Skip to content

demo: AI Assistants - #26076

Closed
thykel wants to merge 86 commits into
devfrom
ai_assistant_demo
Closed

demo: AI Assistants#26076
thykel wants to merge 86 commits into
devfrom
ai_assistant_demo

Conversation

@thykel

@thykel thykel commented Oct 15, 2026

Copy link
Copy Markdown
Contributor

demo-only offshoot of #25976

thykel and others added 30 commits October 12, 2026 22:23
@thykel @claude
Add Subject and Description automation actions and a strict mode
e31e8a7 
The Automations update service drops actions that fail validation and
applies the rest. Callers that need all-or-nothing pass strict: true.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the ai_assistance flag, permission and journal cause
114e077 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add AI assistant, skill, run and finding models
758fe29 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the skill registry and the health check tree rules
36c3c5a 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the health check relation and staleness rules
18fcca9 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the health check rules for linked pull requests
d27e52a 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the run service, the run controllers and the run dialog
bba4f13 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the AI Assistance menu to the work package toolbars
9a75fd8 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the AI review to the health check
4975f52 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Apply proposed changes from AI findings
44fc072 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add admin pages for AI assistants and their skills
1c95afe 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Document the AI Assistance concept and the review split
038443d 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add the AI Assistance chat on work packages
9feb8f1 
One private conversation per user and work package. Answers can propose
changes, which the user applies like the fixes of a finding. The prompt
is an instance setting, not code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add custom prompt presets, a comment-only scope and a copy button
897403a 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Use "AI review" in the interface
57783c5 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Rename the AI judgement to AI review, and move shared proposal classe…
5f7dc95 
…s to AI::Assistance

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Clean up AI Assistance after review
6cf7eef 
- Seed QA assistance with Check health, Chat and Generate test cases, with prompts
- Share the LLM call, the event log, the user locale, polling and the feature guard
- Move lost references, comment-only and the proposal state into AI::Proposal
- One list of changeable attributes (ChangeOptions::ATTRIBUTE_KEYS)
- Run Comment after Notify, so Notify cannot replace a proposed comment
- Neutral apply errors for chats, and a translated review state in Details
- Remove unused code, columns (uuid, error_message) and the failed run status
- Use Primer utilities instead of inline styles where they exist
- Share the admin form views, validate the skill type in the model
- Update the concept doc and outdated specs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Show a spinner while the AI review runs
e3a078f 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Hide AI Assistance without an LLM, and do not call a partial check co…
6943bff 
…nsistent

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Call the health check "Refresh work package" in the interface
a6c7872 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Send the project name and related meeting agendas to the model
3e68882 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Document the OpenProject work for a coding agent integration
8250225 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Fix two AI Assistance specs after the cleanup
61159ef 
The run service spec created a skill with an unknown key, which the skill_key
validation now refuses. The run spec called terminal?, which the cleanup removed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Add personal AI assistants in My account
58eda62 
Users with the global permission manage_own_ai_assistants create their own
assistants and skills at My account > AI assistants. A personal assistant shows
only in its owner's menu, after the global ones, and uses the default model of
the LLM connection. It is deleted with its owner.

The admin list shows every assistant with its owner and filters by owner. The
admin and the My account controllers share their actions and views.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Delete AI Assistance records with their user
dd69bc4 
Add the associations on User that Danger asks for, as for AI text transform
runs, and cover the deletion in the Principals::DeleteJob integration spec.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Accept the pull_requests area in the AI review
462f382 
The rules report pull request findings in that area, but the review schema
did not offer it to the model.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Simplify AI Assistance after the second review
e340620 
- Render the menu items with the button, and remove the items endpoint.
- Remove AI::Assistance::Locale and ModelChoice: User.execute_as sets the
  locale, and the model choice is AI::Assistant#effective_model_id.
- Rules return attribute hashes instead of a Finding value object.
- Comments use journals.internal_visible, as the jobs run as the user.
- Share the skill lookup and the dialog responses between the controllers,
  and show apply errors of findings as a flash, as in chats.
- One pages helper for the admin and My account controllers.
- Remove finished_at on runs, unused proposal helpers and skill type hooks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Let a group own an AI assistant
8971222 
The owner of an assistant is now a principal: none for a global assistant, a
user for a personal one, or a group, whose direct members see it in their
menu. An admin chooses the group when creating the assistant, and the list
filters by Global, Groups and Personal. Owned assistants are deleted with
their principal.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Seed a sample group assistant in the development data
dbcdb94 
"Comms assistance" belongs to the Marketing & Communications department and
has two presets, Draft an announcement and Check the wording, and a chat.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Seed one sample assistant for each owner in the development data
2330d7d 
PM assistant is global, COMMS assistant belongs to the Marketing &
Communications department, and Tom's assistant is a personal assistant of the
admin user. The development seeder replaces the group-only one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
thykel and others added 24 commits October 15, 2026 13:25
@thykel @claude
Fix the ERB layout of the AI skill form
17940c0 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Show the typed AI skill prompt in the full prompt preview, and rename…
1a6fa5a 
… the prompts

The full system prompt preview on the skill page now changes while the
admin types the skill prompt (ai-assistance--prompt-preview), as it
already follows "Use OpenProject MCP tools".

"System prompt" named only one part of the system prompt. The skill field
is now "Skill prompt", and the assistant's "Instructions" field is
"Assistant prompt". The column names stay.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Move the AI Assistance button into the app header
578d590 
The button is now an icon in the app header, next to the other header
buttons, instead of a labelled button in the work package toolbars. It
follows the work package in the URL: a split view (also on boards and
calendars) or the full view, and it hides on other pages.

It uses the extension points of a module, so it can move into a plugin
as it is: a view hook (view_layouts_base_top_menu), a dynamic Stimulus
controller, and Angular services through the plugin context for the
refresh after Apply. The Angular toolbar slot is removed, and the
work package toolbars are back to the dev version.

The menu answers an id that the user cannot see with an empty frame.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Link AI journal entries to their chat
0e87f3a 
The journal entry of an applied proposal now names the user who
applied it instead of "the user". For the owner of the chat, it also
links to the chat, which opens in the chat dialog. Chats stay private:
other users see the text without the link.

A chat whose skill is no longer offered (inactive, or its conditions
no longer match) now opens read-only, from the link or from History.
It has no question field and no Apply or Dismiss buttons.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Group the AI Assistance menu by assistant
3b16049 
Each assistant gets a plain heading in the menu, instead of a label
with its name on every skill. Personal assistants come first, then
group assistants, then global ones.

The menu has a fixed size (480 px wide, up to 600 px high), so its
position no longer changes with the length of the skill texts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Give each AI skill an icon, and drop the assistant icon
982a2b4 
A skill has an icon, chosen from a list of Octicons. Without a choice,
it uses the icon of its type, as before. The menu and the history show
it.

The skill form shows the icon as a button joined to the Label field.
The button opens a menu with each icon and its name, and shows the
chosen icon (the ai-assistance--icon-picker Stimulus controller).

The assistant icon was only shown in the assistant list. It is
removed, with a migration, and the list shows the owner instead: a
person, a group or everyone. The seeded skills get icons.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Separate History from the skills in the AI Assistance menu
9c28369 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Do not start a second AI answer while one is running
8d22432 
- A unique index allows one pending answer per chat, also for two
  requests at the same moment. The service answers the index error
  with the usual "wait for the answer" message.
- Starting a skill while the user's last start of it on the work
  package is still answering opens that chat. An advisory lock per
  user, work package and skill covers two clicks at the same moment.
- The reply job has a GoodJob concurrency key per answer. A run after
  a crashed worker still runs.
- An answer stores its result only while it is still pending. A
  result after the time limit is dropped and logged as late.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Offer AI Assistance on every page
23a6f75 
The header button now shows on every page for users who have "Use AI
Assistance" in a project. Its menu follows the place in the URL: the
open work package, else the project of the page, else no place.

- A skill is offered on "Work packages" (as before) or "Everywhere".
  On a work package, every matching skill is offered. On other pages,
  only Everywhere skills, without proposals. Their context is the user
  and, on a project page, the project. An Everywhere skill cannot be
  limited to types or statuses, and a project or role condition hides
  it off matching projects.
- The permission is "Use AI Assistance" in the project of the place,
  or in any project on pages without one.
- A chat belongs to a work package, a project, or neither
  (AI::Assistance::Place). History lists all the user's chats with
  their place, and each opens with its own context and permission.
- The routes move from /work_packages/:id/ai_assistance to
  /ai_assistance, with the place as a query parameter.
- The seeded Chat skill is offered everywhere.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Show AI assistants as sections with their skills
7783fb9 
The assistant list shows one section per assistant, laid out like the
project attributes: global assistants first, then group and personal
ones. The header names the owner with an icon (globe, group, user) and
links to the group or user. Each skill is a row with its icon, a link,
where it is offered, and a menu to edit or delete it.

The owner filter and the owner column are gone. Skills saved before
"Offered on" existed read it as "Work packages".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Link AI assistant owners to their public profiles
48aa972 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Copy AI assistants and skills
d018cc2 
An assistant is copied with all its skills, for the same owner. A
skill is copied into its own assistant, after its last skill, with
"(copy)" in its label. A copy is inactive, so users see no duplicate
in the menu, and it opens on its edit page. A second copy of an
assistant gets a numbered name, because names are unique per owner.

The copy is in the menus of the assistant list, and on the assistant
page, next to Delete and for each skill. It works in Administration
and in My account.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Rework the AI Assistance seed data
b2beab1 
- The OP assistant (basic data) offers only Chat, everywhere.
- The admin's personal assistant is "My assistant", with Refresh work
  package.
- The COMMS assistant offers "Assign to least busy dev": it proposes
  the team developer with the fewest open work packages as the
  assignee.
- The development data adds Kabiru, Bruno, Judith and Ihor Developer
  to Web & IT and to the COMMS team, names the admin "Tom Admin" (also
  a team member), and adds COMMS to the demo project, so that its
  members can be assignees.
- The MCP tool search_users takes a group, by id or name, and finds
  the members of a group that the user can see. The skill uses it.
- Skills can use the "people" icon.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Rename the seeded OP assistant to OPAssistant
0e67caa 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Count work per person, and seed the COMMS developers' workloads
6c99add 
- count_work_packages and search_work_packages take a list of
  assignee ids. Grouped by assignee, the count lists every requested
  person, also with no work packages, and gives each group its id. A
  model no longer calls again to find people who are missing because
  they have nothing.
- "Assign to the least busy developer" (renamed from "Assign to least
  busy dev") counts the team's open work in one call.
- The development data gives the COMMS developers open work packages
  in the demo project (Kabiru 3, Ihor 2, Bruno 1, Judith 0), so the
  skill has one answer.
- The Refresh work package description no longer names pull requests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Show the owner icon of each assistant in the AI Assistance menu
7b0b409 
The section heading of each assistant shows its owner kind, as in the
assistant list: a globe, a group or a person.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Log AI tool results and the model's text, and stop repeats early
3ca7731 
Each tool call event now keeps the full result, and Details shows it
next to the arguments. The model's text on a turn that calls tools is
logged as a "thought" event, so a loop can be explained. After one
identical call, the next one gets "answer now" instead of three more
rounds.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Put the owner icon after the assistant name in the AI Assistance menu
af809f1 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Log the model's reasoning, and answer now after a repeated tool call
924c090 
The model's reasoning on a tool-calling turn comes in a separate
field, not in its text, so the "thought" event now logs it too.

In about half the runs, the model repeats its last tool call before it
answers, with and without the answer schema. Only a result that says
"answer now" ends that, so the first repeat now says it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Dim the owner icon in the AI Assistance menu headings
f15dbe9 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Dim the owner icon with opacity
42e3cec 
In OpenProject's theme, color-fg-subtle is as dark as the heading
text, so the icon did not look dimmed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Ask AI answers to name people, not guess pronouns
e8f316e 
A pronoun guessed from a name can be wrong, so the output rules of
every skill ask the model to refer to people by name.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Merge remote-tracking branch 'origin/dev' into ai-assistance/health-c…
52f389f 
…heck

dev adds a migration with the version 20261012120000, which the AI
assistants migration had. The AI one is now 20261012120001.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@thykel @claude
Seed the AI Assistance demo on every fresh install of this branch
8f294ad 
Temporary, for previews, and not meant for dev:

- The development people (departments, users, Tom Admin, the COMMS
  team and the AI assistants) are seeded also outside development.
  They need no factory_bot, unlike the rest of the development data.
- After the demo data, the COMMS developers' workloads and a Refresh
  work package demo are seeded: the conference work package with the
  comments and meeting outcomes that its subject, description and
  fields do not show yet.
- Pull previews switch on the ai_assistance feature flag. The LLM
  connection is still configured by hand.
- The root seeder spec is skipped, because its counts no longer match.

A fresh seed without development data was checked: all assistants,
the team, the workloads and the Refresh case are there.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions

github-actions Bot commented Oct 15, 2026 •
edited
Loading

Copy link
Copy Markdown

Deploying openproject with ⚡ PullPreview

Field Value
Latest commit 8f294ad
Job deploy
Status 🗑️ Preview destroyed
Preview URL Destroyed

View logs

@github-actions

github-actions Bot commented Oct 15, 2026

Copy link
Copy Markdown
2 Warnings
⚠️ Attention developer & reviewer: Files with potential user references found:

  • db/migrate/20261012120001_create_ai_assistants.rb (migration with user reference)

Please make sure:

  1. You've added proper relationships (has_many, belongs_to, …) and their inverse in your models
  2. You deal with model destruction dependencies: dependent: :destroy or dependent: :delete_all
  3. You add behavior and tests for the Principal::DeleteJob (app/workers/principals/delete_job.rb)
  4. You replace references to users with deleted user in Principal::ReplaceReferencesService (app/services/principals/replace_references_service.rb) by adding to the replacements initailizer (config/initializers/replace_references_service.rb) for the core, or using the replace_principal_references helper in modules/plugins.
  5. You test the above behaviors with an integration test (e.g., like this one to confirm deletion of users is possible.

This helps prevent dangling database objects when users are deleted and resulting bugs.
⚠️ Attention developer & reviewer: migration looks up an index by hardcoded name:

  • db/migrate/20261015130000_add_one_pending_answer_per_ai_chat.rb

remove_index ..., name: and rename_index MAY fail when the live name differs from
the canonical one (for example, old databases pgloader prefixes MySQL indexes with idx_<oid>_; pre-squash
databases may still use the original migration name).

Please use Migration::Utils instead:

require Rails.root.join("db/migrate/migration_utils/utils")

class SomeMigration < ActiveRecord::Migration[8.0]
  include Migration::Utils

  def up
    remove_index_on :workflows, "wkfs_role_type_old_status", %w[role_id type_id old_status_id]
    rename_index_on :workflows, "wkfs_role_type_old_status", "wkfs_role_type_variant_old_status",
                    %w[role_id type_id old_status_id]
  end
end

Those helpers match the given name, or the name with pgloader prefix, then fall back to columns.

Generated by 🚫 Danger

@thykel thykel closed this Oct 16, 2026
@github-actions github-actions Bot locked and limited conversation to collaborators Oct 16, 2026
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@thykel