GitHub is an important external ledger and discovery surface, but not the runtime substrate.
GitHub serves six roles in the current system:
- remote audit and synchronization target for this repository
- task-agent collaboration surface for draft PRs and issue comments
- Windows-first CI surface for the tracked task validation baseline
- GitHub Pages distribution layer for public docs, search-engine discovery, and AI-readable landing pages
- technology-radar discovery surface for scanning open-source repositories and releases
- hosted-app identity provider for the first web control-plane login flow
- public feedback surface for Hosted Alpha onboarding, comparison, and first-run signal
- private runtime state in
.local/ - raw memory records
- radar evidence payloads stored in SQLite
- local overrides and credentials
- docs and ADRs
- prompts and skills
- governance, task, and radar configs
- GitHub workflow and template files under
.github/ - Pages configuration, machine-readable discovery files, and public benchmark docs
- radar experiment artifacts under
docs/radar/experiments/ - radar promotion logs under
docs/radar/promotions/ - radar adoption registry updates under
configs/radar/adoptions.yaml
- push working branches
- inspect branch-scoped PR state before PR sync
- create or update draft PRs
- inspect target issues before comment sync
- add issue comments when the task plan calls for them
- present PR and Issue templates that reinforce task briefs, verification, and privacy-safe reporting
Task-agent GitHub sync is now recovery-aware.
- draft PR sync remains branch-scoped, so a resumed run updates the existing draft PR instead of creating a fresh one for the same head branch
- PR sync now performs an observe-tier preflight read after push so
task inspect,task status, and recovery output can state whether resume will create a draft PR or update an existing open PR - issue comments use a stable per-run dedupe marker so resume can reuse an already-created comment if the first attempt failed after the remote write
- planner prompts can see a bounded snapshot of recent open issues and PRs when GitHub is configured
- issue-comment sync performs an observe-tier preflight read so closed issues are rejected before the external-write step
task inspect, CLI status, andGET /runs/{run_id}/recoveryexpose whether a run is resumable, what approval is required next, the likely side-effect risk, and the current remote target state- this is still operator-triggered recovery, not autonomous retry logic
The current public CI lane is intentionally narrow.
- runner:
windows-latest - Python:
3.11 - command source of truth:
configs/task/validation.yaml - default shared profile:
baseline - execution helper:
tools/run_validation_suite.py - preview proof-data check:
npm run web:preview:check - public-site build check:
python -m mkdocs build --strict - hosted web-app build check:
npm run web:build
This keeps the public automation aligned with the runtime boundary while also ensuring the Pages distribution layer stays buildable.
Phase 5 closeout keeps the validation story explicit.
- the public CI lane plus
tools/run_validation_suite.pycover the no-credential tracked runtime baseline npm run web:preview:checkcovers the tracked hosted-app preview proof-data sample used by the public app preview pathpython -m skylattice.cli doctor authis a read-only operator preflight for live capability gaps, not a replacement for CIpython tools/run_authenticated_smoke.py --provider githuband--provider openaiare opt-in read-only checks outside the public CI lane- if a credential-dependent check is not run, record it as
skippedorblockedinstead of treating a green CI run as proof that the live adapter path passed - current onboarding-feedback issues remain a supporting signal for docs and positioning work, not a release gate for Phase 5 operational closure
- issues
#2,#3, and#4are now also explicit Hosted Alpha acceptance inputs for first-run friction, comparison clarity, and post-release onboarding feedback
The Pages site is intentionally public-facing and read-only.
- English root pages are canonical
- Chinese mirrors live under
/zh/ robots.txt,sitemap.xml,llms.txt, andllms-full.txtare tracked artifacts- the site should help search engines and AI answer systems understand the repo, not replace runtime truth
- search repositories via the GitHub API
- read repository metadata and latest release metadata
- optionally push promotion artifacts to
origin/main
The radar does not scrape webpages and does not treat GitHub as its only memory store.
For the current radar MVP, successful promotions may push directly to main, but only when:
- promotion gates pass
- changed paths are whitelisted
- rollback metadata is recorded
- the change is represented by tracked artifacts instead of hidden runtime mutation
This is intentionally narrow and does not authorize broader automatic codebase rewriting.