read the same queue. Counts are the map's, a floor for what the 150-day read finds; see [rem-investigation-correctness.md](rem-investigation-correctness.md). ## Windows: only what is new | The person | Mode | Window read | |---|---|---| | Never investigated, or investigated but still unfinished and not in the last 7 days, with nothing newer | full | 150 days (`--days`) | | Mail arrived after the page was last investigated | update | the days since that investigation | | Investigated since their last mail | — | not in the queue | "Last investigated" is the exact time this stage started that person's investigation (`.state/people/investigated.json`), or, for a page investigated another way, the date on its status line. So mail that arrives later the same day is still new. ## What is new since the last run `correspondents_since()` lists each connected mailbox **once** since the previous run (metadata only: dates, senders, recipients) and matches the addresses to people pages, instead of searching the server once per person. The newest date per person goes to `.state/people/activity.json`, the cursor to `.state/people/refresh.json`; both are owner-only (`0700` directory, `0600` files) and hold no subject or body. ## Where the network is used Nothing here gives the model the network. Every notebook run is confined (Codex `--sandbox workspace-write`, no network; Claude Code `acceptEdits`), so the model cannot call `co gmail` or `co outlook` itself. Everything that needs the network is our code, before the model: the listing above, and the investigation's gather — the private archive `co rem init` saved first, then the server for mail it lacks, bodies and attachments — which #1942 writes into the files the one turn searches with `rg`, `sed` and `ls`. Giving the run the network would reopen what the confinement closed (`runner.harness_flags`); it is not proposed. ## The daily round: four runs, two jobs (#1723) Each scheduled run maintains first, as before. Then: - **The first run of the local day finishes unfinished pages,** people, projects and organisations together, **most recent activity first,** within a reserved share of the day's calls (8). A person is one investigation over their window. A project or organisation takes the rest of the share and ends the portion, as before. - **Every later run follows what is new:** the people the listing found with mail since the previous run (each read over only the days since their last investigation), and the projects with new messages you typed (`co rem projects write`'s path, `write_pages(since=…)`: an update from the new messages, or a first write only for a project active since the last run). At most 5 pages, one call each, within the day's call cap. A run with nothing new calls no model and says `nothing_new`. Both stop starting pages at the weekly budget or the 70% floor, and both record in `co rem logs` how many pages are left. Which run is which is decided by what already ran today, not by the clock, so a machine asleep at the first slot still gets its unfinished portion from whichever run comes first. The schedule has six times by default; the first is the unfinished portion and the other five follow new material, which costs nothing when nothing arrived. ## Measured One real run on 2026-09-30, on this machine, against a temporary notebook holding one fresh page for the owner's busiest recent correspondent, the same person as #1850's baseline (157 mails in the 90-day map). Runner Codex, `gpt-6-luna`, `co rem investigate people --limit 1` (a full investigation, 150 days), after #1942. | | Before (#1850 / #1884, digest path) | This run (#1942's evidence files) | |---|---|---| | Material | 157 mails, 1.78M characters | 265 mails (129 Outlook, 136 Gmail) with 99 attachments, 2 coding messages; 1.85M characters in 366 files | | Model calls | 31 (30 digest pieces + 1) | 1 | | Input tokens | 8.2M | 1.93M (1.76M of them cached) | | Output tokens | 503k | 56k | | Time | 3 h 24 min | 15 min 3 s, gather included | | Codex week | did not move off 8% | 25% → 26% | The accepted page has all 11 sections, 88 citations over 17 sources, and 2 lines left `Unknown` (phone, company). The role on the page appears in only 2 of the 265 mails and was found and cited. No labelled phone number appears in anything the person wrote in the window, so `Phone: Unknown` is the right answer. An update for the same person reads only the days since this run. For comparison, a people-only prototype of this stage (an evidence folder per person with quoted threads capped at 4,000 characters, and a skill that told the agent to stop after about 30 tool calls) wrote the same person's page in 0.67M input tokens and 5 min 40 s of model time, with 52 citations. It was dropped so that there is one investigation path, #1942's; the gap is mostly the quoted threads and attachments #1942 keeps whole.