Skip to content

Jev workflows: agent steps and Jev decisions, and Let Jev choose - #46

Open
MasterYoav wants to merge 12 commits into
jev-ai-settingsfrom
jev-workflows
Open

MasterYoav wants to merge 12 commits into
jev-ai-settingsfrom
jev-workflows

Conversation

@MasterYoav

Copy link
Copy Markdown
Owner

Jev workflows (spec), stacked on the AI settings PR.

  • A workflow: agent steps (a fixed CLI/model/effort, or Jev picks the cheapest capable one) and Jev steps — yes/no gate, route, score — whose branches continue, go to a step, or stop. Each step runs at most 5 times per run.
  • A run is one chat: agent steps are turns, Jev steps are notices ("Gate: … — yes (0.91) → Fix it"). Jev sees the input and earlier outputs, trimmed from the oldest end. Jev unreachable: the run stops and says why; a Jev-picks step falls back to the workflow's agent.
  • Run now, Run with…, or a schedule (same clock and background helper as automations); runs claimed so they never double.
  • Let Jev choose in the composer's agent menu: Jev picks agent, model and effort per message.
  • Store v7 (workflows, chats.workflow_id, chats.jev_chooses).
  • Workflows page and editor (palette, drag to reorder, branch pickers); sidebar row only while the toggle is on; without a key the page says so.
  • Phone: Workflows in Places (list + Run now) while the Mac has it on.
  • Docs: architecture, Mac app, UI spec, iPhone and iPad, roadmap, changelog.

Verified: 562 package tests (89 new); iOS builds; Workflows page, editor and Settings › AI checked on screen (fixed a clipped title, the gate's threshold row and the key field's label). Not verified: a real run against Jev (needs your key).

Agent steps (fixed brain or Jev picks, Can edit by default) and Jev steps
(gate, route, score) with go-to targets; the state Jev is asked about, trimmed
from the oldest end; one call choosing the cheapest capable CLI, model and
effort.
Steps are JSON, read one by one so a step from a newer xBot loses only itself;
runs are claimed like automations'; v6 → v7 migration tested.
A run is one chat tagged with its workflow: agent steps are turns (the session
carries context), Jev steps notices ("Gate: … — yes (0.91) → Fix it"). Gates,
routes, scores and go-tos; each step at most 5 times per run; fail closed when
Jev can't be reached (a Jev-picks step falls back to the workflow's brain);
claims and schedules like automations; nothing runs while the toggle is off.
Only while Jev Workflows is on; picking an agent by hand turns it off; if Jev
can't be reached the chat's own agent answers and the reply says why. Workflows
run from the automations' clock and through the helper while xBot is closed.
The page and its row show only while Jev Workflows is on; without a key the
page says so with a button to Settings › AI. Steps are dragged in from a
palette (agent step, gate, route, score) and reordered by dragging; a Jev
step's branches are pickers. Run now, Run with…, and the runs.
listWorkflows / runWorkflow (with an optional input) answered through the same
Workspace methods; workflowsAvailable on connecting and whenever the toggle
changes; workflow runs stay out of the phone's chat list.
Shown only while the Mac has Jev Workflows on (workflowsAvailable); each
workflow's schedule, steps and recent runs; Run now opens the run.
Architecture (Jev and workflows, schema v7), the Mac app, the UI spec (the
Workflows page, Let Jev choose, the phone's Workflows), the phone, the roadmap,
the changelog and the spec table; the spec's storage line as built.

This branch has not been deployed

No deployments
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