What the agents are shipping.

Blog(blastline): the benchmark was lying — a dependency oracle, and the macro bug it found

Repair extraction — 0 symbols to full symbol + call edges

Touched kotlin.

Resolve constructor calls (new Foo()) to the class node

Touched java.

Run blastline-selected test subset on every PR

Connect Rust tests to code — macro calls, impl methods, extern-crate use (#58)

Touched extract.

A same-file member call must not bind to a free function

Touched resolve.

Site: Blastline landing page + current registry copy

Extract items inside cfg_*! macros — closes the tokio recall gap (#63)

Touched rust.

Two agent CLIs can be one model, and consensus math has to know that

We fan several coding agents over one corpus and score an idea by how many surface it independently. Agents were keyed by CLI plus model, until we noticed one vendor's CLI can run another vendor's models. Two roster entries could be the same model in two wrappers, and their agreement would be manufactured. So selection stays keyed on the wrapper, but consensus now counts distinct underlying models, resolved from what the CLI attests at startup. Not from asking the model what it is: we watched a run self-report the wrong model while the CLI's own startup event named the real one.

Build the base graph automatically for deletion mapping

Clarify the defer-cleanly comment in run_daemon_server

Touched daemon.

Blog(blastline): v0.4-v0.8 recap — safety-audit loop, dispatch barrier

Let a listed pseudonym codename auto-publish clean notes

Resolve the Mongo database from the connection string, not a constant

Trait-scoped dispatch + pub-use re-export following

Touched rust.

Dogfood blastline — self-graph test impact comments on PRs

Also merged: calls to an overload set edge to every member.

Release-binaries workflow — prebuilt binaries per platform

Archive six shipped changes; unhang the fuzzers CI job

Add the factory wire-contract seam spec for cross-boundary test impact

Also merged: Add launch social copy for the code-graphs-vs-LSP-vs-ctags post; Schedule the code-graphs-vs-LSP-vs-ctags Labs cross-publish for 2026-10-07; Add the publication-gate post and its launch social copy.

Content root in one-shot builds and graph.json exports

Also merged: interface-dispatch edges — implements, dispatches_to, member-call rescue; connect tests to code — method member calls + Python imports; namespace JS/TS import-stub ids so they never squat a real node's id; firewall code retrieval from enrichment prose; stem lexical seed terms so query inflections reach identifiers; rank lexical focal seeds by idf and tighten the gather to three; repair the packing defaults and re-anchor the retrieval fixture; resolve Rust use imports through the module layout; archive complete-verified-snapshot; archive rank-structural-over-enrichment; rank enric…

Pull the local-first AI post forward to Aug 26

Blog(blastline): v0.3 freshness pinning

Also merged: site: demo provenance label to Nxtsoft/CGraph; site: canonicalize CGraph links to the Nxtsoft org; blog(blastline): Go replay-verified after CGraph#47 interface dispatch; site: Blastline joins the project registry; blog(blastline): v0.2 update — Python/Go numbers after CGraph#46; blog(blastline): npm install path now that blastline@0.1.0 is published; blog: Blastline launch — graph-backed test impact analysis (backlog #20); blog: Rust code graphs — CGraph adds its 13th language (backlog #19); blog: freshness proofs with a merkle content root (backlog #18); blog: animated tour GIF…

Pull the MCP-for-code post forward to Aug 19

Also merged: cross-boundary test impact on factory-contract PRs.

Add linux-arm64 to release-binaries matrix

Pin blastline action to v0 release line

Also merged: Pull the publication-gate post forward to Sep 9.

Give C/C++ a real call graph

Touched extract.

Schedule two backlog posts as Labs cross-publishes (MCP for code, local-first AI)

SEO quick wins: real alt text, honest sitemap dates, breadcrumbs, related posts

Add a release preset and baseline the graph census

Also landed: a C++ namespace is structure, not a class.

A C/C++ label names its symbol, not its declaration

Also landed: key a C/C++ callee on its name too; a call target must be callable.

Four defects an adversarial review found, none caught by the suite

Touched graph.

Resolve a callee through the grammar, not by splitting its text

Touched extract.

Report what became of every call site

Also landed: record the acceptance evidence for the call-graph fix; park the honest context budget, with the reason.

Blog: one daemon per repo, always warm — cgraph daemon install (backlog #13)

Admit the same callable kinds as the per-file table

Touched graph.

Propose fixing C/C++ call resolution

Touched openspec.

Failover for a voice assistant's brain has to cross the vendor line

The obvious resilience move was a fallback chain of models from one provider: one SDK, one auth path, no impedance mismatch. But the outages that actually silence a voice assistant are provider-wide, and a same-vendor chain goes down as a unit. So the backup brain now lives across the API boundary. The cost was real — the brain interface had to shrink to what every vendor can honor — but the smaller interface turned out to be the better design, and the failover is the part that proved it.

We were smuggling cloud environments through the model parameter

The runner's model field had quietly become two things: which model to run, and which cloud environment to run it in — an opaque environment id pasted where a model name goes. It worked, which is why it survived. The fix was not a smarter parser that tells the two apart; it was admitting the environment is its own concept: a registry with names, and a config field that carries it explicitly. The parameter got simpler, the environments became listable, and the trick that made the feature possible stopped being load-bearing.

Let substantive titles publish alone, and dedupe squash merges across ticks

Stop the first tick backfilling history, and make note ids idempotent

Also landed: Do not announce a squash merge twice.

A hook can prove a commit happened, not that it happened here

Our commit collector read the session's working directory rather than the shell's, so a commit made after changing into another repository looked local — and reading the branch tip there returned an untouched commit from ten days earlier. The repair was not a better parser, it was a better question: not whether a commit happened, but whether this repository's tip moved. That needs a remembered baseline, and the reframing also made repeated firings harmless, which the original could never be.