Technology Radar¶
The technology radar is Skylattice's bounded external-learning workflow.
Goal¶
Turn GitHub open-source change into durable, reviewable local capability through this path:
discovery -> scoring -> knowledge increment -> experiment -> promotion -> rollback
Sources¶
Current live source set:
- GitHub repository search
- GitHub repository metadata
- GitHub latest release metadata
No webpage scraping, browser automation, or multi-source aggregation is in scope yet.
The runtime now expresses this through a stable RadarDiscoverySource boundary so future providers can fit without rewriting RadarService.
Tracked provider intent now lives in configs/radar/providers.yaml, so future provider rollout starts as a reviewable config change instead of a hidden runtime assumption.
Candidate and evidence inspection now also carries a provider-neutral identity contract beside the current GitHub-shaped fields, so future providers do not need to masquerade as GitHub repositories just to enter the radar pipeline.
Candidate Selection¶
Tracked source config lives in:
configs/radar/providers.yamlconfigs/radar/sources.yamlconfigs/radar/scoring.yamlconfigs/radar/promotion.yamlconfigs/radar/schedule.yaml
The current defaults favor:
- recent repositories
- recently active repositories
- strong star signal
- topic overlap with
agent,memory,evals,developer-tools,coding-agent,openai, andgithub
Experiment Shape¶
Each spike is repo-contained and intentionally narrow.
Artifacts:
- experiment branch:
codex/radar-<slug>-<runid> - experiment note:
docs/radar/experiments/<slug>-<runid>.md - promotion log:
docs/radar/promotions/<slug>-<runid>.md - behavior change registry:
configs/radar/adoptions.yaml
Local Scheduling¶
Skylattice does not run a resident scheduler inside the repo.
- tracked schedule intent lives in
configs/radar/schedule.yaml skylattice radar schedule showexposes the tracked schedule viewskylattice radar schedule render --target windows-taskrenders Windows-first local registration detailsskylattice radar schedule runresolves one tracked schedule into the same boundedradar scanpath- radar-scheduling.md is the operator runbook for Windows task registration, first-run checks, and weekly-cycle validation
This keeps schedule behavior reviewable without promoting GitHub or an internal queue to runtime truth.
Promotion Rules¶
A candidate can promote only when:
- the worktree is clean
- the base branch is
main - the experiment passed validation
- the candidate cleared the promotion threshold
- the weekly promotion cap has not been exceeded
- all changed paths are inside the tracked allowlist
- freeze mode is off
Behavior Change Mechanism¶
The first behavior change is deliberately simple and auditable:
- promotion writes a tracked adoption entry
- future radar scans read that adoption registry through provider-neutral identity first, with legacy
repo_slugmatching kept as a fallback - candidates matching adopted tags receive a small scoring boost
That means the system changes future behavior through Git-reviewed config, not hidden heuristics.
Rollback¶
Each promotion stores:
promotion_idsource_branchbase_commitexperiment_commitmain_commitrollback_target
Rollback is explicit and Git-native. It reverts the promotion commit chain and records the rollback in ledger and episodic memory.