agentforge-py 0.2.1__py3-none-any.whl

agentforge/templates/_shared/docs/runbooks/18-add-hybrid-search.md

@@ -0,0 +1,78 @@
+# 18 — Add hybrid search (BM25 + vector)
+
+> **Goal:** combine keyword (BM25) and semantic (vector) recall
+> so queries that hit a specific term don't get out-voted by
+> "close-but-wrong" embeddings.
+> **Time:** ~10 minutes.
+> **Prereqs:** runbook 08 (retrieval wired).
+
+## TL;DR
+
+```yaml
+# agentforge.yaml
+retrieval:
+  mode: hybrid               # one of: vector | hybrid | bm25
+  embedder:
+    driver: openai
+    config: {model: text-embedding-3-small}
+  vector_store:
+    driver: postgres         # native FTS via tsvector + ts_rank_cd
+    config:
+      dsn: $POSTGRES_DSN
+      table: docs
+      hybrid:
+        alpha: 0.6           # weight on semantic (0=BM25-only, 1=vector-only)
+  top_k: 8
+```
+
+## Step by step
+
+1. **Pick a vector store with native hybrid.** Every shipped
+   driver supports `lexical_search` as of v0.2:
+   - `postgres` — Postgres `tsvector` + `ts_rank_cd`
+   - `sqlite` — SQLite FTS5 + `bm25`
+   - `neo4j` — Neo4j fulltext index + `score`
+   - `surrealdb` — SurrealDB `DEFINE ANALYZER` + `SEARCH ANALYZER ... BM25`
+   - `in_memory` — pure-Python BM25 (good for tests)
+2. **Set `retrieval.mode: hybrid`.** The `Retriever` then runs
+   the vector pass and the lexical pass in parallel and fuses
+   the result lists via Reciprocal Rank Fusion (RRF).
+3. **Tune `alpha`.** Higher `alpha` weights vector recall;
+   lower weights BM25. Start at 0.6 and bisect against your
+   golden-set evaluator.
+4. **(Optional) Add a reranker** on top — see runbook 17.
+   Hybrid → rerank is the canonical RAG quality stack.
+5. **Index your corpus.** Postgres / SQLite need a one-time
+   index creation (the v0.2 migration framework handles this
+   — see runbook 20).
+
+## Variations
+
+- **bm25 only** — for keyword-dominant corpora (logs, code).
+  `mode: bm25` skips the embedder entirely.
+- **Tenant-scoped hybrid** — pass `tenant_id` through the
+  retriever; both passes filter on it before fusion.
+- **Async indexing** — large corpora benefit from a background
+  job that bulk-inserts then refreshes the lexical index.
+
+## Troubleshooting
+
+| Symptom | Cause | Fix |
+|---|---|---|
+| Hybrid returns same results as vector | `alpha=1.0` or BM25 index missing | tune `alpha` and confirm the lexical migration applied |
+| Latency 2x vector-only | sequential passes | upgrade `agentforge-memory-*` to v0.2 — the passes run in parallel |
+| `lexical_search not supported` | older driver version | upgrade the driver package |
+| Top BM25 hits dominate | low `alpha` + short queries | raise `alpha` toward 0.7 |
+
+## Related
+
+- Runbook 08 — Add memory + retrieval
+- Runbook 17 — Add a reranker
+- Runbook 20 — Apply schema migrations
+- Feature spec: `docs/features/feat-022-hybrid-search.md`
+- Feature spec: `docs/features/feat-025-neo4j-vector-store.md`
+
+<!-- agentforge:end-managed -->
+
+<!-- agentforge:custom -->
+<!-- agentforge:end-custom -->

agentforge/templates/_shared/docs/runbooks/19-add-graphrag.md

@@ -0,0 +1,83 @@
+# 19 — Add GraphRAG (graph-augmented retrieval)
+
+> **Goal:** enrich top-k retrieval with N-hop neighbours from a
+> graph store, so an answer about "Alice" also pulls in
+> "Alice's manager" and "Alice's recent project".
+> **Time:** ~15 minutes.
+> **Prereqs:** runbooks 08 + 18. A populated `GraphStore`.
+
+## TL;DR
+
+```yaml
+# agentforge.yaml
+retrieval:
+  mode: hybrid
+  embedder:
+    driver: openai
+    config: {model: text-embedding-3-small}
+  vector_store:
+    driver: postgres
+    config: {dsn: $POSTGRES_DSN, table: docs}
+  graph_expansion:             # the new bit
+    graph_store: people_graph  # named graph store from modules.graph_stores
+    max_hops: 1
+    relations: [reports_to, works_on]
+    score_decay: 0.5            # neighbour score = parent_score * decay
+  top_k: 12
+modules:
+  graph_stores:
+    people_graph:
+      driver: neo4j
+      config: {uri: $NEO4J_URI, user: neo4j, password: $NEO4J_PASSWORD}
+```
+
+## Step by step
+
+1. **Decide which entities you'll expand.** Pick a small set
+   of relation types (`reports_to`, `cited_by`, ...). Expanding
+   on everything destroys precision; one or two hops is enough.
+2. **Populate the graph.** Either pre-build it (ETL job) or use
+   the agent itself: the `agentforge_core.contracts.graph`
+   surface has `add_triple` / `query_neighbors`.
+3. **Add a `graph_expansion:` block** under `retrieval:`. The
+   `Retriever` runs vector / hybrid first, then expands each
+   top-k hit's entity via the named `GraphStore`, applies the
+   `score_decay` per hop, and re-ranks the merged list.
+4. **Set `max_hops` conservatively.** Start with `1`; `2` is
+   for "explain why" queries where context chains matter.
+5. **Tune `score_decay`.** `0.5` halves a neighbour's score
+   versus its parent at hop 1; lower decay = neighbours
+   compete; higher decay = neighbours augment without
+   overwhelming.
+
+## Variations
+
+- **Composes with anything else.** GraphRAG is post-retrieve,
+  so it works with `mode: vector` or `hybrid`, with or without
+  a reranker (rerank happens after expansion).
+- **Sparse expansion** — only expand the top-1 hit
+  (`expand_top_n: 1`) when you want extra context only for the
+  best match.
+- **Different graph per tenant** — name several graph stores
+  in `modules.graph_stores` and switch via dotted-path env
+  overrides (`AGENTFORGE_RETRIEVAL__GRAPH_EXPANSION__GRAPH_STORE=tenant_graph`).
+
+## Troubleshooting
+
+| Symptom | Cause | Fix |
+|---|---|---|
+| No neighbours returned | entity IDs in vector store don't match graph node IDs | confirm the `id` field matches across stores |
+| Latency 3-4x | hop 2+ on a wide graph | drop `max_hops` to 1, narrow `relations` |
+| Neighbour scores too high | `score_decay` ≥ 0.8 | lower toward 0.5 |
+| Wrong neighbours | over-broad `relations` list | pin to the specific relations the query cares about |
+
+## Related
+
+- Runbook 08 — Add memory + retrieval
+- Runbook 18 — Add hybrid search
+- Feature spec: `docs/features/feat-023-graphrag-hybrid.md`
+
+<!-- agentforge:end-managed -->
+
+<!-- agentforge:custom -->
+<!-- agentforge:end-custom -->

agentforge/templates/_shared/docs/runbooks/20-apply-schema-migrations.md

@@ -0,0 +1,92 @@
+# 20 — Apply schema migrations
+
+> **Goal:** evolve a persistent store's schema (Postgres /
+> SQLite / Neo4j / SurrealDB) without losing data or
+> hand-running SQL on production.
+> **Time:** ~10 minutes for the first migration.
+> **Prereqs:** runbook 08 (a persistent store wired in).
+
+## TL;DR
+
+```bash
+# 1. Generate a migration file from the template.
+agentforge db migrate new "add_user_role_column"
+
+# 2. Edit the generated migration:
+# .agentforge/migrations/postgres/20260514T101200__add_user_role_column.sql
+# Add your DDL/DML under -- up
+
+# 3. Apply pending migrations.
+agentforge db migrate up
+
+# 4. Confirm.
+agentforge db migrate status
+```
+
+## Step by step
+
+1. **Pick the target store.** Migrations are per-driver. The
+   v0.2 framework ships:
+   - Postgres (`.agentforge/migrations/postgres/*.sql`)
+   - SQLite (`.agentforge/migrations/sqlite/*.sql`)
+   - Neo4j (`.agentforge/migrations/neo4j/*.cypher`)
+   - SurrealDB (`.agentforge/migrations/surrealdb/*.surql`)
+2. **Generate a migration.** `agentforge db migrate new
+   "<description>"` writes a timestamped file under the
+   appropriate driver directory.
+3. **Author the `-- up` block.** Use idempotent DDL where you
+   can (`CREATE INDEX IF NOT EXISTS ...`); the migrator tracks
+   applied state in a `__agentforge_migrations` table.
+4. **Parameterise dimension-sensitive schemas.** Postgres
+   `vector(N)` and SurrealDB `HNSW DIMENSION N` should use
+   `${VECTOR_DIM}` so the same migration adapts to
+   3-small (1536) vs 3-large (3072) without forking:
+
+   ```sql
+   -- up
+   CREATE TABLE docs (
+     id text primary key,
+     text text,
+     embedding vector(${VECTOR_DIM})
+   );
+   ```
+
+   Set `${VECTOR_DIM}` via `agentforge db migrate up
+   --var VECTOR_DIM=1536` or in `agentforge.yaml`
+   under `db.migrations.variables`.
+5. **Run `up`.** `agentforge db migrate up` applies every
+   pending migration in timestamp order, in a transaction
+   per file (where the driver supports it).
+6. **Verify with `status`.** `agentforge db migrate status`
+   prints applied + pending counts.
+
+## Variations
+
+- **CI gate.** Add `agentforge db migrate status
+  --fail-on-pending` to your deploy pipeline to refuse a
+  rollout if migrations are unapplied.
+- **Multiple stores.** Run `migrate up` per-store; the CLI
+  detects available drivers via your config.
+- **`init_schema` shim** — pre-v0.2 agents that called
+  `MemoryStore.init_schema()` still work; you can opt-out by
+  setting `db.init_schema: false` and switching to migrations.
+
+## Troubleshooting
+
+| Symptom | Cause | Fix |
+|---|---|---|
+| `MigrationConflict` | two devs added migrations with the same timestamp | regenerate one to bump the timestamp |
+| Migration applied but data missing | non-idempotent up that crashed mid-flight | manually fix; record entry in `__agentforge_migrations` |
+| `${VAR} not substituted` | variable not declared | pass `--var VAR=value` or add to `db.migrations.variables` |
+| Dim mismatch on `vector(N)` | upgrading embedding model | new migration that recreates the column with the new dim |
+
+## Related
+
+- Runbook 08 — Add memory + retrieval
+- Runbook 18 — Add hybrid search
+- Feature spec: `docs/features/feat-024-schema-migrations.md`
+
+<!-- agentforge:end-managed -->
+
+<!-- agentforge:custom -->
+<!-- agentforge:end-custom -->

agentforge/templates/_shared/docs/runbooks/21-use-streaming-guardrails.md

@@ -0,0 +1,82 @@
+# 21 — Use streaming guardrails (sentence-window)
+
+> **Goal:** stream tokens to your end-user while still running
+> the output guardrails — so PII or unsafe content never
+> reaches the client mid-stream.
+> **Time:** ~5 minutes.
+> **Prereqs:** runbooks 11 + 20 (chat sessions + guardrails).
+
+## TL;DR
+
+```yaml
+# agentforge.yaml
+modules:
+  chat:
+    session:
+      safety_mode: sentence-window     # the new bit
+```
+
+```python
+async for chunk in session.stream(turn):
+    # `chunk.content` is sentence-sized — never per-token.
+    # Output validators have already vetted (and possibly
+    # redacted) the text before you see it.
+    await client.send(chunk.content)
+```
+
+## Step by step
+
+1. **Decide your latency / safety trade-off.** Three
+   `safety_mode` settings on `ChatSession`:
+   - `buffer-then-stream` (default) — agent finishes, output
+     validators run once on the full answer, the chat layer
+     segments the result for the wire. Safe, smooth — but the
+     user sees zero tokens until the run finishes.
+   - `sentence-window` — each token is buffered until a
+     sentence boundary; the completed sentence runs through
+     `OutputValidator.check_output`; the validated text emits
+     downstream. Mid-latency, safe.
+   - `stream-then-redact` — currently aliases sentence-window;
+     v0.3 may add inline-regex redaction without buffering.
+2. **Set `modules.chat.session.safety_mode: sentence-window`**
+   in `agentforge.yaml`. `build_chat_session_from_config`
+   forwards the value into `ChatSession(safety_mode=...)`.
+3. **Make sure your `OutputValidator`s are sentence-friendly.**
+   Validators like `pii_redact_basic` work per-sentence
+   naturally. Validators that need the full answer (e.g.,
+   policy-level "did the model commit to anything risky?")
+   should stay on `buffer-then-stream`.
+4. **Test against a known leak.** Drive the agent with a
+   prompt that should produce a redacted string mid-output
+   (e.g., an API key); assert your sentinel never appears in
+   any streamed `ChatChunk`.
+
+## Variations
+
+- **Custom boundary regex.** Subclass `ChatSession` and
+  override `_make_window_buffer` to tweak the sentence
+  segmentation (English `.!?` default).
+- **Stream-then-redact (v0.3).** When inline regex redaction
+  lands, `stream-then-redact` will skip buffering for
+  validators that don't need cross-token context.
+- **Multi-language.** Today's segmentation is English-centric;
+  CJK / Arabic users should stay on `buffer-then-stream`
+  until v0.3 multi-language support ships.
+
+## Troubleshooting
+
+| Symptom | Cause | Fix |
+|---|---|---|
+| Stream looks delayed | sentence boundaries far apart in long unpunctuated text | the 200-char hard cap fires automatically; check the validator isn't slow |
+| Redacted text appears partially | a validator that mutates token-by-token | switch the validator to a sentence-level check |
+| `GuardrailViolation` stops the stream | violator surfaced from a per-sentence check | document the policy to the user; surface as a polite refusal chunk |
+
+## Related
+
+- Runbook 11 — Add safety guardrails
+- Feature spec: `docs/features/feat-020-chat-agents.md`
+
+<!-- agentforge:end-managed -->
+
+<!-- agentforge:custom -->
+<!-- agentforge:end-custom -->

agentforge/templates/_shared/docs/runbooks/README.md.tmpl

@@ -0,0 +1,68 @@
+# {{ project_slug }} — runbook index
+
+This directory carries the AgentForge runbooks for
+**{{ project_slug }}**. Every runbook is task-oriented: one job,
+one ~5-minute read, ready-to-copy code.
+
+## Day-1 reading
+
+| # | Title |
+|---|---|
+| 01 | [Set up a new agent](./01-set-up-new-agent.md) |
+| 02 | [Add a tool](./02-add-a-tool.md) |
+| 05 | [Write prompts](./05-write-prompts.md) |
+| 06 | [Test your agent](./06-test-your-agent.md) |
+
+## When you need it
+
+| # | Title |
+|---|---|
+| 03 | [Add a pipeline task](./03-add-a-pipeline-task.md) |
+| 04 | [Pick a reasoning strategy](./04-pick-reasoning-strategy.md) |
+| 07 | [Debug a run](./07-debug-a-run.md) |
+| 08 | [Add memory / persistence](./08-add-memory.md) |
+| 09 | [Add MCP servers](./09-add-mcp.md) |
+| 10 | [Add evaluators](./10-add-evaluators.md) |
+| 11 | [Add safety guardrails](./11-add-safety-guardrails.md) |
+| 12 | [Add observability](./12-add-observability.md) |
+| 13 | [Configure multi-provider](./13-configure-multi-provider.md) |
+
+## Retrieval & RAG
+
+| # | Title |
+|---|---|
+| 17 | [Add a reranker](./17-add-reranker.md) |
+| 18 | [Add hybrid search (BM25 + vector)](./18-add-hybrid-search.md) |
+| 19 | [Add GraphRAG (graph-augmented retrieval)](./19-add-graphrag.md) |
+
+## Streaming & schema
+
+| # | Title |
+|---|---|
+| 20 | [Apply schema migrations](./20-apply-schema-migrations.md) |
+| 21 | [Use streaming guardrails (sentence-window)](./21-use-streaming-guardrails.md) |
+
+## Operations
+
+| # | Title |
+|---|---|
+| 14 | [Deploy your agent](./14-deploy-your-agent.md) |
+| 15 | [Upgrade your agent](./15-upgrade-your-agent.md) |
+| 16 | [Configuration reference](./16-configuration-reference.md) |
+
+## Conventions
+
+- Every runbook is managed by the framework — opening it for
+  edit, you'll see a `<!-- agentforge:end-managed -->` marker.
+  Add project-specific notes below it; `agentforge upgrade`
+  preserves your section.
+- AI assistants read `AGENTS.md` at the project root before
+  suggesting changes. The runbooks are the long-form companion
+  to that file.
+- `agentforge docs` opens any runbook by topic (e.g.
+  `agentforge docs add-tool` jumps to runbook 02).
+
+<!-- agentforge:end-managed -->
+
+<!-- agentforge:custom -->
+<!-- agentforge:end-custom -->

agentforge/templates/code-reviewer/.env.example

@@ -0,0 +1,8 @@
+# Credentials for {{ project_name }}.
+{% if llm_provider == 'bedrock' -%}
+AWS_REGION=us-east-1
+{% elif llm_provider == 'anthropic' -%}
+ANTHROPIC_API_KEY=
+{% else -%}
+OPENAI_API_KEY=
+{% endif %}

agentforge/templates/code-reviewer/.gitignore

@@ -0,0 +1,7 @@
+.env
+.venv/
+__pycache__/
+*.pyc
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/

agentforge/templates/code-reviewer/README.md

@@ -0,0 +1,12 @@
+# {{ project_name }}
+
+{{ description }} Reviews a diff and emits structured `SimpleFinding`s.
+
+```bash
+uv sync
+cp .env.example .env
+python -m {{ project_slug | replace('-', '_') }} "review the diff at /path/to.patch"
+```
+
+Drop in evaluator graders (faithfulness / correctness) once you
+have rubric examples — see the commented stub in `agentforge.yaml`.

agentforge/templates/code-reviewer/agentforge.yaml

@@ -0,0 +1,23 @@
+agent:
+  name: "{{ project_slug }}"
+  model: "{% if llm_provider == 'bedrock' %}bedrock:us.anthropic.claude-sonnet-4-5-20250929{% elif llm_provider == 'anthropic' %}anthropic:claude-sonnet-4-5{% else %}openai:gpt-4o{% endif %}"
+  system_prompt: |
+    You are a careful code reviewer. For each issue, emit a
+    SimpleFinding with severity, category, file, line, and a
+    concrete recommendation. Stay grounded in the diff — don't
+    invent issues that aren't present.
+  budget:
+    usd: 3.0
+  max_iterations: 30
+
+modules:
+  evaluators:
+    # Wire the geval-based judges once you have rubric examples.
+    # - name: faithfulness
+    #   config:
+    #     sources_field: "retrieved_docs"
+    []
+
+logging:
+  level: "INFO"
+  format: "text"

agentforge/templates/code-reviewer/copier.yml

@@ -0,0 +1,34 @@
+# Copier template: code-reviewer agent (feat-011).
+#
+# Reviews code diffs / PRs and emits SimpleFinding issues. Wires
+# a file_read tool by default so the reviewer can pull source
+# context. Add evaluator graders (faithfulness, correctness) as
+# you build out the rubric.
+
+_min_copier_version: "9.4.0"
+_answers_file: ".agentforge-state/answers.yml"
+_templates_suffix: ""
+
+project_name:
+  type: str
+  help: "Human-readable project name."
+  default: "Code Reviewer"
+
+project_slug:
+  type: str
+  help: "Project slug (kebab-case)."
+  default: "code-reviewer"
+
+llm_provider:
+  type: str
+  choices:
+    Bedrock (AWS): bedrock
+    Anthropic (direct): anthropic
+    OpenAI: openai
+  default: bedrock
+
+description:
+  type: str
+  default: "A code-reviewing AgentForge agent."
+
+_template_name: "code-reviewer"

agentforge/templates/code-reviewer/pyproject.toml

@@ -0,0 +1,18 @@
+[project]
+name = "{{ project_slug }}"
+version = "0.0.0"
+description = "{{ description }}"
+requires-python = ">=3.13"
+dependencies = [
+    "agentforge-py",
+{%- if llm_provider == 'bedrock' %}
+    "agentforge-bedrock",
+{%- endif %}
+]
+
+[build-system]
+requires = ["hatchling>=1.27"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/{{ project_slug | replace('-', '_') }}"]

agentforge/templates/code-reviewer/src/{{project_slug.replace('-', '_')}}/__init__.py

@@ -0,0 +1,5 @@
+"""{{ project_name }}: {{ description }}"""
+
+from {{ project_slug | replace('-', '_') }}.main import run_agent
+
+__all__ = ["run_agent"]

agentforge/templates/code-reviewer/src/{{project_slug.replace('-', '_')}}/main.py

@@ -0,0 +1,32 @@
+"""Entry point for {{ project_name }}.
+
+Reviews a diff. The agent's tools include `file_read` so it can
+pull source context beyond what the diff shows. Outputs are
+typically `SimpleFinding`s (feat-008) — see `agentforge.findings`.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import sys
+
+from agentforge import Agent
+from agentforge.tools import file_read
+
+
+async def run_agent(task: str) -> str:
+    async with Agent(tools=[file_read]) as agent:
+        result = await agent.run(task)
+        return str(result.output)
+
+
+def main() -> None:
+    if len(sys.argv) < 2:
+        print('Usage: python -m {{ project_slug | replace("-", "_") }} "<task>"')
+        sys.exit(1)
+    output = asyncio.run(run_agent(" ".join(sys.argv[1:])))
+    print(output)
+
+
+if __name__ == "__main__":
+    main()

agentforge/templates/docs-qa/.env.example

@@ -0,0 +1,8 @@
+# Credentials for {{ project_name }}.
+{% if llm_provider == 'bedrock' -%}
+AWS_REGION=us-east-1
+{% elif llm_provider == 'anthropic' -%}
+ANTHROPIC_API_KEY=
+{% else -%}
+OPENAI_API_KEY=
+{% endif %}

agentforge/templates/docs-qa/.gitignore

@@ -0,0 +1,7 @@
+.env
+.venv/
+__pycache__/
+*.pyc
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/

agentforge/templates/docs-qa/README.md

@@ -0,0 +1,14 @@
+# {{ project_name }}
+
+{{ description }} Answers questions from documentation / source.
+Outputs `NarrativeFinding`s — markdown prose with citations.
+
+```bash
+uv sync
+cp .env.example .env
+python -m {{ project_slug | replace('-', '_') }} "how does the auth flow work?"
+```
+
+For real RAG, install `agentforge-memory-sqlite` (or postgres),
+wire `modules.memory.driver` and `modules.retriever` in
+`agentforge.yaml`, and index your corpus.

agentforge/templates/docs-qa/agentforge.yaml

@@ -0,0 +1,19 @@
+agent:
+  name: "{{ project_slug }}"
+  model: "{% if llm_provider == 'bedrock' %}bedrock:us.anthropic.claude-sonnet-4-5-20250929{% elif llm_provider == 'anthropic' %}anthropic:claude-sonnet-4-5{% else %}openai:gpt-4o{% endif %}"
+  system_prompt: |
+    You answer questions using only the supplied documentation /
+    source. Emit a `NarrativeFinding` with the answer as `body`
+    (markdown ok), every citation in `references` (path:line or
+    URL), and `category="answer"`. Never speculate beyond the
+    sources.
+  budget:
+    usd: 2.0
+  max_iterations: 30
+
+# When you have a corpus to index, wire memory.driver: sqlite (or
+# postgres) and modules.retriever to enable RAG.
+
+logging:
+  level: "INFO"
+  format: "text"

agentforge/templates/docs-qa/copier.yml

@@ -0,0 +1,31 @@
+# Copier template: docs-qa agent (feat-011).
+#
+# Answers questions from documentation / source. Emits
+# NarrativeFinding (feat-008) — markdown prose with citations.
+# Add a vector store + Retriever once you have a corpus to index.
+
+_min_copier_version: "9.4.0"
+_answers_file: ".agentforge-state/answers.yml"
+_templates_suffix: ""
+
+project_name:
+  type: str
+  default: "Docs QA"
+
+project_slug:
+  type: str
+  default: "docs-qa"
+
+llm_provider:
+  type: str
+  choices:
+    Bedrock (AWS): bedrock
+    Anthropic (direct): anthropic
+    OpenAI: openai
+  default: bedrock
+
+description:
+  type: str
+  default: "A docs / source Q&A AgentForge agent."
+
+_template_name: "docs-qa"

agentforge/templates/docs-qa/pyproject.toml

@@ -0,0 +1,18 @@
+[project]
+name = "{{ project_slug }}"
+version = "0.0.0"
+description = "{{ description }}"
+requires-python = ">=3.13"
+dependencies = [
+    "agentforge-py",
+{%- if llm_provider == 'bedrock' %}
+    "agentforge-bedrock",
+{%- endif %}
+]
+
+[build-system]
+requires = ["hatchling>=1.27"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/{{ project_slug | replace('-', '_') }}"]

agentforge/templates/docs-qa/src/{{project_slug.replace('-', '_')}}/__init__.py

@@ -0,0 +1,5 @@
+"""{{ project_name }}: {{ description }}"""
+
+from {{ project_slug | replace('-', '_') }}.main import run_agent
+
+__all__ = ["run_agent"]