Orchestrator skill improvements: model pinning, billing-aware delegation, and a haiku Explore override#16
Merged
Conversation
… verifier guidance Grounded in Claude Code docs (sub-agents) and recent community findings: - Pin the model alias on every spawn: since Claude Code v2.1.198 the built-in Explore/Plan/general-purpose subagents inherit the main-session model (Explore Opus-capped on the Claude API), so un-pinned background searches bill at the orchestrator's tier; a user-level 'Explore' agent with 'model: haiku' shadows the built-in. - Prefer model aliases over pinned full IDs; drop per-agent reasoning effort to low for cheap-tier recon and mechanical work. - Distinguish what delegation buys per billing model: per-token savings on the API vs. quota-bucket arbitrage on subscriptions, where context reprocessing dominates and delegation can raise total tokens. - Close non-trivial work with a fresh-context verifier that only refutes, never fixes; add guardrails against duplicating workers' slices and against routing security work to frontier classifiers. Applied to all four harness copies and the README summary. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk
is_up_to_date built its expected content with a single newline before the marker, but the installer appends the marker with a leading blank line. Every re-run therefore reported managed files as 'Updated' and rewrote them even when current. Build the expected bytes by reproducing the exact install path (cp + marker append) instead. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk
…uidance - Start delegations at the cheapest plausible tier; after two failed attempts, escalate one tier or take the work back — never a third retry at the same tier. - Enforce read-only subagent roles via per-agent tool allowlists where the harness supports them, rather than prompt text alone. Generated skill copies regenerated with tools/generate. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk
Since Claude Code v2.1.198 the built-in Explore subagent inherits the main-session model (Opus-capped on the Claude API), so spontaneous background searches bill at the daily-driver tier. A user-level agent named Explore is the documented override; this one pins exploration to haiku at effort low with read-only tools. - Add .claude/agents/Explore.md - Teach setup to install .claude/agents/ to ~/.claude/agents/ for Claude Code (same marker/idempotency/prune flow as commands) - Document the agent, the inheritance gotcha, and the CLAUDE.md-loading trade-off in the README Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk
rlorenzo
changed the title
Greptile SummaryUpdates the orchestration guidance across supported harnesses, adds a low-effort Haiku Explore override for Claude Code, and improves setup installation idempotency.
|
| Filename | Overview |
|---|---|
| .claude/commands/efficient-orchestration.md | Revises the canonical orchestration guidance and correctly scopes Claude model aliases as a harness-specific example. |
| .codex/skills/efficient-orchestration/SKILL.md | Regenerates the Codex skill with conditional alias guidance, resolving the previously reported cross-harness issue. |
| .claude/agents/Explore.md | Adds a read-only Explore override pinned to Haiku with low effort. |
| setup | Adds installation support for Claude agents and corrects managed-file byte comparison for idempotent reruns. |
| README.md | Documents the Explore override, installation behavior, model inheritance, and context-loading trade-off. |
Reviews (2): Last reviewed commit: "refactor(orchestration): frame all Claud..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
1363 -> 898 words. Merge the overlapping Pattern / Handoff packets / Subagent stop conditions sections into one 'Run it' list, deduplicate the tier-matching rule, fold the verifier and vetting guidance together, and tighten the usage-limit and pause/resume prose. Every directive from the longer version is retained. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk
…eneric The skill body is generated verbatim into all four harness copies, so the Claude Code v2.1.198 reference read as a harness-specific instruction in the Codex/Copilot/Antigravity skills. Rephrase it as the generic principle (built-ins may inherit your main-session model) with Claude Code as the example, and drop the version number from the runtime prompt — it stays documented in the README, where the Claude-only Explore agent is described. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk
…eneric principles The skill body ships verbatim to Codex/Copilot/Antigravity, so Claude-specific mechanics must read as examples, not instructions: - Tier ladder: mark the Claude ordering as an example and restore the 'check a models command or settings panel' pointer for other harnesses. - Model pinning: 'pin an explicit model' is the universal rule; alias preference applies only where the harness has aliases (Claude does). - Usage checks: restore the generic 'usage/status command if one exists' fallback between ccusage and rate-limit-as-cap. Other skills audited: no Claude-specific references outside efficient-orchestration. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk
Owner
Author
|
@greptile-apps review The earlier alias-portability finding was addressed in 45fc19d; please re-review the latest commit. Generated by Claude Code |
There was a problem hiding this comment.
Pull request overview
This pull request updates the efficient-orchestration skill guidance across all harness copies, adds a Claude Code user-level Explore agent override to prevent unintended high-tier billing for background searches, and fixes setup idempotency so managed files don’t get rewritten (and reported as “Updated”) on every run.
Changes:
- Fix
setup’s managed-file comparison to match the installer’s exact “copy + blank line + marker” bytes, restoring true idempotency. - Add installation support for Claude Code agents (
.claude/agents/→~/.claude/agents/) and introduce a read-onlyExploreoverride pinned tohaiku/effort: low. - Refresh the
efficient-orchestrationdocs across Claude/Codex/Copilot/Antigravity variants and document the new agent behavior in the README.
Reviewed changes
Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.
Show a summary per file
| File | Description |
|---|---|
setup |
Fixes marker comparison to prevent perpetual rewrites; adds installation of .claude/agents into ~/.claude/agents. |
README.md |
Documents Claude Code agents support, the Explore override, and updates uninstall guidance. |
.claude/agents/Explore.md |
Adds a read-only Explore agent definition pinned to haiku / low effort to override built-in inheritance behavior. |
.claude/commands/efficient-orchestration.md |
Updates the Claude Code command content to the improved orchestration guidance. |
.codex/skills/efficient-orchestration/SKILL.md |
Regenerates the Codex skill copy with the updated orchestration guidance. |
.copilot/skills/efficient-orchestration/SKILL.md |
Regenerates the Copilot skill copy with the updated orchestration guidance. |
.antigravity/skills/efficient-orchestration/SKILL.md |
Regenerates the Antigravity skill copy with the updated orchestration guidance. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
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
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
Improvements to the
efficient-orchestrationskill grounded in the official Claude Code sub-agents docs and recent community measurements, plus a new Claude CodeExploreagent override and asetupidempotency fix found while testing.Skill updates (all four harness copies, regenerated via
tools/generate)haiku/sonnet/opus) over full model IDs, and use per-agenteffort: lowfor recon and mechanical work.New: Claude Code
Exploreagent override.claude/agents/Explore.md—model: haiku,effort: low, read-only tools. A user-level agent namedExploreis the documented way to shadow the built-in and undo the v2.1.198 inheritance change for spontaneous searches.setupgains a Claude-only install target:.claude/agents/→~/.claude/agents/, reusing the existinginstall_commandsflow (marker, prompts, idempotent updates, stale pruning).Fix:
setupfalse "Updated" on every runis_up_to_dateexpected a single newline before the source marker while the installer appends a blank line, so every re-run reported all managed files as "Updated" and rewrote them. The comparison now reproduces the exact bytes the install path writes.Testing
shellcheck -x setupclean;markdownlint-cli2clean;tools/generate --checkin sync; all 46 bats tests pass../setuprun (stubbedclaude/gh, throwaway$HOME): fresh install → idempotent re-run shows "up to date" → hand-edited file shows "Updated" once → "up to date" again.🤖 Generated with Claude Code
https://claude.ai/code/session_019H18Vt8ToMYzD9gX64nKnk