session: Merkle-rooted multi-turn Q&A REPL with Bates ledger

`arborist session` is an interactive multi-turn Q&A REPL where every
turn (or fork) mints one node in a per-session SQLite-backed tree.
Each node carries a stable Bates id (`<sid>-<6-digit>`) and folds into
a Merkle subtree-hash chain; the root node's subtree_hash is the
session_root.

Tree shape lets:

- **Forks** happen implicitly: `/cd <bates>` to a prior node, ask
  again → sibling under that parent. Branch points (≥2 children)
  surfaced by `/branches`.
- **Page-refresh caching** stay cheap: a client tracking
  (bates → subtree_hash, body) only refetches subtrees whose hash
  changed. Sibling subtrees that didn't change are byte-identical
  → cache-equivalent. Same property git pack-protocol and IPFS MFS
  use.
- **Audit-chain verification** be per-session and independent: each
  session db has its own session_audit_events with event_hash =
  sha256(prev_hash || canonical_body). `make session-chain-check`
  walks all sessions; 0 breaks each = intact.

Wire:

- arborist/qa/session.py — Session class, Bates minting, Merkle
  recompute on O(depth) insert, audit chain, helpers (list, render,
  resolve <bates|seq|label>).
- arborist/cli.py — `session` subcommand: REPL + --list / --tree
  / --chain-check / --gc / --json flags. Ancestor-titles → retrieval
  keywords (parsed from cited-pointer lines in answer_text) flow down
  the branch via policy["retrieval_keywords"].
- arborist/qa/providence_query.py — honor policy["retrieval_keywords"]:
  augment FTS5 retrieval query without touching cache_key (mirrors
  legacy --retrieval-keywords discipline, #000001).
- Makefile — `make session [SID=...]`, `make session-list`,
  `make session-tree SID=...`, `make session-chain-check`,
  `make session-gc SESSION_KEEP=N`.
- docs/sessions.md — schema, Merkle conventions (portability for
  non-Python consumers), REPL command reference.
- tests/test_session.py — 15 tests: create, resume, add_node, fork
  via cd, branches, root determinism, audit chain (intact + tampered),
  resolve, list, render, sibling-invariance of subtree_hash.

Storage: ~/.arborist/sessions/<sid>.db (self-contained — no FK into
main store). Answers live in providence_cache keyed by cache_key;
session only carries conversation shape. Cache hits stay live across
sessions. Bounded growth via --gc.

Phase 1 scope: tree + Merkle + Bates + retrieval-keyword flow.
NOT in Phase 1: LLM-side conversation_history (threading prior Q&A
into the LLM prompt + conversation_hash). A bare-pronoun follow-up
("who created him?") gets the right retrieval today but the LLM may
still UNGROUNDED because it sees only the new question as user
message. Folding conversation_history into the prompt + cache_key's
conversation_hash dimension is the natural Phase 2.

197 tests pass.
This commit is contained in:
russell@unturf.com 2026-06-01 16:58:08 -04:00
parent 21d2d2774a
commit 508fec6975
No known key found for this signature in database
6 changed files with 1541 additions and 9 deletions

205
docs/sessions.md Normal file
View file

@ -0,0 +1,205 @@
# Sessions — Merkle-rooted conversation trees
Multi-turn Q&A with branching, where the tree itself is Merkle-rooted.
Each turn (or fork) mints a node with a stable Bates id and folds into
a per-session subtree-hash chain; the root node's `subtree_hash` is the
`session_root`, a single hash that pins the entire conversation state.
## On-disk shape
One SQLite file per session at `~/.arborist/sessions/<sid>.db`:
```sql
CREATE TABLE session_meta (
session_id TEXT PRIMARY KEY,
root_bates TEXT NOT NULL, -- always seq=0
session_root TEXT NOT NULL, -- subtree_hash of root node
current_bates TEXT NOT NULL, -- REPL "you are here" pointer
next_seq INTEGER NOT NULL,
created_at TEXT NOT NULL,
updated_at TEXT NOT NULL
);
CREATE TABLE nodes (
bates TEXT PRIMARY KEY, -- "<sid>-<6-digit>"
seq INTEGER NOT NULL, -- monotonic per session
parent_bates TEXT REFERENCES nodes(bates),
question TEXT NOT NULL,
cache_key TEXT NOT NULL, -- → providence_cache
audit_mode TEXT NOT NULL,
created_at TEXT NOT NULL,
label TEXT, -- optional human name
node_hash TEXT NOT NULL, -- sha256 of canonical body
subtree_hash TEXT NOT NULL -- Merkle hash over self + children
);
CREATE TABLE session_audit_events (
seq INTEGER PRIMARY KEY AUTOINCREMENT,
event_type TEXT NOT NULL, -- session_init | node_added
bates TEXT,
prev_root TEXT NOT NULL,
session_root TEXT NOT NULL,
body TEXT NOT NULL, -- canonical JSON
event_hash TEXT NOT NULL, -- sha256(prev_hash || body)
created_at TEXT NOT NULL
);
```
The session db is **self-contained**: no foreign keys into the main
store. Answers live in `~/.arborist/qa.db:providence_cache` keyed by
`cache_key`; the session only carries the **shape** of the conversation
(parent links, Bates, hashes). Cache hits across sessions stay live.
## Bates discipline
`<sid>-<6-digit-seq>`. `sid` is `YYYYMMDD-HHMMSS-<4-hex>`; `seq` is
monotonic per session, never reassigned, never re-used. `seq=0` is a
synthetic root with empty question — collapsed in tree rendering, kept
so all real questions descend from a single node.
## Merkle hashes
Two derived hashes per node:
```
node_hash = sha256(canonical_json({
parent_node_hash, bates, question, cache_key,
audit_mode, created_at, label,
}))
subtree_hash = HashCombine(
leaf_hash(node_hash),
children_root,
)
```
Where `children_root` is built bottom-up by `HashCombine` over the
children's `subtree_hash` values, sorted by hash (deterministic),
with the standard arborist odd-element rule (self-duplicate).
Sorting children by `subtree_hash` (content-addressed) keeps the
combine order independent of insertion timing — two sessions that
land the same nodes in different order produce the same root.
## Insert path is O(depth)
When a new node `N` is inserted under parent `P`:
1. Compute `N.node_hash`.
2. `N.subtree_hash = leaf_hash(N.node_hash)` (new node is a leaf).
3. Walk `N → P → P.parent → … → root` recomputing each ancestor's
`subtree_hash`.
4. Update `session_meta.session_root` to root's new `subtree_hash`.
5. Append one `session_audit_events` row whose `event_hash` chains
off the previous event.
A sibling subtree that didn't change keeps its `subtree_hash`. This
is the property that makes page-refresh caching cheap.
## Caching for page refresh
A UI client tracks `(bates → subtree_hash, body)` locally. On refresh:
1. Client sends current `session_root`.
2. Server returns either "still current" or the new root + a list of
`subtree_hash` deltas along the changed paths.
3. Client walks top-down, fetching only nodes whose `subtree_hash`
differs from cache.
Same shape as git's pack-protocol or IPFS MFS: content-addressed
sibling subtrees are byte-identical → cache-equivalent.
## Audit chain
`session_audit_events` is a hash-chain mirror of `arborist.audit_events`,
scoped per session. Every state change emits one event whose `event_hash`
is `sha256(prev_event_hash || canonical_body)`. Verify with
`arborist session --chain-check [--sid SID]` (or
`make session-chain-check`).
`event_hash` and `session_root` are independent integrity signals:
`session_root` answers "what does the tree look like now?";
`event_hash` answers "is every state change in order intact?". Both
moves on every insert, but for different reasons.
## Forking
There is no `/fork` verb. Forking is implicit: `/cd <bates>` moves the
"current" pointer to any node in the tree; the next question becomes a
child of that node. Branch points are nodes with ≥2 children — listed
by `/branches` or in the tree header.
## Retrieval-side context flow
When a turn is asked under parent P, the REPL gathers the source titles
the LLM **cited** in P and P's ancestors (parsed from each ancestor's
stored `answer_text` for `[E# | <title> | …]` pointers) and passes them
into `providence_query` via `policy["retrieval_keywords"]`. The keyword
string augments FTS5 retrieval ONLY — it never enters `cache_key`,
`question_hash`, or `governance_policy_hash` (mirrors the discipline of
the legacy `--retrieval-keywords` flag, see ticket #000001).
This makes pronoun follow-ups ("who created him?") still surface the
prior turn's primary source. **It does not solve the LLM-side antecedent
problem** — the model receives only the new question as the user
message, so a bare pronoun question may still fail to ground. Threading
prior Q&A pairs as `conversation_history` into the LLM prompt (+ folding
that history into `conversation_hash`) is a future enhancement.
## CLI
```
arborist session # new session, REPL
arborist session --sid <id> # resume
arborist session --list [--json] # list sessions
arborist session --tree --sid <id> # print tree
arborist session --chain-check # verify audit chain (all)
arborist session --gc N # keep N most recent
```
REPL commands:
| command | effect |
|---|---|
| `/tree` | print the tree |
| `/branches` | list branch points (Bates with ≥2 children) |
| `/cd <bates\|seq\|label>` | move current pointer |
| `/back` | cd to parent of current |
| `/label <name>` | name current node |
| `/show <ref>` | dump stored info for a node |
| `/sid` | print session id |
| `/root` | print session_root |
| `/help` | help |
| `/quit` | exit (or Ctrl-D) |
| anything else | new question, child of current |
Make targets:
```
make session [SID=...] # interactive REPL
make session-list # list all
make session-tree SID=... # print one tree
make session-chain-check # verify audit chain
make session-gc # keep SESSION_KEEP=100 (default) most recent
```
## Bounded growth
`make session-gc` (or `arborist session --gc N`) keeps the N most-recent
sessions, deletes the rest. Each session is a single SQLite file →
trivially deletable, shareable, archivable.
## Portability
The on-disk schema has no FK dependency on the main arborist store.
A non-Python consumer that wants to read a session db needs only:
1. SQLite3 with FTS5 (not required — sessions don't use FTS).
2. The canonical-JSON convention: sorted keys, no whitespace, UTF-8.
3. Three hash primitives:
- `sha256(bytes) → hex`
- `leaf_hash(bytes) = sha256(0x00 || bytes)`
- `hash_combine(L, R) = sha256(0x03 || L || R)`
Those are the same conventions as the proxy.unturf.com Go merkle and
the rest of arborist (see `arborist/merkle.py`).