chore: wire strap harness — ledger, regression flow, agent context

.gitignore

@@ -5,3 +5,4 @@ scripts/.env
 luacov.stats.out
 luacov.report.out
 .luacov.stats.*.out
+.serena/

.strap/.gitignore

@@ -0,0 +1,2 @@
+index.db
+runs/

.strap/flows/regression.toml

@@ -0,0 +1,26 @@
+# nyan.nvim regression flow: lint + plenary suite; on red, an agent gets the
+# failure plus relevant lessons and must make `make lint && make test` pass.
+
+[[step]]
+name = "lint"
+kind = "run"
+cmd = "make lint"
+on_fail = "fix"
+
+[[step]]
+name = "test"
+kind = "run"
+cmd = "make test"
+on_fail = "fix"
+
+[[step]]
+name = "fix"
+kind = "agent"
+agent = "claude"
+cmd = "claude -p --allowedTools 'Edit Read Glob Grep Bash' --max-turns 25"
+tags = ["regression", "lua", "nvim"]
+verify = "make lint && make test"
+timeout_secs = 900
+
+[trigger]
+on = ["manual", "agent-finished"]

.strap/lessons/msqks7ac.md

@@ -0,0 +1,11 @@
++++
+id = "msqks7ac"
+tags = ["lua", "nvim", "debugger"]
+confidence = 0.6
+helped = 0
+harmed = 0
+created = "2026-08-12T21:02:24.228836Z"
+last_confirmed = "2026-08-12T21:02:24.228836Z"
++++
+
+no lua adapter exists for mcp-debugger (strap debug); plugin lua runs inside nvim anyway — debug via the 2s plenary suite plus nvim --headless -c 'lua ...' state dumps; OSV via a generic-DAP MCP tool is the upgrade path if step-debugging ever matters

.strap/tools.toml

@@ -0,0 +1 @@
+# managed tools — see `strap tool add`

AGENTS.md

@@ -0,0 +1,25 @@
+# nyan.nvim
+
+Neovim plugin (Lua). Plenary test suite, stylua formatting, luacheck static
+analysis.
+
+## Build and test
+
+```bash
+make test            # plenary suite under headless nvim (~2s)
+make lint            # stylua --check
+make lint-luacheck   # luacheck static analysis
+make format          # apply stylua
+make ci              # full local CI parity: lint + luacheck + test
+```
+
+## Harness
+
+This project uses strap (per-project agent harness). Start tasks with
+`strap lesson search "<topic>" --json`; record non-obvious learnings with
+`strap lesson add "..." --tag lua --tag nvim`; close the loop with
+`strap lesson feedback <id> --helped|--harmed`. See `strap --help`.
+
+The `regression` flow (lint + tests, agent fix on red, `make lint && make test`
+as ground truth) runs after each Claude Code session via the Stop hook in
+`.claude/settings.json`; run it manually with `strap flow run regression`.

CLAUDE.md

@@ -0,0 +1 @@
+@AGENTS.md