dsh-context-graph 0.6.2 → 0.11.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +563 -0
- package/README.md +252 -97
- package/README.zh.md +82 -32
- package/cordis.patch.yml +10 -6
- package/index.js +7 -432
- package/package.json +11 -12
- package/src/client/client.js +880 -0
- package/src/core/builder.js +210 -0
- package/src/core/chat.js +34 -0
- package/src/core/graph.js +134 -0
- package/src/core/index.js +14 -0
- package/src/core/ledger.js +108 -0
- package/src/core/match.js +377 -0
- package/src/core/render.js +195 -0
- package/src/core/session.js +446 -0
- package/src/core/text.js +220 -0
- package/src/core/time.js +147 -0
- package/src/core/tree.js +39 -0
- package/src/host/builder.js +416 -0
- package/src/host/command.js +155 -0
- package/src/host/config.js +50 -0
- package/src/host/memory.js +175 -0
- package/src/host/meter.js +75 -0
- package/src/host/page.js +186 -0
- package/src/host/plugin.js +46 -0
- package/src/host/recall.js +227 -0
- package/src/host/runtime.js +599 -0
- package/src/store/chats.js +236 -0
- package/src/store/files.js +181 -0
- package/src/store/ledger.js +65 -0
- package/src/store/messages.js +88 -0
- package/src/store/migrate.js +143 -0
- package/graph.js +0 -496
- package/lib/client.js +0 -126
- package/store.js +0 -81
- /package/{update.js → src/host/update.js} +0 -0
package/CHANGELOG.md
ADDED
|
@@ -0,0 +1,563 @@
|
|
|
1
|
+
# Changelog
|
|
2
|
+
|
|
3
|
+
All notable changes to dsh-context-graph. Versions follow semver; each minor version from 0.7.0 on
|
|
4
|
+
is one development cycle (audit, issues, fixes, independent review), documented in
|
|
5
|
+
[docs/cycles](docs/cycles).
|
|
6
|
+
|
|
7
|
+
## 0.11.0 — 2026-11-22 — the memory tree
|
|
8
|
+
|
|
9
|
+
The memory has always been a tree (workspace, chats, topics, the messages they came from); now it
|
|
10
|
+
is drawn as one, on a page of its own. The picture moves from the plugin's page in Plugins to a
|
|
11
|
+
Memory tree entry in the sidebar, next to Automation tasks. Its techniques come from projects big
|
|
12
|
+
companies publish under free licences; no code was copied and no dependency added (credits in the
|
|
13
|
+
README).
|
|
14
|
+
|
|
15
|
+
**Its own page** (#85)
|
|
16
|
+
- Memory tree in the sidebar's panel list opens a page in the main column, laid out like dsh's own
|
|
17
|
+
first-level pages. The plugin's page in Plugins keeps the savings and says where the tree is;
|
|
18
|
+
`/context-graph map` points there for the chats it leaves out.
|
|
19
|
+
|
|
20
|
+
**A tree, not circles** (#86)
|
|
21
|
+
- Root on the left, one column per level. The layout keeps D3's tidy-tree rules (a parent centred
|
|
22
|
+
on its children, siblings one row apart, another parent's children half a row further); each
|
|
23
|
+
column is as wide as its widest label, as in AntV G6's compactBox, so labels are no longer cut to
|
|
24
|
+
fit a circle; links are D3's horizontal curves.
|
|
25
|
+
- The newest chats start open while their topics fit 40 rows (the idea of ECharts'
|
|
26
|
+
initialTreeDepth); a closed chat shows how many topics it holds. Topics are grouped by kind, so
|
|
27
|
+
each colour forms one band; chats and the workspace are rings.
|
|
28
|
+
- Search opens the chats it matches and fades the other topics; the kind buttons fade the other
|
|
29
|
+
kinds. The tree scrolls natively; zoom scales the drawing.
|
|
30
|
+
- Host: the `map` route becomes `tree` (chats with their topics and the links among them), and
|
|
31
|
+
`src/core/layout.js` gives way to `src/core/tree.js`.
|
|
32
|
+
|
|
33
|
+
**The last level** (#87)
|
|
34
|
+
- Opening a topic reads it once and adds its newest stored messages as its children (`user t7 ·
|
|
35
|
+
...`); a message leads back to its topic. A topic on its way shows `…`, one that failed `!`.
|
|
36
|
+
- A chat or topic you choose shows its details beside the tree: a topic's gist, links, sources and
|
|
37
|
+
id; a chat's marks, dates and savings.
|
|
38
|
+
|
|
39
|
+
**Focus, motion, keyboard** (#88)
|
|
40
|
+
- Pointing at a node, or focusing it by keyboard, fades everything but its family: ancestors,
|
|
41
|
+
descendants and the topics it links to (ECharts' emphasis focus).
|
|
42
|
+
- What opens grows out of its parent and what moves eases to its place (D3's collapsible tree;
|
|
43
|
+
Material 3 timing: 300 ms, emphasized easing). Nothing moves with reduced motion on.
|
|
44
|
+
- The tree is a WAI-ARIA tree (the W3C pattern Microsoft Fluent UI and GitHub Primer follow): one
|
|
45
|
+
tab stop, Up/Down, Right/Left to open, close or step in and out, Home/End, Enter/Space.
|
|
46
|
+
|
|
47
|
+
**Review** (#89, #90)
|
|
48
|
+
- A linked topic opened from the details panel opens its chat once, so the selection is drawn; the
|
|
49
|
+
chat can still be closed. (#89)
|
|
50
|
+
- A tree showing more than 400 rows (a broad search in a big workspace) draws what enters at once
|
|
51
|
+
instead of starting an animation per node. (#90)
|
|
52
|
+
|
|
53
|
+
Checks: 320 tests in three time zones; every rule of the tree was broken on purpose and a test
|
|
54
|
+
failed each time (49 mutants; one more showed a check that could never fail, which was removed). The motion was
|
|
55
|
+
checked in Chromium: transitions on transform and on a link's path, and the grow-in keyframes,
|
|
56
|
+
interpolate as intended.
|
|
57
|
+
|
|
58
|
+
## 0.10.0 — 2026-11-17 — memory for the way dsh is used
|
|
59
|
+
|
|
60
|
+
Cycle 4. Memory now works for every way dsh drives a chat, not only typed prompts: `/goal` rounds,
|
|
61
|
+
scheduled reminders, plan mode and the model's questions. A new chat pays nothing for memory it
|
|
62
|
+
does not need, a question finds its topic however it is worded, and less is stored and rewritten.
|
|
63
|
+
The audit behind it loaded the plugin into dsh's own runtime; that harness is now part of the suite.
|
|
64
|
+
|
|
65
|
+
**Input dsh drives** (#69, #70, #71)
|
|
66
|
+
- `/goal` rounds, scheduled reminders and late answers to the model's questions count as prompts
|
|
67
|
+
and as turns: a round is matched by its objective (once per goal revision), a reminder by its own
|
|
68
|
+
prompt, and both count toward `buildEveryTurns`. A goal changed in an off stretch, a reminder
|
|
69
|
+
scheduled there and an answer to a question asked there stay private, rounds included. (#69)
|
|
70
|
+
- The plan you approve in plan mode is indexed as the model's text, and your answers to its
|
|
71
|
+
questions (in time or late) as yours, with their question; both are part of a recalled turn. (#70)
|
|
72
|
+
- Turn ids (`t7`, `c3t7`) are dsh's own turn numbers: a steer stays in its turn, a goal round has
|
|
73
|
+
its own. (#71)
|
|
74
|
+
|
|
75
|
+
**Indexing** (#42)
|
|
76
|
+
- A chat that stops short of the next 4-turn mark is indexed once it goes quiet, or at once when a
|
|
77
|
+
new chat of its workspace starts; a turn starting in it cancels that.
|
|
78
|
+
|
|
79
|
+
**Tokens** (#76, #77, #80, #81, #84)
|
|
80
|
+
- The chat list (about 200 tokens, then re-sent as cached input with every later request) rides
|
|
81
|
+
only with a chat's first topic lines or a question about earlier work; a first message about
|
|
82
|
+
something else gets nothing. (#76)
|
|
83
|
+
- `recall("2026-11-01")` and the automatic day digest share a 1,000-token budget among the day's
|
|
84
|
+
chats (up to about 3,600 before); what does not fit is named. (#77)
|
|
85
|
+
- Everyday messages with a time word and "we <verb>" ("what should we do today", 我们明天要讨论的)
|
|
86
|
+
no longer get a day digest of about 360-470 tokens. (#80)
|
|
87
|
+
- A day's digest lists the chats worked on that day first, not the ones that merely span it. (#81)
|
|
88
|
+
- More ways to name a past day: two days ago, a week or 2 weeks ago (that calendar week), last
|
|
89
|
+
night, last Monday, nov 3, November 3rd, 三天前, 一周前, 两个礼拜前; 大前天 is three days back,
|
|
90
|
+
上上周 the week before last, 上上个月 the month before last. (#84)
|
|
91
|
+
|
|
92
|
+
**Quality** (#73, #74, #75)
|
|
93
|
+
- A file or command output the model read that lists topic ids no longer hides those topics, and
|
|
94
|
+
it is no longer scanned (the slowest per-message step in long chats). (#73)
|
|
95
|
+
- An exact lowerCamel identifier (`loadViews`) finds its topic the way snake_case does. (#74)
|
|
96
|
+
- Plurals fold onto their singulars on both sides (tests, queries, branches, caches); compounds
|
|
97
|
+
stay exact. On the corpus 37 reworded questions now find their topic (0.9.0: 4), with no false
|
|
98
|
+
hit and no block for 19 more everyday messages. (#75)
|
|
99
|
+
|
|
100
|
+
**Speed and storage** (#72, #78, #82, #83, #58)
|
|
101
|
+
- The first recall of an ordinary chat reads no dsh log. (#72)
|
|
102
|
+
- Ledger entries keep only counters that are not zero and lists that are not empty; meta.json no
|
|
103
|
+
longer stores the workspace path. (#78)
|
|
104
|
+
- `/context-graph map` lists the newest 40 chats (about 40 KB at most in the session log; 1 MB at
|
|
105
|
+
1,000 chats before) and points to the plugin's page for the rest. (#82)
|
|
106
|
+
- The ledger is one small file per chat, so a turn end writes only its own chat's file: 1.7 ms
|
|
107
|
+
instead of 10.2 ms at 1,500 chats. 0.9.0's ledger.json is only read. (#83)
|
|
108
|
+
- Forks indexed by 0.4-0.6 are found with one dsh session listing per run; one whose fork point is
|
|
109
|
+
not known yet stays out of matching while it is learned in the background. (#58)
|
|
110
|
+
|
|
111
|
+
**Data safety** (#79)
|
|
112
|
+
- A damaged file that cannot be moved aside (a sync tool or a scan holds it) is no longer reported
|
|
113
|
+
as set aside and read as missing: a memory-off chat is not indexed because of it, and the file is
|
|
114
|
+
never overwritten with no copy kept.
|
|
115
|
+
|
|
116
|
+
**Tests**
|
|
117
|
+
- test/e2e/cycle4-real-dsh.test.js boots dsh 0.2.1's own runtime (cordis, sessions, agent loop,
|
|
118
|
+
session query, commands, `/goal`) with a scripted provider and the plugin: a chat indexed at
|
|
119
|
+
dsh's turn end with one call at the lowest effort, a second chat that gets the topic and recalls
|
|
120
|
+
it and its turn, off then on through dsh's command service, and a `/goal` chat. Skipped where dsh
|
|
121
|
+
is not installed.
|
|
122
|
+
|
|
123
|
+
**Found by the independent reviews of this change, and fixed** (first review: 3 angles, 15 findings, 12 confirmed by reproduction and the other 3 fixed as well, plus 2 the fixers left open, closed after the merge; second review: 3 angles, 14 findings, 13 confirmed by an independent reproduction, 1 of them high; third check, of those fixes: 3 angles, 12 findings, each reproduced twice, 3 of them medium. Every fix has a test that fails without it)
|
|
124
|
+
|
|
125
|
+
Privacy
|
|
126
|
+
- A fork that resumes a `/goal` set while an ancestor's memory was off keeps it private: a fork
|
|
127
|
+
takes its ancestors' off stretches up to the fork point as its own first gaps, and a fork made
|
|
128
|
+
before 0.10.0 takes them before its next index. (K0)
|
|
129
|
+
- Under dsh's `ptc` preset, tool calls made from `run_code` count like direct ones: a reminder
|
|
130
|
+
scheduled while memory was off stays private, and plans and answers are read. (K1)
|
|
131
|
+
- Only a plan you approved is indexed; a plan sent back to planning is not, and the words you sent
|
|
132
|
+
it back with are read as your answer. (R1)
|
|
133
|
+
- A chat driven only by `/goal` gets its fingerprint from the goal's objective, so a later app run
|
|
134
|
+
does not list every dsh session on its first message; a listing that failed is not retried on
|
|
135
|
+
every message. (K10)
|
|
136
|
+
|
|
137
|
+
Turn ids
|
|
138
|
+
- Turn ids stored up to 0.9.0 (which counted typed prompts) are never shown beside dsh's numbering:
|
|
139
|
+
a chat's store is converted from its log at its next build or topic recall and marked
|
|
140
|
+
(msgs.json `v: 2`); until then its sources name no turn, on the plugin's page either. (K2, K9)
|
|
141
|
+
|
|
142
|
+
Tokens and indexing
|
|
143
|
+
- The chat list rides with a chat's first block on whichever turn it comes, so a chat that opens
|
|
144
|
+
with a greeting still gets it with its first topic lines or question about earlier work. (K3)
|
|
145
|
+
- A chat with pauses is indexed once quiet at most once between two regular indexes, after 30
|
|
146
|
+
minutes instead of 10: 5 indexing calls instead of 12 on a 12-turn chat paused after every turn
|
|
147
|
+
(0.9.0: 3). (R0)
|
|
148
|
+
- A msgs.json that cannot be read (a damaged file another program holds) reads as no stored
|
|
149
|
+
source: the block still goes out and a recall says no source is stored, instead of failing with
|
|
150
|
+
a raw error that names local paths. (K11)
|
|
151
|
+
|
|
152
|
+
Matching and time
|
|
153
|
+
- Stories that name a day ("the restaurant we tried last night was amazing, find me a similar
|
|
154
|
+
one") get no day digest, and "what did I do wrong" is no memory question. (K4)
|
|
155
|
+
- 会话, 会议, 开会 and 想法 no longer read as plans, so a question about another 会话 gets its
|
|
156
|
+
digest again. (K5)
|
|
157
|
+
- Windows, Teams, Sheets, Pages and Actions are never folded onto window, team, sheet, page and
|
|
158
|
+
action. (K6)
|
|
159
|
+
- A pasted base64 blob or key is one plain word, not a camelCase identifier with a thousand parts:
|
|
160
|
+
3-5 ms instead of 80-117 ms at 10,500 topics, and no false hits. (K7)
|
|
161
|
+
- 两天前端 (the frontend these two days) and 三天前后 are no longer days ago, "1.5 weeks ago" is not 5
|
|
162
|
+
weeks ago, and a month date that is the subject ("the may 1 launch last week") gives way to the
|
|
163
|
+
time phrase. (K8)
|
|
164
|
+
- A lowerCamel brand name (macBook, deepSeek, mongoDB) counts double only where the topic writes
|
|
165
|
+
it as code too. (R2)
|
|
166
|
+
|
|
167
|
+
Second review (of the fixes above and how they meet the rest)
|
|
168
|
+
- Turn recall of another chat judges privacy from that chat's log, as indexing does: in a chat
|
|
169
|
+
indexed by 0.9.0, the rounds of a goal set while memory was off and resumed after `on` are never
|
|
170
|
+
served, also in a fork made then. (high)
|
|
171
|
+
- A fork holds its parent's topics that cite a plan, an answer or a goal round, so they are not
|
|
172
|
+
sent to it again (about 100-160 tokens, then re-sent as cached input with every later request);
|
|
173
|
+
a chat begun with `/goal` is found as an ancestor by its objective, and its fingerprint no
|
|
174
|
+
longer changes when a prompt is typed later.
|
|
175
|
+
- After one quiet index, a new chat of the workspace still indexes a chat's last turns; only a
|
|
176
|
+
quiet index that indexed counts as the one between two regular indexes.
|
|
177
|
+
- "the bug we fixed yesterday is back" and "the port we picked last week, what was it" ask about
|
|
178
|
+
earlier work again; a story needs a word of how it went ("was amazing"), read within its clause,
|
|
179
|
+
so a pasted 100 KB transcript no longer costs about 100 ms.
|
|
180
|
+
- After a failed listing of dsh's sessions, forks indexed by 0.4-0.6 are held back and learned
|
|
181
|
+
from their own logs instead of being offered for the rest of the run.
|
|
182
|
+
- A damaged fork meta keeps its ancestors' off stretches next to the gaps it reads back.
|
|
183
|
+
- What dsh drives from an off stretch is found in the log by every reader instead of being stored
|
|
184
|
+
as gaps, so a daily reminder no longer fills the 50-gap cap and closes shared turns.
|
|
185
|
+
- 3天前端口, 3天前后端 and 2天前后台 are days ago again; "on nov 10" is the day asked about whatever
|
|
186
|
+
else the message names when it is past; identifiers up to 10 humps
|
|
187
|
+
(findAllByCustomerIdAndCreatedAtBetween) are code.
|
|
188
|
+
|
|
189
|
+
Third check (of the second review's fixes)
|
|
190
|
+
- A steer typed into a round of a goal set while memory was off never becomes the chat's title.
|
|
191
|
+
- A damaged meta whose off stretches could not be read back keeps a goal set in them private
|
|
192
|
+
after `on`, in indexing and in turn recall.
|
|
193
|
+
- "what did we decide yesterday about the launch on dec 3?" asks about yesterday: an upcoming date
|
|
194
|
+
no longer wins over the day asked about.
|
|
195
|
+
- A request that points back at earlier work ("the plan we agreed on yesterday, write it up")
|
|
196
|
+
asks about it; "good", "bad" or "perfect" said of your own work is no story.
|
|
197
|
+
- 这两天 or 连续两天 before 前后端 is a duration, not days ago; a mixed-case random id with a digit
|
|
198
|
+
is one plain word.
|
|
199
|
+
- After its quiet index, a chat is indexed 5 minutes after a new chat starts unless you go back to
|
|
200
|
+
it first (a side question costs nothing), and a quiet index that failed counts as the one, so
|
|
201
|
+
a model that keeps failing is not called at every pause.
|
|
202
|
+
- A failed listing of dsh's sessions is tried again after 30 seconds; a fork with a long first
|
|
203
|
+
prompt reads its ancestors' fingerprints once, not once per chat.
|
|
204
|
+
- A message's "pages" or "actions" no longer finds a topic about a page or an action (tried in the
|
|
205
|
+
second review): "github actions" met "Undo action history".
|
|
206
|
+
|
|
207
|
+
## 0.9.0 — 2026-11-17 — see what memory does
|
|
208
|
+
|
|
209
|
+
Cycle 3. You can now see what memory does without reading code: a savings line on the
|
|
210
|
+
`/context-graph` card, and a page in Plugins with each workspace's totals and a map of its memory.
|
|
211
|
+
Nothing here adds tokens to a chat or slows a message.
|
|
212
|
+
|
|
213
|
+
**What memory saved, and what it cost** (#35)
|
|
214
|
+
- Every chat keeps a small ledger: notes and recall results sent, their cached re-sends with every
|
|
215
|
+
later request, the recall tool's share of each request, and the indexing calls as dsh reported
|
|
216
|
+
them, against what used topics stood in for (their newest messages as stored, each once per
|
|
217
|
+
chat). A topic counts as used when the model recalled it or its answer says two of the topic's
|
|
218
|
+
words that the prompt did not bring, once per chat. Cached input is weighed at 0.1 and output at 4, in input-token equivalents; a loss
|
|
219
|
+
is shown as a loss.
|
|
220
|
+
- The card's second line: `Savings (estimate): saved about 8,500 tokens in this chat · saved about
|
|
221
|
+
41,000 tokens in this workspace`; its body shows every term, so the number can be checked.
|
|
222
|
+
- `.context-graph/ledger.json` holds a few counters per chat, written at most once per turn, off
|
|
223
|
+
the critical path; it survives restarts.
|
|
224
|
+
|
|
225
|
+
**The plugin's page** (#36)
|
|
226
|
+
- Plugins → Context Graph: each workspace with memory (from dsh's workspace registry and the
|
|
227
|
+
folders used in this run), its saved / cost / net, and a map: chats as circles, topics as dots
|
|
228
|
+
coloured by kind and sized by links, links as lines; hover for a label, click for the gist, the
|
|
229
|
+
chat, the links, the first lines of the sources and the id to paste; search and kind filters;
|
|
230
|
+
drag and zoom. Light and dark follow dsh's own theme; English and Chinese.
|
|
231
|
+
- Data comes from three read-only routes registered through dsh's connection registry, which only
|
|
232
|
+
answers the app's owner; a workspace is named by a key, never a path, so nothing outside the
|
|
233
|
+
known memory folders can be read. The map is laid out by the host (deterministic), so the
|
|
234
|
+
browser only draws; it sends labels only, gists and sources load when a topic is opened.
|
|
235
|
+
|
|
236
|
+
**Docs** (#37)
|
|
237
|
+
- README: the page, how the estimate works and where it errs, the ledger on disk; the cost table
|
|
238
|
+
points at the live numbers, and test/core/docs.test.js fails when the code and the README's
|
|
239
|
+
figures disagree.
|
|
240
|
+
|
|
241
|
+
**Found by the cycle-3 audit of 0.8.0, and fixed** (2 lenses, 13 findings, 11 confirmed by reproduction; #59-#68)
|
|
242
|
+
- What you type while memory is off is never indexed or shared, also after `on`: `off` remembers
|
|
243
|
+
where memory went off, and `on` turns that stretch into a gap the builder never reads and turn
|
|
244
|
+
recall never serves. Turns from before `off` that were not indexed yet still are. (#59)
|
|
245
|
+
- A fork of a private or memory-off chat starts private or memory-off, and is titled from its own
|
|
246
|
+
first prompt and its own dsh titles, never the parent's. (#60)
|
|
247
|
+
- A topic id the indexing model invents for a new topic can no longer overwrite another new topic
|
|
248
|
+
of the same answer (and take its source). (#61)
|
|
249
|
+
- One chat whose topics file cannot be read (locked by a sync tool, no permission) is skipped
|
|
250
|
+
instead of taking memory away from the whole workspace. (#62)
|
|
251
|
+
- The read cache holds far more than any workspace needs: at 1,500 chats a message went from about
|
|
252
|
+
1 s to under 150 ms. (#63)
|
|
253
|
+
- The `on` card says whether the chat is shared or stays private. (#64)
|
|
254
|
+
- A turn that was running when `on` ran is indexed when it ends. (#65)
|
|
255
|
+
- Matching uses a word index per chat and counts a topic's evidence only over the words it holds
|
|
256
|
+
(same results): a 16 KB paste at 150 chats takes about 35 ms instead of up to 1.3 s, and an
|
|
257
|
+
ordinary message about 8 ms instead of 15. (#66)
|
|
258
|
+
- A topic recall shares its budget: a short prompt leaves its unused share to the long answer, so
|
|
259
|
+
answers that fit are not cut. (#67)
|
|
260
|
+
- The recall tool definition is about 80 tokens instead of 106, sent with every request. (#68)
|
|
261
|
+
|
|
262
|
+
**Found by the independent reviews of this change, and fixed** (first review: 3 angles, 22 findings, 21 confirmed by reproduction and the 22nd fixed as well, plus 3 found while fixing them and 3 by a review of the page fix; second review, of those fixes: 3 angles, 16 findings, 14 confirmed and the other 2 fixed as well; third review, of the second round's fixes: 3 angles, 10 findings, 9 confirmed and the 10th fixed as well; fourth review, of the third round's fixes: 2 angles, 4 findings, 3 confirmed, all minor, and the 4th fixed as well; a last adversarial check of those fixes: 5 findings, all minor, fixed)
|
|
263
|
+
|
|
264
|
+
Privacy (#59, #60)
|
|
265
|
+
- A turn begun while memory was off stays private as a whole, also the part of its answer written
|
|
266
|
+
after `on`. `on` while a turn is still running, or while messages you typed during off are
|
|
267
|
+
still queued in dsh's inbox, takes effect when the chat is idle: at the turn's end, after a
|
|
268
|
+
stopped or failed turn, after an app restart, or when the queued message is removed. A later
|
|
269
|
+
`off`, even one sent as the turn ends, cancels it; dsh's own notices in the inbox do not hold it.
|
|
270
|
+
`off` itself holds at once, also while its write waits for a file another program holds: a
|
|
271
|
+
turn that ends meanwhile is not indexed, and a fork made, a recall asked or an `on` sent
|
|
272
|
+
meanwhile sees memory off; a second `off` keeps it held if the first one cannot write.
|
|
273
|
+
- Nothing from an off stretch becomes a chat's title: not a prompt typed there, not a dsh title
|
|
274
|
+
made from one, not after a damaged meta is repaired (it keeps the title it can read back and
|
|
275
|
+
takes none from the log). The first prompt, used only to match forks to their ancestors, is kept
|
|
276
|
+
as a 16-character fingerprint instead of text; metas from 0.8.0 are converted on their next write.
|
|
277
|
+
- Memory turned off under 0.8.0 (no off point recorded) or a meta recovered after damage: `on`
|
|
278
|
+
treats everything after the last index as typed while off, and a second `off` no longer moves
|
|
279
|
+
the off point forward. A damaged meta keeps its off stretches or, when they cannot be read back,
|
|
280
|
+
serves none of its turns to other chats, up to where memory is turned on again (forks made and
|
|
281
|
+
turns typed after that are not held back); a second damage keeps what the first one closed. Past 50 off stretches the two oldest merge; none is
|
|
282
|
+
dropped.
|
|
283
|
+
- A fork is private from its first message when its parent is, and memory-off when memory was off
|
|
284
|
+
at the fork point: its parent was off then, or the point lies in one of its off stretches (for a
|
|
285
|
+
fork of a fork, also its grandparent's), or the parent's damaged meta lost its stretches. It then
|
|
286
|
+
has no title until `on`. Its parent is the one dsh names, followed up dsh's chain of forks, so a
|
|
287
|
+
look-alike chat (a sibling fork, another chat that starts with "hi") no longer passes its flags
|
|
288
|
+
on. A fork takes no title from its parent: not dsh's "<title> (1)" rename of a new fork, not a
|
|
289
|
+
title made from inherited prompts, and a fork indexed by 0.8.0 or earlier drops the title it got
|
|
290
|
+
then at its next write (keeping the one that write gives it). A fork gets its folder when first seen (a few hundred bytes).
|
|
291
|
+
- While memory is off, the card no longer counts the turns typed since as waiting to be indexed,
|
|
292
|
+
and it shows the chat's own flags even when its topics cannot be read.
|
|
293
|
+
|
|
294
|
+
The savings estimate (#35)
|
|
295
|
+
- "Used" means the answer says two of the topic's words that the prompt did not bring, counted as
|
|
296
|
+
matching counts words: one Chinese word once (also when it sits next to the prompt's words), an
|
|
297
|
+
exact identifier of the label twice. An answer that only repeats the question no longer counts,
|
|
298
|
+
and one that states the topic's fact does.
|
|
299
|
+
- A used topic stands in for its newest cited messages (three at most) at their stored size (a
|
|
300
|
+
source an older version kept whole counts as if cut the same way), each message once per chat
|
|
301
|
+
however many topics cite it; what memory sent instead (the note, the recall result) is counted
|
|
302
|
+
once, as cost. Counting every cited message at its full length could show one recall as 78,000
|
|
303
|
+
tokens saved.
|
|
304
|
+
- A turn is judged by its own answer, also when the next message arrives while it is being
|
|
305
|
+
indexed; a stopped turn's topics are not judged against the next turn's answer; uses survive a
|
|
306
|
+
failed index or an unreadable ledger and are credited at the next turn end. Blocks a compaction
|
|
307
|
+
removed stop being charged as re-sent, also mid-turn and with memory off. A subagent forked from
|
|
308
|
+
a chat is charged to it: the blocks it carries and the recall tool.
|
|
309
|
+
- An unreadable ledger.json (a sync tool holding it) costs the savings figures, not the card or
|
|
310
|
+
the page: both say the savings are unavailable for now instead of showing zeros.
|
|
311
|
+
|
|
312
|
+
The page (#36, #62, #65)
|
|
313
|
+
- A late map or topic answer no longer replaces the one you asked for since, and Close drops a
|
|
314
|
+
topic still on its way; each topic opens a fresh panel (no leftover link or "Copied"), "Copied"
|
|
315
|
+
shows only when something was copied, and a topic that cannot be read says so.
|
|
316
|
+
- A large workspace's map fills its box (the view takes the box's proportions, from the newest
|
|
317
|
+
row) and drags at the pointer's speed; a map that failed to load can be asked again. Chat labels
|
|
318
|
+
fit their slot: the chat number, marks and hidden-topic count stay whole (the marks as 🔒 private
|
|
319
|
+
and ⏸ memory off where words do not fit) and hover shows the full text. The page states the rule
|
|
320
|
+
the estimate follows.
|
|
321
|
+
- The page skips a chat or a whole workspace it cannot read and leaves out chats without a chat
|
|
322
|
+
number (an unreadable meta); a repeated number (a copied folder) reaches the same chat, the one
|
|
323
|
+
with topics, on the map, on a topic and through recall (a copy that cannot be read is skipped,
|
|
324
|
+
and the asking chat never answers in its place);
|
|
325
|
+
a topic whose sources cannot be read still opens.
|
|
326
|
+
- A turn-end build that finds the chat busy keeps the running turn for the next one.
|
|
327
|
+
|
|
328
|
+
Files
|
|
329
|
+
- On Windows a write waits up to about 5 s for a file another program holds open (a sync tool, a
|
|
330
|
+
virus scan) instead of failing after 250 ms.
|
|
331
|
+
|
|
332
|
+
## 0.8.0 — 2026-11-16 — lean memory tree
|
|
333
|
+
|
|
334
|
+
Cycle 2. A six-angle audit of 0.7.0 (new chat, relevance, indexing call, drill-down, speed,
|
|
335
|
+
correctness) found 42 problems; 33 were confirmed and filed as #38-#56, next to the open token and
|
|
336
|
+
relevance issues from cycle 1 (#23-#31, #33, #34). This release makes memory a tree a new chat
|
|
337
|
+
reads from the top: one line per chat, then only the topics a message is about, then a topic's
|
|
338
|
+
source, then the whole turn from dsh's own log. Measured on the audit's fixed corpus (6 chats,
|
|
339
|
+
60 topics, English and Chinese, 20 questions and 20 everyday messages):
|
|
340
|
+
|
|
341
|
+
| | 0.7.0 | 0.8.0 |
|
|
342
|
+
|---|---|---|
|
|
343
|
+
| hits that answer the question (precision) | 40% | 100% (held-out set: 91%) |
|
|
344
|
+
| questions whose topic is found | 20/20 | 20/20 |
|
|
345
|
+
| everyday messages that get a block | 7/20 | 0/20 |
|
|
346
|
+
| tokens injected over the 40 messages | 8,235 | 3,045 |
|
|
347
|
+
| first-message chat list, 50 chats | 632 tokens | under 200 |
|
|
348
|
+
| recall tool definition, every request | about 210 tokens | about 100 |
|
|
349
|
+
| a turn recall, worst case on real logs | 18,633 tokens | about 2,000 |
|
|
350
|
+
| per-message cost, 150 chats / 9,000 topics | 186 ms | about 19 ms (median) |
|
|
351
|
+
|
|
352
|
+
**Only relevant memory is sent** (#23, #43, #39, #31, #51)
|
|
353
|
+
- A topic must earn its place: each prompt word in its label or aliases counts 2, in its gist 0.5,
|
|
354
|
+
and it needs a label word and a score of 3. Rarity only breaks ties. Filler words ("can you
|
|
355
|
+
help", 我们, 的, hiragana endings) no longer count. (#23)
|
|
356
|
+
- The same subject in several chats takes one slot; a lower hit stays only if it matches a word
|
|
357
|
+
the higher ones did not; a topic already in the chat's context leaves its slot empty instead of
|
|
358
|
+
pulling in a weaker one. (#23, #43)
|
|
359
|
+
- A topic's source is included only for a hit with three label words that is clearly ahead, and
|
|
360
|
+
only up to 300 tokens (budgets are now in tokens, so Chinese no longer costs 3.5x). (#39)
|
|
361
|
+
- Topics the model fetched with `recall` count as already shown; ids are read from topic lines
|
|
362
|
+
only, so the example ids in a header can no longer hide a real topic. (#31, #51)
|
|
363
|
+
|
|
364
|
+
**The top layer** (#27, #26, #38, #41, #40)
|
|
365
|
+
- The indexing call also writes a one-line summary of the chat (no extra call). A new chat's
|
|
366
|
+
first message gets one line per earlier chat that has topics, the ones sharing words with the
|
|
367
|
+
message first, within 200 tokens. (#27, #38)
|
|
368
|
+
- The list is sent once per chat: not again after an app restart while it is still in the chat's
|
|
369
|
+
context. (#26)
|
|
370
|
+
- Chats use dsh's own title (model-made, or your rename) instead of the first prompt. (#41)
|
|
371
|
+
- A chat resumed for a few turns counts as active today even before it is indexed again. (#40)
|
|
372
|
+
- Quick chats that never reach an index no longer get a folder. (#38)
|
|
373
|
+
|
|
374
|
+
**Time** (#24, #25)
|
|
375
|
+
- A named time ranks topics from then first instead of hiding everything else; an exact older match
|
|
376
|
+
still wins. A day digest is sent only for a question about the past ("what did we do yesterday",
|
|
377
|
+
"上次聊了什么"), not for "today I want to learn rust". Day digests list only topics from that day. (#24)
|
|
378
|
+
- 目前天气, 登上月球 and 这周期 are no longer read as dates; years (2025年11月1日) and invalid dates
|
|
379
|
+
(2026-02-31) are handled; ranges are whole calendar days across daylight-saving changes; words in
|
|
380
|
+
any alphabet (café, Привет, full-width API) are searchable. (#25)
|
|
381
|
+
|
|
382
|
+
**The indexing call** (#28, #44, #45, #46, #53)
|
|
383
|
+
- The prompt starts with fixed rules (identical for every build of every chat), then the chat's
|
|
384
|
+
topics in id order, then the span, so the provider can cache everything but the new text. (#28)
|
|
385
|
+
- It runs at the route's lowest reasoning effort, picked the way dsh's own title generator does;
|
|
386
|
+
no effort is sent for routes without one. (#44)
|
|
387
|
+
- A failed span is retried as it was, not grown by every turn since; an answer cut off by the
|
|
388
|
+
token cap is retried with its first half; after three unusable answers only that span is
|
|
389
|
+
skipped, and the next build waits the normal 4 turns, not 16. (#45, #53)
|
|
390
|
+
- Every human prompt is read: assistant text is shortened (2,000, 800, then 300 characters) before
|
|
391
|
+
anything else, and a span too long for one call is indexed over several, oldest turns first. (#46)
|
|
392
|
+
|
|
393
|
+
**Drill-down** (#34, #47, #29, #30, #48)
|
|
394
|
+
- `recall("c3t7")` reads turn 7 of chat 3 from dsh's own session log, so the deepest layer costs no
|
|
395
|
+
storage. Only inside what that chat shares: turns it has indexed, and never turns a fork
|
|
396
|
+
inherited. (#34)
|
|
397
|
+
- A turn is the prompt, the answers and one line of tool use (`[tools: read_file ×3, bash]`), about
|
|
398
|
+
2,000 tokens at most; tool output is left out. (#47)
|
|
399
|
+
- A topic recall gives its newest 3 cited messages, 1,200 characters each, each naming its turn,
|
|
400
|
+
and says which older turns hold more. (#29)
|
|
401
|
+
- The recall tool definition is half its size (the `depth` option is gone). (#30)
|
|
402
|
+
- `recall("c3")` past 40 topics names the rest instead of stopping silently. (#48)
|
|
403
|
+
|
|
404
|
+
**Speed** (#33, #49)
|
|
405
|
+
- Files are read through a cache keyed by their stat, each topic's words are computed once, and
|
|
406
|
+
topics that cannot reach the floor are skipped before their evidence is counted: about 19 ms per
|
|
407
|
+
message at 150 chats and 9,000 topics. Shared cached data is frozen, so no reader can
|
|
408
|
+
change it by accident. (#33)
|
|
409
|
+
- A message naming a time is matched once, not three times; `map` loads the workspace once;
|
|
410
|
+
`recall` of a topic or chat loads only that chat; forks are no longer re-detected every turn. (#49)
|
|
411
|
+
|
|
412
|
+
**Correctness** (#50, #52, #54, #55, #56, #57)
|
|
413
|
+
- A chat whose meta became unreadable reads as private with memory off right away, not only after
|
|
414
|
+
its next write. (#50)
|
|
415
|
+
- Forks are handled by content: a topic of any chat (parent, grandparent) whose every source is
|
|
416
|
+
word for word in the fork's own log is not sent to it, and a chat the fork holds entirely is not
|
|
417
|
+
listed. (#16, #52)
|
|
418
|
+
- A damaged file with any amount of trailing junk is repaired in one pass. (#54)
|
|
419
|
+
- `stats` in a new fork counts only the fork's own turns. (#55)
|
|
420
|
+
- The v0.1-v0.6 backup is removed 7 days after its last write, not after the first upgrade. (#56)
|
|
421
|
+
- `npm run test:zones` really runs the suite in UTC, Pacific/Auckland and Los Angeles; in Git Bash
|
|
422
|
+
`TZ=... npm test` silently kept the local zone. (#57)
|
|
423
|
+
|
|
424
|
+
**Found by the two independent reviews of this change, and fixed** (first: 4 angles, 24 findings, 22 confirmed by reproduction; second, on those fixes: 2 angles, 11 findings, 9 confirmed)
|
|
425
|
+
- Evidence is counted per word of the message, not per token: one Chinese word (数据库) or one
|
|
426
|
+
hyphenated word no longer passes as two or three label words, so Chinese and English prompts need
|
|
427
|
+
the same evidence; a word already counted adds nothing when it shows up again inside a longer word
|
|
428
|
+
(in either order); a katakana or Hangul word is one word however long; an exact identifier counts
|
|
429
|
+
double, while bits of code in a pasted traceback count only as gist words, after the plain words
|
|
430
|
+
of the message, so naming a file first no longer weakens the question; "ci/cd" or "read/write" are
|
|
431
|
+
plain words; very long messages need slightly more evidence.
|
|
432
|
+
- A topic already shown no longer hides its replacement: a decision that was superseded, or a
|
|
433
|
+
newer copy of the same subject that says something else, is sent; the newest copy speaks for a
|
|
434
|
+
subject even when it would not pass the gate alone; "same subject" and "says the same" compare the
|
|
435
|
+
text itself, so digits and "not" count ("retries 3" is not "retries 5", "Python 2 support" not
|
|
436
|
+
"Python 3 support"); on equal evidence the topic the message names exactly comes first and is not
|
|
437
|
+
cut by its equal, and a lead must be ahead of every candidate; `/context-graph find` lists every
|
|
438
|
+
copy.
|
|
439
|
+
- A named time reorders the topics after they are chosen, so it can never push an exact older match
|
|
440
|
+
out; a day digest answers only questions about shared earlier work, which must name "we" or "you"
|
|
441
|
+
("before lunch today", "did we win", "the last time I smoked", 今天做了红烧肉, 和朋友聊过 get
|
|
442
|
+
nothing); a fork's digest leaves out the chats it holds word for word.
|
|
443
|
+
- `recall` budgets are tokens everywhere: a topic's source is about 1,000 tokens in any language, a
|
|
444
|
+
turn read from stored messages (no dsh log) keeps the 2,000-token turn budget, and cuts are
|
|
445
|
+
placed by adding up each character's cost; no cut leaves half of an emoji or rare character.
|
|
446
|
+
- The chat list is sent on a chat's first turn (or after `on`) and never again in that chat, after
|
|
447
|
+
a restart or a compaction.
|
|
448
|
+
- A subagent started by a subagent follows the chat that started the task; a chain that cannot be
|
|
449
|
+
followed fails closed.
|
|
450
|
+
- A fork compares stored messages only for chats that share its first prompt (its possible
|
|
451
|
+
ancestors), instead of reading every chat's `msgs.json` on its first message.
|
|
452
|
+
- `recall` of a chat number shared by two chats (possible in pre-0.7 data) finds the other chat,
|
|
453
|
+
never a private one; two readers of a folder being upgraded from 0.4-0.6 both see its topics; a
|
|
454
|
+
build never moves the index past what it read.
|
|
455
|
+
- Forks indexed by 0.4-0.6 (whose topics were drawn from inherited messages) hold those back: a
|
|
456
|
+
fork opened again records its fork point, and `recall` asks dsh for it first, so neither its
|
|
457
|
+
turns nor its topics expose the chat it was forked from. One never opened again keeps offering its
|
|
458
|
+
inherited topic lines to matching until either happens (#58).
|
|
459
|
+
- A config block written for 0.7 still applies (`inlineMaxChars`, `surfaceMaxChats`).
|
|
460
|
+
- Every fix in this release is mutation-checked: breaking it makes a test fail.
|
|
461
|
+
|
|
462
|
+
**Changed**
|
|
463
|
+
- Config: `inlineMaxChars` and `surfaceMaxChats` are replaced by `inlineMaxTokens` (300),
|
|
464
|
+
`listMaxTokens` (200, with an optional `listMaxChats`), `sourceMaxMessages` (3), `sourceMaxTokens`
|
|
465
|
+
(1000) and `turnMaxTokens` (2000); the old keys in a 0.7 block are still read.
|
|
466
|
+
- `/context-graph find` lists every topic with at least one matching label word (up to 10).
|
|
467
|
+
- `meta.json` gains `summary` and `firstPrompt`; `turns` is no longer written.
|
|
468
|
+
- Tests: 124 (was 62), including the fixed relevance corpus, the three-zone runner, and a mutation
|
|
469
|
+
check: each of the 57 fixes in this release, broken on purpose, makes the suite fail.
|
|
470
|
+
|
|
471
|
+
## 0.7.0 — 2026-11-15 — safe and structured
|
|
472
|
+
|
|
473
|
+
Cycle 1. A six-angle audit of 0.6.2 found 59 problems; 55 were confirmed by independent
|
|
474
|
+
reproduction and filed as issues #10-#37. This release fixes everything that could lose, corrupt or
|
|
475
|
+
mix up stored memory, or let one chat's text act as instructions in another (milestone v0.7.0), plus
|
|
476
|
+
the storage duplication (#32). An independent four-angle review of the change then found 23 more
|
|
477
|
+
problems in the new code itself; all are fixed here too. Token, relevance and speed work follows in
|
|
478
|
+
0.8.0; the savings display and the visual map in 0.9.0.
|
|
479
|
+
|
|
480
|
+
**Stored memory cannot be lost, mixed up or corrupted**
|
|
481
|
+
- Concurrent writes can no longer corrupt files (unique temp names, a retried rename on Windows,
|
|
482
|
+
each chat's writes serialized). (#10)
|
|
483
|
+
- A damaged file never takes the workspace down and is never silently overwritten: trailing junk
|
|
484
|
+
(what 0.6 could leave) is repaired; an unreadable file is set aside as `<file>.corrupt-<time>`;
|
|
485
|
+
a chat whose meta was unreadable keeps its number and index position and fails closed (private,
|
|
486
|
+
memory off) because the lost file may have said so. (#10)
|
|
487
|
+
- `recall` never writes, which removes a lost update that could drop a chat's newest topics. (#10)
|
|
488
|
+
- Chat numbers are claimed with an exclusive create, safe across dsh windows. (#11)
|
|
489
|
+
- Each cited message is stored once per chat (`msgs.json`, keyed by seq) instead of once per topic,
|
|
490
|
+
and a build commits text, then topics, then its index position. (#12, #32)
|
|
491
|
+
- Turns that dsh compaction hid before they were indexed are still indexed (the builder reads the
|
|
492
|
+
durable log). (#13)
|
|
493
|
+
- Answers citing messages as `"m0"`, or mentioning `"nodes": [` before the real JSON, are read
|
|
494
|
+
correctly; an answer whose topics cite nothing usable is retried, an honest empty answer means
|
|
495
|
+
"nothing worth remembering", and after three unusable answers in a row the span is skipped instead
|
|
496
|
+
of being re-sent, larger, every few turns. (#14)
|
|
497
|
+
- Graph merges resolve links to earlier topics by label, use only real `nN` ids (a topic labelled
|
|
498
|
+
`constructor` no longer collides with `Object.prototype`), repair `nextId`, and cap labels (80
|
|
499
|
+
chars), gists (200) and aliases (12 x 40, newest first). (#15)
|
|
500
|
+
|
|
501
|
+
**Older memory comes back, safely**
|
|
502
|
+
- Memory written by 0.1-0.3 (one `graph.json` per workspace) is usable again; a graph that covered
|
|
503
|
+
several sessions keeps each session's messages apart. (#17)
|
|
504
|
+
- Upgrades never delete or cut anything on the spot: the new data goes to `topics.json` and
|
|
505
|
+
`msgs.json`, migrated text is kept in full, and the old files stay unchanged as a backup for 7
|
|
506
|
+
days. An older copy of the plugin still running elsewhere can no longer wipe upgraded memory; topics
|
|
507
|
+
it adds meanwhile are merged in. An interrupted upgrade simply runs again. On real data the same
|
|
508
|
+
memory takes 3,826 bytes instead of 9,015 once the backup is gone.
|
|
509
|
+
|
|
510
|
+
**Security**
|
|
511
|
+
- Every block carrying text from another chat says once that it is quoted data, not instructions,
|
|
512
|
+
and exact source text sits in a fence it cannot close. (#18)
|
|
513
|
+
|
|
514
|
+
**Behaviour**
|
|
515
|
+
- A forked chat starts its index after the history it inherited (detected from dsh's
|
|
516
|
+
`inheritedEventCount`, whether or not dsh records the parent), and is not sent the parent's topics
|
|
517
|
+
whose sources it already holds; parent topics citing later messages still reach it. (#16)
|
|
518
|
+
- `/context-graph on` answers at once and indexes in the background; `build` never takes the
|
|
519
|
+
agent's maintenance slot, never splits a running turn's question from its answer, and stops when
|
|
520
|
+
the command is cancelled. `stats` shows how the last index went. (#19)
|
|
521
|
+
- With memory off a chat gets no memory blocks and `recall` reads only its own turns; a subagent
|
|
522
|
+
follows the chat that started it. The update notice is never added to a chat; `stats` shows it. (#20)
|
|
523
|
+
- The meter counts only conversation requests as conversation; compaction and title calls have
|
|
524
|
+
their own line. (#21)
|
|
525
|
+
|
|
526
|
+
**Compatibility**
|
|
527
|
+
- Works on every dsh 0.2.1 build from rc.1 on, and on 0.2.2; refuses 0.3 prereleases until tested
|
|
528
|
+
(`engines.dsh >=0.2.1-rc.1 <0.3.0-0`). (#9, #22)
|
|
529
|
+
|
|
530
|
+
**Structure**
|
|
531
|
+
- Code moved into layers: `src/core` (pure), `src/store` (disk), `src/host` (dsh wiring),
|
|
532
|
+
`src/client` (browser). See [docs/architecture.md](docs/architecture.md). Tests mirror the layout:
|
|
533
|
+
62 tests, run in three time zones.
|
|
534
|
+
|
|
535
|
+
Contributors: William20100221 (author), Claude (Anthropic; audit, fixes, review).
|
|
536
|
+
|
|
537
|
+
## 0.6.2 — 2026-11-04
|
|
538
|
+
- Cap the day list sent with a message that names a time.
|
|
539
|
+
|
|
540
|
+
## 0.6.1 — 2026-11-04
|
|
541
|
+
- Fix seven bugs found in the 0.6.0 review (#1-#7): builds in chats that never ran a command, index
|
|
542
|
+
position lost on restart, retry storms, oversized first builds, repeated topics, wrong texts,
|
|
543
|
+
flags undone by a finishing build.
|
|
544
|
+
|
|
545
|
+
## 0.6.0 — 2026-11-04
|
|
546
|
+
- Chats are shared with the workspace by default; `/context-graph unshare` keeps one private.
|
|
547
|
+
|
|
548
|
+
## 0.5.0 — 2026-11-04
|
|
549
|
+
- Memory on in every chat by default; the plugin switch on the Plugins page turns it off everywhere,
|
|
550
|
+
`/context-graph off` for one chat. Chat dates are local days so recall by date matches.
|
|
551
|
+
|
|
552
|
+
## 0.4.0 — 2026-11-02
|
|
553
|
+
- One graph per chat with a share flag, time-first matching, recall by chat and day; folding removed.
|
|
554
|
+
|
|
555
|
+
## 0.3.0 — 2026-11-02
|
|
556
|
+
- dsh-native look: icon, English and Chinese title and description, Chinese detail page, command card.
|
|
557
|
+
|
|
558
|
+
## 0.2.x — 2026-11-02
|
|
559
|
+
- 0.2.3: update check against npm. 0.2.2: builder survives thinking models. 0.2.1: the command
|
|
560
|
+
declares its arguments. 0.2.0: folding of old turns (removed in 0.4.0).
|
|
561
|
+
|
|
562
|
+
## 0.1.0 — 2026-11-02
|
|
563
|
+
- First version: graph memory of past conversations for DeepSeek Harness.
|