cairnobs/docs/compliance/license-audit-report.md

License compliance audit report — Phase 6

This report is a strong first pass, not a legal opinion. It should be reviewed by actual legal counsel before the project is publicly released, pitched to customers, or used as the basis for any compliance claim. Nothing in this document should be represented to a third party as legal advice or a certified compliance determination.

Scope and goal

Every component in the monorepo — /agent, /transport, /ingest, /storage, /api, /web, /cli, /deploy, and enterprise/ — is licensed AGPLv3, with no separate commercial-license carve-out anywhere in the project. This audit inventories every third-party dependency across every language ecosystem, classifies each for AGPLv3 compatibility, resolves or explicitly flags anything that doesn't resolve cleanly, and stands up CI enforcement so this doesn't need to be manually re-audited from scratch every time a dependency changes.

This is a compatibility audit, not a "replace every permissively licensed library" exercise. MIT, Apache-2.0, BSD, and ISC dependencies are all fine to depend on from AGPLv3 code. The concern is dependencies with licenses that are genuinely incompatible, impose unaccounted-for obligations, or aren't open source at all (source-available licenses like BSL/SSPL, anything with a field-of-use or non-compete restriction).

Methodology

1. Inventory every dependency, direct and transitive, in every language ecosystem present in the repo:
   • Rust (agent workspace: sentry-agent, sentry-parser; search): cargo-deny (cargo deny list --format tsv), installed fresh for this audit (cargo install cargo-deny --locked).
   • Go (api, ingest, alerting, enterprise, deploy/operator, terraform, proto, hack/benchmark-fixture, hack/windows-fixture — every module with real third-party dependencies; cli, hack/webhook-sink, and hack/alert-load-test are stdlib-only, confirmed by inspecting their go.mod files, not scanned): go-licenses (google/go-licenses, installed via go install ...@latest), go-licenses csv ./... per module.
   • npm (web): license-checker (npx license-checker --json).
   • Vendored/non-manifest content: a repo-wide grep for copy/attribution markers ("adapted from", "copied from", "stackoverflow", stray copyright headers, "vendored") — zero hits. A separate pass for binary/asset files (fonts, icons, images) outside node_modules/build output — two font files and one SVG found, see below. Docker base images referenced in docker-compose.yml were pulled out and reviewed manually, since they're not a language-ecosystem dependency any of the three scanners above would see.
2. Classify every distinct license found into: (a) clearly compatible as a dependency, (b) requires a closer look, (c) actually incompatible or non-open-source. Every (b)/(c) result is cited with real reasoning below, not asserted.
3. Named risk areas (Redpanda's licensing history, ClickHouse client libraries, Phase 5 font/asset files, Kubernetes Operator tooling) checked explicitly, against primary sources (actual license files at the actual pinned versions/tags), not general recollection.
4. Remediation options recorded for every (c)/unresolved-(b) item — fix, isolate, or flag for a business decision. Nothing was silently swapped.
5. Own license declarations audited: root LICENSE file, per-ecosystem manifest license fields, and a repo-wide check for any file still claiming a license other than AGPLv3.
6. enterprise/ relicensed to AGPLv3, closing the former commercial-license carve-out project-wide.
7. CI enforcement stood up so this audit's findings don't silently go stale.

Inventory summary

Full machine-readable inventory: /docs/compliance/license-inventory.csv and .json (columns: component, dependency, version, license, direct_or_transitive, ecosystem, flagged, flag_reason, classification).

Ecosystem	Component-scoped rows	Unique dependencies	Flagged (unique)
Rust (cargo)	440	—	4
Go	255	—	9 (segmentio/asm + 8 HashiCorp MPL-2.0 packages)
npm	75	—	3
Docker base images	3	3	1 (Redpanda)
Font/vendored assets	3	3	1 (favicon)
Total	776	502	23 (~4.6%)

Rows double-count dependencies shared across multiple components by design (e.g. pgx appears once per Go module that imports it) — that's what makes the component column meaningful. All 502 unique dependencies resolved to classification (a) except one: Redpanda, classification (c), recorded below with remediation options rather than resolved unilaterally.

Classification results

774 of 776 rows (all but Redpanda and the favicon asset) classify as (a): clearly compatible. No dependency in this project's tree required a genuine (b)-category deep-dive that didn't resolve cleanly — every item that looked ambiguous at first pass (dual-licensed with a copyleft option, a non-standard license file, an unfamiliar SPDX identifier) turned out, on actual inspection, to resolve to (a) once read carefully. That's a real result of doing the reading, not an assumption going in — recorded below per item so the reasoning is checkable.

Dual/multi-licensed crates electing a permissive branch (Rust)

Four Rust crates carry an SPDX OR expression that includes a copyleft/less-common option alongside a permissive one. Standard practice for a disjunctive license grant is that the downstream user elects whichever listed option they prefer — we elect the permissive branch in every case, incurring zero copyleft obligation:

Crate	License expression	Elected	Note
r-efi	MIT OR Apache-2.0 OR LGPL-2.1-or-later	MIT/Apache-2.0	Build-dependency only (via tonic-build→prost-build→tempfile→getrandom), for a UEFI target this project doesn't build for — doesn't even ship in a release artifact.
fastdivide	MIT OR zlib-acknowledgement	MIT
htmlescape	Apache-2.0 / MIT / MPL-2.0	Apache-2.0/MIT	MPL-2.0 would also have been fine on its own merits (see below).
ryu	Apache-2.0 OR BSL-1.0	Apache-2.0	BSL-1.0 here is the Boost Software License — permissive, unrelated to Redpanda's Business Source License below despite the shared abbreviation.

MPL-2.0 (Go, npm, and one Rust option above)

MPL-2.0 is file-level (weak) copyleft: modifications to MPL-covered files must stay available under MPL if distributed, but combining MPL-covered code into a larger differently-licensed work — including an AGPLv3 work — does not require the larger work to relicense (MPL 2.0 §3.3, "Distribution of a Larger Work"). Pre-cleared as category (a) per this audit's own scope definition. Found in:

• Go (/terraform only, HashiCorp's Terraform provider SDK and its own dependencies): go-plugin, go-uuid, terraform-plugin-framework, terraform-plugin-go, terraform-plugin-log, terraform-registry-address, terraform-svchost, yamux. Lower risk still than the general case: /terraform is its own standalone Go module producing a Terraform provider plugin binary, not linked into any core service.
• npm: axe-core (dev-only, used for the Phase 5 accessibility sweep, never shipped), lightningcss + its platform-specific native binary (a transitive dependency of Vite's CSS pipeline — build-time only, never bundled into web's shipped static output).

segmentio/asm — tool detection gap, not a real license question (Go)

go-licenses reported Unknown for every sub-package of github.com/segmentio/asm (a transitive dependency via the ClickHouse Go driver). Its actual LICENSE file (read directly from the module cache) is headed "MIT No Attribution" — SPDX MIT-0, a permissive MIT variant that drops the attribution requirement. The auto-detector's regex didn't recognize that non-standard heading text. Confirmed by reading the file, not assumed; carried as an explicit CI ignore with this citation (see the policy doc) rather than silently added to the general allow-list.

Named risk areas (task 3)

• ClickHouse client libraries (github.com/ClickHouse/clickhouse-go/v2, github.com/ClickHouse/ch-go): Apache-2.0, matching the server itself. No divergence.
• Kubernetes Operator tooling (sigs.k8s.io/controller-runtime, k8s.io/client-go, k8s.io/apimachinery): Apache-2.0 (one forked sub-package, apimachinery/third_party/forked/golang, is BSD-3-Clause — also fine). The concern about "generated boilerplate's license headers" turned out not to apply: deploy/operator/api/v1alpha1/ zz_generated.deepcopy.go is, despite its name, hand-written, not actually produced by controller-gen (no kubebuilder/controller-gen binary was available when it was built — disclosed in the file's own doc comment and in /deploy/README.md). There's no real upstream-generated boilerplate to check for header drift against.
• Phase 5 font/asset files: see below.

Redpanda — classification (c), recorded for a business decision

The pinned Redpanda image (docker.redpanda.com/redpandadata/redpanda:v24.2.7, docker-compose.yml and transport/) ships under BSL 1.1 (Business Source License), confirmed against the actual licenses/bsl.md file at that tag (github.com/redpanda-data/redpanda, tag v24.2.7) — not assumed from general familiarity with Redpanda's licensing history, which the audit brief specifically warned has shifted over time.

Key facts, verified against primary sources:

• Not OSI-approved open source. BSL is explicitly source-available, matching this audit's own category-(c) definition.
• Change Date: 4 years from each version's release date, after which that version's Licensed Work converts to Apache-2.0. v24.2.7 was released 2024-10-11 (confirmed via the GitHub Releases API) — its Change Date is ~2028-10-11. As of this audit, it has not yet converted.
• The restriction is narrow: BSL's Additional Use Grant permits any use except offering the Licensed Work as a "Streaming or Queuing Service" to third parties (defined as a commercial offering letting third parties create topics in the Licensed Work, e.g. a hosted Kafka broker product). This project's docker-compose.yml uses only plaintext core Kafka-protocol functionality — no RCL-gated enterprise features, no tiered storage, no SASL/RBAC — squarely within the permitted grant as an internal transport layer.
• No AGPL linking/compatibility issue. Sentry never links against Redpanda's code; it's consumed purely over the Kafka wire protocol, the same relationship as ClickHouse and Postgres. AGPLv3's copyleft doesn't reach across a network-protocol boundary to unrelated, separately-licensed software you merely talk to.
• The genuinely open question: Phase 6 relicenses enterprise/ to AGPLv3 specifically so that anyone, including competitors, can legally self-host or fork Sentry — including offering it as a network service, per AGPLv3's own terms. If a third party does that using the bundled docker-compose.yml (which pulls this BSL-licensed Redpanda image), does their deployment trip BSL's Streaming-or-Queuing-Service restriction? Sentry's ingest pipeline creates fixed internal topics, not per-end-user topics exposed for direct third-party production or consumption — so this is very likely not a Streaming-or-Queuing-Service under BSL's own definition. But this is a business/redistribution judgment call about a hypothetical third party's use, not a pure technical compatibility question this audit can close unilaterally.

Remediation options (recorded per task 4's requirement)

1. Accept as-is. Document the reasoning above; Sentry's own use is clearly within BSL's permitted grant, and the third-party-SaaS scenario is a reasonable-but-unverified reading, not a known violation. Lowest effort, zero functional change.
2. Swap to Apache Kafka (Apache-2.0, genuinely OSI open source). apache/kafka (KRaft mode, no ZooKeeper needed as of Kafka 3.x) is wire-protocol-compatible with everything transport/ingest/search already speak. Real tradeoff: Redpanda was originally chosen partly for its lightweight single-binary footprint (docker-compose.yml runs it with --smp=1 --memory=1G --overprovisioned, tuned for a resource-constrained local/homelab deployment per CLAUDE.md's stated deployment targets); Kafka's JVM-based broker has a materially larger minimum memory/startup footprint. This is a real regression for the project's stated "docker-compose for local/homelab" use case, not a drop-in swap with no cost.
3. Stop bundling a pinned broker image at all. transport/ already has no application code of its own — it's a thin docker-compose wrapper and topic-provisioning script. Document Kafka-API compatibility as the requirement and let the operator supply their own broker (self-installed Apache Kafka, their own separately-licensed Redpanda, or anything else wire-compatible). This moves the redistribution question out of this project's own docker-compose.yml entirely, at the cost of a rougher out-of-the-box local dev experience (an extra manual setup step instead of docker compose up).

Decision recorded 2026-08-16: option 1, accept as-is. No code or deployment change was made as a result — Redpanda stays pinned at v24.2.7 in docker-compose.yml/transport/, under BSL 1.1, as a disclosed and accepted risk rather than an unresolved one. This decision should be revisited if the project's redistribution posture changes materially (e.g. an official hosted/managed offering of Sentry itself, which would make the third-party-SaaS reading in this section Sentry's own situation rather than a hypothetical third party's).

Non-license finding: favicon.svg

web/src/lib/assets/favicon.svg is SvelteKit's own default project scaffold logo (<title>svelte-logo</title> — the sv create/create-svelte starter icon), never replaced with an original mark during Phase 5's redesign. Not a license-compatibility blocker — Svelte's own project assets are MIT-licensed — but it's unauthored, third-party-branded content shipping as this product's own favicon, caught by the same "grep for anything that looks copied" pass this audit's task 1 asked for. Recorded as an action item (replace with an original Sentry mark), not a compliance blocker; not fixed here since it's a design task outside this phase's scope, not a licensing one.

Own license declarations (task 5)

Before this audit: no root LICENSE file existed anywhere in the repo — not at the root, not in enterprise/. The only license declarations were prose statements in CLAUDE.md/docs/architecture.md and correct license = "AGPL-3.0-only" fields in the two Rust workspaces' Cargo.tomls. web/package.json had no license field at all (npm's tooling reported the package itself as UNLICENSED as a result). web/static/fonts/LICENSE.txt was a paraphrase describing the Overpass font's license, not the actual OFL-1.1 license text.

Fixed:

• Added /LICENSE — the verbatim, unmodified AGPLv3 text from gnu.org/licenses/agpl-3.0.txt, byte-for-byte, not paraphrased.
• Chosen convention, applied consistently: one root LICENSE file governs the whole monorepo, plus a license field in every ecosystem manifest that supports one (Cargo.toml's license/license.workspace, now confirmed correct; package.json's license, added: "AGPL-3.0-only"). Go has no manifest-level license field — the standard convention (and what go-licenses itself looks for) is the root LICENSE file, which now exists. Deliberately not adopting per-file SPDX header comments across the monorepo's several thousand source files: headers are an FSF best-practice recommendation, not a legal requirement once a correct root LICENSE plus copyright ownership is established, and retrofitting them here would be a huge mechanical change for very little incremental legal value over what's now in place. Recorded as a deliberate choice, not left half-done.
• Corrected web/static/fonts/LICENSE.txt from a paraphrase to the actual, complete, unmodified OFL-1.1 text (fetched from github.com/googlefonts/overpass, the actual repository these font files were fetched from per the file's own prior note) plus the correct copyright statement, with the original context note (self-hosted vs. CDN) preserved as a clearly separated project note, not mixed into the license text itself.
• No accidental license mismatch from copied code was found — the repo-wide attribution-marker grep in the methodology section came back empty, and no vendored directories exist.

enterprise/ relicensing to AGPLv3 (task 6)

This is a deliberate business-model choice, recorded plainly so it isn't rediscovered as a surprise later: the project is no longer pursuing commercial-license revenue from the former enterprise/ features (SSO, multi-tenancy/RBAC, audit logging). As of Phase 6, anyone, including competitors, can legally self-host or fork those features under AGPLv3's terms. AGPLv3's source-sharing obligation applies to network use (anyone interacting with a modified version over a network is entitled to its source) — it does not impose any payment obligation, and nothing in this project gates functionality behind a license key or entitlement check.

Confirmed no license-gating logic exists: a repo-wide grep for license-key/entitlement-check patterns (license.?key, entitlement, commercial.?key, paywall, IsLicensed, and similar) across enterprise/'s Go source returned zero hits. There was never a functional paywall to remove — Phase 4's enterprise/ split was always an architectural/import-boundary separation, not a runtime license check, so this task's "if any such gating exists, flag it explicitly, it needs to come out" condition doesn't apply here — confirmed, not assumed.

Changes applied:

• Every prose reference to enterprise/ as "commercial license" or "commercial-licensed" across the repo was updated. Present-tense claims (code comments, README.md files describing current state, CLAUDE.md's non-negotiable constraints) were corrected outright. Historical, phase-specific documents (docs/phase-4-isolation-design.md, docs/phase-4-rbac-design.md, docs/phase-4-runbook.md, and the relevant parts of CLAUDE.md's and docs/architecture.md's Phase 4 sections) were given forward-pointing corrections — "commercial license at the time this was written; AGPLv3 as of Phase 6" — rather than rewritten as if the commercial-license period never happened, matching this project's existing convention for superseded claims (e.g. Phase 3's tenant_id gap, corrected inline by Phase 4's section rather than edited out of Phase 3's).
• enterprise/README.md, enterprise/Dockerfile, enterprise/cmd/enterprise-auth/main.go: relicensing statement corrected.
• hack/check-tenant-boundary.sh and every doc describing it (docs/architecture.md, docs/security/threat-model.md, the two phase-4-*-design.md docs): the import-boundary check itself is kept — it still enforces a real, valuable architectural property (core builds and deploys standalone with zero multi-tenant mechanism present, tenant identity resolution stays server-side) — but every description of why it exists was reframed from a licensing reason to an architectural one, since both sides now carry the same license.
• Final repo-wide grep for commercial confirms every remaining occurrence is one of the above corrections (explicitly framed as historical/superseded), not a live claim. No file in the repo claims a license other than AGPLv3 for Sentry's own code, as of this audit.

Ongoing enforcement (task 7)

.github/workflows/license-compliance.yml — this repo's first CI workflow file (several docs already said "enforced in CI" about hack/check-tenant-boundary.sh, but no CI system had actually been wired up yet; fixed as part of this task rather than left as a second gap next to the one this task asked about). Four jobs, one per ecosystem plus the architectural boundary check, all real commands verified locally against this repo before being written into the workflow (not guessed):

• Rust: cargo-deny-action running cargo deny check licenses against agent/deny.toml and search/deny.toml — both verified passing locally with the policy's real allow-list.
• Go: go-licenses check ./... --allowed_licenses=... per module with real dependencies — verified passing locally for every listed module, including the segmentio/asm and HashiCorp-MPL-2.0 cases.
• npm: license-checker --onlyAllow "..." — verified passing locally.
• Architectural boundary: hack/check-tenant-boundary.sh, now actually wired into CI instead of only documented as if it were.

Full policy, including exactly what's auto-allowed, what needs manual review, and what's rejected outright: /docs/compliance/license-policy.md.

Not verified: the workflow YAML itself has not been run through a real GitHub Actions execution in this environment (no way to trigger that here) — every individual command it invokes was verified locally with real exit codes, but the workflow file's syntax and job wiring should be confirmed on the first real PR that triggers it, the same "written but not run against the live thing" caveat this project applies to its other CI-adjacent and Docker-gated claims.

What's resolved vs. what's still open

Per the audit brief's explicit gate: this phase is not "done" while a (c) or unresolved-(b) item has no recorded resolution. As of the Redpanda decision below, every item has one.

Item	Status
All 774 permissive/MPL-2.0/OR-resolved dependencies	Resolved — classification (a), no action needed.
segmentio/asm (Go)	Resolved — confirmed MIT-0, carried as a cited CI ignore.
favicon.svg	Flagged, not a compliance blocker — action item recorded (replace with an original mark), not license-gating.
Redpanda (BSL 1.1)	Resolved — decision recorded 2026-08-16: accept as-is (option 1). No code change; the BSL exposure is a disclosed, accepted risk, not an unresolved one.
enterprise/ relicensing	Resolved — applied throughout the repo, confirmed via repo-wide grep.
Root LICENSE / manifest declarations	Resolved — added and corrected.
CI enforcement	Resolved — workflow written and every command verified locally; the workflow file itself untested end-to-end (disclosed above).

Every task-2 (c)/unresolved-(b) item now has an explicit resolution — fixed, isolated, or, for Redpanda, flagged and decided. Phase 6's exit criteria in CLAUDE.md are updated accordingly.

Legal disclaimer (repeated, deliberately)

This audit and the policy derived from it are a strong first pass — real primary sources were checked for every named risk area and every flagged item, not guessed from memory. They are not a substitute for review by actual legal counsel, which is recommended before the project is publicly released, pitched to customers, or used as the basis for any compliance claim — particularly the open Redpanda question above, which is exactly the kind of judgment call outside counsel exists to make.