Route Claude 5 models by role and review effort - #3
Conversation
|
@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: 03aaeab6ef
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".
| plus eligible independent review/challenge; Fable 5 = Claude-side interactive | ||
| chief for architecture/risk/RCA, never routine dispatch/review-throughput. |
There was a problem hiding this comment.
When Claude is used for architecture, risk, or RCA, this assigns Fable chief authority even though the manifesto reserves design and red-line decisions for Sol, §2 permits only one active chief, and the §5 routing table assigns architecture/risk/convergence exclusively to the Sol T3 role. This can create two competing decision owners or replace Sol without the required phase handoff; keep Fable in a bounded advisory/challenge role rather than calling it a chief.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 42f2f3a. The routing now makes the boundary surface-specific: Fable remains chief on the Claude surface, but has no chief role in a Codex run; its advice/challenge is non-binding, Sol remains the sole active chief and decision owner for that run, and Fable is never an executor or review-throughput target. The Claude twin was intentionally unchanged because its Fable-chief semantics remain correct on the Claude surface.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 42f2f3a941
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".
| disabled; include any auto-loaded instruction files in authorized context. | ||
| Read/Grep/Glob access stays within the authorized chief-bound impact cone. | ||
| Invoke | ||
| `claude -p --model claude-opus-5 --effort low --allowedTools "Read,Grep,Glob" --output-format json < /absolute/path/to/review-prompt.txt` |
There was a problem hiding this comment.
In Claude CLI, --allowedTools permits listed tools without prompting; it does not restrict the process to those tools. Consequently, this headless review can still expose tools such as web access or subagent delegation, contradicting the preceding no-web/no-subagents boundary and potentially sending data beyond the authorized impact cone. Use the CLI's tool-availability restriction (for example, --tools "Read,Grep,Glob") in addition to the appropriate permission configuration.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 7ab7312. The canonical invocation now pairs --tools "Read,Grep,Glob" (available built-ins) with --allowedTools "Read,Grep,Glob", adds --safe-mode and --no-session-persistence, and keeps web/subagents/MCP/write tools unavailable. The policy also distinguishes flag-enforced tool identity from prompt-enforced path scope; an out-of-cone read invalidates the review evidence.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7ab731234c
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".
| failure. If no authorized provider is available, defer. For an | ||
| eligible Claude lane, use `claude-opus-5` first: run the cheap/fast first pass |
There was a problem hiding this comment.
When an authorized Claude lane is a mechanical sweep or bounded implementation, this instruction routes it to Opus 5 and then imposes a read-only review invocation, contradicting the Haiku/Sonnet assignments immediately above. The README's equivalent rule is correctly limited to “an eligible Claude review”; qualify this as a review/challenge lane so workers are not silently promoted to Opus or prevented from writing.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in 7608f12. The Opus-first read-only invocation now applies only to an eligible independent Claude review/challenge lane; Haiku mechanical sweeps and Sonnet bounded implementation/tests/debugging retain their execution routing. The same scope is mirrored in the installed skill and global policy.
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7608f12e69
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".
| prevents resume. These flags enforce tool identity, not path scope; prompt/ | ||
| brief policy binds Read/Grep/Glob to the authorized chief-bound impact cone, | ||
| and any out-of-cone read invalidates the review evidence. Invoke | ||
| `claude -p --model claude-opus-5 --effort low --tools "Read,Grep,Glob" --allowedTools "Read,Grep,Glob" --no-session-persistence --safe-mode --output-format json < /absolute/path/to/review-prompt.txt` |
There was a problem hiding this comment.
When Anthropic is authorized for only the named impact cone, this invocation still grants Read, Grep, and Glob access to the entire working tree and relies solely on prompt compliance for path isolation. An accidental out-of-cone read can therefore disclose prohibited repository data before the evidence is invalidated, and the command provides no enforcement boundary. Run the reviewer against a filesystem projection or sandbox containing only authorized paths instead of treating prompt policy as access control.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in ee29c37. The default review path is now tool-less stdin (--tools "") and contains only the minimum authorized/redacted question, diff/context, and evidence contract, so it has no repository file access. Read/Grep/Glob may be enabled only inside an OS/filesystem sandbox or projection exposing the authorized cone, with paired tool flags. The policy explicitly rejects prompt-only path restrictions as access control.
|
@codex review |
|
Codex Review: Didn't find any major issues. What shall we delve into next? Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
|
@codex review |
|
Codex Review: Didn't find any major issues. More of your lovely PRs please. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
1 participant
What changed
bounded implementation/tests/debugging, Opus 5 for senior execution and
eligible independent review, and Fable 5 as the Claude-side interactive
chief.
provide only non-binding architecture/risk/RCA advice or a named,
non-binding challenge to Sol.
broader focused cone and high only when findings or semantic risk require a
thorough pass.
authorized/redacted question, diff/context, and evidence contract.
the authorized cone; prompt-only path restrictions are not access control.
review/challenge lane.
authorized review/challenge-only fallback.
code, design, or contract decisions; the focused Terra review and targeted
cross-model challenge remain non-shareable lanes.
Why
The previous policy pinned Claude code/review work to Opus 4.8. Anthropic now
positions Opus 5 as the successor at the same API price, with stronger
code-review precision/recall and useful low-effort accuracy. The routing now
uses that improvement without making every review a high-effort pass.
Official references:
Validation
git diff --checkpython3 -m unittest discover -s tests -v— 20/20 passedReview record
7878bf9ee29c37surface ownership, Claude review isolation, and public README summary; no
rollover or implementation behavior changed.
invocation, and evidence-scope gaps were corrected.
author/reviewer independence, lane collapse, Fable/Sol surface ownership,
review-lane scoping, and enforceable filesystem isolation.
03aaeab: Fable/Sol ownership ambiguity, fixed in42f2f3a.42f2f3a:--allowedToolsdid not restrict available tools,fixed in
7ab7312.7ab7312: Opus-first wording captured execution lanes, fixedin
7608f12.7608f12: prompt-only path bounds could expose unauthorizedfiles, fixed with a tool-less default/sandboxed-browse exception in
ee29c37.0d865e93df6d0b06f6895fc5a6883e0fe923e3cdcd3164adba13acaaddea7081.