Repository navigation
v1.8.38
A UI task now names its design before any code, however you ask. The design rule lived only in /plan and in a file quick tasks are told not to read, so "add a settings screen" stopped at the design gate under /plan and was simply built when asked in plain words. AGENTS.md now carries one line: before editing user-visible UI in production directories, name the design artifact (a link or a committed docs/design/<screen>.md); with none, say so and offer to draft the wireframe first.
- If you have a design file, nothing changes: the agent builds from it.
- If you have none, expect one extra exchange per new screen. The agent drafts the wireframe itself once you say yes.
- Text tweaks, backend work and files under
tools/,scripts/ortests/are not stopped.
Measured on Claude Sonnet in headless downstream sessions: 11 of 14 new-screen tasks with no design stopped first, against 1 of 3 without the line. It raises the rate; it does not guarantee it. A screen added inside an existing page was the weak case, Haiku ignored the line in its one run, and other platforms are not measured. The line governs process, not appearance: a wireframe the agent drafts for itself still produced a bare screen. Spec: docs/specs/ui-design-on-any-entry.md.
No test runner no longer means no way to ship. With no runner, /test asked feature, architecture-change and hotfix for your sign-off and then still refused the receipt that /ship requires. Your recorded sign-off is now the acceptance, and the Work Log says the evidence is a manual trace. A red linter, or a red suite where one exists, still blocks.
Contradictions fixed.
/shipno longer asks ADRs forstatus: frozen; ADRs areaccepted./spec-intakenames the post-ship statusshipped.- The guardrails call the tiny-fix reading mode
Lite, like every other file. AGENTS.mdrule 8 asks for a spec only from the tiers that need one./handoff§6 lets completed## Evidencemove to the overflow archive; gate receipts never move./reviewno longer sends a Not Ready verdict to/plan, a route the validators reject./testcited twoAGENTS.mdsections that do not exist.
Removed, because nothing read it.
verify_agent_evidence.pyand its test. It was never deployed.- The state values
Context Budget tier,FinalizedandMode: Normal | Fast Lane. - Nine keys in
.agent/config.yaml(raw_intake_cleanup, four undercluster, four underdownstream_capabilities). The validators hard-code those values; editing the keys never had an effect. - Eight citations, in deployed files, of paths that are never deployed.
Downstream delta:
- Core files updated on deploy:
AGENTS.md;.agent/config.yaml;.agent/rules/{engineering_guardrails,state_machine}.md;.agent/workflows/{ask-local,handoff,plan,review,shared-contracts,ship,spec-intake,test}.md; four guides under.agentcortex/docs/guides/;.agentcortex/metadata/trigger-compact-index.json; the version banners. - If you edited
.agent/config.yaml, deploy replaces it and keeps yours as.agent/config.yaml.acx-local. That is existing behaviour; compare the two after upgrading. - No validator line is added or removed, and no verdict changes.
Checked downstream before release: three projects installed from 1.8.37 (plain, with a knowledge base, with a design reference and a custom skill) were upgraded. validate.sh gave the same result before and after in each (fail=0), nothing was removed, and the custom skill, its metadata and the design file kept their hashes.