Repository navigation
feat(crossreview): zcode reviewer - #2
Conversation
ZCode ships its agent CLI inside the desktop app (resources/glm/zcode.cjs run under Electron's Node) and has no standalone binary of its own, so the template drives a `zcode` launcher on PATH that forwards to the bundle. The prompt travels as an argument (--attach takes media only), which turns the koda exceptions into ARG_AGENTS; --mode plan is the read-only mode, and the model comes from ~/.zcode/v2/provider_config.json since there is no --model flag. Version 2.1.0.
The signed-in probe reaches the model call and dies on Z.ai's
POST /api/auth/z/login ("Z.ai user information is invalid", identical for a
valid OAuth token, direct or proxied); document the failure mode and keep
the status checked, not verified.
…API-key setup The signed-in probe now reaches the model API and the server answers "[1113][Insufficient balance or no resource package. Please recharge.]" with HTTP 429, which the hints called a rate limit. Give it its own hint, and document the personal API-key provider (personalModelIds, not builtinModelIds) as the zcode setup that bypasses the account's broken coding-plan key exchange.
The reviewer's own first review of this branch: README still called Koda the only brief-as-argument exception, and the expected zcode blocker (project_token_login_failed) had no hint. Verified both against the code and fixed; the one-word probe (9s) and a real 19 KB review (4m29s) through the template make zcode Verified on 0.17.0.
Two reviewers approved with changes; verified and fixed all five findings: the macOS bundle layout was unmarked guesswork, the argument-limit warning suppressed the template-drift check for ARG_AGENTS, nothing pinned --mode plan in the zcode row, entry_warnings had no test, and desktop updates silently break the launcher setup. Rejected the doubts about the marker, the argv substitution, the models column and stale versions as already covered.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
Kept both reviewers and main's takes_brief_as_argument predicate in place of my ARG_AGENTS set, with the template-drift check still keyed on the known-issue warnings only. Version 2.2.0: main released 2.1.0 for Antigravity. Main's antigravity tests were green only on POSIX; fixed their Windows legs so the suite runs everywhere: the detection assertion follows the platform's template column, the plain-command test accepts PowerShell 5.1's UTF-16 redirects and CRLF argument handover, and the run tests drive a real executable instead of a .cmd shim, which the runner refuses to feed a brief.
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 12cd69f771
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| qwen qwen qwen - {bin} -m {model} --approval-mode plan -o text < {brief} > {out} cmd /d /c --% {bin} -m {model} --approval-mode plan -o text < {brief} > {out} | ||
| kimi kimi kimi - {bin} --quiet --plan -m {model} < {brief} > {out} cmd /d /c --% {bin} --quiet --plan -m {model} < {brief} > {out} | ||
| antigravity agy agy models {bin} --model {model} --mode plan -p "$(cat {brief})" > {out} {bin} --model {model} --mode plan -p "$(Get-Content -Raw -Encoding UTF8 {brief})" > {out} | ||
| zcode zcode bigmodel - {bin} --prompt "$(cat {brief})" --mode plan --no-color > {out} {bin} --prompt "$(Get-Content -Raw -Encoding UTF8 {brief})" --mode plan --no-color > {out} |
There was a problem hiding this comment.
When this template is executed verbatim in Windows PowerShell 5.1 through the documented no-Python workflow (SKILL.md:225-234), the bare > transcodes ZCode's stdout into UTF-16LE instead of preserving its UTF-8 bytes. The revised test at tests/test_crossreview.py:406-409 explicitly tolerates that conversion, but UTF-8 consumers receive NUL-filled or corrupted review text. The Python runner's argv bypass avoids this problem; literal template execution does not. Keep stdout byte-preserving in this path too: AGENTS.md requires the PowerShell column to hand redirection to cmd.exe specifically to prevent this re-encoding.
AGENTS.md reference: AGENTS.md:L50-L52
Useful? React with 👍 / 👎.
12cd69f to
7770cf0
Compare
feat(crossreview): zcode reviewer
feat(crossreview): zcode reviewer
What
Add zcode — the agent CLI bundled inside the ZCode desktop app (
resources/glm/zcode.cjs, run under Electron's Node) — as a reviewer:scripts/agents.tsv: the zcode row,zcode --prompt "$(cat {brief})" --mode plan --no-color. The prompt travels as an argument (like koda and antigravity):--attachaccepts media only and stdin is not a prompt.scripts/crossreview.py: argument-brief agents are recognised from the template shape (takes_brief_as_argument); hints for an empty API balance and for the ZCode token-exchange failure; the template-drift check is no longer suppressed for argument-brief agents.references/agents.md: a zcode section — thezcodelauncher recipe (POSIX shim; Windows needs a real.exe, source included), the three setup steps the desktop hides (the bundle'sprovider/zcode-builtin.json,zcode login, a personal API-key provider in~/.zcode/v2/provider_config.json), pitfalls and recovery commands.tests/: table rules for argument-brief agents,--mode planpinned in the zcode templates, anentry_warningsdrift case; the merged antigravity tests got their Windows legs fixed (PowerShell 5.1 redirect encoding, the runner's batch-shim refusal).Verification
project_token_login_failed) even after a successfulzcode login; the personal API-key provider bypasses it.