Files
frameledger/FrameLedger.Rules.targets
poli0981andClaude Opus 5 1ed07c08c6 feat(safety): the Agent seeds the rules file the guard reads (#27)
Nothing in this repository ever wrote the file the hard gate reads, so on any
machine that had not hand-installed one the guard answered `RulesUnreadable` for
every title. Correct fail-closed behaviour, and also the whole story: the first
real injection's opening refusal was exactly this.

Measured end to end: remove the file and the guard says `RulesUnreadable`; run the
Agent and the guard reads its rules and reaches check 1. Four Agent branches
exercised against the machine's real rules location, with its own file backed up
and restored by hash — already-current, installed-from-nothing, idempotent second
run, corrupt-file-replaced.

## Four decisions, each replacing something the design review disproved

- **Provenance, not `rulesVersion`.** The first design replaced the installed file
  when the packaged seed was strictly newer. Measured against this repository's own
  history, `31825cf` and `ba5355e` both changed the `anticheat` block without
  touching `rulesVersion`, and `a4a2c63` — the only commit that bumped it — changed
  the block not at all. The rule would have delivered NONE of the two changes it
  existed for. The seeder records a hash of what it installed instead.
- **`ReplaceFileW`, not `MoveFileEx`.** Measured: `MoveFileEx(REPLACE_EXISTING)`
  returns `ERROR_ACCESS_DENIED` against a reader holding the guard's exact share
  mode. Its backup parameter also makes `05_DETECTION`'s "the last valid copy is
  kept" a mechanism rather than a sentence.
- **Validated by the guard's own parser.** `FlGuardCheckRules` — buffer in,
  `ParseResult` out, no path parameter, so §S3/§S21's "the rules source is not
  selectable" is untouched. `DetectionRulesFile` never reads the `anticheat` block
  (§S15), so validating with the managed reader would have checked everything
  except the half the hard gate consumes, and could have installed a document the
  guard then refuses for every title while the seeder logged success.
- **A usable file we did not write is left alone.** Safe only because the floor is
  generated from the shipped blocklist: a rules file can ADD and cannot remove.
  Under the narrow floor the same rule would have handed permanent control of the
  blocklist to whoever created the file first. An UNUSABLE file is replaced whoever
  wrote it — there is nothing to clobber and nothing else repairs it.

## Mechanics

Random temp name in the DESTINATION directory (same volume, and `%TEMP%` is the
inherited-environment vector §S21 closed), opened `FileShare.None`, flushed
through to disk before the swap, and a reparse-point check on the directory chain.

## The caller is real

`FrameLedger.Agent/Program.cs` was `return 0`. It now runs the seeder and reports
the outcome, and a no-op seeder fails 8 Application tests and 2 Infrastructure
tests — proven by making it one.

## One regression caught by an existing test

`RuleFixtureCorpusTests` walks up from the test binary looking for a directory
called `rules`, and this puts one in the test's own output. The walk stopped
immediately and every fixture went missing. Re-keyed on `FrameLedger.slnx`.

## §S20 does not close

Rewritten to the residual rather than deleted: the HTTPS feed does not exist so
FR-7.3 is unmet; the guard still reads no version so binary and data have no
handshake, and `13_CI_CD`'s claim that rules-publish checks the bump was false and
is corrected; `trustedSigners` is the one allow-widening field a foreign file
still controls, inert until §S19(b) wires the signer half; and this is what makes
the Vulkan layer's §S2 self-scan reachable for the first time.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-04 16:59:00 +07:00

49 lines
2.1 KiB
XML

<Project>
<!--
Ships rules/detection-rules.json beside the importing project's output, as
`rules\detection-rules.json`.
§S20: the guard reads its blocklist from Local AppData and nothing in this
repository ever put a file there, so on any machine that has not hand-installed
one the guard answers RulesUnreadable for every title. The first half of fixing
that is having the seed in a build output at all — before this file, it was in
no csproj, no props and no targets.
Imported by exactly two projects:
- FrameLedger.Agent — the sole host (§S18 blocker 3, ratified),
and the only thing that seeds.
- FrameLedger.Infrastructure.Tests — asserts the shipped copy is the file CI
validated, and drives the seeder end to end.
NOT imported by FrameLedger.App, for the same reason it does not carry the
guard DLL.
-->
<PropertyGroup>
<FlSeedRules>$(MSBuildThisFileDirectory)rules\detection-rules.json</FlSeedRules>
</PropertyGroup>
<ItemGroup>
<None Include="$(FlSeedRules)">
<Link>rules\detection-rules.json</Link>
<CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
</None>
</ItemGroup>
<!--
An ERROR, not a warning, and no Exists() condition on the item above.
FrameLedger.Guard.targets warns, because the native build may legitimately not
have run yet on a fresh clone. This file is different: rules/detection-rules.json
is checked in, so its absence is not a state anyone builds through — it means
the repository is wrong. Copying the Guard.targets warn-half here would be the
gate-that-cannot-fail shape: the seeder would ship, find nothing to seed, and
report success.
-->
<Target Name="FlErrorOnMissingSeedRules" BeforeTargets="Build"
Condition="!Exists('$(FlSeedRules)')">
<Error Text="rules/detection-rules.json is missing at $(FlSeedRules). It is checked in; a build without it would produce an Agent that seeds nothing and says it succeeded." />
</Target>
</Project>