What's wrong with it / where's the bug / why is it slow
Primary mode: decision tables (where to look, what to flag, which tool) + BAD/GOOD pairs. Anti-pattern catalogs live in the other two skills; this skill cross-references them and frames each pattern as "flag if you see this" rather than "don't output this."
How to use this skill
Reviewing a PR / diff? — Read §1 (Rules), §3 (Workflow), §7 (Checklists). Scan the diff against each checklist.
Debugging a specific bug? — §4 (Workflow), §8 (Playbook — find the symptom).
Profiling performance? — §5 (Workflow), §9 (Playbook — pick the right tool), §10 (Common pitfalls).
Production ops / live tracing / distributed debugging under load — this skill covers the review side; production forensics is separate.
Subskills — deep inspection references
The core SKILL.md carries the always-loaded rules, severity guidance, workflow, checklists, and a tight playbook. For extended depth on a specific area, load the matching subskill:
Cross-skill references: for WHAT to write (implementation) → elixir-implementing. For WHY it's shaped that way (architecture) → elixir-planning.
1. Rules for Reviewing, Debugging, and Profiling Elixir (LLM)
ALWAYS separate severity from correctness. A review finds many things; not all are worth blocking on. Use §6 to classify each finding as block / request-change / suggest / nitpick.
ALWAYS start debugging with the smallest-scope tool — IO.inspect or dbg before reaching for tracing, profiling, or Observer. Most bugs are found by inspecting values at suspect points.
ALWAYS start profiling by measuring, not guessing. Use :timer.tc for one-off timing, Benchee for comparisons, mix profile.* for finding the slow function. Never "optimize" without evidence.
ALWAYS use System.monotonic_time for duration measurement — never System.system_time (NTP sync causes jumps).
NEVER use :dbg or :erlang.trace in production — no safety limits, can crash nodes under load. Use :recon_trace or Rexbug with explicit message count limits.
ALWAYS read the symptom before reading the code. What does "slow" mean — p50, p99, tail? What does "crash" mean — which error, which stack frame, how often? A vague symptom wastes time.
ALWAYS suggest the idiomatic refactor when flagging an anti-pattern. "This is bad" without "do this instead" is low-value feedback.
ALWAYS cross-reference elixir-implementing and elixir-planning when flagging a finding — point reviewers to the section that explains why the idiomatic form is better.
PREFER letting the supervisor restart a crashing process over adding defensive rescue clauses. Before suggesting new error handling, ask whether the existing supervision tree already handles it.
NEVER flag style issues that mix format or mix credo would catch. Reviews are for things tools miss: architecture, idioms, subtle correctness, testability, performance.
ALWAYS flag missing tests and @spec on public functions — these are easy to add, compound in value over time, and are the strongest signal that a change was thought through.
PREFER benchmark-driven refactor suggestions over intuition. "This could be faster" is weak; "this is 40% slower in the linked Benchee run" is actionable.
ALWAYS check the composition shape on diffs that pipe ok/error or thread state. Before flagging a with-chain, identify which composition mechanism the operation NEEDS: pipeline (transform), railway (with, sequential dependency), accumulating reduce (independent validations), threading-builder (Multi/Conn/Socket), protocol/behaviour (dispatch). The wrong shape is the most common architectural smell on diffs that "just look long" — and it's invisible without naming the shape. Cross-reference: elixir-implementing §5.10 (composition patterns), elixir-planning §4.7 (composition design vocabulary).
ALWAYS check building-block axes when reviewing a module marked @moduledoc "Building block ..." or claimed pure. Walk the seven-axis checklist (input closure, determinism, @spec, totality, side-effect freedom, errors-as-values, input guard) for every public function. A building-block module that calls Logger, Repo, Phoenix.PubSub, :telemetry.execute, Application.get_env, DateTime.utc_now, or :rand.uniform inside a function body fails axis 5 (effect-free) or axis 1 (input-closure) and is silently lying. Cross-reference: elixir-planning/building-blocks.md (the seven-axis checklist), Archdo CE-54..57 (BlackboxQuadrant, UntestedBuildingBlock, EffectLeak, UnguardedBuildingBlock).
2. The Three Modes of Inspection
The three modes of this skill are not interchangeable. Each has a different question, different tools, different output.
Mode
Question
Primary artifact
Primary output
Review
"What's wrong with this diff?"
PR / branch diff
Comments with severity + suggested fix
Debug
"Why is this bug happening?"
Failing test, crash report, observed misbehavior
Root cause + the minimum fix
Profile
"Why is this slow / heavy?"
Production telemetry, benchmark, complaint
Identified bottleneck + measured improvement
They share one method: read existing code, form hypotheses about what is wrong, verify with evidence (static — read the code; dynamic — run it with inspection). They differ in what "wrong" means and what "evidence" means.
Choose the right mode:
Pre-merge, proactive → Review
Post-merge, something's broken → Debug
Post-merge, something's slow → Profile
Mixing modes wastes time: full review on a bug report, ad-hoc debugging on a PR diff, micro-optimization without measurement.
3. Review Workflow (PR / diff review)
3.0 Layered review — run mechanical tools first
A human review should only contain findings a tool can't produce. Before opening the diff, run (or confirm CI ran) the full mechanical stack:
Zone-of-Pain / Martin metrics, dead code, cyclic dependencies
Out of human-review scope — delegate
Golden rule: if a tool flags it, don't re-flag it. Your review capital is scarce; spend it on architecture, idiom, subtle correctness, and testability that the tools CAN'T see.
Every review finding should answer: "why wouldn't a linter catch this?" If the answer is "it could", delete the finding.
3.1 Step-by-step
Read the PR description. What is the stated intent? Does the diff match?
Read the tests first. Tests describe behavior; code describes implementation. If the tests are missing or weak, that's the first finding.
Scan the diff for architectural smells (§7.1) — wrong layer, cross-context Repo calls, framework references from domain. These are block-severity by default.
Scan for correctness issues (§7.2–§7.5) — control flow, pipelines, error handling, pattern matching.
Scan for OTP/process issues (§7.6) — unsupervised work, blocking callbacks, call vs cast, state shape.
Scan for security issues (§7.8) — String.to_atom on user input, SQL injection via interpolation, leaked secrets in logs, unvalidated external data.
Scan for testing gaps (§7.7) — missing tests, tests that check implementation details, Process.sleep for async behavior.
Scan for documentation gaps — missing @spec / @doc / @moduledoc on public API.
Scan for performance (§7.9, §10) — obvious quadratic patterns, N+1 queries, GenServer bottlenecks, string concat in loops.
Style inconsistency mix format / credo would catch
Do NOT flag. Let the tools handle it
mix credo finding already present
Do NOT flag in the PR — raise it separately or add to baseline
Taste preference with no evidence of harm
Do NOT flag — save capital for real findings
Security issue
Flag. Severity: block
Performance issue with no measurement
Flag as question: "have you measured this?" — do not block on speculation
Performance issue with measurement
Flag. Severity depends on impact
3.3 What to check in every review
Tests exist for the new behavior, and they pass
Tests assert behavior, not internal calls
@spec and @doc on every new or modified public function
No @moduledoc missing on public modules (@moduledoc false is fine for internal)
No cross-context Repo calls (see §7.1)
Cross-context struct destructure (%OtherContext.SomeStruct{field: f} = value in a caller's function head, OR value.internal_field reaching into another context's data) — flag unless that struct is documented as a public data structure (Plug.Conn / Date / Ecto.Changeset style). If it's an "opaque handle" (MapSet / Ecto.Query style), the caller should use accessor functions or operations, not pattern-match the fields. See elixir-planning architecture-patterns.md §4.12. Severity: request-change if the struct's owning context has documented fields as private; suggest if undocumented (and propose deciding the opacity policy).
No String.to_atom(user_input) (see §7.8)
No try/rescue for expected failures (see §7.4)
No Process.sleep in tests (see §7.7)
Idempotent for retried operations (Oban workers, webhook handlers — see §7.6)
Timeouts set where appropriate (see §7.9)
Composition shape matches operation intent — with for sequential dependency, accumulating reduce for independent validations, threading-builder for state accumulation, pipeline for transforms. Wrong shape on a 5+ step chain is the #1 architectural smell on diffs that "just look long" (see §7.12)
Building-block modules are pure — modules whose @moduledoc claims building-block status (or that the diff intends to be one) pass the seven-axis checklist: no Logger/Repo/PubSub/telemetry/ets/persistent_term writes inside function bodies; no hidden inputs (Application.get_env/DateTime.utc_now/:rand.uniform); no raises (use ok/error); narrow input domain (see §7.13)
Subject-first arg order on public functions — first arg is the data being transformed, options last. Backwards (def fun(opts, data)) breaks pipe composition (see §7.12)
LiveView handle_event/3 does not block — synchronous HTTP / heavy computation freezes the LV process; wrap in start_async/3 + handle_async/3 (see §7.14)
4. Debugging Workflow (finding a specific bug)
4.1 Step-by-step
Reproduce the bug — a bug you can't reproduce can't be fixed with confidence. If you can't reproduce, the first goal is to find a reproduction. Try: exact inputs from the report, similar inputs, boundary conditions, concurrent access.
Write the failing test (TDD-for-bugs) — the reproduction IS a regression test. See elixir-implementing §3.6.
Read the error — full stack trace, module, function, line. Is it an exception (raise), exit signal (:exit), crash (supervisor restart)?
Form a hypothesis about which code path produces the bug.
Verify with evidence — add IO.inspect/dbg at the hypothesized point. Run. Is the value what you expected?
If the hypothesis is wrong, widen the inspection outward. Walk backwards through the call chain.
For process-level bugs (timeouts, mailbox buildup, crashes), use :sys.trace, :observer, or Process.info.
For concurrent/flaky bugs, check test async isolation, shared state, timing assumptions.
Once found, consider whether other places have the same bug. A single bug often has family members.
Commit the failing test alongside the fix so the bug is guarded against regression.
4.2 Debug decision — which tool
Question
Tool
Load for detail
What is this value at this point?
IO.inspect/2 or dbg/1
—
What does this pipeline produce at each step?
dbg() at the end of the pipeline
—
Can I pause execution and inspect?
IEx.pry/0 (needs iex -S mix)
—
What messages is this GenServer receiving?
:sys.trace(pid, true)
—
What is this process's state right now?
:sys.get_state(pid)
—
How long is this process's mailbox?
Process.info(pid, :message_queue_len)
—
What's the memory / CPU use of each process?
:observer.start() (GUI) or :recon.proc_count/2
—
What pattern of calls is hitting this module in prod?
:recon_trace.calls(...) with a message limit
—
Why does this test fail intermittently?
See §8.5 (flaky tests)
—
4.3 Start small, escalate only when needed
First reach: IO.inspect / dbg. Most bugs found here.
Next: IEx.pry or break! to pause execution.
Next: :sys.trace or :sys.get_state for OTP processes.
Next: :observer for system-wide view.
Last resort: :recon_trace or Rexbug for production tracing (with message limits!).
NEVER skip straight to tracing/profiling when a single IO.inspect would show the bug.
5. Profiling Workflow (finding a performance problem)
5.1 Step-by-step
Define what "slow" means. p50, p95, p99? Per-request? Whole-pipeline? Steady-state or cold-start? A vague "slow" wastes time.
Measure before changing anything. Set a baseline: Benchee, :timer.tc, :telemetry histogram. Without a baseline you cannot prove you made it faster.
Profile to find the bottleneck — which function / process is actually slow? Don't guess.
Form a hypothesis — why is this slow? Common causes in §10.
Apply one change at a time. Measure after each. If the change didn't help, revert it.
Verify the improvement at the same level the symptom was observed (e.g., if users complained about request latency, measure request latency — not just the inner function).
Keep the measurement in CI as a regression guard, or at minimum document the baseline and improvement.
5.2 Profile decision — which tool
Need
Tool
Notes
One-off duration of a block
:timer.tc(fn -> ... end)
Returns {microseconds, result}. No warmup — misleading alone
Compare two implementations
Benchee.run(%{...})
Warmup, statistics, memory
Find which function in a call tree is slow
mix profile.fprof
High overhead, detailed — use for dev, not prod
Time spent per function
mix profile.eprof
Moderate overhead — profile tests or dev requests
Call counts only
mix profile.cprof
Low overhead — which functions are called most
Unified modern profiler
mix profile.tprof
OTP 27+, lower overhead — prefer for large code bases
Always warm up before measuring. The BEAM JIT (OTP 24+) needs several runs to reach steady state. Benchee handles this; :timer.tc does not.
Always measure at the right granularity. Micro-benchmarking a function that's called once per hour is wasted effort. Macro-benchmark the request / job / pipeline first, then zoom in.
Never optimize without profiling first. You'll optimize the wrong thing.
Beware the heisenberg effect. Heavy profilers (fprof, especially under tracing) change timing characteristics — microbenchmark results may not match production.
Production profiling must be bounded. Message-limit every trace. A runaway trace can exhaust a node.
6. Severity Classification
Every review finding needs a severity. Without it, the reviewer defaults to "everything is equally important" — which devalues every comment.
length(list) > 0; if/else for structural dispatch; try/rescue for ok/error flow; test asserting mock was called
Suggest (nice to have, author's call)
Refactor that would improve readability / maintainability
Extract a helper; prefer for comprehension over reduce; use Map.new/2 instead of Enum.reduce into %{}
Nitpick (take-it-or-leave-it)
Style, naming, comment wording — things that don't affect behavior
Prefer current_user to user; this docstring could be clearer
Question (need info)
Clarification needed before you can assess
"Have you measured this?" / "Why not use existing X?"
Rules:
Each block-severity finding must be fixed before merge.
Prefer suggest over request-change when you're unsure. Request-change is a stronger claim.
Explicitly label nitpicks as "(nit)" so the author can safely ignore them.
Don't block on taste. Block on facts.
Don't block on things the author didn't touch. File a separate issue.
Severity by PR size:
Small PR (<100 LoC) — pick your battles; probably only block-severity comments + a handful of suggests
Medium PR (100–500 LoC) — full review; classify everything
Large PR (500+ LoC) — ask for a split first. Reviewing large PRs well is hard; reviewing them badly is worse than not reviewing them
7. Review Checklists — What to Flag
Depth: For the consolidated catalog of named anti-patterns organized by category (code / process / Ecto / architecture / testing / security / config), each with a BAD/GOOD pair — load anti-patterns-catalog.md. Use that when you spot something "off" and want to name it; use this §7 when you want to scan a diff systematically by area.
Each subsection is a scanning checklist. Read left to right: "if you see this" → "flag it because" → "suggest this" → "severity". Links to the owning skill sections explain why for the author.
7.1 Architectural review
Full reference: elixir-planning §14. Flag these if you see them in a diff.
If you see...
Suggest instead
Severity
Why — see
Directories like lib/my_app/models/, services/, helpers/
Business logic in a controller action, LiveView handler, or CLI handler
Move to a context function; interface translates + delegates + formats
Block
planning §14.3
Plug.Router route block calls another Plug module's call/2 directly with raw opts (e.g. get "/", do: Handler.call(conn, []))
Pre-initialize at compile time: @handler_opts Handler.init([]) and get "/", do: Handler.call(conn, @handler_opts). Bypassing init/1 drops opt normalization — a latent trap that breaks silently the moment init/1 stops being a no-op
One owns it; the other reads through owner's API or uses PubSub
Block
planning §14.11
Repo.preload(:other_context_association) across contexts
Ask the owning context for assembled data
Request-change
planning §14.12
One context's internal module called from another context
Go through the owning context's public API
Request-change
planning §6.4
defdelegate pass-through in a context
Fine if pure pass-through; flag if the context should add telemetry/logging
Nitpick
planning §6.7
Business logic in a GenServer handle_call
Extract to pure module; GenServer delegates
Request-change
planning §14.6
Feature being added to a growing "god context"
Does it belong? Consider split criteria
Suggest
planning §6.2
Introducing an umbrella split for "it feels big"
Keep single-app; add contexts
Request-change
planning §14.7
Adding a new inter-context call via PubSub before trying direct function calls
Can direct calls work? Escalate only when justified
Suggest
planning §9.9
Adding Oban / GenStage / event sourcing without a triggering problem
Use the simplest mechanism; escalate only when needed
Request-change
planning §9.9
Missing @moduledoc on a public module
Add @moduledoc or explicit @moduledoc false
Request-change
implementing §8.5
Missing @spec on a new public function
Add @spec
Request-change
implementing §6.10
@impl true implementation with no explicit @spec
Add @spec on the implementation too — @impl links to the behaviour spec but doesn't substitute
Suggest
implementing type-and-docs rule 1
Behaviour callback overloaded with a "reflection" atom (e.g. execute(:list_instructions, a, b))
Give reflection its own callback (instructions/0, describe/0)
Request-change
planning §4.9
Union type variant with {:tag, nil} payload sentinel
Use bare atom: :tag — don't carry a nil payload
Nitpick
implementing type-and-docs §Union types
Public @spec uses a loose type (atom(), map(), [term()]) where a named @type is already defined in scope
Reuse the named alias: MyMod.instruction(), MyMod.t(), [MyMod.entry()]
Suggest
implementing type-and-docs rule 7
@moduledoc / @doc asserts a behaviour (binds to both X and Y, rejects Z, accepts ranges A..B) not exercised by a test
Add a test that pins the claim, OR update the doc to match the code. Stale docs mislead worse than missing docs do
Request-change
implementing type-and-docs rule 13
Two moduledocs that describe the same subsystem disagree (e.g., App.Application says "binds IPv4 only" while Plugs.RequireLoopback says "binds IPv4 + IPv6")
Reconcile both moduledocs against the code as one atomic change. Cross-file contradictions are Rule 13 distributed — harder to spot, same defect
Block
implementing type-and-docs rule 14
Plug's init/1 / call/2 uses Plug.opts() when the plug actually accepts specific keys
Define a narrow @type opts :: [...] and use it in both specs
Suggest
implementing type-and-docs §Plug signature
New feature is a library candidate but uses Application.compile_env
For library code, use runtime get_env or config-via-args
Block (if library)
planning §10.3
Public context function's moduledoc claims atomicity ("both X and Y, or neither") across a DB write AND a non-DB side effect (process-registration, PubSub broadcast, external API call), but the code just does them sequentially and doesn't roll back
Wrap both in Ecto.Multi.run/3 so the DB row rolls back on side-effect failure, OR remove the atomicity claim and document the recoverable orphan state. Either way: the moduledoc must match the code — this is Rule 13 applied to a specific anti-pattern
Request-change
implementing type-and-docs rule 13
Moduledoc references a future milestone ("M9 will add Broadway" / "in M12 this becomes atomic") after that milestone has already merged
Rewrite the claim in the present tense describing current behaviour. Milestone-references belong in commit messages and a plan doc, not in each module's @moduledoc where they silently decay into lies
Suggest during a rollout; Request-change three milestones after the referenced one merged
implementing type-and-docs rule 13
LiveView handle_info handler for a PubSub message reloads the whole collection from the DB to derive a counter / display state that could be updated from the message content + existing assigns
Change the broadcast payload to carry everything the subscriber needs (e.g., {:device_state_changed, id, old, new} not {:device_state_changed, id, new}) so the LV can update assigns in O(1) instead of re-querying. Never pay N queries per message
LiveView stream with a row template that reads from a separate assign (@device_states, @cluster_status, @presence), and a handle_info that changes that sidecar assign to update the row
Stream row DOM is only re-emitted on stream_insert/3,4 — the sidecar change does NOT redraw the row. Move the state onto the stream member (virtual schema field, list_with_state/1 context fn) and re-insert on every change. This is architectural: may require a schema change. See implementing §5.9, planning §14.14
Block
phoenix-liveview + implementing §5.9
Repo.insert_all(schema_or_"table", rows) where a row field is the high-level form of a type that needs dumping: Ecto.UUID.generate/0 → 36-char hex instead of 16-byte binary; atom for Ecto.Enum column; raw DateTime for :utc_datetime_usec without truncation; custom Ecto.Type without dump/1
insert_all bypasses casts by design. Pass the raw DB representation: Ecto.UUID.bingenerate/0, Atom.to_string/1 for enum columns, DateTime.truncate(dt, :microsecond), MyType.dump!(value). Or use the schema-module form (Repo.insert_all(Schema, rows)) so Ecto dumps known fields
Block (silent Postgrex encode errors at runtime)
implementing §10.6.1
from b in "string_source", ... (schemaless query) selecting a :utc_datetime* column, and downstream code treats the result as DateTime.t()
String-source queries return NaiveDateTime (no schema metadata to lift to UTC). Either use a schema-bound query, or normalize at the context boundary: DateTime.from_naive!(n, "Etc/UTC"). Downstream DateTime.diff/3 etc. will crash on NaiveDateTime
Request-change
implementing §10.6.2
GenServer with boot-time subscription via :net_kernel.monitor_nodes/1,2, Phoenix.Tracker, Phoenix.PubSub.subscribe/2, or Process.monitor/1 against a long-lived target, without replaying current state in handle_continue/2
Subscriptions only deliver future events. Existing peers / rows / children are invisible. Use {:ok, state, {:continue, {:retro_scan, Node.list()}}} and fire the same handler. Pattern applies across :net_kernel.monitor_nodes, Process.monitor on DynamicSupervisor children, Phoenix.Presence on an already-populated topic
Request-change if the GenServer drives business logic from these events; Suggest if telemetry-only
§8.9 retro-scan template
Event or command struct (under Events/Commands namespace, defimpl Commanded.Event.Handler, or persisted via Oban / EventStore) without a :version/:schema_version/:event_version field, no @version attribute, AND no defimpl Commanded.Event.Upcaster for the type
Pick ONE convention per project: inline :version field (small projects) OR defimpl Commanded.Event.Upcaster, for: MyEvent (Commanded idiom — one module per event type, upcaster transforms older persisted shapes). The "module-per-version" pattern (V1.OrderPlaced / V2.OrderPlaced separate modules) is NOT what Commanded recommends. See elixir-implementing §5.16.
Request-change
implementing §5.16, §1 #36
Event/command split into separate V1.<Name> / V2.<Name> modules with an upcaster bridging them
Consolidate into a single event module per type; let defimpl Commanded.Event.Upcaster, for: <Name> fill in defaults for the missing fields when older persisted shapes are read. The single-module shape is what Commanded's docs prescribe — quoted from Commanded.Event.Upcaster: "Upcaster changes any historical event to the latest version, consumers... only need to support the latest version."
Suggest (refactor toward Commanded idiom)
implementing §5.16
7.2 Control flow review
Full reference: elixir-implementing §7.1. Flag these patterns when you see them.
If you see...
Suggest instead
Severity
if is_map(x) and Map.has_key?(x, :type) or similar shape-discriminating if
Multi-clause function with pattern matching
Request-change
if opts[:flag] do ... else ... end with value-returning branches
case Keyword.get(opts, :flag, false) do true -> ...; false -> ... end
Suggest
Nested case (2+ levels) on ok/error results
with chain
Request-change
with containing a single clause
Plain case
Nitpick
if user != nil do ... if user.name != nil ... (nil-check cascades)
Multi-clause on shape: def greet(%{name: name}) when is_binary(name), do: ...; def greet(nil), do: ...
Request-change
cond without a true -> default branch
Add explicit default; otherwise CondClauseError on fall-through
Split into distinct tags: {:wrong_type, v} + {:out_of_range, v, range}
Request-change
Public module mixes raise and {:error, _} for same failure class (e.g. describe/1 raises but execute/4 returns {:error, _})
Pick one per boundary. Safe name returns ok/error; ! variant raises
Request-change
with chain validates data inputs before dispatch key, forcing dummy data for reflection calls
Validate dispatch key first; reflection paths then bypass data validation
Suggest
with chain mixes sentinel-valued short-circuits (:noop, :skip) and {:error, _} tuples, with an else clause :noop -> :noop; other -> other passing everything else through untyped
Give :noop / :skip explicit else clauses separately from {:error, _}=e -> e, OR split into two functions: a "should-we-run" predicate on top, a with chain only for actual work. Current shape lets future {:error, _} returns leak unreviewed into callers
Suggest
7.6 Process / OTP review
Full reference: elixir-implementing §9, elixir-planning §8.
If you see...
Suggest instead
Severity
spawn(...) or spawn_link(...) for long-running work
Supervised Task (Task.Supervisor.start_child/2) or GenServer
Business logic in handle_call/handle_cast/handle_info
Extract to pure module; callback delegates
Request-change
GenServer.call(pid, :msg) without explicit timeout
Pass timeout arg explicitly (default 5000ms is often wrong)
Request-change
GenServer.call to a variable pid without catch :exit
Wrap with try ... catch :exit, _ -> {:error, :down} — the process may die mid-call
Request-change
Hardcoded timeout literal like GenServer.call(pid, msg, 30_000)
Extract to @default_X_timeout module attribute
Suggest
Registry + DynamicSupervisor under :one_for_one
Dedicated :one_for_all sub-supervisor so both restart together when Registry dies
Block
GenServer with large state (> ~100KB)
Move large data to ETS / :persistent_term
Request-change
GenServer just wrapping a map (get/put)
Use ETS directly (no serialization bottleneck)
Suggest
Registry-registered process called as GenServer.call(:name, msg)
Use {:via, Registry, {Reg, name}} or helper via/1
Block
Unsupervised Task.async(fn -> ... end) |> Task.await/2 in a GenServer callback
Task.Supervisor.async_nolink/3 + handle :DOWN
Request-change
Agent-per-entity (simulating objects)
Pure functional modules for thought concerns; processes only for runtime concerns
Request-change
Missing catch-all clause in handle_info
Add def handle_info(msg, state) do Logger.warning(...); {:noreply, state} end
Request-change
Missing @impl on behaviour callback
Add @impl true (or @impl BehaviourModule)
Request-change
Raw %{} / Map.put on a struct (silently accepts typos)
Struct update %{struct | field: val}
Request-change
No format_status/1 on GenServer holding secrets
Implement format_status/1 to scrub tokens/passwords
Request-change
GenServer calls :net_kernel.monitor_nodes(true) in init/1 without replaying Node.list() in handle_continue/2
Node connections that formed before this GenServer started are invisible — monitor_nodes only delivers future events. Retro-scan pattern: {:ok, state, {:continue, {:retro_nodeup, Node.list()}}} then fire the same handler. Same pattern applies to Process.monitor/1 against a pre-existing DynamicSupervisor's children
Request-change if the GenServer drives business logic from nodeup; Suggest if audit-only
Synchronous call into another OTP app's API inside init/1 (e.g. :fuse.install/2, shared :ets.new/2, :persistent_term.put/2 for boot config)
Defer to handle_continue/2. Listing the dep in extra_applications gives a partial order within the current app but not across apps — a GenServer in app A doing B.init_table/1 in its own init/1 can race with app B's own boot
Request-change
Async closure (Task.async, Task.Supervisor.start_child, Task.async_stream, Task.Supervisor.async_nolink) calls Logger.<level> or :telemetry.execute/3 inside the closure body, but does NOT call Logger.metadata/0,1 to restore parent metadata (and does not receive an explicit context struct as an argument)
Capture parent metadata before the closure (md = Logger.metadata()), restore inside (Logger.metadata(md)); for cross-Oban / cross-node hops, pass an explicit TraceContext struct. See elixir-implementing §5.13 / §7.12 and elixir-planning §11.7.
Request-change
Oban worker's perform/1 calls Logger.<level> or :telemetry.execute/3 without Logger.metadata/1 set first from the job's args
Worker hydrates metadata from args at the top of perform/1: Logger.metadata(trace_id: args["trace_id"], order_id: args["order_id"]). Producer side (the Oban.insert/2 caller) writes the trace IDs into args from its own Logger.metadata().
Request-change
:telemetry.execute/3 from inside an async closure where the metadata map does NOT include trace_id (or the project's correlation key)
Add the correlation ID to the telemetry metadata explicitly. Telemetry handlers run in the emitting process; cross-process re-publish (e.g. to a remote sink) does not preserve the emitter's Logger.metadata.
Request-change
7.7 Testing review
Full reference: elixir-implementing §3–§4.
If you see...
Suggest instead
Severity
New public function without a test
Add tests — block until done
Block
Test using Process.sleep before asserting on async behavior
assert_receive pattern, timeout
Request-change
Test asserting a specific internal function was called
Test observable behavior, not call sequence
Request-change
use MyApp.DataCase with async: false for no clear reason
Use async: true — Ecto sandbox supports it
Suggest
Test calls handle_call directly
Test via the client API
Request-change
Mox.stub where expect is appropriate (must-be-called behavior)
Use expect — stub gives no verification
Request-change
Mocking a module the project owns (Accounts, Pricing)
Test directly; mock only system boundaries
Block
No setup :verify_on_exit! with Mox
Add setup :verify_on_exit! in test module
Request-change
Test using assert result == {:ok, ...} (stringified)
assert {:ok, %User{...}} = result (pattern match, better failure)
Suggest
Factory creating hardcoded unique values
Use sequence(:email, &"user-#{&1}@x.com")
Request-change
Test with 10+ lines of setup for a pure function test
The function needs a boundary — pure tests shouldn't need this much setup
Suggest (architectural)
Test that "sometimes" fails
Never leave flaky tests — find the root cause
Block
Parametrized test loop using @attr bad rebinding to smuggle the loop variable into each generated test
Use a single test with for loop inside the body (rich assertion messages), or unquote(Macro.escape(bad)) in the test body (which IS inside a quote)
Suggest
Canonicalizer / validator / parser function with only example-based tests
Add StreamData property tests. Tables miss adversarial edge cases — port suffixes, case variants, IPv4-mapped IPv6, trailing whitespace, Unicode homoglyphs
Suggest
Duplicated fixture data / lookup maps across multiple tests in the same file (e.g., same valid_pairs = %{...} copied into two tests)
Extract to a @module_attribute or a setup callback that returns it in the context
Suggest
7.8 Security review
Depth: For the full security audit checklist (input validation/injection, authn/authz, logging, crypto, Phoenix/Ecto-specific, dependency & operational concerns), load security-audit-deep.md.
Commonly overlooked. Flag any of these.
If you see...
Suggest instead
Severity
String.to_atom(user_input) or Jason.decode!(json, keys: :atoms)
String.to_existing_atom/1, or decode to string keys and convert explicitly
Block (atom table exhaustion)
SQL built with string interpolation (raw query)
Parameterized query via Ecto.Query or Ecto.Adapters.SQL.query/3,4 with args
Block (SQL injection)
User-supplied data passed to :erlang.binary_to_term/1,2
Use :safe mode or prefer JSON / well-defined format
Block
Untrusted HTML interpolated into a template
Use LiveView's default safe rendering; only raw/1 with known-safe content
Block (XSS)
Secrets, tokens, passwords in logs or error messages
Filter via Logger.metadata, format_status/1, or redact explicitly
Block
Secrets in config/config.exs or config/dev.exs
Move to config/runtime.exs + env vars
Block
Secret committed in a migration / seed / test file
Validate path; use Path.expand/1 + check it stays within a safe root
Block (path traversal)
External HTTP call without timeout
Set receive_timeout; default may be infinite
Request-change
Endpoint without CSRF protection (non-API)
Phoenix's :protect_from_forgery plug
Block
Cookie without secure, http_only, same_site
Set them in endpoint config
Block (production)
No rate limiting on auth / password-reset / OTP endpoints
Add rate limiting (Plug, ex_rated, Hammer, or custom)
Request-change
Auth check in the controller, not in the context
Auth is cross-cutting — a plug or the context's public API should enforce it
Request-change
7.9 Performance review (scan level)
Flag obvious issues here. For measured performance investigation, use §9 (Profiling Playbook).
If you see...
Suggest instead
Severity
N+1 query pattern: Enum.map(xs, fn x -> Repo.get!(...) end)
Repo.preload/2,3 or batch query
Block
Enum.reduce(xs, "", &(&1 <> &2)) in a hot path
IO list pattern
Block (O(n²))
GenServer on the hot path serving reads
ETS :public + read_concurrency: true
Request-change
length/1 used for "is non-empty" checks
Pattern match [_ | _] vs [] (O(1))
Request-change
Enum.sort followed by Enum.take(1) or Enum.take(-1)
Enum.min, Enum.max, Enum.min_by, Enum.max_by
Request-change
Large data in Application.get_env on a hot path
:persistent_term
Suggest
Ecto.Repo per-request query building with identical queries
Consider compile-time query module or prepare: :named
Suggest
Repeated Jason.decode! of the same JSON in a loop
Decode once outside the loop
Request-change
Task.async + Task.await for bounded parallelism
Task.async_stream with max_concurrency
Suggest
Missing index on a queried foreign key
Add index in migration
Request-change (if queried often)
Streaming large files via File.read!
File.stream! + Stream.*
Request-change
Patterns worth request-change even WITHOUT measurement: some code shapes make the performance impact obvious from the diff alone — don't block on "have you measured this?" for these. Flag them as request-change with a brief why:
Application.get_env inside a tight loop or per-request hot path (read once at boot into :persistent_term / module attribute).
Enum.* on a large known-to-be-large list where only early results are consumed (Enum.take(..., k) without Stream.*).
Repo.get(...) inside Enum.map on a caller-supplied list (classic N+1).
Jason.decode! of the same JSON string in a loop (decode once outside).
For p95/p99 latency, memory-over-time, GC pressure — those still need measurement. But for algorithmic-shape issues visible in the diff, measurement is the author's responsibility at merge time, not the reviewer's gate.
7.10 Configuration review
Full reference: elixir-implementing §8.6, elixir-planning §10.
If you see...
Suggest instead
Severity
System.get_env(...) in config/config.exs
Move to config/runtime.exs
Request-change
Application.compile_env in a library
Runtime get_env or accept config via options
Block (if library)
Missing default in Application.get_env(:app, :key)
Provide default (get_env(:app, :key, default)), or use fetch_env! if required
Request-change
Config value read on every call (hot path)
Cache in module attribute (app), :persistent_term (library hot path), or Application.compile_env (app)
Suggest
Application code uses Application.get_env for a value that's truly frozen at compile time (no runtime.exs override, no test put_env)
Switch to Application.compile_env — Dialyzer sees the concrete type, missing-key crashes at compile, recompile triggers on config change. Don't blindly switch: if config/runtime.exs or any test overrides the key at runtime, compile_env silently freezes the default and breaks those flows. Verify both paths before changing
Suggest
config/runtime.exs parses an env var with String.to_integer/1, String.to_atom/1, etc. on raw input
Wrap with explicit validation: case Integer.parse(val) do {n, ""} when n in range -> n; _ -> raise "VAR_NAME must be X, got: #{inspect(val)}" end. A raw conversion exception at boot gives ops a stacktrace instead of a message
Request-change
runtime.exs splits a comma-separated env var and String.to_atom/1 on each element (NODEPULSE_NODES="a@h1,b@h2" → list of atoms)
Even for operator-controlled inputs this is unbounded atom creation on typos. Validate each element against a regex like ~r/^[a-z][\w]*@[\w\-.]+$/i BEFORE converting, OR cap the list size, OR use String.to_existing_atom/1 with a raise-on-unknown fallback. A CI pipeline accidentally generating a list of 10k node names can permanently exhaust the atom table
Block if from CI/untrusted; Request-change if strictly operator-controlled
Hardcoded URLs / credentials / secrets
Move to config + env var
Block
Test config imported into runtime code
Keep config/test.exs isolated; production should never import test config
Block
Application.get_env/fetch_env!/compile_env scattered across many modules (> ~3 files outside the config accessor module)
Centralize in MyApp.Config — every module routes config reads through zero-arg accessors. In an umbrella, one Config module per deployable. Grep Application.get_env|fetch_env|compile_env lib/ should only hit the Config module
Request-change
== comparison of a secret (API token, HMAC digest, verifier code, dev-path password) coming from an untrusted source
Plug.Crypto.secure_compare/2 — == is variable-time and leaks byte-position via timing. For cleartext password verification specifically, use the password lib's verifier (Bcrypt.verify_pass/2) — it's already constant-time
Block
7.11 Ecto migration review
Migrations are a distinct artifact class. The bugs are bigger (production data loss, broken rolling deploys) and the lint stack is thinner (no mix format or credo rule catches most migration issues). Review with an explicit checklist:
Reversible — has an explicit def up / def down, or def change that Ecto can auto-reverse. Irreversible migrations (e.g. execute("...") with no down version) must include a comment explaining why.
No schema module inside the migration — migrations must NOT alias MyApp.Schema. Schema modules evolve over time; the migration's semantics must stay pinned. Use raw SQL or a migration-local schema.
New foreign keys are indexed — add :thing_id, references(:things) without a matching create index(:table, [:thing_id]) is a deadlock waiting to happen (FK lookup table-scans).
NOT NULL columns have defaults OR use two-step backfill — adding add :field, :string, null: false to a populated table crashes the migration. Two-step pattern: (1) add nullable + backfill; (2) later migration adds NOT NULL.
Wrapped queue-library migrations are pinned — Oban.Migration.up(version: 12) (not up()) so future Oban versions don't silently add migrations when you re-run.
Precision-sensitive columns match the schema — :utc_datetime_usec column → schema declares timestamps(type: :utc_datetime_usec). Mismatches cause round-trip bugs that are invisible until production.
Concurrent index creation for large tables — create index(..., concurrently: true); ALTER TABLE locks long enough to stall writes otherwise. Requires @disable_ddl_transaction true + @disable_migration_lock true.
execute("...") with user-like strings is parameterized safely — migrations run in a privileged context; SQL injection via migration is rare but devastating. Prefer execute(&up/0, &down/0) with structured code over string interpolation.
Data backfill goes through a separate script or Oban job for large tables — in-migration Repo.update_all(...) on a 50M-row table will time out or lock production. Consider: migrate schema first, backfill in a follow-up job.
Rolling-deploy compatible — during deploy, old code runs against the new schema (and vice-versa). Columns can't be renamed in one step: (1) add new; (2) deploy code reading both; (3) backfill; (4) deploy code reading only new; (5) drop old.
Block-severity: missing down, NOT NULL without backfill, schema module aliased in the migration, unindexed FK. Request-change: precision mismatch, un-pinned Oban migration. Suggest: adding @moduledoc explaining the non-obvious parts.
7.12 Composition shape review
Full reference: elixir-implementing §5.10 (composition patterns), elixir-planning §4.7 (composition design vocabulary). The most common architectural smell on diffs is the wrong composition shape for the operation — not bad code, just the wrong primitive. Six mechanisms cover most production code: Pipeline (transform), Stream (lazy pipeline), Threading-builder (Multi/Conn/Socket — subject type fixed, state accumulates), Railway (with for sequential ok/error), Protocol/Behaviour (dispatch), Process (concurrent state). Mixing them in one operation, or picking the wrong one, is what makes diffs "look long" without being clearly bad.
Diagnosis approach when reviewing a chain: name the shape the code uses (rebind chain, with chain, manual recursion, eager Enum). Then name the shape the operation needs (independent validators? sequential dependency? lazy stream?). Mismatches are the findings.
If you see...
Suggest instead
Severity
Archdo rule
with chain (≥2 <-) inside validate_* / import_* / bulk_* / check_* function whose success body combines 2+ bound values into a struct/map
Replace with error-accumulating reduce — short-circuit gives bad UX (user fixes one field, resubmits, sees next error). Wrap each validator in a reducer that collects errors into a list; return {:error, errors} if non-empty. Cross-ref elixir-implementing §5.10.6, elixir-planning §4.7.2.
Request-change for form/UX-facing flows; Suggest for internal
6.95 ShortCircuitOverAccumulating
case fetch() do {:ok, v} -> {:ok, transform(v)}; {:error, _} = e -> e end (verbose Result.map)
with {:ok, v} <- fetch(), do: {:ok, transform(v)}, OR a Result.map/2 helper if the shape recurs in 3+ places. The case shape buries the intent.
Suggest
6.96 ResultMapOpportunity
Public 2-arg function with first arg named opts/options/config and last arg looking like the data subject (named data/list/map/subject or destructured %Schema{} = subject)
Reorder to subject-first: def fun(data, opts). The stdlib (Enum.map(coll, fn), String.replace(s, pat, rep)) is rigorous; backwards order silently breaks every downstream pipeline. Renaming arguments is a 5-minute refactor that pays off forever. Cross-ref elixir-implementing §5.10.10.
Request-change in domain APIs; Suggest in adapters
6.97 PipeSubjectPosition
2+ levels of nested Map.update / Map.put lambdas reaching into a known structure
update_in(state, [:a, :b], &fn) / put_in(state, [:a, :b], v). Nested-lambda form obscures the path; update_in says it in one line and composes. Caveat 1: Map.update/4 with a non-trivial default (Map.update(req, :body, %{name => v}, &Map.put(&1, name, v))) cannot be replaced by put_in — put_in requires the path to already exist. Use update_in only when the path is guaranteed populated, OR keep Map.update/4. Caveat 2: Archdo 6.98 has FP class — flagging "two Map.put calls where the second is in the value-arg of the first" includes cases where they operate on DIFFERENT maps (Map.put(outer, :k, Enum.reduce(xs, %{}, fn _, acc -> Map.put(acc, ...) end))). Read the diagnostic, verify both Map.puts actually mutate the same structure. Cross-ref elixir-implementing §5.10.9.
Suggest
6.98 NestedMapUpdateAsUpdateIn
Pipeline starting with File.stream! / Repo.stream / Stream.resource / Stream.unfold followed by 3+ eager Enum.* steps
Replace intermediate Enum.* with Stream.*; only the terminal step (reduce/count/into) must be eager. Each eager step materializes the whole intermediate list, defeating the lazy producer. Order-of-magnitude memory difference on large inputs. Cross-ref elixir-implementing §5.10.11.
Request-change if input may be large; Suggest otherwise
Enum.reduce(xs, init, fn h, acc -> transform(h, acc) end) — or one of reduce_while / map_reduce / flat_map_reduce / Enum.sum / Enum.frequencies depending on shape. Manual recursion is only needed for non-fold shapes (tree traversal, mutual recursion). Cross-ref elixir-implementing §5.10.12.
Suggest
6.100 ManualRecursionAsReduce
3+ consecutive subject = Mod.fun(subject, ...) rebindings to the same name (Multi/Conn/Socket/Changeset builder pattern)
Thread with pipes: Mod.new() |> Mod.fun1(...) |> Mod.fun2(...). The rebind form is imperative-flavored — forces the reader to verify each line refers to the previous binding. Cross-ref elixir-implementing §5.10.13.
Suggest
6.101 BuilderPatternNotThreaded
Public to_X/1 (encoder: to_string, to_json, to_url, to_iodata) without a matching from_X / parse_X / decode_X in the same module
Either add the inverse (with a round-trip property test asserting decode(encode(x)) == x), OR if the output is for human display only, implement String.Chars / Inspect instead — those don't imply a round-trip contract. Cross-checked against production: 6.102 has high FP rate — three classes of legitimate one-way encoders that should NOT be flagged: (1) external-API serializers (to_stripe, to_intercom, to_segment) — output is for an HTTP request, the receiving system sends back its own shape; (2) lossy projections (to_minor_units :: Money.t() -> integer — drops currency; to_stripe_currency :: Money.t() -> String.t() — drops amount); (3) stdlib decoder wrappers (to_date! — Date.from_iso8601 exists; to_atom — String.to_existing_atom exists). Skip the diagnostic if the encoder fits any of these classes. Cross-ref elixir-planning §4.7.5, architecture-patterns.md §4.12.8.
Suggest (low confidence — high FP class)
6.102 EncoderWithoutDecoder
Module with defstruct + smart constructor (validate/parse/build/new returning {:ok, %__MODULE__{}}) + functions taking %__MODULE__{} as input
Consider phantom types: split into %UnverifiedEmail{} (raw input) and %Email{} (validated). Consumers signed for %Email{} are statically guaranteed validated. Defensive answer: keep the single struct + add a property test asserting validate's invariant. Cross-ref elixir-planning/architecture-patterns.md §4.12.8.
Suggest (judgment-soft); Request-change only if validation is non-trivial AND consumers should not accept raw input
6.103 PhantomTypeOpportunity
with chain ≥7 <- clauses in one function
Architectural drift signal — investigate, don't auto-flag. 2–4 healthy, 5–6 approaching limit, 7+ deserves a look. Genuine exceptions to the "split it" advice (verified against production code): a deliberate filter cascade where every <- is a "should we proceed?" predicate (e.g., social-feed filtering — Pleroma's streamer.ex:198 runs 11 filter checks; splitting would harm clarity); a threading-builder disguised as with where every step is %{error: nil} = subject <- step(subject) (e.g., logflare's logs_search.ex:60). Read the chain before flagging. Cross-ref elixir-implementing §5.10.4.
Suggest (with note "verify this isn't a deliberate filter / state-machine")
n/a (judgment)
withelse clause that handles success values, OR catch-all else clauses re-raising errors (else x -> x)
Bare with (no else) is the LCO-safe railway form — errors propagate transparently. Add else ONLY to translate error shapes for callers; never to handle success. Catch-all else adds nothing the bare form doesn't already do. Cross-ref elixir-implementing §5.10.1, §5.10.2.
Request-change for new code; Suggest for legacy
n/a (idiom)
Building-block-flavored function (per axes 1–6) with capabilities read inline (DateTime.utc_now, :rand.uniform, Application.get_env, :persistent_term.get)
Pass the capability as an argument. For 2+ capabilities, bundle into a ctx struct threaded through the call chain. The orchestrator resolves and provides the capability. Cross-ref elixir-implementing §5.10.8, elixir-planning rule 22c.
Request-change for modules claiming building-block status; Suggest otherwise
n/a (related to 5.74)
7.13 Building-block module review
Full reference: elixir-planning/building-blocks.md. Modules that claim building-block status (via @moduledoc) — or modules that the diff intends to make building-block-shaped — get reviewed against the seven-axis checklist. Each axis is binary: pass or fail.
The seven axes:
Axis
Means
Failed when...
1. Input closure
Output is a function of arguments only
Reads Application.get_env, DateTime.utc_now, :rand.uniform, :persistent_term.get, Process.get, :ets.lookup (anything not in the parameter list)
2. Determinism
Same inputs → same outputs
Same as axis 1 — usually overlap
3. @spec
Public functions have type contracts
Missing @spec on a public function
4. Totality
Function returns for every input in the documented domain
Raises on inputs the spec accepts; or accepts inputs not documented
raise for an expected failure (e.g., not-found, invalid format); use ! variant if raise is intended
7. Input guard
Input domain narrowed by guards / pattern match / changeset / Ecto.Enum
Public function accepting term() or unconstrained map() without runtime validation
What to flag in review:
If you see...
Severity
Archdo rule
Module with @moduledoc "Building block ..." (or matching @moduledoc "building-block ...") AND a Logger.* / Phoenix.PubSub.* / Repo.* / :telemetry.execute / :ets.insert / :persistent_term.put call inside any function body
Block — the moduledoc is lying. Either remove the side effect (extract to orchestrator) OR remove the building-block claim. Pure logic + side effects in one module is the textbook anti-pattern. Cross-ref elixir-planning/building-blocks.md §3.1 axis 5.
5.74 InlineEffectInBuildingBlock
Building-block module with expensive call (Regex.compile!, Jason.decode!, :crypto.hash, DateTime.from_iso8601) on a literal argument inside a function body
Suggest — hoist to a module attribute (@valid_re Regex.compile!("^[a-z]+$")). The expensive work should run once at compile time, not every call. For boot-time large values use :persistent_term.put. Cross-ref elixir-implementing §9.13, elixir-planning §4.7.8.
5.75 MemoizeOpportunity
Building-block module with public functions that lack @spec
Request-change — building-block status implies a contract. @spec IS the contract.
n/a (axis 3)
Building-block module with public function that takes an unconstrained term() / any() without guard or pattern-match narrowing
Request-change — input guard is axis 7. A function whose input domain is unbounded cannot be property-tested.
n/a (axis 7)
Building-block module with raise on an expected failure path (not bang variant)
Request-change — return {:error, reason}. Raises break axis 6 (errors-as-values). The ! variant is for the explicit-raise contract.
n/a (axis 6)
Module that mixes obvious building-block functions (pure transforms) with obvious orchestrator functions (DB writes, PubSub, HTTP) — no @moduledoc classification, no boundary
Suggest — split into MyApp.Pricing (building block) + MyApp.Pricing.Workflow (orchestrator). Mixing the two in one module fails Archdo's Blackbox analyzer (CE-54..57) and makes property-testing impossible. Cross-ref elixir-planning rule 11b, building-blocks.md §3.5.
Suggest (architectural)
New module added without classification (no @moduledoc indicating building_block / orchestrator / interface)
Suggest — explicit classification clarifies what the module is FOR. The @moduledoc doesn't have to literally use those words, but the description should make the type clear ("Pure pricing functions", "Orchestrates checkout flow", "LiveView for the cart").
Suggest
Severity calibration: a building-block claim is a contract. Breaking it is request-change minimum, block if the claim is load-bearing for property tests or downstream reasoning. A module without the claim isn't held to these rules — only flagged if it would obviously benefit from being one.
7.14 LiveView async / CPS review
Full reference: phoenix-liveview/SKILL.md "Async as Continuation-Passing Style". A LiveView is a single GenServer per user session. Synchronous work in handle_event/3 freezes the session — clicks queue, the UI feels unresponsive, and a single slow API call hangs the whole user.
Wrap with start_async/3 + handle_async/3. The handler returns {:noreply, start_async(socket, :name, fn -> work end)} and a separate handle_async(:name, {:ok, result}, socket) callback receives the result. LV stays responsive throughout.
Block (production smell)
5.76 InlineHttpInLiveViewEvent
handle_event/3 calling a context function that internally performs HTTP (e.g. MyClient.fetch_user(id) which wraps Tesla.get, Github.get_repo(token, name)) without start_async. The context call hides the HTTP underneath but the LV process still freezes for the request duration
Same fix as the direct-HTTP row: wrap the context call in start_async/3. Archdo rule 5.76 has a known false-negative class for indirect HTTP — it only detects direct calls to HTTP-lib modules (Tesla.*, Req.*, etc.). A reviewer reading the actual code must trace context calls that perform I/O underneath. Verified production case: logflare's vercellogdrains_live/index.ex:120 calls Vercel.Client.delete_log_drain(drain_id) synchronously — wraps Tesla internally, not flagged by Archdo, but same UX freeze.
Block (production smell)
5.76 (FN class)
handle_event/3 calling a long-running Repo.* query (full-table scan, complex aggregation) without start_async
Same fix as above. Repo.transaction / Repo.all on slow queries blocks the LV identically to HTTP.
Request-change
n/a
mount/3 doing blocking work without connected?(socket) guard
Move to if connected?(socket), do: assign_async(socket, :data, ...). Mount runs twice (HTTP then WebSocket) — blocking work runs twice. The assign_async pattern handles both cleanly.
Request-change
n/a (LiveView rule 1)
start_async lambda capturing the full socket
Capture only what the lambda needs: id = socket.assigns.user_id; start_async(socket, :load, fn -> Accounts.get(id) end). Capturing the socket pulls all assigns into the spawned process; if any contain large data (streams, big lists), memory bloats.
Request-change for LV with large assigns; Suggest otherwise
n/a
assign_async/start_async result rendered without <.async_result> wrapper for loading/error states
Use <.async_result :let={data} assign={@data}> with <:loading> and <:failed> slots. Direct rendering shows nothing during load, breaks on error.
Add the failure clause: def handle_async(:name, {:exit, reason}, socket) — otherwise Process.exit from the spawned task is unhandled. Crash flows back as an error tuple, not an unhandled exit.
Request-change
n/a
Why this matters more than other "async" review: LiveView's single-process-per-session shape makes blocking visible to the user immediately. A 2-second HTTP call in a controller is a 2-second response; a 2-second HTTP call in a LiveView handle_event is a 2-second freeze where every other click queues. The blast radius is one user, but the perception is "the app is broken."
8. Debugging Playbook — by symptom
Depth: For expanded symptom → diagnosis flow with concrete investigation steps, classification tables for crash reasons, and specific fixes, load debugging-playbook-deep.md.
Add validating clause at the top or a fallback clause
:timeout exit from GenServer.call
Callee slow or busy
Profile callee; extract work to Task; increase timeout only after understanding why
Infinite supervisor restart loop
Child crashes immediately on startup; max_restarts exceeded
Look at init/1; check for missing config; guard against startup with no DB
Tools to reach for:
# 1. Read the full stack trace — Elixir shows it in Logger output
# 2. Check the process's last known state
:sys.get_state(pid)
# 3. Trace the process to see messages arriving
:sys.trace(pid, true)
# ... reproduce ...
:sys.trace(pid, false)
# 4. See what's in the mailbox right now
Process.info(pid, :message_queue_len)
Process.info(pid, :messages) # actual messages (copy — heavy!)
8.2 Memory growth / leaks
Symptom: memory usage growing over time, OOM eventually
Pool the work (Task.async_stream with max_concurrency)
Top-N processes by memory:
:recon.proc_count(:memory, 20) # Top 20 by memory
:recon.proc_count(:message_queue_len, 20) # Top 20 by mailbox length
:recon.proc_count(:reductions, 20) # Top 20 by CPU work
Symptom: a GenServer is falling behind; its mailbox is growing
Cause
Symptom
Fix
Work per message is too slow
message_queue_len growing steadily
Offload heavy work to Task; use handle_continue for post-reply work
GenServer is the bottleneck for reads
High call rate, reads serialized
Move reads to ETS
Sync call blocking the callee
Callee itself blocked on another slow call
Make outer call async via :noreply + GenServer.reply/2
Producer sending faster than consumer processes
Classic PubSub / event-stream overload
Migrate to GenStage / Broadway for backpressure
Process.send_after flood
Scheduled messages piling up during lag
Coalesce timers; cancel before re-scheduling
Investigation:
# Is the mailbox growing over time?
for _ <- 1..5 do
IO.inspect(Process.info(pid, :message_queue_len))
Process.sleep(1000)
end
# What kind of messages are piling up?
Process.info(pid, :messages) |> elem(1) |> Enum.take(20)
8.4 Slow response / timeout
Symptom: requests or GenServer.call timing out, or response times high
Likely cause
Investigation
Fix
N+1 queries
Enable Ecto.DevLogger; count queries per request
Batch with Repo.preload or explicit query
External HTTP call without timeout
Check adapters (Req, Finch, HTTPoison) for receive_timeout
Set timeout explicitly
GenServer serialization bottleneck
message_queue_len growing
ETS for reads; PartitionSupervisor for shards
Missing DB index
EXPLAIN ANALYZE the slow query
Add index in migration
Connection pool exhaustion
:telemetry for [:ecto, :repo, :query] with :queue_time
Increase pool_size; reduce per-request queries
Slow serialization (JSON, term_to_binary)
Profile at the serialization call site
Stream-encode; use :erlang.term_to_iovec/1,2
Inner timeout exceeding outer
Check timeout cascade — outer must be > middle > inner
assert_in_delta with tolerance, or inject a clock module
Async mode but test uses named GenServer
Either use async: false or use a unique name per test
Factory sequence collision
Use sequence/2 for unique fields
Test pollution via :persistent_term
Set + on_exit to erase, or use a per-test namespace
DB migration race in parallel CI
Run migrations serially before parallel tests
Find which test flakes:
# Run in a repeatable order to isolate
mix test --seed 0 --max-failures 1
# Identify which earlier test pollutes the later one
mix test --trace --seed 12345 # pick a seed that reproduces
8.6 Dialyzer warnings
Symptom: Dialyzer flags contracts, missing callbacks, unreachable patterns, etc.
Warning pattern
Usual cause
Fix
Function X has no local return
Function always raises, no branch returns
Add a return path, or mark with @spec returning no_return()
