activegraph 1.0.0__py3-none-any.whl

activegraph/packs/diligence/prompts/document_researcher.md

@@ -0,0 +1,25 @@
+---
+version = "1.0.0"
+---
+You are a diligence researcher answering one question about a company.
+
+Workflow:
+1. Call `fetch_company_docs` with the company name to obtain a list of
+   relevant source documents.
+2. Identify the most promising 1-2 documents for the question at hand.
+3. Call `summarize_document` on each to extract its summary and key
+   facts.
+4. From the extracted facts, formulate claims that answer the question.
+   For each claim:
+   - Write the claim in one sentence, verbatim where possible.
+   - Assign a confidence score (0.0-1.0) calibrated to evidence
+     strength. Below 0.5: rumor or weak inference. 0.5-0.7: plausible
+     but unverified. 0.7-0.9: well-supported by the document. 0.9+:
+     directly stated, primary source.
+   - Provide a verbatim evidence_quote — the exact sentence in the
+     document that supports the claim.
+
+Do not invent claims. If a document does not address the question,
+return an empty claims list and a summary that says so.
+
+Return a single ResearchFindings object per the schema.

activegraph/packs/diligence/prompts/memo_synthesizer.md

@@ -0,0 +1,35 @@
+---
+version = "1.0.0"
+---
+You are synthesizing the final diligence memo for a company.
+
+The memo MUST have exactly these sections, in order:
+
+1. **summary**: 2-4 sentence executive summary. State the company,
+   the thesis question, and the bottom-line conclusion (or "open" if
+   unresolved).
+
+2. **thesis_questions_addressed**: a list, one entry per research
+   question with shape `{question, status, claim_ids}`. `status` is
+   "answered" or "unresolved". `claim_ids` lists the ids of claims
+   that address the question.
+
+3. **key_claims**: a list of the most material claims. Each entry has
+   shape `{claim_id, text, evidence_ids}`. Every claim entry MUST
+   include at least one evidence_id — uncited claims are not
+   permitted in the memo.
+
+4. **open_contradictions**: a list of unresolved contradictions, each
+   with shape `{contradiction_id, claim_a_text, claim_b_text}`. If
+   the run produced no contradictions, leave this list empty AND set
+   `contradictions_note` to the string "no contradictions found".
+
+5. **risks**: a list of the identified risks, each with shape
+   `{risk_id, title, severity, description}`. At least one risk MUST
+   be listed.
+
+Read the graph view carefully — the data you need is all there. Do
+not invent claims, evidence, contradictions, or risks. Use only
+what the graph contains.
+
+Return a single MemoBody object per the schema.

activegraph/packs/diligence/prompts/question_generator.md

@@ -0,0 +1,25 @@
+---
+version = "1.0.0"
+---
+You are generating the initial research questions for an investment
+diligence run on a single company.
+
+Read the triggering event (a `company.created` event) and the view of
+the graph (which contains the company object). Produce between
+min_questions and max_questions distinct research questions that, if
+answered, would constitute a thorough due-diligence picture of the
+company.
+
+Cover at least:
+- Financial performance (revenue, margins, trajectory)
+- Product (what they sell, how it's differentiated)
+- Customers (segments, concentration, retention)
+- Market (size, growth, competitive dynamics)
+- Risks (regulatory, technological, executional)
+
+Questions should be:
+- Specific and answerable, not vague
+- Each focused on one concern (no compound questions)
+- Phrased so a researcher knows what document/source to consult
+
+Return only the questions as a JSON list per the schema.

activegraph/packs/diligence/prompts/risk_identifier.md

@@ -0,0 +1,26 @@
+---
+version = "1.0.0"
+---
+You are identifying material risks for a company based on its
+accumulated claims.
+
+You receive the graph view containing the company, its claims, the
+documents those claims were derived from, and any detected
+contradictions. Read all of it before responding.
+
+For each risk you identify:
+- Title: short noun phrase (e.g. "customer concentration", "regulatory
+  exposure to FDA action").
+- Description: 1-3 sentences explaining the risk in concrete terms,
+  grounded in the observed claims.
+- Severity: "low", "medium", or "high".
+- related_claim_texts: list the verbatim text of the claims that
+  contribute to this risk. The handler will resolve these back to
+  claim ids; exact text match required.
+
+Identify at least one risk per company. If a contradiction was
+detected, it is itself a risk (severity at least "medium") and must
+appear in your output. Material risks only — do not list every
+hypothetical concern.
+
+Return a RiskList object per the schema.

activegraph/packs/diligence/settings.py

@@ -0,0 +1,60 @@
+"""Diligence pack settings. CONTRACT v0.9 #15.
+
+Every field has a default so `runtime.load_pack(diligence_pack)`
+without `settings=` works for the quickstart. The killer demo
+overrides explicitly to demonstrate the typed-injection API.
+"""
+
+from __future__ import annotations
+
+from pydantic import BaseModel, Field
+
+
+class DiligenceSettings(BaseModel):
+    """Configuration for the Diligence pack.
+
+    Accessed by behaviors in three forms (CONTRACT v0.9 #7):
+      1. Typed parameter injection (primary):
+         `def claim_extractor(event, graph, ctx, out, *,
+                              settings: DiligenceSettings): ...`
+      2. `ctx.settings.confidence_threshold_for_review`
+      3. `ctx.pack_settings("diligence")` for cross-pack lookups.
+    """
+
+    llm_model: str = Field(
+        default="claude-sonnet-4-5",
+        description="The Claude model used by all diligence LLM behaviors.",
+    )
+    max_documents_per_company: int = Field(
+        default=5, ge=1,
+        description="Cap on documents fetched per company. Bounds tool cost.",
+    )
+    max_claims_per_document: int = Field(
+        default=20, ge=1,
+        description="Cap on claims extracted per document. Bounds LLM cost and "
+                    "downstream evidence-linking work.",
+    )
+    confidence_threshold_for_review: float = Field(
+        default=0.7, ge=0.0, le=1.0,
+        description="Claims with confidence ≥ this threshold participate in "
+                    "contradiction detection.",
+    )
+    min_questions: int = Field(
+        default=8, ge=1,
+        description="Minimum number of thesis questions the generator produces.",
+    )
+    max_questions: int = Field(
+        default=15, ge=1,
+        description="Maximum number of thesis questions. Generator is one-shot "
+                    "in v0.9 (CONTRACT v0.9 #16); adaptive in v1.0.",
+    )
+    auto_approve_memos: bool = Field(
+        default=True,
+        description="When False, memo writes go through the memo_approval "
+                    "policy (CONTRACT v0.9 #15 / #9). True for quickstart.",
+    )
+    auto_approve_risks: bool = Field(
+        default=True,
+        description="When False, risk writes go through the risk_approval "
+                    "policy. True for quickstart.",
+    )

activegraph/packs/diligence/tools.py

@@ -0,0 +1,124 @@
+"""Diligence pack tools. CONTRACT v0.9 #15.
+
+Three pack-scoped tools backed by recorded fixtures. Production users
+swap these for real implementations:
+  - fetch_company_docs: would call a real fetcher (SEC, news APIs, etc.)
+  - search_filings: would query EDGAR or a similar service
+  - summarize_document: an LLM-backed summarizer with caching
+
+The fixtures live in `activegraph.packs.diligence.fixtures` per
+CONTRACT v0.9 #18 (fixtures ship with the pack).
+"""
+
+from __future__ import annotations
+
+from decimal import Decimal
+from typing import Optional
+
+from pydantic import BaseModel
+
+from activegraph.packs import tool
+
+
+# ---------------------------------------------------- I/O schemas
+
+
+class FetchCompanyDocsInput(BaseModel):
+    company_name: str
+    max_results: int = 5
+
+
+class DocumentRef(BaseModel):
+    title: str
+    url: str
+    summary: str = ""
+
+
+class FetchCompanyDocsOutput(BaseModel):
+    documents: list[DocumentRef]
+
+
+class SearchFilingsInput(BaseModel):
+    company_name: str
+    keyword: str = ""
+
+
+class SearchFilingsOutput(BaseModel):
+    filings: list[DocumentRef]
+
+
+class SummarizeDocumentInput(BaseModel):
+    url: str
+    max_words: int = 100
+
+
+class SummarizeDocumentOutput(BaseModel):
+    summary: str
+    extracted_facts: list[str] = []
+
+
+# ---------------------------------------------------- tool registration
+
+
+@tool(
+    name="fetch_company_docs",
+    description="Fetch a list of source documents for a company. "
+                "Returns up to max_results documents with title, URL, "
+                "and a one-line summary.",
+    input_schema=FetchCompanyDocsInput,
+    output_schema=FetchCompanyDocsOutput,
+    cost_per_call=Decimal("0.005"),
+    timeout_seconds=15.0,
+    deterministic=False,
+)
+def fetch_company_docs(args: FetchCompanyDocsInput, ctx) -> FetchCompanyDocsOutput:
+    """Backed by fixtures in v0.9. Production: real fetcher."""
+    from activegraph.packs.diligence.fixtures import lookup_company_docs
+
+    docs = lookup_company_docs(args.company_name, max_results=args.max_results)
+    return FetchCompanyDocsOutput(
+        documents=[DocumentRef(**d) for d in docs],
+    )
+
+
+@tool(
+    name="search_filings",
+    description="Search a company's regulatory filings by keyword. "
+                "Returns matching filings as document refs.",
+    input_schema=SearchFilingsInput,
+    output_schema=SearchFilingsOutput,
+    cost_per_call=Decimal("0.003"),
+    timeout_seconds=15.0,
+    deterministic=False,
+)
+def search_filings(args: SearchFilingsInput, ctx) -> SearchFilingsOutput:
+    """Backed by fixtures in v0.9. Production: EDGAR query."""
+    from activegraph.packs.diligence.fixtures import lookup_filings
+
+    filings = lookup_filings(args.company_name, args.keyword)
+    return SearchFilingsOutput(
+        filings=[DocumentRef(**f) for f in filings],
+    )
+
+
+@tool(
+    name="summarize_document",
+    description="Summarize a document at a given URL and extract key facts.",
+    input_schema=SummarizeDocumentInput,
+    output_schema=SummarizeDocumentOutput,
+    cost_per_call=Decimal("0.002"),
+    timeout_seconds=30.0,
+    deterministic=False,
+)
+def summarize_document(args: SummarizeDocumentInput, ctx) -> SummarizeDocumentOutput:
+    """Backed by fixtures in v0.9. Production: LLM summarizer."""
+    from activegraph.packs.diligence.fixtures import lookup_summary
+
+    result = lookup_summary(args.url)
+    return SummarizeDocumentOutput(
+        summary=result.get("summary", "")[: args.max_words * 8],  # rough char cap
+        extracted_facts=result.get("facts", []),
+    )
+
+
+TOOLS = [fetch_company_docs, search_filings, summarize_document]