Reports

Working notes and investigations that don't belong to a single work item's spec — deeper dives, surveys, and reference write-ups produced along the way.

• GitHub queries — the GitHub API queries the poller / webhook ingress relies on, and where they live in the CLI.
• Vendor SDKs vs. binaries — whether the-loop should stop shelling out to claude, cursor-agent and gh in favour of the vendors' SDKs, and what each swap would actually cost.
• Status labels & dashboards — the label taxonomy the-loop defines, why the labels sit unused today, and how they feed a GitHub Projects kanban without a bespoke dashboard.
• Is harness-config.yaml required? — every reader of the per-repository harness config, how each key is enforced, why the file cannot be removed, and which keys belong in the CLI config or nowhere (issue-352).
• The end-to-end Slack test series — one work item driven from Slack against a live deployment, operator-simulated, each run validating the previous run's fixes and recording its own findings (redacted; the per-item follow-ups live in followups/):
  • 2026-09-19 — the first run: 13 interventions and two hours; the findings became issue-393, #395, #396, #397.
  • 2026-09-20 (run 2) — on 19.7.0: five messages and 50 minutes; the N1–N3, O7, O9 findings, fixed in PR #402.
  • 2026-09-20 run 3 — on 19.8.3 with zero setup deltas: the P1 (typed selection grammar) and P2 (close path vs. the session's endgame) findings, fixed in issue-405.
• The architecture of the-loop (CLI) — how the CLI is put together, the exact the-loop start / the-loop execute flows, which guardrails keep design behind requirements-approval (and which are code), and how the harness and the daemon talk — with component and sequence diagrams (issue-354).