Skip to content

v6.25.0

Choose a tag to compare

@github-actions github-actions released this 12 Nov 19:44
· 1086 commits to v6 since this release

Added

  • Optional GitHub API budget pacing for the governor loop. A polling hive with many repositories could spend its whole hourly GitHub budget in the first half of the hour and then make no GitHub calls until the reset, which stopped reviews, merges, hand-offs and fix dispatches; the conserve and critical modes only react once few calls are left. The new governor.api_budget_pacing (default off) measures the hive's own charged core calls per minute over the last 10 minutes and, when that rate would use up the calls left above github.api_reserve before the window resets, lengthens the eval interval and the sweep workers' intervals by rate × minutes to reset ÷ (remaining − reserve), never past governor.eval_interval_max_s and never past the reset. It only lengthens an interval, holds a factor until the need moves by more than 25% or clears by a 20% margin (logged once per change as github api budget pacing changed), and leaves webhook and review wakes alone. /api/status (api_budget_pacing) and GET /api/config/governor/general-advanced (api_budget_pacing_state) report enabled, active, factor, effective_interval_s, burn_per_min and reason; Settings → General has the toggle and the live state. The admin MCP gains governor.general_settings, which previews and confirms changes to the General settings (the eval interval, the loop pacing knobs and the pacing switch) through PUT /api/config/governor/general-advanced, sending only the fields that change and refusing a stale preview. What is and is not paced is listed in src/docs/operator-reference/runtime-and-quota.md (API budget pacing).

Fixed

  • Hive no longer logs the "no pluk pattern set for cli=…" warning on every agent attach; it is logged once per backend per process (e.g. codex), so it no longer trips the repeated-error self-report.
  • Fixed repeated "pluk run directory setup failed" warnings when an existing /var/run/pluk log or command directory owned by another uid already has the shared group access.
  • Reviewer autoscale (review.reviewers.autoscale) now scales an agent that can take the reviews it is counting, and follows a change of agent. With no agent named it used to pick the first agent in name order with role: reviewer. The L5/L6 packs ship two of those, the swarm's reviewer and the escalated-PR adjudicator, and adjudicator sorts first, so a hive with review.reviewer_agents: [reviewer] grew adjudicator-2 against a backlog only reviewer could work, then reported blocked by at-max while the PRs kept waiting. The default is now the first review.reviewer_agents entry the review swarm dispatches to, or with no list the first agent its review keyword scan admits; if there is none, autoscale stays off and says why. A named agent must be one the swarm dispatches to as well: PUT /api/config/review refuses any other with a 400 that gives the reason, and a hive.yaml that already names one still loads, with autoscale off and one reviewer autoscale: not scaling warning. Changing the agent (or turning autoscale off and on) used to do nothing until the next scaling decision: the old agent's extra replica kept running, the new agent inherited the old one's replica count, and nothing was logged. The next governor cycle now puts the old agent back on its declared replicas, starts the new agent from its own declared count clamped into [min, max], starts the controller's streaks, cooldown and draining state again, and logs reviewer autoscale: lane agent changed with both agents. GET /api/config/review gains a reviewer_autoscale block (enabled, agent, effective_agent, reason), shown under Scale reviewers with the backlog as Scaling reviewer and returned by settings_read section review on the admin MCP. The entrypoint pre-creates replica users for every agent that could be the lane, so moving it between reviewers needs no restart.

SBOMs: SPDX JSON, generated by Syft against the linux/amd64 image digest, attached below for hive, hive-contributor, and hive-hub. They cover linux/amd64 only — see releases.md for why linux/arm64 is not scanned separately.

NOTICE: third-party license attribution for the Go module dependencies compiled into these images, attached below. Does not cover base-image OS packages or the Node.js/tmux layers built from source — see releases.md for the exact boundary.