Prism runs your coding agents from typed, checked Effect workflows.
Claude Code, Codex, Grok, Kimi, Amp, OpenCode, and more, driven from TypeScript: schema-checked answers, repair loops, and a run ledger you can replay.
npm install -g @skastr0/prism
·
npm
·
v0.8.2 · macOS and Linux
- No types. You ask two agents to review a commit, from a script or a chat. Each answers in its own prose, and you read it or scrape it.
- No checks, no retries. A "revise" with no findings flows straight into your next step. A malformed reply or a flaky exit, and you start over by hand.
- No resume, no record. The script dies halfway and every finished task runs again at full token cost. Afterwards nothing says which agent answered what, on which model, in how long.
A task is a harness CLI, a prompt, an Effect Schema the answer must decode into, and checks the answer must pass. A workflow is an Effect program over tasks. Prism runs each task through the agent CLI you already have, with its own login, and records every run.
| Prism is | Prism is not |
|---|---|
| Workflows you write in TypeScript with Effect | A drag-and-drop pipeline builder |
| A runner that drives the agent CLIs you already use | An SDK that calls model APIs |
Typed: a wrong harness setting fails typecheck, a wrong answer fails the schema |
Prompt-and-hope JSON scraping |
| A local ledger: every run recorded, finished tasks replayed on rerun | A hosted service |
1. Install.
npm install -g @skastr0/prism2. Write a workflow. Two agents review the current commit in parallel and must answer in the same shape. Save it as review.workflow.ts:
import { execSync } from "node:child_process";
import { Effect, Schema } from "effect";
import { defineTask, defineWorkflow, type WorkflowTaskWorkerOptions } from "prism";
const commit = execSync("git rev-parse --short HEAD").toString().trim();
const Review = Schema.Struct({
verdict: Schema.Literals(["ship", "revise", "block"]),
findings: Schema.Array(Schema.String),
});
const review = (id: string, worker: WorkflowTaskWorkerOptions) =>
defineTask({
id,
prompt: `Review commit ${commit} in this repo (git show ${commit}). Is it safe to ship?`,
output: Review,
cacheKey: `review-${commit}`,
worker,
finish: {
maxRepairs: 1,
criteria: [
{
name: "a non-ship verdict names its findings",
check: ({ output }) =>
output.verdict !== "ship" && output.findings.length === 0
? Effect.fail(new Error("verdict without findings"))
: Effect.void,
repairPrompt: () => "Name the findings behind your verdict.",
},
],
},
});
export default defineWorkflow({
name: "review-commit",
run: (wf) =>
Effect.all(
[
wf.runTask(review("codex", { worker: "codex-cli", permission: "sandbox-read-only" })),
wf.runTask(review("claude", { worker: "claude-code", model: "sonnet", permission: "restricted", restrictedTools: ["Read", "Grep", "Bash(git show:*)"] })),
],
{ concurrency: "unbounded" },
),
});3. Typecheck it. Settings a harness can't honor fail here, before any tokens are spent:
$ prism workflow typecheck review.workflow.ts
Workflow typecheck passed: review.workflow.ts
Put permission: "sandbox-read-only" (a Codex mode) on the Claude task and you get:
Type '"sandbox-read-only"' is not assignable to type 'ClaudeWorkflowPermissionMode | undefined'.
4. Run it from inside a Git repo (Codex refuses to run outside one), with Codex and Claude Code logged in. A live run spends their tokens and has no built-in cost cap:
$ prism workflow run review.workflow.ts
{
"runId": "ee259bf3-204f-4b02-8614-9fa56868bc5b",
"tasks": [
{ "id": "codex", "status": "completed", "cached": false,
"output": { "verdict": "revise", "findings": [
"README.md:117 [MEDIUM] The new prose promises a backup whenever a file is repaired. Source changes also produce a `repair`, but `src/sync/plan.ts:279-280` sets `backup: false` for them; ..." ] } },
{ "id": "claude", "status": "completed", "cached": false,
"output": { "verdict": "ship", "findings": [ "327d206 is docs-only: ...", ... ] } }
]
}
5. Read it back. Run it again and both tasks replay from the ledger in under 2 seconds. The trace shows what ran:
$ prism workflow runs trace ee259bf3-204f-4b02-8614-9fa56868bc5b
✓ workflow.run · review-commit · 32.9s
└─ ✓ workflow.program · 32.9s
├─ ✓ workflow.task · codex · 32.9s
│ └─ ✓ task.executor · attempt 0 · codex-cli · 32.8s
└─ ✓ workflow.task · claude · 24.3s
└─ ✓ task.executor · attempt 0 · claude-code sonnet · 24.3s
To rehearse for free, pass --mock-output mock.json (answers keyed by task id). Mocked answers still go through the schema and the checks:
$ prism workflow run review.workflow.ts --mock-output mock.json # codex: "revise", no findings
❌ Workflow run failed: workflow task codex failed finish criterion 'a non-ship verdict names its findings': verdict without findings
output |
An Effect Schema. The answer must decode into it; a bad reply gets up to 2 repair prompts, then the task fails with a typed error. |
finish |
Checks on the decoded answer: plain code, or a judge that returns pass, continue, fail, or escalate. maxRepairs sets how many retries a failed check gets. |
worker |
The harness, model, and permission mode. Each harness only accepts the permission modes it can enforce. |
cacheKey |
Finished results are cached by task id, key, and a hash of what ran. Change the prompt or schema and only that task runs again. |
| the ledger | Every run, task, event, and span in a local SQLite store: prism workflow runs list, summary, trace. |
Workflows can be a list of tasks or any Effect program: Effect.all for fan-out, Effect.result so one failed agent doesn't sink the rest. Judges, phases, named task configurations, detached runs, resume, and OTLP export are in docs/workflows.md.
prism harnesses lists 14:
| Harness | ID | Runs tasks | Installs skills and agents |
|---|---|---|---|
| Claude Code | claude-code |
yes | yes |
| Codex CLI | codex-cli |
yes | yes |
| OpenCode | opencode |
yes | yes |
| Grok Build | grok |
yes | yes |
| Kimi Code | kimi-code |
yes | yes |
| Amp Code | amp-code |
yes | yes |
| Antigravity CLI | antigravity-cli |
yes | yes |
| Oh My Pi | omp |
yes | yes |
| Devin CLI | devin |
yes | yes |
| Cursor | cursor |
yes | yes |
| Hermes Agent | hermes |
yes | skills and tools |
| Pi | pi |
no | yes |
| Amp Orb | amp-orb |
yes | skills into a hosted checkout |
| Amp Runner | amp-runner |
yes | no |
Details per harness: docs/lowerer-capability-matrix.md.
The same CLI installs one set of agents, skills, rules, tools, and hooks into every harness above, each in its own format. Keep a plugin directory in Git and run prism refresh: a second run writes nothing, a file you edited by hand is backed up and repaired, and a file Prism never wrote is refused.
prism init my-standards --with-agent --with-skill --typescript
prism plan my-standards --all
prism refresh my-standards --harness claude-code,codex-cli,opencode
prism doctorTools written once run from any harness through prism tools invoke: docs/tools-architecture.md. Pinned skills from other repos: docs/third-party-skills.md.
Several agents on one piece of work, each in its own harness, answering the same typed contract. The agents also share one toolbox: Quasar ships its tools as a Prism plugin. More at castro.engineer/projects/prism.
| Package | What it is |
|---|---|
@skastr0/prism |
The CLI, with binaries for darwin-arm64, darwin-x64, linux-arm64, linux-x64 |
@skastr0/prism-sdk |
Core contracts and codecs |
@skastr0/prism-packager |
Compile a plugin into harness-native files without the CLI |
bun install
bun run build
bun testMore in CONTRIBUTING.md. Changes: CHANGELOG.md.
Report suspected vulnerabilities privately. See SECURITY.md.
MIT. See LICENSE.