dsh-rerun-env-local
v0.1.2
Published
Frozen-environment provider materializing local per-attempt workspaces with a recorded toolchain fingerprint
Maintainers
Readme
dsh-rerun-env-local
Environment provider on the local machine: one private directory per attempt
holding the materialized workspace, a private DSH_HOME, and the session root
the harvest reads from.
The private DSH_HOME is the load-bearing part. Without it an attempt inherits
the operator's profiles, settings, credentials, and hook scripts, and "the
harness under test" silently becomes "the harness under test plus whatever this
laptop has configured" — exactly the confound the framework exists to remove.
- id: rerun-env-local
name: 'dsh-rerun-env-local'
config:
root: .rerun/work
passEnv: [DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL]Model Experience
None. This package never contributes to a model request: it runs in the orchestrator process, and the model calls it cares about happen inside a separate attempt subprocess running a different build of DSH entirely.
KV Cache effect
None; it neither assembles nor sends a provider request.
Known Limitations and Deferred Work
- It cannot declare its
PATH. Attempts inherit the operator's, so the agent can reach any binary on the machine while the fingerprint records only the probed tools. The provider reportspathDeclared: falseand hashes thePATHso differing toolchains are at least detected — an undeclared input cannot be controlled, but it can be kept off the same trend line. This is why it is the development provider anddsh-rerun-env-containeris the one to measure with. - It does not isolate the network. The environment is reproducible and
private, not sandboxed; a task whose command reaches the internet is
measuring the internet. Use
dsh-rerun-env-containerwhen that matters. - Concurrency is not resource-isolated. Eight attempts on one host share a
page cache, a CPU governor, and a
/tmp, so wall-clock figures are only comparable within a machine. - The workspace diff needs
giton the host. Without it the diff is reported as unavailable rather than computed another way.
