Aleph

/aleph - External Memory Workflow

Core rule: keep whole contexts out of the prompt. Return only focused slices or compact derived results.

This plugin bundles the Aleph MCP launcher, /aleph skill, aleph-expert agent, and an install-check hook. It assumes the aleph executable is already installed and available on PATH.

Note: tool names may appear as mcp__aleph__load_workspace_manifest in some clients.

The 5-Phase Loop

1. Load

Pick the correct front door:

• Large repo or codebase: load_workspace_manifest(...)
• Single large file: load_file(...)
• Inline or generated content: load_context(...)

Repo-scale default:

load_workspace_manifest(paths=["src", "tests"], context_id="repo")
rg_search(
  pattern="FastAPI|APIRouter|router\\.",
  paths=["src", "tests"],
  load_context_id="routes"
)
load_file(path="pyproject.toml", context_id="pyproject")

Single-file default:

load_file(path="/absolute/path/to/large.log", context_id="log")
search_context(pattern="ERROR|WARN", context_id="log")
peek_context(context_id="log", start=1, end=60, unit="lines")

Do not start repo work by reading files one by one when load_workspace_manifest(...) is the right first move.

2. Orient

• Use search_context(...) to find relevant regions
• Use peek_context(...) to inspect small ranges
• Use semantic_search(...) for meaning-based lookup
• Use chunk_context(...) when navigability matters
• Use rg_search(...) to sweep repo trees quickly

Search before peeking. Pull only the slices you need.

3. Compute

• Use exec_python(...) for heavier analysis with ctx bound in the sandbox
• Aleph defaults to output_feedback="full"; exec_python(...) is not print-only
• In full mode Aleph can return stdout, stderr, error text, and a rendered return value
• If output volume becomes distracting, optionally use configure(output_feedback="metadata")
• Use built-in helpers such as search, peek, lines, chunk, extract_*, semantic_search, and cite
• Store compact derived values in variables like summary, counts, matches, or result
• Retrieve only those derived values with get_variable(...)
• Treat get_variable("ctx") as blocked for plugin workflows; use bounded slices or compact derived variables instead

Example:

exec_python(code="""
matches = search(r"ERROR|WARN")
counts = {"matches": len(matches)}
summary = f"{counts['matches']} matching lines"
""", context_id="log")
get_variable(name="summary", context_id="log")

4. Recurse

Real recursion helper signatures:

sub_query(prompt, context_slice=None)
sub_query_batch(prompt, context_slices, limit=None)
sub_query_map(prompts, context_slices=None, limit=None, parallel=True)
sub_aleph(query, context=None)

Runtime guidance:

• Use sub_query_batch(...) for one prompt over many slices
• Use sub_query_map(...) for distinct prompts, keeping parallel=True unless you need sequential execution
• Use configure(sub_query_share_session=true) when nested agents need access to parent contexts
• For depth 3+, use configure(sub_query_timeout=300, sandbox_timeout=300)

5. Converge

• Use evaluate_progress(...) when the answer is not yet stable
• Loop back through orient and compute when confidence is low
• Use summarize_so_far(...) if the trajectory is getting long
• Use finalize(answer=..., confidence=..., context_id=...) when done