fix(examples): make example 61's CLI usage deterministic - #166
Merged
Conversation
ling-senpeng13
force-pushed
the
fix/example-61-deterministic-cli
branch
2 times, most recently
from
August 18, 2026 22:54
9185c44 to
aaf97ce
Compare
ling-senpeng13
marked this pull request as ready for review
August 18, 2026 22:57
The chained GitHub pipeline failed in a retry loop: models wrapped compound commands in 'bash -c' (rejected — bash not allowlisted), the instructed step-3 command itself started with a VAR=$(...) assignment (also rejected), the coder had no file-writing command in its allowlist, and models invented placeholder cwd paths. Add shared TOOL RULES to every CLI stage (no shell wrappers, first word must be allowlisted, no placeholder cwd), rewrite the instructed commands to be validator-legal, and allowlist echo/printf for the coder.
ling-senpeng13
force-pushed
the
fix/example-61-deterministic-cli
branch
from
August 19, 2026 01:01
aaf97ce to
f2af89d
Compare
kowser-orkes
approved these changes
Aug 20, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Example 61's pipeline failed in a retry spiral: models wrapped compound commands in
bash -c(rejected — not allowlisted), the instructed step-3 command itself started withTMPDIR=$(...)(also rejected by first-token validation), the coder had no file-writing command, and models invented placeholdercwdpaths. The repeated rejections tripped the circuit breaker and Conductor's 256 KB workflow-variables limit.Changes (prompting-only, no SDK code):
TOOL RULESblock for all four CLI stages: pass compounds directly withshell=true, neverbash -c/sh -c, first word must be allowlisted, never inventcwdmktemp -das its own step; compounds lead withgh)echo/printffor the coder,catfor the read-only stagesConsidered and rejected: allowlisting
bash. It would eliminate the wrapper rejections, but validation is first-token-only, sobash -c '<anything>'makes the allowlist decorative — and this example exists partly to demonstrate the allowlist as a guardrail. The better long-term fix is SDK-side: unwrapbash -cand validate the inner command.Verified: repeated runs with zero command rejections (previously 10+ per run); one full end-to-end success (issue → commit → real PR on the sandbox repo). Note: deep swarm-stage reliability on Anthropic models also depends on a server-side history fix, outside this PR's scope.