Pullable, inspectable agent workflows for real jobs. Each recipe includes the instructions, starter artifacts, safety boundary, and smallest useful check.
git clone https://github.com/Software-Sausage/recipes.git sausage-recipes
cd sausage-recipes/recipes/portable-agent-repository
./verify.shThat credential-free check should end with Verification passed. To benchmark
an agent, copy the input fixture as described in this recipe's README, open the
directory in Codex, Claude Code, or another repository-aware agent, and paste
the task from PROMPT.md.
| Recipe | Produces | Status |
|---|---|---|
| Portable agent-ready repository | One shared instruction source, a fixed coding fixture, and credential-free verification | Cross-harness baseline plus a measured Claude Code/Opus 5 comparison with and without Ponytail |
| Research paper | A source ledger, LaTeX manuscript, build evidence, and adversarial review log | Editorial |
| Obsidian Operator sampler | A clean-vault daily, meeting, person, project, task, and procedure loop | Fixture-verified sampler |
| Slack to Mattermost | A fixed cost model, operating assumptions, privacy boundary, and migration checklist | Fixture-verified decision kit |
| Self-hosted vs. managed forge | A metadata-only repository inventory, ownership matrix, privacy boundary, and migration checklist | Fixture-verified decision kit |
| CRM admin and exit cost | An editable cost model and reconstruction-focused exit test | Fixture-verified decision kit |
| Automation platform reliability and exit | A workflow inventory, failure drills, and clean-instance reconstruction checks | Fixture-verified decision kit |
| MCP server preflight | A provenance record, least-privilege ledger, dependency evidence, and negative-test plan | Fixture-verified decision kit |
| 1Password to Bitwarden migration | A secret-free inventory, access design, import reconciliation, plaintext-export controls, and rollback gate | Fixture-verified migration kit |
| Notion to Outline or AppFlowy | A content inventory, license boundary, destination pilot matrix, fidelity checks, operating-ownership gate, and rollback plan | Fixture-verified migration kit |
| Google Drive to Nextcloud | An export-authority record, file and sharing inventory, conversion pilot, operating-ownership gate, reconciliation, and rollback plan | Fixture-verified migration kit |
| Spec to production proof | A linked specification, implementation ledger, browser evidence, and rollback gate | Fixture-verified workflow kit |
| AI workflow regression test | Frozen cases, pinned model and prompt settings, comparable runs, and a reviewed release gate | Fixture-verified evaluation kit |
| Open-source coding-agent benchmark | A fixed repository fixture, comparable run manifests, blind scores, and a job-specific agent decision | Fixture-verified benchmark kit |
| Document parser fidelity test | Side-by-side extraction evidence for MarkItDown and Docling with a source-linked claim ledger | Fixture-verified decision kit |
| Browser performance regression | Reproducible browser conditions, before-and-after traces, functional checks, and a reviewed fix | Fixture-verified workflow kit |
| Safe dependency upgrade | A scoped upgrade, dependency and lockfile evidence, scan findings, runtime proof, and rollback record | Fixture-verified workflow kit |
| Three-agent code-change gate | A bounded implementation, deterministic checks, separate correctness and security reviews, and a human ship decision | Fixture-verified workflow kit |
| Multi-agent experiment gate | A one-agent baseline, fixed multi-agent trial, comparable evidence, and a keep-or-stop decision | Fixture-verified evaluation kit |
The recipes are free under the MIT license. They do not replace source verification, institutional policy, or qualified human review.
More workflows and independent software guidance: softwaresausage.com.
Get one recipe a week by email through the Software Sausage newsletter. Signup is confirmed by email, every issue has a one-click unsubscribe, and the list is never sold.
Connect https://softwaresausage.com/api/mcp, call the read-only
find_agent_recipes tool, then read the returned
sausage://recipes/{slug} resource for the complete run ledger. Setup and
privacy boundaries are documented at
softwaresausage.com/mcp.
The pinned v0.12.0 benchmark records harness and model versions, elapsed time, reported cost, file scope, failed commands, verifier output, and limitations. Codex/GPT-5.6-sol, Claude Code/Opus 5, and OpenCode/local Qwen3 Coder passed; OpenCode/local Devstral Small 2 stopped on its own failing verifier, and Gemini CLI was blocked before model execution. It is a small portability test, not a universal model leaderboard.
Release v0.8.0 adds three blinded runs per arm of Claude Code 2.1.259 with Opus 5. Both arms passed 3/3; Ponytail 4.8.4 reduced median time 29.1% and reported list cost 28.8% on this fixture while median code size remained equal. Inspect the protocol and every artifact.
The Codex CLI comparison uses the same fixture and three runs per arm. Both arms again passed 3/3; Ponytail reduced median implementation size from seven to three nonblank lines but increased median time and tokens. The different result is why add-ons must be tested per harness and job.
Request a recipe with the job, required artifact, inputs, constraints, and a verification method. Do not submit vendor-sponsored ranking copy or unverified benchmark claims.
Optional one-time support is available through the repository sponsor button. Contributions never affect recipe inclusion, comparison, or ranking.
Need a business-software decision instead? The free shortlist comes before an optional $39 private report.
