Repository navigation
v1.7.0
This release grows the Agents trace viewer from a flat Gantt into a genuine trace waterfall — a session narrative strip, prompt-segmented bands, nested sub-agent rows, and a where-did-the-time-go breakdown — and corrects a class of cost double-counting for resumed/forked sessions so headline spend and every drill-down now reconcile exactly. It also adds Claude Sonnet 5 and legacy pre-Opus-4.5 pricing, ships prebuilt release binaries, and hardens claudit serve's memory footprint.
Changed
- Reported cost drops for anyone with resumed or forked sessions. A resumed or forked session replays the prior transcript into a new JSONL file — same
message.idand identical usage, but a freshuuid/sessionId— and the old per-file dedup billed every replayed generation once per file it appeared in. Spend is now deduplicated bymessage.idat every aggregation point, so if you resume or fork sessions your totals will read lower than before (they were previously inflated). This is a correction, not a regression — see the three Fixed entries below for the mechanics. Single-session raw traces still show a forked session's full replayed context.
Added
- Timeline is now a trace waterfall. The flat Gantt became a nested, prompt-segmented waterfall over three changes:
- Prompt bands. Each prompt's turns render inside a labeled full-height band, so you can see which prompt a stretch of activity belongs to. Segmentation reuses the same boundaries as the Conversation lens, so the two always agree.
- Nested sub-agents. Sub-agent rows now re-parent under the exact turn that spawned them (via
parent_tool_use_id), with connector elbows and real depth indent — grandchildren nest one level deeper. - Progressive disclosure. Rows collapse by default (agent bar → turn spans → tool sub-spans), dropping a 13-agent / 754-tool session from ~1,300 always-on SVG nodes to ~210. Expansion state survives live repaints, and a cross-lens jump auto-discloses its target.
- Session narrative strip. Above each session's waterfall, one clickable row per prompt segment shows the prompt snippet, turn/tool counts, spawned-agent chips with a pass/fail glyph, duration, and cost. Clicking scrolls the waterfall to that prompt's band and flashes it. Sub-agents attribute wholly to the segment that spawned them (spawn chain walked to the root prompt, time-containment as fallback); the outcome glyph reads the same
error_countthe waterfall's error pips do, so the strip can never disagree with it. - Where-did-the-time-go rollup. Each Timeline session card gains a stacked generation / tool / idle time bar (colors shared with the waterfall segments), and a new Insights → Time panel breaks the same split down per session (graph scope) or per agent, top-10 with an overflow note. Sub-agent waterfall rows also gain a pass/fail outcome chip.
- Full turn reasoning and message on demand. The detail drawer's Reasoning and Message fields get the same show-full affordance tool I/O already has — the untruncated thinking and narration stream from the session JSONL on request, and stay sticky across live repaints. Static reports degrade to the snippet-only note.
- Pricing: Claude Sonnet 5.
claude-sonnet-5andclaude-sonnet-5[1m]added at the introductory $2/$10 base rates (in effect through 2026-08-31; a comment flags the revert to $3/$15 on 2026-09-01). Verified against the pricing page on 2026-07-03. - Pricing: legacy pre-Opus-4.5 rates. Sessions from before ~late 2025 were pricing at $0 because the exact-match table had no entries for the dated legacy ids. Added Opus 4.1/4 ($15/$75), Sonnet 4 ($3/$15), Haiku 3.5 ($0.80/$4), and the retired 3.x models at their last published rates (marked historical).
- Prebuilt release binaries. Every
v*tag now attaches cross-compiled archives for darwin / linux / windows on amd64 / arm64 to the GitHub Release — no localgo buildrequired. net/http/pprofendpoints on the serve mux. Registered explicitly on the server mux (nothttp.DefaultServeMux); loopback-only bind is already the default and the report already exposes prompt text, so this adds no new exposure class.
Fixed
- Resumed/forked replays no longer double-count in the headline. Dedup by
message.idat each cost/token aggregation point — report, serve, diff, and thewatchrolling-totals andwatch --allcombined panels. On the real corpus this removed exactly the duplicated volume (−53M tokens, −291 turns). Legacy single-line transcripts with an emptymessage.idcan't be keyed and always count. - Drill-downs now reconcile with the deduped headline. The per-session (
BuildSessionTimelines) and Agents (BuildAgentGraph) views still summed straight from raw turns, so expanding either showed spend that didn't match the headline (+291 turns / +$36 on the corpus). A new deterministicReplaySet— the lexicographically-smallest source file is canonical, the rest are flagged replays — is now shared by both builders; every drill-down reconciles to the penny. - Per-session attribution is now deterministic across views. The headline aggregator deduped by first-
Add-order, but snapshot turn order is non-deterministic (concurrent parse), so a forked session's cost could land on a different session in the Cache-by-session view than in the drill-downs, and shift run-to-run. The aggregator now dedups via the sameReplaySetthe drill-downs use, making per-session/project attribution stable and consistent everywhere. (Also disclosed in the Tools guide: per-tool cost attribution overlaps by design — a multi-tool turn counts under each tool, so the column sums to more than total spend.) claudit servememory no longer balloons on large corpora. Two leaks: the render cache evicted by entry-count only while its(query, section, generation)keys stranded every older-generation entry as permanently-unreachable dead weight — on a corpus where the Agents payload is ~186 MB per entry, a handful of stranded entries pushed RSS past 3 GB. The cache now tracks total resident bytes and evicts to a byte budget (default 256 MiB) in addition to the count cap, and each store eagerly sweeps entries from older generations. Additionally, turn thinking/text and prompt text are now capped to 2000-rune snippets in the snapshot and on the wire (the dominant share of that 186 MB payload), with the full text one request away. And serve mode now sets a 1.5 GiB soft memory limit (debug.SetMemoryLimit) so Go returns cached-payload headroom to the OS; an explicitGOMEMLIMITis respected.--bindvalues that already carry a port are rejected up front.--bind 127.0.0.1:8791used to be silently composed into127.0.0.1:8791:8787and fail at listen time with a confusing too-many-colons error (and a false non-loopback warning). It's now validated right after flag parsing and points you at--port. The listen address is composed withnet.JoinHostPort, so bare and bracketed IPv6 hosts both keep working.- Client-canceled API requests are no longer logged as errors. A browser fires one request per report section and cancels them on reload/navigation; that surfaced as
context.Canceledlogged at ERROR with a 500 written to an already-closed socket. Those are now swallowed — no ERROR log, no 500. - SDK badge standardized. The SDK-origin pill is pinned to the top-right of the Tree/Timeline/Conversation left session menus and dropped from the Timeline Gantt header (it was shown twice); the Insights token-composition bar is reordered to the canonical input → output → cache-write → cache-read.