@alessandroraffa/tangyr 4.1.0 → 5.0.0

package/README.md

@@ -363,6 +363,26 @@ acceptedLoss:
 
 `--ignore-errors` remains the blunt instrument: it suppresses the exit code for every error entry, including ones nobody has looked at yet.
 
+Not every error entry is structural. `rule-not-delivered` (below) reports a rule that did not reach a target on this run, which is usually fixable — by giving the target an instructions path of its own, or by not enabling it. Accept it when you have decided to live with it, not to quiet it.
+
+## Rules that did not arrive
+
+A rule reaches a target either as a file in that target's own rules slot or as a section in the instructions document it reads. Several targets resolve their instructions to one project-root `AGENTS.md` and write it in turn, so a rule folded into that file survives only for whoever wrote last.
+
+After every target has written, the run measures what is actually on disk and says so — one line per target, the count first, and the rules that did **not** arrive named:
+
+```text
+Rules reaching each target (3 declared, measured on disk):
+  claude-code  3/3
+  copilot      3/3
+  codex        0/3  missing: style, testing, review
+               its rules are folded into AGENTS.md, which cline wrote last
+  opencode     3/3
+  cline        3/3
+```
+
+Nothing is printed when every rule arrived; `--verbose` shows the block either way. The same fact reaches `--loss-report <path>` as a `rule-not-delivered` entry per rule per target, at `error` severity, so a run that loses a rule exits `3` rather than reporting success.
+
 ## Exit codes
 
 The full exit-code taxonomy, in numeric order. Every code below `5`
@@ -404,9 +424,9 @@ reaches them).
 ## Supported tools
 
 - **Claude Code** — instructions (`AGENTS.md` plus the `CLAUDE.md` shim), agents, skills, commands, rules, output styles, hooks, settings, and MCP, placed in scope-resolved Claude Code paths (fanned out across every configured global root — see [Multiple Claude Code global roots](#multiple-claude-code-global-roots))
-- **GitHub Copilot** — instructions (`AGENTS.md` plus the always-read `.github/copilot-instructions.md` mirror; project scope only, since Copilot has no global instructions slot), agents, prompts (compiled from commands), skills, hooks (`~/.copilot/hooks` or `.github/hooks`), and MCP. Hook matchers are rewritten to Copilot's own lowercase tool names (`bash`, `view`, `create`, `edit`, `grep`) — an untranslated `Bash` matcher never fires
-- **OpenAI Codex** — `AGENTS.md`, agent TOML files, rules (`.rules` files for `exec-policy`, appended to the instructions file as guidance otherwise), hooks (`hooks.json` plus the shared `~/.agents/hooks` script bridge; macOS and Linux only), and MCP in `config.toml`
-- **OpenCode** — instructions (`AGENTS.md`), agents, commands, skills (OpenCode-local plus the shared `~/.agents/skills` link), and `opencode.json` (MCP, permissions, custom provider), in XDG config paths. Hooks are **not** delivered: OpenCode exposes hooks only as plugin code, so every hook event is dropped and recorded as a `hook-as-plugin` loss
+- **GitHub Copilot** — instructions (`AGENTS.md` plus the always-read `.github/copilot-instructions.md` mirror; project scope only, since Copilot has no global instructions slot), agents, prompts (compiled from commands), rules (`.github/instructions/<name>.instructions.md`, each carrying the rule's path scope as an `applyTo` glob; project scope only, since that is a repository path with no user-level equivalent), skills, hooks (`~/.copilot/hooks` or `.github/hooks`), and MCP. Hook matchers are rewritten to Copilot's own lowercase tool names (`bash`, `view`, `create`, `edit`, `grep`) — an untranslated `Bash` matcher never fires
+- **OpenAI Codex** — `AGENTS.md`, agent TOML files, rules (`.rules` files for `exec-policy`, appended to the instructions file as guidance otherwise), skills and commands (both as skill directories at Codex's documented skills path — `~/.agents/skills` globally, `.agents/skills` inside a repository — because Codex deprecated its custom prompts in favour of skills), hooks (`hooks.json` plus the shared `~/.agents/hooks` script bridge; macOS and Linux only), and MCP in `config.toml`
+- **OpenCode** — instructions (`~/.config/opencode/AGENTS.md` globally; `.opencode/AGENTS.md` inside a project, declared in the `instructions` field of `opencode.json`, because the project-root `AGENTS.md` is written by four other targets in turn and only the last writer's content survives), agents, commands, skills (OpenCode-local plus the shared `~/.agents/skills` link), and `opencode.json` (MCP, permissions, custom provider), in XDG config paths. Hooks are **not** delivered: OpenCode exposes hooks only as plugin code, so a declared hook event is dropped and recorded as a `hook-as-plugin` loss
 - **Cursor** — instructions, subagents, `.mdc` rules, commands, skills (Cursor-local plus shared), `hooks.json` plus the shared hook-script bridge (matchers rewritten to Cursor's tool names — `Shell`, `Read`, `Write`, `Grep`), `mcp.json`, and `permissions.json`
 - **Cline** — instructions (`AGENTS.md`, project scope only), rules, workflows (compiled from commands), skills, and MCP. Agents and hooks are **not** delivered: Cline has no committable persona file (`persona-not-representable`) and exposes hooks only as `@cline/sdk` plugin code (`hook-as-plugin`); both are recorded as losses rather than silently skipped