Repository navigation
index_status(verbose) reports the live repo head_sha, not the indexed generation — misleading freshness signal #1559
Description
Activity
Thank you @PsyTech-Rob. Confirmed on current
main: verboseindex_statuscallsadd_git_context_json()against the liveroot_pathat request time, while the published generation does not persist the checkout identity it indexed. The reportedhead_shacan therefore look current even when the graph represents an older checkout.#1561 proposes the corresponding persisted-generation and bounded live-checkout evidence. This is routed priority/high; because it already has a linked implementation PR, it intentionally skips separate epic grouping.
Thank you, @PsyTech-Rob! The freshness poll you describe is exactly the client pattern that turns a live
head_shainto a false "up to date". This is covered by #1561 from @tmonestudio. That PR records the checkout the graph was actually built from at publication time, and reports anindex_status.freshnessobject: a verdict of current, stale or unknown, together with the reasons and a recommended action. That's the "indexed vs live" signal you asked for, and it explains why. I'm keeping this issue linked to #1561 and will close it when that lands.
Environment: v0.10.2, Windows 11, daemon mode,
--mode full.Symptom
index_status(verbose=true)returnsgit.head_sha(and branch info) read from the live repository, not from the indexed generation. Sequence observed:dbbec50…).34d6c6f…lands; the watcher has not yet reindexed — node count and graph content are still T0, andcheck_index_coveragestill reportsmetadata.generation= T0.index_status(verbose=true)already reportshead_sha=34d6c6f….A freshness poll keyed on
head_shatherefore declares "index up to date" while the graph is stale. Also mildly surprising: non-verboseindex_statusomits thegitblock entirely, so the same poll silently compares against an empty field forever.Suggestion
Expose the indexed generation's commit alongside the live one — e.g.
git.head_sha(live) +git.indexed_sha(what the graph was built from; coverage already exposesmetadata.generationas a timestamp). That gives clients a truthful staleness signal:indexed_sha != head_sha⇒ stale.