Process Optimization: Operate DCRH (Scan → Triage → Remediate)
Analysis of the process documented in process-operate-dcrh.md, using the current-state / waste / future-state / impact framework.
1. Current state
| Step | Who | Typical wait/handoff |
|---|---|---|
| Threat model | Security analyst | None — self-contained, read-only |
| Scan (static or pipeline) | Analyst or engineer | Handoff into Triage; pipeline path also waits on sandbox setup if not already done |
| Triage | Security analyst | Handoff to Security engineer for patching |
| Patch | Security engineer | Handoff to target's engineering owner for review/merge |
| Act on verdict | Target's engineering owner | Often the longest wait — competes with the owning team's own roadmap |
The process as documented is already fairly lean (5 steps, matching DCRH's own detect → verify → report → patch shape) — the waste below is mostly about operational friction around the steps, not the steps themselves.
2. Waste identified
- Waiting — sandbox setup treated as a per-run step instead of a standing service.
setup-and-sandbox.mdis a one-time-per-host runbook, but if a team runs DCRH from ephemeral CI runners or fresh VMs each time, every scan pays the full gVisor + image build cost. This is the single largest time sink for teams that don't keep a persistent sandbox host. - Rework — skipping the small first wave. The SOP and runbook both call out running
--runs 3before scaling to--runs 20+, but nothing enforces it. Skipping straight to a large batch on an un-calibrated prompt/target risks burning the token budget on a wave that needed a--max-turnsadjustment discovered only after the fact. - Rework — patching without the re-attack tier, then treating it as done.
--no-reattackis documented as a fast-iteration flag, not a shipping flag, but nothing technical stops a "fast" patch from being merged as final. This is a process-adherence gap, not a tooling gap. - Handoff — triage-to-patch and patch-to-review are both manual handoffs with no
shared queue.
/triage's component-owner routing helps identify who should review, but there's no tracked list of "verified findings awaiting a patch attempt" or "verified patches awaiting human review" — both currently live only as files on disk (TRIAGE.json,results/<target>/<timestamp>/). - Over-processing — re-deriving
focus_areason every run of a stable target. If a target's attack surface hasn't changed, running--auto-focus(which re-runs recon) every time repeats workconfig.yaml's hand-writtenfocus_areascould skip. - Manual work — no CI integration documented. The repo supports recurring scans (per its own "Looking Forward" section), but this process document, like the underlying repo, assumes a human manually invoking each CLI step. Teams that want daily/weekly cadence currently have to build that scheduling themselves.
3. Future state (recommendations)
- Stand up a persistent sandbox host (or a small pool) rather than re-running
setup_sandbox.shper invocation. Low effort, removes the largest recurring wait for any team running more than one scan. - Make the small-first-wave step a hard gate in tooling, not just documentation — e.g.
a thin wrapper script that refuses
--runs > 5unless a prior smaller run'sresult.jsonfiles exist for the same target. Medium effort (abin/wrapper script, not a harness code change). - Track
--no-reattackpatches distinctly — tag them inPatchVerdictoutput or a simple checklist so "fast iteration" fixes can't silently become "shipped" fixes without someone consciously accepting the lower verification bar. Low effort. - Add a lightweight shared queue for triage→patch and patch→review handoffs — even a
checked-in
QUEUE.mdper target, or a project-tracker integration if the team already has one, closes the "living only as files on disk" gap without requiring new infrastructure. Low effort, medium impact for teams working multiple targets in parallel. - Prefer hand-written
focus_areasinconfig.yamlfor stable targets, reserving--auto-focusfor first scans or after a target's attack surface changes materially. Zero engineering effort — a documentation/practice change, already technically supported. - Add a scheduled-run wrapper (cron, or the team's existing CI) around
bin/vp-sandboxed run <target> --auto-focus --streamfor targets under active development, feeding results into the existing/triage --autocross-run aggregation. Medium effort, directly enables the repo's own recommended "incorporate scans into your SDLC" adoption path.
4. Estimated impact
| Change | Time saved per cycle | Effort to implement |
|---|---|---|
| Persistent sandbox host | Removes ~5–15 min setup per scan session (image build + gVisor checks) | Low |
| Enforced small-first-wave gate | Avoids re-running a mis-calibrated large batch (potentially hours + token spend) | Medium |
--no-reattack tagging |
Prevents shipping under-verified fixes (risk reduction, not time) | Low |
| Shared triage/patch queue | Reduces status-chasing between analyst/engineer/owner | Low–Medium |
Stable-target focus_areas |
Skips a recon agent invocation per run (minutes + tokens) | Zero (practice only) |
| Scheduled scans | Converts ad hoc scanning into steady-state coverage — not a per-cycle time save, but closes the "we only scan when someone remembers" gap | Medium |
5. Implementation plan
- This week: Adopt the practice changes with zero engineering cost — hand-written
focus_areasfor stable targets, explicit--no-reattacktagging discipline in patch PR descriptions. - This sprint: Stand up the persistent sandbox host; write the small-first-wave gate wrapper script.
- Next sprint: Evaluate whether the team's existing project tracker can host the
triage→patch→review queue, or whether a simple
QUEUE.mdconvention is sufficient given current scan volume. - Ongoing: Once scan cadence stabilizes, add the scheduled-run wrapper and fold its output into the same triage/patch loop.
Related: process-operate-dcrh.md · runbooks/.