@joe-broadhead/opencode-mem 0.13.1 → 0.14.0
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 +99 -0
- package/package.json +1 -1
package/CHANGELOG.md
CHANGED
|
@@ -3,6 +3,105 @@
|
|
|
3
3
|
The format follows Keep a Changelog 1.1.0: one `## [x.y.z] - date` section per
|
|
4
4
|
release.
|
|
5
5
|
|
|
6
|
+
## [0.14.0] - 2026-11-18
|
|
7
|
+
|
|
8
|
+
### Added
|
|
9
|
+
|
|
10
|
+
- **The routines engine is built, and off by default.** `RECORDS_ENGINE` chooses its mode. `off`, the default, writes
|
|
11
|
+
exactly what 0.13.1 wrote: no marker, no lease, no extra column. `ingest` numbers each event, hashes its content and
|
|
12
|
+
assigns its lane inside the ingest transaction, with no model call at all. `shadow` adds the scheduler's leases,
|
|
13
|
+
batches and prompts, and reads the cohort `RECORDS_ENGINE_SPACES` names into a ledger without proposing anything.
|
|
14
|
+
`on` is reserved for the release that turns the engine live.
|
|
15
|
+
- **What runs whatever the flag:** each space's meta and its migration check are cached by the file's change time, so a
|
|
16
|
+
tick that opens every space no longer re-reads them; `RECORDS_MAX_OPEN` sets how many space files one process keeps
|
|
17
|
+
open; the disk check and the page it raises run as before.
|
|
18
|
+
- **A space's handle is reference-counted across an await.** A caller that holds a space across a model call marks the
|
|
19
|
+
handle, and the last release closes it. The `RECORDS_CLAIMS` enablement checklist needs this before the claims store
|
|
20
|
+
can hold a space open across a call.
|
|
21
|
+
- **A person's erasure deletes their core audit and knowledge rows.** After the
|
|
22
|
+
records purge and every memory layer, the records service's `person.eraseCore`
|
|
23
|
+
step calls the core's `POST /v3/ops/erase/user` with the id the gateway sends
|
|
24
|
+
as `x-tdai-user-id` (control's `upstream_user_id`, or `users.id` when the row
|
|
25
|
+
has none) and the same core key and instance header the layer purge uses. It
|
|
26
|
+
is idempotent, and its counts are stored in the erasure's report as
|
|
27
|
+
`core.erase_user` before the erasure is reported, so a replay after a crash
|
|
28
|
+
keeps them.
|
|
29
|
+
- **The core rows never hold the rest of an erasure.** The records purge, the
|
|
30
|
+
personal space's purge and the shared spaces run whatever the core answers,
|
|
31
|
+
and the erasure finishes `done` with the core part reported beside it.
|
|
32
|
+
- A failure (a 5xx, a timeout, an unreadable answer) retries in the
|
|
33
|
+
background on its own clock: five minutes, doubling to an hour, never
|
|
34
|
+
stopping (`core.erase_user.ok: false`, `attempts`, `next_attempt_at`).
|
|
35
|
+
- A core without the route (404, an older core) completes with a recorded
|
|
36
|
+
gap (`supported: false`, "not supported by this core"), is not retried, and
|
|
37
|
+
logs `memory.erasure.core_unsupported` once, ids only, for an alert.
|
|
38
|
+
- The operator runbook says what the core part deletes, which id it uses and
|
|
39
|
+
what each answer means:
|
|
40
|
+
[docs/operations/account-erasure.md](docs/operations/account-erasure.md).
|
|
41
|
+
- **`space-purge --core-purged-offline` purges a container whose core part ran
|
|
42
|
+
offline.** A shared container's core rows cannot be read back through the
|
|
43
|
+
API: a scope's L0 query answers 422 without isolation, the agent list answers
|
|
44
|
+
401, and a read that says zero proves nothing. With this flag the script reads
|
|
45
|
+
no core layer at all, and proves the core empty with the two records the
|
|
46
|
+
coordinator asked for:
|
|
47
|
+
- `--core-purge-report <path>`: the offline purge's own `report.json`, in the
|
|
48
|
+
VM's root-only purge copy folder (`backups/purge/<team>-<UTC>/` under the
|
|
49
|
+
ocmem root) and mounted read-only into the container that runs the script;
|
|
50
|
+
nothing is run from a laptop. It must
|
|
51
|
+
say `verdict: "done"`, `verified: true`, not a restore and not a dry run,
|
|
52
|
+
for this container, and be made within 24 hours (the copy folder's UTC
|
|
53
|
+
stamp). Its `verified` is the purge's post-check: zero of the team's rows in
|
|
54
|
+
every table of every database it walked, no lines or scope folders left,
|
|
55
|
+
and both FTS integrity checks passing.
|
|
56
|
+
- The core's count route, `/v3/ops/purge/team` `dry_run: true`, for every
|
|
57
|
+
service the volume lists (from `/v3/ops/orphans`: the default and each
|
|
58
|
+
`instances/<id>`), never only `--core-services`. Every table, vector, JSONL
|
|
59
|
+
line and scope folder count must be zero; a non-zero one stops with
|
|
60
|
+
`cell_not_empty`, naming the table.
|
|
61
|
+
- The journal records the report's path and SHA-256 and the route's counts.
|
|
62
|
+
A resume re-checks the report; while the cell step is pending it refuses one
|
|
63
|
+
that changed, and after control's purge it accepts a fresh report, so a
|
|
64
|
+
delete that waited past the report's 24 hours is unblocked by one more
|
|
65
|
+
`ocmem purge-team`. The hub's wikis and code graphs keep the released path,
|
|
66
|
+
and the flag is refused for a personal space.
|
|
67
|
+
- It runs on the VM: the core, its routes and the keys are only there, so the
|
|
68
|
+
report is mounted read-only into the container that runs the script. No
|
|
69
|
+
image ships `scripts/`; before a release carries it, a hash-approved
|
|
70
|
+
scratch copy is used, as the 18 November run did. Both are in
|
|
71
|
+
[docs/deployment.md](docs/deployment.md#a-container-whose-core-part-ran-offline).
|
|
72
|
+
- The runbook says how to run it:
|
|
73
|
+
[docs/deployment.md](docs/deployment.md#purge-and-delete-a-space).
|
|
74
|
+
|
|
75
|
+
### Fixed
|
|
76
|
+
|
|
77
|
+
- **A collection failure that names no machine no longer overwrites a named machine's report.** A collector too old to
|
|
78
|
+
send `host` can collect for the same person on its own schedule, from another machine, and fail there on every tick.
|
|
79
|
+
When the standing report names a machine, that failure is kept beside it with the fixed label
|
|
80
|
+
`unnamed: an older Memory` (`elsewhere: {host, at, error}`), and the standing row — `ok`, `at`, `error`, `accepted`,
|
|
81
|
+
`failed_since`, `more`, `host` — is untouched, so Sources no longer shows a source as failing between the reading
|
|
82
|
+
machine's runs. With no named machine standing, such a failure is the binding's, as before.
|
|
83
|
+
|
|
84
|
+
### Notes
|
|
85
|
+
|
|
86
|
+
- **Leave the engine off in production.** At full scale (2,000 active and 10,000 idle spaces, five pairs, 18 November
|
|
87
|
+
2026), `off` keeps ingest p99 within decision 12 of 0.13.1: +0.48% at 300 characters and +1.94% at 1,000, against
|
|
88
|
+
a 3% target. `ingest` adds 3.2 ms per 100-event page at 300 characters, within its 4 ms target, but 9.53 ms at
|
|
89
|
+
1,000 characters, outside it. Keep `RECORDS_ENGINE` unset until a later release brings `ingest` within the target
|
|
90
|
+
at both sizes.
|
|
91
|
+
- With `RECORDS_ENGINE` unset or `off`, the consent path the 0.13.1 release shipped is unchanged: the collector's
|
|
92
|
+
plugin requests, `plugins.cancel`, the request cap, the account grammar, `needsConsent` and the console behave
|
|
93
|
+
exactly as they did.
|
|
94
|
+
- **`routines.status` gains an `engine` key.** It is an additive field: `null` while the engine is off, and otherwise
|
|
95
|
+
the engine's pauses and its version pauses. A client that reads the status by its known keys is unaffected.
|
|
96
|
+
- **The `classify` tier is the engine's own.** The engine's classifier lease reports its call to the guard's
|
|
97
|
+
`classify` tier; the routines' trait client is unchanged and keeps the guard's default tier, so nothing the daily
|
|
98
|
+
does moved.
|
|
99
|
+
- **After a rollback to 0.13.1, a move into a space 0.14.0 touched is refused.** 0.13.1's `space-ops` decides whether a
|
|
100
|
+
space is empty from its `meta` rows and does not know the `additive_digest` row 0.14.0 writes, so an otherwise empty
|
|
101
|
+
space reads as holding content and a move into it is refused. It fails safe: nothing is moved and nothing is lost.
|
|
102
|
+
`docs/operations/records-rollback.md` says what an operator does — the move's target simply is not empty to the older
|
|
103
|
+
build, and the one row can be removed after the rollback.
|
|
104
|
+
|
|
6
105
|
## [0.13.1] - 2026-11-18
|
|
7
106
|
|
|
8
107
|
### Added
|