Repository navigation
Releases
NetOps 1.0.11
NetOps 1.0.11 follows NetOps 1.0.10. It removes the remaining conversion paths for data from before NetOps 1.0 (Auditor migrate-store and migrate-suppressions, Admin plans of 0.x), returns a failed Junos change at once with a plain commit that the device timer cannot undo, makes the release signing tool print its whole check table when the release manifest lacks a key, limits the release tooling to the 1.0.x series and describes an installation without sudo rights.
Changes in 1.0.11
Removed
- Auditor no longer converts stores and suppression files of the format before NetOps 1.0: the operator commands
auditor migrate-storeandauditor migrate-suppressionsare removed, and the Auditor command line refuses them (invalid choice, exit status 2). A store without a schema version or with a schema version other than 2 is refused before any write, also before the switch to WAL, and its file stays unchanged:store schema version N, expected 2; this store format is not supported, orstore holds audit tables without a schema version, expected 2; this store format is not supported. The Auditor MCP server refuses such a store withthis store format is not supportedinstead of pointing to the conversion. A suppression file of version 1 is refused withunknown suppression file version 1, expected 2. Stores and suppression files written by NetOps 1.0.x have version 2 and are not affected. Covered by automated tests only; no installed instance with an old store was checked. - Admin refuses a plan written by Admin 0.x: the content-only
plan_sha256of plans and journal records before 1.0.10 is accepted only from Admin 1.0.x, so such a plan is refused withplan content does not match its plan_sha256. Plans and journal records of 1.0.x before 1.0.10 stay readable for verify, recover and undo. Covered by automated tests only.
Changed and fixed
- Admin Junos: the immediate return after a failed write or readback now commits the inverse lines with a plain commit (
netops-<id>-revert) instead ofcommit confirmed. Measured live on vJunos-switch 26.2R1.7: the expiry of acommit confirmedtimer returns the configuration before the newestcommit confirmed, not before the first one. In 1.0.10, an immediate return interrupted before its confirmation (process killed, unreadable snapshot, unproven confirmation) therefore let the timer bring the change back; Admin reportedrevert-failedorunknown, neverreverted, but the port stayed changed.admin recovertakes the own-revertrow as the end of the timer. The Admin configuration guide adds a manual check afterunknownorrevert-failed. Verified with automated tests on measured device outputs and one live run of the operator command line on the lab switch (forced readback failure, returned withreverted). - Release signing tool (
release/sign/netops-release-sign.ps1): a release manifest without a key, for example without the arm64 platform or one of its files, no longer stops-CheckOnlywith a PowerShell error. The checks of that key are FAIL rows with the reason, the other checks still run, the whole table andOVERALL: FAILare printed, and nothing is signed. A missing checksum list, more or less than one release manifest, an unreadable manifest and an unsupported version are FAIL rows as well; without a verified checksum list no other file is read, andOVERALL: FAILis printed without a checksum. The tool signs only the 1.0.x series from 1.0.11 on; 1.0.10, 1.0.9, 1.1.0 and 0.x are refused withthis signing tool signs the netops 1.0.x series from 1.0.11 on, not <version>. Verified with the signing tool tests. The-CheckOnlyruns in PowerShell 7.4 and Windows PowerShell 5.1 on the 1.0.10 release files used a development state of the tool that still accepted 1.0.10: the complete files passed every row, and a copy without the arm64 platform or without its gate file printed the FAIL rows andOVERALL: FAIL. The tool in this release refuses the 1.0.10 files. No signature was made. - Release tooling: the candidate manifest, the placeholder gate and the source gate accept only the product versions 1.0.x, and the daily CVE workflow stops before any download when the latest release tag is not 1.0.x. The standalone QA packaging scripts of the component releases before 1.0 (release artifacts, SBOM) and the checks that only they used are removed; the AST, SSH, catalog, privacy, import and profile checks of the components are kept.
- Tests: a new contract test of the optional Admin adapter hooks and new response redaction tests against a frozen copy of the 1.0.8 proxy; the Admin MCP test no longer uses a deprecated FastMCP attribute; the Helper public-source check no longer reports a vendor name inside a hyphenated field name; the Auditor coverage baseline follows the removed code. The unused Helper constant
ENVELOPE_VERSIONis removed, and the message of the Admin and Auditor MCP servers without fastmcp names the complete NetOps image instead ofrequirements-mcp.txt. Product behaviour unchanged. - Documentation: quickstart (Docker set up by the host administrator; step 4 without sudo:
jqfrom the host administrator and cosign as a program of the own account in~/.local/bin, checked withsha256sum;jqas a prerequisite of step 9; the Compose messages on standard error explained before the step 8 block; the CA export hint ofinitandconnectionis the terminal form; files withoutVerified OKare not a verified release), README (the same notes on sudo and Compose, ExtremeXOS port admin state and PoE cycle in the Admin scope), Admin configuration (Junos immediate return, manual check afterunknownorrevert-failed), verified support (Junos measurement). The cosign block without sudo was run once on Debian 13 lab machines (amd64 and arm64) under an account that also had sudo rights; the block does not use them. - Update version metadata, the 1.0.10 → 1.0.11 upgrade pin and the upgrade guide (Upgrade from 1.0.10).
The classifier replay baseline (components/netops-helper/tests/fixtures/output_status_baseline.json) holds 1057 captured texts.
1.0.11 has no MCP API change; the empty declaration is release/acceptance/api-changes-1.0.10-1.0.11.json.
Security
The security-related changes of 1.0.11 make returns and refusals stricter; they do not add a command or write path. A failed Junos change that Admin returns at once is no longer brought back by the device timer when the return is interrupted: the inverse is committed with a plain commit, which ends the own timer. The signing tool never signs when a check fails, also when the release manifest lacks a key, and signs only the 1.0.x series from 1.0.11 on. The Auditor refuses a store in an old format before any write instead of converting it. Admin Junos writes still go through the existing preview, plan_sha256, enrollment, journal and audit controls. These notes do not assert the result of the review or scan of the 1.0.11 release candidate.
Known dependency findings
1.0.11 changes no dependency, lock file or image base. The findings and their assessments are those listed in the 1.0.10 release notes unless the release scan shows otherwise. The release scan of both architectures and the assessments covered by the signed checksum list are part of the candidate acceptance, and these notes do not assert their result.
Upgrade from 1.0.10
Follow Upgrade from the previous release:
- Check the 1.0.10 instance, stop it and create its backup and state manifest.
- Download and verify the 1.0.11 release files. Prepare a separate instance directory and a new Compose project with the same server name and port, and download the 1.0.11 image.
- Restore the backup into an empty target volume and compare it with the source manifest before the first start. Keep the original instance and backup.
- Start the new instance. The first start converts only the recorded product version from 1.0.10 to 1.0.11 under the instance lock and writes an
instance_startaudit record withupgradefrom 1.0.10 to 1.0.11. A second start skips conversion. CA, keys, token, role data and connection details are unchanged.
Do not run init on the restored state. Only 1.0.10 is accepted as the previous release; upgrade 1.0.9 to 1.0.10 first.
For rollback, stop 1.0.11, preserve its audit records written after the backup, and start the unchanged 1.0.10 instance; see Rollback.
What changes for operators:
- Auditor:
auditor migrate-storeandauditor migrate-suppressionsare gone; scripts that call them fail. A store or suppression file in the format before NetOps 1.0 is refused without a change (see above). A store that still needs the conversion must be converted withauditor migrate-storeof 1.0.10 before its backup. - Admin: a plan written by Admin 0.x is refused with
plan content does not match its plan_sha256; run a new preview. - Admin Junos: no change of the device entry is needed. After a Junos operation of 1.0.10 that ended
revert-failedorunknown, check the port as described in the manual check of the Admin configuration. - Enrollments made with 1.0.10 stay valid after the upgrade: an enrollment is bound to the enrollment protocol and to the SHA-256 of every Admin profile file, and 1.0.11 changes neither, so a device needs
admin enrollagain only when its device entry, its accounts, its firmware or the Admin notification or status file settings change.
Known limitations
- Admin Junos: the immediate return with a plain commit was run live once (a forced readback failure); an immediate return interrupted after its commit and an unreadable snapshot after it are verified with the simulated device ...
NetOps 1.0.10
NetOps 1.0.10 follows NetOps 1.0.9. It brings Arista EOS, FortiADC, Cisco IOS, IOS XE, NX-OS and Juniper Junos into the common Admin, makes FortiOS config_diff line-based, binds Admin plan_sha256 to the device and firmware, runs the Auditor store in WAL mode, tightens the ExtremeXOS PoE cycle safeguard budget and recover, and hides typed Cisco-style hashes and whole account values in responses while keeping everything that 1.0.8 hid hidden.
Changes in 1.0.10
Added
- Admin supports Arista EOS (cEOSLab 4.36.1F) through the shared Admin path: preview with
plan_sha256, enrollment, journal and audit, check-account readback, undo, recover and the existing MCP tools (no new tool). Operations: interface description, admin up/down and access VLAN, written line by line throughconfigure terminalunder the device-sidecommit timerof an empty EOS configuration session that Admin commits before the change; the timer reverts an unconfirmed change. Admin needs privilege 15 withoutenableand a read-only check account without write rights; see Admin configuration. A change of the admin state of a port, in both directions, is refused on trunks/uplinks, LAG members, routed ports, management interfaces, ports with an LLDP neighbour or learned MAC addresses. Measured on one lab switch (GNS3 cEOSLab, 2026-11-15, tasks 301 and 304): 13 of 13 end-to-end CLI scenarios; the MCP tools were not called live. A later fix (the same port checks also for admin up) is covered by automated tests only and was not repeated live. The configuration sessioncommit timerreturn was measured at 120 s after a missed confirmation; behaviour after a device reload,aaa authorizationvariants other than the lab one andenablepassword use are not measured. - Admin supports FortiADC (load-balance pool member state, add and remove) through the shared Admin path: preview with
plan_sha256, enrollment (admin enroll --probe POOL:MEMBER:REAL_SERVER:PORT), journal, audit, check-account readback, undo, recover and doctor. FortiADC 7.6.7 has no on-device timed return, so an unconfirmed write is reverted by Admin itself with an inverse write in the same run, or byadmin recoverafter an interruption; if the Admin process or the management path is lost after a write, nothing on the device reverts the change until recover runs. A virtual server must keep a healthy enabled member (no_healthy_member). Measured on FortiADC-KVM 7.6.7 build0676 (lab VM220, 2026-11-16, task 306): 13 of 13 end-to-end scenarios including SIGKILL and recover; the configuration hash was equal before and after. A timed return throughalert-policywas found in documentation only and was not measured. Auditor rules are not evaluated for FortiADC in Admin (audit_unevaluated). The olderpool-member*CLI keeps its output and is marked deprecated. - Admin refuses an ExtremeXOS PoE cycle whose safeguard budget is too short: the cycle needs 175 s plus
confirm_margin_secondsbelowsafeguard_seconds, checked from the configuration in preview/apply and again against the switch clock before the power is cut (safeguard_margin_insufficient; unreadable clock:device_clock_unreadable). With the defaults (180 s, margin 45) a PoE cycle is now refused; setsafeguard_secondsto at least 221. The 175 s is a conservative budget from measured read times, not a measured duration. admin recoverof an interrupted ExtremeXOS PoE cycle with the safeguard still installed confirms the cycle only when the journal and fresh reads prove it (power off, on, configuration equal to the state before); otherwise the safeguard fires and the result isreverted(PoE cycle not proven). Covered by automated tests with measured device outputs; no live cycle was run.- Admin supports Cisco IOS 15.x and IOS XE 17.x (platforms
iosandiosxe, one adapter) through the shared Admin path: preview withplan_sha256, enrollment, journal and audit, check-account readback, undo, recover and the existing MCP tools (no new tool). Operations: interface description and admin state, and the access VLAN on the switch images, written line by line under the device revert timer (configure terminal revert timer), whose expiry returns the whole running configuration; confirmation isconfigure confirm, and a failed write or readback is returned at once withconfigure revert now. The profiles list exact versions, measured on GNS3 lab images only: IOSv 15.9(3)M12, IOSvL2 15.2(20200924:215240), IOL XE 17.18.2 and IOL-L2 XE 17.18.2; no hardware platform was measured. The device needs the configuration archive (archive_not_configured), the write account privilege 15 at login (Admin never sendsenable) and a separate privilege-1 check account; see Admin configuration. The timer is treated as Admin's own only with proof from Admin's own session, its time left is counted on the monotonic clock of Admin (the device clock is never read), and waiting andadmin recoverread the timer state.safeguard_secondsis rounded up to whole minutes, at least 2 (2–15). Measured in the GNS3 lab on 2026-11-16: the operator command line scenarios on all four nodes (source3dd813e) and, after later fixes (sourcef0e8b0d), on IOSvL2 Gi1/0 and IOL XE Et0/2: enrollment self-test, apply with--expect-plan-sha256and undo, a forced readback failure returned withconfigure revert now, SIGKILL andadmin recover; the running configuration was equal before and after. The MCP tools were not called live. - Admin supports Cisco NX-OS 9.3(12) (platform
nxos) through the shared Admin path: preview withplan_sha256, enrollment, journal and audit, check-account readback, undo, recover and the existing MCP tools (no new tool). Operations: interface description, admin state and access VLAN. NX-OS 9.3(12) has no commit timer, so the safeguard is a checkpoint with a one-timefeature schedulerjob that rolls the whole running configuration back at its minute;feature schedulermust be enabled once by the switch administrator (scheduler_feature_disabledotherwise), and Admin never enables or disables it. A change that Admin finds unverified is returned at once byconfigure replaceof a copy of the running configuration saved before the change. Admin removes only objects whose creating line its own session saw accepted, removes an unverified or interrupted installation at once (the pending schedule first) and never reads the device clock. With the default margin of 45 s,safeguard_secondsmust be at least 180. Write account measured with rolenetwork-admin, check account withnetwork-operator; see Admin configuration. Measured on a Nexus 9000v (N9K-C9300v, virtual) with NX-OS 9.3(12) in the GNS3 lab on 2026-11-16: the operator command line scenarios with sourcea36fd05and, after later fixes (source7af8285), 9 of 9 scenarios on Ethernet1/48:doctor, enrollment self-test, apply with--expect-plan-sha256and undo, a forced readback failure returned byconfigure replace, SIGKILL andadmin recover; the normalized running and startup configurations were equal before and after. The MCP tools were not called live. - Admin supports Juniper Junos 26.2R1.7 on ELS switches (platform
junos) through the shared Admin path: preview withplan_sha256, enrollment, journal and audit, check-account readback, undo, recover and the existing MCP tools (no new tool). Operations: interface description, admin state and access VLAN by VLAN name; the profiles list26.2R1.7as an exact version, so every other Junos release is refused before any write. Admin changes the configuration only inconfigure exclusiveunder a device-sidecommit confirmed 10safeguard: an empty safeguard commit, the change committed withcommit confirmed 10(which restarts the timer) and, after the readback of the separate check account, a confirmation commit; an unconfirmed timer returns the whole configuration, including a commit of someone else in the window. The timer is treated as Admin's own only with proof from Admin's own session (its own commit rows inshow system commit, recorded as the head of its timer), a timer of someone else is never committed over, and its time left is counted on the monotonic clock of Admin (the device clock is never read). A failed write or readback is returned at once by the inverse lines of the plan committed under the own timer, then confirmed only when a fresh snapshot equals the state before; this holds also when the first read after the own change commit failed and a second read proves that commit. Waiting andadmin recoverdecide by the commit history only; recover of an interrupted operation sends no commit and waits for the expiry. The write account needs the login-class permissionsconfigure,viewandview-configuration(write_account_privilegeotherwise), the check account the built-in classread-only; see Admin configuration. Measured on vJunos-switch 26.2R1.7 (ELS, GNS3 lab, test port ge-0/0/2) on 2026-11-16: the operator command line scenarios with sources5cb02e9,2a78465and65dbe31(task 319:doctor, enrollment self-test, description apply with--expect-plan-sha256and undo, admin down and up, access VLAN and its undo, SIGKILL andadmin recover, refusalsvlan_missingandport_management_interfacewithout a write) and, after later fixes (source4699b25), enrollment self-test, description apply and undo, a forced readback failure returned by the immediate inverse return about 4.5 minutes before the expiry, and SIGKILL withadmin recover; the configuration was equal before and after both runs. A later fix (source6a39724...
NetOps 1.0.9
NetOps 1.0.9 follows NetOps 1.0.8. It adds configuration captures and config_diff for ExtremeXOS, Cisco and Arista, FortiOS IPv6 route lookup and multi-VDOM reads, NX-OS vPC and EOS/EXOS MLAG views, ExtremeXOS optics and chassis summaries, Wi-Fi AP backhaul correlation and FortiADC schedule-pool triage. The Auditor gains findings_summary and collection while NetOps runs; Admin gains expected_plan_sha256, admin_undo_preview, the measured FortiOS firewall service custom profile and ExtremeXOS port admin-state and PoE cycles.
Changes in 1.0.9
Added
- Helper can correlate a tunneled Wi-Fi client with its AP/radio and the AP's switch backhaul. The switch port belongs to the AP, not to the tunneled client; missing station, VLAN or forwarding evidence remains explicit.
- Configuration captures add narrowly gated ExtremeXOS, Cisco IOS, IOS XE, NX-OS and Arista EOS full views. Each requires explicit configuration access, a supported live identity and device permission for the fixed command. Allowlisted visible lines survive; credentials and unclassified lines become markers. Device refusal returns
device-refused, never an empty snapshot. See identities and account requirements. config_diffcompares snapshots on those five platforms with a shortest line edit script (Myers). Limits of 3000 changed lines, bounded comparison work and configuration memory refuse oversized comparisons withconfig_diff_limit_exceeded. Hidden content is not compared (hidden_lines_compared: false); changes to hidden-line counts do not invalidate an otherwise comparable pair. Incompatible scope, firmware or profile returnsconfig_snapshot_incomparable. FortiOS/FortiADC retain positional comparison.- ExtremeXOS optics/DOM and chassis summaries expose the measured metrics and partial evidence; physical DOM evidence is limited to X440G2 and does not qualify X435 optics or virtual hardware sensors.
- FortiOS
ipv6_route_lookupadds scoped IPv6 destination lookup with explicit identity and parsing limits; it does not add IPv6 routes to Linux normalization. - NX-OS vPC and EOS/EXOS MLAG summaries and members accompany the fixed
ssh_readcatalogue queries. They are notswitch_readtables; active peer topologies remain documentation-derived where captures only show inactive, empty or refused states. - FortiOS multi-VDOM catalogue, normalized-table and routing/VPN/health triage reads verify the session VDOM and context, label
data_scope, and fail closed on ambiguous identity, context or interface ownership. Only the enrolled current VDOM is read. Device-wide session counters stay device-wide; multi-VDOMconfig_readwithredaction="minimal"andtriage_flowremain unsupported. A per-VDOM administrator logged in through an interface of its VDOM reads that VDOM directly once the device proves the login context, and reads beyond its VDOM returnvdom_admin_global_scope_unavailable(verified live on the FortiGate-60F 7.6.8 build3744 on 2026-11-15, task 222, source 7bec0b7, 112/112 checks passed). Fullconfig_readreads the global/VDOM structure as visible to the account (config_scope: device-account-visible,vdom-coverage-not-verified), without selecting a VDOM context or assigning the capture to a single VDOM. See scope and refusals. - FortiADC SLB triage follows
schedule-listthrough schedule-pool and schedule-group views, returning sourced candidate pools. Statistics may name an old stored pool, so the result does not claim an active pool, historical selection or backend health from those references. Schedule-window activity is not calculated; content routing remains unsupported. - Auditor supports
collect,runandstatusthroughdocker compose execwhile NetOps runs. A private store lock serializes writers; MCP reads see a committed state. Collection/run write preflight and result audit records. Configuration and imports still require the stopped instance, and commands share container resource limits. - Auditor
findings_summaryreturns brief last-run findings across devices with severity/state/rule filters, bounded output, truncation and counts before/after filtering. Without a rule filter its totals match per-devicelist_findingstotals under the same severity/state filters. It has no paging or historical detail.audit_statusandlist_findingsalso returnsuppressions_configured; it isfalsein 1.0, which has no supported way to supply a suppression file. - Admin
expected_plan_sha256optionally binds apply to the approved preview: a fresh substantive plan mismatch is refused withplan_changedbefore writes. It does not replay preview commands and does not apply to schema transactions. - Admin
admin_undo_previewand CLIadmin undo --previewshow checks and impact without applying undo; actual undo remains an operator action. After an undo has run, undo preview is refused withundo_already_done; repeating the identical operator undo returns its stored result. - Opt-in FortiOS
firewall service customsupports create/update/delete and undo, measured end-to-end on 7.6.8 build3744 (cli-script) and 8.0.2 build0251 HA (webhook). The profile is measured and no longer provisional. Referenced/protected services are guarded. Policy enable/disable continues to use the existing policystatusupdate. Service safeguard POST rollback, SIGKILL/recovery and whole-management-path loss remain unmeasured for this table. - ExtremeXOS port
admin-stateupdates and operator-only PoE cycles use the persistent UPM safeguard with operations gated to ExtremeXOS 33.7.1.6 (patch level and model are not checked); measured only on 33.7.1.6-patch1-11 on X435-8P-4S/X440G2. Unreadable LLDP/FDB usage refuses disabling withport_usage_unreadable; a non-PoE port returnspoe_not_capable(on the switches task 184 still gotunspecified_refusal, fixed in task 188 and covered by automated tests, not repeated live). Device commands were measured on X435/X440G2 (task 171); Admin disable/undo, safeguard revert after loss of the check account and a PoE cycle ran on the two switches on 2026-11-14 (task 184, source 72c1fca; 9 of 10 scenarios passed), with the check account reading the PoE state. Real powered-device recovery time, thedeliveringstate andport_usage_unreadableon a switch remain unmeasured.
Changed and fixed
- Use the current Compose
no-new-privileges=truesyntax. Refresh the pinned Python/Alpine base digest and compatible dependencies (FastMCP 4.1.3, PySNMP 7.1.32, cryptography 50.0.2). This does not claim that every dependency vulnerability is fixed or that release scans/signatures are complete. - Release assessment matching preserves package version and type, preventing another version's assessment from being overwritten during merging.
- FortiOS multi-VDOM reads refuse an enrolled VDOM with
vdom_not_foundonly whenget system statuscounts exactly one VDOM (the session's own), otherwise stillvdom_scope_required, and theauthadvice of a FortiOS target withdiagnostic_vdomsadds that a per-VDOM administrator must log in through an interface of its own VDOM, on every read path including table reads (switch_read,locate_endpoint). - FortiOS flow/policy triage accepts an unknown source port without inventing one, and reports ambiguity when the available evidence cannot choose a policy.
- Preserve the earlier response-redaction protections, including Unicode whitespace/line-boundary cases. Failed secondary reads in normalized-table joins retain their error classification. CDP recognizes record boundaries, advertised totals, non-Cisco remote-port IDs and phone TLVs; Linux stateless/stale FDB entries normalize as dynamic. An Auditor store mismatch message identifies the Auditor workflow.
- Reconcile support documentation with the carried-forward read-only EXOS acceptance of 1.0.8 and restrict FortiSwitch claims to measured 224E-POE operations and topology. Correct endpoint dates, EXOS naming, the
lag_summaryrequirement and the measured SLB durations (about 61/71/81 seconds: healthy HTTP virtual server, a DOWN member, HTTPS/TLS virtual server; FortiADC VM220, 2026-11-10). These are not upper bounds: a virtual server withschedule-listtook 81–86 s (task 198, 2026-11-14). Document Linux VXLANdstFDB andsit0link forms asunparsedcoverage limitations. - Update version metadata, the 1.0.8 → 1.0.9 upgrade pin and the upgrade guide (Upgrade from 1.0.8).
The classifier replay baseline (components/netops-helper/tests/fixtures/output_status_baseline.json) holds 1057 captured texts.
The MCP API changes of these items (new tools, new arguments and changed role responses) are declared in release/acceptance/api-changes-1.0.8-1.0.9.json.
Security
The security-related changes of 1.0.9 tighten what NetOps returns and which scope it reads; they do not add an arbitrary command or write path. Response redaction keeps every character that the released 1.0.8 chain hid, including Unicode whitespace and line-boundary cases; where the current chain cannot reproduce that floor, for example for a JSON document nested in a string value, the output is withheld instead of returned. The new ExtremeXOS, Cisco and Arista configuration captures return only allowlisted visible lines: credentials and unclassified lines become markers, NX-OS and EOS captures also hide recognizable password hashes and Cisco type 7 values (including a value followed by punctuation), and config_diff does not compare hidden content. A FortiOS per-VDOM administrator reads only its own VDOM once the device proves the login context, global and cross-VDOM reads return vdom_admin_global_scope_unavailable, a request for another VDOM is refused with vdom_scope_required (or vdom_not_found when the device proves that VDOM absent), and...
NetOps 1.0.8
NetOps 1.0.8 follows NetOps 1.0.7. It adds CDP neighbours and Linux tables to switch_read, makes locate_endpoint read Linux neighbour and bridge tables, fixes FortiOS route and VLAN parsing, hardens the Auditor store and the two HTTP response readers, makes the Admin Doctor read the journal once per request, and documents the minimum requirements and the platforms on which the installation procedure was run. The product remains one container with Helper, Auditor and Admin behind one audited HTTPS MCP endpoint. The syslog audit format and receiver requirements are unchanged.
Changes in 1.0.8
switch_read(table="cdp_neighbor")returns the CDP neighbours of Cisco IOS, IOS XE and NX-OS from the opt-incdp_neighborsquery: local port, full Device ID, neighbour port, management IP, platform and CDP capabilities as text (not LLDP letters). Unrecognized lines and incomplete blocks lowercoverage; a target without the enabled query is refused before any transport.- Linux gains
switch_readtables from its existing catalogue queries:interface,ip_interface,route(IPv4),neighbor_l3andmac_entry(bridge_fdb). A freshuname -averifies Linux before every read. MTU, MAC and state of one interface need the opt-ininterface_linkquery with the interface enrolled, thenswitch_read(table="interface", interface=...). The Linuxaddressescommand and description stay unchanged from 1.0.7 (ip -brief address): IP and prefix come from this brief output; enabledlinksadds MAC and administrative state and enabledroutesadds the IPv4 default gateway in the sameswitch_read. Unsupported route lines are reported byrouteand do not discard a recognized default gateway fromip_interface. Linux has no VLAN or LLDP table. Unknown lines lowercoverage. - For Linux targets, fresh
ssh_readcalls also verify identity withuname -abefore the requested query (one extra SSH connection, except for thekernelquery, whoseuname -ais the probe itself). An invalid Linux identity is refused withproduct_identity; responses includenormalizedandstatus. The identity probe is not charged separately to the per-target rate limit: a freshssh_readcosts one slot; cached continuation pages cost none. - New optional fields:
interface.mac,interface.mtu,ip_interface.macandip_interface.gateway(the IPv4 default gateway via the same interface on Linux whenroutesis enabled). They are null on platforms that do not supply them and do not lowercoveragethere. Declared inrelease/acceptance/api-changes-1.0.7-1.0.8.json. locate_endpointreads the Linuxneighborsquery (IPv4 and IPv6) onrouterandfirewalltargets andbridge_fdbonswitchtargets. A neighbour proves no edge: when a Linux host has the endpoint as a neighbour but no matching bridge port, the result staysincomplete(bridge_port_not_observed; without any switch readerno_enrolled_switch_reader) unless another enrolled switch locates the endpoint, and missing LLDP evidence leaves a portunverified.- FortiOS routes with recursive next hops (BGP) and summary or Null routes are parsed, the route code legends of FortiOS 7.6 and 8.0 are accepted, and
setlines directly inside a nested configuration node of the interface configuration (such asconfig ipv6) no longer make VLAN entries unparsed. - A FortiGate Wi-Fi client in tunnel mode (the SSID default) is expected not to be learned on the switches between its access point and controller because its traffic is carried in CAPWAP to the FortiGate. This is consistent with the lab result (
not_found_in_checked_data), but the SSID mode and cause were not confirmed. Locally bridged SSIDs were not tested. Usewifi_clientswith the client MAC orwifi_triageto find the access point and radio, then locate the access point. compose.yamltakes the CPU limit fromNETOPS_CPUS(default 2). Docker refusescpus: 2on a host with one CPU; setNETOPS_CPUS=1there. See Minimum requirements.- Wi-Fi REST (Helper) and the FortiOS REST collector (Auditor) reject a chunk-size line that is not plain hexadecimal (
+61,6_1,0x61, blanks,-0) before any body is processed. The Python 3.14.10 standard-library parser still accepts these spellings (CVE-2026-103995), so NetOps checks the line itself; valid chunks, chunk extensions and the byte budget behave as before. Chunk-size lines must end with CRLF; bare LF and whitespace before;(for example61 ;ext) are refused. Regression tests cover both readers. - The Auditor store validates the type of every expected schema object (table, index) when it is opened, migrated or read read-only. A view or other object in place of an expected table is refused with a message naming the object; the schema NetOps creates is unchanged. A store missing an expected table or index is refused by the read-only server with a message naming the object; open it writable (
auditor runorauditor status) to restore it. - The local part of Admin
doctor(also run bydoctor --local) validates and parses each journal record once per request instead of three times. Order, rejection of a corrupted journal and the fresh evaluation of the time budgets are unchanged, nothing is kept between requests, andapplystill reads the journal itself. - Response redaction spans now follow
sort_keysordering and tuples. A direct serialization test of these shapes left a marked synthetic value unmasked on the 1.0.7 code; no shipped tool response path producing these shapes was identified in the review. Regression tests also pin the value-lessunset passwordexception and the audit class of a failedswitch_read. - Wi-Fi triage documentation states what a failed or empty last signal sample leaves in
chain,missingand the verdict, how the 15-read REST reservation behaves, and when an old client error is only a historical notice. - Release tooling: a fixable High or Critical scan finding passes the release gate and the daily rescan only with a dated assessment that records that the fixed version is not yet released, with a source link. Version metadata, the 1.0.7 → 1.0.8 upgrade pin and the build lock (
setuptools84.2.1, build time only) are updated; the base image and all runtime packages are those of 1.0.7. CI runs the four-role engine regression and keeps the full logs. The classifier replay baseline now holds 1029 captured texts.
What was checked on devices and lab images
These results come from development builds before the release candidate; device reads used read-only accounts, while Admin used its separately enrolled accounts. They do not qualify a later rebuilt image.
- CDP: the output of
show cdp neighbors detailof Cisco IOS (two images), IOS XE and NX-OS lab images returned the correct local and remote port, neighbour name and management IP incdp_neighbor(4, 4, 3 and 3 neighbours), compared with the CLI. - Linux: five tables (
interface,ip_interface,route,neighbor_l3,mac_entry) of an Ubuntu 24.04.5 host (kernel 6.8.0-147-generic, iproute2 6.1.0), with 3, 6, 4, 17 and 13 rows, matched independent CLI reads, and six further checks of the per-address default gateway also passed. That build readip_interfacefromip address show; the releasedip -brief addressform is covered by the CLI captures of 2026-11-13 below. A neighbour's NUD state can change between reads, so IP, MAC and interface were compared. The forwarding database of that host held only static entries; in the tests a bridge port from a synthetic dynamic FDB row is returned asunverifiedwith the resultincomplete, never as a located edge. - Linux capture replay follow-up (2026-11-13): read-only CLI output from Ubuntu 24.04.5 (kernel 6.8.0-147-generic), a clean Debian 13.7 cloud VM (6.12.111+deb13-cloud-amd64) and Proxmox VE on Debian 13.7 (7.0.14-28-pve) supplies 23 captures and 168 independent manual records. The three
linux_*_batch_b.jsonscenarios are replayed through the real proxy request/response paths and engine, with MAC/prefix/gateway joins on each host (26 tests). Proxmox bridge and bridge-port metadata are captured; VLAN metadata has unit-test coverage only because no VLAN was present, and bond metadata is not recognized and lowerscoverage. The clean Debian VM had nobridgeexecutable, so its FDB capture expects rc 127 /transport_error. This is replay of fresh CLI output, not a new live MCP qualification. - FortiOS routes and VLANs: the live reads of a FortiGate 80F (FortiOS 8.0.2) and a FortiGate 60F (FortiOS 7.6.8) first showed the unparsed route and VLAN shapes that this release fixes. The fix is checked by replaying the stored CLI captures (32 and 13 routes, 7 and 9 VLANs, no unparsed line); the 80F VLAN capture is a reconstruction from the independent
get system interfaceoutput, not a rawshow system interface. These replays are not a new live run. - Wi-Fi: on a FortiGate with two access points,
wifi_clients,wifi_aps_live,wifi_rogueandwifi_triagematched the independent REST and CLI reads. The client's own MAC was in neither of two switch forwarding databases, which is consistent with the tunnel-mode behaviour above; the SSID mode itself was not recorded. The failed-sample variants of the Wi-Fi triage documentation are covered by tests, not by a live run. - Admin Doctor:
doctor --localand fulldoctorpassed on runtime 1.0.8 configured for a FortiGate 60F in the G3/G4 live round; the local command checks local state without contacting the device. - HTTP reader coverage: the G2 Wi-Fi round exercised the Helper REST reader containing the chunk-size validation check, on dev-1.0.8 against a FortiGate with two access points. Successful ordinary REST reads are documented; the evidence does not record whether the responses were chunked. Rejection of malformed chunk-size lines is covered by tests, not a live-d...
NetOps 1.0.7
NetOps 1.0.7 follows NetOps 1.0.6. It adds read-only Wi-Fi and flow diagnostics, switch/router profiles and normalized tables, improves Auditor coverage reporting and Admin diagnostics, and fixes paged-output secret disclosure. The product remains one container with Helper, Auditor and Admin behind one audited HTTPS MCP endpoint. The syslog audit format and receiver requirements are unchanged.
-
Correlate late ARP/ND MACs against saved endpoint MAC tables and reuse auxiliary captures; expose per-alias read caps, cache failed hostname supplements and avoid ownership retries after transport failures. Preserve FortiSwitch all-VLAN configuration with explicit FortiLink context, clear NX-OS connected/local IP gateways and return only active Junos detail routes, consistently with whole-table reads.
-
Clarify that Auditor exit 3 is driven by required rule coverage, not every export visibility limitation: networking can be actionable while an unrelated user-area limitation can permit exit 0; retain separate export metadata checks.
-
Correct the public SLB triage duration to the measured HTTP, DOWN-member and HTTPS/TLS scenarios (about 61/71/81 seconds), explain timeout headroom and optional progress, and refresh flow-session and SSH Wi-Fi join-time provenance notes.
-
Wi-Fi triage distinguishes exact VAP ids from broadcast SSIDs, preserves current REST and matching event ids, and reports ambiguous SSID profiles as missing evidence instead of a false VAP-down fault.
-
Wi-Fi neighbor filtering, ordering and summary signal now use only AP/radio observers with channel overlap; summaries name the observers of that signal and full reads retain all observations.
-
Scale endpoint location to 64 enrolled targets with eight concurrent device readers, a shared deadline and explicit unfinished-target coverage while preserving audited read-only verdicts.
-
Prevent device-output ReDoS in Junos error counters, NX-OS identity, switch tables and wireless parsers with linear token validation and adversarial regressions.
-
Response redaction now masks complete community values using original SNMP context while preserving the 1.0.6 secrecy floor, avoids redundant passes, and reports full-capture device refusals as safe
-32602errors. -
The new OSPF neighbor table uses lowercase canonical
statevalues (full,2way,exstart, etc.). EOS2 WAYS, NX-OSTWOWAYand Junos2Wayall normalize to2way. -
Preserve the full 1.0.6 helper → proxy → worker redaction chain before terminal cleanup and repeat it afterwards; bound ANSI string controls to one line and 4096 characters, retaining the payload of unfinished controls.
-
Name the real cause of every failed SSH device read (
unreachable,timeout,connection_refused,host_key_unavailableonly for a host that answered without a usable key,host_key_mismatch,auth,disconnectedand the other fixed classes) with fixed advice inerror_code/errorofssh_read,schema_read,switch_read,locate_endpoint,config_read,fortios_diagnosticsincluding the FortiADC recipes and the triage tools, and in the Helper auditdetail, instead ofhost_key,ssh_unavailableorRole operation failed. The client's closing message alone decides the class and anylegacy_sshadvice, so a login banner, even one namingno matching, never turns a refused login intonegotiation(also for thesftp_statreason). Every resolver failure of the target name (EAI_NONAME,EAI_NODATA,EAI_FAIL,EAI_AGAINand the others) isdnsby its type, before any text. -
Parse every member of real EXOS sharing groups, attach a safe transport cause to failed normalized reads and their audit, and distinguish enrolled LLDP neighbors with
endpoint_not_seen_on_neighbor_target. -
Permit FortiADC real-server object reads, distinguish local triage refusals from device reads, give health-check follow-ups enough time, normalize EXOS load-sharing groups and document measured SLB durations and call queuing.
-
Optional
helper.fortios_rest_portselects management HTTPS independently of SSH
(default 443, whole number 1–65535); resolver, egress, socket and HTTPS use it,
enrollment refuses a port outside TCP egress, andtarget_scopereports it.
Match FortiGateadmin-sportwhen 443 serves SSL VPN. The Wi-Fi REST guide
retains the administrator checklist and recommends the verified 80F 8.0.2
build0251 minimum:wifi read+loggrp read, other groupsnoneincluding
authgrp, allcli-*disabled; all-read is an alternative. Omitset vdom
for single-VDOM; use it only for multi-VDOM provisioning. Token generation
requires an administrator with API-user rights. -
Wi-Fi triage selects significant history instead of the newest five events, reports absent MACs with history and a cause or unknown, and samples signal three times with min/max/last. It separates client link/signal from shared radio utilization/retries, keeps unknown measurement ages null, and documents per-target serialization and the 15-read REST triage reservation. Successful warning-level handshakes are not authentication failures.
-
Wi-Fi diagnostics now prefer read-only REST wireless history, continue when history is unavailable, enrich client identity and filter exact hostnames. Configurable signal/SNR heuristics include explanatory findings; both triage views show AP/radio RF evidence and explicitly unknown common measurement age with a separate wired/LLDP gap. Link RX/TX rates, traffic and client retry/discard are distinct; relevant neighbor summaries deduplicate BSSIDs and full observations remain available. Client timelines show coverage and downgrade errors older than the last observed association, even if a later snapshot fails. See Wi-Fi REST for the complete contract.
Changes in 1.0.7
-
Accept the real NX-OS trunk forwarding continuation
noneafter the disabled-VTP notice, preserving allowed VLANs and native tagging inport_vlan. -
fortios_diagnosticsnow defaults to a 10-second timeout (previously 5 seconds), allowing FortiADCnslookupto complete a cold DNS cache lookup. -
Hide SNMP communities and auth/priv credentials across FortiOS/FortiADC/EXOS, retain credential keywords and numbered PSK labels without weakening the 1.0.6 redaction floor, and accept real FortiOS minimal syntax with safe
redaction-unverifiableclient errors. -
Normalize multiword EOS and other OSPF neighbor states with five preserved live captures, add field/hint guidance to invalid proxy arguments, and strip ANSI escapes and C0 terminal controls except newline/tab from device output before hashing, paging and client delivery.
-
Parse Junos day/week route ages and wrapped IPv6 prefixes, explain a missing Junos interface_errors selector with -32602 and an enrolled-interface hint, and recognize FortiOS log search progress as status metadata.
-
Fix day-old cEOS LLDP ages, EXOS connected next-hop semantics and IOS/XE error-counter fallback to separately authorized whole
show interfaces, with incomplete empty results and anonymized real-output regression coverage. -
Add
port_vlanaccess/trunk/routed membership tables, read-only EXOS VLAN detail, IOS trunk-aware VLAN membership and endpoint VLAN tagging evidence. -
switch_read(table="interface_errors", interface=...)adds cumulative port error/discard counters and available flap evidence from IOS/XE, NX-OS, EOS, EXOS and Junos, with explicit source coverage and no new permissions. -
FortiOS
switch_readexplains unsupported tables, normalizes IPv4 lookup/table routes and adds enrolled FortiSwitch link/speed, cumulative errors/drops and explicitly configured VLAN membership to endpoint location, verified by replay of read-only FortiGate 80F 8.0.2 build0251 captures. -
IOS / IOS XE empty IPv4 route lookups fall back to the authorized routing table and select the longest matching prefix.
-
locate_endpointrecognizes an enrolled target's own interface from local/interface ARP or already-read interface addresses and returnsown_interface_of_targetwith target/interface evidence. Unresolved transit MACs may check enabled IPv4 ARP within the existing 40-read/60-second budget; GNS3 replay resolves both router IP and MAC in 28 reads. -
Table interface names and filters share canonical long forms: EOS Ma0/Management0, Et/Ethernet, Po/Port-Channel and NX-OS Eth/Ethernet.
-
locate_endpointreads ARP/ND first for an IP and all MAC tables before auxiliary switching reads, then LLDP only on switches holding the MAC and interface, port VLAN, LACP and VLAN reads only for ports without an LLDP switch neighbour, so the seven-node GNS3 lab fits in 28 of the 40 reads. -
locate_endpointfinds the own interface of EOS, NX-OS, Junos, EXOS and FortiOS targets in their enabledip_interfaceread when nothing else located the endpoint, treats a FortiAP LLDP neighbour (B,W,R,Sor a FortiAP/FAP name) as an AP, and on FortiLink reads nothing behind trunks or the switch's owninternalport and the port state and VLAN configuration only for the edge, so the IT-case AP and FortiSwitch MAC need 8 and 3 reads (about 45 and 20 seconds at 5 s per FortiOS read) instead of reaching the 60-second deadline. -
locate_endpointtreats an LLDP neighbour without capabilities whose system name is the host name (sysName, without case or domain) of another enrolled target as that switch, so two enrolled EXOS in a row givetransiton their link andlocatedat the edge port instead ofincomplete; the own addresses of both lab EXOS areown_interface_of_targeton their read-only outputs. -
switch_readtableip_interface(interface, IPv4 address, prefix length, SVI VLAN, admin/oper state, VRF where printed) for IOS, IOS XE, NX-OS (mask from the connected route), EOS, EXOS, Junos and FortiOS (new read-only templatesystem_interfaces:get system interface), with an optionalinterfacefilter; an IOS/IOS XE target declared as router rea...
NetOps 1.0.6
NetOps 1.0.6 follows NetOps 1.0.5. It fixes eight defects in the Helper, the SNMP worker, routing triage and the Auditor and refreshes the container base image. It adds no tool, changes no tool definition and no catalogue, and has no breaking change. The product remains one container with Helper, Auditor and Admin behind one audited HTTPS MCP endpoint. The syslog audit format and receiver requirements are unchanged.
Changes in 1.0.6
- Helper redaction completely redacts quoted secrets that contain escaped quotes; an unclosed quote is redacted to the end of the line (#7.1).
- SNMPv3 authPriv
snmp_getworks in the shipped image; the worker no longer refuses it before the Helper policy (#7.3). - Routing triage (
triage_routing) correlates the next-hop ARP entry with the egress interface (#7.2). - Auditor text output labels not-evaluated findings
(not-evaluated)instead of(gone)(#7.8). - Auditor refuses to write history when the platform changes under the same alias; use a new alias or a separate store (#7.4).
- Auditor finding detail for a carried finding names the run that holds its evidence (#7.6).
- Auditor REST collector refuses an answer with ambiguous framing: conflicting or non-numeric
Content-Length, orContent-Lengthtogether withTransfer-Encoding(#7.5). - Auditor
schema-check --full-snapshot --catalog-vdomsrefuses, before writing the report, a catalog coverage above the reader limit of 1600 entries instead of producing a report the reader rejects (#7.7). - The container base image is refreshed to the current
python:3.14.10-alpine3.24digest. This fixeszlibCVE-2026-4738 (Critical) andopensshCVE-2026-123158 (High), which were reported for the 1.0.5 image after its release; the other scan findings of 1.0.5 remain, because no fixed packages exist upstream yet (#7.9, see Known limitations). - NetOps 1.0.6 upgrades the state of NetOps 1.0.5; see the upgrade procedure below.
See the CHANGELOG for the authoritative list.
Upgrade from 1.0.5
Follow Upgrade from the previous release:
- Check the 1.0.5 instance, stop it and create its backup and state manifest.
- Download and verify the 1.0.6 release files. Prepare a separate instance directory and a new Compose project with the same server name and port, and download the 1.0.6 image.
- Restore the backup into an empty target volume and compare it with the source manifest before the first start. Keep the original instance and backup.
- Start the new instance. The first start converts only the recorded product version from 1.0.5 to 1.0.6 under the instance lock and writes an
instance_startaudit record withupgradefrom 1.0.5 to 1.0.6. A second start changes nothing. CA, keys, token, role data and connection details are unchanged.
Do not run init on the restored state. State of any other product version is refused; there is no skipped-release upgrade (1.0.4 must first be upgraded to 1.0.5).
For rollback, stop 1.0.6, preserve its audit records written after the backup, and start the unchanged 1.0.5 instance; see Rollback. A direct downgrade of a converted volume is not supported.
Known limitations
- Base image CVEs without an upstream fix (#7.9). The scan of 2026-11-07 (Grype 0.121.2, database v6.1.10, amd64) of the refreshed image reports the same findings as the 1.0.5 image rebuilt with
--pull: 0 Critical, 4 High, 11 Medium, 2 Low. Python 3.14.10 is the latest stable 3.14 release, and Alpine 3.24 offers no fixedsqlite-libsorbusybox, so the findings cannot be removed today:python3.14.10: CVE-2025-15367 (Medium, fixed in 3.15.0a6), CVE-2026-116566 (Medium, fixed in 3.16.0), CVE-2026-103995 (Low, fixed in 3.16.0).sqlite-libs3.53.4-r0, no fixed version: High CVE-2026-119586, CVE-2026-119379, CVE-2026-119374, CVE-2026-119373; Medium CVE-2026-119378, CVE-2026-119375, CVE-2026-119371, CVE-2026-119376, CVE-2026-119591, CVE-2026-119372; Low CVE-2026-119377.busybox,busybox-binsh,ssl_client1.37.0-r31, no fixed version: CVE-2025-60876 (Medium).
- Impact.
netops-1.0.6-vulnerability-assessments.jsonassesses each finding against the advisory text and the NetOps code. CVE-2025-15367, CVE-2026-103995 and CVE-2025-60876 are assessed as not exploitable in NetOps from its code, and so are four findings assessed on 2026-11-07: CVE-2026-119586 (a flaw of thesqlite3command-line shell, which NetOps never runs), CVE-2026-119379 (the session extension's changeset API, which Python'ssqlite3module does not expose), CVE-2026-119377 (ansqlite3_set_authorizer()bypass, while NetOps sets no authorizer and accepts no SQL from a client) and CVE-2026-116566 (a null byte in a host name passed tosocket.getaddrinfo(): every host name comes from the operator's private configuration or is validated first, never from an MCP client or a device; the Auditor REST collector and the Admin ntfy server do not reject a null byte, so this rests on that trust boundary). The eight remainingsqlite-libsfindings are assessed as not exploitable in NetOps under its threat model: CVE-2026-119374, CVE-2026-119373, CVE-2026-119378, CVE-2026-119375, CVE-2026-119371, CVE-2026-119376, CVE-2026-119591 and CVE-2026-119372. NetOps submits no SQL that reaches them; the only path is a crafted store file in the Auditor's private state or in a restored backup, and whoever can plant such a file already controls the instance's state, which is outside the threat model. They remain present in the image and are not fixed; checking the schema object types of an opened store is planned hardening. - Next step. The base image will be refreshed and the scan repeated as soon as fixed packages are available, probably in the next release. Until then 1.0.6 ships with these findings.
Acceptance
The acceptance of 1.0.6 is proportional to its change and consists of these checks:
- The unit tests of the repository, including the state version, previous-release upgrade and start tests,
release/acceptance/testsandrelease/sign/tests, and the acceptance suites of the candidate in GitHub Actions on amd64 and arm64. release/acceptance/upgrade-check.shfrom 1.0.5 to 1.0.6. It compares whole tool definitions and role responses per JSON pointer againstrelease/acceptance/api-changes-1.0.5-1.0.6.json, which declares no change; the check passes only if none is undeclared.- An installation on Debian 13 and on Fedora Server 44 with SELinux enforcing, and the upgrade from a restored 1.0.5 state with the rollback to the unchanged 1.0.5 instance, each by following the README and the quickstart.
- An independent review of the changes since 1.0.5.
- The release signing tool with its documentation, checksum, registry and vulnerability checks on the candidate.
The results are the acceptance reports (acceptance-*.json) covered by the signed checksum list. The maintainer signs only a release whose required reports pass; a candidate without that signature is not a release.
Verify this release
cosign verify-blob --bundle netops-1.0.6-SHA256SUMS.sigstore.json \
--certificate-identity 325383355+radek-cerny-soukr@users.noreply.github.com \
--certificate-oidc-issuer https://github.com/login/oauth netops-1.0.6-SHA256SUMS
sha256sum -c --ignore-missing netops-1.0.6-SHA256SUMSNetOps 1.0.5
NetOps 1.0.5 follows NetOps 1.0.4. It restricts SNMP to SNMPv3 authPriv, adds literal filtering of configuration captures, four read-only triage tools, the FortiOS log recipe logs with a time window, FortiADC pool member add and remove commands and 90 new read queries for FortiOS and FortiADC. The product remains one container with Helper, Auditor and Admin behind one audited HTTPS MCP endpoint. The syslog audit format and receiver requirements are unchanged.
Breaking changes
snmp_getaccepts only SNMPv3 authPriv (SHA-256 authentication, AES-128 privacy). SNMP v1, v2c, noAuthNoPriv and authNoPriv are refused before any network access. Targets that used a v2c community must enroll an SNMPv3 authPriv credential before upgrading. The tool name and input schema are unchanged; the tool description states the new requirement.target_scopereports the enrolled SNMPv3 security in the newsnmp_securityfield (authPrivornull).
Changes in 1.0.5
- New read-only triage tools
triage_slb,triage_health,triage_routingandtriage_vpngive one diagnostic overview each, built on the existing enrolled read functions. The complete product exposes 32 tools with Admin enabled (28 in 1.0.4) and 26 with Admin disabled (22 in 1.0.4). config_readgains literalcontainsandcontextfiltering over the redacted capture with fingerprinted continuation (filter_sha256,content_sha256), without a new SSH session, for FortiOS minimal captures and FortiOS and FortiADC full captures.config_diffof two full captures returnsbase_captureandother_capture.fortios_diagnosticsgains the FortiOS recipelogs, which a target runs only afterlogsis enrolled in itsdiagnostic_recipes, and the optional inputssinceanduntil(default null) for it. Its description now also covers the FortiADC event, session and persistence recipes.- Admin adds the FortiADC operator commands
pool-member-addandpool-member-remove(command line only, no MCP tool) besidepool-member: add creates one new pool member always disabled, remove deletes one disabled or maintain member and never the last one. Both need a FortiADC system administrator write login and end with exit code 0 (confirmed), 3 (rejected), 4 (rolled_back) or 7 (unknown); see FortiADC pool member. target_scopeaddsconfiguration_limits(optional enrolled limits for configuration bytes, records, SSH idle and command timeouts and memory, with fixed upper bounds) andsnmp_security.read_query_cataloggrows from 290 to 380 queries (FortiOS +58, FortiADC +32), among them three MCLAG switch queries (switch_mclag_consistency,switch_mclag_icl,switch_trunk_status). No query is removed. Among the new switch queries,switch_dhcp_bindingsreadsdhcp-snooping statusandssh_readreturns its parsed bindings (MAC address, VLAN, client IP, port) asdhcp_bindings, whileswitch_dhcp_snoopingreads the snooping settings (dhcp-snooping database);switch_mac_limitandswitch_source_guardstate when an empty result is meaningful.totalis 380 andcatalog_sha256ise172c53c708c03035865aae3b4cf3600c7637617e315560d7b44d22c733275e1.- Four typed FortiADC object views (virtual server, pool, health check, WAF profile) are available through
config_read; the Auditor can audit an exported FortiADC configuration and applies eleven further CIS FortiGate rules. See the CHANGELOG for the complete list including Admin profile changes and fixes. - The container base image moves to Python 3.14.10 on Alpine 3.24, and
pyasn1andjoserfcare updated, which fixes the vulnerabilities listed inCHANGELOG.md.
All other tool definitions and role responses are unchanged between 1.0.4 and 1.0.5; the upgrade check compares them field by field against release/acceptance/api-changes-1.0.4-1.0.5.json.
Upgrade from 1.0.4
Follow the upgrade procedure in the quickstart:
- Check the 1.0.4 instance, stop it and create its backup and state manifest.
- Download and verify the 1.0.5 release files. Prepare a separate instance directory and a new Compose project with the same server name and port.
- Restore the backup into an empty target volume and compare it with the source manifest before the first start. Keep the original instance and backup.
- Start the new instance. The first start converts only the recorded product version from 1.0.4 to 1.0.5 under the instance lock and writes an
instance_startaudit record withupgradefrom 1.0.4 to 1.0.5. A second start changes nothing. CA, keys, token, role data and connection details are unchanged.
Before upgrading, enroll SNMPv3 authPriv credentials for every target that uses snmp_get. Do not run init on the restored state. State of any other product version is refused; there is no skipped-release upgrade.
For rollback, stop 1.0.5, preserve its audit records written after the backup, and start the unchanged 1.0.4 instance. A direct downgrade of a converted volume is not supported.
Acceptance
release/acceptance/upgrade-check.sh from 1.0.4 to 1.0.5 compares whole tool definitions and role responses per JSON pointer against the declared changes; none is undeclared and none is stale. The BREAKING entry for snmp_get is the only one.
Verify this release
cosign verify-blob --bundle netops-1.0.5-SHA256SUMS.sigstore.json \
--certificate-identity 325383355+radek-cerny-soukr@users.noreply.github.com \
--certificate-oidc-issuer https://github.com/login/oauth netops-1.0.5-SHA256SUMS
sha256sum -c --ignore-missing netops-1.0.5-SHA256SUMSNetOps 1.0.4
NetOps 1.0.4 follows NetOps 1.0.3. It adds conservative FortiOS configuration captures and comparisons, bounded diagnostics without a schema registry, a paginated query catalogue and the first FortiADC Helper read. The product remains one container with Helper, Auditor and Admin behind one audited HTTPS MCP endpoint. The syslog audit format and receiver requirements are unchanged.
Changes in 1.0.4
- Helper adds
config_readandconfig_difffor explicitly enrolled FortiOS configuration sections. Captures verify the live identity and single-root VDOM context, redact configuration values incrementally and keep bounded snapshots and comparisons in process memory. Completed captures and comparisons can be paged without rereading the device. Opaque comparisons can report changes to hidden values without returning those values. See configuration captures. - Configuration streaming handles the initial FortiOS exec prompt and the matching final prompt; unexpected framing is refused. The proxy allows the bounded configuration operation to finish within its existing tool deadline.
- The existing
fortios_diagnosticstool runs enrolled ping, traceroute, sniffer, session and flow recipes without a schema registry. It verifies an exact measured hardware/version/build identity, a single root VDOM and interface ownership before starting. Explicit VDOM selection and multi-VDOM remain refused. Recipes retain their limits, timer interruption and verified cleanup. FortiGate-VM64 with FortiOS 8.0.2 build 0251 is an explicit identity; this does not enable other 8.0.x models or builds. read_query_catalogadds platform and query-prefix filters, offset/limit pagination and a page-data byte bound. Follownext_offsetto read all results and restart ifcatalog_sha256changes. A default call returns a page rather than the complete catalogue.- FortiADC is a separate Helper platform.
ssh_readverifies the live product before returning catalogue output; the first qualified function issystem_statuson FortiADC-KVM 7.6.7 build 0676. This does not qualify FortiADC Auditor collection or Admin writes. - The Helper exposes 15 tools, including the two configuration tools; the complete product exposes 28 tools with Admin enabled and 22 with Admin disabled.
- Admin profiles refuse FortiOS 8.0.2 build 0251 before installing a safeguard or applying a change. Its device-side configuration rollback is not qualified. See verified support.
Upgrade from 1.0.3
Follow the upgrade procedure:
- Check the 1.0.3 instance, stop it and create its backup and state manifest.
- Download and verify the 1.0.4 release files. Prepare a separate instance directory and a new Compose project with the same server name and port.
- Restore the backup into an empty target volume and compare it with the source manifest before the first start. Keep the original instance and backup.
- Start the new instance. The first start converts only the recorded product version from 1.0.3 to 1.0.4 under the instance lock. Verify the version, health, existing role data, CA, token and audit with the original clients.
Do not run init on the restored state. The syslog receiver and its part archive stay unchanged. State of any other product version or state schema is refused: upgrade 1.0.2 to 1.0.3 first with the 1.0.3 procedure, and follow the earlier release procedures in order for older instances. There is no skipped-release upgrade or downgrade.
For rollback, stop 1.0.4, preserve its audit records written after the backup as described in Rollback, and start the unchanged 1.0.3 instance. Changing the new instance's image back to the old digest does not restore its state.
Documentation and release process
The quickstart makes failed Compose checks and partial audit exports explicit, limits the upgrade audit tail to five records and keeps the backup and restore variable blocks distinct. Step 6 of the quickstart starts in ~/netops, so it also works in a new shell, and the upgrade check describes the audit tail by the history of the instance. Test tooling bounds signing-test subprocesses, validates worker mode and retains unittest outcomes. The process-profile acceptance check accepts a health check that the container runtime has started but not yet handed to the container shim only with full proof of that runtime parent and only for 200 ms, proves the parent of a short-lived runtime helper before the helper's own record disappears, and names the reason when it fails. The source gate refuses a README that describes another release and workflows that name another candidate branch, and CI no longer repeats the checks that the release candidate workflow runs on candidate branches. These changes do not alter device permissions.
Acceptance
NetOps 1.0.4 adds Helper functions and is accepted in proportion to its change: CI suites and engine tests of the candidate image in GitHub Actions on amd64 and arm64, a blind installation by the README and the quickstart on Fedora Server 44 and Debian 13 on amd64, a blind upgrade from 1.0.3 to 1.0.4 on amd64, read-only checks of the new Helper functions on FortiGate-60F with FortiOS 7.6.8 build 3744, FortiGate-VM64 with FortiOS 8.0.2 build 0251 and FortiADC-KVM 7.6.7 build 0676, and an independent code review. Ubuntu, Raspberry Pi OS, arm64 hosts and rootless Docker were not installed by a blind run for 1.0.4. The acceptance reports covered by the signed checksum list record the results.
Known limitations
- Configuration captures require explicit scope, pinned SSH identity and a device account with CLI get/show permission. Diagnostics additionally require diagnose/execute permission; inventory opt-in does not grant device privileges. CLI configuration can remain disabled.
- Configuration values are withheld conservatively. Captures and comparison handles are bounded and do not survive a process restart. The fixed limits in Helper configuration apply; this release adds no policy override for them.
- Bounded diagnostics accept only the explicit identities in the implementation and require a single root VDOM. Reaching the time limit with no packet output is not proof of forwarding or destination reachability.
- FortiADC qualification covers the documented Helper
system_statusread. Auditor support and live Admin pool-member changes are not qualified. - FortiOS 8.0.2 build 0251 Admin changes are refused. Measurements of another model, build or release do not establish support.
- An interrupted converting first start can leave converted state without an audit record naming the upgrade; the quickstart explains recovery and the remaining evidence.
- Existing process isolation, audit persistence and syslog retry limitations remain as described in security. Device output remains sensitive after credential redaction.
Install
Follow the quickstart. Download and verify the signed release files, configure the bind address, port and server name, then initialize and start a new instance. An upgrade uses the restore procedure above.
Verify this release
cosign verify-blob --bundle netops-1.0.4-SHA256SUMS.sigstore.json \
--certificate-identity 325383355+radek-cerny-soukr@users.noreply.github.com \
--certificate-oidc-issuer https://github.com/login/oauth netops-1.0.4-SHA256SUMS
sha256sum -c --ignore-missing netops-1.0.4-SHA256SUMSThe checksum list covers compose.yaml, env.example, the release manifest, which records the image index and per-architecture digests, the acceptance manifest, the per-architecture SBOMs, vulnerability scans and scan gate results, the vulnerability assessments and the acceptance reports (acceptance-*.json); the checksum list itself is the complete list of covered files. env.example already pins the image by that index digest. A download is verified only after both commands succeed.
NetOps 1.0.3
NetOps 1.0.3 follows NetOps 1.0.2 and upgrades the state of an instance of NetOps 1.0.2. It is a corrective release of the same single product: one container image that runs the Helper, Auditor and Admin roles behind one audited HTTPS MCP endpoint. Apart from the audited response delivery controls in the gateway, the runner SSH diagnostics of the Helper (the negotiation message and the classification of failures), the status wait of the health check and the recorded product version, the product code is unchanged. A login banner of the runner that does not reproduce an OpenSSH client message no longer changes the category of a runner SSH failure that the Helper reports, and a failed remote command is never reported in an SSH category. The syslog audit format and the receiver requirements of 1.0.2 stay as they are.
Changes in 1.0.3
- Gateway: after an audited response and its audit commit are recorded, its delivery has one absolute deadline of 60 seconds that progress does not reset. An observed client disconnect, a timeout or cancellation, or a failed send during the delivery leads to an attempt to record
interrupteddelivery withautomatic_retryfalse andoperation_may_have_completed, followed by aconnection_endwith deliveryinterrupted, and the operation slot is released. See Response delivery. - The Helper SSH connection to a runner host keeps the message of the OpenSSH client about a failed algorithm negotiation (client log level
INFOinstead ofERROR) and reports it asssh_negotiation:SSH algorithm negotiation with the runner failed. Update the runner SSH service or review its allowed algorithms.The message names neither the host nor the offered algorithms; the algorithms that the client offers are unchanged. An SSH category (ssh_host_key,ssh_negotiation,ssh_authentication,ssh_connection) is reported only when the SSH client itself fails with exit status 255; any other exit status comes from the remote command and is reported asremote_execorssh_transport, as in 1.0.2. The category is decided by the last line of the client standard error other than the closingDisconnected from ...line of OpenSSH, matched against the exact OpenSSH client messages, and the Helper keeps the last 64 KiB of that output instead of the first, so a login banner of the runner, which theINFOlevel prints, neither changes the category nor pushes the real error out unless its last line reproduces such an OpenSSH message. - The health check
netops healthqueries the roles at the same time and waits up to 5 seconds instead of 2 in total for their status answers, so single runs of the container healthcheck no longer fail while the Helper redacts an SSH output of 2 MB to 16 MiB on a slow CPU; a role that does not answer within these 5 seconds still fails the check, within the 10-second timeout of the container healthcheck. - Upgrade the state of NetOps 1.0.2: the first start of 1.0.3 converts its state under the instance lock and changes only the recorded product version, in the same way as the upgrade from 1.0.1 to 1.0.2.
Response delivery
The gateway validates the buffered ASGI response before committing its response audit. An invalid start/body/completion sequence takes the existing operation-failed path: interrupted outcome, an audited 503 response and no committed-before-send delivery record.
After a valid response and its audit commit are recorded, its ASGI send sequence has one absolute deadline of 60 seconds for all parts, including delivery-task cleanup and the final send. Progress does not reset it. This deadline does not cover admission waiting, operation execution, audit persistence, or direct gateway error and health responses. Previously a blocked send could keep its operation slot without this deadline.
During body delivery the gateway watches the ASGI receive channel for http.disconnect because the shipped uvicorn h11 send can return silently after a disconnect. An observed disconnect, timeout, cancellation, receive-channel failure or send failure triggers an attempt to record interrupted delivery with automatic_retry false and operation_may_have_completed, followed by connection_end with delivery interrupted. These records require a writable audit. The operation slot is released even when recording the interruption fails. An observed http.disconnect returns after that audit tail without raising an application exception; timeout, cancellation and unexpected channel errors still propagate.
Body writes are limited to 32 KiB, and an additional nonterminal send checks backpressure after the last payload. The sole receive watcher is stopped and awaited before the terminal send, so normal response completion is not mistaken for a disconnect. A response that fits in the transport buffer, or a close after the monitored phase, can go unobserved. Completion of the ASGI send sequence does not prove that the client received or read the response. Inspect audit pending, the recorded response and the device before repeating an operation; never retry writes automatically.
Release process
- The privacy gate of the release reports (
release/report_gate.py) also decodes canonical standard and URL-safe base64 text, within shared limits for decoding layers and bytes, before it searches the reports for private markers and private paths. Decoded content that is not UTF-8 text is searched without its NUL bytes, which covers UTF-16, and by its printable runs of at least 8 characters. The public host-name checks do not decode base64, so Docker manifest and configuration JSON embedded in the Grype reports raises no false host findings. - The platform smoke check of the release acceptance (
release/acceptance/platform_smoke.sh) installspython3together withca-certificatesandcurlon hosts with apt before its checks. - Each architecture SBOM is validated offline against the complete, pinned CycloneDX 1.7 JSON schema before a release candidate is written; the SBOM must list components and its root container must match the exact image name and architecture digest.
- The source privacy gate (
release/check_source.py) skips only its domain check for the five vendored CycloneDX 1.7 schema files inrelease/schema/cyclonedx-1.7/, including the SPDX license schema, and only while each file matches its pinned SHA-256; every other check still applies to them. - Open placeholders in the public documentation (
release/placeholder_gate.py): the candidate workflow refuses to build a release candidate while its release notes are missing, its changelog has no section of its version or its public documentation carries a placeholder other than the acceptance placeholder of that version (acceptance-X.Y.Z) and the release-date placeholder (release-date) in the heading of its changelog section. These two are filled after the acceptance of the candidate digests, so the candidate is not rebuilt for them. The signing tool (release/sign/netops-release-sign.ps1) takes the source checkout of the release with-SourceDirand refuses to sign while its Markdown still carries any placeholder, its release notes are missing or the changelog section of its version has no date.
Documentation
- Quickstart: script forms of steps 7 and 8 (
-Tand< /dev/nullalso forupandps, its own block forca,token showandconnection); step 7 names theContainer ... Creating/Createdlines ofdocker compose runas Compose progress on standard error; step 9 counts the tools of an authenticatedtools/listovercurland shows how to find the step 9 requests in the audit. - Quickstart upgrade: the check lists the
instance_startrecords with a readyjqfilter, the variable block setsumask 077so thatnetops-ca.pemstays private in a new shell, the hint about-pthatconnectionprints does not apply in the directory of the new instance, and the directory of the previous instance is found from its Compose label. - Document the upgrade from 1.0.2 to 1.0.3 in the quickstart with the structure of the upgrade from 1.0.1 to 1.0.2, including an instance of 1.0.2 that was itself upgraded from 1.0.1 (
~/netops-1.0.2, projectnetops-102). - Helper tools and README:
ftp_listis refused in the shipped profile before any device contact; its engine checks the path againstsftp_rootsonly lexically (see Known limitations). - Admin configuration: how NetOps ends an Admin FortiOS operation with
unknownandsafeguard_kept(it reads no firing record from FortiOS and waits untilsafeguard.fire_atplus 20 seconds), and how to prove on the device that the safeguard stitch fired, with the measured firing delay of 0.1 to 5 seconds on a FortiGate 80F with FortiOS 8.0.2 and a FortiGate 60F with FortiOS 7.6.7. - README: the covered assets of the signed checksum list are named by kind, and the checksum list itself is the complete list.
Acceptance
NetOps 1.0.3 is a corrective release and was accepted in proportion to its change: CI suites and engine tests of the candidate image in GitHub Actions on amd64 and arm64, a blind installation by the README and the quickstart on Fedora Server 44 and Debian 13 on amd64, a blind upgrade from 1.0.2 to 1.0.3 on amd64 and an independent code review. Ubuntu, Raspberry Pi OS, arm64 hosts and rootless Docker were not installed by a blind run for 1.0.3.
Upgrade from 1.0.2
Follow Upgrade from the previous release. The upgrade has four steps:
- Check the 1.0.2 instance, stop it and back it up as described in Before the upgrade.
- Download and verify the 1.0.3 release files, then prepare a clean 1.0.3 instance in a new directory, with the same server name and port and a
.envwith the 1.0....
NetOps 1.0.2
NetOps 1.0.2 follows NetOps 1.0.1 and upgrades the state of an instance of NetOps 1.0.1. It is a corrective release of the same single product: one container image that runs the Helper, Auditor and Admin roles behind one audited HTTPS MCP endpoint. Apart from one security fix in the Core SFTP banner handling, two operator hints and the recorded product version, the product code is unchanged. The syslog audit format and the receiver requirements of 1.0.1 stay as they are.
Security fix
- The Core
sftp_stattook the first stderr line that began withConnected toas the end of the SSH login banner. A device whose banner contained such a line, for exampleConnected to the lab network, could make every call withsftp_rootsfail and an empty directory unreadable. A banner that also contained a forged OpenSSH client message letsftp_statlist a directory reached through a symbolic link out of an enrolled root (metadata only: names, sizes and modes, not file content). This needs an SSH server whose banner the attacker controls and an existing symbolic link below the root. NetOps 1.0.2 takes the last line exactlyConnected to <host>.for the target host, which the client prints after authentication and therefore after any banner. Banner text before that line is ignored; if the line is missing, the whole stderr is treated as client output. Regression tests cover both banners, and the fix was checked against OpenSSH 10.0p2sftp-server.
Changes in 1.0.2
- The CA export hints printed by
init,connectionandca rotatenow name the complete commanddocker compose run --rm netops ca > netops-ca.pem. They lacked the redirection, so a copied command printed the CA to the terminal instead of writing the file that the client configuration uses. - Upgrade the state of NetOps 1.0.1: the first start of 1.0.2 converts its state under the instance lock and changes only the recorded product version, in the same way as the upgrade from 1.0.0 to 1.0.1.
- Rebuild the quickstart into numbered steps, each ending with a check, after blind installation tests of 1.0.1 on Fedora Server 44 and Ubuntu: Docker Engine from Docker's repository for every host family, SELinux support in the Fedora Docker daemon, account access through the
dockergroup or rootless Docker, cosign from the official Sigstore 3.x release, a download block that stops at the first failed download, a separate step for an MCP client on the same host, and a check of the running instance without an MCP client. - Document that the Core
sftp_statrefuses paths below a root (fail-closed) on embedded SFTP servers that send no permissions in their stat answer or returnREALPATHin another form; the path equal to a root works.
Acceptance
NetOps 1.0.2 is a corrective release and was accepted in proportion to its change: CI in GitHub Actions on amd64 and arm64, a blind installation by the README and the quickstart on Fedora Server 44 and Debian 13 on amd64, a blind upgrade from 1.0.1 to 1.0.2 on amd64 and an independent code review. Ubuntu, Raspberry Pi OS, arm64 hosts and rootless Docker, whose quickstart commands are newly written in 1.0.2, were not tested by a blind run for 1.0.2; the product code on these paths is unchanged since 1.0.1, whose acceptance covered them.
Upgrade from 1.0.1
Follow Upgrade from the previous release. The upgrade has four steps:
- Check the 1.0.1 instance, stop it and back it up as described in Before the upgrade.
- Download and verify the 1.0.2 release files, then prepare a clean 1.0.2 instance in a new directory, with the same server name and port and a
.envwith the 1.0.2 image. - Restore the backup into it as described in the same section of the quickstart.
- Start it; the restore block of the quickstart ends with the start. The first start of 1.0.2 converts the restored state of 1.0.1 under the instance lock.
A syslog receiver configured for 1.0.1 stays unchanged. State of any other product version or state schema is refused, there is no upgrade that skips a release and no downgrade: upgrade an instance of NetOps 1.0.0 to 1.0.1 first, as described in the 1.0.1 release notes, including its syslog receiver change. The 1.0.1 instance, its directory and its state volume stay unchanged. Rollback: stop the 1.0.2 instance, export its audit records written after the backup as described in Rollback and start the original 1.0.1 instance.
Known limitations
- The separate Helper SSH proxy to a runner host supports both keyboard-interactive and password authentication. When a server offers both methods, one connection with an incorrect password can cause two failed authentication attempts, which may count under the runner account's lockout policy.
- Authenticated requests with bodies above 64 KiB or uncertain body framing are processed one at a time across roles, so they can queue behind another large request. Other authenticated requests can use the remaining slots within the shared limit of 16.
- The 60-second deadline covers only the receipt of the request body: a client that does not read a large response keeps its operation slot until the response is sent, so such clients can occupy the shared limit of 16.
- 1.0.2 upgrades only NetOps 1.0.1; it does not upgrade or migrate 0.x installations or data.
- The roles are isolated as processes and files inside one container, not against the host administrator or the container runtime.
- An incomplete local audit write blocks further operations until
netops audit recoverruns while the instance is stopped. A full syslog queue blocks new operations, and an interrupted send can deliver records to the receiver more than once. - The opt-in FortiOS
managed_switch_stackingandmanaged_switch_poequeries have no positive result on stacking- or PoE-capable hardware. Virtual device images do not establish physical hardware behavior. - The process-profile acceptance check uses a nominal 50 ms sampling interval and can miss processes that start and exit between observations; the 1.0.1 release notes give the exact attribution rule.
- When the converting first start of 1.0.2 stops after the conversion, for example because a role refuses its configuration or the start is interrupted, no audit record names the upgrade. The only indirect evidence is the version in the last 1.0.1 records and in the next
instance_startof 1.0.2, and only with audit history kept from before the upgrade; see Refused first start. - The SHA-256 of a syslog record detects damaged or incomplete part archives. It does not prove the authenticity of data that an attacker controlled before the record was complete.
- Audit events suppressed before NetOps sends them cannot be detected from the receiver archive.
Details: verified support and security.
Install
Follow the quickstart. In short: download compose.yaml, env.example, netops-1.0.2-manifest.json, netops-1.0.2-SHA256SUMS and netops-1.0.2-SHA256SUMS.sigstore.json from this release, verify them, set the bind address, port and server name, then run init and start the service.
Verify this release
cosign verify-blob --bundle netops-1.0.2-SHA256SUMS.sigstore.json \
--certificate-identity 325383355+radek-cerny-soukr@users.noreply.github.com \
--certificate-oidc-issuer https://github.com/login/oauth netops-1.0.2-SHA256SUMS
sha256sum -c --ignore-missing netops-1.0.2-SHA256SUMSThe checksum list covers the release manifest, which records the image index and per-architecture digests, the SBOMs, the vulnerability scans and their assessments, and the acceptance reports of this release. env.example already pins the image by that index digest. A download is verified only after both commands succeed.
Security
Every release candidate is built once in GitHub Actions and the same digests are tested, accepted and published without a rebuild. Each architecture is scanned; a fixable High or Critical finding blocks the release and every other finding, of any severity, ships with a written assessment (netops-1.0.2-vulnerability-assessments.json). SBOMs cover the application, its Python dependencies and the operating-system packages. The only security fix in 1.0.2 is the Core SFTP banner fix described under Security fix.