Chat-loop: tell the model how long since its last reply

[sysread]

SYNOPSIS

Extend the per-turn <datetime> tag with a since_last_response="..." attribute so the model can tell a live continuation from a thread the user revived hours or days later.

PURPOSE

The chat-loop already injects <datetime local="..." utc="..." zone="..." /> outside the <user_message> fence, so the model can answer "what year is it?" without guessing. It still has no signal for whether the user is mid-thought ("you just answered") or reviving an old thread ("about 3 days have passed"), and the user shouldn't have to spell that out every turn. Without it the model picks the wrong register on revival turns - charging straight back into the conversation when a brief reorientation would land better.

DESCRIPTION

Existing: buildDatetimeTag(tz) in chat-loop.ts builds the <datetime> tag once per round, formatted in the user's journalTimezone with local / utc / zone attributes. tagLastUserMessage prepends the tag outside the <user_message> fence on the latest user turn. The system prompt's DATETIME_BLOCK teaches the model that the tag is authoritative for clock questions.

Changed:

• new formatRelativeDuration(elapsedMs) - coarse, bucketed elapsed-time formatter: just now (<2min) / a few minutes / about N minutes / about an hour / about N hours / yesterday / about N days / about N weeks / about N months / over a year. Negative / NaN inputs collapse to "just now" so server-vs-browser clock skew doesn't surface as "in 4 seconds".
• buildDatetimeTag(tz, lastAssistantTimestamp?) picks up an optional anchor; on a parseable timestamp it appends since_last_response="..." to the existing tag. Omitted otherwise, so opening turns and unparseable timestamps ship the unchanged 3-attribute tag.
• ChatLoopOptions.lastAssistantTimestamp? plumbs the anchor in. Chat.svelte walks its messages array from the end for the latest role==='assistant' row that isn't in pendingDeleteSet (regenerate-from-here rows are about to be replaced and shouldn't count as "your last reply") and passes its created_at. Recomputed at call time so an auto-retry after a 429 picks up assistant rows persisted on earlier rounds.
• DATETIME_BLOCK in chat-prompt.ts gains a paragraph teaching the model to read the new attribute as a register-calibration signal, with explicit guardrails: do NOT quote the elapsed string back at the user verbatim, do NOT thank them for the gap.

How it closes PURPOSE: mid-thread turns now carry e.g. <datetime ... since_last_response="about 22 hours" />, so the model can decide between picking up mid-thought and re-orienting briefly without the user having to flag the gap. Synthetic ephemeral injections (intuition / context-recall <think> blocks) aren't persisted and therefore not eligible anchors - the semantic is "how long since you last actually replied to the user?", not "since any assistant-role row appeared on the wire."

Notes:

• Tag is recomputed every round (matching the existing buildDatetimeTag recompute), not once at send-time, so a 30s+ multi-tool turn doesn't ship a stale value.
• Attribute is omitted on opening turn (no prior assistant) and on unparseable timestamps, so a fresh thread never carries a misleading "just now". Existing chat-loop tests that anchor on the bare <datetime ... /> tag continue to pass unchanged.
• Buckets are intentionally fuzzy. The model uses this for register, not arithmetic, and "yesterday" / "about 3 days" obviously read as a stale-thread revival.
• Pure additive change to the wire shape; no DB schema change, no user-visible UI change. Updated docs/dev/chat.md accordingly; no docs/user/ change needed.

Test plan

• mise run check green (1069 tests, 0 svelte-check errors, 0 lint warnings)
• new describe('formatRelativeDuration', ...) block covers every bucket boundary including the singular "about an hour" branch, negative-ms clock skew, and NaN
• new chat-loop integration tests cover the three observable shapes: omitted on opening turn, present with bucketed value when anchored, omitted when the anchor doesn't parse
• manual smoke: open a thread, reply on it, leave it overnight, send another message, confirm the model's tone shifts toward reorientation rather than picking up mid-stream

---

Generated by Claude Code