## What & why
Production reports of policy enforcement "hanging" traced to
large/many-page documents: the watermark step's flatten-to-image
(`convertPDFToImage`) on a 500+ page PDF takes minutes, exceeding both
the client poll cap and the backend per-step timeout. This makes the
slow case graceful instead of looking broken, and makes load-shedding
non-fatal.
### Poll runs to completion (no false "hang")
The client poll loop used a flat ~150s cap that was **shorter than the
backend's 300s per-step timeout**, so it abandoned long-but-healthy runs
mid-flight. The budget is now sized to the backend's real worst case —
`stepCount × per-step timeout + grace`, learned from the first status
report — so the client always polls long enough to surface the run's
**actual** terminal state (success or the backend's real error) rather
than a misleading client-side timeout.
### Per-step progress
The activity feed now shows `Enforcing… · step n/m` (from
`currentStep`/`stepCount`), so a slow step shows movement instead of a
dead spinner.
### Soft-retry on queue rejection
Under load the shared `JobQueue` rejects runs ("queue full"), which
previously surfaced as a hard failure needing a manual Retry. The
backend now tags that rejection with a stable `POLICY_QUEUE_FULL`
errorCode; the client treats it as transient backpressure and
**auto-retries the file in place** with exponential backoff (≈4s→64s, ~2
min), shown as a soft "Busy — retrying…" row, falling back to the manual
Retry only once the retry budget is spent.
## Testing
- **Frontend unit tests** (30 pass across the policies suite), including
a new `usePolicyAutoRun.retry.test.tsx` that drives the real controller
orchestration (poll → `POLICY_QUEUE_FULL` → relabel → backoff → in-place
re-dispatch), plus poll-budget, step-progress, and activity-feed relabel
cases.
- **Backend** `PolicyEngineTest` case asserting a queue-rejected run
carries the `POLICY_QUEUE_FULL` code.
- Typecheck clean on all three flavors (proprietary/saas/core); prettier
+ spotless clean.
- Poll-budget + progress + real-error surfacing were also verified live
end-to-end against a 599-page run (survived past the old cap, showed
step progress, reported the backend's real 300s-timeout failure,
recovered after a simulated network drop).
## Not included (follow-ups)
- The underlying flatten-to-image cost itself (bounded-memory/streaming
flatten, revisiting `convertPDFToImage` default and the 300s timeout) —
the real perf fix, deliberately out of scope here.
# Description of Changes
> [!warning]
> **Do not** squash this on merge. It should be merged via a merge
commit
Fixes conflicts in `pgvector_store.py`.
Also since codespell is failing, add comments to ignore the errors in
`sync_en_us_spelling.py`
---------
Co-authored-by: Ludy <[email protected]>
Co-authored-by: Anthony Stirling <[email protected]>
`PolicyController` was annotated `@PremiumEndpoint` (requires a
Pro-or-higher server license). Policies don't need a server-license
gate:
- On SaaS the server runs in Pro mode, so the check is always satisfied
anyway — it gates nothing in practice.
- Access is already governed by team scoping (#6632) plus the per-user /
guest gates.
So the annotation is dead weight and misleading. This removes it (and
its import) — a 2-line change.
## Verification
- `:proprietary:compileJava` succeeds; spotless clean. No other premium
gate on policy classes.
## Problem
We're getting 402s when an AI **agent** (chat) run hits the free
allowance / spending cap, but the frontend handles them poorly and never
pops the usage-limit modal.
The agent runs its tool calls **server-side** (loopback HTTP via
`PolicyExecutor`), so the 402 never reaches the `apiClient` interceptor
that pops the modal for direct calls. It was caught by the generic
tool-failure handler and flattened into a `CANNOT_CONTINUE` reason
string (`"The /api/v1/… tool failed: 402…"`), streamed as a `result`
event, and rendered as a scary chat bubble. This is the same gap the
policy auto-run path bridges (#6626) — one layer up.
## Fix
**Backend** (`proprietary`)
- `AiWorkflowResponse` gains `errorCode` + `errorSubscribed`.
- `AiWorkflowService` detects a downstream 401/402 entitlement sentinel
in its three tool-exec catch sites (`onToolCall`, `runPlan`,
`onConvertMarkdown`) and surfaces the structured code (+ `subscribed`)
on the terminal response instead of the raw failure text.
- Factored the 401/402 body extraction `PolicyEngine` already had into a
shared `DownstreamEntitlementError` util so the two server-side paths
can't drift.
**Frontend**
- New `usageLimitBridge` (`PAYG_LIMIT_REACHED_EVENT` +
`dispatchPaygLimitReached`) generalises the previously policy-only
bridge. Proprietary can't import the saas modal API (layering), so
server-side limit hits broadcast a window event the saas
`UsageLimitModalHost` opens the modal from. Migrated the policy path
onto it.
- `ChatContext` fires the matching modal (free → subscribe, subscribed →
raise cap) on the limit result **and** on a direct 402, replacing the
raw reason with a brief friendly line
(`chat.responses.usage_limit_reached`).
No Python engine changes — the charge/402 happens on the Java tool
endpoint that Java itself calls.
## Test plan
- [x] `:proprietary:compileJava` + `spotlessCheck` clean
- [x] `AiWorkflowServiceTest` + `PolicyEngineTest` green
- [x] eslint, proprietary + saas typechecks clean
- [ ] Manual: drive an agent run over the limit → brief line in chat +
the right modal (free vs cap)
> Note: proprietary test compilation is currently blocked on the
pre-existing `InitialSecuritySetupTest` 6-arg ctor break (unrelated,
tracked separately); verified locally by temporarily patching it.
## Problem
A policy ran over a file but the owner's **free usage was never
consumed**.
A policy run executes on a **background virtual thread**
(`PolicyEngine.submit` → `asyncExecutor`), and Spring's
`SecurityContextHolder` is thread-local — so the worker thread has no
identity. When `PolicyExecutor` → `InternalApiClient.post` resolves the
tool-call API key via `UserService.getCurrentUsername()`, it finds
nothing and falls back to the **`INTERNAL_API_USER`** key. The loopback
tool calls then authenticate as that system account, so
`PaygChargeInterceptor` attributes the charge to *its* team (or none) —
the real owner's free grant is untouched. Folder-watch / scheduled
triggers are even further removed (fired from a background watch loop
with no request context at all).
The charging *mechanism* was fine (AUTOMATION, multipart,
`openProcess`); only the **attribution** was wrong.
## Fix
Propagate the acting identity onto the worker thread using the
**audit-principal MDC key** that `UserService.getCurrentUsername()`
already reads as its documented async fallback (the same mechanism used
for other async jobs). No new plumbing through the executor.
- **`runPolicy`** (stored policies — covers triggers *and* manual
`runWith`) → bill the **policy owner**. `Policy.owner` is the username
stamped at creation, so `getApiKeyForUser(owner)` resolves it.
- **`submit`** (ad-hoc Automate/AI one-offs) → bill the **submitting
user**, captured on the request thread (it doesn't survive the hop to
the worker otherwise).
With the principal set, `InternalApiClient` dispatches each tool call as
that user → the interceptor resolves the right team → free grant draws /
Stripe meters correctly.
## Tests
`PolicyEngineTest`:
- `runPolicyDispatchesToolCallsAsTheOwner` — asserts MDC
`auditPrincipal` == the policy owner at the moment
`InternalApiClient.post` is invoked.
- `adHocRunDispatchesToolCallsAsTheSubmittingUser` — asserts it's the
submitting user for an ad-hoc run.
`:proprietary:test` + `:saas:test` + spotless green; coverage gates met.
## Heads-up (not in this PR)
Once attributed, **automatic folder-watch / scheduled runs consume free
grant (or bill) per file** — set up once, runs forever. That's
automation-is-billable working as intended, but a set-and-forget policy
can drain an allowance fast, so it may warrant a per-policy cap or a
heads-up in the UI. Flagging for a product decision.
## Summary
Pay-as-you-go (PAYG) billing for Stirling-PDF SaaS. Manual PDF editing
stays free forever; only **automation, AI, and API** usage is metered.
Every team gets a **one-time lifetime free grant** (default 500 PDFs)
before any billing; past that, a team adds a card and pays per metered
document, with a self-set monthly spending cap.
This branch combines and supersedes the in-flight BE (#6574) and FE
(#6579) work plus the SaaS edge functions (Stirling-PDF-SaaS PR, now on
`v3`), hardened into a single reviewable feature after a pre-merge
dead-code/security review.
## Billing model
- **Always free:** manual / JWT web-tool usage is `BYPASSED` — never
metered, no matter where it's triggered.
- **Billable categories:** `AUTOMATION`, `AI`, `API`.
- **One-time lifetime free grant** (`pricing_policy.free_tier_units`,
default 500): never resets, survives subscribing. It gates unsubscribed
teams (billable API calls hard-stop with a 402 once exhausted) and
decides the free-vs-paid split of every job.
- **Subscribed:** paid documents (beyond the grant) are metered to a
Stripe Billing Meter; an optional monthly spending cap degrades billable
categories when reached.
- **Dedup:** the same file pushed through several steps within a
workflow window counts **once** (lineage join), so API/AI chaining on
one file isn't double-charged.
## What's included
**Database** — Flyway migrations `V11`→`V21` with matching Supabase
twins: pricing policy + per-team sidecar (`payg_team_extensions`:
subscription id, Stripe customer, free-grant counter), append-only
`wallet_ledger`, shadow charges, subscription-state RPCs (`V14`), audit
logs (`V15`), billing category (`V16`), one-time lifetime free grant
(`V19`), launch-grant seed (`V20`), drop of the unused
`wallet_category_summary` view (`V21`).
**Charge pipeline** — `PaygChargeInterceptor` (open/join a process,
split the free grant, write the ledger DEBIT), `JobChargeService`
(consume the grant under a row lock, restore it on a first-step refund,
meter only the paid portion on completion), `StaleJobCloser` fallback
(idempotent close → meter).
**Entitlement** — `EntitlementService` (per-team cached snapshot:
grant-gated for free teams, monthly-cap-gated for subscribed) +
`EntitlementGuard` (401 `SIGNUP_REQUIRED` / 402 `FEATURE_DEGRADED` /
`PAYG_LIMIT_REACHED`).
**Metering** — `PaygMeterReportingService` writes a durable
`payg_meter_event_log` row around every POST to the `meter-payg-units`
edge fn (pending → posted/failed); `PaygMeterReconcileScheduler` retries
unposted events under the same idempotency key inside Stripe's 24h dedup
window.
**Billing facts** — `TeamBillingService` reads the synced `stripe.*`
mirror (subscription window, per-document rate; the unsubscribed-team
estimate resolves the rate by Price `lookup_key = plan:processor`).
**Wallet API** — `PaygWalletController`: `GET /api/v1/payg/wallet`,
`PATCH /api/v1/payg/cap`.
**Frontend** — PAYG Plan page (two-card free layout + subscribed views),
`useWallet`, upgrade modal with lazy-loaded Stripe Embedded Checkout and
a shared `SpendCapControl`, customer-portal link, 402/401 interceptor
toast, en-GB i18n. (Per-member usage shows each teammate's spend; the
activity feed is behind a flag until polished.)
**SaaS edge functions** (`Stirling-PDF-SaaS` `v3`) —
`create-checkout-session`, `create-payg-team-subscription`,
`create-customer-portal-session`, `meter-payg-units`,
`payg-subscription-webhook`, `stripe-sync`, plus the stripe-sync
`migrate` + scoped-`backfill` scripts. All price lookup is DB-driven (no
`STRIPE_PAYG_PRICE_ID_*` env vars).
## Release prerequisites (prod)
1. Apply Flyway migrations (`V11`→`V21`) and the Supabase migration
twins.
2. Stripe Sync Engine: run `stripe-sync:migrate`, then a **scoped**
backfill — `product`, `price`, `customer`, `subscription` only (not
`all`, which rate-limits).
3. Register 2 PAYG webhook endpoints (each its own signing secret):
`stripe-sync` (product/price/customer/subscription `.*`) and
`payg-subscription-webhook` (`customer.subscription.created`/`.deleted`
drive state; `.updated` + `invoice.*` observed). Keep the legacy
`stripe-webhook` only if credits/self-hosted flows still run.
4. Stripe Billing Meter: `event_name = payg_doc_units`, value key
`processed_documents`.
5. Env: `PAYG_METER_ENDPOINT` + `SUPABASE_EDGE_FUNCTION_SECRET`
(backend); the webhook signing secrets (edge fns). The default pricing
policy must point at the PAYG Stripe Price(s); `V20` seeds
`free_tier_units = 500`.
## Testing
- `:saas:test` green, `:saas:spotlessCheck` clean, edge-fn Deno tests
green, FE saas typecheck clean (the remaining errors are pre-existing
`proprietary/*` + `prototypes/*`, untouched here). Cucumber shadow-mode
suite + CI workflow included.
## Pre-merge review
An independent dead-code/security pass came back **clean on security**
(team-derived authz / no IDOR, leader-only cap mutation, no
billing-category downgrade, dev/mock hooks gated to
`import.meta.env.DEV` + `/dev/`, no secrets/injection, fail-open
metering by design). The dead/unwired code it flagged has been removed
in this branch (unenforced sub-cap control, an unused JDBC DAO + its
view, dead methods).
## Follow-ups (tracked, not blocking)
- **Enforce per-member sub-caps** — the control was removed because it
read for display but never gated a request; the per-member usage display
and `cap_units` column are retained for when enforcement is wired.
- **API/AI chaining billing model + `ProcessType` enum** — confirm
same-file dedup covers API chaining; define per-tool AI charging; decide
whether the unused enum values stay.
- **Activity feed** — hidden behind a flag until the meter-event surface
is polished.
---------
Co-authored-by: Reece <[email protected]>
Added the create agent. Use [these
prompts](https://github.com/Stirling-Tools/Stirling-PDF-SaaS/blob/main/docgen/backend/default_templates/sample_prompts.md)
to test or try your own :)
Here’s the one I use
```
Hey, I need to generate an employee expense report for reimbursement.
Company: Summit Consulting Partners Company address: 88 Riverside Plaza, Suite 1400, New York, NY 10069 Accounting department email: [email protected]
Employee details:
* Employee Name: Michael Tran
* Employee ID: EMP-1047
* Department: Client Services
* Report Date: January 20th, 2026
* Reporting Period: January 5th, 2026 – January 16th, 2026
* Manager Approver: Laura Simmons
Trip purpose: Client onsite meetings with Atlantic Energy Solutions in Boston, MA.
Expense items:
* Flight (NYC to Boston roundtrip) — $325.40 — January 5th, 2026 — Airline ticket
* Hotel (3 nights at Harborview Hotel) — $822.75 — January 5th-8th, 2026
* Taxi from airport to hotel — $48.00 — January 5th, 2026
* Client dinner (3 attendees) — $186.20 — January 6th, 2026
* Parking at JFK Airport — $72.00 — January 5th-8th, 2026
* Breakfast (per diem not used) — $18.50 — January 7th, 2026
* Uber to client office — $22.10 — January 7th, 2026
* Printing + presentation materials — $46.90 — January 8th, 2026
* Lunch with client — $39.75 — January 8th, 2026
* Office supplies (notebooks, pens) — $27.60 — January 10th, 2026
* Mileage reimbursement (client visit in NJ, 42 miles @ $0.67/mile) — $28.14 — January 14th, 2026
* Team lunch meeting (internal) — $64.30 — January 15th, 2026
Reimbursement method should be direct deposit.
Add a notes section stating: "All receipts attached. Expenses are business-related and comply with company travel policy."
```
---------
Co-authored-by: Anthony Stirling <[email protected]>
# Description of Changes
* Improve typing of API (breaking change but unreleased, frontend also
updated in this PR)
* Add ownership concept to policies
* De-AI the comments
* Update the `task dev:saas` rule to spawn the engine as well
# Description of Changes
Add new triggers:
- Schedule (fires every X amount of time)
- Folder watch (fires whenever the OS tells us a folder has a new file
in it; on Mac this is technically a 2s schedule but that's just how Java
implements it)
Add new sources:
- Folder (reads from this directory)
Add new sinks:
- Inline (stores in FileStorage)
- Folder (stores in specified directory)
Still want to do S3 buckets and web hooks and stuff, but they can come
in a future PR. I'm hoping this should make it sufficient to be able to
integrate with processing folders frontend etc. I've also changed it so
that policies can have multiple sources and triggers at once, which
seems like it might be useful.
Adds an optional MCP server (proprietary module) that exposes Stirling's
PDF operations and AI capabilities to MCP clients. Off by default, zero
footprint when disabled.
### What
- New `/mcp` endpoint: streamable-HTTP + JSON-RPC 2.0; 8 tools
(describe_operation, pages/convert/misc/security category tools, AI,
upload, download).
- Runs real operations over an internal loopback; results returned
inline as base64 (small) or by fileId (large).
### Auth (two modes)
- OAuth2 resource server: RFC 9728 protected-resource metadata, RFC 8707
audience binding, JWKS, `mcp.tools.read/write` scopes; binds each token
to a provisioned Stirling account.
- API-key mode: reuses Stirling per-user `X-API-KEY` (no IdP needed).
### Security
- Per-user file ownership in FileStorage: async/queued writes scoped to
the submitting user; legacy/owner-less files stay readable.
- Admin allow/block list controls which operations are exposed.
- Python engine gated behind a shared secret (`X-Engine-Auth`).
- MCP filter chain is isolated and cannot weaken the main app's
security.
- Hardened: no upstream error-body leakage, log injection sanitized,
fileId path/sidecar enumeration blocked.
### Config / footprint
- Disabled by default (`mcp.enabled=false`); all beans
`@ConditionalOnProperty`.
---
## Checklist
### General
- [ ] I have read the [Contribution
Guidelines](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/CONTRIBUTING.md)
- [ ] I have read the [Stirling-PDF Developer
Guide](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/DeveloperGuide.md)
(if applicable)
- [ ] I have read the [How to add new languages to
Stirling-PDF](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/devGuide/HowToAddNewLanguage.md)
(if applicable)
- [ ] I have performed a self-review of my own code
- [ ] My changes generate no new warnings
### Documentation
- [ ] I have updated relevant docs on [Stirling-PDF's doc
repo](https://github.com/Stirling-Tools/Stirling-Tools.github.io/blob/main/docs/)
(if functionality has heavily changed)
- [ ] I have read the section [Add New Translation
Tags](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/devGuide/HowToAddNewLanguage.md#add-new-translation-tags)
(for new translation tags only)
### Translations (if applicable)
- [ ] I ran
[`scripts/counter_translation.py`](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/docs/counter_translation.md)
### UI Changes (if applicable)
- [ ] Screenshots or videos demonstrating the UI changes are attached
(e.g., as comments or direct attachments in the PR)
### Testing (if applicable)
- [ ] I have run `task check` to verify linters, typechecks, and tests
pass
- [ ] I have tested my changes locally. Refer to the [Testing
Guide](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/DeveloperGuide.md#7-testing)
for more details.
# Description of Changes
Add a backend for running any multi-step PDF operations. This is
designed to be used for the upcoming Policies feature, along with
anything else that will require automated running of PDF operations,
like the Automate tool or Processing Folders.
The implementation is not complete. I've tried to get all the
infrastructure in there so that we can add in whichever triggers we need
in the future (like cron triggers or watching folders on disk) but
currently it just supports manual triggering of the policy.
The basis of this work was the operation running from the Stirling
Engine, which this PR removes in favour of this new system. The only
currently accessible frontend way to test this work is to ask the AI
chat to execute multiple operations on a PDF, but I've also extensively
tested with direct API calls to make sure that the policies work and
persist properly.
### To test
- Ask the agent to “list all the things you can do and put them in a
markdown table”. I know we’re explicitly asking it for markdown, but I
don’t want to update the system prompt to ask it to make tables when
necessary because it’ll probably turn everything into a table, not sure
though, we can test in future.
- Notice how the loading is different
- Notice how the user chat is in a bubble but the agent chat is flat
(super standard design practice in AI tools, and looks much better when
the agent outputs mardown, expecially tables and needs room to do so)
- Ask it to do something different, then close the chat, and see that
the agent is marked as running and has a green outline and a green dot.
- Play around with resizing the chat to make it bigger/smaller
Open to any and all criticisms on any of the design choices, and of
course the usual, code etc.
Resizing
<img width="1572" height="812" alt="Screenshot 2026-06-01 at 2 47 53 PM"
src="https://github.com/user-attachments/assets/ec0ac1d0-01da-4025-bf7e-eea4eb544181"
/>
Loading (cool animation not visible through screenshot obviously)
<img width="559" height="141" alt="Screenshot 2026-06-01 at 2 53 41 PM"
src="https://github.com/user-attachments/assets/99f0b1f5-1719-4d78-8947-21b142293052"
/>
Removed bubbles for agent chat (maybe controversial, let me know) and
markdown now renders properly again
<img width="654" height="1060" alt="Screenshot 2026-06-01 at 2 55 01 PM"
src="https://github.com/user-attachments/assets/445f0889-a632-4751-9a16-f80ae388c632"
/>
# Description of Changes
Change Stirling Engine to support deleting documents automatically. This
happens both on user logout and after an amount of time specified by the
Java when ingesting a document (allowing for personal documents to have
short lifetimes but org documents to be left in the db with no expiry
date). Also sets up an [ACL
policy](https://en.wikipedia.org/wiki/Access-control_list) for the
documents so the database knows which users have access to which
documents. This is not fully implemented in the Java, so currently all
docs are treated as having a single owner, the uploader, but
theoretically when we need to support org storage, we shouldn't need to
change the db schema.
# Description of Changes
Two narrowly-scoped hardening changes to the credits engine.
## 1. CreditService — move Stripe meter call to `afterCommit`
The Stripe metered-usage call sits inside the surrounding
`@Transactional`, holding the `user_credits` row lock for the duration
of an HTTP round-trip to Supabase. Under load this starves concurrent
debits; a transient Stripe blip rolls back a (correct) free-credit
consumption and forces the caller to retry.
The Stripe call now runs in a `TransactionSynchronization.afterCommit`
hook — DB commits first, Stripe fires immediately after. If Stripe fails
after commit, we log + increment a new `credits.stripe_report.failures`
counter; the idempotency key is stable, so a manual replay recovers
without double-charging.
Applied to both `consumeCreditBySupabaseId` and
`consumeCreditWithWaterfall`.
**Dead-code removed:**
- Unreachable UUID fallback for MDC `requestId` — `CorrelationIdFilter`
already guarantees the key on every request.
- The `"Unable to report usage to Stripe"` `RuntimeException` and its
catch block — the afterCommit refactor eliminates the throw path.
- `StripeRollbackOnFailureTest` — pinned the rollback-on-Stripe-fail
behaviour this refactor replaces.
## 2. `@AutoJobPostMapping` — build-time lint for `resourceWeight`
`UnifiedCreditInterceptor` multiplies `resourceWeight` into the per-call
charge. An endpoint that falls through to the annotation default
produces a charge derived from a value nobody chose.
- Annotation default flipped from `1` to `Integer.MIN_VALUE` (sentinel).
Both runtime readers (`UnifiedCreditInterceptor`, `AutoJobAspect`)
already clamp into `[1, 100]` so behaviour is unchanged.
- New `AutoJobPostMappingWeightTest` scans the classpath and fails the
build if any method leaves the sentinel.
- Initial run caught 11 endpoints relying on the default. Explicit
weights now declared, chosen by comparing to peer endpoints:
- `EditTextController` — LARGE
- `EmailController#sendEmailWithAttachment` — SMALL
- `ConvertPDFToMarkdown` — MEDIUM
- `AttachmentController` (extract/list/rename/delete) — SMALL × 4
- `ConvertImgPDFController` (cbr/cbz ↔ pdf) — MEDIUM × 2, LARGE × 2
## Tests
- `StripeUsageIdempotencyKeyTest` — pins the `(supabaseId, overage,
requestId)` idempotency key shape so Stripe always dedupes a retry.
- `StripeAfterCommitOrderingTest` — pins that `afterCommit` fires after
commit and NOT on rollback.
- `AutoJobPostMappingWeightTest` — the lint itself, plus a self-check
that the classpath scan finds at least 10 `@AutoJobPostMapping` methods
(guards against the lint passing vacuously).
Build verified: `ENABLE_SAAS=true ./gradlew :stirling-pdf:test
:saas:test`.
---
## Checklist
### General
- [x] I have read the [Contribution
Guidelines](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/CONTRIBUTING.md)
- [x] I have read the [Stirling-PDF Developer
Guide](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/DeveloperGuide.md)
(if applicable)
- [ ] I have read the [How to add new languages to
Stirling-PDF](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/devGuide/HowToAddNewLanguage.md)
(if applicable) — no translation changes
- [x] I have performed a self-review of my own code
- [x] My changes generate no new warnings
### Documentation
- [ ] I have updated relevant docs on [Stirling-PDF's doc
repo](https://github.com/Stirling-Tools/Stirling-Tools.github.io/blob/main/docs/)
— internal-billing change, no public docs impact
- [ ] I have read the section [Add New Translation
Tags](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/devGuide/HowToAddNewLanguage.md#add-new-translation-tags)
— N/A
### Translations (if applicable)
- [ ] Not applicable
### UI Changes (if applicable)
- [ ] Not applicable
### Testing (if applicable)
- [x] I have run `task check` (via `./gradlew :stirling-pdf:test
:saas:test` with `ENABLE_SAAS=true`) — passes
- [x] I have tested my changes locally
# Description of Changes
## What & why
Customers using ADFS (or any generic OIDC provider that doesn't emit
`email`) hit `Attribute value for 'email' cannot be null` during OAuth2
login with no visibility into what claims the provider actually sent.
The only available remedy was guessing at
`security.oauth2.useAsUsername` until something worked.
This PR adds a new opt-in `security.oauth2.debugLogging` flag (default
`false`). When enabled, `CustomOAuth2UserService` logs:
- All ID token claims (sorted, with values)
- All UserInfo endpoint claims (if any)
- The merged attribute key set Spring exposes to `getAttribute()`
- The value the configured `useAsUsername` actually resolved to
- A **`Hint:`** line listing the claim keys present in the token that
map to a valid `UsernameAttribute` enum value — i.e. exactly what the
operator could put in `useAsUsername` to make login work
Logged at `INFO` on the success path and `ERROR` on failure (inside the
existing `catch (IllegalArgumentException)` block that throws
`OAuth2AuthenticationException`). The block is wrapped with a `[OAUTH2
DEBUG] ... [/OAUTH2 DEBUG]` banner and ends with a PII warning so
operators don't leave it on in production.
Default off → zero observable change for anyone not actively
troubleshooting.
## Files changed
| File | Why |
|---|---|
| `app/common/.../ApplicationProperties.java` | New `debugLogging` field
on the `OAUTH2` config class with javadoc warning about PII |
| `app/core/src/main/resources/settings.yml.template` | Documents
`oauth2.debugLogging` so it appears on next startup |
| `app/proprietary/.../security/service/CustomOAuth2UserService.java` |
Emits the claim dump + suggestion hint when the flag is on |
|
`app/proprietary/.../security/service/CustomOAuth2UserServiceDebugLoggingTest.java`
(new) | Unit test: mocks the OIDC delegate, asserts off-path is silent
and on-path emits the dump with the right Hint contents |
## End-to-end verification
Ran the bundled `testing/compose/docker-compose-keycloak-oauth.yml`
Keycloak realm, configured `security.oauth2.useAsUsername: mail`
(Keycloak emits `email`, not `mail`) and `provider: demarest` (matches
the original customer bug report). Triggered the OAuth flow at
`http://localhost:8080/oauth2/authorization/demarest` and confirmed:
- The ERROR-level dump fires with the full 19-claim ID token decoded
- `-- Value at 'mail' : <NULL — this is why login fails>` correctly
identifies the missing claim
- `-- Hint:` correctly suggests `[email, family_name, given_name,
preferred_username]` (the four keys present that map to valid
`UsernameAttribute` values)
- Auth still fails with the original `OAuth2AuthenticationException` —
no change to control flow, just added diagnostic logging
Unit test (`CustomOAuth2UserServiceDebugLoggingTest`) covers both
branches.
## Reviewer notes
- **No new public APIs.** The flag is config-only; no servlet endpoints
exposed.
- **PII is logged when the flag is on.** This is the whole point —
operators need to see the claims to fix their config — but it's gated,
defaults off, and the dump self-documents with a `WARNING: ... Set
security.oauth2.debugLogging=false once troubleshooting is complete.`
footer.
- **Why log everything, not just sub/email?** Because the operator
doesn't know in advance which claim they actually want. ADFS uses `upn`
in some configs and `preferred_username` in others; Azure AD uses `oid`;
the customer here had neither. Dumping the full set is the only way to
make the diagnostic self-service.
- **Out of scope for this PR (follow-ups):**
- The `UsernameAttribute` enum doesn't include `upn` / `unique_name`
(common ADFS claims). If the customer's token only has `upn`, the Hint
will be empty even though the operator can see `upn` in the dump. Worth
a separate PR to extend the enum.
- The known-provider validator in `Provider.java` (rejects e.g.
`useAsUsername: mail` for `provider: keycloak` at startup) bypasses our
diagnostic for those provider names. ADFS customers using `provider:
<name>` fall into the `default` branch so are not affected — but it's a
sharp edge worth documenting.
---
## Checklist
### General
- [x] I have read the [Contribution
Guidelines](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/CONTRIBUTING.md)
- [x] I have read the [Stirling-PDF Developer
Guide](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/DeveloperGuide.md)
(if applicable)
- [ ] I have read the [How to add new languages to
Stirling-PDF](https://github.com/Stirling-Tools/Stirling-PDF/blob/main/devGuide/HowToAddNewLanguage.md)
(if applicable) — N/A, backend-only change
- [x] I have performed a self-review of my own code
- [x] My changes generate no new warnings
### Documentation
- [ ] Doc-repo update (if functionality has heavily changed) —
diagnostic flag is self-documenting via the `settings.yml.template`
comment and the in-log warning; happy to add a doc-repo entry if
reviewers want one
- [ ] Translation tags — N/A
### UI Changes (if applicable)
- [ ] N/A — backend-only
### Testing (if applicable)
- [x] Unit test added (`CustomOAuth2UserServiceDebugLoggingTest`)
covering on/off paths and Hint correctness
- [x] End-to-end verified locally against bundled Keycloak compose with
intentionally misconfigured `useAsUsername`
- [x] Full `:proprietary:test` suite passes
# Description of Changes
Adds storage in the database for full document content alongside the RAG
content (and changes the service to `DocumentService` instead of
`RagService`). Then adds a generic capability that should be usable by
any agent (currently just used by the Question Agent) which allows the
agent to pull out the full contents of the doc, chunks it into various
sections that will fit in the context window, and then processes them in
parallel to create an intermediate result, and then processes the
intermediate result into a final answer. It will re-chunk as many times
as necessary to get the content small enough for the actual answer to be
analysed (I've tested on PDFs ~3500 pages long, which is well above the
context limit and requires maybe 3 rounds of compression to get an
answer).
The new full doc analysis stuff is heavier than the RAG lookup so both
remain. The agents should use RAG for targeted info and the chunked
reasoner for info that requires reading the full doc.
# Description of Changes
Hooks up the (alpha) PDF Editor backend to the AI engine Edit Agent via
an intermediary API which is easier for the agent to call. It suffers
from all the same issues that the PDF Editor does in actually editing
the text, but should also benefit from any fixes to that.
It also adds protection against the underlying tools misbehaving by
hanging, and fixes a hanging bug in the PDF Editor.
---------
Co-authored-by: EthanHealy01 <[email protected]>
# Description of Changes
Have the Java send a list of enabled endpoints to the AI engine so it
can intelligently respond to the user that the tool does exist but is
disabled on the server so it can't acutally run the operation, instead
of the current behaviour where it sends the API call back and then 503
errors because the execution fails when the URL is disabled.
<img width="380" height="208" alt="image"
src="https://github.com/user-attachments/assets/5842fb2e-2e55-45a5-8205-25515636daae"
/>
---------
Co-authored-by: EthanHealy01 <[email protected]>
# Description of Changes
Flesh out the RAG system and connect it to the PDF Question Agent so it
can respond to questions about PDFs of an extremely large size.
I'd expect lots more work will need to be done to finish off the RAG
system to really be what we need, but this should be a reasonable start
which will let us connect it to tools and have the ingestion mostly
handled automatically. I'm leaving file deletion and proper file ID
management to be done in a future PR. We also need to consider whether
all tools should retrieve content exclusively via RAG, or whether it's
beneficial to have tools sometimes fetch the direct content and other
times fetch it from RAG.
A diagram of the expected interaction is as follows:
```mermaid
sequenceDiagram
autonumber
actor U as User
participant FE as Frontend<br/>(ChatPanel)
participant J as Java<br/>(AiWorkflowService)
participant O as Engine:<br/>OrchestratorAgent
participant QA as Engine:<br/>PdfQuestionAgent
participant RAG as Engine:<br/>RagService + SqliteVecStore
participant V as VoyageAI<br/>(embeddings)
participant L as LLM<br/>(Claude / etc.)
U->>FE: types "Summarise this PDF"<br/>(PDF already uploaded)
FE->>J: POST /api/v1/ai/orchestrate/stream<br/>multipart: fileInputs[], userMessage
Note over J: ByteHashFileIdStrategy<br/>id = sha256(bytes)[:16]
J->>O: POST /api/v1/orchestrator<br/>{ files:[{id,name}], userMessage }
O->>L: route via fast model
L-->>O: delegate_pdf_question
O->>QA: PdfQuestionRequest
loop for each file
QA->>RAG: has_collection(file.id)
RAG-->>QA: false
end
QA-->>O: NeedIngestResponse(files_to_ingest)
O-->>J: { outcome:"need_ingest", filesToIngest:[...] }
Note over J: onNeedIngest
loop per file
J->>J: PDFBox: extract page text
J->>O: POST /api/v1/rag/documents<br/>(long-running timeout)
O->>RAG: chunk + stage documents
O->>V: embed_documents (batches of 256)
V-->>O: embeddings
O->>RAG: add_documents
O-->>J: { chunks_indexed: N }
end
Note over J: retry with resumeWith=pdf_question
J->>O: POST /api/v1/orchestrator
Note over O: fast-path to PdfQuestionAgent
O->>QA: PdfQuestionRequest
Note over QA: build RagCapability<br/>pinned to file IDs
QA->>L: run(prompt) with search_knowledge tool
loop up to max_searches
L->>QA: search_knowledge(query)
QA->>V: embed_query
V-->>QA: query vector
QA->>RAG: search(vector, collections=[file.id])
RAG-->>QA: top-k chunks
QA-->>L: formatted chunks
end
Note over QA: once budget spent,<br/>prepare() hides the tool
L-->>QA: PdfQuestionAnswerResponse
QA-->>O: answer
O-->>J: { outcome:"answer", answer, evidence }
J-->>FE: SSE "result"
FE->>U: assistant bubble
```
# Description of Changes
Add an extra parameter to every agent to receive the conversation
history in addition to the current message. This will make it possible
to answer followup questions from the AI without needing to give full
context in your message.
# Description of Changes
Redesign AI engine so that it autogenerates the `tool_models.py` file
from the OpenAPI spec so the Python has access to the Java API
parameters and the full list of Java tools that it can run. CI ensures
that whenever someone modifies a tool endpoint that the AI enigne tool
models get updated as well (the dev gets told to run `task
engine:tool-models`).
There's loads of advantages to having the Java be the one that actually
executes the tools, rather than the frontend as it was previously set up
to theoretically use:
- The AI gets much better descriptions of the params from the API docs
- It'll be usable headless in the future so a Java daemon could run to
execute ops on files in a folder without the need for the UI to run
- The Java already has all the logic it needs to execute the tools
- We don't need to parse the TypeScript to find the API (which is hard
because the TS wasn't designed to be computer-read to extract the API)
I've also hooked up the prototype frontend to ensure it's working
properly, and have built it in a way that all the tool names can be
translated properly, which was always an issue with previous prototypes
of this.
---------
Co-authored-by: Anthony Stirling <[email protected]>
Co-authored-by: EthanHealy01 <[email protected]>
# Description of Changes
Adds a streaming endpoint to the Java AI orchestrator
(`/api/v1/ai/orchestrate/stream` in addition to the existing
`/api/v1/ai/orchestrate`). This allows the caller to get updates of what
stage of orchestration is being run at the time so UIs can give the user
feedback.
Also contains some dubious Gradle changes to suppress errors coming from
Spotless, when it crashes in Google stuff. I'm not sure if that's
appropriate to add, feel free to ask for changes in review.
Reece BrowneandGitHub
James BruntonandGitHub
ConnorYohandGitHub
Anthony Stirling
dependabot[bot]andGitHub
