SSI

One command takes a request to a reviewed draft PR. A small script, not the model, decides what comes next, so it remembers where it was and repairs itself after an interruption. It never merges, closes an Issue or deploys.

/ssi Fix: total() ignores percentage discounts

The flow

flowchart LR
  A[1 Analyze] --> B[2 Confirm] --> C[3 Reproduce<br/>bugs only] --> D[4 Plan<br/>reuse first] --> E[5 Implement<br/>draft PR] --> F[6 Review<br/>Codex or sub-agent] --> G[7 Visual GIF<br/>UI only] --> H[8 Land<br/>CI green] --> Z([Draft PR<br/>ready for you])
  F -- blocker --> E
  G -- wrong --> E
  H -- CI red --> E
  A -.-> Q{{One question<br/>to you}}
  C -.-> Q
  D -.-> Q
  E -.-> Q
  F -.-> Q

It runs on its own and asks one question only when it must: unclear request, no reproduction, protected code touched, a big diff, a review or CI that keeps failing, or an action outside its mandate.

What happens inside

sequenceDiagram
  participant You
  participant Skill as Model (skill)
  participant Engine as ssi engine (script)
  participant World as git + PR
  You->>Skill: /ssi <request>
  loop until done or a question
    Skill->>Engine: ssi next
    Engine->>World: read branch, PR, recorded proof
    Engine-->>Skill: the one next action, or a stop
    Skill->>Skill: do the work (tests, code, review, GIF)
    Skill->>Engine: ssi record (the proof)
  end
  Engine-->>You: draft PR ready, or one question

If the notes and the repo disagree, the repo wins. Lost local notes are restored from a block in the PR's progress comment. A new commit that touches proven files makes that proof stale and it is redone.

What is different

Comparisons are about design, not quality: the projects below are good, and this one borrows from them.

| Common in prompt-only workflow skills | SSI | |---|---| | The model remembers the phase | A script derives the phase from git, the PR and recorded proof | | You call each step or skill yourself | One entry point; it switches phases itself | | Stopping is a judgement call | Stops and thresholds are computed (diff size, protected paths from the real diff, retry budgets) | | An interrupted run starts over | It resumes, restores notes from the PR and redoes only stale proof | | Review is the same model re-reading its work | Codex if it works, else a sub-agent on another model; the report says whether it was independent | | Output tuned for experts | One short template: where we are, what happened, what is next, at most one question; optional adhd style | | Remote side effects are ad hoc | Keyed and idempotent: no duplicate Issue or comment; evidence GIF on a ssi-assets branch |

Install and use

Needs Node 20+, git, and gh logged in. Optional: Codex CLI (review), Chrome or Playwright and ffmpeg (GIF).

cp -R skills/ssi ~/.claude/skills/ssi          # the folder is self-contained
claude --plugin-dir .                          # or load the repo as a plugin

Settings live in ssi.config.json (repo) and .ssi/config.local.json (yours); ssi config prints them. Optional status band and phase map: mods/ssi-cockpit.

Inspired by

• mattpocock/skills and jsmastery-pro/skills: engineering practices adapted into the phase guides.
• DietrichGebert/ponytail: the "smallest change that works" lens in planning.
• ayghri/i-have-adhd: the reader-friendly message rules.
• openai/codex-plugin-cc: how Codex is called (installed binary, read-only, optional model) and the review report shape.
• obra/superpowers: the brainstorm and design process used to build this.

Exact versions and licences: skills/ssi/NOTICE.md.

Read more

• docs/philosophy.md: why it is built this way.
• docs/design.md: how it works, as a reference.
• docs/runs.md: the real runs and what they found.

Status

Early. 134 engine tests, three real runs, six rounds of Codex review (the last round's fixes are not re-reviewed yet). MIT, see LICENSE.