Semlith 0.26.0

The whole portal, free, in the binary

Thirteen pages, no plan, no key, no lock and no price. Everything the v4 design
draws now renders from your own store, and everything under it answers from the
terminal and over MCP as well — three new commands, three new tools, and a page
for each.

Impact. The graph read backwards: from a symbol, every caller and every
file that reaches it, breadth first to a hop limit, each edge carrying the
support class the store already recorded. semlith impact, semlith_impact
and a page of its own. This left the free product in 0.13.0 to be sold and
never was; the monetization hold of 2026-09-15 made the binary free whole, so
it is back with no key of any kind.

Trace. A chain turned into evidence: the answer sentence, the hops, and one
supporting source line per hop, each marked a supporting fact or a candidate to
corroborate, with a control that copies the block for a review. semlith trace,
semlith_trace, and the panel. Nothing re-walks the graph — it reads the chain
semlith path produced, so the two cannot describe one path differently.

The path finder, drawn. semlith path has answered from a terminal since
0.14.0 and had no page. It has one now, with the seams, the Prefer verified edges and Strict toggles, the refusal that names the rule, and the
hypothesis line on a chain that is one.

Map. The store's subsystems as a list: label, member count, three hubs with
file and line, and the strongest edge out of each. Label propagation over the
settled calls and imports edges, seeded by name order and capped, so the
same store gives the same communities twice. No new dependency, no model, no
second picture beside the canvas.

Inside the index. The Index page is now what the design calls it, and
carries the language mix, chunks by month indexed, and Graph health: call edges
by support class as one bar, the top unresolved call targets, how many names
carry several definitions, and how many languages carry edges — read from the
counts semlith stats reports rather than recomputed, so the page and the
terminal cannot disagree.

The ledger, session by session. A table of every agent session: when it was
last seen, who it was, how many reads, net tokens after refunds, what that
would cost at a model you pick, and the tier of the figure. Filters by client
and tier, sort, pagination, and export as Markdown, CSV or JSON of exactly the
rows on screen.

Session replay. What the agent did after each answer — read the whole file,
grepped anyway, or edited — read from this machine's Claude Code transcripts.
Off until the Privacy page turns it on, because those files belong to another
program, and nothing it reads leaves the machine.

Reports. Five, generated here: retrieval savings, AI access audit, change
brief, index health and knowledge gaps, each as Markdown, CSV, JSON or
print-styled HTML. semlith report and semlith_report write the same bytes
the page exports. PDF is your browser's own print of the HTML; no PDF writer
ships in the binary.

Cloud. One page describing an optional hosted service, in its
not-connected state and nothing else. This binary has no cloud command and
opens no connection to any host.

Four defects, all of them older than this release

• #122, and this one had survived a whole retrieval release: a query
  describing a picture ranked a source file above the picture. The image list
  was fused on the flat curve while a question's vector list is steep, so a
  confident CLIP match could not reach a passing text chunk's score however
  well it matched — measured at 0.049180 against 0.093317 on the four-shape
  fixture. The image list is now steep for every shape and the confidence rides
  in the weight, which leaves an unconfident image scoring what it always did.
• #120: POST /api/index recorded whatever path it was handed as a root of
  the store before the run applied the boundary, so the refusal that
  docs/security.md has promised since 0.20.0 could never happen. The boundary
  is read first now, and a request naming no store indexes the path into the
  store home::resolve picks for it rather than into whichever store happened
  to be writable.
• #121: a forwarded semlith_index reported a skipped count with no
  reasons while the same call answered in process named every one. One call in
  two places may not say two different things about it.
• #124: a repair note printed its backticks as characters. Fixed where the
  string is written, so the terminal stops printing them too.

Numbers

Nothing in this release touches the retrieval path, so the 0.25.0 figures were
re-measured on this binary rather than restated. Three runs on the
sealed thirty, median of three, spread zero:

	0.25.0	0.26.0
hit@8	29 / 30	29 / 30 (96 %)
hit@3	27 / 30	27 / 30 (90 %)
hit@1	24 / 30	24 / 30 (80 %)

Every identifier question in the split is in the top three — 11 of 11, which
is how many the sealed thirty holds; the 0.25.0 record said 12 of 12 and no
question file has changed since, so that figure was wrong rather than this
one. Wrong-yes is 0. By class at k=1 / k=3 / k=8: concept 7 / 10 / 12 of 13,
identifier 11 / 11 / 11 of 11, multi-hop 6 / 6 / 6 of 6.

tools/list is 5 840 bytes, about 1 460 tokens, for sixteen tools — against
4 473 bytes and about 1 119 for thirteen.

Performance, measured on the reference laptop with the scale test run on its
own: warm query p50 13.6 / 13.3 / 141.4 ms at 700 / 7 000 / 70 000 chunks,
peak resident memory while indexing 246 MB, idle resident memory 151 MB at
700 chunks and 150 MB at 70 000 — flat across a hundredfold corpus, which is
the promise that opening a store loads no vectors. Searching grows instead,
by 241 bytes per chunk. The idle watcher costs 0.01 s of CPU over 60 s and an
edit becomes searchable in 1.29 s.

Run the measure suite the way cargo test runs it by default and three heavy
tests share one process: the same 7 000-chunk query reads 13.3 ms alone and
47.2 ms beside the others, and two resident-memory assertions fail on the
contention rather than on the code. Every figure above was taken with
--test-threads=1.

The tool list an agent pays for once per session grows with the three new
tools, and the gate on it moved from 1 120 tokens to 1 600. The Agents page
states what this binary's own list measures rather than a number written down
when it was last checked.

Install

macOS and Linux

curl -fsSL https://raw.githubusercontent.com/semlith/semlith/main/install.sh | sh

Windows

irm https://raw.githubusercontent.com/semlith/semlith/main/install.ps1 | iex

Or download the archive for your platform below, unpack it, and put
semlith on your PATH. Intel macOS has no prebuilt binary: ONNX
Runtime no longer publishes x86_64 macOS builds.

Verify a download against SHA256SUMS before running it.