Context Wiki

A working directory of markdown docs, grouped into named profiles, that you can push into a model's context automatically (proxy injection) or write once into an agent's own config file (export).

What it does

The wiki is a plain folder of .md files (backend/wiki.py, default <ROOT>/wiki, overridable via the wiki_dir config key). Each doc is a single markdown file; names are sanitized (_safe_name()) so a doc can never write or read outside that directory — no path separators, no ./...

Docs are grouped into profiles: a profile is a name plus an ordered list of doc filenames plus a description, stored in the wiki_profiles config key. compose(profile_name) turns a profile into one block of text: it reads each doc in the profile's list order, skips empty ones, and wraps each with a ## <doc-name-without-.md> heading before joining them with blank lines. compose() returns "" for an unknown or empty profile name — there is no error path, only no context.

Each model can have its own active profile (wiki_active, a {model_id: profile_name} map set via set_active()). active_profile(model_id) looks up that model's entry, or returns "" if none is set. This is the piece that makes injection per-model: two models can carry different context, or none at all.

There are two ways the composed text reaches an agent:

Proxy injection. Every request that passes through LlamaForge's Anthropic shim (/v1/messages, streaming and non-streaming) or its OpenAI-compatible proxy (/v1/chat/completions) is composed fresh from that request's model and active profile, then merged into the system prompt: _inject_anthropic_system() prepends the composed text to the Anthropic system field (a string, or a list of blocks, where it becomes a new first text block) or sets it; _inject_openai_system() prepends it to the first system message, or inserts a new one. Both helpers are a no-op and return the request body untouched when compose() returns an empty string — no active profile means no system-prompt change, not an empty block.
Agent-file export. export_agent_file(path, composed) writes the composed text into a target file (e.g. ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.pi/AGENTS.md) inside a marker-delimited region:  ... . If those markers already exist in the file, only the text between them is replaced, so hand-written content above and below survives repeated exports. If they don't exist, the region is appended. Any pre-existing marker text inside a doc itself is defused ( becomes ) so a doc body can't spoof the delimiters. Before any write, the original file is copied once to <path>.llamaforge.bak — the backup is write-once, so re-exporting never overwrites the true original.

How to use it

Open the Context tab.
In the Context docs card, pick a document, or click New to create one, write it in the text box and click Save doc. Delete removes the selected doc.
In the Profiles card, group docs into a profile: give it a Profile name, pick the Documents in profile, and click Save profile. compose() concatenates the docs in list order.
In the Active profile per model card, choose a Model and a Profile and click Set active. Leaving a model on (none) means it gets no injected context.
For agents that call the dashboard's own /v1/messages or /v1/chat/completions endpoints (on panel_port), the active profile's docs are injected automatically on every request — no further action needed. Clients that talk to the llama.cpp router directly bypass LlamaForge, so nothing is injected for them.
For agents that read a static context file instead (Claude Code's CLAUDE.md, Codex's AGENTS.md, pi.dev's AGENTS.md), use the Export to agent file card: choose the Agent and the Profile, then click Export. Leave Project path blank to write the agent's global file (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md or ~/.pi/AGENTS.md), or give it the path of the file to write instead. The composed profile goes into that file's marker region.

Screenshot

Context tab

Reference

Concept	Source	Behavior
Doc storage	`backend/wiki.py` `_dir()` / `list_docs()` / `read_doc()` / `write_doc()` / `delete_doc()`	Flat directory of `.md` files, default `<ROOT>/wiki`, overridable via `wiki_dir` config key.
Name safety	`backend/wiki.py` `_safe_name()`	Rejects names containing `/` or `\`, or equal to `.`/`..`; always resolves to a basename inside the wiki dir.
Profiles	`backend/wiki.py` `get_profiles()` / `save_profile()` / `delete_profile()`	Stored in config key `wiki_profiles`: `{name: {docs: [...], description: ...}}`.
Compose	`backend/wiki.py` `compose(profile_name)`	Joins each listed doc as `## <name>\n\n<text>`, in profile order; skips blank docs; returns `""` for unknown/empty profile.
Active selection	`backend/wiki.py` `active_profile()` / `set_active()`	Config key `wiki_active`, a `{model_id: profile_name}` map.
Anthropic injection	`backend/routes.py` `_inject_anthropic_system()`, used by `_anthropic_messages()` (routes.py) and `_anthropic_stream()` (server.py)	Prepends composed text to the Anthropic `system` field; no-op when composed text is empty.
OpenAI injection	`backend/routes.py` `_inject_openai_system()`, used by the `/v1/chat/completions` handler in `server.py`	Prepends composed text to the first `system` message, or inserts one; no-op when composed text is empty.
Export	`backend/wiki.py` `export_agent_file()`	Marker-delimited region `<!-- llamaforge:start -->` / `<!-- llamaforge:end -->`; replaces the region if present, appends if not; backs up the original once to `<path>.llamaforge.bak`.

Troubleshooting

If a model's requests don't seem to carry your wiki docs, confirm a profile is actually set active for that specific model id — activation is per-model, not global, and an unset model silently gets no injection. If export doesn't seem to change the target file, check that the file wasn't already up to date (re-exporting the same composed text replaces the marker region with identical content) and that <path>.llamaforge.bak was created next to it on first export.

See also Connect an Agent for the endpoints and one-click configs that route agents through the Anthropic shim or the router, and Theming & Accessibility for the dashboard's appearance settings.