session_bootstrapper¶
mode: subagent — always the first execution step, before @literature_scout.
What it does¶
Turns a data_dir plus the user's task into a ready session_dir containing
a validated session_context.json, by calling
bootstrap_session — one tool call does
all the deterministic work. It writes no Python and reads no binary files.
Not this agent's job: bootstrap is registry-agnostic. Any match against
a registered dataset (reuse) or first-time registration is opened as a
file-backed human gate by ingest_to_session
and handled by @data_ingestor, never here. No
registry state is inherited across independent user tasks because none is
written at bootstrap time.
Inputs¶
data_dir— the directory holding the raw catalogue and any ancillary files (masks, n(z), systematics maps,DATA_DESCRIPTION.md).user_task— the user's task string verbatim, including every structured section (## probes:,## inference priors:,## inference sampler:,## inference scale cuts:,## inference gate:,## plots:, etc. — see Writing TASK.md). This must not be summarised or paraphrased: the inference tools read these sections back out ofsession_context.json["task"]at a much later pipeline step, and a dropped section there silently produces a wrong auto-generated config instead of failing loudly.survey=— passed through when the task orDATA_DESCRIPTION.mdnames a known family (see Catalogue-type profiles).
A single-purpose worker¶
Like every subagent, it does exactly one job and returns its result as text —
it never orchestrates, dispatches, or @mentions another agent. If it's
blocked, it emits one NEEDS_INPUT:/ERROR: line (see
Platform conventions),
never a free-form question.