docs: add GRA-374 experiment artifact

Gradata/docs/research/gra-374-multi-cli-install-success-rate.md

@@ -0,0 +1,96 @@
+# GRA-374: multi_cli_install_success_rate experiment
+
+Status: evaluated — DISCARD
+Window: 2026-05-12 to 2026-05-19
+Primary owner: gradata-eng
+Related issues: GRA-374, GRA-55, GRA-1161, GRA-1163
+
+## Objective
+
+Define and evaluate one quick 7-day experiment for Gradata's multi-agent installer:
+can the installer reach a useful cross-CLI success rate after adding bidirectional
+hook capture for Codex, Hermes, and OpenCode?
+
+## One measure
+
+`multi_cli_install_success_rate`
+
+Definition:
+
+- Numerator: CLI targets whose install plus correction-capture smoke test passes.
+- Denominator: CLI targets tested.
+- Target set: Claude Code, Codex, Cursor, Hermes, OpenCode.
+- Success threshold: at least 4 of 5 CLIs pass, i.e. >=80%.
+
+Formula:
+
+```text
+multi_cli_install_success_rate = passing_cli_targets / tested_cli_targets
+```
+
+## Baseline
+
+Before GRA-55, only Claude Code had confirmed bidirectional support.
+
+| CLI | Baseline install/capture state |
+| --- | --- |
+| Claude Code | PASS |
+| Codex | FAIL — capture broken |
+| Cursor | FAIL conservative — unverified |
+| Hermes | FAIL — capture broken |
+| OpenCode | FAIL — capture broken |
+
+Baseline: 1/5 = 20%.
+
+## Measurement runbook
+
+For each target CLI:
+
+```bash
+gradata install --agent <cli>
+# Record exit code.
+
+# Make one test correction in that CLI.
+
+gradata list-corrections --last 1h --format json
+# PASS if at least one fresh correction from that CLI appears.
+```
+
+Decision rule:
+
+| Result | Decision |
+| --- | --- |
+| >=4/5 pass | KEEP — continue multi-CLI installer investment |
+| 2-3/5 pass | PARTIAL — file per-CLI debug issues |
+| <2/5 pass | DISCARD — re-scope before more onboarding investment |
+
+## Evaluation result
+
+Verdict: DISCARD.
+
+Reason: the prerequisite implementation, GRA-55, was not actually merged. Code audit
+showed Codex, Hermes, and OpenCode still had injection-only hook registration and no
+post-tool/session-end capture path. Cursor remained unverified, so the conservative
+cross-CLI result stayed at the baseline.
+
+| CLI | Injection | Capture | Evaluation status |
+| --- | --- | --- | --- |
+| Claude Code | yes | yes | PASS |
+| Codex | yes | no | FAIL |
+| Cursor | yes, via MCP | unverified | FAIL conservative |
+| Hermes | yes | no | FAIL |
+| OpenCode | yes | no | FAIL |
+
+Final result: 1/5 = 20%.
+
+## Follow-up
+
+The hypothesis was not disproven; the prerequisite implementation did not ship.
+Follow-up issue GRA-1163 was filed to re-implement post_tool and session_end hooks
+for Codex, Hermes, and OpenCode with an explicit verification gate before completion.
+
+## Paperclip artifact note
+
+This document exists because the original GRA-374 completion was reverted by the
+artifact monitor: it had a useful in-thread experiment spec but no durable merged
+artifact URL. This file is the durable artifact for the experiment spec and evaluation.