Skip to content

refactor(skills): consolidate Genie into 14 focused workflows - #2902

Merged
automagik-genie merged 3 commits into
devfrom
nmstx/skills-lightweight-body
Sep 15, 2026
Merged

automagik-genie merged 3 commits into
devfrom
nmstx/skills-lightweight-body

Conversation

@automagik-genie

Copy link
Copy Markdown
Contributor

What changes

Genie shipped 25 overlapping skills with repeated orchestration policy, old model-tuning instructions, and obsolete resources. This reduces the catalog to 14 distinct workflows, with domain audits loaded on demand and Orca-specific execution guidance attached to the core lifecycle.

refine accepts only --for openai and --for claude. The switches select prompting guidance, never a runtime model. The skill and both concise provider guides link the official GPT-6 Astra prompting guidance, OpenAI prompt engineering, Claude Fable 5.1 guide, and Claude best practices, checked 2026-09-15. Invalid/model-specific selectors stop before reading or writing input.

The shared Astra/Fable agreement is to retain concrete intent, boundaries, runtime contracts, independent review, and required evidence; remove duplicated policy, fixed rituals, decorative personas, and guidance that the current runtime already supplies. A Claude Fable 5.1 coworker launched through Orca assessed every skill/resource, agreed the consolidation, implemented a disjoint scope, and independently reviewed the Astra-owned changes. Astra reviewed the Fable-owned changes. The agreement itself was refined through the refine workflow.

Across the entire shipped skills/ tree (entrypoints, metadata, templates, and references), this reduces 42,811 words on dev to 16,263 and 291,224 bytes to 115,582 (about 62% less text). The original 25 entrypoints become 14; 72 files become 47.

Every original skill

Original Disposition
brainstorm Keep design exploration and digest-backed independent review; replace readiness scores and persona cards with concrete decisions.
wish Keep canonical executable-plan template/schema; share it with conditional Orca planning.
work Keep dependency execution, ownership, reviews and delivery; load Orca coordination only when selected.
review Keep independent artifact/PR review; absorb seven focused audit lenses.
fix Keep the single repair-budget and cause-based escalation contract.
genie Keep concise routing and persisted-state resume; ordinary requests bypass planning.
refine Keep provider-guided prompt rewriting with exactly two switches and safe input/output boundaries.
report Keep diagnosis as the default; create an issue only when requested.
docs Keep audit/author distinction and documentation/DX verification.
council Keep independent scrutiny of consequential decisions with relevant lenses.
dream Keep approved-wish batch orchestration and verified merge/QA reporting.
quick Keep the one-hour dev-delivery contract and repository-required independent review.
omni Keep the distinct channel-wiring flow, explicit route provider and real round-trip verification.
genie-hacks Keep registry lookup/contribution; remove stale recipes and unsafe cache cleanup.
architecture Merge into review architecture lens.
code-quality Merge into review code-quality lens.
dx-docs Merge into review DX lens and docs.
perf Merge into review performance lens.
qa Merge into review test-quality lens.
repo-hygiene Merge into review repository-hygiene lens.
supply-chain Merge into review security/supply-chain lens.
trace Merge into report investigation.
genie-orca-wish Merge into wish Orca mode.
genie-orca-work Merge into work Orca mode/reference.
genie-orca-review Merge into review Orca mode.

Resource review also removes duplicate native-surface guidance, brainstorming persona cards, obsolete Orca migration/retro scripts, and the historical fixed-percentage skills:audit script. Machine-consumed wish fields and both self-contained design-digest helpers remain. Metadata and distribution tests derive their inventory from the shipped tree.

Upgrade behavior

The pinned skills installer does not remove retired skill names. After replacement skills are verified, genie update now atomically moves previously recorded, unchanged retired directories into an owner-only GENIE_HOME/state-backups/skills-retirement-* tree before publishing the new install record. Edited, unverified, symlinked/redirected, and outside-home copies are preserved with a manual-review notice. An unverified replacement set in one home preserves its retired copies for manual review while other homes can converge. No verified replacement anywhere, or an archive failure, retains the previous record for retry. Cross-filesystem archive moves fail with the original intact; manual skills.sh installs without a Genie record require manual review of old names.

Validation

  • bun install --frozen-lockfile — passes; 293 installs / 364 packages, unchanged lockfile.
  • bun run check — all static gates pass; the final full suite reports 2,049 pass, 1 skip, and two five-second timeouts in unchanged tests (plus the timeout-killed process error). Both timed-out tests pass on a normal-settings targeted retry in 1.28 seconds total. The preceding full run passed 2,048 tests before the final three added regression cases; all 58 current installer tests pass (typecheck, Biome, Knip, skill/wish/complexity/bundle gates and full tests).
  • Focused final review checks — 58 installer tests and 41 release-docs tests pass; design evidence and wish validation tests pass.
  • npx -y skills@1.5.23 add "$PWD" --list plus inventory parity — exactly the 14 shipped skills.
  • bash scripts/build-binary.sh --platform linux-x64-glibc — 34 MB tarball; source, staging and extracted-payload smoke/version checks pass. All 47 skill/resource files match the reviewed source bytes; compiled CLI reports 5.260915.1 from a neutral directory.
  • Independent reviews: Fable reviewed the core skills, provider prompts, installer and integration changes; Astra reviewed Fable's remaining skills/resources. Both resolved their blocking findings. Fable reviewed the three preceding refinement commits chronologically and the final working-tree changes; the final two commits separate the already-reviewed installer and skill scopes.

Two existing complexity warnings remain in unchanged functions. Only the native Linux tarball was built locally; cross-platform release promotion is outside this PR.

Scope notes

This builds on the two commits from #2901 rebased onto dev and the provider-only refinement change. All changes are delivered in this new PR; no merge or release is included. The separately maintained documentation site still has older skill pages and is not changed by this repository PR; the root and shipped skill READMEs document the new routes and upgrade behavior.

@coderabbitai

coderabbitai Bot commented Sep 15, 2026 •

Copy link
Copy Markdown

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: a8bd1248-8aa4-4e44-8a8c-74a9d3a5a947

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1f7a3c7105

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

// Allow a symlinked HOME, but never follow a redirected agent home below it.
const parentMatches =
contained && realpathSync(dirname(target)) === join(realpathSync(context.home), dirname(mirrored));
if (!parentMatches || expected === undefined || computeSkillDirDigest(target) !== expected) {

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Revalidate the skill after the archival rename

When another process edits or replaces a retired skill after this digest check but before renameSync, the check authorizes the old snapshot while the later rename moves the new user-owned directory out of the live skill home. This can occur during concurrent agent or manual skill edits, violating the promise that modified copies remain in place; bind the check to the directory's physical identity or reverify the parked directory and restore it on mismatch.

AGENTS.md reference: AGENTS.md:L30-L30

Useful? React with 👍 / 👎.

@automagik-genie
automagik-genie force-pushed the nmstx/skills-lightweight-body branch from 1f7a3c7 to ed6028f Compare September 15, 2026 11:41
@automagik-genie
automagik-genie merged commit a5f489a into dev Sep 15, 2026
21 checks passed
@automagik-genie
automagik-genie deleted the nmstx/skills-lightweight-body branch September 25, 2026 04:51
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants