Isolint

Lint your AI harness so weak small models can actually run it. Finds the phrasing that breaks Minimax 2.5, Nemotron 3, Mistral 7B, and local models — then rewrites it with a smart model.

You author a harness (opencode modes, Claude Code agent files, Cursor rules, plain .md skills) with a frontier model. You want to run it on a cheap small model. The logic is fine — but the prose is full of phrases that only a frontier model knows how to interpret: should, when relevant, one of the usual categories, the table above, creative, leveraged. Small models drop clauses, invent items, ignore soft imperatives, and blow past taste words.

Isolint fixes that:

┌─────────────────┐           ┌─────────────────────┐           ┌─────────────────┐
│  Your harness   │  isolint  │  Lint report + diff │  --fix    │  Same harness,  │
│  (.md files)    │ ────────▶ │  of every phrase    │ ────────▶ │  small-model-   │
│                 │           │  a 7B model misses  │           │  safe prose     │
└─────────────────┘           └─────────────────────┘           └─────────────────┘

Isolint also ships an Isomorphic Plan engine (isolint plan / run) — a large model emits a strict JSON plan; a small model executes it with schema validation. The linter uses this engine internally for its LLM-assisted rules. Use it directly when you want a fully-deterministic pipeline instead of markdown prose.

Contents

• Why
• Architecture
• The linter
• Plans
• Install
• CLI
• Examples
• Programmatic API
• Tests

Why

Most "agent" systems push all complexity into runtime prose: long system prompts, sprawling instructions, taste-based validation. 7B-class models collapse under that weight — not because the logic is wrong, but because the prose is ambiguous.

The default recommended preset scans for **28 deterministic rule patterns

• 5 LLM-assisted rules**, each targeting one concrete small-model failure mode. An optional performance preset adds 18 advisory rules for harness overhead: repeated instructions, oversized examples, redundant contracts, low-value prose, saturated emphasis, mode-conditional branches in shared prefixes, cross-file duplication, and other avoidable token/latency costs. Every finding is a fixable phrase. Every fix preserves intent and markdown formatting — and every LLM rewrite is re-linted before being applied so bad fixes never ship.

Architecture

src/
├── lint/        # Markdown harness linter (deterministic + LLM-assisted)
│   ├── rules/   # Individual rules (soft-imperative, taste-word, ...)
│   ├── fix.ts   # Deterministic fix engine + LLM rewriter
│   └── ...
├── schema/      # Strict Plan schema (the contract)
├── planner/     # Large-model → Plan (JSON)
├── runtime/     # Small-model executor (validates every step)
├── providers/   # OpenAI / OpenRouter / Ollama / Mock
├── util/        # JSON extraction, logger
└── cli/         # isolint lint | verify | plan | run | validate

The Linter (the practical win)

If you use Claude Code / Codex / Cursor to author a harness and then run it on minimax / nemotron / local models, the #1 failure mode is not logic — it's prose the small model can't parse. The linter catches that.

What it flags

Every rule targets a concrete failure mode in weak models. All rules run against fenced code blocks, inline code, and HTML comments are skipped.

| Rule | Catches | Severity | | --- | --- | --- | | soft-imperative | should, could, might, consider, ideally — weak models drop these | warn | | vague-quantifier | some, several, a few, many, most without a number | warn | | taste-word | creative, engaging, passionate, leveraged, cutting-edge, etc. | warn | | ambiguous-deictic | the table above, as mentioned below — position-based refs | info | | double-negation | don't forget to not skip… — weak models flip the sign | warn | | long-sentence | Sentences > 35 words — weak models drop clauses | info | | pronoun-no-antecedent | It, This, They at start of block after a list/heading | info | | enum-without-list | one of the usual categories without the list inline | warn | | word-count-target | write 100 words — weak models count tokens poorly | info | | implicit-conditional | when relevant, if appropriate, as needed | warn | | trailing-etc | A, B, C, etc. — unclosed set, weak models invent items | warn | | heading-without-imperative | Mode headings without an action verb | info | | nested-conditional | Multiple if / unless / except in one sentence | warn | | multiple-output-formats | "return JSON and a summary" in one step | warn | | placeholder-leftover | TODO, FIXME, <insert X>, [INSERT X] — weak models echo scaffolding | warn | | output-format-no-example | "return JSON" with no example, schema, or field list nearby | warn | | numbered-step-gap | 1. … 2. … 4. … — weak models inherit the gap and skip the step | warn | | step-without-verb | ## Step N body that opens with a noun, not an imperative verb | info | | undefined-step-reference | Prose mentions Step 5 when only 3 steps exist across the harness | warn | | missing-file-reference | Read cv.md when cv.md doesn't exist in the repo | warn | | context-budget | Harness file over the prose-length budget — weak models drop the middle | info / warn | | dangling-variable-reference | $input.X / $steps.Y.output with no declared source — weak models hallucinate the value | warn | | invalid-json-fence | ```json fence body that doesn't parse as JSON — weak models copy the malformed shape | warn | | heading-hierarchy | Skipped heading levels (# A → ### B) — weak models use depth as structure | info | | stale-link-reference | [label](./missing.md) where the target isn't in the repo | warn | | table-column-mismatch | Table rows with different cell counts than the header | warn | | mixed-list-marker | One list mixes - and * markers — weak models split on the change | info | | frontmatter-schema | Harness frontmatter missing required fields (Claude Code description, Cursor globs/alwaysApply) | warn |

Plus five LLM-assisted rules (opt-in via --llm) that use a smart model with JSON-mode output for checks that need judgment:

| Rule | Catches | | --- | --- | | llm-atomicity | Instructions bundling multiple actions into one sentence | | llm-implicit-context | Phrases that rely on context not present in the file | | llm-unexplained-schema | Output contracts in prose with no example / schema nearby | | llm-tone-drift | Authoritative imperatives followed by casual phrasing in the same section | | llm-implicit-assumption | Sentences that reference entities with no prior definition |

Quick start

# Scan any directory of .md / .mdc / .mdx files — reliability + performance by default
npx @razroo/isolint lint /path/to/harness

# Reliability rules only (opt out of performance)
npx @razroo/isolint lint /path/to/harness --preset recommended

# Performance rules only (advisory, info-severity)
npx @razroo/isolint lint /path/to/harness --preset performance

# See how many tokens your harness costs per turn (shared prefix + per-mode + per-agent)
npx @razroo/isolint cost /path/to/harness

# Just your opencode modes / Claude Code agents / Cursor rules
npx @razroo/isolint lint .opencode/skills .opencode/agents modes .cursor/rules

# Only lint files changed since main (great for PR CI)
npx @razroo/isolint lint . --since origin/main --fail-on warn

# JSON for CI / GitHub annotations (SARIF)
npx @razroo/isolint lint ./modes --format sarif > lint.sarif

# Auto-fix everything the rules can fix deterministically + via LLM rewrites
export OPENROUTER_API_KEY=sk-or-...
npx @razroo/isolint lint ./modes --fix --llm --large anthropic/claude-3.5-sonnet

# Dry-run: see the diff, don't write files
npx @razroo/isolint lint ./modes --fix --llm --diff

.md, .mdc (Cursor rules), and .mdx are all picked up by default. YAML (---) / TOML (+++) frontmatter at the top of any file is skipped.

When run inside a git repo, isolint honors .gitignore (plus .git/info/exclude and any global git excludes) so build output like generated CLAUDE.md / AGENTS.md / .cursor/rules/ copies don't get linted. Pass --no-gitignore to disable.

Suppressions

Any file can silence findings with HTML comments:

<!-- isolint-disable-next-line taste-word -->
These words (leveraged, cutting-edge, passionate) are ATS red flags.

<!-- isolint-disable -->
This whole block is documentation about banned words, not an instruction.
- leveraged
- utilized
- spearheaded
<!-- isolint-enable -->

No rule ids = suppress all rules on the target line(s).

Config (.isolint.json)

Both .isolint.json and the legacy .isomodel-lint.json are read. Shape:

{
  "extends": ["recommended"],
  "rules": {
    "pronoun-no-antecedent": "off",
    "long-sentence": "warn"
  },
  "ignore": ["docs/archive/**", "*.draft.md"],
  "options": {
    "long-sentence.max_words": 40,
    "taste-word.extra": ["bespoke", "revolutionary"]
  },
  "skip_spans": {
    "quoted_strings": true,
    "quoted_strings_max_chars": 40
  },
  "custom_rules": [
    {
      "id": "no-acme-brand",
      "pattern": "\\bAcme\\s+Corp\\b",
      "severity": "warn",
      "message": "Use 'ACME Inc' instead of 'Acme Corp'"
    }
  ]
}

custom_rules[] lets a team codify its own banned phrases without patching isolint. Each spec takes id (must not collide with a built-in rule), pattern (regex source), optional flags (default gi), optional severity (default warn), and message. Invalid specs are logged to stderr and skipped — one bad pattern never breaks a whole run.

Presets:

• recommended — deterministic reliability rules only; safe for CI.
• strict — recommended + all five LLM-assisted rules; requires --llm.
• performance — 18 advisory deterministic rules for harness efficiency.

Default (no config, no --preset): recommended + performance runs together. Performance findings are info severity so CI exit codes under the default --fail-on error are unaffected. Set {"extends": ["recommended"]} in .isolint.json or pass --preset recommended to keep only reliability rules.

Combine performance with either reliability preset, either via config:

{
  "extends": ["recommended", "performance"]
}

Or directly on the command line with --preset (repeatable, or comma-separated):

npx @razroo/isolint lint . --preset recommended --preset performance
# equivalent to:
npx @razroo/isolint lint . --preset recommended,performance

--preset overrides the config's extends when set — useful for one-off runs without editing .isolint.json. Valid values: recommended, strict, performance.

With --fix --llm, the performance preset can rewrite duplicated output contracts and redundant schema prose. perf-style-tone-overhead also has a deterministic fix for simple trailing tone/style suffixes.

The performance preset adds:

• perf-repeated-instruction-block
• perf-example-heavy-section
• perf-duplicated-output-requirement
• perf-step-restates-prior-step
• perf-low-value-prose-section
• perf-shared-prefix-budget
• perf-large-example-in-shared-prefix
• perf-long-runbook-in-shared-prefix
• perf-redundant-schema-prose
• perf-structured-output-explanation
• perf-style-tone-overhead
• perf-mirrored-agent-spec
• perf-rationale-in-shared-prefix — Why: / Historical note: / dated incident narratives in always-loaded files
• perf-emphasis-inflation — saturated MUST / NEVER / CRITICAL density (weak models ignore saturated emphasis)
• perf-cross-file-duplicate-block — same paragraph copy-pasted verbatim across ≥2 harness files
• perf-dense-prohibition-list — 3+ consecutive Do not X. Never Y. Must not Z. sentences that should be a bullet list
• perf-conditional-mode-branch-in-shared-prefix — When the orchestrator dispatches an \apply`…` branches that belong in the mode's own file
• perf-nested-conditional-chain — sentences chaining 3+ if/when/unless conditions that weak models can't track

isolint cost — what does my harness actually cost per turn?

Lint tells you what's wasteful; cost tells you how much you're paying.

# Quick view: always-loaded baseline + per-mode + per-agent breakdown
npx @razroo/isolint cost /path/to/harness

# Fail CI if the always-loaded cost exceeds a budget
npx @razroo/isolint cost /path/to/harness --budget 10000

# Machine-readable, for scripts / dashboards
npx @razroo/isolint cost /path/to/harness --format json

The command reports cost per tool load group because different tools load different subsets of the shared-prefix files — you pay one tool's bundle per turn, not all of them summed together:

• Claude Code loads CLAUDE.md.
• AGENTS.md convention (opencode / Codex CLI / Zed) loads AGENTS.md + modes/_shared.md + .opencode/instructions.md when those exist.
• Cursor loads .cursor/rules/*.mdc (frontmatter-aware alwaysApply filtering is not yet implemented, so the total is a ceiling).
• Per-mode files (modes/<name>.md) load only when that mode runs.
• Per-agent files (.claude/agents/, .opencode/agents/, iso/agents/) load only when the orchestrator dispatches.

iso/instructions.md is authoring source — it compiles to the tool-specific files at build time. When a tool's file isn't in the repo, iso stands in as the compiled-content equivalent so each tool's cost reflects what you actually pay at runtime.

For shared-prefix files, the text report breaks down the biggest sections so you can see exactly where the budget is going, then shows the per-tool totals:

Shared-prefix files (section breakdown)

    ~8,072 tokens   iso/instructions.md    4,679 words
      ~920   § Hard Limits — NEVER exceed these numbers            549w
      ~615   § Subagent Routing — which agent for which task       311w
      ~578   § Validation State Lags Behind Actual Field State     337w
    ~4,087 tokens   modes/_shared.md       2,184 words

Per-tool always-loaded cost (pick the tool you actually run)

  Claude Code (CLAUDE.md)                           ~8,072 tokens / turn
      ~8,072   iso/instructions.md
    note: CLAUDE.md not in repo; iso/instructions.md stands in as the compiled content.

  AGENTS.md convention (opencode / Codex / Zed)     ~12,159 tokens / turn
      ~8,072   iso/instructions.md
      ~4,087   modes/_shared.md
    note: AGENTS.md not in repo; iso/instructions.md stands in as the compiled content.

  Cursor (.cursor/rules)                             ~8,072 tokens / turn
      ~8,072   iso/instructions.md
    note: .cursor/rules/ not in repo; iso/instructions.md stands in as the compiled content.

  Worst-case tool: AGENTS.md convention at ~12,159 tokens / turn

Per-mode context (loads when that mode runs)
    ~4,549 tokens   modes/apply.md                2,630 words
    ...
  Worst case (worst tool + heaviest mode): ~16,708 tokens / turn

Token estimates use chars ÷ 4 — the same heuristic perf-shared-prefix-budget uses, so the numbers line up with the findings. Actual per-provider cost depends on the tokenizer.

CI guard. Put this in a workflow step to catch regressions when someone adds a 500-word "just one more thing" to a shared file:

- run: npx @razroo/isolint cost . --budget 10000

--budget exits non-zero (exit 1) when the worst-case tool total exceeds N tokens — i.e. the most expensive tool group's bundle, not a naive sum across tools.

Other flags: --no-sections hides the per-section breakdown, --no-gitignore disables the git allowlist (by default cost and lint both skip files git ignores — generated build output like CLAUDE.md from the agentmd / iso tooling isn't counted).

Working with agentmd

agentmd is a structured-markdown dialect for authoring LLM agent prompts: a # Agent: <name> H1, explicit ## Hard limits / ## Defaults sections, and rule items shaped as - [H1] claim with an indented why: rationale underneath. The dialect compiles down to the plain AGENTS.md / CLAUDE.md / Cursor rules that tools actually load.

isolint supports agentmd natively:

• Auto-detected. A file with a top-level # Agent: <name> heading is treated as agentmd. No config flag needed.
• Rationale stays. In plain harness prose, Why: paragraphs in a shared-prefix file are overhead — perf-rationale-in-shared-prefix flags them. In agentmd the rationale is load-bearing (the model uses why: to judge edge cases), so that rule skips agentmd files. You can keep rich why: on every rule without fighting the linter.
• Everything else applies. Every other rule — soft-imperative, taste-word, trailing-etc, long-sentence, the cross-file duplication and emphasis-inflation checks, all of it — runs normally on agentmd files. The dialect changes rationale semantics, not prose-quality expectations.
• Mixed harnesses work. One lint run can have plain-prose mode files alongside agentmd-authored agents; each file is judged by the dialect it's actually in.
• Generated outputs are skipped (with .gitignore honored by default). If you author with agentmd and compile to AGENTS.md + .cursor/rules/ + .opencode/agents/, only the source file is linted — not the N generated copies. That keeps findings in the file you actually edit.

If you're using agentmd and the linter is flagging your why: paragraphs, the most likely cause is that the H1 isn't # Agent: <name> (the detector looks for that exact shape). Rename the H1 and isolint will recognise the dialect on the next run.

Plan files

For JSON Plan files, use isolint validate --perf instead of the markdown linter. That path runs plan-specific performance checks such as:

• repeated or restated steps
• instructions that duplicate expected_output
• schema details repeated in prose
• structured outputs that also ask for explanation
• tone/style guidance on structured outputs
• long low-signal step instructions

isolint plan uses the same checks during generation. If the model emits a schema-valid plan with plan-performance findings, the planner retries with the formatted findings as repair feedback until the plan is clean or it runs out of attempts.

Skip spans

Rules never fire inside these spans:

| Span | Default | Why | | --- | --- | --- | | Fenced code blocks | on | Example input/output, not prose | | Inline code (`word`) | on | The word is being named | | HTML comments | on | Author notes | | Short double-quoted phrases (≤ 40 chars) | on | Avoid "leveraged", "cutting-edge" — words being named, not used. Full-sentence quoted directives (>40 chars) still lint. | | YAML/TOML frontmatter | on | Opencode modes, Claude Code agents, Cursor .mdc rules — structured metadata, not prose instructions | | > blockquotes | on | Usually example input/output or quoted prose, not harness instructions |

soft-imperative additionally skips findings inside questions (sentence ending in ?) — What story should they tell? is not an instruction.

Disable any span via skip_spans in config.

Section-aware severity

Every finding is tagged with its enclosing ## Heading. You can mute or downgrade rules inside specific sections:

{
  "section_severity": {
    "Examples": "off",
    "Notes": "info",
    "Changelog": "off"
  }
}

Keys are case-insensitive. Useful for docs-heavy harness files where ## Examples contains intentional bad prose. The text reporter groups findings by section so CI logs stay navigable.

Sentence context

Every finding carries its containing sentence — the tokenizer correctly handles abbreviations, filenames, decimals, URLs, and ellipses. This powers three things:

1. Text reports show in: "<sentence>" so you can act without opening the file.
2. --fix --llm coalesces multiple findings in the same sentence into one rewrite — no more conflicting edits.
3. JSON/SARIF expose sentence and section fields so downstream tools (Claude Code, review bots) have the full context.

Cross-file awareness

Three rules look beyond a single line or file:

• undefined-step-reference aggregates every ## Step N / ## Block X heading across every file in the lint set. A reference is flagged only if it isn't defined anywhere in the scanned harness. Multi-file harnesses that keep shared Blocks in _shared.md work out of the box.
• missing-file-reference reads the repo file list and flags any filename in prose (cv.md, profile.json, schema.yaml) that doesn't exist. Basename-matched, so references to cv.md from modes/apply.md resolve against data/candidates/cv.md.
• context-budget counts prose words (excluding code fences and frontmatter) per harness file. Override thresholds via "context-budget.info_words" / "context-budget.warn_words".

isolint verify — real-model empirical bridge

The simulator is deterministic. To see how the rewrites move the needle on your actual target model, use verify:

npx @razroo/isolint verify \
  --harness modes/classify.md \
  --input tests/sample.json \
  --small mistralai/mistral-7b-instruct \
  --large anthropic/claude-3.5-sonnet

verify runs the original harness through the small model, applies --fix --llm rewrites via the large model, runs the fixed harness through the small model again, and reports: simulator fragility before/after, harness size before/after (chars, words, approximate prompt tokens, sentences, performance findings), char-delta in output, JSON-validity before/after, and the raw model outputs for side-by-side inspection. First time the "small models don't break" claim can be validated empirically on the actual target.

Proof the fixes work

The tagline is "rewrites harnesses so small models don't break." A simulator codifies known 7B-class failure modes (drops should, invents items past etc., skips when relevant conditionals, loses the middle of long sentences, hallucinates dangling $input.X refs) and produces a fragility score — the fraction of instructions a weak model would break on.

The regression-sim test suite asserts that for every fixture, running isolint --fix strictly decreases the fragility score. Measured on the built-in fixtures:

| Fixture | Before fix | After fix | | --- | --- | --- | | classify | 0.75 (3 fail / 1 follow) | 0.00 (0 fail / 4 follow) | | multi-step | 0.75 (3 fail / 1 follow) | 0.25 (1 fail / 3 follow) |

CI fails if a rule change regresses either fixture. For the first time, the "so small models don't break" claim is a number, not an assertion.

Validated rewrites

Every LLM rewrite is re-linted before being applied. A rewrite is accepted only if:

1. The rule that triggered the fix no longer fires on the new text.
2. No new rules fire on the new text.
3. Markdown structure is preserved (heading, list, code-fence, inline-code, bold, and link counts all match).

If a rewrite fails validation, the linter retries once with explicit feedback about what went wrong. If the retry also fails, the fix is skipped and reported in the fix summary — no mangled prose ever lands on disk. This is what makes --fix --llm safe to run in CI.

Three mechanisms work together to improve rewrite quality:

• Few-shot examples per rule — each rule ships canonical bad → good pairs that are included in the prompt when that rule fires. Grounds the model in the intended fix direction instead of hoping it guesses. Every example is also a self-test: CI asserts that bad triggers the rule and good doesn't.
• Self-consistency sampling — 3 candidates are generated in parallel per attempt; the validator scores each; the lowest-problem candidate wins. Tunable via samples_per_attempt (default 3; set to 1 to disable).
• Feedback-driven retry — a rejected rewrite is fed back to the model with its specific validation problems ("rewrite still violates X"; "markdown structure changed"). The retry is targeted, not a cold reroll.

Rewrite stats

isolint lint --fix --llm --stats prints a per-rule table of rewrite outcomes — candidates, first-try accepts, accepts after retry, rejects, and an overall accept percentage. Low accept rates are a signal that the rule's examples or message need work.

Real-world result

Run against JobForge's modes/ directory (19 files, 1900+ lines of opencode harness prose):

$ npx @razroo/isolint lint /Users/you/JobForge/modes
...
19 files scanned — 54 warnings, 6 infos

Every finding points at a concrete phrase that a 7B model will handle worse than a Claude-class model. --fix --llm rewrites each one while preserving intent and markdown formatting.

Offline demo (no API key)

npx tsx examples/lint-demo/run.ts

Runs lint + fix end-to-end against a sample file with a mock rewriter.

Plans (the compiler mode)

The Plan format

A Plan is a pure data object (see src/schema/plan.ts):

• input_schema — JSON Schema for the user input.
• steps[] — ordered, atomic steps. Each step has:
  • instruction — a single imperative task.
  • inputs — explicit references: $input or $steps.<id>.
  • constraints — machine-checkable rules (length, regex, enum).
  • expected_output — text, enum, json+schema, or list.
  • failure_handling — retry | repair | fallback | fail.
• final_output — which step (or composition of steps) is the result.

Every plan is validated against src/schema/plan.ts before execution. Every step output is validated against its expected_output. Nothing is trusted.

The Runtime

For each step, the runtime:

1. Resolves inputs ($input, prior step outputs).
2. Builds a minimal prompt — instruction, labeled inputs, constraints, explicit output format. No persona preamble, no hidden reasoning.
3. Calls the small model.
4. Validates the output:
   • Output kind (text/enum/json/list) + schema.
   • Constraints (must_match / must_not_match regex).
5. On failure, follows the step's failure_handling:
   • retry — same prompt.
   • repair — append validator errors, ask for a fix.
   • fallback — return a deterministic value.
   • fail — abort the plan.

Install

If you just want the CLI, run it directly with npx:

npx @razroo/isolint lint ./modes

Or add it to a project:

npm install -D @razroo/isolint

If you're developing this repo itself:

npm install
npm run build

Requires Node ≥ 18.17.

Configure

Copy the checked-in example env file if you want local defaults:

cp .env.example .env

Preferred env vars:

OPENROUTER_API_KEY=sk-or-...
ISOLINT_LARGE=anthropic/claude-3.5-sonnet
ISOLINT_SMALL=mistralai/mistral-7b-instruct

Legacy ISOMODEL_LARGE / ISOMODEL_SMALL are still accepted for compatibility.

Force provider selection with --provider openrouter|openai|ollama|custom. OpenRouter and OpenAI work out of the box. Groq, Together, vLLM, and other OpenAI-compatible endpoints work via --provider custom --base-url.

CLI

lint and verify are covered above because they are the main linter workflows. The commands below are for the plan/runtime pipeline.

isolint plan

Ask the large model to generate a Plan for a new task.

npx @razroo/isolint plan \
  --task "Extract purchase orders from emails into {po_number, vendor, total}" \
  --out plans/po-extract.json

After writing the file, plan also runs the advisory plan-performance checks and prints any overhead findings immediately.

isolint run

Run an existing Plan through the small model.

npx @razroo/isolint run \
  --plan examples/multi-step-reasoning/plan.json \
  --input examples/multi-step-reasoning/input.json

Pipe stdin:

cat ticket.json | npx @razroo/isolint run --plan plan.json --input -

isolint validate

Schema-check a plan without running it.

npx @razroo/isolint validate --plan examples/cold-email/plan.json

Run the schema check plus advisory performance analysis:

npx @razroo/isolint validate \
  --plan examples/cold-email/plan.json \
  --perf

CI / local gate for bundled examples:

npm run check:plans

Examples

Three ready-to-run plans live in examples/:

| Example | What it shows | | ------------------------- | --------------------------------------------------------------- | | cold-email/ | Structured paragraphs with regex constraints (subject/hook/value/ask) | | data-extraction/ | Strict-schema JSON extraction + numeric computation over prior steps | | multi-step-reasoning/ | Support-ticket triage: classify → severity → root cause → next action |

Offline demo (no API key)

npx tsx examples/offline-demo/run.ts

Runs the multi-step-reasoning plan end-to-end against a mock small model. Proves the full loop: input validation → per-step execution → schema validation → final output composition.

Real run

npx @razroo/isolint run \
  --plan examples/data-extraction/plan.json \
  --input examples/data-extraction/input.json \
  --small mistralai/mistral-7b-instruct

Programmatic API

import { Planner, Runtime, createProvider } from "@razroo/isolint";

const plan = await new Planner(
  createProvider({ model: "anthropic/claude-3.5-sonnet" }),
).generate({ task: "Classify tickets and propose a next action" });

const result = await new Runtime(
  createProvider({ model: "mistralai/mistral-7b-instruct" }),
).run(plan.plan, { ticket: { subject: "...", body: "..." } });

console.log(result.final);

Design constraints (enforced)

• No long runtime prompts. All complexity lives in the Plan.
• No vague instructions. The planner system prompt forbids "be creative", "engaging", "appropriate", etc.
• No world-knowledge dependence. Every step must cite $input or a prior step as its fact source.
• No taste-based validation. Constraints must be regex, length, enum, or schema.
• Deterministic by default. Runtime temperature defaults to 0.

Tests

npm test

The suite covers lint rules, suppressions, fix engine, diff output, schema validation, JSON extraction, runtime retry/repair, fallback handling, and input-schema enforcement.

License

MIT