Repository navigation
test(e2e): prove Keychain secrets by reference end to end - #2612
Conversation
Add the bep_store_reference_end_to_end case to session-e2e.sh. It stores a value with min secret set, references it in a box spec, checks that an HTTPS echo upstream receives the stored value, and checks that no box file or variable holds it. The live proxy gate takes no_sign_in for cases that need no GitHub sign-in.
|
Important Review skippedBot user detected. To trigger a single review, invoke the
|
|
Reviewed: 1 finding, none blocking at Address the comments on the diff. Clobber: none seen. |
There was a problem hiding this comment.
arch-review: CONDITIONS · 2 conditions · the full ruling
The new e2e case follows Gatehouse §6.10, but two things are open. The audit check accepts a record without the store, and §6.10 line 849 says the record carries store and id. C1 makes the check require keychain:<id> exactly. The leg reported NOT RUN, and a NOT RUN leg is not a pass. C2 asks for a macOS run on this head. The --upstream flags on a host-store set conflict with architecture.md line 513. That question is an escalation, not a condition on this change.
Ruled at 51654d5 against gominimal/arch@109541a48c31 by claude-opus-5-5.
…ecrets-by-reference-end-to-end
The audit check accepted a bare identifier or any <store>:<id>. It now takes the store as an argument and requires <store>:<id> exactly, as StoreAdmit::identifier() writes it.
|
Round 1 of 10 on a failed check suite, pushed as
Verify Review threads the coder judged:
Not shown on this box. The case stores an API key with Notes from the agent: bep_store_jsonl_check now takes the store as $3 and requires identifier == store + ':' + id exactly. I checked the helper against sample records: keychain:abc passes, and a bare abc fails. The failed checks in the directive: the ci-macos-success log shows only the aggregator's 'exit 1'. The e2e job excerpt shows other cases (net expose) and does not show this case's output, so I cannot find a cause in this case from the excerpt. A person must read the full macOS e2e log for the bep_store_reference_end_to_end group. I cannot run the macOS lane from this Linux box. |
There was a problem hiding this comment.
arch-review: CONDITIONS · 2 conditions · the full ruling
The audit check now requires keychain:<id>, as §6.10 line 849 and StoreAdmit::identifier() require. That resolves the baseline point. Two points stay open. The only e2e leg reported NOT RUN, so the proof is not shown yet. Run the case on the macOS lane at this head (C1). The PR body states a span gap but names no issue. File the issue and name it in the PR body (C2).
Ruled at cc15f9f against gominimal/arch@109541a48c31 by claude-opus-5-5.
An interrupted run left the case's Keychain item in the developer's real Keychain, because only the case's own paths called sr_cleanup. The EXIT trap now runs `min secret rm` for SR_ID. The rule needs no step there: it lives under $WORK, which teardown removes.
|
Round 2 of 10 on the runtime's review findings, pushed as
Verify Review threads the coder judged:
On the macOS lane, a person can run ./scripts/session-e2e.sh bep_store_reference_end_to_end. The case stores an API key with Notes from the agent: Thread C3: teardown now runs |
There was a problem hiding this comment.
arch-review: CONDITIONS · 3 conditions · the full ruling
The teardown fix conforms. The case checks the redemption that Gatehouse §6.10 lines 844 and 849 require. The case is NOT RUN at this head, and it is the task's only proof. Run it on the macOS lane and record the pass (C1). The box spec writes the store reference as a broker scope, but §6.10 line 842 and BEP-032 write source = "store". File the tree gap and mark the case as an interim (C2). File the span gap that the PR body declares, or amend the plan line (C3).
Ruled at 1ec9e15 against gominimal/arch@109541a48c31 by claude-opus-5-5.
grep -r exits 2 on an unreadable file even after it lists matches. The case now fails on a match first, and reports grep's exit status and errors when the search does not finish. The comment also says that the default echo service is a public third-party host.
|
Round 3 of 10 on the runtime's review findings, pushed as
Verify Review threads the coder judged:
On macOS, a person can run ./scripts/session-e2e.sh bep_store_reference_end_to_end. It stores an API key with Notes from the agent: Revise round 3. On Linux the case reports NOT RUN, because the Box Egress Proxy runs on macOS only. The macOS lane must run it, and the PR needs its 'Keychain secret by reference OK' line. bash -n and shellcheck -S warning pass on the script. |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 2 conditions · the full ruling
The case conforms to §6.10 lines 844 and 849 and architecture.md line 501. This head fixes the last code point: the box scan reports matches when grep exits 2, and it still fails closed. Two points stay open. C1: the case reported NOT RUN, so the proof is open until the ci-macos lane reports it OK on this head. C2: file an issue for the missing store and identifier fields on the box_connection span, and name it in the PR body.
Ruled at e1774c6 against gominimal/arch@109541a48c31 by claude-opus-5-5.
There was a problem hiding this comment.
minimal-review — approval withheld
Verdict COMMENT — APPROVE withheld: 13 unresolved threads — minimal-arch[bot] scripts/session-e2e.sh, minimal-arch[bot] scripts/session-e2e.sh, minimal-arch[bot] scripts/session-e2e.sh:12188, minimal-arch[bot] scripts/session-e2e.sh:12188, minimal-arch[bot] scripts/session-e2e.sh:12361, …. No findings.
Panel: 3 of 3 personas ran (correctness, security, supply-chain).
The tests lane shellchecks every script. The store-reference box scan passes $? and $s in single quotes to the box shell, so it needs the directive. The comments also say that the seed spells the store reference as a broker scope for now, and that the connection spans carry no store or identifier.
|
Round 4 of 10 on a failed check suite, pushed as
Verify Review threads the coder judged:
On macOS, a person can run ./scripts/session-e2e.sh bep_store_reference_end_to_end. The case stores an API key with Notes from the agent: The failed tests lane came from shellcheck SC2016 on the box scan line. all_scripts_pass_shellcheck runs scripts/lint-shell.sh. A shellcheck directive fixes it, and lint-shell.sh now passes. ci-linux-native-success is the aggregator of that lane. The case is still NOT RUN off macOS. This box has no macOS lane, so the macOS evidence is still owed by ci-macos. I cannot file issues because I have no gh credential. A person must file two issues: (1) GrantSource in crates/boxes/src/core/bep.rs must accept source = "store" with store and id, per BEP-032. (2) The box_connection span in crates/bep/src/proxy.rs must carry the store and the identifier, as the task's observability line asks. Name both issues in the PR body as residuals. The script comments now describe both gaps, without issue numbers. |
There was a problem hiding this comment.
minimal-review — approval withheld
Verdict COMMENT — APPROVE withheld: 13 unresolved threads — minimal-arch[bot] scripts/session-e2e.sh, minimal-arch[bot] scripts/session-e2e.sh, minimal-arch[bot] scripts/session-e2e.sh:12188, minimal-arch[bot] scripts/session-e2e.sh:12188, minimal-arch[bot] scripts/session-e2e.sh:12369, …. No findings.
Panel: 3 of 3 personas ran (correctness, security, supply-chain).
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 3 conditions · the full ruling
The case checks what Gatehouse §6.10 asks for a Keychain store reference. Three points stay open. The seed writes the reference as a broker scope, and line 842 spells it source = "store". The comment calls it an interim but links no issue. C1 fixes that. The e2e leg reported NOT RUN, and NOT RUN is not a pass. C2 asks for the macOS-lane result. The span gap in the comment also needs a linked issue (C3).
Ruled at 1a6585a against gominimal/arch@109541a48c31 by claude-opus-5-5.
|
/revise Round 5 of #2612, on the arch conditions at
Generated by Claude Code |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The compare scopes the bep binary check to the no_sign_in case and the proxy stop to this run's listener socket. Signed-in BEP cases fail again when no proxy comes up. Both changes conform. The store case proves §6.10 lines 844 and 849 by hard failures, but its macOS leg still reports NOT RUN. C1 stays open until the lane in #2657 prints Keychain secret by reference OK.
Ruled at 06f522a against gominimal/arch@109541a48c31 by claude-opus-5-5.
A refused store reference no longer records store and identifier on the connection span, so the span does not read as an injection that did not happen. The span test now covers a refused reference.
|
Round 6 of 10 on the runtime's review findings, pushed as
Verify Outside the task's
Converted to a draft: something above needs a person before review. Review threads the coder judged:
On the macOS lane, a person stores an API key with Notes from the agent: crates/bep/src/forward.rs is outside the task's layers. Earlier rounds on this branch changed it already (commit 00f7b06), and this round fixes a finding in that change. The e2e test line ran and printed NOT RUN, because the case runs on macOS only. The macOS leg of bep_store_reference_end_to_end is still owed by the lane that #2657 stages. |
|
@cla-assistant check Generated by Claude Code |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The compare conforms. The box_connection span now names the store and the identifier only after an injection. The audit record still names the reference and the decision on admit and on refusal, as Gatehouse §6.10 line 849 requires. The new test reaches inject_store and proves the refusal span. One item stays open: the macOS end-to-end leg reports NOT RUN. C1 keeps it open until the run on #2657 prints Keychain secret by reference OK.
Ruled at c7e45a7 against gominimal/arch@109541a48c31 by claude-opus-5-5.
|
Round 7 of 10 on the runtime's review findings: the coder changed nothing. Revise round 7: the one open thread is judged invalid, so this round changes no file (noop). I did not run Review threads the coder judged:
|
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
This head merges main into the head that the last ruling covered, and the PR content does not change. The connection span names the store and the identifier only after an injection, as AT25 requires. The macOS leg of bep_store_reference_end_to_end is still NOT RUN. C1 keeps it open until the run on #2657 prints Keychain secret by reference OK. There is one new escalation. Step 1 of the case uses min secret set --upstream --header on the Keychain, and the architecture min secret reference refuses those flags for a host store. A person must decide which document changes.
Ruled at b4c5dc1 against gominimal/arch@42d214dc8ac8 by claude-opus-5-5.
|
Revise round 8 ran against |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The change conforms to AT25 and the min box audit field names. The span names the store and the identifier and never the value. The macOS leg of bep_store_reference_end_to_end is NOT RUN. C1 keeps the proof open until that lane prints Keychain secret by reference OK on #2657. One escalation is open: design line 517 and spec BEP-079 disagree about min secret set --upstream for a host store. That question is not this PR's to fix.
Ruled at 363ad7e against gominimal/arch@42d214dc8ac8 by claude-opus-5-5.
|
Revise round 8 ran against |
|
Owner ruling on the arch review's design question: spec 24 BEP-079 wins. The conflicting Arch condition C1 stays carried on #2657, as before. Generated by Claude Code |
There was a problem hiding this comment.
arch-review: PASS_WITH_CONDITIONS · 1 condition · the full ruling
The change conforms to AT25 and Gatehouse §6.10. The box_connection span names the store and the identifier only after an injection, and never the value. The e2e case for this proof reports NOT RUN on this head. A NOT RUN leg is open, so C1 stays a condition. Run the case on the macOS lane of #2657, and record the Keychain secret by reference OK line there. This head is a clean merge of main, and the ruling on 363ad7e7 stands.
Ruled at 6f61210 against gominimal/arch@42d214dc8ac8 by claude-opus-5-5.
norrietaylor
left a comment
There was a problem hiding this comment.
Approve at 6f612100.
- The e2e case
bep_store_reference_end_to_endchecks each AT25 mitigation that a box-side test can show. The value never enters the box, and the audit record names the store, the id and the authority, never the value. inject_storerecordsstoreandidentifieron thebox_connectionspan only after an injection that happened.- The script labels its three interims (broker-scope spelling, proxy restart, the
no_sign_ingate) and links #2656 and #2746. - arch-review rules PASS_WITH_CONDITIONS at this head. Its only condition, C1, is the macOS run of this case. By the owner's ruling, C1 is deferred to #2657.
- The base merge of
mainatbea4ea72is clean and touches none of this PR's files. Every lane is green, including the macOS e2e.
Generated by Claude Code
norrietaylor
left a comment
There was a problem hiding this comment.
Babysitter approval at 6f61210: every merge gate passes at this head (arch PASS at 6f61210, zero open threads, every lane aggregator green). Squash title test(e2e): prove Keychain secrets by reference end to end and the squashed message pass commitlint. Queued for squash merge.
Generated by Claude Code
…anges (gominimal#2767) * feat(bep): re-certify the published root on a control request bep serve takes recertify_root on its control socket. It re-issues the root for the current store rules, publishes it to the identity file, and replies only after both, with ok or the reason. The CONNECT path stays the fallback for a rule file edited by hand. * feat(minimal): ask a running proxy to re-certify after secret set min secret set sends recertify_root on the proxy's control socket after it writes a store rule. When the proxy did not re-issue or publish its root, it says the rule is saved and names the upstream that a box created before then does not reach. With no proxy running it prints nothing new. * test(e2e): activate the store-reference box against the running proxy min secret set now re-certifies the running proxy's root, so the case drops the bep stop and daemon restart that gominimal#2612 kept for gominimal#2746. It checks that the same proxy serves before and after the set. * test(minimal): classify the re-certify control connection min secret set now opens the egress proxy's control socket, so the connection inventory records that site as ungated. --------- Co-authored-by: minimal-foundry[bot] <327719912+minimal-foundry[bot]@users.noreply.github.com>
T29 — Keychain secrets by reference, end to end
slice: S3
kind: integration
covers: none
size: M
est: 80 code, 120 tests
executor: agent
difficulty: needs care
model: standard
diagnostics:
min box audit <box>shows each injection of the run by store and identifierobservability: one connection span per injected request, with the store and identifier, never the value
layers:
- e2e:
- scripts/session-e2e.sh: one case: set a secret, add a rule, reference it in a box spec, see the upstream receive it, and find it nowhere in the box
test: ./scripts/session-e2e.sh bep_store_reference_end_to_end
produces: none
consumes: none
proof: Store an API key with
min secret, reference it in a box spec, and the box's requests to its upstream carry it while the key never enters the box.depends on:
- blocked by #2280
- blocked by #2287
- blocked by #2288
- blocked by #2292
- blocked by #2325
Proof
On a macOS lane with the Box Egress Proxy,
./scripts/session-e2e.sh bep_store_reference_end_to_enddoes these steps. It stores a random value withmin secret set <id> --upstream httpbin.org --header X-Minimal-E2E-Probeand references<id>in a box spec. It shows thatmin box speclists the reference with its upstream. It shows that the upstream receives the stored value (the box compares only hashes), and it finds the value in no variable and no file in the box. It also shows thatmin box auditrecords the admitted injection by store and identifier, and that neither the audit nor bep.log holds the value. On this Linux box the case reports NOT RUN, so the end-to-end proof is not delivered here.Notes from the agent: The proof is not delivered on this box. The case needs macOS, the Keychain, a live bep and MINVMD_GVPROXY_BIN, so here it reported NOT RUN. The macOS lane must run it before anyone takes the proof as shown. Design choices: (1) The upstream is an HTTPS echo service, E2E_BEP_STORE_ECHO_URL, with https://httpbin.org/headers as the default. This is the only way to see the upstream receive the value without a real third-party key. The echo returns the value to the box, but only inside a shell pipe. The box prints only a hash and writes the echo to no file and to no exported variable. The case then searches the box files for the value's shape, not for the value, so the search does not bring the value into the box. (2) bep_live_gate takes an optional no_sign_in second argument, because a store reference needs no GitHub sign-in. Every other caller keeps its current behaviour. (3) The box declares packages base, coreutils and ca-certificates, and the case expects curl from the shell stack. The case fails with 'no curl in the box' if curl is not there. (4) Observability gap: the box_connection span in crates/bep/src/proxy.rs has no store or identifier field. Only a DEBUG line 'store value injection' in crates/bep/src/forward.rs names them. Th…[truncated]
Layer exception
The epic owner accepted these files outside T29's layers on 2026-10-27, for arch condition C3 (round 5):
crates/bep/src/proxy.rs: thebox_connectionspan recordsstoreandidentifier, never the value.crates/bep/src/forward.rs: passes the store and identifier of a redeemed store reference to that span.Tests
./scripts/session-e2e.sh bep_store_reference_end_to_endDeferred: the epic owner ruled on 2026-10-29 that #2657 carries the macOS leg of
bep_store_reference_end_to_end. #2657 stages bep in the macOS CI testbed. The lane run there must printKeychain secret by reference OK, not NOT RUN, and is recorded on #2657.Verify
min task run agent-verifyexited 101 under rustc 1.98.1, which the repository does not pin (1.99.0 in rust-toolchain.toml); the PR is not a draft on that alone — CI, which honours the pin, is the gate.Closes #2294
Generated by Claude Code