The suite's supabase-js clients followed redirects by default, so a 3xx from the sandbox gateway would have carried the apikey header to another origin. `noRedirectFetch` (`redirect: 'error'`) is passed as `global.fetch` of both client factories; a regression test with two loopback servers proves REST, auth and Storage calls stop at the 302. Live proof j and a second 5/5 run are in the audit log; no production client changes.
18 KiB
PLAN — bugfix: pnpm e2e:avatar on develop can no longer reach DEV
- date: 2026-11-02 | flow: bugfix | branch: bugfix/avatar-e2e-sandbox-only (off develop
6287e5b) - revision 3: three challengers (correctness, robustness, simplicity: CONCERNS each, no BLOCKER), then one confirmation challenger (robustness, CONCERNS(5)); every MAJOR folded in as a named change
- worktree: /tmp/claude-1000/-home-bchanot-Documents-faunosteo/573c6c77-7294-4d96-a518-83afaafef8d1/scratchpad/bugfix-avatar
- reference (safe runner already built): feature/v1-2-anatomical-sheets at
a15c35c, worktree /tmp/claude-1000/-home-bchanot-Documents-faunosteo/e641af9a-743e-46d0-9804-99667836040e/scratchpad/v1-2 (commitscdd43ac,2a84e37,3c141c9,f924ad0)
Bug
On develop, pnpm e2e:avatar = bash scripts/avatar-e2e.sh:
- reads API_URL, ANON_KEY, SERVICE_ROLE_KEY from
supabase status -o env(the DEV stack), lines 12-23; - starts the worker container on
supabase_network_faunosteoagainsthttp://supabase_kong_faunosteo:8000(DEV Kong) with a 5 s sweep and a 0 s quarantine grace; - runs
packages/supabase/tests/e2e/avatar-flow.e2e.tswith those keys: creates a confirmed account on DEV, signs in as the seededbob@faunosteo.localwith the DEV password written in the suite (LocalDev-only-password-1), uploads, andexpeditePurges()moves EVERY due date of DEV'sprivate.object_purge_queueto now;psqlgoes tosupabase_db_faunosteoby default (suite :29-31, :175, :215). Root cause: neither the runner nor the suite knows an authorised target; URL, keys and container come from the environment or from DEV defaults. Same family as the sheet-assets incidents of 2026-11-01 (fixed on develop by6287e5b).
Decisions already taken (constraints)
- Owner brief 2026-11-02 §1: minimal port of the V1-2 runner if its dependencies exist on develop, else an
explicit refusal; no DEV access (read or write), no implicit key recovery through
supabase status, no fallback to DEV URL / database / accounts / buckets; refusal before any connection when the configuration is absent or invalid; refusals proven without any IP connection; independent security review; merge into develop only (never V1-2). - Owner brief §3: a JWT issuer check is not a signature check; never present it as full cryptographic authentication nor as the only protection; strict destination restriction + sandbox isolation stay mandatory; check that redirects, inherited variables and ambiguous configurations cannot leave the sandbox.
- Scope answer (gated 2026-11-02): avatar only;
cmd_sheet_assets_e2eis NOT ported (pnpm e2e:sheet-assetskeeps refusing on develop until the V1-2 merge). - LRN-110: absolute verified cwd for every command, no
cd X && cmd &, quoted heredocs, env validated inside the script; agents never run the old scripts (develop'sscripts/avatar-e2e.sh, the main checkout's or old-958228a's copies). - Each ported block is copied from the V1-2 text (same bytes) where develop has no conflicting need; this does NOT make the later V1-2 → develop merge conflict-free (see "Merge expectations").
Dependencies checked on develop: scripts/db-sandbox.sh (network, db/auth/rest/storage/mail containers,
keys signed with issuer db-sandbox :80-81, load_secrets :50, wait_for :63, with_sandbox_env :229);
the worker's default run logs "event":"worker_start" (crates/worker/src/run.rs:121, main.rs:21-29:
run takes no argument on develop); the image is static-debian13:nonroot with USER nonroot
(crates/worker/Dockerfile:37-39); apps/app/e2e/support/sandbox.ts:111 creates its accounts with random
passwords; packages/supabase/src/assets.ts identical on both branches.
NOT ported (V1-2 only, with the reason): SB_SEED_PW / seed_passwords / SANDBOX_SEED_PASSWORD
(develop's packages/supabase/tests/sandbox/workspaces.sandbox.ts:197-199 signs in to
owner@faunosteo.local with the seed password: rotating it would break pnpm db:sandbox test),
throwaway_password, verify-e2e / cmd_verify_e2e (needs run --jobs sheet-verify, V1-2 worker only),
test args pass-through, cmd_sheet_assets_e2e (scope answer).
Fix plan (files in the bugfix worktree)
-
scripts/db-sandbox.sh— port from V1-2a15c35c, copied by function name, bytes identical:- usage header: insert after develop line 16 the V1-2
avatar-e2e [args]three lines only (V1-2 :20-22); the usage fallback becomessed -n '2,26p'(header grows by 3 lines); - variables: after
MAIL=…insert V1-2 :47-51 verbatim (GATEWAY=,WORKER=faunosteo_dbsandbox_worker_verify,AVATAR_WORKER=,GATEWAY_IMAGE=nginx:1.31-alpine,WORKER_IMAGE=faunosteo_worker:dev); the PORT line becomes V1-2 :52 (PORT_GATEWAY=57800appended).WORKER=is inert on develop; it letsdownremove a verify worker left by an interrupted V1-2 run (shared Docker names) and keeps the line identical; - the
REPO="$(realpath "$PWD")"check: V1-2 :58-62 (comment,REPO=,case, pattern,esac) replace develop :45-47 (a STATE inside the repository reached through a symlink no longer passes; the port adds$STATE/worker.XXXXXXfiles holding the sandbox service key); cmd_down: V1-2 :255 verbatim ("$AVATAR_WORKER" "$WORKER" "$GATEWAY" "$REST" …);- after
with_sandbox_env: V1-2 :270-332 (GATEWAY_CONF,start_gateway,start_worker,print_worker_log,worker_cleanup,worker_preflight) and :342-359 (cmd_avatar_e2e, the comment above it through the closing brace, the suite invocation :356-358 included); NOT :333-341 (cmd_verify_e2e) nor :361-376 (cmd_sheet_assets_e2e); - dispatcher:
avatar-e2e) shift; cmd_avatar_e2e "$@" ;;betweene2e-shellandsql. bash -n $W/scripts/db-sandbox.shandbash $W/scripts/db-sandbox.sh(usage, exit 1; absolute path, never a barepnpm db:sandbox) prove the file parses and the header range is right.
- usage header: insert after develop line 16 the V1-2
-
scripts/avatar-e2e.sh— deleted (git rm). -
package.json—"e2e:avatar": "bash scripts/db-sandbox.sh avatar-e2e". -
packages/supabase/tests/e2e/avatar-flow.e2e.ts— replaced by the V1-2 file (same bytes): guard at module load:SUPABASE_URL === 'http://127.0.0.1:57800',AVATAR_E2E_DB_CONTAINER === 'faunosteo_dbsandbox_db', both keys' payloadiss === 'db-sandbox', elsethrowbefore any client is created; clients built on the constant, never on the env URL;psqlon the constant container; a random password per run; two accounts created per run and deleted after;expeditePurges(userId)scoped to the account's objects;describe.skipIf(!configured)gone: a wrong environment is an error, not a skip. The guard checks the issuer, not the role and not the signature (brief §3): a role mix-up fails at the first admin call against the sandbox; a forgediss: db-sandboxkey passes the guard and fails at the sandbox's GoTrue. -
packages/supabase/vitest.e2e.config.ts,scripts/dev-env.sh— the comment lines (V1-2 text). -
Docs: only the lines this fix makes false (dead
scripts/avatar-e2e.shreferences, DEV-target instructions), worded so they are true on develop (removal date 2026-11-02, no "V1-2.3", noverify-e2e):CLAUDE.md: thedb:sandboxline (addavatar-e2eonly), thee2e:avatarline, the two folder-structure lines (tests/e2e, scripts list), the "Storage or worker change" workflow line;docs/development/STORAGE.md§6: the DEVdocker runkeeps its place with the V1-2 comment ("the worker beside the local stack, for the manual photo flow (IDENTITY.md); the e2e suite refuses this stack"), thepnpm --filter @faunosteo/supabase test:e2eline is DELETED (it would always throw), thee2e:avatarline becomes the sandbox alias;docs/development/SUPABASE.md§3.2 block and §6 DEV-acceptance note; §5.1 table: a develop-only row forpnpm db:sandbox avatar-e2ethat describes the gateway on its own (no reference to a verify-e2e row);docs/development/WORKSPACE.md,docs/development/TEST_STRATEGY.md,README.md:114,docs/development/ENVIRONMENTS.md:38,docs/development/IDENTITY.md:130(the manual photo flow points at the STORAGE §6docker runline, not at the e2e container: it sits on the sandbox network and is removed at exit).- NOT touched (still true after the fix):
docs/development/RUST.md:178,BROWSER_TESTS.md:103,SUPABASE.md:38,CHANGELOG.md(neither6287e5bnor the V1-2 commits added a line).
-
(owner, checkpoint 2026-11-02)
packages/supabase/tests/e2e/noRedirectFetch.ts:export const noRedirectFetch: typeof fetch = (input, init) => fetch(input, { ...init, redirect: 'error' }); the suite'sanonClient()andserviceClient()addglobal: { fetch: noRedirectFetch }(supabase-js handsglobal.fetchraw to auth-js and throughfetchWithAuthto REST and Storage; no layer setsredirect, our wrapper sets it last); regression testtests/no-redirect-fetch.test.ts(twonode:httpservers, 302 from A to B, auth + REST + Storage calls fail, B sees nothing,apikeyonly at A, control plain fetch reaches B); live proof j under strace. The suite now differs from V1-2 by those lines: V1-2 takes the same change at integration (merge rule amended).
Merge expectations (V1-2 → develop, later)
Develop's db-sandbox.sh is the merge-base text (9135d73); the port changes base regions V1-2 changes
differently. Expected conflicts: the usage header, the function block (V1-2 has cmd_verify_e2e and
cmd_sheet_assets_e2e around cmd_avatar_e2e), the dispatcher, the usage sed range. Rule: take V1-2's
file whole; check with an empty git diff <V1-2 tip> -- scripts/db-sandbox.sh after the merge. The suite,
package.json, dev-env.sh, the vitest comment and the ported variable / down / REPO lines merge
cleanly, except the suite: V1-2 adopts the noRedirectFetch lines (owner decision 2026-11-02). Doc lines reworded for develop conflict with V1-2's: keep V1-2's text (same rule as the
sheet-assets lines in TODO).
Proof (log committed as .claude/audits/2026-11-02-avatar-e2e-sandbox-only.md on this branch)
Who runs what: the executor applies the code and the static checks (contract criteria 1-5, 8); the LIVE proofs
below (a, c-i, the positive run) and the audit log are run by the orchestrator in the main loop, never by a
sub-agent (incident 1 of 2026-11-01 was an executor's backgrounded cd).
W=/tmp/claude-1000/-home-bchanot-Documents-faunosteo/573c6c77-7294-4d96-a518-83afaafef8d1/scratchpad/bugfix-avatar.
Every command names $W explicitly: bash "$W/scripts/db-sandbox.sh" …, corepack pnpm --dir "$W" …,
"$W/packages/supabase/node_modules/.bin/vitest" …; NEVER a bare pnpm e2e:avatar / pnpm db:sandbox /
pnpm worker:image (the shell's cwd resets to the main checkout, which still holds the DEV runner).
Pre-flight before any live command: test ! -e "$W/scripts/avatar-e2e.sh" and grep -F '"e2e:avatar": "bash scripts/db-sandbox.sh avatar-e2e"' "$W/package.json"; docker ps -a --filter name=^faunosteo_worker_e2e$
(a leftover of the old DEV runner; read-only; the owner removes it if present); the DEV containers' start
times recorded (docker inspect -f '{{.State.StartedAt}}', read-only) and compared at the end.
Keys used in c, d, e, h, i are FORGED unsigned JWTs built in the shell (base64url(header).base64url(payload).x),
never read from ~/.cache/faunosteo-db-sandbox/secrets.sh and never the DEV keys.
Trace: strace -f -e trace=connect,sendto,sendmsg -o <file> on bash "$W/scripts/db-sandbox.sh" … or on the
vitest binary directly, never through pnpm (its own process would be inside the trace). Criterion for a-e
and i: no AF_INET / AF_INET6 in the trace, except where stated.
a. bash "$W/scripts/db-sandbox.sh" avatar-e2e with FAUNOSTEO_SANDBOX_DIR pointing at an empty temp dir
(no secrets file) → db-sandbox: not running, exit 1, no connect, no container started;
b. the suite (vitest run --config vitest.e2e.config.ts in $W/packages/supabase) with no environment →
throws at load, no connect (contract oracle 5, positive control on curl 127.0.0.1:9);
c. the suite with SUPABASE_URL=http://127.0.0.1:54321, AVATAR_E2E_DB_CONTAINER=supabase_db_faunosteo and
two forged JWTs saying iss: supabase-demo → throws at load, no connect;
d. the suite with the sandbox URL and forged iss: db-sandbox keys but AVATAR_E2E_DB_CONTAINER=supabase_db_faunosteo
→ throws, no connect;
e. the suite with the sandbox URL and container and a key whose payload is not JSON → throws, no connect;
i. the suite with the sandbox URL and container, a forged anon key iss: db-sandbox and a forged service
key iss: supabase-demo, then the swap → throws, no connect. This is the only case that isolates the
issuer comparison itself (c stops at the URL, d at the container, e in the JSON.parse catch).
h. (sandbox NOT running, nothing listens on 57800) the suite with the sandbox URL and container and two
forged JWTs saying iss: db-sandbox → the guard passes, beforeAll fails with ECONNREFUSED; strace shows
connects to 127.0.0.1:57800 only. The brief §3 point made visible: the issuer check routes, it does not
authenticate; whoever holds 127.0.0.1:57800 would receive the keys the environment supplies.
f. (after the setup below, sandbox up from this branch, image present) bash "$W/scripts/db-sandbox.sh" avatar-e2e with 127.0.0.1:57800 held by a foreign listener (python3 -I -m http.server bound to
127.0.0.1:57800) → already in use, exit 1, docker ps shows no gateway and no avatar worker.
EXCEPTION, by construction: the probe (exec 3<>"/dev/tcp/127.0.0.1/$PORT_GATEWAY") is one
connect(AF_INET, 127.0.0.1:57800); the criterion for f is "the only AF_INET connect is 127.0.0.1:57800;
none to :54321 / :54322 or the LAN address". Flagged to the owner. Known limitation (not exercised): a port
held by another run's faunosteo_dbsandbox_gateway is torn down by worker_preflight before the probe
(shared container names and default STATE dir across worktrees); fails closed, sandbox only.
Setup (destroys the current sandbox and the V1-2 build of the worker image; the V1-2 worktree belongs to an
earlier session at its checkpoint, no run in progress; listed at the approval gate): corepack pnpm --dir "$W" worker:image (develop's crates), bash "$W/scripts/db-sandbox.sh" down, bash "$W/scripts/db-sandbox.sh" up fresh.
Positive run, ONCE, under strace, WITH hostile exports in the shell: SUPABASE_URL=http://127.0.0.1:54321,
SUPABASE_ANON_KEY and SUPABASE_SERVICE_ROLE_KEY forged (iss: supabase-demo),
AVATAR_E2E_DB_CONTAINER=supabase_db_faunosteo, SUPABASE_NETWORK=supabase_network_faunosteo,
SUPABASE_INTERNAL_URL=http://supabase_kong_faunosteo:8000, WORKER_IMAGE=faunosteo_worker:local (the old
runner's knobs; the script's constant wins) → bash "$W/scripts/db-sandbox.sh" avatar-e2e 5/5; trace
criterion: no connect to :54321 / :54322 or the LAN address; every other AF_INET connect named in the log
(expected: 127.0.0.1:57800 from the suite, the /dev/tcp probe). A watcher started beforehand in the
background with absolute paths (until docker inspect faunosteo_dbsandbox_worker_avatar; do sleep 1; done)
records, DURING the run and with filtered templates only: docker port faunosteo_dbsandbox_gateway 8000/tcp
(= 127.0.0.1:57800), the worker's networks ({{range $k,$v := .NetworkSettings.Networks}}{{println $k}}{{end}}
= faunosteo_dbsandbox_net) and {{range .Config.Env}}{{println .}}{{end}} | grep '^SUPABASE_URL='
(= http://faunosteo_dbsandbox_gateway:8000). NEVER a raw docker inspect or .Config.Env in the log: the
service key sits there. After the run: both containers gone, DEV start times unchanged.
What the isolation rests on (stated in the log with file:line, not "only one network": NET is a plain
bridge that can route to the host and the LAN where DEV's Kong publishes): the worker's URL is fixed to the
gateway container in the 600 env file, crates/worker/src/http.rs:31 max_redirects(0) and :33 proxy(None),
the gateway's upstreams are the sandbox containers by name, the sandbox keys are signed with a secret DEV
does not hold; scripts/dev-env.sh:23 exports FAUNOSTEO_DEV_API_URL into every child of db-sandbox.sh,
the suite included, and nothing reads it.
Residuals (LOW, for the owner, no code change here, fixed in V1-2 first then ported): the suite's supabase-js
clients use Node's global fetch (redirect: follow; HTTP(S)_PROXY honoured when NODE_USE_ENV_PROXY=1):
no endpoint the suite calls redirects today, a proxy would receive sandbox keys only; worker_env is read by
the EXIT trap before start_worker assigns it (an inherited value would be rm -f'd on an early die);
every docker call follows DOCKER_HOST / the current context.
Static: grep -rn "supabase status\|54321\|supabase_kong\|supabase_db_faunosteo\|supabase_network_faunosteo" "$W/scripts/db-sandbox.sh" "$W/packages/supabase/tests/e2e" → nothing;
git -C "$W" grep -n avatar-e2e.sh -- ':!.claude' → nothing (the historical contract of 2026-10-13 under
.claude/ keeps its citation).
Hand-off (in the log and to the owner): end with bash "$W/scripts/db-sandbox.sh" down; the V1-2 session
rebuilds faunosteo_worker:dev (pnpm worker:image there) and runs pnpm db:sandbox up fresh before its next
sandbox run (the develop secrets file has no SB_SEED_PW, V1-2's load_secrets refuses it: fails closed).
Copies that STILL reach DEV after the merge: the main checkout /home/bchanot/Documents/faunosteo (branch
feature/v1-1-clients-animals, merged, holds scripts/avatar-e2e.sh and the old e2e:avatar line: pnpm e2e:avatar there targets DEV until that checkout moves to develop and the branch is deleted with
gitflow.sh delete), and the worktree old-958228a (same two scripts). Agents never run them.
Gates
bash -n,pnpm typecheck,pnpm lint,pnpm --filter @faunosteo/supabase test:run(PGlite) green in the worktree.- fresh verifier on the contract; fresh security-auditor on the diff (MODE: gate), both before commit.
- merge into develop with
gitflow.sh finishonly after the owner's go (brief §1 authorises it after the verifications succeed); never V1-2.