Reach two more harnesses, and stop a database pin from drifting

.gitignore

@@ -5,3 +5,8 @@ research/
 *.db-journal
 *.db-wal
 *.db-shm
+
+# Generated by `make -C doc` from engram.texi.
+doc/engram.info
+doc/engram.html
+doc/engram.pdf

CHANGELOG.md

@@ -13,6 +13,45 @@ follows [Keep a Changelog](https://keepachangelog.com/); versions follow
 
 ### Added
 
+- **OpenClaude** is a supported harness (the eighth). It is a Claude Code fork
+  with its own config root: commands go to `~/.openclaude/commands/`, the MCP
+  registration is read from `~/.openclaude.json`, and its transcripts are read
+  by the existing Claude Code reader — the record types the fork adds are
+  recognised rather than counted as format drift.
+- **Antigravity** now gets a plugin at `~/.gemini/config/plugins/engram/`
+  (`plugin.json` plus one `skills/engram-<name>/SKILL.md` per command). It has
+  no slash-command directory at all; `agy plugin validate` reports a plugin's
+  own `commands/` as "converted to skills", so engram writes skills directly.
+- `install` reports `db_origin` (`override` / `registered` / `env` / `default`)
+  alongside the database it pins, so the relative-`engram.db` fallback is
+  visible rather than silent.
+
+### Fixed
+
+- **`install` now detects a stale database pin.** When a generated command
+  points at a different database than the harness currently registers, the
+  drift is reported on both the file and the harness before being corrected.
+  Previously the two could diverge indefinitely: if a harness's registration
+  moved after `install` ran, its `/engram-*` commands and its engram MCP tools
+  read different stores with nothing to say so.
+- **`find_git_root` requires a working tree, not merely a `.git` entry.** A
+  directory must contain `.git/HEAD`; a `.git` file (worktree or submodule
+  pointer) also counts. An empty `.git` directory in a shared location — e.g.
+  `/tmp/.git` — previously captured every path beneath it, so `save-chat` would
+  resolve its project root there, create `chat/`, and edit that directory's
+  `.gitignore`.
+- Harnesses with no writable surface each state their own reason instead of
+  sharing one sentence that described none of them precisely.
+
+### Changed
+
+- `HarnessSpec` models its command surface as an enum —
+  `CommandSurface::{Markdown, Plugin, None}` — replacing `commands_dir`,
+  `command_file`, and the `command_frontmatter` bool. Antigravity's surface
+  differs in artifact *shape*, not just in whether a header is read.
+
+### Added
+
 - **`flake.nix`** — Engram is now consumable as a Nix flake input
   (`github:Spacecraft-Software/Engram`), exposing `packages.default`,
   `packages.engram`, `apps.default`, `checks.default`, and `default`/`docs`

README.md

@@ -106,7 +106,7 @@ engram rule add --id skill-description-1000 \
 # Read what's in effect.
 engram rule list
 
-# Render into AGENTS.md and CLAUDE.md at the project root.
+# Render into AGENTS.md at the project root.
 engram rule sync
 
 # Withdraw one when it stops applying, then re-sync.
@@ -134,7 +134,7 @@ engram rule purge --id old-policy --yes                      # delete a retired
 ```
 
 Everything outside them is preserved verbatim, so the block can sit inside a
-hand-written `CLAUDE.md` indefinitely. The rendered block is a pure function of
+hand-written `AGENTS.md` indefinitely. The rendered block is a pure function of
 the rules — no generation timestamp — so re-running `sync` with unchanged rules
 writes nothing at all. That makes it safe in a `SessionStart` hook, a pre-commit
 gate, or a CI check (`engram rule sync --dry-run` reports `updated` if someone
@@ -214,7 +214,7 @@ storage failure. New routes should keep following that pattern.
 > touches anything outside the database. Target paths come from the server
 > process's own working directory and never from caller input, so there is no
 > path-traversal surface — but combined with the no-auth posture it means any
-> local process can rewrite that project's `AGENTS.md` and `CLAUDE.md`. The
+> local process can rewrite that project's `AGENTS.md`. The
 > CLI's `--file` override is deliberately not exposed over HTTP. Weigh this
 > before binding the server anywhere but `127.0.0.1`.
 

doc/engram.texi

@@ -601,7 +601,7 @@ explicit).  @code{--dry-run} previews the deletion.
 engram rule sync [--scope @var{id}] [--file @var{path}]@dots{} [--dry-run]
 @end example
 
-Render the scope's rules into @file{AGENTS.md} and @file{CLAUDE.md}
+Render the scope's rules into @file{AGENTS.md}
 (or the given @code{--file} targets, repeatable).  Only the region
 between the engram sentinels is rewritten.  With @code{--dry-run},
 reports what would be written without touching any file.
@@ -992,7 +992,7 @@ deletes.
 
 @item rule_sync
 @cindex rule_sync tool (MCP)
-Render a scope's rules into @file{AGENTS.md} and @file{CLAUDE.md} at
+Render a scope's rules into @file{AGENTS.md} at
 the project root.
 
 @item save_chat
@@ -1099,7 +1099,7 @@ unknown rule, @code{200} with the outcome on success.
 
 @item POST /v1/rules/sync
 @cindex POST /v1/rules/sync
-Render the scope's rules into @file{AGENTS.md} and @file{CLAUDE.md}.
+Render the scope's rules into @file{AGENTS.md}.
 This is the only route that writes outside the database; targets
 derive from the server process's working directory, never from caller
 input.
@@ -1175,9 +1175,10 @@ Four invariants worth not breaking:
 @item
 @cindex sync, as delivery
 @emph{Sync is the delivery mechanism, not an export.}  A row in SQLite
-never reaches a model's context.  Rendering into @file{AGENTS.md} and
-@file{CLAUDE.md} --- files harnesses auto-load --- is what makes a rule
-take effect.
+never reaches a model's context.  Rendering into @file{AGENTS.md} ---
+the file harnesses auto-load, and that Claude Code reaches through the
+@code{@@AGENTS.md} import in its @file{CLAUDE.md} (Steelbore Standard
+§5.7) --- is what makes a rule take effect.
 
 @item
 @emph{The rendered block is a pure function of the rules.}  No
@@ -1264,24 +1265,48 @@ installed, and never creates a directory for one that is not.
 Support is uneven, and the manual would rather say so than imply
 otherwise:
 
-@multitable @columnfractions .22 .26 .22 .30
+@multitable @columnfractions .22 .26 .26 .26
 @headitem Harness @tab Transcripts @tab Commands @tab Hooks
 @item claude-code @tab read @tab written @tab written (opt-in)
+@item openclaude @tab read @tab written @tab ---
 @item codex @tab read @tab written @tab ---
 @item opencode @tab --- @tab written @tab ---
+@item antigravity @tab --- @tab plugin (skills) @tab ---
 @item qwen @tab --- @tab --- @tab ---
 @item goose @tab --- @tab --- @tab ---
-@item antigravity @tab --- @tab --- @tab ---
 @item copilot-cli @tab --- @tab --- @tab ---
 @end multitable
 
+@cindex openclaude
+@cindex forks, harness
+OpenClaude is a fork of Claude Code with its own configuration root.  Its
+transcripts match Claude Code's format down to the record keys, so one
+reader serves both; the records the fork adds are recognized rather than
+counted as drift.  Its MCP registration lives in @file{~/.openclaude.json},
+the analogue of @file{~/.claude.json} --- not in
+@file{~/.openclaude/settings.json}, which holds unrelated settings.
+
+@cindex antigravity, plugin
+@cindex skills
+Antigravity is the one harness with no slash-command directory at all.
+Its extension surface is @dfn{skills}, packaged in @dfn{plugins}, and a
+plugin's own @file{commands/} directory is reported by
+@command{agy plugin validate} as ``converted to skills'' --- so a command
+there would become a skill regardless.  Engram therefore writes a plugin
+directly: a manifest plus one skill per command.  The skill frontmatter is
+a different contract from a command file's, but its description is lifted
+from the same template, so the two surfaces cannot disagree about what a
+command does.
+
 Every gap in that table is reported at runtime with its reason attached,
 rather than omitted.  @command{engram install --list} names the harnesses
 with no command surface; @command{engram ingest} fails with exit
 status@tie{}2 and an explanation when asked to read a harness it cannot.
-A feature that quietly does nothing on four of seven harnesses would read
-as broken; one that says which four, and why, is merely honest about its
-reach.
+A feature that quietly does nothing on three of eight harnesses would read
+as broken; one that says which three, and why, is merely honest about its
+reach.  Each of those three states its own reason: unsurveyed session
+storage and an absent command directory are different problems, and one
+sentence shared between them described neither.
 
 @section Formats Engram does not control
 

src/cli.rs

@@ -460,14 +460,14 @@ EXAMPLES:
         #[arg(long)]
         scope: Option<String>,
     },
-    /// Render the scope's rules into AGENTS.md and CLAUDE.md.
+    /// Render the scope's rules into AGENTS.md.
     ///
     /// Rewrites only the region between the engram sentinels, leaving the rest
     /// of each file untouched. Running it twice with unchanged rules is a
     /// no-op, so it is safe to wire into a hook or a commit gate.
     #[command(after_help = "\
 EXAMPLES:
-  # Render into AGENTS.md and CLAUDE.md at the project root.
+  # Render into AGENTS.md at the project root.
   engram rule sync
 
   # Read-only check: reports 'updated' if the block is stale or hand-edited.
@@ -481,7 +481,7 @@ EXAMPLES:
         #[arg(long)]
         scope: Option<String>,
         /// Target file, repeatable. Relative paths resolve against the project
-        /// root. Defaults to AGENTS.md and CLAUDE.md.
+        /// root. Defaults to AGENTS.md.
         #[arg(long = "file")]
         files: Vec<std::path::PathBuf>,
         /// Report what would be written without touching any file.