feat: better-harness weekly self-review + Worker overview dashboard page - #12
Conversation
…-mediated fleet integration Bake the pinned QoderAI/better-harness CLI into the openclaw-base, hermes, and qwenpaw images (Node 22 via NodeSource for the Python images; npm install --omit=dev since the project is npm-managed) and ship a builtin better-harness skill in every agent template (Manager, OpenClaw/CoPaw/Hermes/OpenHuman/QwenPaw workers, Team Leader). Each agent runs a deterministic, read-only `better-harness harness source` evidence collection at most once per 7 days (run-weekly.sh cadence guard, state persisted to MinIO so it survives restarts), authors findings, and reports them to the Manager over Matrix/shared storage — never the human user. The weekly trigger is native to each runtime: QwenPaw and Hermes worker daemons run a supervised run_better_harness_weekly_loop background task; OpenClaw/CoPaw/ Manager/Team Leader invoke run-weekly.sh from their heartbeat/session-wake. The Manager aggregates all agents' reports via the new harness-integration skill (aggregate-reports.sh) into one consolidated summary, presents a single fleet-wide integration plan to the admin for approval, and applies approved changes to the manager-owned builtin templates / controller-owned config so they propagate to every agent via upgrade-builtins.sh / pushBuiltinSkills. Agents never self-apply hooks/loops/skill edits. Add an AGENTTEAMS_BETTER_HARNESS_ENABLED env gate (default on) propagated to every worker/manager container to disable the weekly review fleet-wide. Tests: pushBuiltinSkills mirrors better-harness for all runtimes (Go); worker weekly-loop unit tests (qwenpaw + hermes); Hermes skill-sync install test; QwenPaw cadence-state persistence test. Live full-stack integration tests are deferred (require an LLM API key and full deployment).
…ry timeline
Add a new Workers tab to the dashboard web UI listing every worker; clicking a
worker opens a drill-down detail with current health (reusing the health-probe
strip), an event/failure timeline, and the worker'"'"'s better-harness weekly
findings (read from the existing MinIO shared/harness-reports/<name>/ space).
Back it with a new controller event-history store: WorkerStatus.RecentEvents is
a bounded ring buffer (newest first, capped at 50) recorded by the health
monitor (health transitions, zombie failures) and the lifecycle handlers
(wake/sleep/ensure-ready), exposed via GET /api/v1/workers/{name}/events.
CRD schema updated and synced to Helm.
Tests: controller unit tests for event capping/ordering, lifecycle recording,
and the events endpoint; proxy allowlist tests for the events/status routes.
Web build verified; Workers tab verified in a browser.
…mediation # Conflicts: # dashboard/web/src/style.css
…refix The AI provider management feature (merged from main) generated route names with the retired hiclaw-<name>-route prefix, breaking the AgentTeams rename-contract check (helm-lint). Rename to agentteams-<name>-route across the gateway client, Higress driver, gateway handler, and their tests.
![kilo-code-bot[bot]](https://avatars.githubusercontent.com/in/2193792?s=60&v=4){kind=small}
| *out = new(string) | ||
| **out = **in | ||
| } | ||
| if in.ServiceEnabled != nil { |
There was a problem hiding this comment.
CRITICAL: BackendRuntime deepcopy removed while field is still *string
WorkerSpec.BackendRuntime is declared as *string (pointer type) at types.go:239, but its deep-copy logic has been removed in this hunk. Without this block, DeepCopyInto will shallow-copy the pointer — both the original and the copy share the same underlying *string. In the Kubernetes controller-runtime, status updates and reconciliation copies could alias the same underlying pointer, causing data races or unintended mutation of the original Worker spec.
Restore the if in.BackendRuntime != nil { ... } block, or re-run controller-gen object paths=./... to regenerate correctly.
Reply with @kilocode-bot fix it to have Kilo Code address this issue.
There was a problem hiding this comment.
Verified as a false positive, with evidence:
-
The base branch's
zz_generated.deepcopy.gocontained a duplicateif in.BackendRuntime != nil { ... }block insideWorkerSpec.DeepCopyInto(two occurrences - one afterDeployMode, one afterEnv). This hunk removed the duplicate; a valid block remains (now after theEnvblock, matching controller-gen's canonical field order pertypes.go). -
Re-running
controller-gen object paths=./api/v1beta1on this branch produces zero diff - the committed file exactly matches machine-generated output, with exactly oneBackendRuntimedeepcopy block. No pointer aliasing is possible. -
go build,go vet, andgo testpass for theapi,server,controller,service,gateway, andconfigpackages.
The other three findings (unpinned BETTER_HARNESS_REF, missing jq guard, missing keyboard handler) were confirmed and fixed in 658f8cd.
| # better-harness skill for weekly agent self-review. Node 22 (installed above | ||
| # via NodeSource) satisfies the CLI's >=22.20 engine constraint. The project is | ||
| # npm-managed (packageManager: npm), so install with npm, not pnpm. | ||
| ARG BETTER_HARNESS_REF=main |
There was a problem hiding this comment.
WARNING: BETTER_HARNESS_REF=main makes builds non-reproducible
The default main branch is a moving target — git clone --depth 1 --branch main fetches whatever HEAD is at build time. If the upstream QoderAI/better-harness repo introduces a breaking change or force-pushes, all image builds will silently change behavior or break.
Pin to a specific commit SHA or tag (as done for HERMES_GIT_REF=v2026.4.16 in the hermes Dockerfile) to ensure reproducible builds. The same unpinned default appears in hermes/Dockerfile:87 and qwenpaw/Dockerfile:109.
Reply with @kilocode-bot fix it to have Kilo Code address this issue.
| # Collect every findings file (any agent, any date). | ||
| mapfile -t FILES < <(find "${LOCAL_DIR}" -mindepth 2 -maxdepth 2 -name '*.json' 2>/dev/null | sort) | ||
| if [ "${#FILES[@]}" -eq 0 ]; then | ||
| jq -cn '{agents:0, reports:0, findings:[]}' |
There was a problem hiding this comment.
WARNING: jq used without verifying availability
mc is guarded with command -v mc on line 30, but jq is called directly here and at lines 49, 55, and 61. If the container image lacks jq, the script fails with an unhelpful "command not found" error. Add an early guard after the PREFIX validation:
if ! command -v jq >/dev/null 2>&1; then
log "ERROR: jq is required but not found on PATH"
exit 1
fiReply with @kilocode-bot fix it to have Kilo Code address this issue.
| } | ||
| el.innerHTML = workers.map((w) => workerRow(w)).join(''); | ||
| el.querySelectorAll('[data-worker-name]').forEach((card) => { | ||
| card.addEventListener('click', () => onOpen(card.dataset.workerName)); |
There was a problem hiding this comment.
SUGGESTION: Add keyboard handler for Enter/Space on worker cards
The card has role="button" and tabindex="0" (line 106), making it keyboard-focusable, but only a click listener is registered here. Keyboard users pressing Enter or Space won't trigger the detail view. Add a keydown handler alongside the click listener.
Reply with @kilocode-bot fix it to have Kilo Code address this issue.
Code Review SummaryStatus: No Issues Found | Recommendation: Merge Incremental Review (658f8cd)All 4 issues from the previous review are resolved:
No new issues found in the incremental diff. Files Reviewed (6 files)
Previous Review Summary (commit 7d26a2d)Current summary above is authoritative. Previous snapshots are kept for context only. Previous review (commit 7d26a2d)Status: 4 Issues Found | Recommendation: Address before merge Overview
Issue Details (click to expand)CRITICAL
WARNING
SUGGESTION
Files Reviewed (63 files)
Reviewed by mimo-v2.5-pro · Input: 103K · Output: 18.1K · Cached: 785.2K |
- Pin BETTER_HARNESS_REF to commit SHA 4b871d9b in openclaw-base, hermes, and qwenpaw Dockerfiles (init/fetch/checkout pattern; branch/tag overrides still work) for reproducible builds - Guard jq availability in harness-integration aggregate-reports.sh, failing fast with a clear error - Activate dashboard worker cards with Enter/Space in addition to click (role=button cards were keyboard-focusable but not operable) The CRITICAL deepcopy finding was verified as a false positive: controller-gen regeneration produces zero diff; the PR hunk removed a duplicate BackendRuntime block present in the base file.
1 participant
Summary
This PR brings two related features to the fork:
1. Better Harness weekly self-review (96cd091)
QoderAI/better-harnessCLI into theopenclaw-base,hermes, andqwenpawimages (Node 22 via NodeSource for the Python images;npm install --omit=dev).better-harnessskill in every agent template (Manager, OpenClaw/CoPaw/Hermes/OpenHuman/QwenPaw workers, Team Leader).better-harness harness sourceevidence collection at most once per 7 days (cadence guard, state persisted to MinIO), authors findings, and reports them to the Manager — never the human admin.run_better_harness_weekly_loop; OpenClaw/CoPaw/Manager/Team Leader invokerun-weekly.shfrom heartbeat/session-wake.harness-integrationskill aggregates all agents findings and presents one fleet-wide plan for approval.AGENTTEAMS_BETTER_HARNESS_ENABLEDenv gate (default on) for a fleet-wide kill switch.2. Worker overview dashboard page (5c67d1a)
shared/harness-reports/<name>/).WorkerStatus.RecentEventsbounded ring buffer (newest first, capped at 50) recorded by the health monitor (health transitions, zombie failures) and lifecycle handlers (wake/sleep/ensure-ready).GET /api/v1/workers/{name}/events; CRD schema updated and synced to Helm.Verification
go build,go vet, and full tests pass (api, server, controller), including new event tests.vite buildsucceeds; Workers tab verified in a browser.