Skip to content

Commit 692d5a6

Browse files
zpzjzjclaude
andauthored
docs(changelog): add 0.1.2 release notes (#38)
Cover the 14 user-visible commits landed since v0.1.1: skill-upper meta-skill, --runtime override, network_policy egress isolation, mocked MCP servers, batch UploadFiles, OpenSandbox SDK v1.0.1, sibling default workspace, context.git.checkout fix, credential default-path fix, and the runtime/agent cleanup-context fixes. Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
1 parent 970f2f7 commit 692d5a6

1 file changed

Lines changed: 53 additions & 4 deletions

File tree

CHANGELOG.md

Lines changed: 53 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -5,7 +5,25 @@ All notable changes to this project will be documented in this file.
55
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
66
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
77

8-
## [0.1.2] - 2026-05-20
8+
## [0.1.2] - 2026-05-21
9+
10+
### Added
11+
- `skill-upper`: a bundled meta-skill under `skills/skill-upper/` that
12+
drives `skill-up` to scaffold and run evals for a target skill end-to-end.
13+
- `--runtime` CLI flag on `skill-up run` overrides `environment.type`
14+
(`none` / `opensandbox`) without editing `eval.yaml`. The override path
15+
also rejects `--runtime none` when `network_policy` is configured, so
16+
network isolation cannot be silently dropped.
17+
- `network_policy` for sandbox egress isolation: `deny_all` blocks all
18+
egress; `allow_declared` plus `environment.allowed_egress` permits only
19+
the listed FQDN / wildcard domains. Validated against the runtime type
20+
(rejected on `environment.type: none`) and against URL-shaped entries.
21+
- Mocked MCP servers (`mode: mocked`): provisioned as generated stdio
22+
servers via the existing agent MCP install path. Supports the built-in
23+
`filesystem` mock and arbitrary `tool_responses` via `config_ref`.
24+
- `userconfig.LoadFile(path)` helper for callers that need to read and
25+
validate a single config file without consulting discovery layers.
26+
- Coverage badge generation in CI and README badge links.
927

1028
### Changed
1129
- `skill-up init`: `--config <path>` now names a **source** file to read
@@ -16,10 +34,41 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
1634
still writes the commented template as before. This aligns `--config`
1735
with the `run`/`validate` discovery chain semantics
1836
(`embed < user < project < --config`).
37+
- OpenSandbox directory uploads now use the SDK's batch `UploadFiles` in
38+
chunks of 128 files (one multipart request per chunk) instead of N
39+
concurrent single-file uploads. `file_transfer_parallelism` now only
40+
governs directory **download**.
41+
- OpenSandbox SDK upgraded to v1.0.1.
42+
- Default eval workspace is now placed **alongside** the skill directory
43+
(`<skill>-workspace/` as a sibling of `<skill>/`) instead of inside it,
44+
matching the Anthropic eval layout convention. Docs and `--output-dir`
45+
default updated accordingly.
1946

20-
### Added
21-
- `userconfig.LoadFile(path)` helper for callers that need to read and
22-
validate a single config file without consulting discovery layers.
47+
### Fixed
48+
- `context.git.checkout` is now actually honored by the fixture loader.
49+
Previously the field parsed but was never read, so cases ran on the
50+
fixture's current branch. The loader now runs `git switch` (or creates
51+
the branch on an unborn HEAD) before inline `context.files` and
52+
`apply_diff`; a missing branch in a committed fixture fails loudly,
53+
and branch names are validated against shell injection.
54+
- `skill-up run` credential resolver now reads the documented
55+
`~/.skill-up/credentials.yaml`. Previously the run path passed an
56+
empty config path, so only `--api-key`, env vars, and a CWD `.env`
57+
were consulted.
58+
- Post-run session lookup, stdout download, and last-message resolution
59+
(claude_code / codex / qodercli) no longer inherit the canceled run
60+
context, so a timed-out eval case no longer produces two bursts of
61+
misleading "command exited with code -1: ..." log lines for the
62+
cleanup steps. A fresh 30s `sessionCleanupTimeout` is used instead.
63+
- `runtime/none`: when a command is killed because its context was
64+
canceled or timed out, the log now reads "command killed by context"
65+
instead of pretending to be a real `-1` exit with the full bootstrap
66+
script attached.
67+
- `runtime/opensandbox`: when the OpenSandbox SDK call itself fails
68+
(network error, nil execution, timeout before a remote exit code is
69+
reported), the log now identifies it as an SDK-level failure and
70+
skips the misleading script dump. Genuine remote non-zero exits keep
71+
the existing log.
2372

2473
## [0.1.1] - 2026-05-15
2574

0 commit comments

Comments
 (0)