The paper model gained two things while the MCP layer was not looking, and
both were invisible from a tool call.
`paper_sentence.line_break_before` — the flag a paragraph needs so a formula,
a list of items or an author line can start a line of its own — was not in
`SentenceSpec`, and `sentence_inputs` dropped it on the floor: a model could
write the field, watch the call succeed, and get prose back. It is now part of
every write path (`paragraph_write`, `paper_write`, `sentence_add`,
`sentence_update`) and of every read path (`sentence_row` reports it, so a
model can see the lines before moving one). Two bugs fell out of doing it
properly:
* `paragraph_move` rebuilt its sentences without the flag, so moving a
paragraph that carried a formula turned it into prose. It also passed the
*source* sorts through, which interleaved the moved sentences with the
target's own — the merged paragraph read in an order neither writer wrote.
Both are fixed; the tool now appends, as its own description promised.
* `document_text` printed a formula run into the sentence before it. The paper
view renders `$$…$$` as a block, so the seam rules are now the same in both
places, with a comment on each side naming the other.
`PaperDocumentRead.references` and `reference_style` — the per-work 参考文献
list the backend now assembles and formats — were reachable over REST but not
over MCP: `paper_document(format="json")` returned citations and nothing to
point them at, even though docs/MCP.md already described the list. The json
form now carries the entries (index, reference_id, text, missing, citations)
and the style name, so a model quotes the bibliography instead of assembling a
second one out of the same fields.
smoke_mcp.py grows from 76 checks to 99 over two new steps — the line-break
step pins the display-formula rule, the marker toggle and the moving case; the
bibliography step cites one work from two sentences and checks that the text,
the entry and both markers agree on one number.
The paper is written by a model now. A browser is the wrong client for that:
the work is "generate the paper, then put it in", and doing it through a form
means a person retyping what a model already produced. So the same domain is
served over the Model Context Protocol, which Claude Code, Codex and the
DeepSeek Harness all speak.
It is a third front door, not a second implementation. Every tool is three
lines around an `app.crud` call and validates through `app.schemas`, exactly
as the REST routes do, so a rule fixed in the CRUD layer is fixed on both
surfaces and a paper written by an agent is indistinguishable from one written
by hand. What `app/mcp/` adds is only what a model needs and a browser does
not:
- 29 tools, prefixed `paper_` `paragraph_` `sentence_` `template_` `field_`,
because a model picks a tool out of a list by name rather than by reading 29
descriptions;
- results as compact `None`-free JSON, since a tool result is paid for in
context tokens and `PaperRead.model_dump()` carries four counts and two
timestamps into every list row;
- paragraphs addressed by **heading** as well as by position. Storage is
correct as it stands — a sentence remembers the position it sits at, which is
what makes a template switch non-destructive — but nobody writing
"1. Introduction" knows the template places it at `sort = 20`. The server
translates, and refuses with the real heading list when it cannot, so a model
that guessed wrong corrects itself in one retry;
- `paper_write` and `paper_write_text`: one intention, one call. The latter
finds its own sections from Markdown headings or from lines that name a
template heading, and reports every heading it could not place instead of
writing half a paper;
- `sentence_search` across papers, for consistency rather than retrieval — a
paper that says 洪水损失 should not be joined by one that says GUL;
- `paper_delete` refuses once, naming what would go with it. A cascading delete
has no undo in a tool call.
Two transports, one build. `stdio` is what a client spawns — so nothing in the
process may print to stdout, and diagnostics go to stderr. `streamable-http` is
what a client on another machine connects to, optionally behind a bearer token;
binding a non-loopback address disables the SDK's DNS-rebinding allow-list,
because a LAN client sends whatever Host it knows the server by.
Tools register with `structured_output=False` on purpose: inferred from a
`-> str` annotation the SDK publishes a `{"result": ...}` envelope and sends
the JSON twice, once as `structuredContent` and once as text, and clients that
read only one of the two then disagree about what came back.
`scripts/smoke_mcp.py` drives the whole loop through a real MCP client — the
child process and JSON-RPC over stdin/stdout a client actually uses — and runs
unchanged against a running HTTP server via `--url`. 47 checks pass on both
transports; the REST suite still passes its 45.