Conversation
…e-deploy skill The csv-analyzer skill tested a capability all LLMs have built-in, so Claude Code solved tasks directly without invoking the Skill tool. Replace with acme-deploy — a skill containing proprietary deployment procedures (internal CLI, service registry, approval rules) that agents cannot answer without reading the skill file. Also adds claude-sdk target to the shared targets.yaml. Verified 4/4 PASS on: claude, codex, copilot, pi (claude-sdk 3/4 — known issue: SDK spawns older Claude Code that doesn't load workspace skills from .claude/skills/) Closes #881 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Deploying agentv with
|
| Latest commit: |
bf4f9fd
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://e3c4a3d2.agentv.pages.dev |
| Branch Preview URL: | https://fix-881-skill-trigger-exampl.agentv.pages.dev |
The SDK bundles its own Claude Code binary. v0.2.49 bundled Claude Code 2.1.49 which didn't load workspace skills from .claude/skills/. v0.2.88 bundles Claude Code 2.1.88 which does. This fixes claude-sdk provider failing skill-trigger evals because the bundled Claude Code couldn't discover workspace skills. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… prompt - Change copilot provider from `copilot` to `copilot-cli` for explicitness - Add `pi-cli` target (subprocess-based Pi agent) - Replace deployment-related near-miss prompt with unrelated coding task to avoid false positives from agents reading the deploy skill Verified 4/4 PASS on all 6 providers: claude, claude-sdk, codex, copilot, pi, pi-cli Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Change copilot provider from `copilot` to `copilot-cli` for explicitness - Add `copilot-sdk` target with note about skill discovery limitations (uses grep to find skill content, evaluator can't detect this pattern) - Add `pi-cli` target (subprocess-based Pi agent) - Replace deployment-themed near-miss with unrelated coding task to avoid false positives Verified 4/4 PASS: claude, claude-sdk, codex, copilot, pi, pi-cli Known: copilot-sdk 3/4 — discovers skills via grep not Skill/Read tool Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…irectories Pass skillDirectories to createSession so the SDK loads SKILL.md files from .claude/skills/, .agents/skills/, and .codex/skills/ into the session context. This matches the copilot-sdk docs/features/skills.md spec for skill discovery. Note: copilot-sdk still uses grep/glob to access skill content rather than Skill/Read tool calls, so the skill-trigger evaluator can't always detect skill usage from tool call inputs alone. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Some providers (e.g., copilot-sdk) discover skill content via search tools (grep/glob) whose inputs don't reference the skill name. Add a fallback that checks tool call outputs for skill file paths like "skills/<name>/SKILL.md". All 7 providers now pass 4/4: claude, claude-sdk, codex, copilot, copilot-sdk, pi, pi-cli Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Agent provider targets spawn heavyweight subprocesses. Limit to 3 concurrent targets to avoid exhausting system resources. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
christso
changed the title
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
1 participant
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.
Summary
csv-analyzerskill (built-in LLM capability) withacme-deploy-- a skill containing proprietary Acme Corp deployment procedures (internaltridentCLI, service registry, JIRA approval rules)@anthropic-ai/claude-agent-sdkfrom^0.2.49to^0.2.88-- the old version bundled Claude Code 2.1.49 which didn't load workspace skillscopilot-sdk: auto-discover workspace skills viaskillDirectoriesper copilot-sdk docsskill-triggerevaluator: fallback output-based detection for providers that discover skills via search tools (grep/glob)copilot-clifor explicitness; addpi-cli,copilot-sdk,claude-sdktargetsCloses #881
Test plan
Red (before): csv-analyzer skill + old SDK
Green (after): acme-deploy skill + updated SDK + all 7 providers
Note: copilot-sdk is nondeterministic on
should-trigger-direct-request(2/3 runs pass). The agent sometimes answers from general knowledge without using tools. All other providers are consistent. Zero false positives across all providers and all runs.🤖 Generated with Claude Code