Qredence Documentation

Model-authored Python calls these tools from the Sandbox. The Sandbox broker forwards each call to Fleet, which runs it under the Session’s authority. See Sandbox execution.

Semantic and recursive calls

Tool	Description
`llm_query`, `llm_query_batched`	Native DSPy semantic calls on the Sub model. Each batched prompt counts toward `rlm.max_llm_calls`.
`rlm_query`	Run one bounded child RLM over staged inputs. Root only.
`rlm_query_batched`	Run several independent child RLMs and return results in input order. Root only.

Session Workspace

Writes take effect immediately and don’t depend on Turn Commit.

Tool	Description
`list_workspace_files`	List immediate entries in the Session Workspace.
`stat_workspace_file`	Read metadata for one path.
`read_workspace_text`	Read one UTF-8 page with `max_chars` from 1 to 10,000. Continue with `next_cursor` until `eof`.
`read_workspace_text_batch`	Read up to 32 pages in one call, with at most 32,000 characters in total.
`write_workspace_text`	Write UTF-8 text.
`append_workspace_text`	Append UTF-8 text.
`edit_workspace_text`	Replace exactly one unique occurrence of `old` with `new`. Fails if `old` is missing or repeated.
`delete_workspace_path`	Delete one file or empty directory.

Pass expected_sha256 to edit_workspace_text or delete_workspace_path to refuse the change if the file has changed.

Projects

Projects hold deliverables that stay browsable across Sessions. Paths take the form <slug>/<path>, and a leading projects/ is accepted as an alias. Writes take effect immediately.

Tool	Description
`list_project_files`	List immediate entries. Use `path='.'` for the projects root.
`stat_project_file`	Read metadata for one path.
`read_project_text`	Read one UTF-8 page with `max_chars` from 1 to 10,000.
`write_project_text`	Write UTF-8 text.
`edit_project_text`	Replace exactly one unique occurrence of `old` with `new`.
`delete_project_path`	Delete one file or empty directory.

Attachments and Artifacts

Tool	Description
`read_attachment`	Read one authorized Attachment by ID.
`create_artifact`	Stage generated text, Markdown, or JSON as an Artifact candidate.
`publish_workspace_artifact`	Stage an existing Session Workspace text file as an Artifact candidate.

Artifact candidates become Artifacts only when the Turn commits.

Memory

Tool	Description
`read_workspace_memory`	Read the latest cross-Session learnings.
`remember`	Record one learning or preference when the user explicitly asks. Returns the entry ID.
`update_workspace_memory`	Legacy alias of `remember`.
`list_memories`	List entries chronologically in pages. Pass `next_cursor` as `after`.
`search_memories`	Search entries with deterministic lexical ranking. Optionally filter by `category`.
`edit_memory`	Replace one entry in place by ID.
`forget`	Remove one entry by ID.
`propose_memory`	Propose a stable, non-secret learning for promotion after Turn Commit. It doesn’t change memory immediately.

Task and history

Tool	Description
`read_active_task`	Read the Session’s current goal and progress.
`update_active_task`	Record the goal, decisions, paths, completed work, or pending work. Pass the `revision` from `read_active_task`.
`read_session_history`	Read a page of older committed messages.

Skills

Tool	Description
`load_skill`	Load an authorized Skill. Returns `skill_markdown`, `allowed_tools`, and `resources`. `allowed_tools` is guidance, not authorization.
`read_skill_resource`	Read one resource from a loaded Skill.

See Use Skills.

Fleet-RLM

Fleet Prime Agent

Fleet Pi

GEPA Omni

Qredence Skills

Fleet Reasoner

Qredence Plugins

Sandbox tools

Semantic and recursive calls

Session Workspace

Projects

Attachments and Artifacts

Memory

Task and history

Skills

​Semantic and recursive calls

​Session Workspace

​Projects

​Attachments and Artifacts

​Memory

​Task and history

​Skills

Semantic and recursive calls

Session Workspace

Projects

Attachments and Artifacts

Memory

Task and history

Skills