byollm 0.1.0-alpha.7 → 0.1.0-alpha.71

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/README.md CHANGED
@@ -1,5 +1,5 @@
1
1
  > [!WARNING]
2
- > **Alpha (`0.1.0-alpha.7`) — under active development. Don't use this yet.**
2
+ > **Alpha (`0.1.0-alpha.71`) — under active development. Don't use this yet.**
3
3
  >
4
4
  > Install it deliberately: `npx byollm@alpha`, or `npm install byollm@alpha`.
5
5
  >
@@ -11,14 +11,116 @@
11
11
  > npm assigns `latest` on a first publish and won't let it be removed, so a
12
12
  > bare install resolves here too. This notice is the only guard — deliberately
13
13
  > not an npm deprecation, which would read as *abandoned* rather than *early*.
14
- > Ask for `@alpha` explicitly so your lockfile records that you meant to.
14
+ > Ask for `@alpha` explicitly so your lockfile records that you meant to.>
15
+ > **`alpha.15` is a breaking wire change, and it breaks daemons and relays —
16
+ > not app authors.** If you call `app.enqueue(...)` and read results, nothing
17
+ > in your code changes. If you run a daemon or an upstream, every package must
18
+ > move together: a mixed pair refuses on both sides, because both ends parse
19
+ > `.strict()`.
20
+ >
21
+ > What moved, all of it reconciling the frozen `byollm_009` with its code:
22
+ > `JobStub` gains `site` (the site's identity key id) and loses
23
+ > `audienceAllow`; `ResultRequest` gains `leaseId`; `HeartbeatResponse` loses
24
+ > `leases`, which nothing read; `WireErrorCode` gains `not-ready`,
25
+ > `clock-skew` and `forbidden`, and `403` is `forbidden` rather than
26
+ > `unauthorized`. `RESULT_PROVENANCE` is superseded by
27
+ > `PROVENANCE_NAMES_DEVICE`. See `byollm_009` Amendment A.>
28
+ > **`alpha.16` is a breaking wire change — daemons and relays again, not app
29
+ > authors.** `app.enqueue(...)` and reading results are unchanged. All five
30
+ > packages move together: both ends parse `.strict()`, so a mixed pair
31
+ > refuses.
32
+ >
33
+ > What moved, all of it Tier 2 of `cloud_008`: `model`, `backendClass` and
34
+ > `durationMs` come off `ResultRequest` and are sealed **inside** the result
35
+ > envelope as `SealedOutcome = { outcome, ran }` — so a daemon can no longer
36
+ > declare a model it did not sign, and a relay carries neither.
37
+ > `HeartbeatResponse` loses `leases` (nothing read it) and now reports real
38
+ > cancellations instead of an empty list. `WireErrorCode` gains `forbidden`
39
+ > for 403, leaving `unauthorized` at exactly 401. The relay gained a
40
+ > site-plane `cancel` endpoint, honours `stub.deadlineAt`, honours
41
+ > `stub.audience`, and remembers a refusal.>
42
+ > **`alpha.17` is additive** — no wire change. It exports `ReleaseReason`,
43
+ > which `RoutingStore.releaseLeases` names and the package did not export, so
44
+ > the interface was unimplementable outside this repo.>
45
+ > **`alpha.18` is a breaking wire change — daemons and relays, not app
46
+ > authors.** `app.enqueue(...)` and reading results are unchanged. All five
47
+ > packages move together.
48
+ >
49
+ > The **bearer token is gone**: off `PairPollResponse`, off the runner row,
50
+ > off the daemon's pairings file, out of the adapter's schema. It was minted,
51
+ > hashed and stored on two disks and never sent, looked up or compared —
52
+ > `REQUESTS_SIGNED_NOT_BEARER` was enforced by signatures the whole time. If
53
+ > you run the Supabase adapter, apply
54
+ > `20260819000000_drop_runner_token.sql`; `byollm_approve_pairing` now takes
55
+ > one argument. A pairings file written by an older daemon still loads.
56
+ >
57
+ > `model`, `backendClass` and `durationMs` moved **inside** the sealed result
58
+ > (`SealedOutcome = { outcome, ran }`), so a daemon cannot declare a model it
59
+ > did not sign and a relay carries none of them. Writing a `RoutingStore`?
60
+ > `releaseLeases` takes an optional `reason` and `complete` requires
61
+ > `leaseId`, and **an implementation that ignores either still typechecks** —
62
+ > run the store contract tests.>
63
+ > **`alpha.19` is additive on the wire and a behaviour change in every
64
+ > store.** `ResultResponse` gains an optional `duplicate`. Nothing is removed,
65
+ > so an older daemon keeps working — but the *order* two rules are checked in
66
+ > has changed, and a `RoutingStore` implementation must change with it.
67
+ >
68
+ > `complete` now checks **terminal state before holder**, scoped to the device
69
+ > that finished the job: a replay from that device is answered `duplicate:
70
+ > true` with a 2xx, and anyone else gets exactly the refusal they would get
71
+ > for a job that is not terminal. Previously `RESULT_IDEMPOTENT` held only
72
+ > because the lease is nulled on success, so the holder check tripped first —
73
+ > deleting the idempotency branch failed no test. Run the store contract
74
+ > tests; the compiler cannot see this.
15
75
  >
16
76
  > **`alpha.3` is a breaking change.** A config naming `openai-http` with a
17
- > remote base URL and an offer scope wider than `self` is narrowed to `self`
77
+ > remote base URL and an offer scope wider than `private` is narrowed to `private`
18
78
  > until you acknowledge the spend and set a daily ceiling:
19
79
  > `byollm offer <backend> public --cap <cents>`. Local base URLs are
20
80
  > unaffected. `byollm backends` shows the cost class per route.
21
81
 
82
+ <!-- release-note 0.1.0-alpha.21 -->
83
+ > [!NOTE]
84
+ > **`0.1.0-alpha.20` is not a complete release — do not pin it.** Four
85
+ > packages published and `@byollm/server` did not: a Sigstore
86
+ > transparency-log 409 on its provenance attestation. The workflow's
87
+ > "already published" guard correctly refuses to resume a partial publish,
88
+ > so `0.1.0-alpha.71` is that release, whole.
89
+ >
90
+ > If you run the Supabase adapter, `alpha.21` needs
91
+ > `20260819010000_completed_by_lease_id.sql`: alpha.19 shipped §3.6's
92
+ > ordering without the column it stores the grant in.
93
+
94
+ <!-- release-note 0.1.0-alpha.40 -->
95
+ **`byollm install` — stop keeping a terminal open.** The daemon can now run
96
+ under your computer's own supervisor and restart itself if it stops: a launchd
97
+ agent on macOS, a `systemd --user` unit on Linux, a logon task on Windows. All
98
+ user-level — no root, no system directories, and `byollm uninstall` takes it
99
+ away. `byollm status` gained a line saying whether it is actually supervised
100
+ right now, including the state that matters most: installed but not running,
101
+ which looks fine from an app's dashboard and serves nothing.
102
+
103
+ If you are running via `npx`, install properly first (`npm install -g
104
+ byollm@alpha`) — `install` refuses to supervise a copy in npx's cache, because
105
+ npm deletes that directory and the service would fail at some later boot.
106
+
107
+ <!-- release-note 0.1.0-alpha.41 -->
108
+ **`onNoRunner` takes a string.** Your fallback answer is your own value, not
109
+ wire data, and handing back a whole result record for it was ceremony — the
110
+ README's own example got the shape wrong, which is how this was found.
111
+
112
+ ```ts
113
+ const { outcome, fallback } = await job.result({
114
+ onNoRunner: () => runOnHostedModel(transcript),
115
+ });
116
+ ```
117
+
118
+ Whatever you return, `result()` labels it `fallback: true` — the stamp is
119
+ applied by the wait, not taken from you, so an answer that did not run on
120
+ somebody's device cannot be reported as though it did (`FALLBACK_LABELED`).
121
+ Both delivery channels do it, polling and Supabase Realtime. Records still
122
+ work; they just get labelled too.
123
+
22
124
  # `byollm`
23
125
 
24
126
  What end users run. Connects **outbound** to an app you trust, claims only the
@@ -79,12 +181,48 @@ password and never accepts a pasted secret.
79
181
  }
80
182
  ```
81
183
 
82
- A job's `kind` selects a route **you defined**. A job can never name a model, a
83
- URL, a path or a flag — there is no field on the wire for any of them.
184
+ A job's `kind` selects a route **you defined**, and a job can never name a
185
+ model, a URL, a path or a flag — there is no field on the wire for any of
186
+ them. It cannot name one of your services either: a site declares what it
187
+ *needs*, you decide which of your services answers that need, and the job
188
+ carries the site's word for the need rather than yours for the answer.
189
+
190
+ `byollm services` shows what is configured, what is healthy, what each one
191
+ answers, and which is your own default. A service that is down is never
192
+ advertised, so you never get work you cannot run.
193
+
194
+ ## Keep it running
195
+
196
+ By default `byollm connect` and `byollm run` hold a terminal — close the
197
+ window and the device stops serving. Nothing breaks (your pairings live in
198
+ `~/.byollm/pairings.json` and survive), but the device goes quiet without
199
+ telling anyone, and if it is on a team's roster, the first person to notice is
200
+ a teammate whose job did not run.
201
+
202
+ ```bash
203
+ byollm install # keep running in the background, and restart if it stops
204
+ byollm status # says whether it is actually supervised right now
205
+ byollm uninstall
206
+ ```
207
+
208
+ It installs at the user level on every platform — a launchd `LaunchAgent` on
209
+ macOS, a `systemd --user` unit on Linux, a logon task on Windows. No root, no
210
+ system directories: it runs as you, it stops when you say so, and you can read
211
+ every file it wrote. Output goes to `~/.byollm/service.log` on all three.
212
+
213
+ Two things worth knowing:
84
214
 
85
- `byollm backends` shows what is configured, what is healthy, and what is
86
- therefore advertised. A backend that is down is never advertised, so you never
87
- get work you cannot run.
215
+ - **Install `byollm` properly first.** `byollm install` refuses to supervise a
216
+ copy running from `npx`'s cache, because npm deletes that directory without
217
+ warning and the service would stop working at some later boot with nothing
218
+ to show for it. `npm install -g byollm@alpha` first.
219
+ - **On Linux, `systemd --user` stops when you log out** unless lingering is
220
+ enabled. `byollm install` prints the one command for that rather than
221
+ running it — it changes something outside your session, so it is your call.
222
+
223
+ `byollm status` reports three states, not two: running under supervision,
224
+ *installed but not running* (the one that looks fine from an app's dashboard
225
+ and serves nothing), and not installed at all.
88
226
 
89
227
  ## The trust surface
90
228
 
@@ -92,32 +230,69 @@ The meter is the product, and it gets the same care as the loop.
92
230
 
93
231
  ```bash
94
232
  byollm status # what's connected, what's running, what you've done for others
233
+ byollm sites # which sites this device serves, and which are waiting on you
234
+ byollm approve <site> # say yes to a site that asked
95
235
  byollm log # every prompt that has ever run here
96
236
  byollm log --full # the whole text, not the first line
97
237
  byollm pause # stop claiming work
98
238
  byollm resume
99
239
  ```
100
240
 
241
+ ### A site cannot add itself
242
+
243
+ Pairing is with an *app* — a hub, a relay, your own server — and one pairing
244
+ can cover several sites. Which sites arrives on the heartbeat, from the same
245
+ party that routes the work.
246
+
247
+ So a site that turns up after pairing **waits**. It is listed by
248
+ `byollm sites` with its fingerprint, nothing is claimed for it, and it starts
249
+ being served the moment you run `byollm approve <site>`. Compare the
250
+ fingerprint against what the site itself shows you before you do.
251
+
252
+ The reason is narrow and worth stating: the daemon pins each site's keys so
253
+ that the party routing a job cannot choose which key signed it. If that party
254
+ could also *add* a site, it could generate a keypair, announce it, sign its
255
+ own work with it, and every pin check downstream would pass — because the
256
+ list they check against is the thing it wrote. Approving is the one step it
257
+ cannot perform for you.
258
+
259
+ A key that moves under a site you already approved is refused rather than
260
+ replaced, for the life of the pairing — including when the site leaves the
261
+ list and comes back. Rotation is an explicit path, not a silent swap.
262
+
101
263
  Every prompt is appended to `~/.byollm/ingress.log` **before** it executes, so
102
- a job that wedges the machine still leaves a record of what it was. The file is
264
+ a job that wedges the computer still leaves a record of what it was. The file is
103
265
  JSONL, `0600`, and yours to read, grep and delete.
104
266
 
105
- ## Lending your machine to other people
267
+ ## Lending your computer to other people
106
268
 
107
269
  Off by default. A fresh daemon runs your work and nobody else's.
108
270
 
109
271
  ```bash
110
- byollm allow https://your-app.com alice # asks you to confirm, in plain words
111
- byollm allow --list # everyone who can use this machine
112
- byollm offer openai public --cap 250 # share a paid backend, with a ceiling
113
- byollm disallow https://your-app.com alice
272
+ byollm offer qwen team --cap 250 # share a service with your team
273
+ byollm offer qwen private # back to your work only
114
274
  ```
115
275
 
116
- The allowlist is **yours**, checked locally, keyed by `(app, user)`. An app
117
- saying "this runner is allowed" is not enough — your daemon decides. Community
118
- jobs are additionally rate-limited, capped daily, given a tighter resource
119
- budget, and their prompts are reduced to hashes after 7 days so you are not
120
- holding strangers' content indefinitely.
276
+ Who "your team" is lives in the dashboard, not on this machine. Add somebody
277
+ to your roster there and their next job can run here; remove them and their
278
+ next claim fails, including work already queued.
279
+
280
+ `byollm allow` and `byollm disallow` are gone. A device no longer keeps its own
281
+ list of who may use it — that list had to be kept in step with the roster by
282
+ hand, and two lists that must agree are one list and a bug waiting. Both
283
+ commands leave a tombstone pointing at where the decision lives now.
284
+
285
+ What replaced it is stronger than a list. Every teammate's job arrives carrying
286
+ a **grant**: short-lived, single-use, signed by the control plane your device
287
+ pinned when it paired, naming the site, the person, the job and the service to
288
+ run. Your daemon checks that signature itself before it runs anything, so an
289
+ app saying "this runner is allowed" is still not enough — and neither is a
290
+ relay saying it. Jobs from other people are additionally rate-limited, capped
291
+ daily, given a tighter resource budget, and their prompts are reduced to hashes
292
+ after 7 days so you are not holding anybody else's content indefinitely.
293
+
294
+ `public` is not a scope. It was one until 2026-08-26, and `byollm offer <service>
295
+ public` now refuses by name rather than silently doing something else.
121
296
 
122
297
  **Your subscription-backed models are never part of this.** `claude-cli` is
123
298
  locked to your own work — a protocol rule, not a setting you can change.
package/dist/bin.js CHANGED
@@ -1,7 +1,7 @@
1
1
  #!/usr/bin/env node
2
2
  import {
3
3
  main
4
- } from "./chunk-3E2PGDHL.js";
4
+ } from "./chunk-YJ3OSRLT.js";
5
5
 
6
6
  // src/bin.ts
7
7
  process.exitCode = await main(process.argv.slice(2));