hallucitools
Treat Claude Code sessions as inspectable, layered interaction logs. Reads the JSONL transcripts already on disk; or drop your own below — nothing leaves your browser.
Looking for just load-mark-export? halluci · the focused edition is 1/3 the size and does the core well — no LLM tracks, no hypergraph, no word frequency. This page below is the kitchen-sink workshop.
layers
L0 = observable JSONL fields · L1 = structural derivation · L2 = model self-report (hint, never truth) · L3 = your annotation (gold). v0 emits L0 + L1, writes L3.
live gallery
★ good call
? unclear
! notable
↯ disagreement
click #nn to add a note
live analyses · multiple tracks · llm-driven
Beyond structural projections, you can ask an LLM to examine the focused session. Multiple tracks run in parallel, stream token-by-token, and can be cancelled. Bring your own API key — it never leaves your browser except to the provider you select.
configure
Browser-direct fetch · key persisted to localStorage["halluci:llm:key"] · sent only as Authorization header to the chosen provider · never logged · clear with clear all annotations. The via gemini-gate provider option proxies through your own vault at gemini.hyperstitious.org — you authenticate to the vault, the vault holds the upstream key.
install · cli
$ git clone https://github.com/uprootiny/hallucitools $ cd hallucitools $ bb run sessions # list available sessions $ bb run review <uuid> # the same view as above, on real data $ bb run disagreements <uuid> $ bb run mark <uuid> 7 ★ "right call" $ bb run note belief "context window matters more than I thought" $ bb test # 5 tests · 21 assertions · 0 failures