8000 Resolve silo paths from configuration instead of hardcoded directory names by simon-inkie · Pull Request #9 · simon-inkie/inkie-deck · GitHub
[go: up one dir, main page]

Skip to content

Resolve silo paths from configuration instead of hardcoded directory names - #9

Merged
simon-inkie merged 6 commits into
mainfrom
silo-paths-from-config
Aug 18, 2026
Merged

simon-inkie merged 6 commits into
mainfrom
silo-paths-from-config

Conversation

@simon-inkie
@simon-inkie simon-inkie commented Aug 17, 2026 •
Copy link
Copy Markdown
Owner

Four sites were flagged as sharing one root cause: a silo directory name written into the code as a literal, where the surrounding code resolves it from configuration. Reading each one, three were real and one was not, and the two that looked like the same bug turned out to be different bugs.

What changed

1. cron.server.ts - nudge run-state went to a hardcoded directory. The highest-severity of the set, and an outright blocker on the Scheduler against real data. The generated nudge script wrote fired/delivered state to a hardcoded path, and listCrons independently re-derived the same hardcoded path to read it back. Writer and reader agreed with each other and disagreed with the deployment, so on any instance whose agent directories are not at that exact path, state written by a fired nudge was invisible to the UI, and editing a nudge started a second state tree beside the first - splitting the per-fire-slot marker that makes delivery idempotent.

Both halves now resolve from one source, the INBOX_ROOT every other silo-adjacent module already uses. cron-state/ sits beside inbox/ inside each agent's directory, so INBOX_ROOT is the correct root and no new configuration is introduced. Injected through CronDeps like every other side effect here, and resolved lazily so a rejected target still fails as a bad target rather than as a configuration error.

The root is a value rather than a literal now, so it is also shell quoted. Making it configurable puts an operator-supplied path inside the generated script's double quotes, where a $ or a backtick would expand; the old hardcoded literal had no injection surface to begin with, this quoting closes a risk the refactor itself would otherwise have introduced, not one in the code being replaced.

2. review.server.ts - a substring ban the confinement rework missed. Not the same bug as the others. TASK-434 replaced deny-known-bad with allowlist confinement across safe-fs, attachments and projects, for the reason recorded in safe-fs's own header: a substring ban on a silo directory name is a build-era invariant encoded at the wrong layer, and it blocks the very paths production must serve. review.server.ts was not converted with the others.

Two things were wrong with what was left. The banned string is a directory name no deployment is obliged to use, so on most instances it denied nothing while reading like a protection. And resolveReviewRoots ALWAYS appends INBOX_ROOT as an implicit root - so on an instance that did use that name, the ban and the root list contradicted each other outright, and the panel refused every read from the one root it is guaranteed to have.

Containment against realpath'd roots is strictly stronger than a substring test for the traversal and symlink-escape cases the ban was reaching for, so nothing load-bearing goes with it. The tests that asserted the ban are replaced by tests for the property that actually holds the surface: a file inside a silo-shaped root is served, one outside every root is refused, and an escape through an intermediate symlink is refused on the realpath.

3. safe-fs.server.ts - MC_SANDBOX_ONLY could not name what it was guarding. README states FND-04's other half with no directory in it ("a sandbox configuration may never be pointed at a live agent silo"). The implementation could only ever recognise one directory: this project's default silo name joined onto $HOME.

On any deployment keeping its agents elsewhere the flag was not weaker, it was inert - it refused nothing, while .env.example described it as a protection and test/setup.ts set it on every test run.

New optional MC_LIVE_SILO_ROOT names the silo. Unset, the default is exactly the previous behaviour, so no existing instance changes and no existing test moves. Absolute paths only, validated at point of use rather than at boot, because the variable means nothing unless MC_SANDBOX_ONLY is also set and failing a production start over an unused sandbox setting would be the wrong trade.

This also makes the refusal testable for the first time. The single test covering it returned early unless the machine happened to have the default silo directory - which no CI machine does - so it had been passing by skipping. New cases stand a tmpdir up as the live silo and assert the refusal for real: the root itself, a sibling sharing its name prefix, the unset-flag no-op, and a blocked inbox write end to end.

4. create-agent.sh - not the reported bug; a false citation instead. The flagged line is a fallback, not a hardcode: --inbox-root and $INBOX_ROOT both win ahead of it, and when it does fire it warns loudly and prints where the agent went. That design is right and is left alone.

5. review.server.ts - REVIEW_ROOTS was vetted by nothing. Held back from the first pass and closed here, on review. Dropping the substring ban left MC_SANDBOX_ONLY with nothing to say about REVIEW_ROOTS: safe-fs's ROOT_ENV_VARS table is vetted once at boot and INBOX_ROOT is in it, but REVIEW_ROOTS is a fifth, separately parsed root list that was in no table at all. So with the flag set, an operator could name the live silo there and read it back through the review API - wider than the deleted code allowed on the default layout, and a defeat of the new MC_LIVE_SILO_ROOT on any other one. The first pass recorded that in the module header and deferred it. Naming a hole in a comment does not close it.

The deferral was right about the constraint and wrong about there being no home for the fix. Per-root assertNoLiveSiloRoot inside assertContained does land on the autolink path, whose realpath cost a regression test bounds - but root RESOLUTION already happens once per request, not once per path. resolveReviewRoots now vets every root it is about to return, the configured entries and the implicit INBOX_ROOT alike, memoised on the root list plus the two env vars the refusal reads. A running instance pays once; the autolink path pays a string compare, and that regression test still measures 4 realpath calls for 3 blocks. Only a pass is recorded, so a refusal cannot memoise itself into a pass.

The memoisation in that last paragraph was wrong, and is gone. Section 6 below is the second review round finding it and the fix. This paragraph is left standing as written, because it is what this branch actually claimed and what a reviewer had to reproduce a bypass to disprove. The placement it describes - vetting inside resolveReviewRoots rather than only at boot - is unchanged and was confirmed correct.

Boot is the same code under the name validateReviewRoots, called by the root loader beside validateConfiguredRoots, so a sandbox instance pointed at a live silo dies at start with the root named. The per-resolution check is not redundant with it: /api/review is a resource route, and RR7 runs a resource route's own loader only, never the root loader, so on a cold process the boot check alone would not cover the exact request the finding describes.

Two smaller calls. A configured root that does not exist yet is vetted where it sits rather than skipped, because it starts serving the moment the directory appears. And the implicit INBOX_ROOT is vetted here too, even though the boot table carries it, because on that same cold resource-route request the boot table has not run.

What was wrong was the wording. It announced the value as "the value from the application's .env.example", and that file ships INBOX_ROOT empty because there is no default (FND-03). The citation dressed a guess up as a lookup, which is the one thing a warning about guessing must not do.

No behaviour change. Tests now pin the source ordering, since a fallback firing over something the operator had configured is the failure that would actually matter, and nothing covered it - the existing case passes --inbox-root and never reaches the fallback branch.

6. review.server.ts - the vetting was right, memoising its answer was not. Second review round. Codex read the commit above and did not stop at the reasoning: it reproduced a live read out of a silo, in one process, against this branch.

The memo was keyed on the root list plus MC_SANDBOX_ONLY and MC_LIVE_SILO_ROOT, described in that commit as "everything the answer depends on". It is not. assertNoLiveSiloRoot answers from the FILESYSTEM: it returns early when the silo cannot be realpath'd, and both sides of its containment comparison are realpaths. None of that is a function of the configuration, so no key over the configuration can stand in for it. The repro, verbatim in shape:

  1. MC_SANDBOX_ONLY=1, MC_LIVE_SILO_ROOT=<tmp>/silo, REVIEW_ROOTS=<tmp>/silo/agents, and no silo on disk.
  2. resolveReviewRoots() passes - correctly, there is nothing to refuse yet - and records the pass.
  3. Create silo/agents/private.md.
  4. readReviewFile(private.md) returns the contents, off a pass recorded before the directory existed, with the env and the raw root list byte-for-byte unchanged.

Retargeting or replacing a symlinked root after a pass is the same bug wearing a different hat.

The fix is the shape Codex asked for: memoise only while MC_SANDBOX_ONLY is unset, where the refusal is a no-op and a recorded pass cannot go stale, and re-vet on every resolveReviewRoots() call while the guard is active. vetReviewRoots reads the flag only to decide how long an answer stays valid, never to decide what the answer is, so assertNoLiveSiloRoot stays the one authority on the refusal itself and the vetting loop still runs at least once for any distinct root list whatever the flag says. Route-level placement untouched.

The re-vet is bounded at the granularity Codex signed off on - once per root RESOLUTION, not once per path or per block. Every caller here already resolves once per request: linkifyBlockBodies hoists resolveReviewRoots out of its block loop precisely so that cost is per request, and the two read entry points call it once each.

Two regression tests: the reproduction above through readReviewFile, and the symlink-retarget variant. Both fail on the parent commit, and the first fails by returning the silo file's contents rather than by any incidental difference.

The autolink perf regression test needed restating rather than relaxing. It asserted an absolute bound of 4 realpath calls, measured with the memo pre-warmed outside the window, and there is no longer anything to pre-warm. It now measures three blocks against six and asserts the counts are EQUAL - the property it was always about (nothing scales with the number of BLOCKS) stated directly instead of encoded as a constant - plus a bound of 8 on the per-call figure so a fixed cost cannot balloon unnoticed. 8 is the vetting of 2 roots, each realpath'd and compared against a realpath'd silo, plus assertContained's 4 paid once for the one distinct candidate. Without the sharing that second half alone would be 4 per block, so the regression the test was written for still fails it.

Deliberately not done

  • No shared constant for the silo name. There is not one today, and introducing one would only centralise a default that two of the three call sites should no longer have at all.
  • Existing generated nudge scripts still carry the old path. Anything created before this lands keeps writing where it was told to until the entry is edited through the app, which rewrites the script. Worth a note in release notes rather than a migration.

Also worth knowing

  • assertGitClean in review.test.ts and safe-fs.test.ts allowlists paths like app/ and test/, but git status --porcelain reports repo-root-relative paths, which since the monorepo layout are packages/deck/app/.... The allowlist therefore matches nothing and the assertion fires on any uncommitted change under the package. Green on a committed tree, so CI is unaffected, but it makes the suite unusable mid-edit. Left alone as unrelated to this change.
  • The suite is load-flaky, as the root package.json already documents. Two separate runs each produced one failure (skeleton.test.ts on a 5s dynamic-import timeout; one base case), both passing in isolation and on a clean re-run.

Checks

pnpm test (base 41, deck 1846 + 11 skipped, 89 files), pnpm typecheck, pnpm check:boundaries (583 modules), pnpm check:dashes, pnpm check:licence, leak-gate --self-test (30 rules), leak-gate (clean, 367 files) all green.

leak-gate --messages reports two findings, both pre-existing on historical commits (2dc627dc14, c95abce7bc) and identical to the pre-change baseline. None of the commits here appear.

Re-run in a clean clone after the follow-up commit that closes the REVIEW_ROOTS gap: pnpm test (base 41, deck 1845 + 11 skipped, 88 files - the 9 new cases included), pnpm typecheck, pnpm check:boundaries (483 modules), pnpm check:dashes, pnpm check:licence, leak-gate --self-test (30 rules), leak-gate (clean, 366 files) all green. leak-gate --messages still reports the same two pre-existing findings and nothing from this branch.

Re-run in a clean clone after the memoisation fix (section 6): deck pnpm test 1847 passed + 11 skipped across 88 files (2 new cases), pnpm typecheck, pnpm check:boundaries (483 modules, 858 dependencies), leak-gate --self-test (30 rules), leak-gate (clean, 366 files), leak-gate --history (clean, 567 blobs / 447 object paths) all green.

leak-gate --messages, run against a clone of this branch alone, reports one finding: c95abce7bc, the Co-authored-by trailer GitHub's squash-merge UI wrote onto main when PR #1 landed. Byte-identical at the parent commit f209ec1, so this branch introduces nothing - and there is a separate branch already in flight for that rule. The earlier note above says two findings because it was run in a checkout carrying other local branches; the second (2dc627dc14) is not reachable from this branch. Nothing from the commits here appears in either scope.

🤖 Generated with Claude Code

Simon added 4 commits August 17, 2026 22:43
The generated nudge script wrote its fire/delivered state to a directory
named by a hardcoded silo path, and listCrons independently re-derived the
same hardcoded path to read it back. The two agreed with each other and
disagreed with the deployment, so on any instance whose agent directories
are not at that exact path:

  - run-state written by a fired nudge was invisible to the Scheduler UI,
    which is the whole of the DC4 visibility feature; and
  - editing a nudge through the app rewrote the script against the same
    wrong path, so the per-fire-slot marker that makes delivery idempotent
    started over in a second tree beside the first.

Both halves now come from one configured source, the INBOX_ROOT that every
other silo-adjacent module already resolves through. cron-state sits beside
inbox/ inside each agent's directory, so INBOX_ROOT is the correct root and
no new configuration is introduced. Injected through CronDeps like every
other side effect in this module, and resolved lazily so a rejected target
still fails as a bad target rather than as a configuration error.

The root is now a value rather than a literal, so it is also now shell
quoted. It went into the script inside double quotes, where a dollar sign
or a backtick in an operator's path would have expanded.
…missed

TASK-434 replaced deny-known-bad with allowlist confinement across safe-fs,
attachments and projects, for a reason recorded in safe-fs's header: a
substring ban on a silo directory name is a build-era invariant encoded at
the wrong layer, and it blocks the production paths the app is meant to
serve. review.server.ts was not converted with the others and kept the ban.

Two things were wrong with what was left behind. The banned string is a
directory name no deployment is obliged to use, so on most instances the
check denied nothing while reading like a protection. And resolveReviewRoots
ALWAYS appends INBOX_ROOT as an implicit root, so on an instance that did
use that name the ban and the root list contradicted each other outright:
every read from the one root the panel is guaranteed to have was refused.

Containment against realpath'd roots is strictly stronger than a substring
test for the traversal and symlink-escape cases the ban was reaching for,
so nothing load-bearing goes with it. The tests that asserted the ban are
replaced by tests for the property that actually holds the surface: a file
inside a silo-shaped root is served, one outside every root is refused, and
an escape through an intermediate symlink is refused on the realpath.

Records one gap this deliberately does not close. The sandbox live-silo
refusal covers the four roots in safe-fs's ROOT_ENV_VARS table, which
validateConfiguredRoots vets once at boot, and INBOX_ROOT is in it. Nothing
vets REVIEW_ROOTS. The obvious fix lands on the autolink hot path, whose
realpath cost is measured by a regression test, so it wants a
once-per-request home rather than a per-path syscall.
MC_SANDBOX_ONLY is the half of FND-04 that stops a sandbox configuration
being pointed at a live agent silo. README states that rule with no
directory in it. The implementation could only ever recognise one: the silo
directory name this project ships, joined onto the home directory.

On any deployment that keeps its agents anywhere else, the flag was inert.
Not weaker, inert: it refused nothing at all, while .env.example described
it as a protection and test/setup.ts set it on every test run. A guard that
cannot name the thing it is guarding is a comment.

MC_LIVE_SILO_ROOT names it. Unset, the default is exactly what the code did
before, so no existing instance changes behaviour and no test moves. Set, it
points the refusal at whatever a real fleet actually uses. Absolute paths
only, validated where it is used rather than at boot, because the variable
means nothing unless MC_SANDBOX_ONLY is set too and failing a production
start over an unused sandbox setting would be the wrong trade.

This also makes the refusal testable for the first time. The one test that
covered it returned early unless the machine running it happened to have the
default silo directory, which no CI machine does, so it had been passing by
skipping. The new cases stand a tmpdir up as the live silo and assert the
refusal for real, including the root itself, a sibling sharing its name
prefix, the unset-flag no-op, and a blocked inbox write end to end.
…claims

When INBOX_ROOT is unset and no --inbox-root is passed, create-agent.sh
falls back to this project's default silo layout and says so loudly, which
is the right design: an inbox created in the wrong place should be visible
when it happens, not the first time a message goes missing.

The wording undercut it. It announced the value as "the value from the
application's .env.example", and that file ships INBOX_ROOT empty, because
there is no default (FND-03). The citation dressed a guess up as a lookup,
which is the one thing a warning about guessing must not do. It now says
plainly that nothing here knows where the reader's inbox root is and that
the value below is a guess.

No behaviour change, and the fallback value is deliberately left alone. It
is the last of three sources and both of the ones ahead of it already win.
Tests now pin that ordering, since the fallback firing over something the
operator had configured is the failure that would actually matter, and no
test covered it: the existing case passes --inbox-root and never reaches
the fallback branch at all.
Simon and others added 2 commits August 18, 2026 10:56
…atch

Dropping the substring ban from this module left MC_SANDBOX_ONLY with
nothing to say about REVIEW_ROOTS. safe-fs's ROOT_ENV_VARS table is vetted
once at boot and INBOX_ROOT is in it, but REVIEW_ROOTS is a fifth,
separately parsed root list that was in no table at all. So with the flag
set, an operator could name the live agent silo in REVIEW_ROOTS and read it
back through the review API: wider than what the deleted code allowed on the
default layout, and a defeat of MC_LIVE_SILO_ROOT on any other one. The
module header recorded that as a known gap and deferred it. Naming a hole in
a comment does not close it, and it should not have shipped open.

The deferral reasoned that the fix wanted a once-per-request home, because
the obvious placement - assertNoLiveSiloRoot per root inside assertContained
- lands on the autolink path, whose realpath cost linkifyBlockBodies shares
across blocks deliberately and a regression test bounds. That reasoning was
right about the constraint and wrong about there being no home: root
RESOLUTION already happens once per request, not once per path.

resolveReviewRoots now vets every root it is about to return, the configured
entries and the implicit INBOX_ROOT alike, and throws before any of them
reaches a caller. It is memoised on the root list plus the two env vars the
refusal reads, so a running instance pays for it once and the autolink path
pays a string compare; the perf regression test still measures 4 realpath
calls for 3 blocks, unchanged. Only a pass is recorded, so a refusal can
never memoise itself into a pass.

Boot is covered by the same code under the name validateReviewRoots, called
by the root loader beside validateConfiguredRoots, so a sandbox instance
pointed at a live silo dies at start with the offending root named rather
than looking healthy until someone opens the panel. The per-resolution check
is not redundant with it: /api/review is a resource route, and RR7 runs a
resource route's own loader only, never the root loader, so on a cold process
the boot check alone would not have covered the exact request the finding
describes.

Two smaller decisions worth stating. A configured root that does not exist
yet is vetted where it sits rather than skipped: containment skips it today,
but it starts serving the moment the directory appears. And the implicit
INBOX_ROOT is vetted here too even though the boot table already carries it,
because on that same cold resource-route request the boot table has not run.

Tests: eight cases in review.test.ts (silo root itself, a root below it, the
read end to end, a not-yet-created entry, the implicit INBOX_ROOT, the boot
entry point, a sibling sharing the silo's name prefix, and the unset-flag
production cutover), plus one in review-autolink.test.ts pinning that the
render path fails CLOSED - no links into a refused silo, and no throw out of
a render. All seven refusal cases fail without this change.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…disk

The previous commit put the MC_SANDBOX_ONLY live-silo refusal in the right
place and then cached its answer in the wrong shape. The memo was keyed on the
root list plus MC_SANDBOX_ONLY and MC_LIVE_SILO_ROOT, described in that commit
as "everything the answer depends on". It is not. assertNoLiveSiloRoot answers
from the FILESYSTEM: it returns early when the silo cannot be realpath'd, and
both sides of its containment comparison are realpaths. None of that is a
function of the configuration, so no key over the configuration can stand in
for it.

Codex reproduced the bypass in one process, not in theory. Configure a sandbox
instance with MC_LIVE_SILO_ROOT=<tmp>/silo and REVIEW_ROOTS=<tmp>/silo/agents
while the silo does not exist. resolveReviewRoots passes, correctly, because
there is no silo to resolve, and records that pass. Create silo/agents and
private.md afterwards and readReviewFile(private.md) returns the contents, off
a pass recorded before the directory existed, with the env and the raw root
list byte-for-byte unchanged. Retargeting a symlinked root after a pass is the
same bug wearing a different hat.

The fix is the shape Codex asked for: memoise only while MC_SANDBOX_ONLY is
unset, where the refusal is a no-op and a recorded pass cannot go stale, and
re-vet on every resolveReviewRoots() call while the guard is active. The flag
is read in vetReviewRoots only to decide how long an answer stays valid, never
to decide what the answer is, so assertNoLiveSiloRoot stays the one authority
on the refusal itself and the vetting loop still runs at least once for any
distinct root list whatever the flag says. The route-level placement is
untouched: that half of the last commit was right and the finding said so.

The re-vet is bounded at the granularity Codex signed off on, once per root
RESOLUTION rather than once per path or per block. Every caller on this surface
already resolves once per request: linkifyBlockBodies hoists resolveReviewRoots
out of its block loop precisely so the realpath cost is paid per request, and
the read entry points call it once each.

Tests: two regression cases, the reproduction above through readReviewFile and
the symlink-retarget variant Codex named as the same class. Both fail on the
parent commit, and the create-after-pass one fails by returning the silo file's
contents, not by some incidental difference.

The autolink perf regression test needed restating rather than relaxing. It
asserted an absolute bound of 4 realpath calls, measured with the memo pre-
warmed outside the window, which is no longer a thing that can be pre-warmed.
It now measures three blocks against six and asserts the counts are EQUAL,
which is the property it was always about (nothing scales with the number of
BLOCKS) stated directly instead of encoded as a constant, plus a bound of 8 on
the per-call figure so a fixed cost cannot balloon unnoticed. 8 is the vetting
of 2 roots, each realpath'd and compared against a realpath'd silo, plus
assertContained's 4 paid once for the one distinct candidate. Without the
sharing that second half alone would be 4 per block, so the regression the test
was written for still fails it.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
@simon-inkie
Copy link
Copy Markdown
Owner Author

Folio (independent review, round 3, model-diverse) — 2026-08-18

Re-reviewed through 900ca48. The sandbox-only cache gating closes both stale-filesystem cases: create-after-pass and symlink retargeting are re-vetted on the next resolution, and enabling MC_SANDBOX_ONLY cannot reuse an inactive-state cache entry. I also traced file reads, asset reads, resource routes and autolinking through the guarded resolver, plus the other safe-fs call sites; no equivalent stale-pass path remains. Focused tests, typecheck and boundaries pass. I would approve this if GitHub allowed self-review.

@simon-inkie
simon-inkie merged commit 02041a6 into main Aug 18, 2026
4 checks passed
@simon-inkie
simon-inkie deleted the silo-paths-from-config branch August 18, 2026 15:27
simon-inkie added a commit that referenced this pull request Aug 18, 2026
…names (#9)

* fix(cron): resolve the nudge state directory from INBOX_ROOT

The generated nudge script wrote its fire/delivered state to a directory
named by a hardcoded silo path, and listCrons independently re-derived the
same hardcoded path to read it back. The two agreed with each other and
disagreed with the deployment, so on any instance whose agent directories
are not at that exact path:

  - run-state written by a fired nudge was invisible to the Scheduler UI,
    which is the whole of the DC4 visibility feature; and
  - editing a nudge through the app rewrote the script against the same
    wrong path, so the per-fire-slot marker that makes delivery idempotent
    started over in a second tree beside the first.

Both halves now come from one configured source, the INBOX_ROOT that every
other silo-adjacent module already resolves through. cron-state sits beside
inbox/ inside each agent's directory, so INBOX_ROOT is the correct root and
no new configuration is introduced. Injected through CronDeps like every
other side effect in this module, and resolved lazily so a rejected target
still fails as a bad target rather than as a configuration error.

The root is now a value rather than a literal, so it is also now shell
quoted. It went into the script inside double quotes, where a dollar sign
or a backtick in an operator's path would have expanded.

* fix(review): drop the silo-name substring ban the confinement rework missed

TASK-434 replaced deny-known-bad with allowlist confinement across safe-fs,
attachments and projects, for a reason recorded in safe-fs's header: a
substring ban on a silo directory name is a build-era invariant encoded at
the wrong layer, and it blocks the production paths the app is meant to
serve. review.server.ts was not converted with the others and kept the ban.

Two things were wrong with what was left behind. The banned string is a
directory name no deployment is obliged to use, so on most instances the
check denied nothing while reading like a protection. And resolveReviewRoots
ALWAYS appends INBOX_ROOT as an implicit root, so on an instance that did
use that name the ban and the root list contradicted each other outright:
every read from the one root the panel is guaranteed to have was refused.

Containment against realpath'd roots is strictly stronger than a substring
test for the traversal and symlink-escape cases the ban was reaching for,
so nothing load-bearing goes with it. The tests that asserted the ban are
replaced by tests for the property that actually holds the surface: a file
inside a silo-shaped root is served, one outside every root is refused, and
an escape through an intermediate symlink is refused on the realpath.

Records one gap this deliberately does not close. The sandbox live-silo
refusal covers the four roots in safe-fs's ROOT_ENV_VARS table, which
validateConfiguredRoots vets once at boot, and INBOX_ROOT is in it. Nothing
vets REVIEW_ROOTS. The obvious fix lands on the autolink hot path, whose
realpath cost is measured by a regression test, so it wants a
once-per-request home rather than a per-path syscall.

* feat(safe-fs): let MC_SANDBOX_ONLY name the silo it is protecting

MC_SANDBOX_ONLY is the half of FND-04 that stops a sandbox configuration
being pointed at a live agent silo. README states that rule with no
directory in it. The implementation could only ever recognise one: the silo
directory name this project ships, joined onto the home directory.

On any deployment that keeps its agents anywhere else, the flag was inert.
Not weaker, inert: it refused nothing at all, while .env.example described
it as a protection and test/setup.ts set it on every test run. A guard that
cannot name the thing it is guarding is a comment.

MC_LIVE_SILO_ROOT names it. Unset, the default is exactly what the code did
before, so no existing instance changes behaviour and no test moves. Set, it
points the refusal at whatever a real fleet actually uses. Absolute paths
only, validated where it is used rather than at boot, because the variable
means nothing unless MC_SANDBOX_ONLY is set too and failing a production
start over an unused sandbox setting would be the wrong trade.

This also makes the refusal testable for the first time. The one test that
covered it returned early unless the machine running it happened to have the
default silo directory, which no CI machine does, so it had been passing by
skipping. The new cases stand a tmpdir up as the live silo and assert the
refusal for real, including the root itself, a sibling sharing its name
prefix, the unset-flag no-op, and a blocked inbox write end to end.

* fix(base): stop create-agent citing a file that does not say what it claims

When INBOX_ROOT is unset and no --inbox-root is passed, create-agent.sh
falls back to this project's default silo layout and says so loudly, which
is the right design: an inbox created in the wrong place should be visible
when it happens, not the first time a message goes missing.

The wording undercut it. It announced the value as "the value from the
application's .env.example", and that file ships INBOX_ROOT empty, because
there is no default (FND-03). The citation dressed a guess up as a lookup,
which is the one thing a warning about guessing must not do. It now says
plainly that nothing here knows where the reader's inbox root is and that
the value below is a guess.

No behaviour change, and the fallback value is deliberately left alone. It
is the last of three sources and both of the ones ahead of it already win.
Tests now pin that ordering, since the fallback firing over something the
operator had configured is the failure that would actually matter, and no
test covered it: the existing case passes --inbox-root and never reaches
the fallback branch at all.

* fix(review): vet REVIEW_ROOTS against the live silo the ban used to catch

Dropping the substring ban from this module left MC_SANDBOX_ONLY with
nothing to say about REVIEW_ROOTS. safe-fs's ROOT_ENV_VARS table is vetted
once at boot and INBOX_ROOT is in it, but REVIEW_ROOTS is a fifth,
separately parsed root list that was in no table at all. So with the flag
set, an operator could name the live agent silo in REVIEW_ROOTS and read it
back through the review API: wider than what the deleted code allowed on the
default layout, and a defeat of MC_LIVE_SILO_ROOT on any other one. The
module header recorded that as a known gap and deferred it. Naming a hole in
a comment does not close it, and it should not have shipped open.

The deferral reasoned that the fix wanted a once-per-request home, because
the obvious placement - assertNoLiveSiloRoot per root inside assertContained
- lands on the autolink path, whose realpath cost linkifyBlockBodies shares
across blocks deliberately and a regression test bounds. That reasoning was
right about the constraint and wrong about there being no home: root
RESOLUTION already happens once per request, not once per path.

resolveReviewRoots now vets every root it is about to return, the configured
entries and the implicit INBOX_ROOT alike, and throws before any of them
reaches a caller. It is memoised on the root list plus the two env vars the
refusal reads, so a running instance pays for it once and the autolink path
pays a string compare; the perf regression test still measures 4 realpath
calls for 3 blocks, unchanged. Only a pass is recorded, so a refusal can
never memoise itself into a pass.

Boot is covered by the same code under the name validateReviewRoots, called
by the root loader beside validateConfiguredRoots, so a sandbox instance
pointed at a live silo dies at start with the offending root named rather
than looking healthy until someone opens the panel. The per-resolution check
is not redundant with it: /api/review is a resource route, and RR7 runs a
resource route's own loader only, never the root loader, so on a cold process
the boot check alone would not have covered the exact request the finding
describes.

Two smaller decisions worth stating. A configured root that does not exist
yet is vetted where it sits rather than skipped: containment skips it today,
but it starts serving the moment the directory appears. And the implicit
INBOX_ROOT is vetted here too even though the boot table already carries it,
because on that same cold resource-route request the boot table has not run.

Tests: eight cases in review.test.ts (silo root itself, a root below it, the
read end to end, a not-yet-created entry, the implicit INBOX_ROOT, the boot
entry point, a sibling sharing the silo's name prefix, and the unset-flag
production cutover), plus one in review-autolink.test.ts pinning that the
render path fails CLOSED - no links into a refused silo, and no throw out of
a render. All seven refusal cases fail without this change.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* fix(review): a live-silo pass is not memoisable, it answers from the disk

The previous commit put the MC_SANDBOX_ONLY live-silo refusal in the right
place and then cached its answer in the wrong shape. The memo was keyed on the
root list plus MC_SANDBOX_ONLY and MC_LIVE_SILO_ROOT, described in that commit
as "everything the answer depends on". It is not. assertNoLiveSiloRoot answers
from the FILESYSTEM: it returns early when the silo cannot be realpath'd, and
both sides of its containment comparison are realpaths. None of that is a
function of the configuration, so no key over the configuration can stand in
for it.

Codex reproduced the bypass in one process, not in theory. Configure a sandbox
instance with MC_LIVE_SILO_ROOT=<tmp>/silo and REVIEW_ROOTS=<tmp>/silo/agents
while the silo does not exist. resolveReviewRoots passes, correctly, because
there is no silo to resolve, and records that pass. Create silo/agents and
private.md afterwards and readReviewFile(private.md) returns the contents, off
a pass recorded before the directory existed, with the env and the raw root
list byte-for-byte unchanged. Retargeting a symlinked root after a pass is the
same bug wearing a different hat.

The fix is the shape Codex asked for: memoise only while MC_SANDBOX_ONLY is
unset, where the refusal is a no-op and a recorded pass cannot go stale, and
re-vet on every resolveReviewRoots() call while the guard is active. The flag
is read in vetReviewRoots only to decide how long an answer stays valid, never
to decide what the answer is, so assertNoLiveSiloRoot stays the one authority
on the refusal itself and the vetting loop still runs at least once for any
distinct root list whatever the flag says. The route-level placement is
untouched: that half of the last commit was right and the finding said so.

The re-vet is bounded at the granularity Codex signed off on, once per root
RESOLUTION rather than once per path or per block. Every caller on this surface
already resolves once per request: linkifyBlockBodies hoists resolveReviewRoots
out of its block loop precisely so the realpath cost is paid per request, and
the read entry points call it once each.

Tests: two regression cases, the reproduction above through readReviewFile and
the symlink-retarget variant Codex named as the same class. Both fail on the
parent commit, and the create-after-pass one fails by returning the silo file's
contents, not by some incidental difference.

The autolink perf regression test needed restating rather than relaxing. It
asserted an absolute bound of 4 realpath calls, measured with the memo pre-
warmed outside the window, which is no longer a thing that can be pre-warmed.
It now measures three blocks against six and asserts the counts are EQUAL,
which is the property it was always about (nothing scales with the number of
BLOCKS) stated directly instead of encoded as a constant, plus a bound of 8 on
the per-call figure so a fixed cost cannot balloon unnoticed. 8 is the vetting
of 2 roots, each realpath'd and compared against a realpath'd silo, plus
assertContained's 4 paid once for the one distinct candidate. Without the
sharing that second half alone would be 4 per block, so the regression the test
was written for still fails it.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

---------

Co-authored-by: Simon <simon@inkie.ink>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

0