Skip to content

Registry workflows: a person approves what is sent, before it is sent - #69

Merged
0design merged 7 commits into
mainfrom
claude/workflows-real-1101
Nov 3, 2026
Merged

0design merged 7 commits into
mainfrom
claude/workflows-real-1101

Conversation

@0design

@0design 0design commented Nov 1, 2026

Copy link
Copy Markdown
Owner

Registry manifests only (registry/, test/, scripts/); the Core runtime does not change.

  • content-feed, feed-fanout, release-watch, webhook-relay, json-digest, brand-mentions: a human approval-gate (bind: sha256) directly before the send; the send carries the gate's subject, so what is sent is what was approved. feed-fanout gates each lane.
  • json-digest and webhook-relay send the approved object in Core's envelope, so titles and data stay JSON.
  • brand-mentions: the brand is the required input brand; the schedule trigger is removed (a scheduled run has no inputs).
  • price-watch: escalateOn: none — an ordinary day ends at the check without asking anyone; the alert sends the approved rate.
  • Every changed workflow sets settings.model; versions 1.1.0 → 1.2.0.
  • test/registry-real-workflows.test.mjs: offline runs with a stubbed fetch — nothing is sent before the gate, and the body after approval equals the approved subject.

The stage candidate 2026.11.01-registry.39 is built from the same manifests on the Core.49 base (branch claude/workflows-real-1101-core49, archive 631ba858 unchanged). Its price-watch sends the verdict reason, which is the gate subject on Core.49. This PR is the main-line version for the next Registry release; releaseVersion is left as on main.

Draft: not for merge until the next Registry release is planned.

🤖 Generated with Claude Code

0design and others added 7 commits November 1, 2026 21:29
… and do what they say

Candidate 2026.11.01-registry.39 on the Core.49 base (6814eb7). Only registry/, test/ and
scripts/ change, so `npm pack --ignore-scripts` still gives the pinned Core archive
631ba858051dcf08b4dbd204d12c50ed07c5b1901fadf5cdc8d9043905778394.

- content-feed, feed-fanout, release-watch, webhook-relay, json-digest, brand-mentions: a human
  approval-gate (bind: sha256) directly before the send; the send carries the gate's subject,
  so what is sent is what was approved. feed-fanout gates each lane.
- json-digest and webhook-relay send the approved object in Core's envelope: titles and data
  stay JSON instead of text inside a string.
- brand-mentions: the brand is the required input `brand`; the schedule trigger is removed
  (a scheduled run has no inputs and would fail every day).
- price-watch: escalateOn none, so an ordinary day ends at the check without asking anyone;
  the reason states the rate; the alert sends the approved reason.
- content-factory: model set in the manifest. The approval subject is still the check's
  verdict on Core.49 (needs the next Core release); the brief now says so.
- Every changed workflow sets settings.model (no unconfigured model); versions 1.1.0 -> 1.2.0.
- test/registry-real-workflows.test.mjs: offline runs (stubbed fetch) prove nothing is sent
  before the gate and the post-approval body equals the approved subject; all 11 fail on the
  registry.38 bytes. core37/registry-build/export tests follow webhook-relay now being gated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…man gate approves the content past the verdict)

On Core.51 a human gate's subject is the content past agent verdicts, so the alert reads the
rate from the approved subject. The Core.49 stage candidate (branch -core49) keeps the
verdict reason, which is its subject there. content-factory's brief drops the Core.49 note.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nd webhook-relay (the visual editor's insert keeps them in place)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ence) once, then sends each entry from it

On Core.49 lane rows share step ids, so a gate inside the second lane took the previous
lane's delivery as its subject and its sentence never reached the receiver; the title and
link were sent without being shown either. One model call now returns the list, one human
gate approves it exactly, and the fan-out runs over the approved list. price-watch's brief
names the no-key case.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n words in the editor)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t an engine guarantee); price-watch's no-key limit is true for this Core

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ate component

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant