Guildwise HQ

← All releases

Guildwise HQ 1.7.0

Runner 0.1.74. No migrations and no new environment variables. Roles changed: tester v14, planner v3, architect v6, builder v8.

What's new

Faster DevOps queue, prepared worktrees

DevOps requests that a parked card or a working agent is waiting on go first, and a second DevOps shift opens behind a long run, so a short setup request no longer waits behind a 10-minute shift. A parked card says where its request stands in the queue. The runner also installs a project's dependencies in a new card worktree before the agent starts. Submodules are not prepared; the agent asks DevOps for them.

See where team time and money go

The Metrics page has two new tables. Time and cost by role shows shifts, shift time and cost per role, seniority and model. Builder rework by seniority, model and card size shows how often builders' cards came back from review, so you can tell whether a stronger model pays off on large cards. Planners and architects can also mark a card large before its first build, so a senior builder takes it.

Tests-only cards go to the tester

When a card only needs tests, the architect marks it as tests-only and the tester writes them, instead of a builder who is not allowed to touch the test specs. The card is no longer sent to a builder and bounced back.

Full changelog

Added

  • Metrics shows where the time and money go. Two new tables on the Metrics page: Time and cost by role (shifts, cards, shift time, median shift, cost and cost per card, per role, seniority and model) and Builder rework by seniority, model and card size (cards a builder submitted, how many came back from review, rework rounds and the rate). Use them to judge, for example, whether the stronger model pays off on large cards; read a rate together with its card count. New team object on GET /api/metrics.
  • Planner and architect can mark a card large (aom risk large --reason "<why>") before its first build, so a senior builder takes it. They never lower a card's size; the business analyst and the manager still set either size.
  • Cards that only need tests are worked by the tester. The architect marks such a card with aom design --tests-only; its Build stage then goes to a tester (the hire hint names the tester when none is hired) and the tester needs no test environment for it. The server refuses the flag on an investigation-only card and when no active tester may work the project. A runner older than 1.7.0 ignores the flag and the card goes to a builder as before.
  • More hashtags on every social post. Facebook now takes 0 to 5 (was 3), LinkedIn 3 to 8 (was 5), X and Bluesky 1 to 3 (was 2), TikTok 3 to 8; Instagram stays at 5 and YouTube is unchanged. Besides the topical hashtags, each post gets one standing tag and one popular tag (at most two from either kind), rotated by a fixed rule so the same post always gets the same tags. Defaults: standing OpenSource, Free, FreeSoftware, SelfHosted, FOSS; popular AI, Automation, DevOps, Tech, Productivity, AIAgents, GenerativeAI, MachineLearning, Programming, SoftwareDevelopment. On X and Bluesky the standing and popular tags are dropped first when the text does not leave room for them.
  • Two platform settings, blog_writer_hashtags_standing and blog_writer_hashtags_popular, replace those lists; an unset key uses the defaults and an empty list turns that kind of tag off. PATCH /api/platform/blog/writer accepts hashtagPools, and the blog writer and video status responses carry it.
  • Worktree preparation. When the runner creates a card's worktree it installs the project's dependencies (pnpm, npm, dotnet, uv, go and the other toolchain steps) in the run-step sandbox before the agent starts, within 5 minutes, so the agent no longer spends its first minutes or a DevOps request on it. A failed or cut-off preparation does not fail the shift: it is noted on the card and the agent is told what was not prepared. New runner config key prepareWorktrees (default true) turns it off. Submodules are not prepared: a repository with submodules gets one line telling the agent to ask DevOps.
  • Queue position for parked cards. A card parked on a DevOps request says where the request stands ("number 2 of 3 in the DevOps queue" or "running now") in its park messages and in aom card facts.

Changed

  • DevOps requests are served in a better order. Requests somebody waits on (a parked card, an agent still holding a card shift on that card, or a card held at Test deploy) go first; any request older than 30 minutes joins them, so nothing starves; the rest keep their order of arrival. While one DevOps shift has run for 5 minutes or more, a second DevOps shift may start for such a waited-on request of another card, so a short setup request no longer waits behind a 10-minute shift. At most two DevOps shifts run per engineer.
  • Stop-loss: a large card gets three review rounds (was two) before it is parked; its budget is now 4 build passes, 3 review rounds and 25 shifts. An agent (the architect, when there is no CEO) may extend one card's budget at most three times; after that the owner's desk decides, and the stop-loss item shows how often the budget was extended.
  • Builders run the whole unit test command of each package they changed once, after their last change, as their evidence; while working they run only the test files they touch. (Builder v8.)
  • A builder is no longer stopped for editing documentation, package.json, lock files or configuration inside an E2E folder, or for E2E fixtures and snapshots (those count as test content); the builder's handoff checks ignore E2E test content. A tester on a tests-only card is still held to its specs.
  • A read-only role (such as the reviewer) is judged on what it changed during its own shift: files that were already modified when the shift started no longer get it refused.
  • aom question --recommend counts from 0. A number equal to the option count is read as the 1-based number of the last option, and the accept message always says how the number was read; any other wrong number is refused with the valid ones listed.
  • CHANGELOG.md merges with git's union strategy (.gitattributes), so two cards' changelog entries rebase without a conflict.

Fixed

  • Instagram video uploads that ended in a processing failure are retried. A processing failure, a Graph "unsupported request" or "does not exist" answer and an answer marked transient are retried up to 4 attempts with a new container, using the existing back-off of 1, 5 and 30 minutes; an authorization or rate-limit refusal, a server error and invalid media keep their old handling, and nothing is ever published twice.
  • MP4 files are written without edit lists (-use_editlist 0, with the index at the front): the rendered videos and the MP4 copies of help videos that Instagram refused.
  • Org chart: an agent tile is named by what it shows (name, role, status, workstation), followed by where its questions go; the screen-reader name no longer differs from the visible text.
  • A drifted submodule no longer blocks a card. A submodule found off its pinned commit is reported to the agent with the one command that puts it back (git -C <path> checkout <pin>); the runner does not reset it. If the pin is not available locally the agent asks DevOps (aom question --topic environment).
  • The runner no longer runs a program planted in a nested repository or submodule of an agent's worktree: its own git commands stay out of them. Update workstations to runner 0.1.74.
Many hands. One master.

Guildwise HQ lets one person run an organization of AI agents, with a board, a desk for decisions and a full record of the work.

See how Guildwise HQ works