Fork drift report: what I added before merging your latest in (ai-visualizer, backtalk, fullstack-agent)

[aram-cloudstak]

What I've added on my forks (before merging your latest in)

I've been running forks of ai-visualizer, backtalk, and fullstack-agent for a
few days and built a few things on top before pulling your latest changes in
today. Figured I'd write these up in case any are useful upstream, and so
they don't get lost/re-invented. I've already merged your latest into my
forks, so this is just "here's what I had that you didn't" as of that merge.

ai-visualizer

Cost and transcription overlays (core.js, server.py, ai-visualizer.json)
A live overlay on the visualizer face showing session cost and running
transcription text, backed by a new server.py endpoint.

Inbox-queue widget + real billing data (follow-up commit)
Extended the cost overlay to pull actual billing numbers instead of an
estimate, plus a widget showing what's sitting in the inbox queue.

Both are additive UI/overlay work — didn't touch your existing face
rendering paths, so should apply cleanly as a feature branch if you want to
look at it.

backtalk

Singleton lock (backtalk/main.py)
Refuses to start a second backtalk process while one is already running,
so you don't end up with two processes fighting over the same hotkey/mic. A
lock file at .backtalk.lock holds the owning PID; a stale lock left by a
crashed run (PID no longer alive) is detected and reclaimed automatically.
Windows liveness check goes through OpenProcess with
PROCESS_QUERY_LIMITED_INFORMATION rather than os.kill(pid, 0) — on
Windows that call actually maps to TerminateProcess under the hood, so it
isn't safe to use as a read-only liveness probe.

Session resume + playback speed config — built this independently
around the same time you did (my commit 2026-08-18, yours 2026-08-19). Mine
added resume_last_session (opt-in, off by default) and a speed config
key for Kokoro. Yours ended up more complete — it wires the resume id
through the permission-gate/can_use_tool path at connect and falls back
cleanly to a fresh session if a saved resume id is stale, mine didn't do
either. I took your version during the merge; flagging only so you know two
people hit the same gap at once, not because I think there's anything left
to reconcile.

fullstack-agent

meeting-transcriber wired into start.bat as "agent ears"
I run a fourth local tool (meeting-transcriber — dual-track recording +
faster-whisper transcription + summary, not one of your repos) and added it
to the launcher alongside face/hands/voice, gated on a Python 3.13 install
check with a clear skip message if it's not found. Not something you'd want
verbatim since it depends on a tool you don't ship, but flagging the pattern
(a fourth optional leg in start.bat) in case it's useful shape for your own
"one more optional piece" additions.

Not worth a mention

ai-memory-vault and ai-marketing-skills forks only had housekeeping on my
side (CHANGELOG backfill, merge commits) — no feature drift there.

[aram-cloudstak]

One more thing: patterns built on top of ai-memory-vault itself

Separate from the code stuff above — I've also built a fair bit on top of the vault system itself (the memory/notes side, not the tool repos). Nothing here touched your ai-memory-vault repo's actual template files, it's all stuff layered on top in my own vault content and a few standalone tools, but some of it might be useful shape for the template if any of it's relevant to where you're taking that project.

A richer note-type system. Replaced the single project slug field with a proper type taxonomy — person, client/account, project, decision, architecture fact, recurring issue, business opportunity, meeting source, technical standard, lesson-learned, runbook — plus a field for "who this note is scoped to" separate from "who actually employs this person." Useful once you're tracking more than one counterparty (clients, vendors, subcontractors) instead of just your own projects.

A knowledge-graph discipline, not just a notes folder. Meeting notes/emails/transcripts are treated as raw sources, not the canonical record — durable facts get mined out into linked canonical pages (one page per person/client/decision/etc.), with an explicit source-of-truth precedence order when sources disagree, a citation requirement for any real recommendation ("based on [[X]] and [[Y]]"), and a monthly pass flagging stale/undercited notes.

Alias-based entity resolution + a hard rule about uncertain matches. People/client notes carry alias lists so an automated matcher can resolve nicknames back to the canonical page. Match confidence is graded by signal type (an explicit name mention beats a participant-list hit beats a keyword hit), and — this is the part worth stealing — a match driven by keyword overlap alone, with zero name or participant corroboration, is never auto-filed. It either needs a stronger signal or gets surfaced for a human to confirm. Learned this one the hard way after a few false-positive misfiles.

A less lossy archive process. Archiving timestamps the note into a dated subfolder and requires every canonical page that cited it to get a real [[wikilink]] back to it in the same pass — so an archived source never goes dark in the graph, it just moves.

A couple of standalone tools that feed the vault, not part of your stack:

• A local meeting recorder + transcriber + summarizer that files straight into the graph as a dated source.
• A cost dashboard that reconciles an estimated cost (parsed from local session token counts) against actual subscription billing — worth building since subscription billing exposes no real per-dollar API, and the two numbers turned out to diverge by a lot more than I expected.
• A throttled job queue for any automation that spends LLM budget, rate-limited against a rolling spend ceiling instead of a hard cap.
• A tool that mines an IDE's local chat history (Cursor, in my case) into the same per-account log format the meeting recorder uses — reuses the same entity-matching code rather than reimplementing it.

Auto-backup on session end — the vault and all the tool repos get committed (AI-written commit message per repo) and pushed automatically when a session ends, so none of this depends on remembering to commit.

Happy to go into more detail on any of these if useful — didn't want to dump a wall of text if it's off-topic for what you're building toward.