Extend MCP idle timeout to one hour

[justrach]

Summary

• extend the MCP idle timeout from 10 minutes to 1 hour
• keep dead-client cleanup separate from idle timeout with a 1-second stdin HUP poll
• update tests and README copy for the new timeout behavior

Verification

• zig build test
• zig build
• installed binary smoke: closing MCP stdin exited cleanly in ~285ms

[chatgpt-codex-connector]

💡 Codex Review

codedb/src/main.zig

Lines 967 to 968 in 9a70fb4

	const last = mcp.last_activity.load(.acquire);
	if (last == 0) continue;

Avoid tight loop when activity timestamp is unset

The watchdog now sleeps only at the end of the loop, but if (last == 0) continue; bypasses that sleep. This introduces a hot spin path that repeatedly polls and reloads atomics with no delay until last_activity is initialized, which can peg a core if MCP startup is delayed or refactored to initialize activity later. Add a sleep in this branch (or initialize last_activity before starting the watchdog) to keep the 1-second polling behavior.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

Benchmark Regression Report

Thresholds: 10.00% and 50,000 ns absolute delta

NOISE means the percentage threshold was exceeded, but the absolute delta was too small to fail CI.

Tool	Base (ns)	Head (ns)	Delta	Abs Delta (ns)	Status
codedb_bundle	481005	487143	+1.28%	+6138	OK
codedb_changes	54861	53992	-1.58%	-869	OK
codedb_deps	8958	9104	+1.63%	+146	OK
codedb_edit	6124	6109	-0.24%	-15	OK
codedb_find	61830	63170	+2.17%	+1340	OK
codedb_hot	98770	97587	-1.20%	-1183	OK
codedb_outline	231731	238884	+3.09%	+7153	OK
codedb_read	83841	83474	-0.44%	-367	OK
codedb_search	174453	177438	+1.71%	+2985	OK
codedb_snapshot	251022	243525	-2.99%	-7497	OK
codedb_status	216322	218376	+0.95%	+2054	OK
codedb_symbol	59188	57909	-2.16%	-1279	OK
codedb_tree	87990	91863	+4.40%	+3873	OK
codedb_word	67519	66759	-1.13%	-760	OK

[justrach]

Speed comparison for timeout change

Compared current #328 build (67746b5d..., 1h idle timeout) against the immediately previous installed 0.2.579 build (f0a40ecd..., 10m idle timeout).

GitHub bench-regression

Official PR bench artifact passed with all rows OK; no benchmark regression from the timeout/watchdog change.

Tool	Base	Head	Delta	Status
codedb_bundle	481005 ns	487143 ns	+1.28%	OK
codedb_outline	231731 ns	238884 ns	+3.09%	OK
codedb_search	174453 ns	177438 ns	+1.71%	OK
codedb_snapshot	251022 ns	243525 ns	-2.99%	OK
codedb_status	216322 ns	218376 ns	+0.95%	OK
codedb_symbol	59188 ns	57909 ns	-2.16%	OK
codedb_tree	87990 ns	91863 ns	+4.40%	OK

Full artifact had every benchmark row within the 10% / 50,000ns gate.

Local MCP lifecycle smoke

20 local runs per binary, using the installed binaries directly:

Measurement	Current avg / median / p95	Previous avg / median / p95
initialize response	40.2 / 40.6 / 46.9 ms	41.2 / 41.5 / 42.4 ms
tools/list	0.2 / 0.1 / 0.3 ms	0.1 / 0.1 / 0.2 ms
codedb_status	0.1 / 0.1 / 0.4 ms	0.1 / 0.1 / 0.1 ms
normal stdin-close shutdown	1195.5 / 1190.2 / 1278.0 ms	1189.6 / 1188.7 / 1346.9 ms

The lifecycle numbers are effectively unchanged for normal MCP EOF shutdown. The behavioral change in #328 is the policy: idle timeout is now 1 hour, while the dead-client watchdog poll remains independent and is bounded to 1 second.

Remote wiki calls

Both current and previous builds successfully handled backend=wiki calls for chromium policy and axios/axios CVEs. Timings were network-bound and noisy, so I’m treating them as smoke checks rather than regression data.