oh-my-pi-typescript

TypeScript workflows for pi, the coding agent. One pi install brings three skills, two extensions, one slash-command prompt, a settings preset, and a reproducible benchmark into your project.

Part of the oh-my-pi reference set. Distribution philosophy: CLI + skill files only, no MCP servers, no wrappers around tools pi already knows (git, jq, rg, tsc, vitest).

---

What you get

Type	Name	Cost	Purpose
skill	ts-test-runner	573 tokens on-demand	Run a TS test suite once (non-watch). Invariants the agent keeps forgetting.
skill	ts-deps-audit	664 tokens on-demand	Stateful deps audit: outdated / vulnerable / top-size. 7-craft registry.
skill	ts-monorepo-nav	978 tokens on-demand	pnpm / turbo / npm-workspaces cheat sheet with the ...^N pitfall.
extension	ts-pre-commit-guard	always-on	Block git commit tool calls when type-check or tests would fail.
extension	ts-status-line	always-on	Footer status: active tsconfig + tsc version + Node runtime + engines check.
prompt	/ts-review	on invocation	Four-axis code review: types, side effects, tests, API surface.
preset	settings.json	—	compaction.keepRecentTokens: 40000, transport: "sse".
bench task	ts-find-bad-type	5,238 tokens (measured)	Reproducible faux-provider benchmark. See § Measurements.

Total token budget declared (pi.ohmypi.tokenBudget.onDemand): 2,215 tokens — the sum of the three skills when pi invokes all of them in one turn. Extensions do not count toward the on-demand budget; they are loaded as code at startup.

---

Install

# When this package is published:
pi install npm:oh-my-pi-typescript

# During development from a local checkout:
pi install -l file:./reference-packages/oh-my-pi-typescript

Requires @mariozechner/pi-coding-agent (pi's core). The five pi-bundled cores are declared as peerDependencies: "*" so any pi ≥ 0.67 satisfies them without a second install of pi-ai / pi-tui / typebox / etc.

Verify the install:

omp doctor                              # inventory skills / extensions / prompts
omp doctor -f json | jq .conflicts      # 'none' on a clean setup
omp remeasure ./ --check                # confirms SKILL.md budgets are in sync
omp bench ./ --runs 3                   # re-run the bench task locally

---

Using the resources

Skills (invoked by the agent when relevant)

pi autoloads the three SKILL.md files. You can also pull a skill manually:

/skill:ts-test-runner
/skill:ts-deps-audit
/skill:ts-monorepo-nav

ts-deps-audit ships a Node shebang script (scripts/ts-deps-audit) following the terminalcp 7 craft rules for stateful tools:

./scripts/ts-deps-audit summary            # one-shot registry snapshot
./scripts/ts-deps-audit outdated --format json --top 10
./scripts/ts-deps-audit audit --severity high --lines 40 --strip-ansi
./scripts/ts-deps-audit help               # craft #4: discoverable, machine-readable help

Extensions (active while pi runs)

• ts-pre-commit-guard watches every bash tool call. When the agent tries to run git commit it first runs your project's typecheck (or check:tests / check / npx tsc --noEmit fallback) and npm test. On failure it returns { block: true, reason: <tail of output> } so the agent sees why and can fix it in the same turn. Bypass for one call: pi --ts-pre-commit-guard-skip ....

• ts-status-line populates the footer with: the active tsconfig.json (marked with ↑ when discovered in an ancestor directory), the pinned TypeScript compiler version, and the current Node runtime. When package.json.engines.node is declared it is compared against the live Node; the status is tinted success on match, warning on mismatch, dim otherwise. Disable: pi --ts-status-line-skip ....

Prompt /ts-review

/ts-review                      # reviews the staged diff
/ts-review packages/core/src    # reviews a path or git pathspec

Output is a structured Markdown report with a severity-grouped findings table (block / warn / note) plus a Clear section listing what was checked and found OK — so the agent cannot fabricate findings to fill space. See prompts/ts-review.md for the full axis list.

Settings preset

pi.ohmypi.presets["settings.json"] recommends:

{
  "compaction": { "keepRecentTokens": 40000 },  // pi default: 20000
  "transport": "sse"                             // pi default: "sse"
}

pi does not merge these automatically (silent edits to user settings are a non-goal). omp doctor reports drift when your live settings diverge; sync manually when you agree.

---

Measurements

Bench run, omp bench ./ --runs 3 --write, measured on Node 24.13:

metric	mean	stddev	stddev %
totalTokens	5,238	0.00	0.00%
promptTokens	2,234	0.00	0.00%
outputTokens	89	0.00	0.00%
successRate	100.0%	—	—

Reproducibility gate (concept-v0 §4.1): stddev(totalTokens) / mean < 5%. Current result: 0.00%. The faux provider replays a scripted two-step exchange so these numbers are deterministic across machines.

The task lives in bench/task-1.md with the fixture under bench/fixtures/ts-bad-type/. Run it yourself:

omp bench ./ --runs 3                # view mean + stddev
omp bench ./ --runs 3 --write        # stamp pi.ohmypi.bench.lastResult

---

Disabling individual components

• Skill opt-out: delete or .gitignore its skills/<name>/SKILL.md, or add "enableSkillCommands": false to your pi settings to silence the /skill:name commands entirely.
• Extension opt-out: pi --ts-pre-commit-guard-skip ... or pi --ts-status-line-skip ... for a single session; remove the entry from pi.extensions in your package.json / settings to disable permanently.
• Prompt opt-out: delete prompts/ts-review.md or run pi with --no-prompt-templates.

Uninstall the whole package:

pi uninstall npm:oh-my-pi-typescript    # or `pi remove ...`, same command

---

Compatibility

• pi ≥ 0.67.68 (tested; older versions untested).
• Node ≥ 20 (extensions use top-level for/await and ESM; package declares "type": "module").
• TypeScript projects with a reachable tsconfig.json. The extensions remain inert on non-TS projects rather than erroring.

---

Layout

oh-my-pi-typescript/
├── package.json            # pi manifest (pi.ohmypi.{category,tokenBudget,presets,bench})
├── skills/
│   ├── ts-test-runner/SKILL.md
│   ├── ts-deps-audit/{SKILL.md, scripts/ts-deps-audit, bench/task-1.md}
│   └── ts-monorepo-nav/SKILL.md
├── extensions/
│   ├── ts-pre-commit-guard/{index.ts, util.ts, test/*.test.ts}
│   └── ts-status-line/{index.ts, util.ts, test/*.test.ts}
├── prompts/
│   ├── ts-review.md
│   └── test/load.test.ts
├── bench/
│   ├── task-1.md
│   └── fixtures/ts-bad-type/src/app.ts
└── test/manifest.test.ts   # schema + preset + peerDependencies pins

Each extension follows pi loader rule 2 (subdirectory with index.ts) so helper files stay at extensions/<name>/util.ts instead of being mis-detected as separate extensions. Run npm test from the monorepo root to exercise 76 tests across this atom alone.

---

License

MIT