[lukemaj]

What changed: the serving hook now waits up to 2 seconds for a reply (connect up to 1 second; Grok stays at 4), and the OpenCode plugin up to 2 seconds; v0.5.6.
Why: the hook cut moments at 1 s total and 0.1 s connect, and since Cue went live 40 tool results on rocky had no pre-tool moment at all; the user decided to raise the cut-off to 2 s while Cue is new ("first we make it work, then we make it better", toolboxmd/chromeria#248).
So what: pairs with toolboxmd/chromeria#251 (Cue waits 1.5 s); after both merge the hook and the reinstalled OpenCode plugin are deployed on rocky, cavallo and bestia, and the lost-moment rate is measured on chromeria#248.

Part of toolboxmd/chromeria#248.

Change

• bin/contextmd-hook: deadline 1 -> 2 (Grok 4 unchanged), --connect-timeout 0.1 -> 1; header and comment state the cut-off and its reason.
• contextmd/serve/opencode/contextmd.js: the moment request's AbortSignal.timeout 1000 -> 2000 (settle stays 300 ms).
• README.md: the server-down sentence states the new limits.
• Installed hooks already allow 5 s (install.py), checked live on all three machines.

Proof (head 293ba1e25c)

• test_hooks_wait_two_seconds_and_grok_four (replaces the 1 s test): a 1.5 s server now answers claude; at 2.5 s claude fails open in under 2.4 s while grok still answers. Failed at the old limits.
• OpenCodeCutoffTest: a reply taking 1.5 s now reaches the OpenCode user message. Failed at 1 s.
• CONTEXTMD_CHROMERIA=<chromeria with the paired Cue change> python3 -m unittest discover -s tests -t .: OK. (Against an older Chromeria without #240 one contract test fails as expected.)
• versionctl release-check: v0.5.6 OK.

Elon record

• Requirements and who asked: serving cut-off 2 s and connect 1 s (user, 2026-10-13, via the planner on chromeria#248).
• Deleted: nothing; the faster aim stays on #248 and Router: test the input window format on the 0.6B student #20.
• Bottleneck: moments cut at the hook before they reach serving.
• Checked myself: the hook and plugin limits, the installed hook timeouts on rocky, cavallo and bestia, and the runs above.

🤖 Generated with Claude Code

[lukemaj]

What: PASS at 293ba1e25c951469bd4c03bd0db1c885a4bf37f7, with no blocking findings.
Why: The shell hook and OpenCode retain fail-open behavior while allowing replies between one and two seconds.
So what: Proceed with paired chromeria#251; whichever of this PR and contextmd#56 merges second must rebase and advance the version.

• Verified curl connect limit 1 s, total 2 s including connect, Grok total 4 s, OpenCode moment abort 2 s, and both settle requests 300 ms. install.py sets host hook timeout 5 s, so even Grok's moment plus settle fits that envelope. Installed-host 5 s values are the supplied task premise, not rechecked live here.
• Paired server Cue waits 1.5 s for user messages and failures, collects 300 ms after successful calls, connects to Cue within 500 ms, sends a 1.5 s queue deadline, and bounds the background exchange at 2 s. There is no separate earlier overall route deadline.
• PYTHONDONTWRITEBYTECODE=1 CONTEXTMD_CHROMERIA=<exact #251 worktree> python3 -m unittest discover -s tests -t .: 119/119 passed, including 18 production-server contracts. Initial missing-dependency setup error was corrected in the temporary clone before this passing run.
• Holding head tests constant on base b62293180977058c70cb79e0a6b0b3b3bb9198dc: the two new cutoff regressions both fail for the intended reason (empty Claude reply and no OpenCode rule at 1.5 s). The head suite verifies slow-server fail-open and Grok's longer wait.
• docs: add docs/map.md, the ContextMD feature map #56 is still open and also claims 0.5.6. This is merge-order bookkeeping, not a defect in this head.
• These are component deadlines, not a demonstrated hard bound for every real workload. Slow filesystem, projection or authentication work can consume the remaining budget; #248's production timing and post-deploy acceptance remain unverified here.

Read the complete PR diff and body, linked Issue and all prior PR/Issue comments, including inline review comments. No private data or credentials found in the changes. Testing used only my clones/worktrees under /tmp; no main checkout writes, live Chromeria ports, live database access or restart. Reviewer: independent Sol, medium effort.

[lukemaj]

Version-only rebase onto main 8d496f4 (v0.5.6, #56): head 8692515fe3, v0.5.7, from reviewed 293ba1e25c. The patch to bin/contextmd-hook, contextmd/serve/opencode/contextmd.js, README.md and tests/test_serve.py is byte-identical; only VERSION and CHANGELOG changed. Full suite OK against chromeria#251.

[lukemaj]

What: PASS at 293ba1e25c951469bd4c03bd0db1c885a4bf37f7. No blocking findings.
Why: The hook waits 2 s total and 1 s to connect (Grok stays 4 s), the OpenCode plugin aborts the moment at 2 s, and settle stays 0.3 s. Installed hooks stay at 5 s. A slow or missing server still exits 0 with no body.
So what: This can merge. contextmd#56 also claims 0.5.6; whichever merges second rebases and bumps. The paired Cue wait is reviewed on chromeria#251.

Head matches the brief. One commit.

Checked, and consistent with the decision:

• bin/contextmd-hook: deadline=2, Grok 4, curl --connect-timeout 1 -m "$deadline". Off file, missing credential, curl failure, and a non-200 all exit 0. Settle is a separate curl -m 0.3 after the reply is printed, so it is outside the 2 s cut and inside the 5 s hook.
• contextmd/serve/opencode/contextmd.js: moment AbortSignal.timeout(2000), settle 300. plugin_source reads this file. A thrown abort returns null.
• install.py still sets the command hook timeout to 5.

Proof, temp clone, not the main checkout:

• OpenCodeCutoffTest and test_hooks_wait_two_seconds_and_grok_four pass (1.5 s reaches Claude and OpenCode; at 2.5 s Claude is empty in under 2.4 s and Grok still answers).
• Same tests with the hook put back to deadline=1 and --connect-timeout 0.1, and the plugin to AbortSignal.timeout(1000), both fail: OpenCode gets [], Claude gets ''.

Non-blocking:

• contextmd#56 also sets VERSION and CHANGELOG.md to 0.5.6. The second merge conflicts until it rebases.
• Since 16:11Z, 3 of about 1,700 Cue-started pre-tool moments already took over 2 s of server time, so that tail still fail-opens. The 2 s hook is what was asked. Whether the 1.5 s Cue wait stays inside it is the finding on chromeria#251.

No private data. The only token in the diff is the test fixture "t". Not live-verified: I did not start Chromeria or use its ports.

[lukemaj]

What: PASS: 8692515fe3835c632ae1e5c71f3252d4606c7f8e is a pure rebase of 293ba1e25c951469bd4c03bd0db1c885a4bf37f7, with version bookkeeping.
Why: The patch excluding VERSION and CHANGELOG.md is byte-identical, and all 119 tests pass.
So what: The independent rebase check is complete; the planner can use this verdict for the exact head.

Compared each commit against its parent. Both non-bookkeeping patches are 7,666 bytes, SHA-256 c1e2c21ea7636b527133e6fbd9ddc196954c044b01398f962a1710fa4c0148ff; only the VERSION and CHANGELOG.md patches differ. Version is 0.5.7. The new parent is #56's merge commit 8d496f4b7d8b19317de5b133483c5055622a8af6; head-tree differences in AGENTS.md and docs/map.md are inherited from that base.

Read-only review in /tmp/contextmd-review-58.V0Y4HD. CONTEXTMD_CHROMERIA=/tmp/chromeria-review-247.naggDa python3 -m unittest discover -s tests -t .: 119 tests, 36.101 s, OK, no skips. The initial run using the local Chromeria checkout failed during server setup; rerunning with the available dependency checkout above passed. No writes to ~/dev/toolboxmd/contextmd.

[lukemaj]

Rebase check of 8692515 (parent 8d496f4, main after #56) against 293ba1e.

The code patch is byte-identical. Only VERSION and CHANGELOG.md differ: 0.5.6 became 0.5.7, same note. The commit message is unchanged.

python3 -m unittest discover -s tests -t . at that SHA: 119 tests, OK (18 skipped, CONTEXTMD_CHROMERIA unset).

Verdict: pure rebase.