Agents
What each agent is allowed to do, when it last actually ran, and what it costs.
These agents draft; they never decide. Every permission set below excludes push, merge, approve and deploy —
enforced by the GitHub token, not by the prompt. Re-run
scripts/validate-agent-scope.sh after any permission change
and file the result in evidence/.
siglos-depscan-bot
healthyDaily dependency and vulnerability scan
| Last run | 3h ago |
| Last status | no_findings |
| Expected cadence | 24h |
| Model | claude-haiku-4-5-20251001 |
| Permissions | metadata:read, contents:read, security_events:read, issues:write |
| Cannot | push, merge, approve, deploy, read secrets |
siglos-status-bot
not reportingWeekly status report
| Last run | never |
| Last status | — |
| Expected cadence | 168h |
| Model | claude-sonnet-5 |
| Permissions | metadata:read, contents:read, pull_requests:read, issues:write |
| Cannot | push, merge, approve, deploy, read secrets |
API spend, last 30 days
$0.38
use this real figure in the business plan, not an estimate
Runs, last 30 days
5
Failed runs recorded
1
of the last 5
Run history
| Agent | Run | Started | Status | Findings | Tokens | Cost | |
|---|---|---|---|---|---|---|---|
| siglos-depscan-bot | run-0041 | 2026-08-13 12:56Z | no_findings | 0 | 18400/900 | $0.0721 | Output |
| siglos-docreview-bot | run-0040 | 2026-08-13 06:56Z | success | 3 | 31200/4100 | $0.1608 | Output |
| siglos-depscan-bot | run-0039 | 2026-08-12 15:56Z | no_findings | 0 | 18100/870 | $0.0709 | Output |
| siglos-depscan-bot | run-0037 | 2026-08-11 15:56Z | success | 1 | 17900/1600 | $0.0812 | Output |
| siglos-docreview-bot | run-0038 | 2026-08-11 15:56Z | error | 0 | 0/0 | $0.0000 | Credential values did not match their labels - halted before writing. Defect F4, fixed in fix/agent-workflows. |
How agents report
Add a final HTTP Request node to each workflow:
POST http://localhost:8787/api/agent-runs
Authorization: Bearer <AGENT_INGEST_TOKEN>
Content-Type: application/json
{
"agent": "siglos-depscan-bot",
"run_id": "DEPSCAN-2026-08-13",
"started_at": "2026-08-13T06:00:00Z",
"status": "success",
"findings_count": 3,
"model": "claude-haiku-4-5-20251001",
"input_tokens": 4120,
"output_tokens": 830,
"artifact_url": "https://github.com/.../issues/42"
}
The ingest token authenticates the workflow, not a person. Agent runs are recorded with
actor_type=agent in the audit log so machine activity is never mistaken for human activity.