HivePlane
A release and deployment platform with authentication and regression gating mechanisms. It controls workload rollout through regression reports, promotion workflows, drift detection, and production survival thresholds, emphasizing that gates must actually block problems rather than merely annotate them.
Timeline (last 90 days)
The synthetic probe runner ignored the Spec.expected field and only checked whether the run was admitted, so the probe only failed when a run was rejected and could not flag incorrect output.
After the fix, the recovery logic checks for an active quarantine record rather than the certification status enum, and the quarantine record becomes the source of truth for "whether it is quarantined".
HivePlane v0.2.0's drift detector initially worked by comparing the agent's failure counts across different corpus versions.
Drift auto-quarantine resets the certification status to UNCERTIFIED, but the recovery logic requires status == QUARANTINED to execute, so recovery is impossible after quarantine.
When the benchmark set grew from 10 tasks to 20, an agent that originally passed all 10/10 showed 10 failures under the new counting method, causing it to be misjudged as capability decay and automatically quarantined.
The 36/36 test pass record and the complete design documentation have been published in the project repository.
The v0.2.0 field tests used scenarios S4, S5, S8, and S24 to demonstrate the complete loop of drift detection, evaluation, probes, auto-quarantine, and recovery after re-certification.
Drift detection principle: a single exceedance is a warning rather than quarantine, quarantine requires N consecutive failures or a strong signal, and a stable agent is never quarantined due to noise.
After the fix, the probe starts a run, waits for termination, and compares the output with expected, and a mismatch triggers a probe.decay warning.
The fix for D-9 is to return 409 with an explanation, because a constraint violation is a user-facing condition, not a server failure.
21 more timeline events
All Facts (20)
The 36/36 test pass record and the complete design documentation have been published in the project repository.
50%UnverifiedDrift detection principle: a single exceedance is a warning rather than quarantine, quarantine requires N consecutive failures or a strong signal, and a stable agent is never quarantined due to noise.
50%UnverifiedHivePlane v0.2.0's drift detector initially worked by comparing the agent's failure counts across different corpus versions.
50%UnverifiedAfter the fix, the probe starts a run, waits for termination, and compares the output with expected, and a mismatch triggers a probe.decay warning.
50%UnverifiedThe synthetic probe runner ignored the Spec.expected field and only checked whether the run was admitted, so the probe only failed when a run was rejected and could not flag incorrect output.
50%UnverifiedAfter the fix, the recovery logic checks for an active quarantine record rather than the certification status enum, and the quarantine record becomes the source of truth for "whether it is quarantined".
50%UnverifiedDrift auto-quarantine resets the certification status to UNCERTIFIED, but the recovery logic requires status == QUARANTINED to execute, so recovery is impossible after quarantine.
50%UnverifiedThe fix for D-8 is for the endpoint to report the state derived jointly from the record and the tool's current state, rather than reading the record alone.
50%UnverifiedDuring real-world PostgreSQL field testing, HivePlane v0.2.0 found that the bug was not in the logic, but at the seam where data crosses the persistence boundary.
50%UnverifiedThe HivePlane team recorded a defect ledger containing D-1 through D-11, of which five typical defects cover six types of seams.
50%UnverifiedThe fix for D-5 is to base64-encode the ciphertext when writing to Postgres and decode it when reading back, while the model keeps bytes unchanged.
50%UnverifiedThe fix for D-7 is to use a composite key (source_id, workload_id) with an idempotent upsert, so that the second data source produces a second row instead of a collision.
50%UnverifiedThe fix for D-9 is to return 409 with an explanation, because a constraint violation is a user-facing condition, not a server failure.
50%UnverifiedThe fix for #573 is to verify during restore whether the actual schema head matches the manifest, and refuse to execute if they do not match.
50%UnverifiedThe HivePlane team summarized six types of seams that need to be tracked one by one: binary vs JSON, global keys vs tenant scope, id vs uniqueness, state vs record, foreign keys vs 409, manifest head vs schema head.
50%UnverifiedHivePlane v0.2.0's release flow is: recertify, generate a regression report, promote.
50%UnverifiedIn HivePlane v0.2.0, defect #523's default min_production_runs_survived=50 made the certified state unreachable in a field test that ran only once.
50%UnverifiedIn HivePlane v0.2.0, defect #522 caused recertify_and_promote to still allow promotion when the regression report was blocked=True.
50%UnverifiedThe v0.2.0 field test demonstrated the promotion gate blocking regressions via replayable diffs (scenario S3), drift auto-quarantine and recovery (S4/S5), and the regression diff API.
50%UnverifiedThe fix for #522 is to make recertify_and_promote throw PromotionRefusedError to directly refuse promotion when the regression report is blocked.
50%