The assistant and programs
TeXRA has one assistant. You start every task the same way: open the TeXRA panel (or run texra in a terminal), say what you want and on which files, and send it. There is nothing to pick first.
The assistant reads your project, edits files, runs shell commands, searches arXiv, the web and your Zotero library, runs Wolfram and Python computations, works on Lean 4 proofs, and compiles your documents. For a larger job it does one of two things:
- Starts a program: a packaged job with a fixed procedure, such as a whole-document polish or a manuscript audit. The programs are listed below.
- Hands part of the work to a helper agent that it sets up for that job, with only the tools the job needs, and reads the result before relying on it. Several helpers can run in parallel; read Multi-agent workflows.
You can also start any program yourself, by name, from the terminal.
Converse and call tools in a loop until the task is finished.
Revise the input files in fixed revisions; hand back a versioned diff.
A chat loops: the assistant converses and calls tools until the work is done. A document program revises the input files in fixed revisions and hands back a versioned diff.
Starting a program
In a chat: ask in plain words, for example "polish section 2 of main.tex", or name the program: "run the review program on main.tex". The assistant proposes the run. The proposal card shows the program, the instruction and the files; approve it, or select Edit to change them first.
From the terminal:
texra run <program> --input <file>, with an optional--instruction:bashtexra run edit --input paper.tex --instruction "Polish the introduction" texra run review --input paper.tex --instruction "Check every derivation in Section 3"Every program needs at least one
--inputfile. A name that is not a program prints the list of programs. Read the CLI guide for the other flags.
The assistant can start programs only while the Multi-Agent Workflow plugin is on (Settings → Plugins, or /plugins in the terminal chat). It is on by default.
Document programs
These revise or create LaTeX documents. Each runs a fixed number of revisions, compiles and diffs every revised .tex file, and proposes the result for you to accept. The instruction decides how far the program goes and from which point of view it reads.
| Program | What it does | What the instruction decides |
|---|---|---|
edit | Rewrites the documents. | How far to go: a proofread, a polish for clarity and flow, or new and expanded content. With no instruction it proofreads. |
annotate | Reviews the documents by adding \criticize{} comments. It never rewrites. | The point of view, for example a first-time reader. With none it reads as a referee of the mathematics. |
enhance | Improves the substance: makes the changes it can carry out in full and annotates the rest. | What to look at: the mathematics, the logical structure, or the notation. |
apply | Implements the fixes that inline comments from an earlier review ask for. | Which comments to handle. With none it handles them all. |
merge | Merges a partial edit back into the full original document. | Started by the Merge action. |
transcribe | Turns media files into a LaTeX document: handwriting or scans, or audio, by the media file's type. | Anything specific about the content, such as who the speakers are. |
slides | Turns a paper into a Beamer slide deck on the bundled template. | Length, emphasis and audience. |
poster | Turns a paper into a Beamer poster on the bundled template. | Emphasis and audience. |
A point of view is a sentence of the instruction, not a separate program: "read as a first-time reader" for annotate, "check that the notation is consistent" for enhance. A typical loop is annotate, then apply. Read Programs: how they work for the revisions, the optional critic and where the files land.
edit
Best for: proofreading before submission, tightening prose whose math and structure are settled, adding a derivation the draft skips.
Fix typos, grammar and LaTeX syntax. Do not change the technical content or the writing style.Expand the derivation of equation (12), showing every intermediate step.A polish of the sample document produces a color-coded diff of exactly what changed, with additions in blue and deletions in red:
annotate
Each comment is \criticize{comment}{severity}{confidence}, placed where the issue is: severity runs from 5 (invalidates a main claim) down to 1 (cosmetic), with 0 for a point verified correct, and confidence from 1 (a guess) to 5 (certain). The program checks that a revision changed nothing outside its comments.
Read as a first-time reader and mark every place where you lose the thread.enhance
Marks each change it made with a comment that starts with "Implemented:" and each improvement it left to the authors with one that starts with "Suggestion:".
Make the notation consistent across the sections and the appendix.apply
Reads \criticize{}, \todo{}, \comment{} and % comments, applies the fix each asks for, and removes a comment once it is handled. A comment it cannot resolve stays in place.
Handle only the comments with severity 3 or higher.transcribe
For images, converts handwritten or scanned mathematics into LaTeX that matches the notation of the context documents. For audio, writes a transcript with timestamps, speaker labels, and spoken mathematics typeset as equations; audio needs a model with native audio support (read Working with figures). The output is transcript.tex.
slides and poster
Each writes a draft, then reviews and fixes it. slides writes slides.tex, a deck of 10 to 15 slides unless the instruction says otherwise; poster writes poster.tex.
A 20-minute seminar talk for non-specialists: spend more slides on motivation and state the main theorem without its proof.merge
Takes a partially edited document (one that shows only the changed parts and marks the rest "unchanged") and rebuilds the complete document, copying the unchanged parts from your original verbatim. Read Intelligent merge.
Task programs
Each of these hands the job to one helper agent, told the checklist for that job and given the tools it needs, and answers with that agent's report. File edits it makes are approved like any other edit. The instruction says what to work on; --input names the files.
| Program | What it does | Edits files |
|---|---|---|
review | Audits a manuscript: reproduces key derivations, checks limiting cases and dimensions, notation, references, figures, and that code matches the text. | Only if asked |
search | Finds and synthesizes literature and web sources, citing each with its URL, DOI or arXiv ID; it can add papers to Zotero and download arXiv sources. | No |
derive | Carries out an analytical derivation or symbolic computation, with Wolfram Language through wolframscript where it helps, and verifies the result. | Yes |
solve | Attacks a research-level or open problem: surveys what is known, computes small cases, hunts for counterexamples, and writes a proof, disproof or scoped partial result. | Yes |
experiment | Designs, runs and validates a numerical experiment (a simulation, a convergence study, a reproduction of a figure) and reports what it shows. | Yes |
code | Implements a feature, makes a targeted edit or fixes a bug in the project's code, then runs its build and tests. | Yes |
review-code | Reviews a diff or the named files for correctness, clarity and security, and answers with prioritized findings. | No |
simplify | Simplifies working code or LaTeX without changing what it does, then confirms the tests or the build still pass. | Yes |
test | Writes or repairs tests with the project's own framework and runs them. | Yes |
fix-latex | Diagnoses and fixes compilation errors, broken references and citations, warnings and bad boxes so the document builds. | Yes, in place |
latexdiff | Produces a compiled visual diff of two versions of a LaTeX document with latexdiff; the old and new files are the inputs. | Writes the diff |
Examples:
texra run review --input paper.tex --instruction "Confirm every goal the abstract states is delivered"
texra run solve --input problem.tex --instruction "Check the problem's status in the literature first"
texra run experiment --input solver.py --instruction "Sweep the step size and report the observed order of accuracy"
texra run fix-latex --input main.tex
texra run latexdiff --input v1/main.tex --input main.texIn a chat you rarely need to name them: "audit this manuscript" or "find recent papers on tensor-network contraction" is enough, and the assistant decides whether to do the work itself or start the program.
Lean programs
The Lean plugin adds four: prove completes the proofs in .lean files until Lean reports no error and no sorry and the files declare no new axiom; lean-simplify brings them toward Mathlib quality without changing what they state, under the same check; blueprint writes or updates a LeanBlueprint document until leanblueprint web and leanblueprint checkdecls pass; and lean-search finds what Mathlib and the project already have for a statement or an API and says what is missing. Read Lean 4 proofs.
Your own programs
A plugin can carry programs of its own, and the assistant can write a program for itself that it reuses in later conversations. Read Tailoring TeXRA.
If you used 0.40
There is no agent menu any more, and no --agent flag. Ask the assistant for what an agent used to do, or start the matching program: review, search, solve (was prover), derive (was research), experiment (was numerics), code (was coder and engineer), review-code, simplify, test, fix-latex (was latexFixer) and latexdiff. Slide decks and posters are the slides and poster programs; the Lean agents' jobs are prove, lean-simplify (was leanSimplifier), lean-search (was leanSearch) and blueprint. Custom agent files are not read; see Tailoring TeXRA.
Next steps
- Programs: how they work: revisions, the critic, and run storage
- Multi-agent workflows: helper agents and scripts
- Research tools: literature and computation
- Models: choosing a model