Repository navigation
Releases
v0.1.8
- Rewrite every MCP tool definition for agents: each description says what the
tool does, when to use it and which sibling to use instead, its side effects
and what it returns. Every parameter now has a description, and defaults are
declared in the schemas. Tool behavior, names and accepted inputs are unchanged. - Declare MCP tool annotations (
readOnlyHint,destructiveHint,
idempotentHint,openWorldHint); a test checks them against each tool's real
effect on the workspace. The MCP server also publishes short instructions. - Descriptions no longer start with a
[group]tag;evokit toolsstill
reports the group. - Releases are automated: a
v*tag publishes to PyPI, to npm through Trusted
Publishing, to a GitHub Release with the changelog section, and to the official
MCP Registry, then points the DSH plugin index on GitHub Pages at the new npm
version after validating it with DSH's format-1 validator. Addserver.json
for the registry, kept at the package version.
evokit v0.1.7
- Accept every DSH 0.2.x from 0.2.2-rc.1: the bundle's
engines.dshrange is
now>=0.2.2-rc.1 <0.3.0-0instead of the exact0.2.2-rc.1. DSH matches
ranges with prereleases included, so<0.3.0-0keeps 0.3.0 prereleases out.
Install, listing, configuration dump and MCP tool and skill registration were checked on DSH
0.2.2-rc.1 and 0.2.2-rc.4; the 0.1.6 bundle is refused on 0.2.2-rc.4. EVOKIT_MCP_COMMAND,EVOKIT_WORKSPACEandEVOKIT_SKILL_ROOTare optional and default
toevokiton PATH, the directory where DSH starts and the bundle's ownskillsfolder.- The tool inspector overlay also checks that the
onboard,evolveand
metaskills are registered, and falls back toevokiton PATH. - Lead both READMEs with what evokit does, how adoption is gated, the
documented BFCL pilot result and the shortest install path. Add the MCP
Registry ownership marker to the PyPI description. - Add a
dsh-evokitconsole command, identical toevokit, souvx dsh-evokit mcp
starts the MCP server. - Add a root
package.jsondeclaring the same DSH bundle, so plugin catalogs and
Git installs detect it; npm packages are still published fromintegrations/dsh.
Install:
uv tool install dsh-evokit
dsh plugin --profile evokit add dsh-evokit@0.1.7
Pinning @0.1.7 matters only during the first day: DSH's pnpm skips npm versions younger than 24 hours unless they are named exactly. The release files below are the verified alternative; compare them against SHA256SUMS first.
evokit v0.1.6
-
List the author as quei. Documentation and package metadata are English only.
-
Ship the optional accounting relay in the Python wheel. Separate JSON
byte admission from token context and conservatively reserve against both. -
Configure native DSH improver context at 524288 tokens with bounded 32768-token
output; expose validated client-limit construction for other host profiles. -
Add periodic publication and bounded final publication retry, including real
Windows deny-delete recovery without re-executing evaluations. -
Add operator-only circuit reset prerequisites and archived reset evidence;
preserve reservations, settled cost and circuit thresholds. -
Add exactly-once prior-attempt cost/model/wall-time restoration. Missing wall
time is explicitly a lower bound, never presented as complete latency. -
Add domain-validated completed-response caching with explicit stop metadata,
revalidation of legacy records, and quarantined rejected/unvalidated replies.
Transport remains responsible for settlement before content validation.
Install: uv tool install ./dsh_evokit-0.1.6-py3-none-any.whl and dsh plugin --profile evokit add ./dsh-evokit-0.1.6.tgz. Compare files against SHA256SUMS first.
evokit v0.1.5
Changelog
v0.1.5
- Add generic paired-defined secondary comparisons, explicit both-undefined
applicability, count guards, and complete-repeat quality noise estimates. - Anchor configurable primary noise multipliers and optional observed-net floors;
scale the reference for pooled repeat means. Gates over the mean of R>1 repeats
now use threshold/sqrt(R), which is more permissive than v0.1.4. Single-repeat
gates are unchanged; majority aggregation keeps the single-repeat threshold. - Preserve reviewed v0.1.2/v0.1.3 default-noise epochs without rewriting their
stored anchors. Repeat quality guards aggregate task means in task units.
v0.1.4
- Select
deepseek-v4.1-flashfor proposals through a dedicated improver route,
require served builddeepseek-v4-1-flash-261017, and enable high thinking with
a 32,768-token output budget. Preserve target and generator route settings.
Account reasoning as a subset of output and remove DeepSeek cache controls. - Strengthen the evolve playbook with measured detectability, full-diff literal review,
record-cited history, predicted task flips and verified artifact load scope.