What shipped.

The worker now threads each task’s deadline into its GitHub client and re-enqueues on a GitHubRateLimitError using the API’s own retryAfter instead of failing the task outright. A rate-limited request now waits and resumes rather than burning a retry budget on a call that was always going to succeed once the window reset. PR #8259.

Colony’s dependency graph now mirrors to GitHub’s native issue-dependencies API: when an issue is blocked by an unfinished dependency, the block shows up in the GitHub UI as a “blocked by #N” badge instead of living only in Postgres. The change parallels the existing subtask → sub-issue mirroring, and a self-healing reconciler corrects drift between the two. PR #7966.

Four separate code paths could apply the colony:needs-human label without posting a comment explaining what triggered the escalation, leaving operators to reconstruct the reason from pipeline state. All four now post an explanatory comment alongside the label. PR #7709.

Marshal now runs a focused LLM pass to catch semantic conflicts — overlapping changes two concurrent branches make to the same symbols — before merge, not just the textual conflicts Git already catches. High-confidence conflicts route through overlap-deferral instead of merging blind. PR #6103.

Colony’s Strategist subsystem shipped this week. Operators set a charter — goals plus a risk envelope that caps the largest issue size the Strategist can queue without escalation — and the Strategist collects pipeline signals (stalled issues, cost concentration, cycle time) once a week and proposes a snapshot of self-improvement track changes for operator approval.

The full stack landed together: a Postgres-backed charter/snapshot store, a one-shot LLM executor that turns signals into a proposed snapshot, a weekly cron trigger with on-demand override via colony strategize, and a dashboard charter editor for reviewing and approving proposals. PRs #5493, #5496, #5511, #5545.

The executor pipeline now routes each task to a model based on predicted cost and complexity instead of one fixed model per agent, configured through a new hot-reloadable model_routing config block. PRs #5241, #5229.

The webhook receiver now ingests GitHub pull request lifecycle events — opened, synchronized, closed — for pipeline-created PRs, persisting them through the pull-request store instead of waiting on Mayor’s next poll to notice a merge or a force-push. PR #4960.

Operators can now pull a GitHub issue into the pipeline directly from a comment: /colony:enqueue triggers intake without waiting for the next poll cycle or relabeling the issue by hand. PR #4779.

The dashboard’s SSE stream now carries per-task progress, not just issue-level state transitions — PipelineStateView updates as a task runs instead of waiting for the next state change. The monitor service emits issue_status_updates events sourced from the pipeline store’s task lifecycle, and the existing useSSE hook forwards them straight through. PR #4555.

Every in-flight issue now carries a current_status field: a human-readable description of what the pipeline last did, written atomically alongside every state transition. Analyzer, Planner, Developer, Reviewer, Merger, and Sweep each emit a status string at key moments in their execution; operator actions and system events (cost-cap hits, auto-unblocks, dependency resolution) do the same. The status surfaces in the SDK via getIssueStatus() and getIssueEvents(), and in the dashboard IssueDetail view. Issue #3977.

Also shipped this week: the CI repair executor. When a build fails and the failure is diagnosable, the pipeline now re-invokes the Developer with the structured CI failure context instead of escalating to the operator. Issue #3913.

The Colony Cloud dashboard now receives pipeline events over a server-sent events stream instead of polling on a fixed interval. Worker status updates, cost changes, issue state transitions, and alert events push to the browser as they happen. A useSSE hook wraps the stream with automatic reconnect and a polling fallback when the connection is unavailable. A connection indicator in the header shows stream health; the HiveView breathing dot now dims when data is stale rather than just when the connection drops.

IssueDetail shipped a parallel set of improvements: the header bar now surfaces the linked PR, current cost, and relative merge time at a glance. A PipelineRail component renders the issue’s phase history visually, and the view auto-refreshes while an issue is in flight.

The colony CLI now ships bootstrap templates for Rails, Next.js, and Django, each with integration flags (Stripe, Resend, R2) that scaffold a working app the pipeline can operate against from issue one. End-to-end smoke tests cover all five bootstrap templates in CI. PRs #3650, #3651, #3653.

The home page now leads with the category sentence: The orchestration layer for AI coding agents. The receipts band — 1,284 PRs merged, 14m median cycle, $0.91 per merged PR — moved above the problem framing. The primary CTA is Read the receipts →, pointing at the new /built-by-colony/ destination (live data once the public-stats API ships).

Tightened the cost-cap escalation path so that builders spawning subagents bubble their incremental spend up to the parent issue’s budget in real time, instead of at task close. Inspector now sees the spend trajectory and can escalate to operator before the cap is breached. PR #2843.

Hybrid tenancy — control plane managed by Colony, worker pools in your VPC — is now in private preview with a small set of design partners. Public preview Q3 2026. Talk to us on the pilot call if you need stricter data residency than Cloud’s regional default provides.

colony why <issue> now emits a structured diagnostic report for any issue, paired with a new dead-letter table that captures terminal failures with block_reason and block_context so the Sentinel agent surfaces them to operators instead of letting them sit silent. PRs #2459, #2488.

Colony now talks to Anthropic models through Azure AI Foundry, for teams whose enterprise procurement already covers Azure model contracts. The worker config picks the provider per agent; the cost ledger reports through the same pass-through pricing. PR #2154.

Between March 31 and April 1, the pipeline merged 120+ PRs on its own repo in a single 36-hour window — Mayor through Marshal, with Inspector running deterministic checks and an LLM review before every merge. The failure modes the session surfaced are now encoded as recovery logic in docs/production-learnings.md and anchor the proof points the rest of the site leans on.

Several large internal modules — the Builder executor (2,547 lines), the Mayor’s coordination logic, the SDK type surface — were decomposed into focused submodules through self-improvement issues the pipeline filed against itself. Surveyor prompts now require call-site verification before concluding an issue has no work to do. PRs #1357, #1408.