Conversation
Split the bespoke `/api/v1/ai/pdf-comment-agent` into a reusable primitive
plus a composed AI tool, eliminating the name-to-path dispatch plumbing.
New shared primitives (common):
- `PdfAnnotationService` + `AnnotationLocation` / `StickyNoteSpec` records
hold the PDFBox sticky-note logic. Reusable from any caller: Automate
workflows, AI agents, scripts, tests.
New deterministic tool (core):
- `POST /api/v1/misc/add-comments` takes PDF + JSON comment specs and
returns an annotated PDF. Same contract every other `/api/v1/misc/*`
tool follows — bit-identical output for identical input.
Endpoint relocation (proprietary):
- Moved PDF Comment Agent from `/api/v1/ai/pdf-comment-agent` to
`/api/v1/misc/pdf-comment-agent` so it passes the primary dispatch
allowlist regex without any special plumbing. Controller slimmed to
delegate annotation placement to the shared `PdfAnnotationService`.
Removed plumbing:
- Deleted `AgentTool` registry + tests.
- Reverted `InternalApiClient`'s secondary allowlist branch.
- Reverted `AiWorkflowService.resolveToolPath` translation.
- Dropped `AgentToolId.PDF_COMMENT_AGENT` from the engine's agent-tool
registry (math-auditor still uses it pending the Phase B refactor).
Python orchestrator:
- `delegate_pdf_review` replaces `delegate_pdf_comment`. Emits a plan
step targeting `ToolEndpoint.PDF_COMMENT_AGENT` (the standard tool
path) instead of a bare agent id. System prompt now distinguishes
"review" (annotated PDF) from "question" (chat answer).
- `tool_models.py` gains `CommentSpec`, `AddCommentsParams`, and
`PdfCommentAgentParams` with their `ToolEndpoint` entries.
Frontend:
- Tool op config posts to the new `/api/v1/misc/pdf-comment-agent` URL.
Verification:
- Java: `:common:test`, `:proprietary:test`, `:stirling-pdf:test` for
affected modules — all green.
- Python: ruff + pyright clean, pytest 10/10.
- Live: direct add-comments, direct pdf-comment-agent, and orchestrator
dispatch ("Add review comments to this PDF...") all return annotated
PDFs with meaningful sticky notes.
The math auditor is one specialist that produces a Verdict. How that
Verdict is presented to the user (chat answer vs annotated PDF) is the
meta-agent's job, not the specialist's. Before this commit we had two
top-level math delegates and an ``output_mode`` on the math controller —
presentation concerns leaking all the way down.
- **Specialists** produce structured findings. Stateless. No presentation.
- ``MathAuditorAgent`` → ``Verdict`` JSON. Always.
- ``PdfCommentAgent`` → ``PdfCommentResponse``.
- **Meta-agents** own presentation and may delegate to specialists.
- ``delegate_pdf_question`` — math intent → consult math-auditor →
render Verdict as prose (``outcome=answer``).
- ``delegate_pdf_review`` — math intent → consult math-auditor →
project Verdict to ``CommentSpec``s → ``/api/v1/misc/add-comments``.
- Non-math review still uses the composed ``pdf-comment-agent`` tool.
- **Orchestrator top-level** stays simple: question / review / edit. No
math-specific branch.
A delegate can emit an ``EditPlanResponse`` with ``resume_with`` set.
``AiWorkflowService`` runs the plan, captures any structured report (JSON
body or ``X-Stirling-Tool-Report`` header), attaches it as a new
``ToolReportArtifact`` on the next ``WorkflowTurnRequest``, and re-enters
the orchestrator. The delegate's ``_run`` method finds the report in its
artifacts and produces its final output.
- New ``ArtifactKind.TOOL_REPORT`` + ``ToolReportArtifact`` contract.
- ``EditPlanResponse.resume_with`` field (optional
``SupportedCapability``).
- New ``SupportedCapability.PDF_REVIEW`` for the review resume path.
- New ``stirling.agents.math_presentation`` module with
``is_math_intent``, ``extract_math_verdict``, ``verdict_to_prose``,
``verdict_to_comment_specs``, ``verdict_to_add_comments_payload``.
- Orchestrator: dropped ``delegate_math_report`` /
``delegate_math_review`` / the old ``math_auditor_agent`` delegate.
``_run_pdf_question`` / ``_run_pdf_review`` gained math-intent
routing with multi-turn resume. ``_resume`` handles ``PDF_REVIEW``.
- ``tool_models.py``: slimmed ``MathAuditorAgentParams`` back to just
``tolerance`` — no output mode.
- Deleted ``agent_tool_models.py`` — no longer needed.
- ``PdfContentExtractor``: ``ArtifactKind.TOOL_REPORT`` enum value +
``ToolReportArtifact`` record alongside ``ExtractedTextArtifact``.
- ``AiWorkflowService.onPlan``: after executing steps, if the
orchestrator response had ``resumeWith`` set AND we captured a tool
report, build a follow-up ``WorkflowTurnRequest`` with the report
attached as a ``ToolReportArtifact`` and return ``Pending`` — the
existing outer ``while`` loop re-enters the engine.
- ``MathAuditorAgentController``: reverted to a pure specialist.
Always returns ``Verdict`` JSON. No ``outputMode``, no comment
projection, no report header.
- Deleted ``MathAuditorOutputMode`` + ``MathAuditorCommentProjector``
+ its test — presentation is the caller's concern now.
- ``AiWorkflowResponse.report`` (``JsonNode``, optional) from the prior
round kept — it remains useful as a file+metadata channel for tools
like ``pdf-comment-agent`` that surface ``X-Stirling-Tool-Report``
alongside a file body.
- Java ``:common:test :proprietary:test :stirling-pdf:test`` green.
- Python ``ruff`` + ``pyright`` + ``pytest`` (122 tests) green.
- Live end-to-end:
- "Is the math correct?" → ``outcome=answer`` with a prose Verdict
summary.
- "Review the math and leave comments" → ``outcome=completed`` with
annotated PDF, 3 sticky notes per the Verdict's discrepancies.
- "Review this PDF and comment on unclear wording" → unchanged,
uses ``pdf-comment-agent``; ``report`` carries annotation counts
and the agent's rationale.
Previously the tool was registered in the core tool registry and visible
from every build (proprietary/saas/desktop/core). That's wrong for an
experimental AI tool — it should only ship in the prototypes bundle until
it graduates.
## Mechanism
Follow the existing proprietary-overlay stub/shadow pattern (see
`useProprietaryToolRegistry`), one tier deeper:
- `core/types/prototypeToolId.ts` — empty stub (PROTOTYPE_*_TOOL_IDS = []).
- `prototypes/types/prototypeToolId.ts` — real impl listing "pdfCommentAgent".
- `core/data/usePrototypeToolRegistry.tsx` — empty hook.
- `prototypes/data/usePrototypeToolRegistry.tsx` — real hook returning the
registry entry (icon, component, i18n keys, operation config, etc.).
`@app/*` alias resolution:
- Proprietary/saas/desktop/core builds → prototype stubs are empty. ToolId
union excludes "pdfCommentAgent"; tool picker doesn't show the entry;
the tool component isn't imported at all (tree-shakes out of the bundle).
- Prototypes build (`@app/* → prototypes → proprietary → core`) → real
stubs win. ToolId includes "pdfCommentAgent"; tool picker shows it;
bundle includes the component.
## Changes
- Moved (git mv, history preserved):
- `core/tools/PdfCommentAgent.tsx` → `prototypes/tools/PdfCommentAgent.tsx`
- `core/hooks/tools/pdfCommentAgent/*` → `prototypes/hooks/tools/pdfCommentAgent/*`
- New files:
- `core/types/prototypeToolId.ts` (stub)
- `core/data/usePrototypeToolRegistry.tsx` (stub)
- `prototypes/types/prototypeToolId.ts` (real)
- `prototypes/data/usePrototypeToolRegistry.tsx` (real)
- Modified:
- `core/types/toolId.ts` — import PROTOTYPE_* unions, add to
REGULAR/SUPER/LINK_TOOL_IDS, remove "pdfCommentAgent" from
CORE_REGULAR_TOOL_IDS.
- `core/data/toolsTaxonomy.ts` — new `PrototypeToolRegistry` type.
- `core/data/useTranslatedToolRegistry.tsx` — drop direct imports and
registry entry for pdfCommentAgent; spread `usePrototypeToolRegistry()`
alongside the proprietary overlay.
- `proprietary/utils/creditCosts.ts` — loosen `TOOL_CREDIT_COSTS` type to
`Partial<Record<ToolId, number>>` so overlays can contribute build-only
ids without every overlay needing to know the others'. Fallback to
MEDIUM already handled in `getToolCreditCost`.
## Verification
- `npx tsc -p tsconfig.json --noEmit` (proprietary/default build):
zero errors mentioning pdfCommentAgent or our new files. The 13
remaining errors are all pre-existing test-file Axios/Tauri typing
issues unrelated to this branch.
- `npx tsc -p src/prototypes/tsconfig.json --noEmit` (prototypes build):
one pre-existing error in `ChatContext.tsx:284`; `"pdfCommentAgent"`
is in the `ToolId` union as expected.
…into pdf-comment-agent
ConnorYoh
requested review from
EthanHealy01,
Frooodle,
Ludy87,
jbrunton96 and
reecebrowne
as code owners
dosubot
Bot
added
size:XXL
stirlingbot
Bot
added
Java
![aikido-pr-checks[bot]](https://avatars.githubusercontent.com/in/898896?s=60&v=4){kind=small}
Previous round put agent tools under /api/v1/misc/ as a plumbing workaround — InternalApiClient's dispatch allowlist only permitted a fixed set of tool namespaces, and /api/v1/ai/ wasn't one of them. Putting AI tools under misc is semantically wrong. Carve out /api/v1/ai/tools/ as the explicit sub-namespace for dispatchable AI tools and extend the allowlist to include it. /api/v1/ai/orchestrate and /api/v1/ai/health remain off-limits to internal dispatch so a plan step can't re-enter the orchestrator (recursion risk). Net namespaces: - /api/v1/ai/* AI-facing endpoints (orchestrate, health, ...) - /api/v1/ai/tools/* Dispatchable AI tools - /api/v1/misc/* Deterministic primitives Changes: - MathAuditorAgentController @RequestMapping → "/api/v1/ai/tools" - PdfCommentAgentController @RequestMapping → "/api/v1/ai/tools" - InternalApiClient ALLOWED_ENDPOINT_PATH regex adds "|^/api/v1/ai/tools(/[A-Za-z0-9_-]+)+$" alternation. - Python ToolEndpoint: MATH_AUDITOR_AGENT + PDF_COMMENT_AGENT values updated accordingly. - Frontend PDF_COMMENT_AGENT_ENDPOINT constant updated. - Java + Python tests updated; new InternalApiClientTest case asserting /api/v1/ai/tools/* is accepted while /api/v1/ai/orchestrate is still rejected. Also in this commit — unrelated collateral fix: - AddCommentsController + its test updated to match WebResponseUtils. pdfFileToWebResponse's new return type (ResponseEntity<Resource>), which changed in upstream PR #6160 "Migrate stream to resource for stability" (merged via the recent main merge). Live verification: - POST /api/v1/ai/tools/pdf-comment-agent — 200 + annotated PDF + report header. - POST /api/v1/ai/tools/math-auditor-agent — 200 + Verdict JSON. - Orchestrator "review this PDF" → outcome=completed with annotated PDF. - Orchestrator "review the math" → outcome=completed with sticky-note PDF. - Orchestrator "is the math correct?" → outcome=answer with prose summary.
Math-auditor review previously stacked all sticky notes in the right margin because the Python projector emitted fixed coordinates. Now each CommentSpec can supply an anchorText hint; the add-comments tool locates the first matching line on the target page and anchors the icon there, falling back to the caller's coordinates when no match is found. The projection helper passes the discrepancy's stated value (or context when stated is empty) as the anchor, so sticky notes land on the line that was actually flagged.
- Remove engine/src/trace* log dumps that got committed and gitignore trace/trace.* so they stop re-appearing. - Restore frontend/src/core/hooks/useIndexedDBThumbnail.ts to main — the thumbnail-regeneration fix from #6134 was partially reverted here by a bad merge; it's unrelated to this PR. - MathAuditorAgentController javadoc said the endpoint lives under /api/v1/misc/; it was moved to /api/v1/ai/tools/ in c87f24f.
- AiToolInputValidator: new shared validator enforcing the PDF content type and a 50 MB upper bound on uploads. Spring's multipart limit is tuned for the regular tools (2 GB), which is far too permissive once a payload feeds into an LLM token budget. Both PdfCommentAgent and MathAuditorAgent now route uploads through it, replacing the duplicated inline content-type checks. - PdfAnnotationService: drop specs with non-positive width/height (previously produced invisible/zero-size annotations) and with text over 100 000 chars (pathological / accidental document-dump). Matches the existing "skip and log, never throw for one bad spec" convention so every caller benefits, not just AI tools.
- AiEngineClient: surface network IOException from httpClient.send as ResponseStatusException — HttpTimeoutException becomes 504 GATEWAY_TIMEOUT, everything else becomes 503 SERVICE_UNAVAILABLE with the original cause chained. Previously these leaked out as raw IOException and every caller (PdfCommentAgent, MathAuditorAgent, AiWorkflowService, AiEngineController) bubbled them up as an unstructured 500. Adds a package-private constructor that accepts an HttpClient for testing. - pdf_comment agent: JSON-encode request.user_message when building the prompt rather than inlining it between literal triple quotes. The old format was trivially breakable by a user message containing `"""` followed by fake chunk records, which would spoof structure. Chunks were already JSON-encoded; this aligns the user-prompt path.
- New AiToolResponseHeaders class holds the X-Stirling-Tool-Report constant; both PdfCommentAgentController (producer) and AiWorkflowService (consumer) now reference it, dropping the self-acknowledged "must stay in sync" duplication. - pdfCommentAgentOperationConfig.ts now sanitises the filename pulled from the Content-Disposition header, rejecting blank/whitespace values and anything containing path separators before returning it. Hostile or buggy backend can no longer steer the browser save into a parent directory; falls back to <inputName>-commented.pdf.
- math_presentation: add tests for extract_math_verdict — the resume-turn entry point — covering happy-path round-trip from ToolReportArtifact back to Verdict, empty artifacts (first turn), non-math reports being ignored, and graceful degradation on malformed reports. Also cover verdict_to_prose for the pdf_question math path (clean verdict, errors + warnings rendering). - pdf-comment-agent routes: add test that an agent-layer exception surfaces as HTTP 500, so Java's AiEngineClient can map it to 502 instead of mis-applying a spurious success. - AddCommentsController: cover the empty-array case — callers with zero comments to place get a 200 + unmodified PDF instead of having to special-case the no-op.
…nguage Addresses PR comment #7. The math auditor is the one finding issues; turning the structured Verdict into prose or sticky-note text is the consumer's job, and the consumer should answer in the same language the user asked in. No English glue should leak from the engine. - math_presentation.py: dropped verdict_to_prose, verdict_to_add_comments_payload, verdict_to_comment_specs, _comment_body, _severity_label, _comment_subject, _discrepancy_one_liner, _anchor_text_for. Module is now just the language-agnostic helpers (is_math_intent, extract_math_verdict). - PdfQuestionAgent: small math-synth LLM (fast model) takes the Verdict + the user's original question and answers in the same language. Drops verdict_to_prose entirely. - PdfReviewAgent: localiser LLM produces sticky-note subject + text per Discrepancy, in the user's language. Python combines that with deterministic placement geometry (page, x/y stacking, anchor_text) to build the CommentSpec list. No English in the response shape. - EditPlan summaries on first turn / pre-tool turn are emptied so no English UI hint leaks; frontend handles its own loading state. - Tests: math_presentation tests slimmed to the helpers that remain; the placement-geometry and camelCase-serialisation tests moved to the new test_pdf_review.py with the localiser LLM mocked.
The is_math_intent regex was English-only and over-eager — "review the invoices for ambiguous wording" matched on "invoices" and routed prose review through the math auditor. Both problems disappear when the orchestrator's existing top-level LLM (already classifying the request in any language) emits the math-intent decision alongside the routing. - Orchestrator: delegate_pdf_question and delegate_pdf_review now take consult_math_auditor: bool, threaded through to _run_pdf_question and _run_pdf_review (with a False default so the resume-turn fast-path still works). System prompt updated with a language-neutral rule for the flag. - PdfQuestionAgent.orchestrate / PdfReviewAgent.orchestrate take the flag and use it on the first turn. Resume turns are still detected by Verdict-artifact presence and are unaffected. - math_presentation.py: dropped is_math_intent and _MATH_KEYWORDS. Module is now just extract_math_verdict. - Both consumer system prompts now require verbatim quoting of any stated/expected numeric values from the Verdict — guards against the LLM paraphrasing $215,000 as "about $215k". - Tests: dropped the keyword test; added flag-driven first-turn tests for both delegates and prompt-pinning tests for the verbatim rule.
…riant Closes the loop on PR comment #9. Previously the question delegate's return type leaked the union PdfQuestionResponse | EditPlanResponse because the math-intent first turn needed to ask the caller to run a side-quest. J's suggestion was to embed the plan as a nullable member of the question response — a single, clean type with optional pre-work attached. - contracts/pdf_questions.py: PdfQuestionAnswerResponse gains edit_plan: EditPlanResponse | None. Outcome stays ANSWER on this turn; answer is empty string until the resume turn produces it. - agents/pdf_questions.py: orchestrate() returns PdfQuestionAnswerResponse(answer="", edit_plan=...) on the math first turn instead of bare EditPlanResponse. Return type collapses to PdfQuestionResponse. - agents/orchestrator.py: delegate_pdf_question and _run_pdf_question return type drop the | EditPlanResponse — single signature now. - Java side: - New AiWorkflowEditPlan POJO mirrors the engine's EditPlanResponse shape (summary, rationale, steps, resumeWith). - AiWorkflowResponse gains a nested editPlan field. - AiWorkflowService.advance() splits ANSWER into its own branch and dispatches to onAnswer, which checks editPlan and runs the embedded plan if present. onPlan is refactored to delegate to a runPlan helper that takes the plan fields directly so both call sites share the iteration + resume logic. - Tests updated to assert on PdfQuestionAnswerResponse with edit_plan set rather than a bare EditPlanResponse return.
Reverts the orchestrator-flag approach. The orchestrator no longer knows what specialists exist; the question and review delegates each ask their own tiny classifier LLM whether the prompt needs the math auditor. Decision stays language-agnostic (LLM, not regex) and the classifier code is shared so both delegates use the same prompt and shape. - math_presentation.py: new MathIntentClassifier — small LLM (fast model) with a structured bool output. Reused by both consumers. - pdf_questions.py / pdf_review.py: drop the consult_math_auditor parameter from orchestrate(); each instantiates the classifier in __init__ and calls it on the first turn (resume turns short-circuit via the Verdict-artifact check, so no LLM call there). - orchestrator.py: delegate_pdf_question / delegate_pdf_review and the matching _run_* methods drop the flag parameter. System prompt drops the math-routing clause (it's no longer the orchestrator's concern). - Tests updated to mock the classifier where the flag used to be passed. New assertions confirm resume turns don't call the classifier. Cost: one extra small-model LLM call per first turn of Q&A or Review. Acceptable given cleaner separation — the orchestrator stays a pure router and each consumer owns its own intent decision.
…heck ChatContext stamps ToolOperation.toolId="ai-workflow" on files coming back from the AI orchestrator (introduced in #6116). The ToolId union didn't include it, so the assignment was a latent type error — never caught because typecheck:prototypes wasn't in typecheck:all until the recent +1 review-point fix added it. Add "ai-workflow" to PROTOTYPE_SUPER_TOOL_IDS so it satisfies ToolId. Keeps it out of any other build's union (proprietary/core/saas/desktop) and out of the prototype tool registry's user-facing entries — it's a generic marker for orchestrator-produced files, not a pickable tool. Follow-up worth doing later: surface the terminal tool's endpoint path on AiWorkflowResponse so we can record the actual tool that produced the file rather than this generic marker.
jbrunton96
reviewed
Apr 28, 2026
PR comment from James — test directories aren't packages and shouldn't need __init__.py files. Removed the empty ones; switched pytest to the modern importlib import mode so duplicate basenames (e.g. test_routes.py in two different test folders) still collect cleanly without needing package markers. Added ``tests`` to pythonpath so shared conftest helpers keep importing.
PR comment from James: the "If the user is asking a question about the PDF contents (and does NOT want comments written onto the document) use delegate_pdf_question instead" sentence read like an over-correction to something I'd hit in development and added too much weight. The prior sentence already covers when to use delegate_pdf_question.
PR comment from James — the bare ``except Exception`` with a noqa was covering more than necessary. ``model_validate`` only raises pydantic.ValidationError on a malformed payload, so catch that and let anything else propagate.
PR comment from James — Pretty handles model dumping internally and lazily, so calling ``response.model_dump()`` upfront does the work even when the log level is above DEBUG.
PR comment from James — only some of the subagents had been slimmed in the previous round, which read inconsistently. Push the same pattern through PdfEditAgent and UserSpecAgent: each gets an ``orchestrate(OrchestratorRequest)`` adapter that builds the typed request the inner pipeline expects, so the orchestrator's _run_* methods collapse to one-liners. ``handle()`` / ``draft()`` keep their typed request signatures so the direct API routes don't have to change. Will fold these into whatever shared agent interface lands in #6163 once that's settled.
PR comment from James — the previous ``report: dict[str, Any]`` meant any JSON could pass through as a tool report and the meta-agent would only discover problems when it tried to read it. Pin the shape: - ToolReportArtifact is now ``MathAuditorToolReportArtifact`` with ``report: Verdict``. Pydantic validates the payload on receipt — a malformed Verdict is rejected before extract_math_verdict ever sees it, rather than being silently swallowed. - ToolReportArtifact stays as a type alias so callers don't have to know there's only one variant today; lifts into a discriminated union when another consumer-side report appears. - extract_math_verdict shrinks accordingly — no more model_validate, no try/except, just an isinstance check on the typed artifact. - PdfCommentReport added in contracts/pdf_comments.py for the report the pdf-comment-agent surfaces alongside the annotated PDF (annotations applied, instructions received, rationale). Lands as the top-level AiWorkflowResponse.report on COMPLETED — no resume, so it never re- enters the orchestrator as an artifact. - Tests updated: construct artifacts with the typed model directly; the old "graceful degradation on malformed JSON" test becomes a "validation rejects malformed payload at receipt" test.
jbrunton96
previously approved these changes
Apr 30, 2026
Co-authored-by: James Brunton <jbrunton96@gmail.com>
jbrunton96
approved these changes
Apr 30, 2026
reecebrowne
approved these changes
Apr 30, 2026
github-merge-queue
Bot
removed this pull request from the merge queue due to no response for status checks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PDF Comment Agent
AI tool that annotates a PDF with sticky-note comments based on a user prompt. Prototype-gated.
New endpoints
POST /api/v1/ai/tools/pdf-comment-agent— composed AI tool. Takes PDF +prompt. Extracts text chunks, calls the engine for review comments, resolves positions, applies viaadd-comments. Returns annotated PDF +X-Stirling-Tool-Reportheader with{annotationsApplied, rationale}.POST /api/v1/ai/tools/math-auditor-agent— relocated from/api/v1/ai/. Pure specialist, returnsVerdictJSON.POST /api/v1/misc/add-comments— deterministic Java primitive. Takes PDF + JSON comment specs. Reusable from Automate, scripts, any caller.Namespace convention:
/api/v1/ai/tools/*for dispatchable AI tools (on the orchestrator allowlist),/api/v1/ai/*for non-dispatchable AI endpoints (orchestrate, health — recursion-safe),/api/v1/misc/*for deterministic primitives.Orchestrator changes
delegate_pdf_review— plans review annotations, either via math-auditor +add-comments(math intent) or viapdf-comment-agent(prose intent).delegate_pdf_question+delegate_pdf_reviewdetect math intent and consult math-auditor via a multi-turn plan:resume_with=<capability>+ a newToolReportArtifactcarries theVerdictback to the delegate, which projects it to prose (question) or sticky-note specs (review).math_auditor_agentdelegate and theAgentToolname→path registry — agent tools now live under/api/v1/ai/tools/*and dispatch through the standardInternalApiClientallowlist.Test in the UI
Run the prototypes build:
Tool picker → Advanced → Document Review → Add AI comments. Upload a PDF, type a prompt, submit. Review pane shows the annotated PDF with clickable sticky-note icons.
Chat panel — orchestrator routes:
Verify gating: build in proprietary (default) mode — the tool must NOT appear in the picker and the chat orchestrator shouldn't target /api/v1/ai/tools/pdf-comment-agent