Error Message Manager

error-message-manager

Rewrite vague, delayed, or low-context Julia error messages into structured, actionable diagnostics. The goal is errors that tell the user exactly what went wrong, what was expected, what was received, and—whenever a likely fix exists— what to do next. Prefer catching mistakes early (at API boundaries) over letting them propagate into cryptic numerical failures.

Workflow

Step 0 — Offer an Explore agent for multi-file scope

If the user's request covers more than one file — a whole directory, a module, or the entire repo — offer to spawn an Explore agent before doing any file reads yourself. The agent runs all the reads in parallel without flooding the main context, and returns a structured inventory you can act on directly.

When to offer: any time the target is a directory path (e.g. src/) or a vague scope like "the whole package" or "all the source files".

Offer text (adapt as needed):

"This spans multiple files — I'd recommend spawning an Explore agent to survey all throw/@assert/error sites in parallel. It keeps the audit fast and leaves the main context clean for the actual rewrites. Want me to do that?"

Agent prompt to use (fill in <path> and <package_name>):

Audit `<path>` for error-raising patterns. For every `@assert`, `error(`, or
`throw(` site in every `.jl` file:

1. Record: file, line number, exception type (or "bare @assert" / "bare error"),
   and the full message text (including multiline strings).
2. Classify message quality:
   - "good"  — has `$(expr)` interpolation showing the actual received value, and
     is either short (≤3 lines) or already in a `_throw_` helper function
   - "long-inline" — message content is good, but the body exceeds 3 lines and
     the throw is written inline (not in a `_throw_` helper)
   - "vague" — missing a received value, or no Expected/Got structure
   - "missing" — bare `@assert` with no message at all
3. Note whether the site is at an API boundary (user-facing input) or an internal
   invariant (would require a package bug to fire).

Return a markdown table with columns:
  File | Line | Exception type | Quality | Notes (one-line note on what's wrong if vague/missing/long-inline)

Focus only on sites that are "vague", "missing", or "long-inline" — skip "good" ones.

How to use the result: treat the returned table as your working inventory for Steps 1 and 2. You do not need to re-read the flagged files yourself to classify — go straight to reading only the lines that need rewrites (Step 3 onwards).

Step 1 — Audit the target scope

Identify which files or functions to address. If the user named a specific function, start there. If the request is repo-wide, run:

rg -n '(@assert[^(]|@assert\(|error\(string\(|throw\(ErrorException)' src/

Then collect all message-less @assert calls:

rg -n '@assert' src/ | grep -v '"'

Also flag @warn calls that use string concatenation instead of interpolation:

rg -n '@warn\s+string\(' src/

And flag catch blocks that discard the original exception when re-throwing:

rg -n 'catch\s' src/

For each catch hit, check whether the subsequent throw or error call interpolates the caught variable (e.g. $e or sprint(showerror, e)). If it does not, the original exception type and message are silently lost.

For each hit, record: file, line, the condition being checked, and whether it guards user-provided input (API boundary) or an internal invariant.

Step 2 — Classify each site

Use this table to choose the right exception type:

Avoid ErrorException unless there is no better choice.

Type mismatches vs dimension mismatches: an @assert isa(x_mean, AbstractVector{FT}) inside an if isa(x, AbstractMatrix{FT}) branch is checking that the user supplied consistent arguments (matrix ensemble → vector mean), not that two arrays have matching sizes. Use ArgumentError, not DimensionMismatch, for this pattern.

Distinguish API boundary sites (where the user passed something wrong — prefer typed exceptions with actionable messages) from internal invariant sites (where a bug in the package itself would have to exist — bare error(...) with a clear note is fine there).

Loop-body errors: if the throw is inside a for or while loop, treat the loop index and key per-iteration state as required context. Without this, the user sees "matrix is not positive definite" with no idea whether it happened on iteration 2 or iteration 200. Always capture i (or the loop variable) and the state that changed between iterations — the ensemble step count, the parameter vector being updated, the ensemble member index, etc. For nested loops, include both the outer and inner loop variables — the outer variable says which group or batch failed; the inner variable says which element within it failed. See the loop-context example in the Canonical examples section below.

catch e losing the original exception: when a Julia exception is caught and a new one is thrown, the new message must include the original exception. If it does not, the user loses the root cause (e.g. PosDefException, SingularException) and has no way to distinguish a code bug from a numerical issue. Use sprint(showerror, e) rather than $e alone — it formats the exception type and message together:

# anti-pattern — root cause vanishes
catch e
    throw(ArgumentError("Matrix factorization failed."))
end

# correct — root cause preserved
catch e
    throw(ArgumentError("""
Matrix factorization failed.

Caused by: $(sprint(showerror, e))

Suggestion:
    ...
"""))
end

Only suppress the original exception if it is a well-known internal Julia error (e.g. SingularException) and you are intentionally providing a higher-level fallback — and even then, log it at @debug level.

@warn string(...) concatenation: @warn string("...", x, "...") is the warning-side equivalent of error(string(...)) — it's noisy, hard to read, and doesn't benefit from Julia's interpolation. Rewrite as @warn "... $x ...". @warn messages that use structured strings are also easier to grep and suppress selectively.

Double-gated invariants: if a helper is only ever called after the public API has already checked the same condition (e.g., get_vector_of_parameterized is called from construct_prior only when d.args[1] == Symbol("VectorOfParameterized") is true), the check inside the helper is an internal invariant even though it looks like a user-data check. Use a single-line error(...) rather than a full structured ArgumentError:

# internal invariant — the caller already validated this
d.args[1] == Symbol("VectorOfParameterized") || error(
    "Internal error: get_vector_of_parameterized called with non-VectorOfParameterized expression (got $(d.args[1]))",
)

Step 2.5 — Decide: inline or helper?

Before writing the rewrite, decide whether the error belongs inline or should be extracted into a _throw_<what>(...) helper function.

When to extract — pull the error into a helper when either condition holds:

• Length (primary trigger): the message body exceeds 3 lines. Extract unconditionally — single call site, non-loop context, no surrounding complexity required. A full Expected / Got / Suggestion block always crosses this threshold. Even a one-off long block left inline establishes a pattern that makes entire files hard to scan, and accumulates quickly once a few exceptions are made.
• Duplication: the same error shape (same summary line, same Expected / Got / Suggestion skeleton) appears at ≥2 call sites. Extract even when each block is short — the wording drifts silently over time and the call sites collapse to readable one-liners.

Inline is appropriate only for genuinely short messages (≤3 lines) at a single call site. A single summary line, or a summary plus one Got line, is the ceiling for inline. When in doubt, count — if it doesn't fit in 3 lines, extract.

Where helpers go

Default: a ## Error helpers section at the bottom of the source file, above end # module. Keeping helpers near their callers preserves traceability — the reader sees the throw site, jumps to the bottom of the same file, and finds the message without switching files.

Promote to a shared src/ErrorMessages.jl (or the repo's equivalent top-level utility file) only when ≥2 different source files call the same helper. Discover which file to use by reading the top-level module file (e.g. src/PackageName.jl) for its include(...) list — then add include("ErrorMessages.jl") as the first include so every subsequent file sees the helpers without any using/import.

Naming convention

_throw_<what>(positional_required_facts...; kwargs_for_optional_context...)

• Underscore prefix → unexported private helper.
• Verb prefix _throw_ → the function unconditionally raises; callers know there is no return value.
• Suffix describes the failure mode: _dim_mismatch, _missing_keys, _bad_obs_type, _not_iterable.

Signature convention

Pass the facts that are always present as positional arguments (the offending value, the expected vs got summary). Pass optional context as keyword arguments with nothing defaults — especially loop context (index, total, iter, phase). Build optional sections inside the helper by checking isnothing(...). This keeps call sites compact and lets the same helper serve both loop and non-loop contexts (see the Helper with optional loop context canonical example).

Performance: use @noinline

Prefix every helper with @noinline. This prevents Julia from inlining the cold error path into the surrounding hot code, keeping numerical kernels unaffected:

@noinline function _throw_x_not_iterable(x; where::Symbol)
    throw(ArgumentError(...))
end

What NOT to do

• Don't create a catch-all _throw_arg_error(msg::String) — that just shifts the inline triple-quoted block to another file without any DRY benefit.
• Don't use macros (@check_dim(...)) — they're magical and harder to debug than plain functions.
• Don't bundle all context into one opaque context::NamedTuple — explicit kwargs are clearer to call and easier to extend.

Step 3 — Rewrite with the canonical layout

Use this structure for every user-facing exception:

throw(ArgumentError("""
Short one-line summary of the failure.

Expected:
    <what would have been valid>

Got:
    <what was actually received, with interpolated values>

Loop context:
    iteration  = $iter (of $n_iter)
    <key per-iteration state variable> = $(summary_of_state)

Context:
    <surrounding state that helps locate the problem>

Suggestion:
    <most likely fix>
"""))

Section rules:

• Summary: always present; one line; imperative or declarative.
• Expected / Got: strongly preferred for any mismatch check; use $(expr) interpolation to show actual values.
• Loop context: include whenever the throw is inside a for or while loop. Always report the loop index and the key state that varies between iterations (e.g., the EKI step number, the ensemble member index, or the parameter being updated). This is what lets the user reproduce the failure without adding println debugging. Omit for errors that can only fire at a fixed point in the code (before the loop starts or after it ends).
• Context: include when the same error can arise from multiple call sites and naming the calling function or struct helps the user orient.
• Suggestion: include whenever a likely fix exists. Omit rather than write a generic platitude.
• Never dump full matrices or large arrays. Prefer size(x), eltype(x), typeof(x), or a scalar summary statistic.

Step 4 — Move validation early

If the current code lets an invalid input reach a numerical routine before failing (e.g., cholesky on a non-symmetric matrix, inv on a singular one), add an explicit guard at the API boundary:

# Before: error surfaces deep in cholesky
cov_chol = cholesky(C)

# After: check at the boundary, raise immediately
issymmetric(C) || throw(ArgumentError("""
Covariance matrix must be symmetric.

Got:
    size(C) = $(size(C))
    norm(C - C') = $(norm(C - C'))

Suggestion:
    Pass a symmetric matrix, e.g. `C = (C + C') / 2`.
"""))
cov_chol = cholesky(C)

Use || for single-condition guards. For multi-condition guards, use if/throw.

When using || with a multiline triple-quoted throw, the closing )) goes on its own line immediately after the closing """:

condition || throw(ArgumentError("""
Summary line.

Expected:
    ...

Got:
    ...
"""))   # ← closing )) on the line right after the closing """

This is the only layout that keeps indentation correct — triple-quoted strings in Julia do not strip leading whitespace, so indenting the message body would include those spaces in the string.

Step 5 — Preserve domain language

Write messages in terms the user understands, not in terms of internal Julia dispatch or linear algebra internals. For example:

• Say "ensemble member count" not "size(x, 2)"
• Say "parameter covariance matrix" not "the second argument to cholesky"
• Say "observation noise covariance" not "Γ"

Step 6 — Apply rewrites

Edit each site, keeping the surrounding code untouched. Confirm the package still loads:

julia --project -e 'using EnsembleKalmanProcesses'

Step 7 — Add @test_throws tests

Before writing any test, check whether coverage already exists. Grep the matching test/<module>/runtests.jl for the public API function that reaches the rewritten site:

grep -n '@test_throws' test/<module>/runtests.jl | grep '<function_name>'

Three outcomes:

For every site that needs a new test, add it in the matching test/<module>/runtests.jl:

@test_throws ArgumentError multiplicative_inflation!(ekp; s = 2.0)

Use the specific exception type — never bare @test_throws Exception. The test should construct the minimal invalid input that triggers the new error, without duplicating happy-path coverage.

Update existing tests that used the wrong type. If the file already has a @test_throws ErrorException (or any other type) for a site you're rewriting to ArgumentError, update that existing test in the same edit. Leaving a stale @test_throws ErrorException will cause it to pass against the old code but fail once your rewrite lands — or vice versa.

Testing unexported helpers. If the site is inside an unexported helper (e.g., construct_constraint, construct_2d_array), do not import the internal directly. Instead, test through the nearest exported public API function that calls it, using invalid input that propagates to the helper:

# construct_constraint is unexported — test via get_parameter_distribution
no_constraint_dict = Dict("uq_param" => Dict("prior" => "Parameterized(Normal(0.0, 1.0))"))
@test_throws ArgumentError get_parameter_distribution(no_constraint_dict, "uq_param")

This keeps tests coupled to the public contract and avoids brittleness when internal function names change.