docs: reposition as a general RLM harness (not security-specific); unify examples

rlm-kit is domain-agnostic — anything dspy.RLM can do (deep research, an
RSS-digest-to-webhook agent, structured extraction, detection authoring). The
"security tasks" framing was an artifact of the author's first use, not the
kit's scope. Generalize it everywhere; keep the real security PROPERTIES
(sandbox-is-the-boundary, SSRF guard, MCP host-side, SECURITY.md) since those
hold for any task that runs model-written code over untrusted input.

- pyproject: description -> "harness for building tasks on DSPy RLMs"; drop the
  `security` keyword (add `harness`) and the `Topic :: Security` classifier (add
  `Application Frameworks`).
- README/CLAUDE/CONTRIBUTING/__init__/task/fetch/sandbox: intros say "tasks", not
  "security tasks"; add a domain-agnostic positioning line; neutral quickstart.
- examples: SecurityTriage/RateSeverity -> neutral Research/Summarize tasks.
- Unify the two examples dirs: move rlm_kit/examples/harness_run.py to the
  top-level examples/ and delete the in-package rlm_kit/examples/. Examples no
  longer ship inside the importable wheel (conventional layout); the
  `examples/harness_run.py` references in task.py/README are now accurate.

uv build: wheel ships no examples, Summary updated. pytest 156 passed; ruff clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: qazbnm456

CLAUDE.md

@@ -1,7 +1,7 @@
 # rlm-kit — agent guide
 
 `rlm-kit` is a small, reusable scaffold over `dspy.RLM` (Recursive Language Models)
-for building security tasks. A task is a *declaration* — a `RLMTask` subclass with
+for building tasks (of any kind). A task is a *declaration* — a `RLMTask` subclass with
 a `signature`, `output_field`, optional `output_model`, `instructions`, and
 `tools`; retry/validation, sandbox selection, budget caps, and observability are
 inherited. See `README.md` for the full layout and usage.
@@ -124,7 +124,7 @@ One companion rule ships under `.claude/rules/`:
 
 ## Consumer-driven hardening
 
-- This kit is driven by a real downstream consumer (a security task that builds on the
+- This kit is driven by a real downstream consumer (a task that builds on the
   scaffold via an editable path dep). That dogfooding is the design loop: when the consumer
   forces a workaround, log the **reusable** gap and fix it in the kit — do not special-case
   the consumer. Generic mechanics get promoted here via the base/wrap split (a generic base +

CONTRIBUTING.md

@@ -1,7 +1,7 @@
 # Contributing to rlm-kit
 
 Thanks for helping improve `rlm-kit` — a small, reusable scaffold over `dspy.RLM`
-(Recursive Language Models) for building security tasks. This guide is the short
+(Recursive Language Models) for building tasks of any kind. This guide is the short
 version; the deep design rules live in [`CLAUDE.md`](./CLAUDE.md) and the
 extension contract in the README's [**Building a consumer**](./README.md#building-a-consumer).
 

README.md

@@ -1,6 +1,6 @@
 # rlm-kit
 
-A clean, reusable scaffold for building **security tasks** on top of
+A clean, reusable harness for building **any task** on top of
 [DSPy](https://dspy.ai)'s Recursive Language Model module (`dspy.RLM`).
 
 RLMs ([Zhang & Khattab, MIT, arXiv:2512.24601](https://arxiv.org/abs/2512.24601))
@@ -10,6 +10,11 @@ first-party implementation (Khattab co-authored both DSPy and the RLM paper) —
 works with existing Signatures and is optimizer-compatible (GEPA/MIPRO). This kit
 distills the boilerplate around it into one small, opinionated layer.
 
+**rlm-kit is domain-agnostic** — anything `dspy.RLM` can do fits: multi-hop "deep
+research", an RSS-digest agent that posts to a webhook, structured extraction,
+detection authoring, you name it. Security happens to be the author's own first
+use of it, but it isn't the kit's scope.
+
 ## Why this exists
 
 Using `dspy.RLM` directly leaves you re-writing the same plumbing for every task:
@@ -21,19 +26,19 @@ from rlm_kit import RLMConfig, RLMTask, configure
 from rlm_kit.tools import make_schema_validator
 from pydantic import BaseModel
 
-class Finding(BaseModel):
+class Article(BaseModel):
     title: str
-    severity: str
+    summary: str
 
-class TriageDiff(RLMTask):
-    signature = "diff: str -> finding: Finding"
-    output_field = "finding"
-    output_model = Finding
-    instructions = "You are a secure-code reviewer."
-    tools = [make_schema_validator(Finding)]
+class Summarize(RLMTask):
+    signature = "document: str -> article: Article"
+    output_field = "article"
+    output_model = Article
+    instructions = "Read the document and produce a title and a one-paragraph summary."
+    tools = [make_schema_validator(Article)]
 
 configure(RLMConfig.from_env())
-finding = TriageDiff().run(diff=big_patch_text)   # validated Finding
+article = Summarize().run(document=long_text)   # validated Article
 ```
 
 The retry loop, pydantic validation, sandbox selection, and budget caps are all
@@ -290,7 +295,7 @@ and `max_iterations` there so an offline reader reads the real per-run values, n
 ## Building a consumer
 
 `rlm-kit` is the ROLLOUT floor; a consumer is a thin declaration on top of it. `examples/harness_run.py`
-is a minimal worked example — a security-triage task that wires the sub-LM hook, skills, tracing, and
+is a minimal worked example — a task that wires the sub-LM hook, skills, tracing, and
 RL export together. Five steps:
 
 1. **Declare the task.** Subclass `RLMTask`: a `signature`, `output_field`, an `output_model`

rlm_kit/examples/harness_run.py examples/harness_run.pyrlm_kit/examples/harness_run.py renamed to examples/harness_run.py

@@ -32,22 +32,22 @@
 )
 
 
-class Finding(BaseModel):
-    title: str = Field(..., description="Short finding title.")
-    severity: str = Field(..., description="low|medium|high|critical")
+class Note(BaseModel):
+    title: str = Field(..., description="Short note title.")
+    takeaway: str = Field(..., description="the key takeaway, one line")
 
 
 def _non_empty(text: str):
     return None if text.strip() else "empty response"
 
 
-class SecurityTriage(RLMTask):
-    signature = "evidence: str -> finding: Finding"
-    output_field = "finding"
-    output_model = Finding
+class Research(RLMTask):
+    signature = "topic: str -> note: Note"
+    output_field = "note"
+    output_model = Note
     instructions = (
-        "You are a security analyst. Use list_skills/read_skill to consult "
-        "playbooks, then emit a Finding JSON."
+        "You are a research assistant. Use list_skills/read_skill to consult "
+        "reference notes, then emit a Note JSON."
     )
 
     def __init__(self, skills_dir: str, **kw):
@@ -65,15 +65,15 @@ async def main() -> None:
         base_sub, validators=[_non_empty], postprocessors=[str.strip], name="local-sub"
     )
 
-    task = SecurityTriage(
+    task = Research(
         skills_dir=os.getenv("SKILLS_DIR", "./skills"),
         sub_lm=intercepted_sub,
     )
 
     trace_path = "./traces/run.jsonl"
-    with TraceRecorder(trace_path, run_id="triage-001", meta={"task": "triage"}):
-        finding = await task.arun(evidence="...redacted evidence blob...")
-    print(finding.model_dump_json(indent=2))
+    with TraceRecorder(trace_path, run_id="research-001", meta={"task": "research"}):
+        note = await task.arun(topic="...the topic to research...")
+    print(note.model_dump_json(indent=2))
 
     # The same trace doubles as an Agentic-RL dataset source.
     runs = group_by_run(load_events(trace_path))

examples/mini_run.py

@@ -27,18 +27,18 @@
 )
 
 
-class Severity(BaseModel):
-    label: str = Field(..., description="one of: low, medium, high, critical")
-    reason: str = Field(..., description="one short sentence")
+class Summary(BaseModel):
+    title: str = Field(..., description="a short title")
+    gist: str = Field(..., description="one short sentence")
 
 
-class RateSeverity(RLMTask):
-    signature = "advisory: str -> severity: Severity"
-    output_field = "severity"
-    output_model = Severity
+class Summarize(RLMTask):
+    signature = "document: str -> summary: Summary"
+    output_field = "summary"
+    output_model = Summary
     instructions = (
-        "You are a vulnerability triage assistant. Read the advisory text and "
-        "return a Severity JSON. Keep it short."
+        "Read the document and return a Summary JSON (a title and a one-sentence gist). "
+        "Keep it short."
     )
 
 
@@ -47,15 +47,16 @@ async def main() -> None:
     print(f"main={cfg.main_model} sub={cfg.sub_model} interpreter={cfg.interpreter}")
 
     # Keep the loop tiny/cheap.
-    task = RateSeverity(max_retries=2)
+    task = Summarize(max_retries=2)
 
     trace_path = "./traces/mini_run.jsonl"
-    advisory = (
-        "A remote unauthenticated attacker can send a crafted HTTP request to the "
-        "admin endpoint and execute arbitrary OS commands as root."
+    document = (
+        "Recursive Language Models treat unbounded context as a variable in a sandboxed "
+        "REPL and recursively call sub-models over it, instead of stuffing everything into "
+        "one prompt."
     )
-    with TraceRecorder(trace_path, run_id="mini-001", meta={"task": "rate_severity"}):
-        result = await task.arun(advisory=advisory)
+    with TraceRecorder(trace_path, run_id="mini-001", meta={"task": "summarize"}):
+        result = await task.arun(document=document)
 
     print("\n=== RESULT ===")
     print(result.model_dump_json(indent=2))

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "rlm-kit"
 version = "0.2.0"
-description = "A clean, reusable scaffold for building security tasks on DSPy Recursive Language Models (RLM)."
+description = "A clean, reusable harness for building tasks on DSPy Recursive Language Models (RLMs)."
 readme = "README.md"
 requires-python = ">=3.11"
 license = "MIT"
@@ -14,7 +14,7 @@ keywords = [
     "llm",
     "agent",
     "reinforcement-learning",
-    "security",
+    "harness",
 ]
 classifiers = [
     "Development Status :: 4 - Beta",
@@ -24,8 +24,8 @@ classifiers = [
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Python :: 3.13",
     "Operating System :: OS Independent",
-    "Topic :: Security",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Topic :: Software Development :: Libraries :: Application Frameworks",
     "Typing :: Typed",
 ]
 dependencies = [

rlm_kit/__init__.py

@@ -1,4 +1,4 @@
-"""rlm-kit — a clean, reusable scaffold for security tasks on DSPy RLMs.
+"""rlm-kit — a clean, reusable harness for building tasks on DSPy RLMs.
 
 Public surface::
 

rlm_kit/examples/__init__.py

@@ -1,8 +1,8 @@
 """Sandbox / code-interpreter selection — the security boundary of the scaffold.
 
 RLM works by letting the model write and execute Python in a REPL. When that
-REPL is fed half-trusted scraped content (exactly the security use case this
-scaffold targets), the interpreter choice *is* the attack surface.
+REPL is fed half-trusted scraped content (a common case the moment a task pulls
+from the web or any untrusted source), the interpreter choice *is* the attack surface.
 
 Policy:
 

rlm_kit/sandbox.py

@@ -1,13 +1,13 @@
 """The ``RLMTask`` base class — the one abstraction this scaffold exists for.
 
-A security task is declared by subclassing ``RLMTask`` and filling four fields:
-
-    class TriageCommit(RLMTask):
-        signature = "diff: str -> finding: CommitFinding"
-        output_field = "finding"
-        output_model = CommitFinding          # a pydantic BaseModel
-        instructions = "You are a secure-code reviewer..."
-        tools = [make_schema_validator(CommitFinding)]
+A task is declared by subclassing ``RLMTask`` and filling four fields:
+
+    class Summarize(RLMTask):
+        signature = "document: str -> article: Article"
+        output_field = "article"
+        output_model = Article                 # a pydantic BaseModel
+        instructions = "Summarize the document into a title and a paragraph."
+        tools = [make_schema_validator(Article)]
 
 Everything else — building ``dspy.RLM``, choosing the sandbox, budget caps,
 retrying on validation failure, observability — is inherited. A consumer's