@xiangyixie/dsh-superagent 0.10.0 → 0.12.0
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 +286 -88
- package/README.md +571 -39
- package/bin/archive.js +185 -0
- package/bin/doctor.js +84 -0
- package/bin/sync.js +187 -0
- package/bin/workbench.js +98 -0
- package/docs/policy.example.json +38 -0
- package/lib/audit.js +555 -0
- package/lib/cost.js +267 -0
- package/lib/doctor.js +878 -0
- package/lib/guard.js +841 -0
- package/lib/index.js +2031 -1104
- package/lib/inject.js +138 -0
- package/lib/ledger.js +270 -0
- package/lib/memory.js +736 -503
- package/lib/orchestrator.js +1950 -217
- package/lib/policy.js +739 -0
- package/lib/report.js +1068 -733
- package/lib/store.js +873 -402
- package/lib/sync.js +456 -0
- package/lib/templates.js +690 -0
- package/lib/watch.js +810 -458
- package/lib/workbench.js +1119 -0
- package/lib/workspace.js +433 -0
- package/package.json +92 -70
- package/public/workbench.html +836 -0
package/CHANGELOG.md
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
# Changelog
|
|
1
|
+
# Changelog
|
|
2
2
|
|
|
3
3
|
Doc-only releases: npm serves the README from the tarball, so a README edit is
|
|
4
4
|
invisible on the package page until something is published. When the edit
|
|
@@ -8,93 +8,291 @@ along with the next code release. A doc-only version leaves `lib/` and `bin/`
|
|
|
8
8
|
byte-identical to the version before it, and the pinned install line in the
|
|
9
9
|
README is updated to the version being released.
|
|
10
10
|
|
|
11
|
-
## 0.
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
- **
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
11
|
+
## 0.12.0 — 2026-11-15
|
|
12
|
+
|
|
13
|
+
### Added
|
|
14
|
+
|
|
15
|
+
- **Worker safety guard.** Child tool calls are now checked on the host's
|
|
16
|
+
`tools/pre-execute` / `ctx.tools.guard` pipeline. Writes are limited to the
|
|
17
|
+
child's workspace and the OS temp directory; credential stores and
|
|
18
|
+
`.git/config` / `.git/hooks` are refused for reading and writing; system
|
|
19
|
+
directories are refused for writing. Default command rules block disk
|
|
20
|
+
formatting, system scheduler and registry changes, writes into system
|
|
21
|
+
directories, destructive recursive deletes whose target is `/`, `~`, `.`,
|
|
22
|
+
`..`, a wildcard, or outside the workspace, permission nukes, and commands
|
|
23
|
+
that name a credential store. `guard.allowCommands` is an explicit,
|
|
24
|
+
audited override. Blocked calls surface in `superagent_status` and in a
|
|
25
|
+
`## Safety guard` report section. See “Worker safety guard” in the README
|
|
26
|
+
for the full defaults and their rationale.
|
|
27
|
+
- **Permission policy.** `<stateDir>/policy.json` (or `policyFile`) lets a person
|
|
28
|
+
say which paths may be written, which commands may run, which action classes
|
|
29
|
+
need a human, and what the default level is per run. The file accepts `//`,
|
|
30
|
+
`#`, and `/* */` comments, validates against a schema, and reports every
|
|
31
|
+
problem with file, line, and column. It only tightens: deny wins, explicit
|
|
32
|
+
allow entries record what they relaxed, and an invalid policy fails closed. A
|
|
33
|
+
start the policy denies is refused before any run record exists; a start held
|
|
34
|
+
for confirmation needs `approve_policy: true`, which is logged.
|
|
35
|
+
- **Audit trail.** Append-only JSONL at `<stateDir>/audit/audit.jsonl`: tool
|
|
36
|
+
calls, path writes, commands, blocks, approvals, attention decisions, and run
|
|
37
|
+
lifecycle, each with run id, subtask id, actor, timestamp, decision, result,
|
|
38
|
+
and duration. Appends are serialized in-process and locked across processes;
|
|
39
|
+
a torn tail or a damaged line is reported and skipped instead of hiding the
|
|
40
|
+
log. `audit: false` turns the trail off without changing run behaviour.
|
|
41
|
+
- **`superagent_audit`** queries by run, subtask, kind, decision, result, target,
|
|
42
|
+
and time window, and exports the matching events as one JSONL or CSV file.
|
|
43
|
+
- `createAuditor(...).attachGuard(fn)` is the seam for a separate path/command
|
|
44
|
+
guard: it can tighten a decision, never relax one.
|
|
45
|
+
- `superagent_inject(run_id, message, subtask_index?, wait_seconds?)` steers a
|
|
46
|
+
running child instead of waiting for the next prompt. It calls the host's
|
|
47
|
+
documented live channel (`subagents.sendMessage`, which lands in the same
|
|
48
|
+
`Agent` inbox that `steer`/`inject` expose), so the message is claimed at the
|
|
49
|
+
child's next pre-step and becomes part of its next model request. A child
|
|
50
|
+
that had already decided to stop runs one more turn to read it. Receipts
|
|
51
|
+
record `latencyMs` from acceptance to the claim, or a `missedReason` when the
|
|
52
|
+
child is cancelled, disposed, or finishes before another boundary. On a real
|
|
53
|
+
`dsh@0.2.2-rc.1` host the measured delay was 2986 ms with a 4 s child step
|
|
54
|
+
and 6978 ms with an 8 s step, i.e. bounded by the remainder of the running
|
|
55
|
+
step, not by the plugin. A host without that channel degrades to the followup
|
|
56
|
+
queue and the reply labels it as such.
|
|
57
|
+
- **Multi-device sync over git.** `dsh-superagent-sync` turns the state
|
|
58
|
+
directory into a repository and carries it between machines through any git
|
|
59
|
+
remote — no server, no account, no dependency. `init <remote>` writes a
|
|
60
|
+
managed `.gitignore` block (credentials, locks, cancel markers, quarantine
|
|
61
|
+
files, temp files, logs, machine identity) and makes the first commit;
|
|
62
|
+
`pull` fetches and merges, stashing uncommitted local work first and saying
|
|
63
|
+
so; `push` commits local state and pushes, reporting a rejection instead of
|
|
64
|
+
forcing; `status` reports ahead/behind counts, uncommitted files, stashes,
|
|
65
|
+
and conflicted paths. A conflict stops the merge with both sides kept and the
|
|
66
|
+
paths listed — nothing is resolved automatically. Run records were already
|
|
67
|
+
one file per run, so ordinary concurrent work on two machines does not
|
|
68
|
+
collide; the remaining shared surface is an append-only memory scope. The
|
|
69
|
+
README documents the limits: no realtime channel, no cross-machine lock, no
|
|
70
|
+
key distribution.
|
|
71
|
+
- **Local workspace panel.** `workbenchPort` (off by default, loopback only)
|
|
72
|
+
serves an IDE-shaped page from `public/workbench.html`: file tree, text editor,
|
|
73
|
+
diff preview before save, and a run view that lists the files a run changed.
|
|
74
|
+
The plugin reads and edits the filesystem itself, so no host API is involved.
|
|
75
|
+
Every request needs the per-process bearer token in `workbench.json`; absolute
|
|
76
|
+
paths, traversal, symlinks out of the root, the state directory, binaries, and
|
|
77
|
+
files above 512 KB are refused with a reason. Saving requires the hash and
|
|
78
|
+
mtime the editor read, so an external edit is never overwritten.
|
|
79
|
+
- `dsh-superagent-workbench` runs the same panel standalone, with `--root`,
|
|
80
|
+
`--port`, and `--state-dir`; no host is required.
|
|
81
|
+
- Run records now carry `changes` (recorded when the panel is on), which is what
|
|
82
|
+
the panel's run view reads and diffs against `git HEAD`.
|
|
83
|
+
|
|
84
|
+
## 0.11.0 — 2026-11-15
|
|
85
|
+
|
|
86
|
+
Runs became something you can talk back to and shape: instructions and answers
|
|
87
|
+
reach a run that already exists, planned subtasks can depend on each other,
|
|
88
|
+
workers can be isolated in their own worktrees, and the panel and report show
|
|
89
|
+
parallel structure instead of opaque rows. Runs are now priced down to the
|
|
90
|
+
subtask, spending is kept in a ledger, common goals start from a template, and
|
|
91
|
+
an install can check itself. No runtime dependencies were added, and the package
|
|
92
|
+
still ships one file per module from `lib/`.
|
|
93
|
+
|
|
94
|
+
### Added
|
|
95
|
+
|
|
96
|
+
- **Direct mode.** `superagent_start` takes `mode: "direct"`: one worker does the
|
|
97
|
+
whole goal, with no planner, no aggregation, and no self-heal round, and its
|
|
98
|
+
reply becomes the report verbatim. The run record, quota, deadline, cancel,
|
|
99
|
+
budget, milestones, and per-child usage accounting are identical to a planned
|
|
100
|
+
run. `mode: "planned"` is the default and unchanged.
|
|
101
|
+
- **Queueing instead of a hard failure.** `superagent_start` takes
|
|
102
|
+
`queue: true`: when `maxActiveRuns` is in use the run returns
|
|
103
|
+
`status: "queued"` with `queue.position` / `queue.of`, and starts automatically
|
|
104
|
+
when an active run finishes. Waiting runs do not count against
|
|
105
|
+
`maxActiveRuns`, are persisted with an owner heartbeat, start oldest first, and
|
|
106
|
+
renumber when one is cancelled. `maxQueuedRuns` (default 5, 0 disables) caps the
|
|
107
|
+
line and `queueTimeoutMinutes` (default 30) fails a run that never got a slot,
|
|
108
|
+
without spending anything. `queueOnFull: true` makes queueing the default for
|
|
109
|
+
every start; without it a start that does not pass `queue: true` still gets the
|
|
110
|
+
speed quota error.
|
|
111
|
+
- **Relevance-ranked memory recall.** `superagent_memory` takes
|
|
112
|
+
`action: "search"` with `query` and/or `tags`; a new note supersedes the older
|
|
113
|
+
one with the same key, and the planner digest now ranks notes against the run's
|
|
114
|
+
goal instead of only taking the newest ones.
|
|
115
|
+
- `superagent_followup(run_id, message)` appends an instruction to an active
|
|
116
|
+
run. It stays next-prompt by design: the next worker spawn or the aggregation
|
|
117
|
+
carries it, whichever comes first, and subtasks already running or finished
|
|
118
|
+
keep their results. Each instruction records `appliedAt` / `appliedAtStage`,
|
|
119
|
+
or `missedAt` with a reason when the run settled before another boundary ran.
|
|
120
|
+
A paused run keeps the instruction for its next resume; a completed run
|
|
121
|
+
refuses it. A run owned by another process is refused rather than written,
|
|
122
|
+
because that owner's next save would erase it.
|
|
123
|
+
- `superagent_answer(run_id, attention_id?, answer, wait_seconds?)` answers one
|
|
124
|
+
structured attention item. Attention items now carry stable ids (`q1`, `q2`, …)
|
|
125
|
+
from a per-run counter; the answer is recorded on that item and the run
|
|
126
|
+
continues — an active run at its next boundary, a terminal run by reopening so
|
|
127
|
+
its aggregation runs again with the decision. A question retired by a resume
|
|
128
|
+
is archived to `attentionHistory` with `expiredAt`, so a late answer is
|
|
129
|
+
refused as expired instead of being matched to whatever is open now. Unknown
|
|
130
|
+
ids, already answered questions, and an answer whose text is another
|
|
131
|
+
question's option are refused with the id to use instead.
|
|
132
|
+
- **Per-worker workspace isolation.** `workspaceIsolation: true` gives every
|
|
133
|
+
worker its own git worktree on `superagent/<runId>-w<n>`, harvested into a
|
|
134
|
+
commit and merged back into the main checkout one at a time. A conflict aborts
|
|
135
|
+
the merge, keeps both branches, and is reported by status and the wake-up
|
|
136
|
+
message. When isolation is unavailable the run continues in the shared
|
|
137
|
+
directory and says so.
|
|
138
|
+
- **Dependency-aware plans.** The planner may give a subtask an id and a
|
|
139
|
+
`dependsOn` list. The dependent waits (`waiting`) instead of starting, and its
|
|
140
|
+
prompt carries the finished dependency's report; independent subtasks still
|
|
141
|
+
run in parallel. A dependency on an undeclared task is `skipped` with the
|
|
142
|
+
reason, a failure skips everything downstream instead of reporting it as done,
|
|
143
|
+
and a cycle rejects the plan before any subtask starts. Status, notifications,
|
|
144
|
+
and reports distinguish waiting from running.
|
|
145
|
+
- **Parallel structure in the panel and the report.** Runs alive at the same
|
|
146
|
+
moment share a numbered gutter in `watch`, so concurrent work is visible as a
|
|
147
|
+
group, and a waiting row leads with the run holding its slot and its queue
|
|
148
|
+
position. The HTML report lists each worker's retry chain and its share of the
|
|
149
|
+
run's wall clock next to the outcome.
|
|
150
|
+
- **Documentation site.** `docs/site/` holds a quickstart, a comparison page,
|
|
151
|
+
and a single-file landing page; nothing in `lib/` changed for it.
|
|
152
|
+
- **Per-run and per-subtask pricing.** `pricing` in the plugin config maps a
|
|
153
|
+
model to input/output prices per million tokens. Every run and every subtask
|
|
154
|
+
now carries a `cost` with micro-unit totals, and status, the panel, and the
|
|
155
|
+
report print it (`cost ¥0.3519+`). A model absent from the table is reported as
|
|
156
|
+
`unpriced` — never as `0.0000`, and a total that mixes priced and unpriced work
|
|
157
|
+
is printed with a trailing `+` so it cannot be read as complete.
|
|
158
|
+
- **Spend ledger.** `lib/ledger.js` keeps a durable per-run spend record, mounted
|
|
159
|
+
from settled runs so a restart backfills what no tool was watching, and
|
|
160
|
+
`superagent_status` reports the rolling spend and how many runs in the window
|
|
161
|
+
were only partially priced.
|
|
162
|
+
- **Workflow templates.** `superagent_template` lists, shows, and expands a
|
|
163
|
+
reusable goal template (built-ins plus files under the template directory), so
|
|
164
|
+
a common goal is expanded into its steps instead of being retyped. Templates
|
|
165
|
+
are optional: a template directory that cannot be created is logged and the
|
|
166
|
+
run tools keep working.
|
|
167
|
+
- **Install self-check.** `dsh-superagent-doctor` reports, without installing or
|
|
168
|
+
editing anything, whether every imported file is in the `files` whitelist and
|
|
169
|
+
on disk, whether the installed copy matches its published tarball, whether the
|
|
170
|
+
profile registers the bundle row, and whether all nine tools are registered.
|
|
171
|
+
Every failure prints the command that fixes it. `npm test` also asserts that
|
|
172
|
+
`npm pack --dry-run` ships every file reachable from the entry points
|
|
173
|
+
(`scripts/check-pack.mjs`), which is the check the 0.9.x tarball lacked.
|
|
174
|
+
- **Run archives.** `dsh-superagent-archive` (`bin/archive.js`) adds three
|
|
175
|
+
subcommands. `prune` lists terminal runs older than `--older-than <days>`,
|
|
176
|
+
optionally filtered by `--status`, with the files and bytes it would remove;
|
|
177
|
+
it deletes nothing unless `--confirm` is given, and it never touches a run
|
|
178
|
+
that is still active or whose `.lock` another process holds. `export` writes
|
|
179
|
+
one run as a self-contained directory (README with the report, `report.md`,
|
|
180
|
+
`run.json`, per-subtask markdown and JSON, guard records, `usage.json`, and a
|
|
181
|
+
manifest with a SHA-256 per file), as one JSON bundle, or as one offline HTML
|
|
182
|
+
page with the bundle embedded. `import` reads any of those back into the
|
|
183
|
+
store and refuses to overwrite a local run with the same id. `store.js` gained
|
|
184
|
+
`planPrune`, `buildArchive`, `exportRun`, and `importRun`; `prune` gained
|
|
185
|
+
`dryRun` and `statuses` and now returns the plan alongside `removed`. No
|
|
186
|
+
runtime dependency was added.
|
|
187
|
+
- New config: `queueOnFull`, `maxQueuedRuns`, `queueTimeoutMinutes`,
|
|
188
|
+
`workspaceIsolation`, `workspaceRoot`, `pricing`.
|
|
189
|
+
|
|
190
|
+
### Changed
|
|
191
|
+
|
|
192
|
+
- A single completed subtask still short-circuits aggregation, unless the run
|
|
193
|
+
has instructions or answers: then the aggregation prompt is built, because
|
|
194
|
+
otherwise reopening the run would return the old report and silently drop the
|
|
195
|
+
decision.
|
|
196
|
+
- `superagent_resume` reopens a `completed` run only when it holds an answer
|
|
197
|
+
that never reached a prompt; every other completed run keeps its existing
|
|
198
|
+
no-op behaviour. Without this, an answer that could not be applied (quota or
|
|
199
|
+
lock) would be stranded behind a "resume the run" instruction that did
|
|
200
|
+
nothing.
|
|
201
|
+
- Status prints the attention id, any recorded answer, and how many instructions
|
|
202
|
+
were applied, are waiting, or never took effect; notifications lead with the
|
|
203
|
+
question id so `superagent_answer` can be called directly.
|
|
204
|
+
|
|
205
|
+
|
|
206
|
+
|
|
207
|
+
|
|
208
|
+
|
|
209
|
+
## 0.10.0 — 2026-11-15
|
|
210
|
+
|
|
211
|
+
The dispatch, progress, delivery, and memory layer. No runtime dependencies were
|
|
212
|
+
added, and the package still ships one file per module from `lib/`.
|
|
213
|
+
|
|
214
|
+
### Added
|
|
215
|
+
|
|
216
|
+
- **Cross-run memory.** `<stateDir>/memory/<scope>.jsonl` keeps short notes the
|
|
217
|
+
aggregator writes (`decision`, `pitfall`, `next`, `dispatch`, `note`) plus the
|
|
218
|
+
six most recent runs derived from the run store. The digest is bounded
|
|
219
|
+
(12 notes / 1600 characters / 2600 characters total), compacted above 400
|
|
220
|
+
lines or 128 KB, and injected into the planner and aggregator prompts. A newer
|
|
221
|
+
note with the same key supersedes an older one, `forget` retracts it, and
|
|
222
|
+
notes older than 30 days expire. Secrets are redacted before storage; the
|
|
223
|
+
directory carries its own `.gitignore` and is never in the package.
|
|
224
|
+
- Added `superagent_memory` (`list`, `set`, `forget`). Existing tools keep their
|
|
225
|
+
arguments and output; this is the sixth tool.
|
|
226
|
+
- **Layered memory.** `rules` and `profile` scopes are read into planning and
|
|
227
|
+
aggregation ahead of the run scope; `superagent_memory` takes `scope`. A run
|
|
228
|
+
only writes its own scope, so a model cannot overwrite a project rule.
|
|
229
|
+
- **Duplicate dispatch is collapsed.** The same active goal (case and whitespace
|
|
230
|
+
ignored) returns the existing run id with `deduped: true`, under a per-goal
|
|
231
|
+
claim so two processes sharing a state directory cannot both start it. The
|
|
232
|
+
match is scoped to the owning session, so two sessions that ask for the same
|
|
233
|
+
generic goal still get their own runs. `allow_duplicate: true` on
|
|
234
|
+
`superagent_start` (or `allow_duplicate` on the control `start` body) opts
|
|
235
|
+
out. A `wait_seconds` call keeps the `deduped` marker in its result.
|
|
236
|
+
- **Transient backoff.** A rate-limit, quota, timeout, or 5xx failure now waits
|
|
237
|
+
2s, then 4s, then 8s before the next attempt instead of retrying immediately.
|
|
238
|
+
- **Recovery note.** A resumed run tells each worker what failed and instructs it
|
|
239
|
+
to inspect files, `git status`, `git log`, and remote state before repeating an
|
|
240
|
+
externally visible action.
|
|
241
|
+
- **Structured attention.** A worker or the aggregator may append one fenced
|
|
242
|
+
`attention` block. It is stripped from the report, stored as `attention.items`,
|
|
243
|
+
shown as `needs your decision:` in status, and leads the notification.
|
|
244
|
+
- **Attention text is bounded.** Question, options, and reason are collapsed to
|
|
245
|
+
one line and capped at 500 / 200 / 300 characters, so neither status nor a
|
|
246
|
+
notification can be given forged extra lines. An invalid attention block is
|
|
247
|
+
not stripped silently: its raw text stays in the report and status prints
|
|
248
|
+
`attention: <reason>`.
|
|
249
|
+
- **Milestones.** `plan_ready`, `running`, `aggregating`, and `settled` are
|
|
250
|
+
recorded with timestamps and shown in `superagent_status`; nothing is pushed.
|
|
251
|
+
- **Report file.** A terminal report is written atomically to
|
|
252
|
+
`<stateDir>/reports/<runId>.md` (directory carries `.gitignore: *`), recorded
|
|
253
|
+
in `run.artifacts`, and shown in status. `keepRuns` pruning removes it together
|
|
254
|
+
with the run record.
|
|
255
|
+
- **Quiet by default.** Only a delivered, failed, or partially finished run may
|
|
256
|
+
wake the session; a cancelled run never does, and progress is never pushed.
|
|
257
|
+
`maxNotificationsPerHour` (default 3) buffers extra runs into one summary.
|
|
258
|
+
- **Notification outbox.** A refused wake-up is retried with backoff
|
|
259
|
+
(`notifyRetries`, `notifyRetryDelayMs`) and messages for one session stay in
|
|
260
|
+
order. A buffered batch is put back into `pending` when the send fails, so a
|
|
261
|
+
later message still carries those runs; it is dropped only after three failed
|
|
262
|
+
deliveries.
|
|
263
|
+
- **Self-heal before reporting.** A subtask failure that looks transient
|
|
264
|
+
(quota, rate limit, timeout, 5xx) gets `autoResumeRounds` (default 1) more
|
|
265
|
+
rounds, optionally on `childModelFallback`. The failure message now says what
|
|
266
|
+
was tried, and `self-heal:` appears in `superagent_status`.
|
|
267
|
+
- **`memoryMaxEntries` is a compaction trigger.** Exceeding it, 400 lines, or
|
|
268
|
+
128 KB compacts the notebook on the next write.
|
|
269
|
+
- New config: `memory`, `memoryScope`, `memoryMaxEntries`, `memoryMaxAgeDays`,
|
|
270
|
+
`memoryDigestEntries`, `memoryDigestChars`, `maxNotificationsPerHour`,
|
|
271
|
+
`notifyRetries`, `notifyRetryDelayMs`, `autoResumeRounds`,
|
|
272
|
+
`childModelFallback`.
|
|
273
|
+
|
|
274
|
+
### Changed
|
|
275
|
+
|
|
276
|
+
- **A resume is never refused by the concurrency quota.** A `superagent_resume`
|
|
277
|
+
only finishes work that already exists, so it is not counted against
|
|
278
|
+
`maxActiveRuns` and works even when the cap is full; only `superagent_start`
|
|
279
|
+
creates new spend and is subject to the cap. The resume is still
|
|
280
|
+
session-scoped: it may only be issued by the session named in the run's
|
|
281
|
+
`parentSessionId`. `force` (`force` or `override_quota` over HTTP) is still
|
|
282
|
+
accepted so an existing caller keeps working, but it no longer changes
|
|
283
|
+
anything.
|
|
284
|
+
- `superagent_status` gained the milestone line, the attention line, and the
|
|
285
|
+
report file path; `superagent_start` gained `allow_duplicate`; the control
|
|
286
|
+
`start` body gained `allow_duplicate`, and `resume` still accepts `force`.
|
|
287
|
+
|
|
288
|
+
### Fixed
|
|
289
|
+
|
|
290
|
+
- Memory, notification, and report-artifact writes are fail-open or retried, so
|
|
291
|
+
a transient filesystem error cannot turn a finished run into a failed one; a
|
|
292
|
+
report that is already written stays on disk and is listed in status.
|
|
293
|
+
- Test suite: 168 tests, 0 failures; `npm test` and `node --test` discovery
|
|
294
|
+
agree.
|
|
295
|
+
|
|
98
296
|
## 0.9.3 — 2026-11-14
|
|
99
297
|
|
|
100
298
|
Documentation only; no code change.
|