feat: add zcode agent preset and hook installer - #2166
Conversation
|
pengzhou seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account. You have signed the CLA already but the status is still pending? Let us recheck it. |
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
fa81850 to
dd1a158
Compare
| // Keep user hooks, drop any duplicate git-ai entries, then ensure | ||
| // exactly one git-ai zcode hook remains (migrates an existing | ||
| // claude-preset hook to zcode). | ||
| let mut hooks_array = blocks[block_idx] | ||
| .get("hooks") | ||
| .and_then(|h| h.as_array()) | ||
| .cloned() | ||
| .unwrap_or_default(); | ||
|
|
||
| hooks_array.retain(|hook| !Self::is_git_ai_process_hook(hook)); | ||
| hooks_array.push(Self::desired_hook(params)); | ||
|
|
||
| if let Some(block) = blocks[block_idx].as_object_mut() { | ||
| block.insert("hooks".to_string(), Value::Array(hooks_array)); |
There was a problem hiding this comment.
🟡 Old git-ai hook can be left behind when it sits under a different matcher, causing duplicate checkpoints
When updating the ZCode hook config, existing git-ai checkpoint entries are only cleaned out of the block whose matcher exactly equals Edit|Write|Bash (hooks_array.retain(...) at src/mdm/agents/zcode.rs:148-155), so an older entry stored under any other matcher stays and a second entry is added alongside it.
Impact: A user whose ZCode config used a different tool matcher ends up with the same edit reported twice, producing duplicate/misattributed sessions.
Mechanism: matcher-scoped dedup vs. config-wide detection
install_hooks_at locates (or creates) the block with matcher == ZCODE_MATCHER (src/mdm/agents/zcode.rs:129-143) and only strips git-ai hooks from that block's hooks array before pushing the desired hook. A pre-existing git-ai hook under e.g. matcher "*" or "Write|Edit|Bash" is untouched, so both fire on every Write/Edit/Bash tool call — one as checkpoint claude, one as checkpoint zcode.
Note the asymmetry: hook_status (src/mdm/agents/zcode.rs:51-77) scans every block, so such a config is reported as installed/up-to-date, while uninstall_hooks_at (src/mdm/agents/zcode.rs:207-223) does remove git-ai hooks from every block. The Droid installer handles exactly this case by stripping git-ai from all non-target matcher blocks before inserting into the canonical block (src/mdm/agents/droid.rs:138-166).
Was this helpful? React with 👍 or 👎 to provide feedback.
136fa22 to
ad035e5
Compare
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
span: checkpoint preset + install-hooks fid: none level: none test: pass influence: 新增 zcode agent 一等公民支持：git-ai checkpoint zcode 解析 ZCode hook payload（协议与 Claude Code 一致，复用 ClaudeJsonl transcript），git ai install-hooks 将 process 型 checkpoint hook 写入 ~/.zcode/cli/config.json（保留 mcp/plugins 等既有配置）
ad035e5 to
a211c95
Compare
| let commit = repo.stage_all_and_commit("Add AI lines").unwrap(); | ||
|
|
||
| // Dump the checkpoint debug log to stderr so CI failures show exactly what | ||
| // the checkpoint command emitted (preset_name / event_count / requests). | ||
| let log_dir = repo | ||
| .test_home_path() | ||
| .join(".git-ai") | ||
| .join("internal") | ||
| .join("checkpoint-debug-logs"); | ||
| if let Ok(entries) = fs::read_dir(&log_dir) { | ||
| for entry in entries.flatten() { | ||
| if let Ok(content) = fs::read_to_string(entry.path()) { | ||
| eprintln!( | ||
| "=== checkpoint debug log: {} ===\n{}", | ||
| entry.path().display(), | ||
| content | ||
| ); | ||
| } | ||
| } | ||
| } | ||
|
|
||
| // We should have exactly one session record keyed by the zcode agent_id | ||
| assert_eq!( | ||
| commit.authorship_log.metadata.sessions.len(), | ||
| 1, | ||
| "Expected a single session record" | ||
| ); | ||
| let session_record = commit | ||
| .authorship_log | ||
| .metadata | ||
| .sessions | ||
| .values() | ||
| .next() | ||
| .expect("Session record should exist"); | ||
|
|
||
| assert_eq!(session_record.agent_id.tool, "zcode"); | ||
| // Model is extracted from the real transcript fixture copied in the second checkpoint | ||
| assert_eq!( | ||
| session_record.agent_id.model, "claude-sonnet-4-20250514", | ||
| "Session record model should come from the latest checkpoint's transcript" | ||
| ); |
There was a problem hiding this comment.
🟡 New end-to-end test never checks that AI lines are actually credited to the AI
The new end-to-end test commits twice (stage_all_and_commit("Add AI lines") at tests/integration/zcode.rs:56) but never asserts the per-line credit for the committed file, which the repository's mandatory testing rules require after every commit.
Impact: The test can pass even if the newly added agent credits its edits to a human or leaves them uncredited, so a real attribution regression would ship unnoticed.
AGENTS.md mandates line-level assertions after every commit
AGENTS.md states: "ALWAYS use the helper to assert the lines post-commit AND make sure to always assert line-level after EVERY commit for EVERY test you EVER right. This is CRUCIAL." and lists assert_committed_lines / assert_lines_and_blame as the helpers.
tests/integration/zcode.rs:8-97 performs repo.stage_all_and_commit("Initial commit") (tests/integration/zcode.rs:26) and repo.stage_all_and_commit("Add AI lines") (tests/integration/zcode.rs:56), then only asserts on commit.authorship_log.metadata.sessions (session count, agent_id.tool, agent_id.model). No assert_committed_lines/assert_lines_and_blame call exists anywhere in the file, so the actual AI-vs-human line attribution produced by the new zcode preset is never validated.
A fix would add repo.filename("src/main.rs").assert_committed_lines(...) after each commit, expecting "fn main() {}".unattributed_human() for the pre-existing line and .ai() for the two lines written between the checkpoints.
Prompt for agents
AGENTS.md requires that every integration test assert line-level attribution after EVERY commit using the TestRepo helpers (assert_committed_lines / assert_lines_and_blame). The new test test_zcode_e2e_prefers_latest_checkpoint_for_prompts in tests/integration/zcode.rs performs two commits ("Initial commit" and "Add AI lines") but only asserts on the authorship log's session metadata (session count, agent_id.tool, agent_id.model). It never verifies that the lines written between the zcode pre/post checkpoints are actually attributed to AI, which is the core behavior the new preset is supposed to deliver. Add line-level assertions after each commit: after "Initial commit", src/main.rs should have its single line as unattributed_human; after "Add AI lines", the two added comment lines should be attributed to AI. Use repo.filename("src/main.rs") and the lines! macro with .ai() / .unattributed_human() as described in AGENTS.md.
Was this helpful? React with 👍 or 👎 to provide feedback.
|
All Devin Review findings have been addressed in the latest commit (a211c95):
Fork CI is green: Lint & Format (fmt/clippy/doc), Test core on ubuntu/windows/macos, and both Ubuntu integration shards. |
2 participants
Summary
Add first-class support for ZCode (a Claude Code-compatible CLI coding agent) to git-ai:
git-ai checkpoint zcodeparses ZCode's hook payload. ZCode's hook protocol is identical to Claude Code's (same JSON shape, capitalizedWrite/Edit/Bashtool names, Claude-format transcript), so the new preset mirrors the existing Claude preset and attributes work tozcodeinstead ofclaude.git ai install-hooksnow supports ZCode via a new installer that writes the process-style checkpoint hook into~/.zcode/cli/config.json(hooks.events.PreToolUse/PostToolUse, matcherEdit|Write|Bash,type: "process"with checkpoint args in theargsarray). It preserves existing top-level config keys (mcp,plugins, ...) and migrates an existingcheckpoint claudehook tocheckpoint zcode.Changes
src/commands/checkpoint_agent/presets/zcode.rs— newZcodePreset(mirrorsclaude.rs, agent idzcode) + unit testssrc/commands/checkpoint_agent/presets/mod.rs— register thezcodepresetsrc/mdm/agents/zcode.rs— newZcodeInstaller(writes/updates/removes the ZCode hook config) + unit testssrc/mdm/agents/mod.rs— registerZcodeInstallerTesting
Notes
~/.zcode/cli/config.jsonand uses a different hook schema than Claude Code (processtype +argsarray), which is why a dedicated installer is needed.