nicolasacchi

by kb

Provenance-aware code browsing via a kb-code daemon: a PreToolUse hook (kb-code-why) that injects trailer/exact-confidence "who wrote this and why" context right before an Edit/Write touches a line kb-code can attribute, LLM-free and fail-open. A second hook (kb-code-annotations) surfaces the operator's `flag-for-agent` review annotations into the same sessions — a per-file PreToolUse nudge with the exact reply/resolve commands, plus a one-line SessionStart summary of any repo with open flags — turning reader-side review comments into a live human-to-agent dialogue. Ships `kb-code hook install|uninstall|status` to print the settings.json wiring for both hooks and check daemon reachability (never edits settings.json silently), and `kb-code bench-search` — a retrieval-quality/latency instrument over the unified Search-Everywhere box, with a starter query set for the kb repo itself.

by kb

Triage SPA comments (kb-comments/1) from Claude Code. /kb-comments runs the list → edit-source → reply/resolve round-trip for one artifact; /kb-comments-watch is the live SSE loop that surfaces new reviewer comments as they land. The `kb` CLI is the only safe path — the daemon owns the ETag-protected review files. Packaging only; no daemon code.

by kb

Claude Code memory via the kb daemon: automatic recall each turn (UserPromptSubmit), a re-injected memory protocol at session start (SessionStart), verbatim transcript capture when the agent finishes (Stop), and a Stop-hook nudge that suggests /kb-distill when a session committed code but kept no memory — LLM-free, every hook just runs a kb verb. Ships /kb-distill (the episodic→semantic distiller: 1–3 quality-gated, deduped, provenance-stamped memories per session, --dry-run supported), /kb-weekly (the GET /api/sessions/ledger's narrative half — a project's week of sessions/commits/decisions turned into a short review note; facts from the daemon, prose from the agent), /kb-verify (the memory↔code drift sweep: code-citing memories resolved fresh through kb-code's doclens, dated [kb-drift] comments filed on dead citations — one open comment per (memory, path) — plus a born-stale check and kb memory flag escalation when the fact itself is wrong; --dry-run supported), /kb-setup: a guided first-run bootstrap (install check, daemon check, writes a kb.toml with memory + sessions corpora, starts + verifies the daemon), /kb-slate-tidy (kb-slate/1 housekeeping: reads a project's slate board plus history, drops noise the running session or a presumed-ended session authored, merges near-duplicate founds with --supersedes, closes asks whose asker ended — never touches a live other session's now/warn/take/hand), and /kb-slate-distill (turns a closed or rotated slate — current AND archived generations — into 1-5 durable memories, a narrative note, and a dated plan-file Status line).

by kb

The memory 'dream': /kb-reflect consolidates past Claude Code sessions into durable curated facts. Reads session digests, dedups against existing memories (ADD/SUPERSEDE/MERGE/NOOP), and — after you approve a dry-run — writes graded, provenance-linked facts via `kb remember`. The distiller is the live agent (same model, same tokens); the daemon stays model-free. Append-only, human-gated, recall untouched (#10).

by kb

Author kb-indexed HTML research artifacts and keep them true in one flow. Ships the kb-artifact authoring skill (the kb authoring contract), the kb-audit truth-maintenance skill (reconcile a corpus of research docs' status claims against repo truth, --fix flips stale markers in the source), and a /kb-tools command that prints the live `kb` CLI manifest. Packaging only — no daemon code; everything runs the existing `kb` CLI plus `git`/`rg`.