vivac
A tree where every node knows which node it was born from. It exists to answer "why are we here?" months later, when nobody remembers any more.
$ vivac why 11
Why we are here -> t11
------------------------------------------------------------------
g1 vivac 0.1 publishable
A provenance system for work that can answer "why are we
here" months later.
(7 open / 4 closed below)
|
v
t8 Port to Rust in the public repo
When the format stops moving, not before.
(3 open below)
|
v
t11 Redaction guard on write
Security pillar. Goes BEFORE any cloud mode.
^^^ you are here
In parallel, still open (2):
t9 TUI for the maintainer
t10 Migrate from JSON to SQLite
t8 does not close until these close (1):
t11 Redaction guard on write
The problem
When you develop with an agentic AI, work spawns more work. Three hops in, you have lost the thread of what you originally set out to do.
It is not a memory problem: usually everything is written down. It is a provenance problem. What is written does not say what it was born from, and without that edge there is no way to reconstruct why you are where you are.
Measured on a real compiler: the path between the goal and the day's work was six levels deep, spread across a chronologically ordered 8,853-line tracker, 52 planning documents and 21 issues. The structure was temporal, which is exactly the opposite of provenance.
Logbooks, ADRs and issues all store the node. None of them stores the edge. That is how you can have everything written down and still not be able to say where something came from.
How it is used
There are two audiences, and the tool splits in two because of them.
The agent writes. Capture hangs off the seams of the work: you open a node when you start, you close it when you finish. The provenance edge is created on its own, with nobody having to remember to declare it.
The maintainer reads.
And there are safe stops. A vivac is the bivouac partway up a climb: a
coherent state, with the stack frozen and the identity of the code at that
moment. push, pop and park leave one without anybody asking.
restore never touches the working tree. Mixing context navigation with
tree manipulation gives you a branch manager worse than git.
Everything the agent needs to do can be done with no interactive interface,
and every read command accepts --json.
The two edges
It is the distinction that holds the model up, and it came out of seeding two real trees and putting them side by side:
| Question it answers | When it is created | |
|---|---|---|
| born from | where did this come from? | on its own, at every push |
--blocks |
does this stop its parent from closing? | explicitly |
A closed batch of issues with an open finding underneath is correct: the
batch finished and the finding is another thing. An audit marked DONE with
its findings open is a false marker — one of those took 26 days to be
spotted. Same shape, opposite verdict.
That is why vivac done refuses to close with open conditions and lists
what is missing. It is the only rule in the model that rejects an operation,
and it earns that privilege because the case it prevents is measured.
$ vivac done 8
t8 CANNOT close: 1 open closure condition(s)
t11 Redaction guard on write
A run closes with its findings, not with its report.
Closing it anyway leaves a trace: vivac done 8 --force
What it never stores
A provenance tree is a map of where a system is weak and not yet fixed. That forces a few things, and they are not negotiable:
- No keys and no secrets. There is a redaction guard at write time. In doubt it refuses and says why; it never stores in silence.
- No personal data. No email, no name, no home path. The
actoron every event is an opaque identifier. - No file contents. Only paths, references and prose about what was decided. It bounds the blast radius of a leak to what was being worked on, never to what the code is.
- No telemetry. The binary does not phone home.
These rules come from the pillars, which govern by definition: security vetoes, performance budgets, DX judges.
Status
Tier 0 complete. The tree, the two edges, the closure rule, the redaction
guard, the brief with its token budget, the session hooks, the vivacs and the
Anchor with its Git and Null implementations. 73 tests, of which 11 are
the brief specification's contract executed against the real binary.
reconcile is the first of Tier 1. It answers the one question that keeps the
tree honest -- what changed since the tree last looked, and which of it does
no node claim? -- by diffing the anchor's history against the governs globs
the nodes declare. It reports and never writes: it can say nobody claims a
file, and it cannot say which thread that file belongs to.
The brief is deterministic by contract: same log, same --now, same bytes.
The spine — the path from the root to the focus — is never truncated: if it
does not fit the budget it comes out anyway, and the warning says that what is
left over is tree, not render.
Measured on this machine, excluding process startup:
| nodes | push |
brief |
tree |
|---|---|---|---|
| 100 | ~5 ms | ~5 ms | ~5 ms |
| 1,000 | ~11 ms | ~10 ms | ~13 ms |
| 10,000 | ~54 ms | ~63 ms | ~95 ms |
The write budget is p99 < 5 ms and the read budget < 50 ms over 10,000 nodes. In the low hundreds of nodes it holds, and from there up it degrades linearly with the size of the log, which is read whole on every call. That is where SQLite comes in, and now it has a number instead of a hunch.
Not there yet: TUI, search, cascading invalidation, team mode.
0.3.0 does not read a log written by 0.1.x or 0.2.x. The tool was written in Spanish and those releases stored the event fields under Spanish names, which 0.2.x read through aliases. 0.3.0 speaks one language, so it reports those lines as unreadable rather than guessing. If you have such a log, 0.2.1 still reads it.
Hooks
SessionStart injects the brief into the agent's context; Stop leaves an
automatic stop. Stop runs on every turn, not at session close — there is
no end-of-session event — so the stop is only saved if the tree changed since
the previous one: a stop that repeats identically is not a stop, it is a log.
Both stay quiet and exit 0 where there is no .vivac/, so they can be left in
the global configuration without getting in the way of other projects.
Install
From source, cargo install --path . inside the repo.
No daemon, no server and no network. The store is .vivac/, two files.
Licence
MIT OR Apache-2.0, at the option of whoever uses it. The text of each is in
LICENSE-MIT and LICENSE-APACHE.