Planner-audit #9. \`fetchSegment\` previously \`await response.json()\`
on any body BRouter returned. A misbehaving or compromised upstream
could OOM the planner process by returning gigabytes of JSON.
New \`readBodyWithCap()\`:
- Reject upfront when \`content-length\` declares over the cap.
- Stream the body and abort the reader once received bytes exceed
the cap (handles the case where upstream lies about content-length
or omits it).
Cap chosen at 10 MB — real per-segment GeoJSON is <100 KB; even the
longest realistic multi-day route stays well under 2 MB. Above 10 MB
is upstream bug or abuse; we'd rather error.
Applied to both \`fetchSegment\` (GeoJSON path) and \`computeSegmentGpx\`
(GPX path).
Tests: 3 cases (within cap, content-length over cap, streamed body
mid-cap abort).
Addresses planner-audit #8. `listSessions()` had no LIMIT — every call
to `GET /api/sessions` returned every non-closed session and did a
full table scan ordered by last_activity. On a long-running planner
host that's both a memory cliff and a query-plan footgun.
Now: defaults to 50 rows, clamps to [1, 200], accepts `?limit=` for
pagination-light scenarios. `sessions_last_activity_idx` (added in
#438) backs the ORDER BY + LIMIT efficiently.
New \`e2e/journal-planner-save.test.ts\` exercises the full save flow:
1. Seed a routeId + JWT via \`/api/e2e/seed\` (journal).
2. POST to \`/api/sessions\` on the planner with \`callbackUrl\` +
\`callbackToken\` — mirrors what the journal's \`edit-in-planner\`
action does server-to-server.
3. Open the planner session in a real browser.
4. Click \"Save to Journal\". The button POSTs sessionId+GPX to the
planner's \`/api/save-to-journal\` action (Phase A); the action
forwards to the journal callback with the Bearer.
Test 1 asserts:
- The journal's route ends up with geometry (round-trip works).
- The exact JWT string is **never** present in any browser-issued
request body or \`Authorization\` header — i.e. Phase A correctly
keeps the token server-side.
Test 2 asserts:
- A second click on Save reuses the same stored JWT, the journal's
jti consumer rejects it, and the planner UI surfaces the error.
This is the Phase B replay guard exercised end-to-end through the
UI rather than just the API.
Added \`journal-planner-save\` Playwright project (no baseURL — the
test navigates both apps using absolute URLs).
Full repo: pnpm typecheck / lint / test all green (cached).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
After Phase A (#442) moved the journal callback token off the browser,
the token was still replayable on the wire until \`exp\` (7 days). This
PR makes each token strictly single-use.
Changes:
- **\`journal.consumed_jwt_jti\` table** — \`jti TEXT PRIMARY KEY,
consumed_at TIMESTAMPTZ, expires_at TIMESTAMPTZ\`. Picked up by
drizzle-kit push on deploy.
- **\`createRouteToken\` now sets a \`jti\` claim** (\`randomUUID()\`).
- **\`verifyRouteToken\` atomically consumes the jti** via
\`INSERT … ON CONFLICT DO NOTHING RETURNING jti\`. Postgres
serializes the insert, so exactly one concurrent caller wins; the
rest see an empty result and throw \`TokenAlreadyConsumedError\`.
Tokens without a \`jti\` claim (i.e. minted before this PR) are
also rejected — the right call: any in-flight legacy token sitting
in a planner session is replayable, and we'd rather fail-loud than
silently grandfather them in.
- **\`consumed-jti-sweep\` job** — daily 03:45 UTC cron that
\`DELETE WHERE expires_at < now()\`. Keeps the table tiny; offset
from the other purge jobs to spread load.
- **e2e replay test** — \`integration.test.ts\` now exercises a
same-token double-submit and asserts the second returns 401 with
\`/consumed|already/i\`.
UX implication worth flagging: a user who clicks \"Save\" twice (or whose
network retries a failed POST) sees an error on the second attempt.
They go back to the journal for a fresh \"Edit in Planner\" link.
Full repo: pnpm typecheck / lint / test all green (177 + 31 integration).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Save-to-Journal flow had the browser fetch the journal with a
\`Bearer \${callbackToken}\` header. The JWT was visible in DevTools,
exfiltratable via any XSS or browser extension, and the planner's
\`loader\` shipped it down to the client as part of the page payload.
Now:
- **New action**: \`POST /api/save-to-journal\` (\`routes/api.save-to-journal.ts\`).
Body: \`{ sessionId, gpx }\`. The action loads \`callbackUrl\` +
\`callbackToken\` from \`planner.sessions\` (set at /new time when the
user came from the journal), POSTs to the journal server-to-server
with the Bearer, and forwards the response.
- **\`SaveToJournalButton\`**: drops the \`callbackUrl\` + \`callbackToken\`
props. Takes \`sessionId\` only and POSTs to the planner action.
- **\`session.\$id.tsx\` loader**: stops returning \`callbackUrl\` /
\`callbackToken\` to the client. Returns a single \`hasJournalCallback\`
boolean so the button still knows whether to render.
- **\`SessionView\`**: same prop simplification.
Trust model is unchanged: the same \`sessionId\` that grants Yjs
membership grants save authority. Knowing the URL = ability to act.
The action only adds a server-side hop so the JWT never reaches
browser JS.
Phase B (jti single-use enforcement on the journal side) follows in
a separate PR — needs a journal DB column + verifier change.
Full repo: pnpm typecheck / lint / test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses planner-audit #1 (CRITICAL — WebSocket joins unauthenticated)
with the narrow fix. The upgrade handler used to immediately
\`handleUpgrade\` for any \`/sync/<id>\` path, then \`getOrLoadDoc\` *created*
a Y.Doc on demand. An attacker connecting to random sessionIds could:
- exhaust process memory by spinning up Y.Doc objects for sessions
that don't exist (no DB row backed them), and
- resurrect closed/expired sessions by reconnecting (closed rows
were still cacheable to \`docs\`).
Now: the upgrade handler calls \`getSession(sessionId)\` first
(\`SELECT \\* FROM planner.sessions WHERE id = ? AND closed = false\`).
Missing or closed → \`socket.destroy()\` before handshake. DB unreachable
also closes — fail closed; the client reconnects after backoff.
Costs one DB roundtrip per upgrade. Upgrades are rare vs message
volume, so the overhead is negligible.
Note: we are NOT adding a join token. The planner is anonymous-by-design
(see CLAUDE.md \"sessions are anonymous and ephemeral\") — knowing the
URL still equals membership. The check here only guards against
attacker-supplied sessionIds with no backing row.
Full repo: pnpm typecheck / lint / test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses planner audit #5 — a connected client could feed unlimited
waypoints / no-go / notes into the Yjs doc, growing the in-memory map
and the persisted \`sessions.yjs_state\` blob until OOM or DB blowout.
Two guards:
1. **Per-frame**: \`MAX_MESSAGE_BYTES = 256 KB\`. Anything bigger arrives
from a buggy or hostile client — a single waypoint add is hundreds
of bytes. Oversized frame → \`ws.close(1008)\` and drop the message.
2. **Per-doc**: \`MAX_DOC_BYTES = 5 MB\`. After a sync-message apply
succeeds, recompute \`Y.encodeStateAsUpdate(doc).byteLength\`. If it
crossed the cap, mark the session \"quarantined\": close every
connected client (\`1008 policy violation\`), and short-circuit
subsequent handleMessage calls so no further bytes can be applied
and the debounced save can't write an oversized blob to Postgres.
The 5 MB limit is generous — a typical multi-day route's serialized
state is well under 100 KB. The cap exists to make abuse expensive,
not to constrain real use.
Exports \`MAX_MESSAGE_BYTES\`, \`MAX_DOC_BYTES\`, and \`docByteSize\`
for testing. Tests cover:
- constants are positive and ordered
- docByteSize is 0 for unknown sessions
- a realistic 500-waypoint route stays well under the cap
- ~6MB of garbage in a Y.Text trips the cap (smoke test for the guard)
Full repo: pnpm typecheck / lint / test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses planner audit #3 (SSRF via callbackUrl) and #7 (URL-param
size). Two attack surfaces hardened:
1. /new loader — \`callback\`, \`token\`, \`returnUrl\`, \`gpx\` query
params now validated:
- callbackUrl: must be a valid absolute http(s) URL ≤ 2048 chars.
If \`PLANNER_CALLBACK_ALLOWED_HOSTS\` is set (comma-separated),
the host must match — defense-in-depth SSRF guard for self-
hosted instances. Unset = no allowlist (dev / open self-host).
- token: max 2048 chars.
- returnUrl: must be a same-origin path or absolute http(s) URL
≤ 2048 chars. Rejects \`javascript:\`, \`data:\`, and
protocol-relative \`//host\` (which would resolve to a remote
origin on HTTPS pages).
- gpx: ≤ 2 MB encoded.
Invalid input throws 400 from the loader.
2. /session/:id default-export component — \`waypoints\`, \`noGoAreas\`,
\`notes\`, \`returnUrl\` URL params now bounded before
\`JSON.parse\` / use:
- waypoints / noGoAreas: ≤ 50KB each; over-cap returns undefined
(component starts with empty initial state, same as malformed).
- notes: ≤ 10KB.
- returnUrl: ≤ 2KB + same scheme rules as #1.
Pulled the URL validation into \`lib/url-validation.server.ts\` so
both routes (and any future caller) share the same rules.
Tests: \`url-validation.server.test.ts\` (14 cases — schemes,
allowlist, length caps, protocol-relative guards, env parsing).
Full repo: pnpm typecheck / lint / test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The hourly `expireSessions()` cron runs `DELETE FROM planner.sessions
WHERE last_activity < cutoff`. Without an index on `last_activity`,
that's a full table scan growing linearly with the total sessions
ever created (planner sessions are never user-deleted; expiry is the
only churn path).
`drizzle-kit push` picks this up on next deploy.
Mirrors journal PR #5. `fetchSegment()` and `computeSegmentGpx()` in
lib/brouter.ts called `fetch()` with no AbortSignal — a hung or slow
BRouter would stall the request handler indefinitely.
New `lib/http.server.ts::fetchWithTimeout()` (same as the journal's
helper) wraps fetch with `AbortSignal.timeout(30_000)`, composable
with a caller-supplied signal via `AbortSignal.any()`.
Tests: lib/http.server.test.ts (2 cases — timeout abort, happy path).
The three planner-coloring tests wait for the elevation chart's
\`<canvas>\` after seeing the Yjs \"Connected\" text. \`ElevationChart\`
returns \`null\` until \`points.length >= 2\` — i.e. until the BRouter
response (even mocked) has flowed through the Yjs round-trip and
updated \`yjs.routeData\`. The chart module is also lazy-imported
(\`Suspense\`) — cold CI runs need to fetch the chunk before mounting.
Saw the failure on PR #424's CI run: canvas \"element(s) not found\"
after the 10s window, both initial run and retry. Other planner
suites that wait on \`.leaflet-container\` (which renders
unconditionally and isn't behind a lazy boundary) don't see this.
Bumped \`CANVAS_TIMEOUT\` to 20s. Not masking a real regression — the
lazy chunk + Yjs apply is a real serial cost that the test budget
didn't reflect.
The five \`*.integration.test.ts\` files in apps/journal (explore, follow,
demo-bot, notifications, notifications-fanout — 31 tests total) were
gated behind \`EXPLORE_INTEGRATION=1\` / \`FOLLOW_INTEGRATION=1\` /
\`DEMO_BOT_INTEGRATION=1\` / \`NOTIFICATIONS_INTEGRATION=1\`. The unit-test
job doesn't have Postgres, so they correctly skipped there — but no
CI job set the env vars, so they were effectively dead code.
Added a step in the E2E job (which already has Postgres + schema
pushed) that flips all four gates and runs the integration files
serially (\`--no-file-parallelism\` — they share the schema and trip FK
constraints if run in parallel; ~2.5s sequential anyway).
Wiring them up surfaced two real issues, both fixed here:
1. **\`createRoute\` silently dropped \`input.visibility\`** — every
caller passing \`visibility: \"public\"\` (including the demo-bot)
was getting the column's \`private\` default. Spread now mirrors
\`createActivity\`'s pattern. Demo-bot routes have actually been
private in production all this time — they were rendering on the
home feed only via the \`activity_published\` fan-out from the
*activity*, not as visible *routes*.
2. **The demo-bot test's fetch stub was incomplete** — it stubbed
\`.text()\` but the planner-session preflight calls \`.json()\`. The
stub now branches on URL: \`/api/sessions\` returns
\`{ sessionId: 'test-session' }\` JSON, the BRouter call returns the
stub GPX text.
Full repo: pnpm typecheck, pnpm lint, pnpm test all green
(177 unit-test pass + 31 integration pass = 208 total, no skips).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Self-hosted instances no longer inherit the trails.cool flagship Sentry
DSNs. Code paths now read DSN strictly from env — unset = no Sentry
init, no events sent.
Flagship continues to report unchanged: cd-apps.yml and cd-staging.yml
inject the public DSNs as workflow env vars, which feed into:
- runtime via \`infrastructure/app.env\` / \`staging.env\`
(\`SENTRY_DSN_JOURNAL\` / \`SENTRY_DSN_PLANNER\` → compose env per service)
- build time via docker build-arg \`VITE_SENTRY_DSN\` (journal only;
planner has no client Sentry init).
Sentry DSNs are public-by-design (transmitted unencrypted from the
client JS bundle), so embedding them as plaintext workflow env vars
is no worse than the runtime exposure. Forks should replace these with
their own DSNs or remove the workflow env lines to ship Sentry-free.
Files changed:
- apps/journal/server.ts: \`process.env.SENTRY_DSN\` only, no fallback
- apps/planner/server.ts: same
- apps/journal/app/lib/sentry.client.ts: \`import.meta.env.VITE_SENTRY_DSN\`
only; falsy = skip init
- apps/journal/Dockerfile: new \`ARG VITE_SENTRY_DSN\` baked into build
- infrastructure/docker-compose.yml: pass \`SENTRY_DSN\` per service
- infrastructure/docker-compose.staging.yml: same
- .github/workflows/cd-apps.yml: workflow env + build-arg + app.env echo
- .github/workflows/cd-staging.yml: same
Full repo: pnpm typecheck, pnpm lint, pnpm test, journal build all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
infrastructure/docker-compose.yml had \`\${VAR:-default}\` fallbacks for
JWT_SECRET, SESSION_SECRET, and POSTGRES_PASSWORD that silently
substituted known-weak values (\`change-me-in-production\`, \`trails\`)
when the env wasn't set. A misconfigured prod deploy would happily come
up with these defaults — JWT/session forgery + a guessable Postgres
password.
Switched the four critical substitutions to \`\${VAR:?message}\` so
\`docker compose up\` fails loud instead. Mirrors the existing pattern
already used for BROUTER_URL/BROUTER_AUTH_TOKEN and (in staging) for
JWT_SECRET/SESSION_SECRET.
Sites fixed:
- journal: DATABASE_URL (POSTGRES_PASSWORD), JWT_SECRET, SESSION_SECRET
- planner: DATABASE_URL (POSTGRES_PASSWORD)
- postgres: POSTGRES_PASSWORD (container itself)
- exporter: DATA_SOURCE_NAME (POSTGRES_PASSWORD)
- staging: both DATABASE_URLs
Production already provides all three via SOPS-encrypted
secrets.app.env, so this is a defense-in-depth change with no behavior
change on a properly-configured deploy.
Updated infrastructure/.env.example to make SESSION_SECRET explicit
(was missing) and call out the trio as required.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mirrors #422 (fail-loud DB URL) and #430 (health-check pool reuse) for
the planner app, which still had:
- two \`process.env.DATABASE_URL ?? \"postgres://trails:trails@localhost:5432/trails\"\`
call sites that would silently boot a misconfigured prod against
localhost
- a /health handler that opened a fresh postgres client + connection on
every probe (no pool, per-call TCP/TLS handshake)
Both now route through @trails-cool/db's \`getDatabaseUrl()\` (refuses to
start in production if unset / matches the dev default; E2E=true is the
opt-out) and a module-level singleton client (max: 2, idle_timeout: 30).
Full repo: pnpm typecheck, pnpm lint, pnpm test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mirrors PR #429 for the planner app. Each HTTP request now gets a
requestId (inbound X-Request-Id honored, otherwise a fresh UUID),
echoed on the response, and propagated to every downstream log call via
AsyncLocalStorage + pino's \`mixin\`.
Tests: planner logger.server.test.ts (2 cases, same shape as the
journal version).
Full repo: pnpm typecheck, pnpm lint, pnpm test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the i18next peer dep override workaround with a proper single
TypeScript version in the workspace catalog. All 15 packages typecheck
cleanly with TypeScript 6.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Mobile's TypeScript 6 devDep created a second peer-resolution variant of
i18next (i18next(typescript@6)) alongside the web apps' (i18next(typescript@5)).
Under pnpm's hoisted nodeLinker the server and client ended up with different
instances of the i18n singleton, so SSR rendered raw keys and hydration
mismatched.
Pin i18next and react-i18next TypeScript peer deps to 5.9.3 in root overrides
so the entire workspace resolves a single i18next instance regardless of which
app's TypeScript version is active.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The previous handler opened a fresh postgres client (max: 1) on every
call to /api/health and tore it down in the finally block. Under the
prod monitoring cadence (probes every few seconds), that's a fresh
TCP + TLS + auth handshake on every probe, plus connection-table
churn on the Postgres side — fine for the trickle of curl-ish manual
checks, slow-bleed under blackbox monitoring.
Now we cache a module-level singleton postgres client dedicated to
/api/health (max: 2, idle_timeout: 30) and reuse it across calls.
Separate from the app's main DB pool (via @trails-cool/db's createDb)
on purpose — so a starvation event on the main pool doesn't fail the
liveness check and trigger a restart loop.
Full repo: pnpm typecheck, pnpm lint, pnpm test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Every HTTP request now gets a requestId (inbound X-Request-Id header is
honored, otherwise a fresh UUID is minted) and the value is echoed on
the response. The server wraps the request in
\`requestContext.run({ requestId }, ...)\` — an AsyncLocalStorage scope —
so pino's \`mixin\` callback can read it on every log call without the
caller threading it through.
Net effect: \`logger.info({ ... }, \"db error\")\` from a loader, action,
or downstream lib now lands in JSON with a \`requestId\` field, making
cross-handler debugging trivial (\`grep requestId=abc-123\` returns the
full request trace).
Out of scope here:
- Planner gets the same treatment (separate, smaller PR after this lands).
- BRouter / Fedify outbound calls don't propagate the requestId yet —
those are HTTP boundaries where we'd add it as a header, but the
audit value was the in-process trace.
Tests:
- logger.server.test.ts (2 cases — als-bound info tags requestId; no
context = no tag).
Full repo: pnpm typecheck, pnpm lint, pnpm test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Sentry DSNs were hardcoded in journal/server.ts, planner/server.ts,
and journal/app/lib/sentry.client.ts. Self-hosted instances inheriting
the trails.cool flagship DSN would silently ship their errors to our
Sentry account.
Now each init site reads its DSN from env:
- journal/server.ts: SENTRY_DSN (server runtime env)
- planner/server.ts: SENTRY_DSN (server runtime env)
- journal/app/lib/sentry.client.ts: VITE_SENTRY_DSN (build-time bake)
The flagship DSN is kept as the fallback so the production deploy keeps
reporting without an infra change — but self-hosters can:
- set SENTRY_DSN=\"\" / VITE_SENTRY_DSN=\"\" to ship their own builds
without Sentry, or
- set SENTRY_DISABLED=true to skip init entirely at runtime, or
- set SENTRY_DSN=/VITE_SENTRY_DSN= to their own DSN.
Follow-up: a future PR can remove the hardcoded fallbacks once
infrastructure/docker-compose.yml + the cd-apps.yml workflow are wired
to pass SENTRY_DSN explicitly. That requires SOPS edits + workflow
changes I want isolated from this purely-code change.
Full repo: pnpm typecheck, pnpm lint, pnpm test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
verifyLoginCode, verifyMagicToken, and verifyEmailChange previously did
SELECT WHERE used_at IS NULL → UPDATE … SET used_at = now. Two concurrent
verifications could both pass the SELECT and both succeed, accepting the
same single-use token twice.
Collapsed each to a single UPDATE … WHERE … RETURNING * statement.
Postgres serializes row-level locks within an UPDATE, so exactly one
concurrent caller observes a returned row; the rest see an empty array
and get \"Invalid or expired\". The token is also marked used as part of
the same statement — no second write needed.
verifyEmailChange's tertiary email-availability check now runs *after*
the consume; we keep the original semantics where the token is burned
on a clash (the previous code explicitly did the same with a separate
UPDATE).
No behavior change on the happy path. Closes a credential-reuse
window that mattered most for the 6-digit login codes (small search
space, more likely to race).
Full repo: pnpm typecheck, pnpm lint, pnpm test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous \`importOne\` only fetched page 1 of /v1/workouts and errored
with \"not found on page 1\" if the workout wasn't there — silently
breaking import for any workout older than roughly the most recent 30
entries (per_page default). Webhook-driven imports happen to land on
page 1 by definition, so this only bit on user-initiated catch-up
imports of older workouts.
Now we paginate forward, using the \`total / per_page\` returned by page 1
to compute a stop condition, with a \`MAX_PAGES=100\` ceiling so a
misbehaving API can't loop us. We also stop early on an empty page.
Tests:
- new pagination case (workout on page 2, expect 2 fetch calls)
- new \"not found on any page\" case
Full repo: pnpm typecheck, pnpm lint, pnpm test all green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Rollup was warning on 5 modules that were both dynamically and statically
imported. With static importers in the same chunk, the dynamic forms
buy no chunking benefit — they were leftovers from earlier
cycle-avoidance workarounds that no longer apply.
Converted to static:
- @trails-cool/gpx (routes.\$id.server.ts, sync.import.\$provider.server.ts)
- logger.server (boss.server.ts — comment claimed test cycles, but tests pass)
- boss.server (activities.server.ts at two sites)
- connected-services/manager (komoot/importer.ts, wahoo/importer.ts)
- notifications.server (root.tsx)
Kept dynamic: fit-file-parser in sync.import.\$provider.server.ts — it's
heavy and only the FIT ingestion path needs it, no other static
importers exist, so the dynamic actually does chunk-split it.
Build is now warning-free.
Full repo: pnpm typecheck, pnpm lint, pnpm test, pnpm build all green.
192 tests passed (up from 181 — rate-limit test from #424 + others).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
