Unvibe

Unvibe

Quick start

/unvibe                      # analyze recent git changes
/unvibe src/lib/auth.ts      # analyze a specific file or function

Setup

1. Detect the language, build system, and test runner from repo files (package.json, pyproject.toml, Cargo.toml, go.mod, etc.).
2. Do not change files silently. Before creating .unvibe/ or editing .gitignore, ask for explicit approval.
3. If the stack or runner is unclear, pause before Phase 2 and ask whether to make a runnable native exercise or a conceptual exercise.

Mode detection

• Post-session: no args → inspect git diff and git diff --staged; if both are empty, inspect the last commit (git show --stat --patch HEAD).
• On-demand: file/function arg → read that specific code.
• Unclear: ask "Which file or function do you want to unvibe?"

Bundled resources

• references/reference.md — rubrics, priorities, and format adaptations. Consult this when you need the Phase 1 question forms, the difficulty rubric, the Examples-happy-path rule, the reference-grounding rule for external-surface facts, or the per-status notes-body layout.
• assets/templates/ — readme.md, index.md, test_python.py, test_go.go, test_ts.ts. Read the relevant file, fill the <TOKENS>, then Write the substituted file to the exercise dir.
• assets/runners/python.sh, assets/runners/go.sh, assets/runners/ts.sh — copy the matching one to the exercise dir as run.sh, chmod +x. For other stacks, follow the same [-q] [test_name] interface.
• scripts/render_notes.py — render notes.md with status-conditional body sections. Use this in Phase 4 instead of hand-writing the file.
• scripts/update_index.py — update .unvibe/INDEX.md in place from a notes.md frontmatter. Use this after every notes.md write (including re-drills).
• scripts/quick_validate.sh — self-check that a generated exercise dir is well-formed (dir name, README difficulty tag, executable run.sh, test + stub files). Auto-detects the sliced layout (stepN_*/ subdirs) and validates each step as a leaf. Run at the end of Phase 2.

Workflow

Step 1 — Identify candidates

Silently find up to 3 pieces worth owning (1–3, never invent filler to reach 3); see references/reference.md → "Candidate identification priorities" for the selection criteria.

Present them as a numbered shortlist with a short "why this matters" note, then ask the user which one to drill and how confident they feel about it.

Step 2 — Exercise loop (repeat per selected piece)

Phase 1: Explain — ask one focused question about intent, not line-by-line. Evaluate solid → Phase 2, partial → one follow-up → Phase 2, fuzzy → you explain (3–4 sentences) → they repeat back → Phase 2. See references/reference.md → "Phase 1: Evaluation rubric".

Phase 2: Reimplement (LeetCode-style) — right-size first. A piece with more than one responsibility, or one the user was low-confidence on in Phase 1, gets sliced into a short progression of small step-subdirs (pure helpers first, composition last) instead of one oversized stub; a single cohesive idea stays one exercise. See references/reference.md → "Phase 2: Right-sizing — slice large pieces into a progression" for the trigger, the stepN_<slug>/ layout, ordering, and the give-earlier-solutions-forward rule. If the user approves file writes, create the container .unvibe/exercises/<YYYY-MM-DD_HH-MM-SS>_<slug>/. For a single exercise it holds these four files directly; for a sliced one it holds a progression-map README.md and each stepN_<slug>/ leaf holds these four:

1. README.md — Read assets/templates/readme.md, fill the <TOKENS> (<DIFFICULTY> per the rubric, <PIECE_NAME>, <SOURCE_PATH>, <PROBLEM_DESCRIPTION>, <EXAMPLES_BLOCK> happy-path only, <CONSTRAINTS_LIST>, <LANG>, <SIGNATURE_BLOCK>), Write to the exercise dir. Ground external-surface facts before asserting them: any claim about a platform / stdlib / framework / language-spec / third-party API you did not author must be verified with WebSearch/WebFetch and stated as an absolute only if verified; otherwise hedge it. Keep the links out of the README — they belong in the Phase 4 trail. See references/reference.md → "Reference grounding (external-surface facts)". Never write a citation URL you did not actually fetch.
2. exercise.<ext> — write the public signature plus a stub that fails clearly (e.g. raise NotImplementedError, panic("not implemented"), throw new Error("not implemented")). No hints.
3. test_exercise.<ext> — Read the matching assets/templates/test_<lang>.<ext>, fill <PIECE_NAME> and add the hidden edge / failure cases. Every assertion must carry a descriptive message; edge and failure cases live ONLY in this file, never in the README.
4. run.sh — copy the matching assets/runners/<lang>.sh to the exercise dir, chmod +x. Verbose by default; [-q] for quiet; --list + numeric index for exact test selection (TS / Go); <test_name> for pattern match; --help prints the full contract. See references/reference.md → "Phase 2 / 3: run.sh contract".

After all four files exist, run bash scripts/quick_validate.sh <exercise-dir> and address any failure before moving on. If file writes are declined or native tests are not safe to generate, present a conceptual exercise spec in chat instead (see references/reference.md → "Phase 2: Conceptual exercise fallback").

Phase 3: Verify — run the exercise's run.sh. If tests fail, give one conceptual nudge and let the user revise. If the agent cannot run tests, ask the user to paste the output before moving on.

Phase 4: Capture — write per-piece notes that reflect what actually happened, regardless of outcome. Save directly; do not gate on user approval. The notes are a record, not a deliverable — if the user wants something changed they can mention notes.md and you edit it in place. Don't draft-then-confirm.

1. Compose a config dict (date, first_drilled, piece, source, confidence_before/after, status one of owned / partial / explained-only / abandoned, time_minutes, difficulty, tags, re_drills, and the body sections). Save to a temp JSON file. If you grounded any external-surface facts this session, add a Further reading entry to sections holding the vetted trail (one verified link + a one-line anchor each; format in references/reference.md → "Reference grounding"). Omit it when nothing was grounded — the renderer drops the empty section. Never include a link you did not actually fetch.
2. Render straight to the file: python scripts/render_notes.py --config <tmp.json> > <exercise-dir>/notes.md. The script handles status-conditional body layout.
3. Update the workspace index: python scripts/update_index.py --notes <exercise-dir>/notes.md.
4. Tell the user the notes were saved in one line, with the path, so they know where to refine them — but do not ask for confirmation.
5. On re-drill of the same piece: append a new dated section to the existing notes.md (do not overwrite earlier sections), bump re_drills, then re-run update_index.py. See references/reference.md → "Phase 4: Re-drill append pattern".

Step 3 — Session summary

Briefly highlight what changed this session — pieces now owned, still partial, where the exercises live. .unvibe/INDEX.md is the authoritative log; the chat summary just calls out the deltas.

Rules

• One phase at a time: never ask them to explain and reimplement in the same message.
• Never show the original code during Phase 2: they work from memory plus the problem statement and tests.
• Never reveal hidden edge/failure cases in the README — they surface only when a test goes red.
• Hints during Phase 2: one conceptual nudge only, never code.
• Keep generated tests behavioral. Do not assert private implementation details.
• notes.md is saved automatically after each piece — no approval gate, no draft-then-confirm. State the saved path so the user can refine it; if they ask for changes, edit the file in place.