Add agentic workflow enhancements: timeout, timing, truncation, enriched info

[samtalki]

Summary

• Eval timeout: optional timeout parameter kills the worker on hung/infinite code (e.g., eval(code="while true end", timeout=5))
• Execution timing: every eval result now includes elapsed time ([45.2ms] or [1.23s])
• Output truncation: max_output parameter (default 50K chars) prevents context window overflow, preserving head (60%) + tail (40%)
• Enriched info: info tool now shows variable types and sizes (e.g., x::Int64, df::DataFrame (1000, 5)) instead of just names
• Configurable stackframes: max_stackframes parameter on eval controls error trace depth (default 5)

Zero new dependencies, no new MCP tools — all features extend existing eval and info tools. Plugin docs updated.

Test plan

• All 103 tests pass (julia --project=. -e "using Pkg; Pkg.test()")
• New tests for: timeout (sleep+kill+respawn), timing (format_elapsed, result contains timing), truncation (head+tail, marker), enriched info (types/sizes), configurable stackframes
• Manual: eval(code="1+1") shows timing, eval(code="sleep(5)", timeout=1) times out, info shows typed variables

🤖 Generated with Claude Code

[samtalki]

Code review

Found 7 issues:

1. truncate_output uses byte-based string indexing on Unicode strings, will crash with StringIndexError on multi-byte UTF-8 characters (box-drawing chars, Greek letters, math symbols common in Julia output). length(text) returns character count but text[1:n] indexes by bytes. Fix: use first(text, n) and last(text, n).

AgentREPL.jl/src/formatting.jl

Lines 15 to 19 in 4b5d7f0

	return text[1:head_chars] *
	"\n\n... [$(removed) characters truncated] ...\n\n" *
	text[end-tail_chars+1:end]
	end

2. Channel race condition in timeout implementation leaks async tasks. Channel{Any}(1) with two @async producers means the losing task's put! creates an unconsumed item. After timeout kills the worker, the eval @async task continues running and attempts fetch(future) on a dead worker, then put!s the error to an abandoned channel. This also departs from the documented expression-based IPC pattern (CLAUDE.md says "Uses remotecall_fetch(Core.eval, worker_id, Main, expr) instead of closures to avoid serialization issues").

AgentREPL.jl/src/worker.jl

Lines 134 to 153 in 4b5d7f0

	# With timeout: race remotecall against a timer
	result_channel = Channel{Any}(1)
	future = remotecall(Core.eval, worker_id, Main, eval_expr)
	# Race: eval completion vs timeout
	@async begin
	try
	result = fetch(future)
	put!(result_channel, (:ok, result))
	catch e
	put!(result_channel, (:error, e))
	end
	end
	@async begin
	sleep(timeout)
	put!(result_channel, (:timeout, nothing))
	end
	tag, payload = take!(result_channel)

3. get_worker_info() size logic shows misleading length=1 for all scalar variables. For scalars, size(42) returns (), length(()) == 0 is not > 1, so it falls through to "length=$(length(val))" which gives "length=1" for every scalar. Fix: check !isempty(s) before producing size output.

AgentREPL.jl/src/worker.jl

Lines 191 to 197 in 4b5d7f0

	size_str = try
	if applicable(size, val) && !(val isa AbstractString)
	s = size(val)
	length(s) > 1 ? string(s) : "length=$(length(val))"
	elseif applicable(length, val)
	"length=$(length(val))"
	else

4. mode_tool is still registered in server.jl but removed from plugin README and SKILL.md documentation, creating an inconsistency between server and docs. (CLAUDE.md says "Seven tools registered via ModelContextProtocol.jl" and lists mode as tool #7)

AgentREPL.jl/src/server.jl

Lines 74 to 83 in 4b5d7f0

	log_viewer_tool = create_log_viewer_tool()
	mode_tool = create_mode_tool()
	# Create and start the server
	server = mcp_server(
	name = "julia-repl",
	version = "0.5.0",
	description = "Persistent Julia REPL for AI agents - eliminates TTFX",
	tools = [eval_tool, reset_tool, info_tool, pkg_tool, activate_tool, log_viewer_tool, mode_tool]
	)

5. CLAUDE.md Key Functions section is now stale: capture_eval_on_worker(code) signature changed to capture_eval_on_worker(code; timeout=nothing) returning a 4-tuple (value_str, output, error_str, elapsed) instead of a 3-tuple. New functions truncate_output and format_elapsed in formatting.jl are also not listed.

AgentREPL.jl/CLAUDE.md

Line 78 in 4b5d7f0

- **`capture_eval_on_worker(code)`** - Evaluates code on the worker with output capture (`worker.jl`)

6. .mcp.json passes "JULIA_REPL_PROJECT": "${JULIA_REPL_PROJECT}" which may resolve to an empty string when unset, causing server.jl to error with "Cannot activate project: directory '' not found" since empty string passes the !== nothing check but fails isdir(""). (CLAUDE.md documents this as a variable users set manually before launching)

AgentREPL.jl/claude-plugin/.mcp.json

Line 12 in 4b5d7f0

"JULIA_REPL_PROJECT": "${JULIA_REPL_PROJECT}"

7. New timeout tests directly access AgentREPL.WORKER.worker_id for assertions, but the docstring on WORKER in src/types.jl says "Access via ensure_worker!() rather than directly."

AgentREPL.jl/test/test_eval.jl

Lines 85 to 95 in 4b5d7f0

	# Worker should have been killed
	@test AgentREPL.WORKER.worker_id === nothing
	end
	@testset "Next eval after timeout spawns fresh worker" begin
	# After the timeout above killed the worker, a new eval should work
	value_str, output, error_str, _ = AgentREPL.capture_eval_on_worker("42")
	@test error_str === nothing
	@test value_str == "42"
	@test AgentREPL.WORKER.worker_id !== nothing
	end

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}