Files
msd-core/.pr-body-3514.md
Tom Boucher 6badb839a0 fix(#3514): deny internal fetch hosts; disclose unverified integrity (#3516)
* test(#3514): add failing-first denylist and integrity suites

* fix(#3514): deny internal fetch hosts; disclose unverified integrity

* docs(#3514): trust-model, glossary, and changeset entries

* fix(#3514): scope v6 checks to literals; exact pin kinds in prompt

* chore(#3514): backfill changeset pr number

---------

Co-authored-by: sim <sim@local>
2026-08-14 21:34:29 -04:00

74 lines
4.4 KiB
Markdown

## Fix PR
> **Using the wrong template?**
> — Enhancement: use [enhancement.md](?template=enhancement.md)
> — Feature: use [feature.md](?template=feature.md)
---
## Linked Issue
> **Required.** This PR will be auto-closed if no valid issue link is found.
Fixes #3514
> The linked issue must have the `confirmed-bug` label. If it doesn't, ask a maintainer to confirm the bug before continuing.
---
## What was broken
Three hardening gaps at the ADR-1244 D3/D5 edges (epic #1900, finding F21): the capability URL-import fetcher never validated the resolved host (a loopback or cloud-metadata **https** endpoint was fetched like any URL); an `http://` tarball spec failed with a raw `ERR_INVALID_PROTOCOL` instead of a named reason; and an install without an integrity pin produced a consent prompt that did not distinguish verified from unverified content.
## What this fix does
- **Pre-transport fetch gate** (`realHttpsGet` → `assertFetchableUrl`): loopback/link-local/metadata/unspecified hosts (`127/8`, `169.254/16` incl. `169.254.169.254`, `0/8`, `::1`, `fe80::/10`, `::`, IPv4-mapped spellings in both dotted and hex-normalized forms) and `localhost`/`*.localhost` names are denied with a named error **before any bytes leave** — the injected transport is provably never invoked (tests assert call-count 0). The WHATWG URL parser normalizes alternate IP spellings (decimal `2130706433`, hex `0x7f000001`, octal, short forms) to dotted-quad before the gate sees them — locked by tests.
- **Non-https, per the adopted split decision**: `parseSpec` still *classifies* `http://` tarball specs (internal-mirror flows are not broken at parse); the fetcher refuses with a clear https-only reason naming the mirror alternative.
- **Unverified-integrity disclosure**: `evaluateInstallTrust` accepts `integrityPinned` (a verified `--integrity` pin, or a git `#sha:<commit>` ref); the consent prompt renders `content: NO PINNED HASH — staged unverified` when no pin was supplied (and the pinned counterpart when one was). The status is **prompt-only** — deliberately excluded from `disclosureSignature` (consent's content binding is `bundleContentHash`, #1459; tests lock that it can never fire a spurious re-consent). Legacy callers see byte-identical output.
## Root cause
The D3 fetcher was built with a byte cap and a timeout but no host policy — the finding is an edge the pipeline's safe posture simply wasn't applied to. The integrity line was missing because the disclosure was built from the manifest alone; the pin fact lives on the resolve path, and nothing threaded it into the verdict.
## Testing
### How I verified the fix
- Failing-first: `gsd-test` at the tests-only commit (holodeck, linux-node24) — verdict `failed` with exactly the 20 expected failures (12-case denylist matrix, https-reason, trust/lifecycle integrity suites), 34,343 green.
- GREEN: full `gsd-test` matrix at the final HEAD — verdict below.
- Controls lock the negative space: a public host and a private-range mirror still fetch through the gate; `parseSpec` still classifies `http://`; legacy `summarizeDisclosure` output is line-identical.
### Regression test added?
- [x] Yes — added a test that would have caught this bug
### Platforms tested
- [x] macOS
- [ ] Windows (including backslash path handling)
- [x] Linux
### Runtimes tested
- [ ] Claude Code
- [ ] Gemini CLI
- [ ] OpenCode
- [ ] Other: ___
- [x] N/A (not runtime-specific — engine-internal source resolver + trust gate)
---
## Checklist
- [x] Issue linked above with `Fixes #3514` — **PR will be auto-closed if missing**
- [x] Linked issue has the `confirmed-bug` label
- [x] Fix is scoped to the reported bug — no unrelated changes included
- [x] Regression test added (or explained why not)
- [x] All existing tests pass (`npm test`) — full `gsd-test` matrix at final HEAD
- [x] `.changeset/` fragment added — `Security` type
- [x] No unnecessary dependencies added
## Breaking changes
None intended. `http://` tarball URLs already failed at the transport; they now fail with a better message. Denied hosts (loopback/link-local/metadata/localhost) are not legitimate install sources. **Known limits, documented** in `docs/explanation/capability-trust-model.md`: RFC1918 private ranges are deliberately *not* denied (internal mirrors); the check is on the URL's host literal — DNS rebinding is out of scope. Local/unpinned-git installs render the unverified line even though their trust basis is the path/commit (honest, slightly noisy).