Search o'Clock

Your AI keeps fixing bugs with answers from before its training cutoff. It defaults to the old, most-common pattern instead of the current API, slaps on --force / @ts-ignore to make the error disappear, skips the web-search tool that's right there, and burns retry loops on the same stale guess. Bigger models don't fix this — they trained on yesterday's code too.

Search o'Clock intercepts the failure, stamps today's date, and makes the agent research the fix against live, dated sources — pick the most durable one (not the fastest band-aid) — and have an independent second model agree from the evidence before the fix is trusted.

The Claude Code plugin is its own repo / marketplace submission. This repo is the harness-agnostic core so everyone gets results — Codex, Gemini CLI, Cursor, opencode, Goose, Cline, Aider, and any shell.

Why existing fixes don't cut it

• A bigger model still trains on yesterday's code and defaults to the old pattern.
• AGENTS.md / system prompts go stale and the model can ignore them — a prompt is a suggestion, a hook is enforcement.
• An optional /command doesn't fire when you need it — in-prompt good behavior stops transferring the moment the model can act (arXiv:2602.16943).
• Self-review doesn't work — a model checking its own answer invents errors or rubber-stamps. You need a separate model. See docs/WHY.md.

The core: one portable engine

socx is a dependency-light (bash + python3) deterministic engine. Given a failed command + error, it emits a date-anchored, durability-ranked, double-validated research protocol on stdout. No model is in its trigger path; it never blocks; any internal error → exit 0 (fail-open).

socx hook          # read a normalized failure event (JSON) on stdin → protocol on stdout  (for adapters)
socx run -- CMD…   # run CMD; on non-zero exit, print the protocol  (universal, no hooks needed)
socx text | json   # debug render
socx selftest      # exit 0 if it works

Install — pick your integration (precedence: native hook > shell-wrapper > MCP)

Harness	Mechanism	How
Claude Code	native hook ✅	the dedicated plugin — /plugin marketplace add waitdeadai/searchoclock
OpenAI Codex CLI	native hook ✅	adapters/codex-cli/ — PostToolUse fires on non-zero Bash exit
Google Gemini CLI	native hook ✅ (v0.26.0+)	adapters/gemini-cli/ — extension-bundled PostToolUse/AfterTool
Cursor	native hook ✅ (1.7+)	adapters/cursor/ — afterShellExecution in .cursor/hooks.json
opencode	native hook ✅	adapters/opencode/ — tool.execute.after TS plugin (MCP-tool calls excluded, issue #2319)
Goose	native hook ✅ (v1.34.0+)	adapters/goose/ — PostToolUse hook
Cline	native hook ✅ (in-code)	adapters/cline/ — @cline/sdk lifecycle hook
Aider · Crush · Amp · Continue	shell-wrapper ✅	no failure hook → socx-run -- <cmd> / aliases (adapters/shell-wrapper/)
Hermes (Nous) · OpenClaw	shell-wrapper / MCP	gateway/agent, no tool-lifecycle hook → wrap the backend command or MCP
any MCP-capable harness	MCP (supplement)	mcp/ — research-on-failure tool; model-elective, not guaranteed-on-failure

Adapter mechanisms are from each tool's 2026 docs; some hook schemas evolve fast — each adapter folder notes the doc to confirm against. The shell-wrapper works everywhere and needs nothing from the harness, so start there if in doubt.

30-second universal start (any CLI)

git clone https://github.com/waitdeadai/searchoclock-core
export PATH="$PWD/searchoclock-core:$PATH"
# wrap the commands your agent runs:
socx-run -- npm run build
# or alias them so failures always trigger research:
alias build='socx-run -- npm run build'

Configuration

Same knobs as the plugin, read as SOCX_* (or SEARCHOCLOCK_*): ENABLE, SOLUTION_MODE (durable|fast|both), MIN_CANDIDATES, PROBE_ALL, VALIDATE, VALIDATOR_MODEL, VALIDATE_MIN_CONFIDENCE, COOLDOWN_SEC, MIN_INTERVAL_SEC, MAX_PER_SESSION, SEVERITY_MIN, GOAL_SCOPE, STATE_DIR, HARNESS (label shown in the protocol).

Relationship to the Claude Code plugin

The plugin is the polished, marketplace-validated Claude Code adapter. This core is the shared engine + the other-harness adapters. They keep the same protocol, durability rubric, and double-validation so a fix you'd get in Claude Code is the fix you'd get anywhere. adapters/claude-code/ is the reference adapter.

Sister tools

Part of the waitdeadai family of Claude Code governance hooks — searchoclock (the Claude Code plugin), time-anchor, no-sycophancy, llm-dark-patterns. More at waitdead.com.

License

Apache-2.0. See LICENSE.