dsh-ctx-budget 0.2.0 — DSH plugin bundle

Now installable as a DSH plugin. package.json declares dsh.bundle with cordis.patch.yml, and the package ships the ctxbudget Python code it runs.

dsh plugin add --profile web https://github.com/LUgABeEEE/dsh-ctx-budget/releases/download/v0.2.0/dsh-ctx-budget-0.2.0.tgz

Four read-only tools:

• ctx_budget_report — six-layer static-context snapshot (rules, memory, skills, tool surface, sessions, KB)
• ctx_budget_audit — memory duplicates, stale entries, over-limit files
• ctx_budget_check — verify backticked paths and commands in a markdown capability map
• ctx_budget_doctor — home layout and Python runtime check

memory-archive is deliberately not exposed to the model: moving entries stays an explicit human --apply action.

Also in this release: 14 plugin tests under node --test (including a fake-host registration test), a Node 20/22 CI job, and .npmignore so __pycache__ never ships.

dsh-ctx-budget 0.1.0

First public release.

Extracted from a production agent home and generalised: no hard-coded paths, read-only by default, writes require --apply and are backed up first.

Commands:

• doctor — environment check
• capabilities — verify backticked paths/commands in a markdown capability map
• skills — compress SKILL.md descriptions (stdlib frontmatter parser, .bak backups)
• memory-audit — duplicates, stale entries, over-limit files
• memory-archive — score-based tiering with archive files and pointer entries
• probe — real tool surface and cold-start static context from session logs
• baseline — six-layer snapshot

96 unittest cases; standard library only; Python >= 3.10.

Honest limits: tool descriptions in logs are truncated (numbers are a lower bound), system prompts are not persisted, and static context is only measurable on cold-start sessions — otherwise the value is null.