engine-sync 0.4.0

Added

• Plugin errors, warnings and meaningful info are bounded JSON rows in errors.jsonl, warnings.jsonl and info.jsonl under $PLUGIN_LOG_DIR/engine-sync (default ~/.local/state/engine-sync/logs). Capped files move to .1, and a failed logging streak reports once on stderr without changing command results.

Fixed

• A plain release check compares version places without comparing binary checksums; --release still rebuilds all four binaries and verifies the committed sums.

• Publication without supplied private leak terms reports GAP, not success. Permission fixtures remain meaningful when tests run as root.

• Newly introduced tree checks accept an empty baseline against an older release tree that lacks it, while rejecting new exemptions and failed Git enumeration.

• Failed project check and doctor commands write error rows with their diagnostic cause and session.

• The build-warning plugin-log row names its warnings in detail.warnings beside detail.count, instead of only counting them. A set too long for the 4096-byte row keeps as many whole warnings as fit, cuts the next one, and says so in detail.truncated, detail.omitted (warnings not shown) and detail.cut; the row is never dropped for its length.

• A machine-global OpenCode agent keeps the mcp__{server}__{tool} grants of the servers the machine config names in mcp.globalServers: the machine path passed no known servers, so every such grant was dropped as an unknown server. A configured server name renders in OpenCode's tool-part charset (letters, digits, _, -; any other character becomes _), as a project server's does, so a dotted name grants and no configured name reaches an agent file raw. A tool of a server named nowhere, or of a retired one, is named in a warning and its server is denied outright ({server}_*: false): OpenCode enables a tool an agent's map leaves out. This changes the bytes of a machine-global OpenCode agent: each one now denies every configured global server it does not grant ({server}_*: false).

Changed

• Global configuration now lives in ~/.claude/engine-sync.json, with machine destinations under its targets object. Move keys from the former config and targets files into that file; an old file without its replacement is refused visibly, including under an absolute XDG_CONFIG_HOME.
• Generated declarations require machine.sourceGeneratedAgentsDir; the home generated-agent store and its machine.generatedAgentsDir and machine.legacyCodexAgentsDir keys are retired. Shared files likewise require the source generated directory. Machine-only declaration controls in project agent files are refused.
• Optional machine.managedBy stamps machine Codex roles and prompts and OpenCode agents and commands with managed-by: {installer} immediately after the generated header. This changes those output bytes when configured; machine doctor fails with ORPHAN for stamped files whose source is gone. Project mirrors and rendered Claude declarations remain unstamped.
• Command references use chat call and the current chat and harvester MCP server names; configured rewrites determine the corresponding mirror bytes.

engine-sync 0.3.0

Added

• Agent frontmatter carries per-engine settings in an engines: block (Codex model, effort and sandbox; OpenCode model, mode and permission; false skips an engine). One YAML node parser serves every source kind; a header failing strict YAML is retried once with single-line values read as literal strings, and one that still fails is FRONTMATTER-INVALID {file}:{line}, skipped by name.
• Derived machine agents are declaration files: a machine source carrying from: declares one rendered agent whose Claude-native keys and ## Heading body sections replace the base's; retitle, pins, add, unset, rename, prepend and append shape it, and variants pin rewrites a file's pins: block in place. A declaration extends a shared .md below a subdirectory by relative path; a shared file nothing extends is reported SHARED-UNUSED.
• Machine agents are capability-gated: requires is checked against --capability and the targets.json capabilities list, and machine GPT agents install only when asked for.
• machine.sourceGeneratedAgentsDir (source-repository-relative, empty by default) renders declarations beside their sources instead of the home store; links point there, global retires what it owned in the home store, and --uninstall never removes the repository directory.
• Machine doctor is closed-world: INSTALLED, MISSING, STALE, SHADOWED, SKIPPED, BROKEN, STRAY, and a failing ORPHAN row for each rendered file whose declaration is gone. global --adopt moves shadowing hand files to the adopted store; one broken agent is skipped by name and keeps its installed copy.

Changed

• Rendered machine agents carry a role stamp and a mark naming engine-sync global; OpenCode global agents are written under agents/. Project Codex role names are the frontmatter name verbatim; targets.json codexHomes is optional.
• check and doctor see the whole Claude roster for the source collision; check skips a directory holding no declaration. A declaration inside a Claude agents directory engine-sync installs into is a named problem and is never replaced.

Removed

• --gpt-agents: use --capability. Flat codex-* and opencode-* agent keys, the machine.variants carrier (variants.json), machine.modelMap, opencode.agent and abstract: are refused, each naming its replacement: port each variants.json entry to its own {machine.sourceAgentsDir}/{name}.md declaration and each abstract to a shared file it extends.

Fixed

• global --adopt change events carry no path, as docs/interface.md's JSON contract says.
• Reconciliation retires only assets engine-sync owns; a registry link with no receipt that points at an engine-sync-generated file is recognized as ours, so a moved render repoints it, and a moved home store is retired only after the registry links repoint, so no link dangles into it.

engine-sync 0.2.0

Added

• One cached source model, rewrite chain, target registry and artifact reconciler serve project and machine scopes. Added plan, --engine, build-only --dry-run, config show|validate|schema, JSON Schema, permanent goldens and mechanical tree gates with pinned staticcheck and Claude Code validation.
• Configurable machine source/marker directories, command rosters and Codex review model, effort and scope prompts; the standalone defaults contain no integrator protocol or layout. Project and machine policy report their provenance.

Changed

• Unmapped Codex models warn and omit the model pin; configure a model map to retain a pin. The historical agent-name sandbox exemption is removed: use explicit codex-sandbox: inherit when write access is intended. Removing .mcp.json removes its managed MCP fence while preserving hand-written keeper content.
• Layered OpenCode permission patterns retain denies from either layer; a deny wins equal-pattern conflicts and more specific patterns follow broad rules. Missing policies and empty rule sets deny access. Project bash entries no longer erase machine git-write denies.
• Codex keeper MCP regions retain the unbranded legacy script fence syntax readable by 0.1.0 on a fresh machine; the immediately preceding development fence remains accepted for migration.
• Download locks recover dead owners immediately and expire after 120 seconds. Unavailable hooks save a project-scoped deferred refresh; the next equipped hook in that project runs SessionStart recovery. Hook children are bounded and Stop notices include busy-lock details.
• Development builds report a prerelease of the next version (0.3.0-alpha once 0.2.0 is released); the version gate requires a prerelease newer than the plugin manifest. Release builds and four native CI jobs execute assets to verify the reported version matches their names.
• Neutral model maps, adapters, agent preambles, MCP knowledge, retirement lists and generated-store defaults replace integrator-specific defaults. Configure previous values to retain their bytes; changed defaults intentionally alter unconfigured mirrors.
• Global commands render only from machine sources and policy, under machine guards, including during project builds. One resolver selects explicit homes, CODEX_HOME, targets, configured homes and defaults; an explicit empty roster disables Codex homes.
• Configured Go agent-preamble templates now interpolate fields; legacy root adapters retain literal shell-style placeholders. Project command references also use the configured machine roster, so known global slash commands translate in project bodies. Dotted filenames are preserved instead of being mistaken for commands.
• Native OpenCode model pins and reserved permission keys are quoted; absent policies and empty permission rule sets deny access. Agent names, suffixes and exclusions are shared across engines, with full Unicode case/NFC collision checks and path validation. Codex effort aliases map max/ultra to xhigh; TOML control characters are escaped.
• Codex model comments say model instead of tier. Unresolvable-source notices say output kept. Generated markers and rebuild hints name engine-sync; JSONC markers identify configuration and MCP sources. Complete historical headers remain recognized.
• Native ordered OpenCode permissions and agent overrides come from configuration. Lower bare actions survive as leading wildcard rules, JSONC preserves user keys, and a missing command description warns per file. BOM, CRLF and spaced frontmatter fences are parsed; malformed or non-UTF-8 sources fail visibly.
• Generated ownership requires a complete header or a recorded link receipt. Unproven copies and hand-made aliases are preserved, including historical links whose location alone does not prove ownership. Garbage collection preserves foreign children, scratch files and unresolvable outputs. New parents use 0755; keeper modes and contained keeper links survive updates. Replacing changed generated content emits a notice.
• Source reads are limited to 8 MiB and ownership reads to a header; special files fail before open. Canonical bounded roots, ancestor checks, bounded hook locks and partial-write records prevent escapes and repeated obsolete writes.
• The wrapper rechecks cached checksums, backs off downloads, serializes stable-link changes, preserves stdin, rejects invalid Stop JSON and always exits zero for hooks. Dirty flags remain durable throughout Stop stages and concurrent edits survive flag cleanup.

Fixed

• Multi-home retargeting, machine doctor/installer disagreement, silent write/config/discovery errors, invalid dry-run success, contradictory usage and ownership diagnostics, output collisions, template failures, and release/gate producer errors. Baselines may only shrink, including ordinary CI checks; empty enumerations and missing publication inputs report ERROR.
• Dead duplicate CLI/config/reconcile adapters and their tests were removed; surviving OpenCode CLI tests use the production route. Documentation now states current scope, recovery, byte changes and the checks actually enforced.

engine-sync 0.1.0

Added

• The engine-sync binary, extracted from pfm's codex and opencode compilers with their history: codex build|check|agents, opencode build|check|doctor, build, check and doctor over both engines, doctor --machine, global for the machine-level Codex roles, commands and Claude agent links, and hook claude mark|stop|start. Output is byte-identical to pfm's.
• A version guard: a binary refuses a project or Codex home a newer engine-sync last wrote (NEWER-WRITER), and writes to one root are serialized by a lock.
• The engine-sync Claude Code plugin: a PostToolUse hook marks a turn that wrote a Claude source, Stop rebuilds and checks both mirrors and blocks on a failing stage, SessionStart rebuilds a stale mirror and refreshes the machine targets targets.json names. Its wrapper resolves $ENGINE_SYNC_BIN, then the cached binary, then a checksum-verified release download, and keeps a stable bin/engine-sync link that only moves forward.