Guides
Coding tasks & verification
Run coding agents with company + repo context, then independently verify before treating work as delivered.
Coding tasks are how HiveBase supervises coding agents against business intent and repository context, not just a green CI check. Agents propose; verification judges whether the work is actually delivered. Product surface: /tasks/code.
Overview: Tasks. Connected agents: Agents. Repos: GitHub.
Producer ≠ verifier
The agent that produces a change is not the final judge of “done.” HiveBase separates execution from verification so “completed” stays a claim until evidence supports a verification verdict.
- 01Intent + contextBrain + repo + brief
- 02Agent proposesDiff / PR
- 03VerificationIndependent scorecard
- 04VerdictVerified or held
- 05You / merge policyStill gated
- 01
Intent + context
Brain + repo + brief
- 02
Agent proposes
Diff / PR
- 03
Verification
Independent scorecard
- 04
Verdict
Verified or held
- 05
You / merge policy
Still gated
What coding tasks use for context
| Source | Role |
|---|---|
| Company Brain | Why this work exists — customers, decisions, constraints |
| Connected repos | What the code actually is |
| Task brief | Acceptance criteria and scope |
| Agent host | Cursor, Claude Code, Codex, OpenCode (etc.) as runners |
Verification in plain language
Typical verdict shapes (labels may vary in product UI):
| Verdict | Meaning |
|---|---|
| Verified | Checks / evidence support the reviewed head |
| Reviewed | Diff and risks examined; strong executed proof may still be missing |
| Unverified | Evidence too thin — gaps named, not over-claimed |
| Stale | Newer pushes invalidated the prior head; re-verify |
Completed is a claim. Delivered is a verdict.
Trust defaults
- Merge, force-push, and destructive git operations are not silent.
- GitHub writeback follows fail-closed certification — see GitHub.
- Every consequential apply path can leave an action receipt.