Spec · CHG-2026-014916 · Agentic SDLC

change_id: CHG-2026-014916
risk_class: R3
autonomy_tier: A2
controls: ['TRC-01', 'CHG-04', 'HUM-14']
data_classification: internal
originator: olaf@freundcloud.com
agent_identity: svc-agent-platform
model_route: gateway/tier-frontier
supersedes: None

Requirements

R1 — The question belongs to `detect`, not to either consumer

diagnose and propose both raise findings, so putting the lookup in each is the same six lines twice — the duplication CHG-2026-014911 was spent removing, reintroduced in YAML where no linter will find it.

“Is this finding new?” is part of detection. detect runs first, holds the metric, and gains one output, existing_finding: the number of an open Stage 6 issue already tracking this metric, or empty. Both downstream jobs gate on it in their existing if:.

The metric name is added to detect_anomaly.py’s GITHUB_OUTPUT block. It already computes it — result["metric"] — and merely does not emit it.

R2 — Matching is deterministic and narrow

An open issue matches when it carries the stage-6 label and its title contains the metric name. Both proposing paths already title their issues with it.

Deliberately narrow. Title-and-label is inspectable, needs no state, and is wrong in only one direction: an unusual title produces a duplicate, which is today’s behaviour. Anything cleverer — similarity scoring, a model judging sameness — puts a judgement call in detection, which principle 4 forbids.

R3 — Fails open, and says which way it failed

If the lookup errors or the API is unavailable, existing_finding is empty and the finding is raised. A duplicate is an annoyance; a suppressed real anomaly is worse than the problem being fixed.

R4 — Suppression is visible or it is not suppression

A run that suppresses writes to its step summary which issue it deferred to. A silent skip is indistinguishable from a detector that stopped working — the failure that cost three nights under CHG-2026-014914.

R5 — Closing a finding re-arms the detector

Matching is on open issues only. Close the finding and the next breach raises a new one. This is what stops suppression becoming permanent silence, and it makes closing an issue the explicit act of saying “tell me again if this recurs”.

Policy conflicts

None. .github/workflows/ floors this at R3 (policy/risk-classes.yaml), so it is declared R3/A2 — the matrix maximum at R3 in development.

The detect job gains issues: read. It does not gain issues: write: detection looks, and the jobs that were already permitted to write are the ones that write.

Verification

make build test lint gates.
scripts/tests/test_detect_anomaly.py extended: the metric is emitted to GITHUB_OUTPUT.
Dispatch with the tracked finding open → suppressed, summary names the issue.
Close it, dispatch again → a new finding is raised. R5 is the requirement most likely to rot, so it gets tested by doing it.

Spec · CHG-2026-014916