dsh-context-pack

In DeepSeek Harness, model-visible context is part of the agent/session lifecycle. The tension is real: you do not want to dump an entire repository into every prompt, yet you also do not want a model-directed scan to silently pick up .env files or SSH private keys. dsh-context-pack answers this with a single explicit /context-pack command that, upon human trigger, performs a read-only scan, builds a bounded and inspectable pack, and queues it through agent.inject() for the next model step. It never wakes an idle agent and never mutates source files.

Determinism is the design anchor: fixed path ordering, known noisy-directory exclusions, fail-closed sensitive-path filtering (.env, common credential and secret names, SSH private-key patterns), a text-extension allowlist, and layered budgets covering per-file bytes, total bytes, file count, and final injection size. Every included file carries a SHA-256 identity; the whole pack has a digest. Files are included in full only, with no silent truncation. The command returns a compact receipt (pack digest, file count, source bytes, exclusion count), while the actual context pack is consumed later via DSH's injection seam.

This fits teams that need controllable repository context inside DSH workflows: the model should see project structure and key source snippets, but the process must be auditable, reproducible, and guaranteed to stay clear of credential-bearing files. It is not a DLP tool, a sandbox, or a semantic relevance ranker. Its scope is precisely one thing: a just-enough, boundary-clean context pack.