Repository navigation
v1.8.35
The README says what the validator actually checks. In English, 繁體中文 and llms.txt, phase records are described as agent-written: the validator checks that each phase left one, in order, not that the work behind it happened. The secret scan and your test suite are the checks an agent cannot argue past.
Test evidence is tied to the code it tested. .agentcortex/tools/record_test_run.py run --worklog <log> -- <test command> runs your suite without a shell and records the exit code and a fingerprint of the code (runtime state under .agentcortex/context/ excluded). validate.sh and /ship fail a current-branch log whose latest recorded run failed before a PASS receipt, or whose code changed after that run once it is shipping. A feature/architecture-change/hotfix log with a test or ship PASS and no recorded run gets a WARN. It catches stale and failed evidence, not a hand-typed line.
A declared knowledge base is consulted on a plain task (#234). If downstream-capabilities.yaml declares knowledge_sources, AGENTS.md and CLAUDE.md now say to read it at task start and consult the KB before planning or reviewing a change in a KB-routed domain — no /bootstrap needed. Re-running the 2026-10-18 harness, sonnet applied the KB's rule on 6 of 6 covered tasks (0 of 6 before) at the same cost; haiku was unaffected (#235).
acx-* subagents run on your session's model (#236). They pinned opus (reviewer) and sonnet; a haiku session was silently upgraded.
Installers. deploy_brain.sh --ref <tag> (-Ref in PowerShell, or ACX_REF) pins an update to a release. deploy.sh --help prints usage, and an unknown flag exits 2 instead of becoming the target directory; deploy.ps1 --help no longer deploys into the current directory. When your project keeps its own AGENTS.md / CLAUDE.md / GEMINI.md, deploy and both validators now say plainly that governance is not active until the .acx-incoming sidecar is merged.
Rules and docs. engineering_guardrails §8: a standing instruction ("keep going", permission to merge) does not answer an owner decision such as freezing a spec or amending an ADR. security_guardrails §2 reviews LLM output as untrusted input. /retro retires a lesson first when its root cause is now enforced. docs/INSTALL.md pins its CI snippet's actions by SHA and lists how each platform starts a governed task. The repo's own TruffleHog moves to v3.99.2 with its image digest.
Measured, not claimed: a first lifecycle simulation (#239, six chained releases, n=1) found governed and no-framework runs equal in quality and the governed run about 25× the cost on tasks that left no headroom (corrected 2026-10-31 from 53×, a scoring error) — docs/reviews/2026-10-20-lifecycle-sim-pilot.md.
Downstream delta:
- Core files, updated on deploy:
.agent/rules/engineering_guardrails.md,.agent/rules/security_guardrails.md,.agent/workflows/test.md,.agent/workflows/ship.md,.agent/workflows/retro.md,.agentcortex/bin/deploy.sh,.agentcortex/bin/deploy.ps1,.agentcortex/bin/validate.sh,.agentcortex/bin/validate.ps1,.agentcortex/tools/record_test_run.py(new),.claude/agents/acx-*.md,.agentcortex/metadata/trigger-compact-index.json,installers/deploy_brain.*, and the version banners in.agentcortex/docs/. - Scaffold files:
AGENTS.mdandCLAUDE.md. If you edited them, the new version arrives as.acx-incoming— and deploy now tells you governance is not active until you merge it. - New validator line: "test evidence freshness". Silent unless a Work Log records a test run, or a tested-tier log has a PASS receipt with no recorded run (WARN).
deploy.shnow rejects unknown flags with exit 2. A script that passed an unsupported flag used to deploy into a directory named after it.
Checked downstream before release:
- An upgrade from 1.8.34 updated 206 files, added
record_test_run.py, removed nothing and left the project's own files byte-identical; the downstream validator reportedfail=0with the new test-evidence line present andacx-revieweratmodel: inherit.