[debugmcpdev]

Closes #963.

What changed

The DAP connect window is bounded by what it waits for, not by a count. DapConnectionManager.connectWithRetry was 500 ms + 60 attempts × 200 ms ≈ 12.5 s for every adapter, alive or dead, with no relation to the parent's 30 s initialization ceiling. An adapter that needed more than ~12.5 s to listen failed the launch with Failed to connect DAP client: connect ECONNREFUSED while it was still booting, 17.5 s before the parent would have given up. Now:

• The parent's deadline travels in the init payload (ProxyInitPayload.initializationTimeoutMs, written by ProxyManager, validated by the parser like adapterPort).
• connectWithRetry(host, port, { budgetMs, abort }) is bounded by wall clock. In spawn mode the worker passes the parent's deadline + 500 ms (the parent's timer was armed first and its stage-aware message — adapter process spawned (PID n) … DAP connection was never established — is the one a slow boot should get) and an AbortSignal the adapter's exit handler fires with { code, signal, stderrTail }, so a dead adapter ends the wait at once and the error quotes the exit. Connect mode (a remote rdbg/debugpy already listening) and legacy payloads keep the historical 12.5 s window as the default. The 500 ms / 200 ms cadence is unchanged.
• Three exits, each its own message, all keeping the Failed to connect DAP client: prefix.

Found on the way, fixed here: the worker's stdout pipe was never read. ProxyManager spawns the worker with stdio: ['pipe','pipe','pipe','ipc'] and drained only stderr. Outside the server CLI (which sets CONSOLE_OUTPUT_SILENCED=1, inherited by the worker), the worker's winston console transport writes every log line to stdout; once the OS pipe buffer filled (64 KB on Windows) the worker blocked solid — CPU flat, deaf to IPC (terminate ignored, the parent force-killed it), its DAP socket unread, no exception anywhere. The new 14 s integration case was the first test to log that much; it stalled three different ways at a fixed ~13.8 s until the cause was found. The parent now drains stdout like stderr ([ProxyManager STDOUT] at debug, sanitized). Every ProxyManager embedder that is not the CLI was exposed.

Diagnostics. The stdio-client suites' servers now log to logs/tests/stdio/, so their per-session proxy logs land inside the Windows CI diagnostics artifact; before, they went to os.tmpdir() and were lost with the runner — the connect timeline that would have settled #963 was unrecoverable. The adversarial suite's new case prints the proxy log's connect lines in its failure message; ADVERSARIAL_KEEP_TMP=1 keeps its temp dir.

Measured

CI, windows-latest, 2026-10-19 (runs 39336731176 and 39339665769)	failing Ruby case 15.8 s and 15.5 s before ECONNREFUSED, rdbg alive (PID, no exit code); a sibling launch in the same suite succeeded in 12.9 s under the same load; normal boots 3.5–4.8 s
rdbg boot to first accepted connect, this box (Ruby 3.4.9, debug 1.11.0), 10 samples	idle: median 380 ms, max 1909 ms (one cold start); during npm run build: 308–386 ms
Adversarial fixture, listenDelayMs: 14_000, parent deadline 25 s	main: rejects at 12.5 s with ECONNREFUSED; this branch: connects, start() resolves in 15.4 s
Same fixture, listen delays 8 / 11 / 13 s vs 13.5 / 14 s, before the stdout drain	resolved / resolved / resolved vs stalled / stalled — the stall tracked logged bytes, not the step in flight; CONSOLE_OUTPUT_SILENCED=1 or logLevel: warn made it vanish, UV_THREADPOOL_SIZE=64 changed nothing

The runner, not rdbg, is the variable: locally rdbg listens in under half a second even under a build.

Tests

• tests/unit/proxy/dap-proxy-connection-manager.test.ts: past 60 attempts while the budget remains; default budget expires with the "nothing accepted a connection … within 12.5s" message; at least one attempt however small the budget; abort ends the wait at once quoting code/signal/stderr and no attempt follows; a pre-aborted signal makes no attempt; a bare abort() (an AbortError, whose own code is 20) is not read as an exit code.
• tests/proxy/dap-proxy-worker.test.ts: spawn mode passes the deadline-derived budget and a signal; a legacy payload passes no budget; connect mode passes neither; an adapter exit during the connect aborts the signal with the exit and the stderr tail.
• tests/unit/proxy/dap-proxy-message-parser.test.ts, tests/unit/proxy/proxy-manager.start.test.ts: the payload field; the init command carries the manager's deadline (default and injected); the stdout drain (listener attached, lines logged, secret line redacted, trailing partial flushed on end).
• tests/integration/adversarial-dap/adversarial-adapter.test.ts: the 14 s listen-delay case (red on main, green here, ~15 s).

Verification

Gate	Result
npx vitest run --project unit	373 files, 7452 passed, 4 skipped
tests/integration/adversarial-dap/adversarial-adapter.test.ts	7 passed (new case 15.4 s)
Ruby stop-on-entry + JS launch-config stdio suites (real rdbg, real js-debug)	20 passed; proxy logs present under logs/tests/stdio/sessions/
Python integration pair	3 passed
npm run lint, npm run typecheck:all (ratchet 0/0), node scripts/check-docs.mjs, pnpm changelog:check	clean
Dev proxy dogfood after rebuild	see the last comment

Out of scope

The 500 ms / 200 ms cadence; a readiness signal from rdbg's DEBUGGER: banner; the parent's 30 s default; #875 (whose next sighting now gets a proxy log in the artifact).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y3ULnoVSoCZ5t2Y8d98Ctv

[codecov]

Codecov Report

❌ Patch coverage is 98.92473% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/proxy/dap-proxy-connection-manager.ts	98.46%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[debugmcpdev]

Dev proxy dogfood after the rebuild from this branch (Windows, the dev proxy at tools/dev-proxy): a Ruby launch of examples/ruby/fizzbuzz.rb with stopOnEntry: true paused at line 4 (reason: entry, stopOnEntrySuccessful: true), continue_execution then ran it to completion with exit code 0 and get_output held all 15 lines through 15: FizzBuzz plus the fizzbuzz complete stderr line; a Python launch of examples/python/fibonacci.py with stopOnEntry: true paused at its entry and ran to exit code 0. Both proxy workers went through the new connect loop and the drained stdout.