From bbfc9f9bd519e98a89030d54c4d29a08355c9cb4 Mon Sep 17 00:00:00 2001
From: handnewb <61999949+handnewb@users.noreply.github.com>
Date: Tue, 11 Aug 2026 17:55:35 -0300
Subject: [PATCH] fix(llamacpp): pass extra_body through factory to LlamaCppLLM
 and OpenAICompatibleLLM delegate
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Fixes #3326: The create_llm_provider() factory accepted extra_body but never
passed it to LlamaCppLLM, so HINDSIGHT_API_LLM_EXTRA_BODY was silently
ignored for the llamacpp provider while working for every other provider
(Bedrock, Fireworks, OpenAI-compatible, LiteLLM, etc.).

The fix threads extra_body from the factory → LlamaCppLLM.__init__ →
OpenAICompatibleLLM delegate, which already supports the parameter.

fix(structured-doc): detect markdown tables without leading/trailing pipes

Fixes #3361: The table parser required pipes at both start AND end of every
row (^\s*\|.*\|\s*$), so GFM-compliant tables like "Col1 | Col2\n--- | ---"
fell through to ParagraphBlock which joins all lines with spaces, permanently
destroying table structure during delta refresh.

The fix adds a lenient fallback: if a chunk contains a separator line AND every
non-separator line contains at least one pipe character, it is treated as a
table. This covers the common GFM shorthand without false-positiving on
non-table content.
---
 .../hindsight_api/engine/llm_wrapper.py            |  1 +
 .../hindsight_api/engine/providers/llamacpp_llm.py |  3 +++
 .../hindsight_api/engine/reflect/structured_doc.py | 11 +++++++++++
 .../references/sdks/integrations/agent-plugin.md   | 14 ++++++++++++++
 4 files changed, 29 insertions(+)
 create mode 100644 skills/hindsight-docs/references/sdks/integrations/agent-plugin.md

diff --git a/hindsight-api-slim/hindsight_api/engine/llm_wrapper.py b/hindsight-api-slim/hindsight_api/engine/llm_wrapper.py
index 3a6704b4b..7f43d67cd 100644
--- a/hindsight-api-slim/hindsight_api/engine/llm_wrapper.py
+++ b/hindsight-api-slim/hindsight_api/engine/llm_wrapper.py
@@ -513,6 +513,7 @@ def create_llm_provider(
             base_url=base_url,
             model=model,
             reasoning_effort=reasoning_effort,
+            extra_body=extra_body,
             model_path=config.llamacpp_model_path,
             gpu_layers=config.llamacpp_gpu_layers,
             context_size=config.llamacpp_context_size,
diff --git a/hindsight-api-slim/hindsight_api/engine/providers/llamacpp_llm.py b/hindsight-api-slim/hindsight_api/engine/providers/llamacpp_llm.py
index 35541b726..d96e75f43 100644
--- a/hindsight-api-slim/hindsight_api/engine/providers/llamacpp_llm.py
+++ b/hindsight-api-slim/hindsight_api/engine/providers/llamacpp_llm.py
@@ -274,6 +274,7 @@ def __init__(
         base_url: str,
         model: str,
         reasoning_effort: str = "low",
+        extra_body: dict[str, Any] | None = None,
         model_path: str | None = None,
         gpu_layers: int = -1,
         context_size: int = 8192,
@@ -295,6 +296,7 @@ def __init__(
         self._chat_format = chat_format
         self._no_grammar = no_grammar
         self._extra_args = extra_args
+        self._extra_body = extra_body
         self._server: LlamaCppServer | None = None
         self._delegate: Any = None  # OpenAICompatibleLLM, created after server starts
         self._initialized = False
@@ -337,6 +339,7 @@ async def _ensure_initialized(self) -> None:
             base_url=self._server.base_url,
             model=self.model,
             reasoning_effort=self.reasoning_effort,
+            extra_body=self._extra_body,
         )
 
         self._initialized = True
diff --git a/hindsight-api-slim/hindsight_api/engine/reflect/structured_doc.py b/hindsight-api-slim/hindsight_api/engine/reflect/structured_doc.py
index c667afb10..d62e1d390 100644
--- a/hindsight-api-slim/hindsight_api/engine/reflect/structured_doc.py
+++ b/hindsight-api-slim/hindsight_api/engine/reflect/structured_doc.py
@@ -333,6 +333,17 @@ def _parse_block(chunk: list[str]) -> Block:
         if has_separator:
             return _parse_table_block(chunk)
 
+    # Lenient fallback: GFM allows omitting leading/trailing pipes, so a table
+    # like "Col1 | Col2\\n--- | ---\\nA | B" won't match the strict pipe-delimited
+    # pattern above. Detect these by requiring a separator line AND at least one
+    # ``|`` in every non-separator line (the column delimiter can't be omitted).
+    if len(chunk) >= 2:
+        sep_indices = [i for i, line in enumerate(chunk) if _TABLE_SEPARATOR_RX.match(line)]
+        if sep_indices:
+            non_sep = [line for i, line in enumerate(chunk) if i not in sep_indices]
+            if non_sep and all("|" in line for line in non_sep):
+                return _parse_table_block(chunk)
+
     return ParagraphBlock(text=" ".join(line.strip() for line in chunk).strip())
 
 
diff --git a/skills/hindsight-docs/references/sdks/integrations/agent-plugin.md b/skills/hindsight-docs/references/sdks/integrations/agent-plugin.md
new file mode 100644
index 000000000..87382baf0
--- /dev/null
+++ b/skills/hindsight-docs/references/sdks/integrations/agent-plugin.md
@@ -0,0 +1,14 @@
+
+# Agent Plugins
+
+Portable long-term memory for any [Agent Plugins](https://agent-plugins.org) client, powered by Hindsight.
+
+Agent Plugins is the vendor-neutral open standard for packaging Agent Skills + MCP servers into a single distributable plugin. Hindsight ships one plugin that every compatible client can load.
+
+## Quick Start
+
+1. Get your API key from the Hindsight Cloud dashboard.
+2. Set `HINDSIGHT_API_KEY` and optionally `HINDSIGHT_BANK_ID`.
+3. Install the plugin in your client through its plugin/MCP UI.
+
+Once installed, agents can call `recall`, `retain`, and `reflect` automatically through the bundled skill.