werkmeister-core 0.0.0-stage → 0.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +1043 -0
- package/LICENSE +202 -0
- package/NOTICE +6 -0
- package/README.md +39 -2
- package/dist/agent_gen.d.ts +52 -0
- package/dist/agent_gen.js +228 -0
- package/dist/agent_state.d.ts +172 -0
- package/dist/agent_state.js +728 -0
- package/dist/archetypes.d.ts +110 -0
- package/dist/archetypes.js +227 -0
- package/dist/baseline.d.ts +29 -0
- package/dist/baseline.js +92 -0
- package/dist/browser.d.ts +135 -0
- package/dist/browser.js +702 -0
- package/dist/bug_report.d.ts +12 -0
- package/dist/bug_report.js +95 -0
- package/dist/build_caches.d.ts +70 -0
- package/dist/build_caches.js +250 -0
- package/dist/cache.d.ts +23 -0
- package/dist/cache.js +124 -0
- package/dist/caddy.d.ts +71 -0
- package/dist/caddy.js +428 -0
- package/dist/claude_cli.d.ts +54 -0
- package/dist/claude_cli.js +173 -0
- package/dist/claude_md.d.ts +58 -0
- package/dist/claude_md.js +583 -0
- package/dist/colors.d.ts +27 -0
- package/dist/colors.js +17 -0
- package/dist/compose.d.ts +177 -0
- package/dist/compose.js +646 -0
- package/dist/compose_file.d.ts +14 -0
- package/dist/compose_file.js +37 -0
- package/dist/compose_infra.d.ts +33 -0
- package/dist/compose_infra.js +75 -0
- package/dist/config.d.ts +421 -0
- package/dist/config.js +682 -0
- package/dist/daemon_client.d.ts +81 -0
- package/dist/daemon_client.js +239 -0
- package/dist/detect.d.ts +49 -0
- package/dist/detect.js +198 -0
- package/dist/devcontainer_gen.d.ts +22 -0
- package/dist/devcontainer_gen.js +96 -0
- package/dist/docker.d.ts +176 -0
- package/dist/docker.js +597 -0
- package/dist/dot_dir.d.ts +7 -0
- package/dist/dot_dir.js +10 -0
- package/dist/errors.d.ts +35 -0
- package/dist/errors.js +41 -0
- package/dist/feature_routing.d.ts +87 -0
- package/dist/feature_routing.js +133 -0
- package/dist/fixture_fakes.d.ts +53 -0
- package/dist/fixture_fakes.js +127 -0
- package/dist/fixture_workspace.d.ts +82 -0
- package/dist/fixture_workspace.js +289 -0
- package/dist/folders.d.ts +39 -0
- package/dist/folders.js +63 -0
- package/dist/form_schema.d.ts +112 -0
- package/dist/form_schema.js +225 -0
- package/dist/fragment_detect/allocate.d.ts +31 -0
- package/dist/fragment_detect/allocate.js +102 -0
- package/dist/fragment_detect/assist.d.ts +58 -0
- package/dist/fragment_detect/assist.js +243 -0
- package/dist/fragment_detect/compose.d.ts +5 -0
- package/dist/fragment_detect/compose.js +97 -0
- package/dist/fragment_detect/correlate.d.ts +26 -0
- package/dist/fragment_detect/correlate.js +99 -0
- package/dist/fragment_detect/dockerfile.d.ts +4 -0
- package/dist/fragment_detect/dockerfile.js +66 -0
- package/dist/fragment_detect/files.d.ts +15 -0
- package/dist/fragment_detect/files.js +84 -0
- package/dist/fragment_detect/helm.d.ts +7 -0
- package/dist/fragment_detect/helm.js +72 -0
- package/dist/fragment_detect/index.d.ts +20 -0
- package/dist/fragment_detect/index.js +20 -0
- package/dist/fragment_detect/k8s.d.ts +14 -0
- package/dist/fragment_detect/k8s.js +138 -0
- package/dist/fragment_detect/kustomize.d.ts +2 -0
- package/dist/fragment_detect/kustomize.js +58 -0
- package/dist/fragment_detect/package.d.ts +2 -0
- package/dist/fragment_detect/package.js +100 -0
- package/dist/fragment_detect/plan.d.ts +72 -0
- package/dist/fragment_detect/plan.js +206 -0
- package/dist/fragment_detect/precedence.d.ts +25 -0
- package/dist/fragment_detect/precedence.js +113 -0
- package/dist/fragment_detect/render.d.ts +12 -0
- package/dist/fragment_detect/render.js +196 -0
- package/dist/fragment_detect/types.d.ts +53 -0
- package/dist/fragment_detect/types.js +6 -0
- package/dist/fragment_detect/validate.d.ts +5 -0
- package/dist/fragment_detect/validate.js +52 -0
- package/dist/fragment_detect/workspace.d.ts +14 -0
- package/dist/fragment_detect/workspace.js +94 -0
- package/dist/gate.d.ts +59 -0
- package/dist/gate.js +427 -0
- package/dist/generate.d.ts +130 -0
- package/dist/generate.js +1399 -0
- package/dist/git_hooks.d.ts +22 -0
- package/dist/git_hooks.js +81 -0
- package/dist/github_ingest.d.ts +20 -0
- package/dist/github_ingest.js +42 -0
- package/dist/global_config.d.ts +238 -0
- package/dist/global_config.js +700 -0
- package/dist/graduate.d.ts +71 -0
- package/dist/graduate.js +285 -0
- package/dist/harness/autonomous_deny.d.ts +10 -0
- package/dist/harness/autonomous_deny.js +98 -0
- package/dist/harness/claude/claude_md.d.ts +58 -0
- package/dist/harness/claude/claude_md.js +619 -0
- package/dist/harness/claude/cli.d.ts +38 -0
- package/dist/harness/claude/cli.js +173 -0
- package/dist/harness/claude/desktop.d.ts +8 -0
- package/dist/harness/claude/desktop.js +8 -0
- package/dist/harness/claude/hooks_gen.d.ts +141 -0
- package/dist/harness/claude/hooks_gen.js +684 -0
- package/dist/harness/claude/index.d.ts +2 -0
- package/dist/harness/claude/index.js +24 -0
- package/dist/harness/claude/session.d.ts +57 -0
- package/dist/harness/claude/session.js +130 -0
- package/dist/harness/claude/transcript.d.ts +55 -0
- package/dist/harness/claude/transcript.js +375 -0
- package/dist/harness/claude/usage.d.ts +49 -0
- package/dist/harness/claude/usage.js +191 -0
- package/dist/harness/codex/adapter.d.ts +2 -0
- package/dist/harness/codex/adapter.js +23 -0
- package/dist/harness/codex/cli.d.ts +13 -0
- package/dist/harness/codex/cli.js +128 -0
- package/dist/harness/codex/index.d.ts +9 -0
- package/dist/harness/codex/index.js +54 -0
- package/dist/harness/codex/session.d.ts +22 -0
- package/dist/harness/codex/session.js +54 -0
- package/dist/harness/codex/transcript.d.ts +30 -0
- package/dist/harness/codex/transcript.js +184 -0
- package/dist/harness/codex/usage.d.ts +8 -0
- package/dist/harness/codex/usage.js +68 -0
- package/dist/harness/index.d.ts +11 -0
- package/dist/harness/index.js +13 -0
- package/dist/harness/layout.d.ts +17 -0
- package/dist/harness/layout.js +86 -0
- package/dist/harness/mcp/index.d.ts +2 -0
- package/dist/harness/mcp/index.js +24 -0
- package/dist/harness/opencode/hooks.d.ts +11 -0
- package/dist/harness/opencode/hooks.js +183 -0
- package/dist/harness/opencode/index.d.ts +2 -0
- package/dist/harness/opencode/index.js +28 -0
- package/dist/harness/opencode/layout.d.ts +8 -0
- package/dist/harness/opencode/layout.js +17 -0
- package/dist/harness/opencode/session.d.ts +16 -0
- package/dist/harness/opencode/session.js +69 -0
- package/dist/harness/opencode/transcript.d.ts +15 -0
- package/dist/harness/opencode/transcript.js +117 -0
- package/dist/harness/sessions.d.ts +21 -0
- package/dist/harness/sessions.js +81 -0
- package/dist/harness/stub/index.d.ts +2 -0
- package/dist/harness/stub/index.js +26 -0
- package/dist/harness/turn.d.ts +23 -0
- package/dist/harness/turn.js +114 -0
- package/dist/harness/types.d.ts +178 -0
- package/dist/harness/types.js +2 -0
- package/dist/hooks_gen.d.ts +127 -0
- package/dist/hooks_gen.js +624 -0
- package/dist/infra.d.ts +60 -0
- package/dist/infra.js +336 -0
- package/dist/issue_snapshot.d.ts +22 -0
- package/dist/issue_snapshot.js +101 -0
- package/dist/issue_status.d.ts +32 -0
- package/dist/issue_status.js +159 -0
- package/dist/issue_tracker.d.ts +74 -0
- package/dist/issue_tracker.js +177 -0
- package/dist/issue_trackers/github.d.ts +19 -0
- package/dist/issue_trackers/github.js +251 -0
- package/dist/issue_trackers/jira.d.ts +25 -0
- package/dist/issue_trackers/jira.js +290 -0
- package/dist/issue_trackers/types.d.ts +124 -0
- package/dist/issue_trackers/types.js +11 -0
- package/dist/issues_report.d.ts +13 -0
- package/dist/issues_report.js +121 -0
- package/dist/jira-rest.d.ts +95 -0
- package/dist/jira-rest.js +247 -0
- package/dist/k8s.d.ts +111 -0
- package/dist/k8s.js +382 -0
- package/dist/k8s_forwards.d.ts +44 -0
- package/dist/k8s_forwards.js +95 -0
- package/dist/launcher.d.ts +17 -0
- package/dist/launcher.js +29 -0
- package/dist/lifecycle.d.ts +491 -0
- package/dist/lifecycle.js +1369 -0
- package/dist/locks.d.ts +50 -0
- package/dist/locks.js +128 -0
- package/dist/machine_setup.d.ts +516 -0
- package/dist/machine_setup.js +1849 -0
- package/dist/mcp_gen.d.ts +37 -0
- package/dist/mcp_gen.js +127 -0
- package/dist/md_form.d.ts +67 -0
- package/dist/md_form.js +248 -0
- package/dist/migrate.d.ts +32 -0
- package/dist/migrate.js +427 -0
- package/dist/model_tiers.d.ts +29 -0
- package/dist/model_tiers.js +74 -0
- package/dist/notes.d.ts +77 -0
- package/dist/notes.js +278 -0
- package/dist/plugin_check.d.ts +44 -0
- package/dist/plugin_check.js +211 -0
- package/dist/pr_guard.d.ts +30 -0
- package/dist/pr_guard.js +103 -0
- package/dist/pr_lifecycle.d.ts +387 -0
- package/dist/pr_lifecycle.js +2561 -0
- package/dist/pr_rules.d.ts +8 -0
- package/dist/pr_rules.js +99 -0
- package/dist/proc.d.ts +67 -0
- package/dist/proc.js +153 -0
- package/dist/projects.d.ts +32 -0
- package/dist/projects.js +110 -0
- package/dist/registry_sync.d.ts +143 -0
- package/dist/registry_sync.js +523 -0
- package/dist/reporter.d.ts +71 -0
- package/dist/reporter.js +117 -0
- package/dist/resources.d.ts +82 -0
- package/dist/resources.js +179 -0
- package/dist/run_start.d.ts +25 -0
- package/dist/run_start.js +86 -0
- package/dist/safari.d.ts +23 -0
- package/dist/safari.js +217 -0
- package/dist/sandbox.d.ts +873 -0
- package/dist/sandbox.js +2609 -0
- package/dist/scaffold.d.ts +56 -0
- package/dist/scaffold.js +192 -0
- package/dist/scheduler.d.ts +71 -0
- package/dist/scheduler.js +92 -0
- package/dist/scheduler_client.d.ts +18 -0
- package/dist/scheduler_client.js +44 -0
- package/dist/scheduler_state.d.ts +105 -0
- package/dist/scheduler_state.js +385 -0
- package/dist/sdk.d.ts +21 -0
- package/dist/sdk.js +22 -0
- package/dist/secrets.d.ts +42 -0
- package/dist/secrets.js +81 -0
- package/dist/skill_gen.d.ts +65 -0
- package/dist/skill_gen.js +322 -0
- package/dist/slots.d.ts +136 -0
- package/dist/slots.js +407 -0
- package/dist/standards.d.ts +217 -0
- package/dist/standards.js +644 -0
- package/dist/stats.d.ts +138 -0
- package/dist/stats.js +407 -0
- package/dist/status.d.ts +131 -0
- package/dist/status.js +194 -0
- package/dist/strategies/icm.d.ts +13 -0
- package/dist/strategies/icm.js +457 -0
- package/dist/strategies/icm_decisions.d.ts +11 -0
- package/dist/strategies/icm_decisions.js +146 -0
- package/dist/strategies/icm_templates.d.ts +60 -0
- package/dist/strategies/icm_templates.js +512 -0
- package/dist/strategies/index.d.ts +18 -0
- package/dist/strategies/index.js +96 -0
- package/dist/strategies/issue_section.d.ts +3 -0
- package/dist/strategies/issue_section.js +15 -0
- package/dist/strategies/jit.d.ts +19 -0
- package/dist/strategies/jit.js +266 -0
- package/dist/strategies/types.d.ts +70 -0
- package/dist/strategies/types.js +13 -0
- package/dist/telepresence.d.ts +225 -0
- package/dist/telepresence.js +665 -0
- package/dist/temp_home.d.ts +10 -0
- package/dist/temp_home.js +58 -0
- package/dist/test_run.d.ts +74 -0
- package/dist/test_run.js +311 -0
- package/dist/tool_bins.d.ts +31 -0
- package/dist/tool_bins.js +132 -0
- package/dist/usage_snapshot.d.ts +40 -0
- package/dist/usage_snapshot.js +200 -0
- package/dist/web_servers.d.ts +113 -0
- package/dist/web_servers.js +542 -0
- package/dist/workspace_init.d.ts +112 -0
- package/dist/workspace_init.js +836 -0
- package/dist/workspace_prompts.d.ts +120 -0
- package/dist/workspace_prompts.js +719 -0
- package/dist/workspace_skills.d.ts +100 -0
- package/dist/workspace_skills.js +292 -0
- package/dist/workspace_state.d.ts +111 -0
- package/dist/workspace_state.js +278 -0
- package/dist/workspace_token.d.ts +24 -0
- package/dist/workspace_token.js +76 -0
- package/dist/worktrees.d.ts +109 -0
- package/dist/worktrees.js +369 -0
- package/package.json +128 -4
package/CHANGELOG.md
ADDED
|
@@ -0,0 +1,1043 @@
|
|
|
1
|
+
# sandbox-core
|
|
2
|
+
|
|
3
|
+
## 0.89.0
|
|
4
|
+
|
|
5
|
+
### Minor Changes
|
|
6
|
+
|
|
7
|
+
- 93442fd: Publish `werkmeister-core` to npm. Its `werkmeister-core/sdk` entry point is the public SDK and follows semver; every other entry point is internal. Versions stay in lockstep with the CLI.
|
|
8
|
+
|
|
9
|
+
## 0.88.2
|
|
10
|
+
|
|
11
|
+
## 0.88.1
|
|
12
|
+
|
|
13
|
+
## 0.88.0
|
|
14
|
+
|
|
15
|
+
### Minor Changes
|
|
16
|
+
|
|
17
|
+
- 77b4b48: SW-675, SW-668: a wave of workspaces on a small budget no longer deadlocks on the capacity queue, and capacity waits no longer stop for a person.
|
|
18
|
+
|
|
19
|
+
- Backfill: a start that fits goes ahead of a queue head that doesn't. It stops passing once the head is a gate's start or has waited 5 minutes. Gate starts queue ahead of plain `feature up`s.
|
|
20
|
+
- In an autonomous workspace, a `--service` or gate start that has to wait while its feature already runs containers stops them first. Nothing is held while it waits, and it starts them again with the new service. Two gates each holding werkmeister-core no longer wait on each other.
|
|
21
|
+
- A plain `feature up` at the head of the queue may stop idle or waiting-on-a-person features after 5 minutes (`budget.preempt: false` still turns all preemption off).
|
|
22
|
+
- The idle reaper stops a feature whose session has waited on a person's answer for `overrides.idle_stop_minutes`, but only while another feature waits for capacity.
|
|
23
|
+
- Exit 75 means "no capacity yet, rerun": a queued `feature up` (it was 0), a `--service` start that waited 10 minutes (it was 1), and `pr gate` / `pr open` when every failure was a capacity wait (gate reads `retry later (no capacity)`). Scripts that treat any non-zero exit as failure now see a queued `feature up` as one. Each message lists who holds the budget, the head it waits behind, and whether a person is needed.
|
|
24
|
+
- `budget status` lists each running feature's CPUs, memory, session state and running gate, and each waiter's need and wait, in admission order. `--json` adds `running` and `waiters`.
|
|
25
|
+
|
|
26
|
+
## 0.87.3
|
|
27
|
+
|
|
28
|
+
## 0.87.2
|
|
29
|
+
|
|
30
|
+
### Patch Changes
|
|
31
|
+
|
|
32
|
+
- 8a90281: SW-629: a start whose CPUs or memory alone are more than the whole scheduler budget fails at once with `NEVER_FITS`, naming the cost and the budget of the first resource over (`cost 5.0 CPUs > budget 4`) and the fix (raise the engine's CPUs or memory, or `budget.cpus` / `budget.memory` in `~/.werkmeister/config.yaml`, or start fewer services). Before, `feature up` queued it for good and the gate's service starts waited 10 minutes and failed `NO_CAPACITY`, though no preemption could ever make room.
|
|
33
|
+
- aa2e36e: SW-649: an autonomous workspace pre-approves the exact frozen install each of its repos' root lockfiles pins (pnpm, npm, yarn v1, yarn berry with `.yarnrc.yml`, uv) instead of always the fixed npm and pnpm pair. The lockfiles are read from the repo's checkout under `repos_root`, not the workspace's worktree, so a lockfile an agent writes grants nothing. A repo with none of these lockfiles gets no install rule.
|
|
34
|
+
- 238b387: SW-660: a workspace's allow rules now cover the test command exactly as werkmeister-context.md prints it, the repo's appended lint included. A test command of more than one command (an `&&` chain, a `{ …; }` group, `||`, `;`, a pipe) is printed and allowed as one exact `with-temp-home sh -c '…'` line, so every part runs under the temp HOME and the `{ …; }` rows are no longer a shell syntax error; the bare per-segment rules for the wrapped form are gone. One simple command keeps its `with-temp-home <cmd>:*` rule. `validate:` text still gets no rule. Existing workspaces pick the rules up on refresh or `werkmeister machine migrate`, and the headless wake passes the same list.
|
|
35
|
+
|
|
36
|
+
## 0.87.1
|
|
37
|
+
|
|
38
|
+
## 0.87.0
|
|
39
|
+
|
|
40
|
+
## 0.86.4
|
|
41
|
+
|
|
42
|
+
## 0.86.3
|
|
43
|
+
|
|
44
|
+
### Patch Changes
|
|
45
|
+
|
|
46
|
+
- dd1c412: The ICM bug-fix exercise stage collects each manual row's result before the review step, so the review's rerun of `exercise.md` cannot drop a recorded `MANUAL: not run` line.
|
|
47
|
+
- 7162800: The ICM research stage checks a recommended answer that rests on an external API, permission or platform behaviour live and read-only before it goes into `decisions.md`; when it cannot, the option text itself says `unverified`, so the person is asked with the word in front of them.
|
|
48
|
+
- 5d768ab: The ICM exercise step stops before any browser check when `werkmeister feature up` warns that a routed hostname does not resolve, and asks the person to run `werkmeister machine setup`. It never bypasses a certificate warning; a published-port fallback is recorded in `exercise.md` as a deviation only when the person says to skip setup.
|
|
49
|
+
- c06f99d: The ICM exercise stages require each manual-row question to state the URL to open, how to reach this workspace's copy (the `X-Feature` cookie or selector for a routed hostname, or the container's published port), and the visible before → after that means PASS, so the person is not left asking what to look for or checking another workspace's build.
|
|
50
|
+
- 44922a4: The ICM build stages (feature implement, bug-fix fix, refactor implement) start by fetching the base and listing the commits on it that touch the files `spec.md`, `repro.md` or `scope.md` changes, plus the status of any ticket a decided question names. A new commit or ticket change that bears on a decision stops the stage for the person before code is written.
|
|
51
|
+
- 5030975: The ICM plan stages (feature spec, bug-fix reproduce, refactor scope) say code changes stay uncommitted until the propose stage, so a spec plans the work and its checks, not commits. A grouped PR's contracts say the implement stage makes the per-ticket commits.
|
|
52
|
+
|
|
53
|
+
## 0.86.2
|
|
54
|
+
|
|
55
|
+
## 0.86.1
|
|
56
|
+
|
|
57
|
+
### Patch Changes
|
|
58
|
+
|
|
59
|
+
- e8f00dd: `werkmeister pr gate` and `pr open` no longer test every service when the diff touches the repo's own `.werkmeister/services.yaml`. The gate compares the branch's rendered fragment with the one on its base (both through the same registry merge) and touches only the services whose entries differ, so a changed `test:` runs that one service. The `tests` row says so (`fragment: .werkmeister/services.yaml changes <svc>`). A `repo:` key change, or a change to one of the repo's own `setup_services`, still tests every service, and so does a `path_mappings` / `no_service_paths` change that maps a file of the diff differently.
|
|
60
|
+
- ae543b5: `werkmeister pr gate` and `pr open` run the release check (`release_check`, e.g. `changeset status`) before lint and the test commands. A missing changeset fails the release row in seconds and the tests are not run. A running container gets its setup one-shot re-run first. With nothing running, the check uses the first container the tests would start, so a passing gate starts no extra container, unless that service starts from the branch's own settings (`compose.gate.yml`).
|
|
61
|
+
|
|
62
|
+
## 0.86.0
|
|
63
|
+
|
|
64
|
+
### Minor Changes
|
|
65
|
+
|
|
66
|
+
- c8c721b: A repo can declare `format:` in its `.werkmeister/services.yaml` `repo:` block, e.g. `pnpm exec biome check --write {files}`, `uv run ruff format {files}` or `./gradlew spotlessApply`. `werkmeister-context.md` lists it under "Repo commands", with `release_check` for a `ship: tag` repo. The ICM contracts and the generated `w9r-pr` skill no longer name biome, pnpm or changesets as a managed repo's commands. A repo with no `format:` gets no format step; the gate's `lint` still checks formatting.
|
|
67
|
+
|
|
68
|
+
## 0.85.7
|
|
69
|
+
|
|
70
|
+
## 0.85.6
|
|
71
|
+
|
|
72
|
+
### Patch Changes
|
|
73
|
+
|
|
74
|
+
- ee7e229: `.werkmeister/bin/with-fixture-workspace` now works inside a werkmeister container (`docker exec -w /app <container> <ws>/.werkmeister/bin/with-fixture-workspace …`): it runs the CLI at /app, where the container's node_modules is, instead of the host-path copy that failed `ERR_MODULE_NOT_FOUND: dockerode`. From a host checkout with no node_modules it exits 2 and names that form. Rerun `werkmeister workspace init` to regenerate the script.
|
|
75
|
+
- 1f39deb: `with-fixture-workspace --gate --pr-field status=<value>` sets the seeded PR file's `status:` line verbatim (one line, a `# note` included), so a fixture run can check how `pr gate` reads a status form, e.g. `"status=approved # approved via check-in 2026-11-06"`.
|
|
76
|
+
|
|
77
|
+
## 0.85.5
|
|
78
|
+
|
|
79
|
+
### Patch Changes
|
|
80
|
+
|
|
81
|
+
- bb9398d: `werkmeister workspace abandon --help` has its own text, separate from `workspace close`. Help and error messages that pointed at Users guide sections now name the new Concepts, CLI reference, Configuration and Registry fragments pages.
|
|
82
|
+
|
|
83
|
+
## 0.85.4
|
|
84
|
+
|
|
85
|
+
## 0.85.3
|
|
86
|
+
|
|
87
|
+
### Patch Changes
|
|
88
|
+
|
|
89
|
+
- 531e140: SW-551: when `workspace init` (or `workspace new`) can't get a feature slot, it no longer reports plain success. It warns that the container step did not run and gives the next step: free a slot with `werkmeister feature down <feature>`, then rerun `werkmeister workspace init`. It also writes that notice to `.werkmeister/werkmeister-context.md`, so the agent sees it there too, and the next successful init replaces it with the container recipe. The MCP `init-workspace` tool result now includes `containers`.
|
|
90
|
+
|
|
91
|
+
## 0.85.2
|
|
92
|
+
|
|
93
|
+
## 0.85.1
|
|
94
|
+
|
|
95
|
+
## 0.85.0
|
|
96
|
+
|
|
97
|
+
## 0.84.0
|
|
98
|
+
|
|
99
|
+
### Minor Changes
|
|
100
|
+
|
|
101
|
+
- e84d3bb: SW-577: every workspace now records an explicit model, picked by tier. `--model` takes a tier (`small`, `mid`, `frontier`, `advanced_frontier`) or an explicit model; blank means `models.default_tier` (frontier). A tier resolves at create/init through `models:` in `~/.werkmeister/config.yaml`, merged over the built-ins (claude: haiku / sonnet / opus / claude-fable-5-1; opencode: the same as `anthropic/…`), and the resolved model goes to workspace.yml and `.claude/settings.json`, so changing the config affects new workspaces only. Codex has no built-ins: set `models.codex.<tier>`, or `workspace new` on a Codex workspace, and `workspace init` in one that records no model yet, fails naming the key. Run `werkmeister machine migrate` once to write the default model into existing workspaces that record none (only `model`: their rules and hooks are left as they are). `werkmeister machine config` prints the tiers; `workspace list` and the workspace card show each workspace's model.
|
|
102
|
+
|
|
103
|
+
## 0.83.5
|
|
104
|
+
|
|
105
|
+
## 0.83.4
|
|
106
|
+
|
|
107
|
+
## 0.83.3
|
|
108
|
+
|
|
109
|
+
### Patch Changes
|
|
110
|
+
|
|
111
|
+
- 7f0c186: SW-554: a compose regeneration (`feature up`, the pr gate's `startService`, `workspace init`, the gate's compose.gate.yml) run inside a container now refuses with `INSIDE_CONTAINER` instead of writing the container's `/app` paths into the host-read file, which made the install run an empty script and exit 0. Docker's `/.dockerenv` is the test, so unsetting `WERKMEISTER_FEATURE` no longer gets past it; a relocated `WERKMEISTER_CONFIG_FILE` (every test and fixture) still writes.
|
|
112
|
+
|
|
113
|
+
## 0.83.2
|
|
114
|
+
|
|
115
|
+
## 0.83.1
|
|
116
|
+
|
|
117
|
+
## 0.83.0
|
|
118
|
+
|
|
119
|
+
### Minor Changes
|
|
120
|
+
|
|
121
|
+
- 6255c49: SW-487: one workspace state model in werkmeister-core. `WorkspaceStatus.workspaceState` gives each workspace a phase, an activity (thinking, running a tool or tests, a background agent, queued for CPU, paused, containers starting, idle, session gone) and a needs-you list. Each item has what, since, action and source: questions, permissions, plans, stage gates, a dead session, run the postmortem, close it, crash loops and system notices, plus CI failed, conflict, changes requested and merge it from `getWorkspacePRs()`. Hooks now keep one record per session (`.werkmeister/sessions/<id>.json`), so sessions and subagents can't overwrite each other's ask, and a crashed session reads as gone after 30 minutes. The /workspaces pane lists every item with its time and action, and check-in reads the same model. Run `werkmeister machine migrate` to add the new Claude Code hook events to existing workspaces.
|
|
122
|
+
|
|
123
|
+
## 0.82.3
|
|
124
|
+
|
|
125
|
+
### Patch Changes
|
|
126
|
+
|
|
127
|
+
- 1b060f6: SW-421: when a workspace closes out (from the CLI, MCP or the pane), the VS Code extension closes the terminals it opened there within about 3 s of close-out finishing, and frees their pool slots. That covers the `claude <ws>` sessions, including the one that ran the close-out, and the pooled `werkmeister <ws>` ones. It shows a message naming the sessions it closed. A terminal you opened by hand is never touched, and a session whose folder was only moved stays open. Close-out's `/exit it` lines no longer list the sessions the extension closes, and still list every other terminal session in the folder.
|
|
128
|
+
|
|
129
|
+
## 0.82.2
|
|
130
|
+
|
|
131
|
+
## 0.82.1
|
|
132
|
+
|
|
133
|
+
### Patch Changes
|
|
134
|
+
|
|
135
|
+
- 272cd4f: Create prints an info line naming the services it took from the issue text, and a `services` value passed blank (`""` or `[]`) is no longer overridden by that match (SW-413).
|
|
136
|
+
|
|
137
|
+
## 0.82.0
|
|
138
|
+
|
|
139
|
+
### Minor Changes
|
|
140
|
+
|
|
141
|
+
- 4a6a8b3: SW-478: an approval given at the root `/check-in` now reaches a waiting workspace as a file it acts on, with no second prompt. The ICM spec, implement and review stages (feature `spec.md` and `review.md`, bugfix `fix.md` and `review.md`, refactor `review.md`) now start their output with `status: proposed`. Their human check flips it to `decided`, and the router enters the next stage when it reads `decided`, the way 01_research already did. The `check-in` tool's `gates` also lists a `pr-*.md` at `status: draft`. The `check-in` skill ships in the werkmeister plugin. Its gate-file row now covers PR approvals and stage "proceed"s, which before went back to the person to type again, and the reverted SW-309 pane-form row is gone. After upgrading, remove any hand-copied `check-in` skill from your workspaces root's `.claude/skills/`.
|
|
142
|
+
- 28828ce: SW-517: every ICM stage's human check now asks through the harness's question tool, named by the new `HarnessLayout.questionTool`: `AskUserQuestion` on Claude Code, `question` on OpenCode. It asks at most 4 questions per call, in batches when there are more, each standing on its own, and the recommended option first. On `codex` and `mcp`, which have no such tool, it asks one numbered block at the end of the reply. The answer goes into the stage's output as its status, noted `(approved in chat YYYY-MM-DD)`. Autonomous self-checks are unchanged. Reaches a workspace at its next `workspace init`.
|
|
143
|
+
|
|
144
|
+
### Patch Changes
|
|
145
|
+
|
|
146
|
+
- 9e650ec: SW-495: the approval notes (`(approved in chat YYYY-MM-DD)`, `(approved via check-in YYYY-MM-DD)` and their PR-file YAML-comment forms) now come from one export, `APPROVAL_NOTE` in `pr_rules.ts`, which builds the generated approval text. A new test fails when the `w9r-pr` skill, `docs/workspace-lifecycle.md` or the plugin's `check-in` skill spells a note any other way. The generated text is unchanged.
|
|
147
|
+
|
|
148
|
+
## 0.81.1
|
|
149
|
+
|
|
150
|
+
### Patch Changes
|
|
151
|
+
|
|
152
|
+
- 532983f: SW-526: the generated ICM contracts' test step gives the background log pattern without a trailing `echo "exit $?"` (read the exit code from the job's completion notice) and keeps the echo for a foreground call.
|
|
153
|
+
- 7689f33: SW-534: the implement-stage contract (feature and refactor) reruns the code review on the delta when the exercise changes code, and appends the result to review.md.
|
|
154
|
+
- c04d525: SW-535: the 04_propose contract tells the agent to file each follow-up that decisions.md promises, and write its issue key there, before writing the PR file.
|
|
155
|
+
|
|
156
|
+
## 0.81.0
|
|
157
|
+
|
|
158
|
+
### Minor Changes
|
|
159
|
+
|
|
160
|
+
- 4822338: SW-509: the fixture has a GitHub project for `o/<repo>#N` refs, so a seeded issue resolves to the fake gh. `with-fixture-workspace --issue o/app#1` (or `withFixtureWorkspace(fn, { issue })`, or `issue_id` / `group:` in a `--workspace` file) shows `issue start`, `pr open`, `pr status` and `workspace close` moves (`gh issue edit` / `close` / `comment`, for the lead and each member) in `--- gh calls`. Jira keys still resolve to Jira.
|
|
161
|
+
|
|
162
|
+
### Patch Changes
|
|
163
|
+
|
|
164
|
+
- 315ea0c: SW-483: `workspace init` writes `.devcontainer/devcontainer.json` only on a `harness: claude` workspace. On `mcp`, `codex` and `opencode` it writes none (the generated one installs and registers Claude Code only) and removes one it generated before; a hand-written file is kept.
|
|
165
|
+
- 3ef569b: SW-511: the fixture's fake docker logs a gate `exec` as `exec <container> <row command>`, not the whole argv with the inlined `with-temp-home` script, so `--- docker calls` shows which lint or test command ran.
|
|
166
|
+
- 8f25967: SW-520: every git call werkmeister makes drops `GIT_DIR`, `GIT_WORK_TREE` and `GIT_COMMON_DIR` inherited from a calling git (a `rebase --exec` step, a hook), so `git -C <dir>` acts on `<dir>` and not on the caller's repo. This covers `run`/`runAsync`, `runCapture` and every child started with `childEnv()` (handoff commit and push included), the `with-fixture-workspace` fixture, and the stale-build check. The test runners clear the same variables, so a suite started from `rebase --exec` can no longer re-init the real repo.
|
|
167
|
+
|
|
168
|
+
## 0.80.1
|
|
169
|
+
|
|
170
|
+
## 0.80.0
|
|
171
|
+
|
|
172
|
+
### Minor Changes
|
|
173
|
+
|
|
174
|
+
- 24ae868: SW-385: `harness: opencode` (checked on OpenCode 2.2.5). `workspace init --harness opencode` writes AGENTS.md, `.agents/skills`, a plugin in `.opencode/plugins/werkmeister.js` that runs the standards gate, PR guard and agent state, and `opencode.json` with the werkmeister MCP server (scoped by the workspace token) and shell deny rules for shared infra, other workspaces, global installs and `gh pr merge`. Session tools (`claude-send`, transcripts, the Workspaces pane) drive OpenCode over `opencode run` and `opencode session list` / `export`. Session-start context and auto-stop do not run on OpenCode.
|
|
175
|
+
|
|
176
|
+
## 0.79.4
|
|
177
|
+
|
|
178
|
+
### Patch Changes
|
|
179
|
+
|
|
180
|
+
- c9af0ba: SW-503: the grouped-PR stage contracts put each ticket's release entry in that ticket's own commit, so reverting one member keeps the others' notes.
|
|
181
|
+
- 53f0dc8: SW-505: the stage 03 contract says `output/review.md` must be a `| criterion | result | evidence |` table with PASS/FAIL result cells, so a prose review no longer costs a gate pass.
|
|
182
|
+
- 357d007: SW-506: the host test step in the ICM implement contract yields to a service's own `docker exec` recipe; the werkmeister repo's cli `validate:` text now says to build, test and run `with-fixture-workspace` through `docker exec` after `feature up`.
|
|
183
|
+
|
|
184
|
+
## 0.79.3
|
|
185
|
+
|
|
186
|
+
### Patch Changes
|
|
187
|
+
|
|
188
|
+
- fdc8418: SW-500, SW-501, SW-515: the ICM first-stage contract and `w9r-research` now say to run `werkmeister issue start` alone, as its own Bash call, so its allow rule matches. If that call is denied anyway, the agent records the denial, stops and tells the person. When the CLI reads stale, the agent tells the person to apply the printed fix and never runs it itself. `pr-rules` rule 3 runs `werkmeister pr open > .werkmeister/pr-open.log 2>&1` alone. The autonomous run limits allow `issue start`'s status move as well as `pr open`'s. In a grouped workspace every automatic issue move now covers the issue and each `group:` member: In Progress, In Review, Done and abandon. The grouped propose contract checks each member reads Done instead of closing it by hand.
|
|
189
|
+
- 973a6e5: SW-508: standards links ignore `GIT_DIR`, `GIT_WORK_TREE` and `GIT_COMMON_DIR` inherited from a calling git (a `rebase --exec` step, a hook), so they name the standards tree's own repo, not the caller's.
|
|
190
|
+
- 9e36c19: SW-510: a workspace's Claude settings no longer pre-approve `with-temp-home env WERKMEISTER_ROOT=* node --test:*` or its `--no-wait` twin, because `node --test <file>` runs any file, including one the agent just wrote. Both forms are now `ask` rules, and `machine migrate` moves them in existing workspaces. Tests run without a prompt through the registered test commands (`with-temp-home <test command>`), and the generated `werkmeister-context.md` now says so.
|
|
191
|
+
- 4e4fc2a: `werkmeister pr status` now marks a merged PR that sets `skip_release` as shipped (`release: none`) on the first run when its diff has no `.changeset/*.md` and the repo's `release_check` is changesets'. Before this, a test-only PR waited for a release tag until `release: none` was set by hand. A PR with a changeset still waits for its tag (SW-514). `with-fixture-workspace` gains `--pr-field <key>=<value>` to add front matter to the seeded PR file.
|
|
192
|
+
|
|
193
|
+
## 0.79.2
|
|
194
|
+
|
|
195
|
+
### Patch Changes
|
|
196
|
+
|
|
197
|
+
- b5aa231: SW-464: `with-fixture-workspace --workspace <file>` (and `withFixtureWorkspace(fn, { workspace })`) merges a YAML mapping into the fixture's workspace.yml, e.g. `group`, `work_type`, `context_strategy`. Under `context_strategy: icm`, the `--gate` preset writes its PR file, exercise.md and review.md where the icm strategy reads them, so a grouped icm fixture passes `pr gate` with no hand edits.
|
|
198
|
+
- 0ff080d: SW-466, SW-497, SW-467, SW-479, SW-465, SW-468: grouped ICM contracts tell the agent to regenerate a changed generated snapshot (e.g. `strategy-jit.snapshot.json`) inside each ticket's commit, and that commit order across tickets is free (the grouped 02_spec no longer fixes it). The grouped refactor 01_scope lists its steps under each ticket key. Every build stage's review brief rates a finding that can restart or write the host daemon, the real `~/.werkmeister` or another workspace as Must fix. `--group` keys are normalised by the tracker's `parseRef`, and the lead's own key is dropped (a lead-only list writes `group: []`). The `--work-type` help says it selects the ICM stages, and the `workspace init --group` help says "one PR, one commit per ticket".
|
|
199
|
+
- c40c1a9: `with-temp-home` waits at most 90 s for host load (was 10 min), so an agent's foreground call no longer times out before its command starts; `pr gate` keeps its 10 min budget (SW-469).
|
|
200
|
+
- 491f366: SW-470: the fixture parses a `--services` registry only under `--gate`, so a YAML error without it reaches `workspace init`'s own error.
|
|
201
|
+
- 97b0b89: SW-473: `with-fixture-workspace --exec-runs` (and `withFixtureWorkspace(fn, { execRuns: true })`) makes the fixture's fake docker run each `docker exec` command with `sh -c` in the fixture's code repo and return its exit code, so a fixture `pr gate` passes or fails on the registry's `lint` and a service's `test:`. Without it the fake still logs each exec and answers 0. `--exec-runs` must come first, and a workspace's Claude settings get `ask` rules for it (beside the wrapper's allow rule), because it runs the registry's rows on the host; `machine migrate` adds them to existing workspaces. The fixture also blanks `GITHUB_TOKEN`.
|
|
202
|
+
- 1cb2086: SW-475: `pr watch` claims a held wake by rename before appending to it, the way the hook claims it before delivering, so a hook that delivers while `pr watch` writes no longer hands the session the same wake twice.
|
|
203
|
+
- ff8cbb8: SW-476: a held `pr watch` wake (`.werkmeister/wake-nudge*`: the per-session file, a `.claimed` leftover, or the older `wake-nudge.json`) is git-ignored on the workspace record branch, so a record commit never takes it. `feature up` adds the rule to an older workspace and untracks one already committed.
|
|
204
|
+
- c21edc6: SW-482: the fixture's fake gh answers `gh pr view`, and `with-fixture-workspace --merged` (or `withFixtureWorkspace(fn, { merged: true })`) makes it report the PR merged, so a fixture run goes `workspace init → pr open → pr status → workspace close` with exit 0.
|
|
205
|
+
- 4a0a958: SW-488: Resume's open-session guard matches the workspace's harness, so in a `harness: codex` workspace an open interactive `codex` gets a pending "continue" instead of a headless `codex exec resume` beside it.
|
|
206
|
+
- d68fa02: SW-489: `pr watch`'s wake finds the folder a session started in through the workspace's harness, so a `harness: codex` session started in a code worktree records `session_cwd` at `pr open` and is resumed from there, not from the workspace root.
|
|
207
|
+
- e644b9c: PR rules and gate fixes (SW-494 group):
|
|
208
|
+
|
|
209
|
+
- SW-494: a `pr watch` conflict wake tells the session to rebase onto `origin/<base>` (never merge it in), to take the base's copy of a conflicting generated file and regenerate it, and prints the person's `git -C <worktree> push --force-with-lease=<branch>:<old sha>` filled in; `pr-rules.md` rule 5 says the same.
|
|
210
|
+
- SW-493: a chat reply of "continue" counts as approval, beside "decided", "approved" and "proceed".
|
|
211
|
+
- SW-116: `pr-rules.md` says `pr open` runs the gate itself, so `pr gate` is only for checking without opening.
|
|
212
|
+
- SW-155: the PR file's acceptance table takes `MANUAL` for a criterion only the reviewer can check; the gate passes it and lists it.
|
|
213
|
+
- SW-119: a registry `test: none` (or `test: "none: <reason>"`) marks a service with no tests; the gate reports it skipped with the reason.
|
|
214
|
+
- SW-490: `pr gate` fails while a decided question in `decisions.md` promises a follow-up and names no issue key.
|
|
215
|
+
|
|
216
|
+
## 0.79.1
|
|
217
|
+
|
|
218
|
+
### Patch Changes
|
|
219
|
+
|
|
220
|
+
- 965fbfc: SW-406: the werkmeister-core gate tests run faster: the core container gets 3 CPUs, the core suite's test concurrency follows the container's CPU limit (at most 4), and the two slow compose-probe tests wait about 1 s instead of 11 s (`WERKMEISTER_TIMEOUT_SCALE`, test-only).
|
|
221
|
+
- 5794317: SW-450: the pre-PR gate names the services a diff left untested (`tests: not touched, not run: …`), and a repo fragment's `repo:` block takes `path_mappings` for paths no service owns, so the werkmeister repo no longer tests all five services for a change under `apps/werkmeister-mcp`, `plugins`, `.github`, `.vscode` or `.claude`.
|
|
222
|
+
|
|
223
|
+
## 0.79.0
|
|
224
|
+
|
|
225
|
+
### Minor Changes
|
|
226
|
+
|
|
227
|
+
- e033938: SW-384: Codex sessions. A `harness: codex` workspace reads its sessions from `~/.codex/sessions`: the pane and dashboard show Working / Needs you / Done and the transcript, `claude-send` runs `codex exec [resume <id>]`, `claude-resume` and the pane's Continue / New Session open `codex`, and usage reads Codex's token counts. Machine-level asks (Jira fallback, refine) and PR reconcile stay on Claude.
|
|
228
|
+
|
|
229
|
+
### Patch Changes
|
|
230
|
+
|
|
231
|
+
- 2cce430: SW-455: the generated approval text (ICM contracts, the router's one rule, `pr-rules.md`, the `w9r-pr` skill) no longer accepts a relay "only through check-in", a channel the receiving session cannot see. `/check-in` now marks the file it flips with `(approved via check-in YYYY-MM-DD)`, and that note is the approval. A peer message that claims approval while the file is unchanged, or that comes with a flipped status and no check-in note, is refused: the agent leaves the file as found and asks the person.
|
|
232
|
+
|
|
233
|
+
Graduation now also offers a decision whose status line carries an approval note (`status: decided (approved via check-in …)` or `(approved in chat …)`); before, the note hid it.
|
|
234
|
+
|
|
235
|
+
- 3484a52: SW-463: a `pr watch` wake held for an open but idle Claude session now reaches it without anyone typing. Workspace status rows list `heldWakes` (the sessions with a `wake-nudge-<id>.json`), and the werkmeister plugin's band submits one short prompt when its own session has one, no turn is running and Desktop's wrap_up box isn't waiting. The prompt's UserPromptSubmit hook then delivers and claims the wake as before.
|
|
236
|
+
|
|
237
|
+
## 0.78.0
|
|
238
|
+
|
|
239
|
+
## 0.77.0
|
|
240
|
+
|
|
241
|
+
### Minor Changes
|
|
242
|
+
|
|
243
|
+
- d16406b: SW-189: `harness: codex` — a Codex CLI workspace gets werkmeister's hooks in `.codex/hooks.json` (the standards gate reads `apply_patch` paths, PermissionRequest marks needs-you, the pause gate denies without `continue: false`) and the werkmeister MCP server in `.codex/config.toml` with `default_tools_approval_mode = "approve"` and `tool_timeout_sec = 1800`.
|
|
244
|
+
|
|
245
|
+
## 0.76.1
|
|
246
|
+
|
|
247
|
+
## 0.76.0
|
|
248
|
+
|
|
249
|
+
## 0.75.0
|
|
250
|
+
|
|
251
|
+
## 0.74.0
|
|
252
|
+
|
|
253
|
+
### Patch Changes
|
|
254
|
+
|
|
255
|
+
- 7090ac2: SW-353: the generated approval text (ICM contracts, the router's one rule, `pr-rules.md`, the `w9r-pr` skill) now says a message from another session that only says to re-read an output file is not the approval: the file's `decided` or `approved` status is, so the agent acts on it without asking the person again. A message that claims approval with no file change is still not accepted, and a relay that exists only in a message still counts only through check-in.
|
|
256
|
+
- 149111d: SW-434: the generated `issues-report.md` search step also searches by the workspace's name (Jira `text ~ "<workspace>"`, GitHub `--search "<workspace>"`), since other sessions file from the same review and exercise files and an issue filed from a workspace names it.
|
|
257
|
+
|
|
258
|
+
## 0.73.0
|
|
259
|
+
|
|
260
|
+
## 0.72.1
|
|
261
|
+
|
|
262
|
+
## 0.72.0
|
|
263
|
+
|
|
264
|
+
## 0.71.3
|
|
265
|
+
|
|
266
|
+
## 0.71.2
|
|
267
|
+
|
|
268
|
+
## 0.71.1
|
|
269
|
+
|
|
270
|
+
## 0.71.0
|
|
271
|
+
|
|
272
|
+
### Minor Changes
|
|
273
|
+
|
|
274
|
+
- 057962d: SW-388: ICM feature and refactor implement stages, and the bugfix/incident exercise stage, end with a code review step. A fresh subagent runs `ponytail:ponytail-review` over each code worktree's uncommitted diff, every Must fix is fixed in place, and the stage writes `output/review.md` with a PASS/FAIL row per worktree. The bugfix/incident exercise stage now loads the coding standards, since the review may change code there. The PR gate has a new `review` check: it fails when `review.md` is missing or has a row that is not `PASS`, and `skip_review: <reason>` in the PR file skips it. Existing workspaces are unaffected: the gate asks for a review only when a stage's `CONTEXT.md` names `review.md`, which happens once `werkmeister workspace init` re-renders the contracts (init rewrites every stage contract).
|
|
275
|
+
|
|
276
|
+
## 0.70.2
|
|
277
|
+
|
|
278
|
+
### Patch Changes
|
|
279
|
+
|
|
280
|
+
- d99e9ea: SW-237: `werkmeister machine setup` leaves the werkmeister plugin alone when `CLAUDE_CODE_PLUGIN_DIRS` (the shell's, or `env` in `~/.claude/settings.json`) already loads it from a checkout: no marketplace add or install, reported as `loaded from <dir> (dev mode)`. `--check` no longer reports such a machine as missing the plugin, and flags a marketplace copy that overrides the checkout, with the uninstall commands.
|
|
281
|
+
- 924600e: `werkmeister pr status` says why a merged PR in a `ship: tag` repo is not shipped yet: `not shipped yet: no release tag contains <sha>; latest tag <tag> (<date>) predates the merge (<date>)`, or `no tag matches <pattern> yet` when the repo has none. With `skip_release` set it also names the `release: none` fix (SW-389).
|
|
282
|
+
|
|
283
|
+
## 0.70.1
|
|
284
|
+
|
|
285
|
+
## 0.70.0
|
|
286
|
+
|
|
287
|
+
### Minor Changes
|
|
288
|
+
|
|
289
|
+
- dd6aed2: Revert SW-309: agents no longer ask through Markdown forms, and /workspaces cards no longer show the `❓ <file> · N open` line or the Answer sheet. ICM stages go back to their earlier gate files, `decisions.md` goes back to its earlier format, and werkmeister-core drops `parseForm`, `applyAnswers`, `formErrors`, `submitForm` and `openForms`. A replacement is tracked in Jira.
|
|
290
|
+
|
|
291
|
+
### Patch Changes
|
|
292
|
+
|
|
293
|
+
- c4263c6: The `werkmeister-install` one-shot runs `pnpm install --frozen-lockfile` on every run instead of skipping while the lockfile stamps match: pnpm's own up-to-date check takes seconds and restores a package missing under the `node_modules` volume, which the stamps cannot see; the npm path keeps the stamp skip (SW-362). The PR gate runs that one-shot once for a container it keeps: `compose up` of a running service leaves it as is and re-runs the exited setup, so a kept container's rows run on a checked volume. Its `containers` row reads `kept (… already running, setup one-shot re-run)`.
|
|
294
|
+
|
|
295
|
+
## 0.69.0
|
|
296
|
+
|
|
297
|
+
### Minor Changes
|
|
298
|
+
|
|
299
|
+
- 2a72809: The Claude adapter's `ask` helpers (starting-prompt refine, issue distillation, the Jira Atlassian-MCP fallback) run through the Agent SDK with structured output instead of scraping `claude -p`'s stdout: a failed run names the SDK's result subtype. `@anthropic-ai/claude-agent-sdk` is a required dependency of `werkmeister` again (SW-386).
|
|
300
|
+
|
|
301
|
+
## 0.68.0
|
|
302
|
+
|
|
303
|
+
### Minor Changes
|
|
304
|
+
|
|
305
|
+
- 9368784: `werkmeister registry sync --from <worktree> --out <file>` merges that checkout's `.werkmeister/services.yaml` in place of its repo's fragment and writes the registry to `<file>`, so a fragment change can be proven inside the workspace before merge. `~/.werkmeister` is left untouched; `--check` works against `--out` too (SW-281).
|
|
306
|
+
|
|
307
|
+
### Patch Changes
|
|
308
|
+
|
|
309
|
+
- b442dfd: Devcontainer workspaces install the ponytail plugin in `postCreateCommand`, into the shared `~/.claude` volume, so Claude sessions inside the container run with ponytail like host sessions do.
|
|
310
|
+
- 2c525f2: SW-377: the werkmeister-install-check hooks ship with the Claude Code plugin. The plugin's hooks module runs `skills/install-check/check.sh --fix` when an interactive session starts and, after a repair, refuses the person's next typed prompt once; messages from other sessions, background notifications, slash commands and headless sessions are never refused (SW-348). `werkmeister machine setup` retires the hand-copied `~/.claude/settings.json` hook entries and names the leftover `~/.claude/skills/werkmeister-install-check` folder.
|
|
311
|
+
|
|
312
|
+
## 0.67.0
|
|
313
|
+
|
|
314
|
+
### Minor Changes
|
|
315
|
+
|
|
316
|
+
- b0496a9: SW-188: a generic MCP adapter. `werkmeister workspace new --harness mcp` (config `harness`, env `WERKMEISTER_HARNESS`) generates a workspace for any agent harness with MCP and skills: `AGENTS.md` instead of `CLAUDE.md`, skills in `.agents/skills`, no subagents, and hooks reported as unavailable (`hooks: unavailable on mcp …` at init and `up`, the `hooks` MCP tool's new `hooks` field, a note in `AGENTS.md`) rather than silently missing. The harness connects to the host daemon at `http://127.0.0.1:3002/mcp`; `docs/harnesses.md` says how for Codex CLI, Cursor and Copilot CLI.
|
|
317
|
+
|
|
318
|
+
## 0.66.3
|
|
319
|
+
|
|
320
|
+
### Patch Changes
|
|
321
|
+
|
|
322
|
+
- 8efb3ce: A message relayed from another Claude Code session now shows in `claude-transcript` as a user turn, and the new `relay-check` MCP tool says whether a relay sent at a given time was picked up by the workspace's session. When it was not, it returns `not picked up: type <text> in <session>` for the check-in skill to report instead of a success (SW-348).
|
|
323
|
+
|
|
324
|
+
## 0.66.2
|
|
325
|
+
|
|
326
|
+
### Patch Changes
|
|
327
|
+
|
|
328
|
+
- adf5e2a: The generated `.werkmeister/werkmeister-context.md` now marks each host whose service has a `feature_router` with no remote upstream as answering only while a feature runs that service (in Quick Health Checks and Proxy Routing), and tells the reader to check `~/.werkmeister/infra/proxy/features.json` before reading a TLS reset (curl exit 35) with a healthy `w9r-caddy` as an outage.
|
|
329
|
+
- 99e3380: `werkmeister pr gate` now prints `✓ containers: skipped (<container> already running)` when a running feature container means it skips starting a service, instead of omitting the row.
|
|
330
|
+
|
|
331
|
+
## 0.66.1
|
|
332
|
+
|
|
333
|
+
### Patch Changes
|
|
334
|
+
|
|
335
|
+
- c5b8878: `werkmeister feature up` and `workspace init` now render a workspace's `compose.yml` with each of its repos' own `.werkmeister/services.yaml` fragment merged over the global registry, the way the pre-PR gate already tests a branch. A branch that changes a service's `command:`, image, setup or volumes runs that change in its feature containers before merge; other repos' services still come from the registry, which is left untouched (SW-337).
|
|
336
|
+
|
|
337
|
+
## 0.66.0
|
|
338
|
+
|
|
339
|
+
### Minor Changes
|
|
340
|
+
|
|
341
|
+
- 2da3414: SW-341: `werkmeister feature up` (including `--service`) and `werkmeister infra proxy-reload` now resolve every routed hostname after regenerating the Caddyfile and warn about any that does not resolve to 127.0.0.1 on this machine, printing the exact `echo "127.0.0.1 <hosts>" | sudo tee -a /etc/hosts` line and pointing at `werkmeister machine setup`. The command still succeeds.
|
|
342
|
+
|
|
343
|
+
### Patch Changes
|
|
344
|
+
|
|
345
|
+
- d543024: Machine config writes (`werkmeister machine config --write`, `machine setup`, `machine migrate`) refuse inside a workspace container with `INSIDE_CONTAINER`: the container's HOME is the host home and `~/.werkmeister` is bind-mounted, so a write there pinned the host's `werkmeister_root` and `env.sh` to the container's `/app/apps/cli` (SW-354). A config relocated with `WERKMEISTER_CONFIG_FILE` is still written. The `setup_run_pending` test, which did the writing from the PR gate's container, now stubs the writers.
|
|
346
|
+
- a2a7db3: `werkmeister feature up` commits the `.claude/settings.json` and `.werkmeister/claude-permissions.json` it rewrites on the record branch (`chore: refresh Claude settings`), the way `workspace create` commits them, instead of leaving them uncommitted when a werkmeister upgrade added a permission rule since the workspace was inited.
|
|
347
|
+
|
|
348
|
+
## 0.65.0
|
|
349
|
+
|
|
350
|
+
### Minor Changes
|
|
351
|
+
|
|
352
|
+
- 7fa111c: `werkmeister pr status` (and the PR watch pass) now notices a merged PR that changes its repo's `.werkmeister/services.yaml` fragment. By default it prints `registry sync needed: <repo> fragment changed in <url>` and shows it on the workspace card. With `sync_on_merge: auto` in the fragment's `repo:` (or in `services.local.yaml` under `repos.<repo>`), it fast-forwards the clean source checkout, runs `registry sync`, prints what changed and names the running features to bring up again with `werkmeister feature up`.
|
|
353
|
+
|
|
354
|
+
## 0.64.2
|
|
355
|
+
|
|
356
|
+
### Patch Changes
|
|
357
|
+
|
|
358
|
+
- 3c8f294: SW-336: `werkmeister pr gate` (and `--tests`) runs each touched service's test in that service's own container, starting it on demand with its setup service, whatever its `type`. Before, a touched `web` service's test ran in whichever sibling container was running or started (e.g. an `api` service's), with the wrong image and `node_modules`. Lint and the release check still use any one container of the repo.
|
|
359
|
+
|
|
360
|
+
## 0.64.1
|
|
361
|
+
|
|
362
|
+
### Patch Changes
|
|
363
|
+
|
|
364
|
+
- d1f17bc: A workspace whose worktree is the werkmeister repo now pre-approves the built CLI run through its temp-HOME wrapper (`Bash(<workspace>/.werkmeister/bin/with-temp-home <worktree>/apps/cli/werkmeister:*)`), so the auto-mode classifier no longer refuses a CLI-level exercise. The bare wrapper is still never allowed. Existing workspaces pick the rule up with `werkmeister machine migrate`.
|
|
365
|
+
- ce1256e: SW-231: the generated `werkmeister-context.md` tells the exercise step to start each command with the `with-temp-home` script (`with-temp-home env WERKMEISTER_ROOT=… <cmd>`) and names the allow rule `Bash(<script>:*)` that covers it, so a permission rule can match.
|
|
366
|
+
- 3223e06: SW-343: `machine setup` without a TTY (or with `--yes`) now reports the skipped CA trust and /etc/hosts as ⚠ with the command to finish, and ends with "Setup incomplete" instead of "Machine is set up."
|
|
367
|
+
|
|
368
|
+
## 0.64.0
|
|
369
|
+
|
|
370
|
+
### Minor Changes
|
|
371
|
+
|
|
372
|
+
- a34b387: A repo can carry its own standards in `<repo>/.werkmeister/standards/`, laid out like the global standards tree. A workspace that checks out the repo loads them on top of `overrides.standards_dir`, and they are active there without an archetype naming them (`standards_override.remove` still drops one). Their generated `w9r-*-standard` skills link to the repo they came from. `werkmeister standards check` also validates every repo tree under `paths.repos_root` and fails when a repo module reuses a global module's `domain`; the global one wins at load time. `checkStandards()` returns the repo trees it checked as `repoTrees`, and `loadStandardsModules` takes the workspace's repo paths as a second argument.
|
|
373
|
+
|
|
374
|
+
### Patch Changes
|
|
375
|
+
|
|
376
|
+
- 90c30e8: SW-335: `werkmeister machine detect` names the shared install's `node_modules_volume` `<repo>-node-modules` instead of `node-modules`, so two repos in one workspace no longer share one install volume (an existing fragment's name is kept on a re-run). With a lockfile, the install command is npm's clean install or `pnpm install --frozen-lockfile`, so an up can't rewrite it. New docs page: docs/repo-setup.md, on setting up a repo for Werkmeister.
|
|
377
|
+
|
|
378
|
+
## 0.63.0
|
|
379
|
+
|
|
380
|
+
### Minor Changes
|
|
381
|
+
|
|
382
|
+
- 7b92579: `werkmeister pr gate --tests` runs the gate's lint and test rows in the workspace container on the uncommitted tree; the implement stage runs it so container-only failures show before `pr open` (SW-316).
|
|
383
|
+
|
|
384
|
+
## 0.62.1
|
|
385
|
+
|
|
386
|
+
### Patch Changes
|
|
387
|
+
|
|
388
|
+
- 63f4d97: SW-324: `werkmeister issue view` prints the whole description instead of cutting it at 4000 characters without a sign. Where the cap still applies (the briefing that refine and the form pre-fill feed to a model), the text now ends with `[cut: N more characters not shown]`. `issueBriefing` takes an optional `{ full: true }`.
|
|
389
|
+
|
|
390
|
+
## 0.62.0
|
|
391
|
+
|
|
392
|
+
### Minor Changes
|
|
393
|
+
|
|
394
|
+
- 506e04b: SW-309: agents ask a person through Markdown forms. Each question in a form is a `## ` heading with `status: proposed`. Options are `- ( )` (pick one) or `- [ ]` (pick any), with `(recommended)`, `Other:`, and an empty `**Label:**` text box. `decisions.md` uses this syntax, every later ICM stage except propose ends by writing `output/question-approval.md`, and a question that comes up partway through a stage goes in `output/question-<slug>.md`, after which the agent stops. On the /workspaces pane, a workspace card with an open form shows `❓ <file> · N open`. Its Answer button opens the form inside the card as one sheet with one Submit. Submit writes the ticks and text back, changing only those lines and the `status:` line, refuses when the file changed after it was drawn, and asks the root session to relay "gate decided, continue". New in werkmeister-core: `parseForm`, `applyAnswers`, `formErrors`, `submitForm` and `openForms`. A `question-*.md` file never counts as a stage's output.
|
|
395
|
+
|
|
396
|
+
## 0.61.0
|
|
397
|
+
|
|
398
|
+
### Patch Changes
|
|
399
|
+
|
|
400
|
+
- c0b4184: SW-304: the generated human checks, the ICM one rule, `pr-rules.md` and the `w9r-pr` skill now say a direct chat reply from the person ("decided", "approved", "proceed") counts as the approval: the agent flips the status itself and notes it, e.g. `(approved in chat YYYY-MM-DD)`. An approval relayed by another session counts only when it arrives through check-in.
|
|
401
|
+
- 1843168: SW-313: a `pr watch` wake can now fix the PR in a workspace without `autonomous`. The headless turn gets the workspace's autonomous allow and deny rules, and the pass waits for it. Refused calls that left the fix behind (uncommitted edits or unpushed commits) show as `auto-fix blocked: needs approval — …` in the log and as `needs-you` on the workspace card; refusals the agent got around are one `refused and retried` log line. A wake for a session that is still open, idle included, is held (with `needs-you`, `waiting for your open session: …`) instead of forking that session's history.
|
|
402
|
+
|
|
403
|
+
## 0.60.2
|
|
404
|
+
|
|
405
|
+
### Patch Changes
|
|
406
|
+
|
|
407
|
+
- a9d211b: SW-187: Claude Code-specific code now sits behind a harness interface (`harness()` returning a `HarnessAdapter`), with the Claude adapter under `harness/claude/` in core, the MCP server, the CLI and the VS Code extension. That covers the `claude` CLI calls, `CLAUDE.md` and hooks generation, the transcript reader, usage and pricing, session driving, the 4 `claude-*` MCP tools, `detect --assist`'s Agent SDK resolver, the `claude --continue` terminals, `wrap_up` and PR rule 5's Desktop sentence. No behaviour change: generated workspace files and the MCP tool list are identical, and every old export from `werkmeister-core` still resolves.
|
|
408
|
+
|
|
409
|
+
## 0.60.1
|
|
410
|
+
|
|
411
|
+
### Patch Changes
|
|
412
|
+
|
|
413
|
+
- 9238da2: SW-270: `werkmeister workspace close --no-pr "<reason>"` (MCP `close-out` `no_pr`) closes a workspace whose work changed no repo as done, not abandoned. It needs no PR files, `postmortem.md`, and every code worktree clean with no commits ahead of its base, and refuses with the failing repo named otherwise. The issue moves to Done with `No PR needed: <reason>`, which goes in `closeout.md` too. A no-PR workspace closed without the flag now hears about `--no-pr` beside `workspace abandon`.
|
|
414
|
+
- 35c2dd8: SW-286: `pr gate` / `pr open` check `exercise.md` with the PR-file checks, so a malformed exercise table (`PASS: …`, `(reproduces bug)`) fails before any test command runs instead of after them. The generated ICM contracts that write `exercise.md` (03_implement, the bug-fix/incident 03_exercise, the refactor implement stage) now say the result cell is exactly `PASS` or `FAIL` with output in its own column (`| check | result | evidence |`, or `detail` for 03_exercise), that this holds for every table in the file, and that a superseded run goes in prose, not a `FAIL` row. 04_propose says the same for every result-column table in the PR body, and the JIT `w9r-test` skill matches.
|
|
415
|
+
- 316d1d6: `werkmeister issue start` prints a `CLI:` line saying whether the installed werkmeister is current (the latest release, or a link on `main` containing `origin/main`) and the fix when it is stale; 01_research step 0 tells the agent to stop on a stale CLI.
|
|
416
|
+
|
|
417
|
+
## 0.60.0
|
|
418
|
+
|
|
419
|
+
### Minor Changes
|
|
420
|
+
|
|
421
|
+
- 429226a: SW-271: `werkmeister issue start`, step 0 of the first ICM stage and of JIT's `w9r-research`, now also brings each code worktree up to its base. It fetches origin, and a branch with no commits of its own is fast-forwarded to `origin/<base>`. A branch with its own commits, uncommitted changes, or a fetch that fails is left as is, with a `Worktree:` line such as `app: 4 commits behind origin/main (has its own commits, left as is)`. It still exits 0, so a stage never fails on it. A wave workspace created days before its run no longer starts research on stale code.
|
|
422
|
+
|
|
423
|
+
### Patch Changes
|
|
424
|
+
|
|
425
|
+
- c73f9ee: SW-262: the PR gate's host-load wait is one 10 min budget for the whole run, shared by lint, each test and the release check, instead of up to 30 min before each one. Before it waits, `pr gate` / `pr open` says how long it may wait (`waiting up to 10 min for host load …`) and prints a `still waiting …` line each minute; once the budget is spent the remaining checks run at once and their rows say so. `werkmeister internal wait-load` (run by `with-temp-home`) waits up to 10 min the same way. `_shared/pr-rules.md` rule 3 and the 04_propose contract now say to run `pr open` detached with `nohup … ; echo "exit $?" >> .werkmeister/pr-open.log` and read the log to the `exit` line (SW-266, SW-273, SW-298).
|
|
426
|
+
|
|
427
|
+
## 0.59.4
|
|
428
|
+
|
|
429
|
+
### Patch Changes
|
|
430
|
+
|
|
431
|
+
- ab40ce4: SW-280: the issue extractor and brief format now pick and name a service whose container the ticket measures, so `werkmeister feature up` starts it.
|
|
432
|
+
|
|
433
|
+
## 0.59.3
|
|
434
|
+
|
|
435
|
+
### Patch Changes
|
|
436
|
+
|
|
437
|
+
- a68920b: SW-299: the workspace lookup from a container's compose label gives its `docker ps` the teardown's 300 s limit, so a lookup starved on a loaded host no longer reads as "no workspace" (`pause` failed `NO_WORKSPACE`). Test-only: the forward-probe tests give healthy probes 30 s, and pause-resume's fake Engine keeps idle connections open across a blocked event loop.
|
|
438
|
+
|
|
439
|
+
## 0.59.2
|
|
440
|
+
|
|
441
|
+
### Patch Changes
|
|
442
|
+
|
|
443
|
+
- c87d8b7: SW-279: a run whose `WERKMEISTER_CONFIG_FILE` is not the default `~/.werkmeister/config.yaml` (a temp-HOME fixture) no longer reloads or starts the shared `w9r-caddy`. Before, `workspace abandon` / `workspace close` on such a fixture loaded the fixture's Caddyfile into the real proxy and took routing down for every workspace. `reloadCaddy` and `startCaddy` now skip with `Caddy reload skipped (WERKMEISTER_CONFIG_FILE is not the default home)`, as they already did under `WERKMEISTER_FIXTURE=1`, so teardown, `feature up` / `down`, `infra proxy-reload` and the MCP tools all leave the shared proxy alone.
|
|
444
|
+
|
|
445
|
+
## 0.59.1
|
|
446
|
+
|
|
447
|
+
### Patch Changes
|
|
448
|
+
|
|
449
|
+
- db3505a: SW-264: the generated `with-temp-home` exports `WERKMEISTER_HOST=http://127.0.0.1:1` unless the caller already set it, so a run under a temp HOME can no longer reach the real daemon on :3002; daemon calls raise `DaemonUnavailable` and take their local fallback. Existing workspaces pick it up with `werkmeister machine migrate`.
|
|
450
|
+
|
|
451
|
+
## 0.59.0
|
|
452
|
+
|
|
453
|
+
### Minor Changes
|
|
454
|
+
|
|
455
|
+
- 6e4707d: SW-268: `werkmeister machine config` (bare or `--show`) now also prints the whole `config.yaml` as YAML with secrets masked: any value under a key named like `token`, `secret`, `password`, `credential`, `api_key`, `private_key`, `access_key` or `auth` (so `jira.api_token`) prints as `<redacted>`, and empty values stay empty. The autonomous run limit that says not to read the real `~/.werkmeister` now points agents at this view.
|
|
456
|
+
|
|
457
|
+
### Patch Changes
|
|
458
|
+
|
|
459
|
+
- 0b6e9c3: SW-277: test-only. The wrapper npx-bootstrap test runs the wrapper async with a 120 s limit (macOS's first-exec check of its fresh stubs slows with host load), and `infra-retire-legacy`'s `argvOf` names a missing docker-stub line and prints the whole log instead of throwing a TypeError.
|
|
460
|
+
- d7ba2ba: SW-278: `werkmeister pr gate` no longer passes silently with zero tests run when a PR's changed files map to no service (for example, the registry has no `repos:` entry for the worktree). It fails a `tests` row that names the files, unless the PR file sets `skip_tests: { tests: <reason> }`. An empty diff and a docs-only diff (all `no_service_paths`) behave as before. The `cli` validate text now says what a gate fixture needs.
|
|
461
|
+
|
|
462
|
+
## 0.58.0
|
|
463
|
+
|
|
464
|
+
### Minor Changes
|
|
465
|
+
|
|
466
|
+
- 80b6117: SW-249: the generated `.devcontainer/devcontainer.json` passes `CLAUDE_CODE_OAUTH_TOKEN` from the host at every attach (`remoteEnv`, `${localEnv:CLAUDE_CODE_OAUTH_TOKEN}`), so a container started with a `claude setup-token` token set needs no Claude login, and no generated file holds the token. Every workspace's container now mounts one shared `~/.claude` volume, `werkmeister-claude-home`, so you log in once; each container still reaches only its own workspace, because the werkmeister registration lives outside the volume and sends that workspace's token. Close-out never removes the shared volume. The per-workspace `features-claude-home-<workspace>` volumes from 0.54.0 are no longer used; drop them with `docker volume rm`.
|
|
467
|
+
|
|
468
|
+
## 0.57.0
|
|
469
|
+
|
|
470
|
+
### Patch Changes
|
|
471
|
+
|
|
472
|
+
- 1cbdd5d: SW-208: `pr open`, `pr status` and `pr gate` resolve `gh` through `WERKMEISTER_GH_BIN` and `docker` through the new `WERKMEISTER_DOCKER_BIN`, as do `runCapture` and `proc.run`/`runAsync`. A fixture run of the built CLI can therefore fake every gh and docker call. A fake first on PATH is not enough, because `childEnv()` puts `/opt/homebrew/bin` and `/usr/local/bin` ahead of it.
|
|
473
|
+
- 11d002e: SW-257: taking down the pre-rename `sandbox-infra` stack no longer gives up when the `docker ps` / `docker volume ls` check for held containers takes over 10 s on a loaded host; the check now waits as long as the teardown does, so `infra start` and `down` still free the old stack's ports and volumes.
|
|
474
|
+
|
|
475
|
+
## 0.56.0
|
|
476
|
+
|
|
477
|
+
### Minor Changes
|
|
478
|
+
|
|
479
|
+
- d8c9205: SW-160: `werkmeister pr watch` is werkmeister's own auto-fix. While a PR is open, each pass wakes the session that opened it, in one message, when a check on the head commit fails (with the first error lines of `gh run view --job <id> --log-failed`), when a review or comment arrives (quoted), or when the PR conflicts with its base (with the base commits that moved). Each event is claimed once in the PR file (`notified_checks`, `notified_review`, `notified_conflict`). A terminal or VS Code session whose transcript changed in the last 5 minutes is held and woken on a later pass, so a headless turn never runs beside a live one. PR rule 5 no longer turns on the desktop app's Auto-fix (`ccd_pr set_monitor`), and `pr open` no longer prints the `<pr-created>` marker.
|
|
480
|
+
|
|
481
|
+
## 0.55.1
|
|
482
|
+
|
|
483
|
+
### Patch Changes
|
|
484
|
+
|
|
485
|
+
- a1a40a1: SW-260: give `refreshKubeconfig`'s local `kubectl` calls 30 s and 15 s instead of 10 s and 5 s, so a loaded host no longer reports a missing kubeconfig.
|
|
486
|
+
- 378260d: SW-254: a daemon older than the CLI no longer fails with only `Tool scheduler not found`. When the daemon answers that it has no tool the CLI called, the error says the daemon is older than the CLI and names `werkmeister daemon restart` (code `DAEMON_STALE`). Other daemon errors are unchanged.
|
|
487
|
+
- 6cf6e82: SW-259: under `node --test`, the Docker client gets a dead socket unless the test file called `useDockerEngine()` (fixtures.mjs), so a core test can no longer read the host engine's containers by accident, even when `DOCKER_HOST` is set. Outside a test run nothing changes.
|
|
488
|
+
|
|
489
|
+
## 0.55.0
|
|
490
|
+
|
|
491
|
+
### Minor Changes
|
|
492
|
+
|
|
493
|
+
- b64c8ea: SW-186 (breaking): the CLI is grouped by noun — `feature`, `workspace` (`ws`), `pr`, `issue`, `infra`, `machine`, `budget`, `daemon` and `internal` — e.g. `up f` is now `feature up f`, `init-workspace` is `workspace init`, `close-out --abandon` is `workspace abandon <reason>`, and the hook commands are `internal gate|pr-guard|agent-state|session-end`. Every old spelling still runs, with the same arguments, and prints one deprecation line on stderr; they are removed in 0.58.0. `werkmeister --help` lists the groups.
|
|
494
|
+
|
|
495
|
+
After upgrading, run `werkmeister machine migrate` (`--check` first to see what it would change): it rewrites each workspace's hooks, permission rules and `with-temp-home`, replaces the stale `SANDBOX_MCP_TOKEN` in `secrets.env`, and points Claude Desktop at `werkmeister internal mcp`. Upgrade the CLI and the Claude Code plugin together, and restart open Claude sessions.
|
|
496
|
+
|
|
497
|
+
Also: `feature up <f> --service <name>` starts one service (SW-132); `machine config --check` names `--write` only for problems it fixes (SW-162); new `budget status` and `daemon status|install|uninstall|restart`; gen-docs writes one man page per `noun verb`.
|
|
498
|
+
|
|
499
|
+
### Patch Changes
|
|
500
|
+
|
|
501
|
+
- 5816545: SW-236: `werkmeister machine setup` no longer leaves the host daemon (or the pr-watch agent) down when `launchctl bootstrap` races the bootout and fails with 5: it retries up to 5 times, 1 s apart. A re-run now bootstraps an agent whose plist is unchanged but isn't loaded, so "re-run to retry" works.
|
|
502
|
+
- 662dcfd: SW-239: under `node --test`, `sandboxRoot()` (and so `monorepoRoot()`) throws unless the test called `pinWerkmeisterRoot()`, so a new core test can no longer depend on the unpinned root. Outside a test run nothing changes.
|
|
503
|
+
- 1200605: SW-245: `up` now removes a feature with no local services from the waiting queue when it starts it, so the scheduler no longer leaves it waiting and re-ups it on every grant.
|
|
504
|
+
- 3e0b4a0: SW-246: the scheduler writes a feature's `waiting.json` before `queue.json` when it queues it (`admit` waiting, or preempted), so a workspace with no `.werkmeister/` gets the ENOENT error and stays out of the queue instead of showing as waiting in `status` with a session that never learns it is queued.
|
|
505
|
+
- 2237edb: SW-247: the PR gate's exercise and acceptance checks fail on a result cell that is not exactly `PASS`, such as `FAIL — reason` or `skipped`, in any table with a `result` column. Before, only a cell reading exactly `FAIL` was caught.
|
|
506
|
+
|
|
507
|
+
## 0.54.0
|
|
508
|
+
|
|
509
|
+
### Minor Changes
|
|
510
|
+
|
|
511
|
+
- eaa500b: SW-138: `init-workspace` writes `.devcontainer/devcontainer.json`, so a Claude Code agent in a dev container drives the host daemon on 3002 with the workspace's token; see users-guide § Running the agent itself in isolation.
|
|
512
|
+
|
|
513
|
+
## 0.53.0
|
|
514
|
+
|
|
515
|
+
### Minor Changes
|
|
516
|
+
|
|
517
|
+
- c22be3d: The host daemon now owns the scheduler (SW-185): slot leases, the waiting queue, and starting waiters.
|
|
518
|
+
|
|
519
|
+
- **Through the daemon:** the CLI, the stdio MCP server and the dashboard ask the daemon (`POST /api/tool`, machine token) to allocate, read and release slot leases, admit or queue a feature (including preemption), and read the waiting queue. VS Code reaches it through the CLI and the `status` tool. Agents can't call these ops: they aren't on `/mcp`.
|
|
520
|
+
- **Starting waiters:** after a client's `stop` or `down`, the daemon starts the next waiting feature in the background. The CLI prints `The daemon will start the next waiting feature (N waiting)` and no longer starts the waiter's containers itself. The `pr watch` tick still starts waiters.
|
|
521
|
+
- **Fallback:** with no daemon (or no daemon token, as in CI and tests), a client warns once and schedules in its own process against the same `~/.werkmeister/scheduler/` files and locks. The daemon reads those same files, so existing leases and waiters carry over on upgrade with nothing to migrate.
|
|
522
|
+
- **Fix:** `/api/tool` waits up to 15 minutes for a tool instead of the MCP SDK's 60 s, so a long close-out or grant isn't cut off.
|
|
523
|
+
|
|
524
|
+
## 0.52.1
|
|
525
|
+
|
|
526
|
+
### Patch Changes
|
|
527
|
+
|
|
528
|
+
- 29a1404: SW-224: the generated `with-temp-home` waits before its command while the host's 1-minute load is above `budget.host_load`, using the pr gate's wait (same default, same 30-minute cap, then it runs anyway and says so), through the new internal `werkmeister wait-load`. `WERKMEISTER_NO_WAIT=1` skips it; with no `werkmeister` on PATH the wrapper runs at once.
|
|
529
|
+
- a69799c: SW-235: an autonomous workspace's BRIEF.md no longer tells the agent to ask the user about its open questions. `## Open questions` drops the "Ask / Confirm with the user …" sentences that follow each question, and says to decide each one per the run limits and record the choice under `## Recommendations for reviewer`. Interactive briefs are unchanged.
|
|
530
|
+
|
|
531
|
+
## 0.52.0
|
|
532
|
+
|
|
533
|
+
### Minor Changes
|
|
534
|
+
|
|
535
|
+
- e75909c: SW-206: `werkmeister close-out` (shipped and `--abandon`) writes `usage.md` and `usage.json` into the workspace record before the folder is removed, committed and pushed with `closeout.md`. They hold token totals by model, the USD estimate, per-session rows (first and last turn, turns, models), wall-clock time, approximate per-stage timing, and the run's werkmeister version and commit, model, autonomous setting and ticket. `closeout.md` gets a one-line `Usage:` summary. A retry after a failed removal keeps the first snapshot; a workspace with no transcripts still gets one that says so. The `workspace-usage` MCP tool's sessions gain `firstTurn`, `lastTurn`, `turns` and `models`.
|
|
536
|
+
|
|
537
|
+
### Patch Changes
|
|
538
|
+
|
|
539
|
+
- d1fc18a: SW-198: `pr open` prints the full gate output (`<repo>: gate passed` and every ✓/✗ row, the same text as `pr gate`) before it pushes, not only when the gate fails.
|
|
540
|
+
- a8203cc: SW-219: for a repo whose registry entry has a gate `lint` command, the generated `werkmeister-context.md` How to validate test commands end with `&& <lint>`, so the implement stage that runs them catches a format error before `werkmeister pr open`. Repos without `lint` render unchanged; the PR gate's own lint row is unchanged.
|
|
541
|
+
|
|
542
|
+
## 0.51.1
|
|
543
|
+
|
|
544
|
+
### Patch Changes
|
|
545
|
+
|
|
546
|
+
- 5e12dac: SW-209: `pr gate` and `pr open` wait before each lint, test and release check while the host's 1-minute load is above `budget.host_load` (default twice the CPU count, `0` off), for up to 30 minutes, then run anyway and say so; `--no-wait` skips the wait. A failed check ends with the host load, and the generated `with-temp-home` prints the host load when its command fails, so a timeout under load reads as load.
|
|
547
|
+
|
|
548
|
+
## 0.51.0
|
|
549
|
+
|
|
550
|
+
### Patch Changes
|
|
551
|
+
|
|
552
|
+
- a6b8980: SW-205: workspace records no longer push to a stale repo. `werkmeister setup --check` now fails when `repos_root/sandbox-workspace`'s origin isn't `overrides.workspace_repo` (or the `<gh user>/sandbox-workspace` default), naming both URLs; SSH and HTTPS forms of the same owner/repo match. `werkmeister setup` asks before repointing origin to the configured HTTPS URL, keeps the old URL as a remote named after its owner, then fetches; non-interactive runs change nothing and print the commands. It never fetches a mismatched origin. `werkmeister workspace` refuses to create a record worktree until they match.
|
|
553
|
+
|
|
554
|
+
## 0.50.2
|
|
555
|
+
|
|
556
|
+
### Patch Changes
|
|
557
|
+
|
|
558
|
+
- f9a45e5: SW-183: lock waits no longer block the event loop. `sandboxLock` now awaits a timer between retries instead of `Atomics.wait` and always returns a promise, so two concurrent `up` calls through the HTTP MCP server both finish instead of hanging the server (the Caddyfile lock had no timeout) or failing after 30 s (the slot lock).
|
|
559
|
+
|
|
560
|
+
## 0.50.1
|
|
561
|
+
|
|
562
|
+
### Patch Changes
|
|
563
|
+
|
|
564
|
+
- 18d4490: SW-193: a merged PR that needs no release now reaches `shipped`. When the gate's release check runs and reports nothing to bump (changesets' `NO packages to be bumped`), `werkmeister pr open` writes `release: none` into the PR file, and `werkmeister pr status` ships that PR on merge instead of waiting for a release tag that will never contain it, so close-out is no longer stuck. A PR that needs a release still waits for its tag, and `skip_release` still only skips the gate's check.
|
|
565
|
+
- 16e90f7: SW-194: `werkmeister pr gate` and `pr open` fail fast when a PR file's `repo:` is not the code worktree's `git remote get-url origin` (SSH and HTTPS forms, with or without `.git`). The `repo` row names the field and shows both values, where the mismatch used to surface as a `gh pr create` GraphQL error. The Stage 04 contract and PR rules now say to take `repo:` from that command.
|
|
566
|
+
- 5d2bdcf: SW-199: `pr open` with PR files that exist but are not approved now lists each file and its status (`pr-app.md is at status: draft; set status: approved to open it.`) instead of telling the person to write a file that is already there. "No matching PR file … Write …" is kept for a workspace with no PR files.
|
|
567
|
+
|
|
568
|
+
## 0.50.0
|
|
569
|
+
|
|
570
|
+
### Minor Changes
|
|
571
|
+
|
|
572
|
+
- 050e6a6: SW-181: a file-backed scheduler. A slot is leased in `~/.werkmeister/scheduler/leases.json` the moment it is allocated, so back-to-back creates get distinct slots, and `up`, the gate and `pr open` never move a leased slot: they fail naming whoever sits on it. `up` admits a feature only when the budget (Docker's memory minus 1G and its CPUs by default, `budget:` in config.yaml to override) has room for its memory reservations and CPU limits; otherwise it waits in a queue, its session is held like a pause, and it starts when another feature stops or on the `pr watch` tick. A person's `up` may stop an autonomous feature to make room (`budget.preempt: false` turns it off). The gate checks its ports are free before starting anything, `status` shows the queue, and generated compose sets `memswap_limit` equal to each memory limit.
|
|
573
|
+
|
|
574
|
+
## 0.49.0
|
|
575
|
+
|
|
576
|
+
## 0.48.0
|
|
577
|
+
|
|
578
|
+
### Minor Changes
|
|
579
|
+
|
|
580
|
+
- 2209de7: SW-182: idle-stop and the Resume nudge read a session liveness record the agent-state hook writes to `.werkmeister/agent-state.json` (`sessions`, keyed by session id), not Claude Code's session files. The hook now also runs on SessionStart; a workspace picks that up at its next `up`.
|
|
581
|
+
|
|
582
|
+
## 0.47.4
|
|
583
|
+
|
|
584
|
+
## 0.47.3
|
|
585
|
+
|
|
586
|
+
### Patch Changes
|
|
587
|
+
|
|
588
|
+
- 169783d: `werkmeister connect` replaces a host web server that holds its port but can no longer serve (SW-137). An example is an MCP server left running from a deleted worktree that answers 400 to every request. `start()` and `status()` now check the server over HTTP with a JSON POST, instead of trusting a live PID or an open port. The `web-servers-self` test runs on an ephemeral port and kills anything it spawned, so it can no longer start a detached server on the real port 3002.
|
|
589
|
+
|
|
590
|
+
## 0.47.2
|
|
591
|
+
|
|
592
|
+
### Patch Changes
|
|
593
|
+
|
|
594
|
+
- 6de7585: `werkmeister pr gate` and `pr open` run the PR-file checks (stages, acceptance, approved) first and stop before any test command when one fails (SW-150), so a malformed PR file fails in seconds instead of after every registry test command. The acceptance failure message and `pr-rules.md` now state the accepted format: a result cell that is exactly `PASS` or `FAIL`, evidence in its own column.
|
|
595
|
+
|
|
596
|
+
## 0.47.1
|
|
597
|
+
|
|
598
|
+
### Patch Changes
|
|
599
|
+
|
|
600
|
+
- 74e0654: The generated ICM `01_reproduce` and `01_research` contracts ask each decision that renames, keeps or removes a name, fallback or alias for its definition site and the `git log -S` commit that introduced it, and list "remove" as an option when that commit marks a transition or legacy alias (SW-123).
|
|
601
|
+
|
|
602
|
+
## 0.47.0
|
|
603
|
+
|
|
604
|
+
### Minor Changes
|
|
605
|
+
|
|
606
|
+
- 5556fd4: Workspaces ship a temp-HOME test wrapper (SW-124). `init-workspace` writes `.werkmeister/bin/with-temp-home <cmd…>`, which runs a command with a fresh `mktemp -d` HOME and removes it on exit (keeping corepack's cache, so pnpm isn't re-downloaded). `werkmeister-context.md` states each host test command through it, the allow rules pre-approve the wrapped form, and the ICM test step tells agents, and the subagents they brief, to run every test command through it.
|
|
607
|
+
|
|
608
|
+
### Patch Changes
|
|
609
|
+
|
|
610
|
+
- ddf7977: The reload step of generated stage contracts no longer forbids `werkmeister up` outright (SW-135): if the container is not running, the agent runs `werkmeister up` once first, because a never-started container has no `pnpm` and `docker restart` on it loops on "pnpm: not found". `docker restart` is only for reloading a running container. Applies to the feature `03_implement`, bugfix `03_exercise` and refactor `02_implement` contracts of new workspaces.
|
|
611
|
+
- 9b2283a: SW-148: `runCapture("gh", …)` resolves the binary through `WERKMEISTER_GH_BIN`, so `withFakeGh` fakes every core `gh` call and a test can no longer reach the real `gh` past `childEnv()`'s PATH. The fake now covers `api user` and `search prs` and fails closed on any other subcommand.
|
|
612
|
+
|
|
613
|
+
## 0.46.0
|
|
614
|
+
|
|
615
|
+
### Patch Changes
|
|
616
|
+
|
|
617
|
+
- 9273365: SW-136: `werkmeister config` with no arguments is now read-only — it prints the resolved config (what `--show` printed) and changes nothing. Writing `config.yaml`, `env.sh`, `~/.zshenv` and the Desktop MCP config takes `--write` (the wizard on a TTY, as the bare form ran before) or a value to set (`--repos-root`, `--workspaces-root`). `--show` still works; the `--check` hints now say `werkmeister config --write`.
|
|
618
|
+
- 59b5340: `werkmeister up` leaves a workspace's record branch clean (SW-144). The managed `.gitignore` now ignores compose.yml, compose.gate.yml and the runtime state files (`agent-state.json`, `last-exec`, `starts.json`). `up` adds those rules to an older workspace and untracks a committed compose.yml once. init-workspace seeds `.env` before writing settings, so its allow rules already match what `up` writes.
|
|
619
|
+
- d688847: ▶ Start work no longer dead-ends in auto mode (SW-157): when the pane's own `start_session` call is refused, it drafts a message in the hosting session's prompt box asking its model to make the same call with the same arguments, and the person presses Enter to send it. Run titles from `startArgs()` / `werkmeister start-args` are cut at a word with a trailing `…` instead of mid-word.
|
|
620
|
+
|
|
621
|
+
## 0.45.0
|
|
622
|
+
|
|
623
|
+
### Minor Changes
|
|
624
|
+
|
|
625
|
+
- 493cff4: ▶ Start work on the workspaces pane card starts a fresh workspace's run in a Desktop session (SW-153). `werkmeister start-args` and werkmeister-core's `startArgs()` build its title, model, prompt and run limits, the next-wave skill uses them, and an autonomous workspace's BRIEF.md now carries the run limits.
|
|
626
|
+
|
|
627
|
+
## 0.44.4
|
|
628
|
+
|
|
629
|
+
### Patch Changes
|
|
630
|
+
|
|
631
|
+
- 8ba5b70: SW-145: `werkmeister pr open` prints a `<pr-created>URL</pr-created>` line for each PR it opens, so the Claude desktop app counts the session as the opener and `set_monitor` with `auto_fix: true` is allowed. Rule 5 in the generated `pr-rules.md` now says: when `set_monitor` is refused, don't retry; ask the person to turn Auto-fix on from the PR bar, and don't rely on `get_status` to confirm it.
|
|
632
|
+
|
|
633
|
+
## 0.44.3
|
|
634
|
+
|
|
635
|
+
### Patch Changes
|
|
636
|
+
|
|
637
|
+
- 65e4060: SW-118: the PR gate now tests a branch's own container settings. When the branch's `.werkmeister/services.yaml` renders a touched service differently from the workspace `compose.yml` (a new `environment:` entry, `node_modules_volume`, image, …), `pr gate` / `pr open` start that service from a gate-only `compose.gate.yml` rendered from the branch, run its tests there, then recreate it from `compose.yml` (or stop it, if it wasn't running). Before, the tests ran in containers built from the machine registry, so such changes were never tested.
|
|
638
|
+
|
|
639
|
+
## 0.44.2
|
|
640
|
+
|
|
641
|
+
### Patch Changes
|
|
642
|
+
|
|
643
|
+
- ff1cff8: SW-109: when `pr open` (or any lifecycle move) can't read the issue, it retries the read once, and if that fails too the warning now gives the reason (Jira's HTTP status and message, the network error code, the MCP fallback, or `gh`'s error) and the `werkmeister issue update <ref> --status … --comment …` command that makes the move by hand. `werkmeister issue view` prints the same reasons when it can't read an issue.
|
|
644
|
+
- 09d62c8: SW-115: the shipped `mcp-client` archetype no longer projects the `werkmeister` MCP server into a workspace's `.mcp.json`. That file is shared by host and container sessions, and the entry's `http://host.docker.internal:3002/mcp` only resolves in a container, so every host session hit `ENOTFOUND` on start. Host sessions get the same server from the `werkmeister` plugin. The next `init-workspace` prunes the entry from existing workspaces. A container agent opts back in with an archetype of its own (`docs/users-guide.md`, edit 3).
|
|
645
|
+
- c380936: SW-120: the Open PRs tab searches the org from `overrides.github_org` (default: the workspace repo's owner) instead of the stale `sandbox` owner, and fails with a message naming that setting when no org is known rather than searching all of GitHub.
|
|
646
|
+
|
|
647
|
+
## 0.44.1
|
|
648
|
+
|
|
649
|
+
## 0.44.0
|
|
650
|
+
|
|
651
|
+
## 0.43.0
|
|
652
|
+
|
|
653
|
+
### Minor Changes
|
|
654
|
+
|
|
655
|
+
- 3bd6bf5: SW-104 ready work: `werkmeister issue ready [--project <name>] [--json]` lists a project's open work in priority order, from its board's ready column in board rank when it has a board, else the scope's open issues. Issue trackers gain `listReady(board?)`. The `next-wave` skill moves into the werkmeister plugin and reads its tickets from `issue ready` on a machine with projects.
|
|
656
|
+
- 75bdd3d: SW-108: a post-mortem comes before close-out. `close-out` (CLI, MCP, pane, VS Code) refuses while the workspace has no `postmortem.md`, unless `--skip-postmortem "<reason>"` (MCP `skip_postmortem`) gives a reason, which goes in `closeout.md`. The new werkmeister `postmortem` skill gathers the run's problems, agrees a change for each with the person, writes `postmortem.md` and turns each change into an SW issue, standards edit or memory entry. The ICM route table, propose and spike checks, `pr-rules.md` (ICM and JIT) and `w9r-pr` send agents there before close-out, and the handoff commits `postmortem.md` to the record branch.
|
|
657
|
+
|
|
658
|
+
## 0.42.0
|
|
659
|
+
|
|
660
|
+
### Minor Changes
|
|
661
|
+
|
|
662
|
+
- ba48455: SW-105: the `/workspaces` pane's cards group their buttons by what they act on, one labelled row each: `session` (⇆ Open beside, which replaces ↗ Continue in Desktop and falls back to opening the session in place when the app won't split it; Continue in Terminal / VS Code; + New session; 📌 Pin), `containers` (Up / Resume, Stop, Pause, Restart, Down, then Logs and Open feature ↗ while they run) and `work` (Refresh PRs, Open issue ↗, Close out, Abandon…). The `⋯` menu is gone. Logs opens a Terminal window following the containers' logs (new `followLogsInTerminal` in werkmeister-core). The dashboard, MCP App and VS Code are unchanged.
|
|
663
|
+
|
|
664
|
+
## 0.41.0
|
|
665
|
+
|
|
666
|
+
### Minor Changes
|
|
667
|
+
|
|
668
|
+
- 810e41b: SW-104: projects. A `projects:` map in `~/.werkmeister/config.yaml` names tracking strategies — a tracker (Jira project key, or GitHub org and optional repos), an optional board, and its own `issue_statuses` — so one machine can work on a Jira-tracked product and a GitHub-tracked one. A workspace records its project as `project:` in `workspace.yml` (`--project` on `workspace` / `init-workspace`; the create form and wizard ask after the issue, defaulting to the project whose scope holds the ref, else `default: true`), and its issue link, status moves and `pr open` / ship updates go to that project's tracker. `config --check` names invalid projects; `connect` checks each project's tracker; `config` and `setup` keep `projects:` when they rewrite the file. A machine with no projects, and a workspace with no `project:`, behave as before.
|
|
669
|
+
|
|
670
|
+
### Patch Changes
|
|
671
|
+
|
|
672
|
+
- 2290eeb: SW-107: fragment detection puts a package's `typecheck` script before `test` in its detected test command (for pnpm, `pnpm --filter api run typecheck && pnpm --filter api test`), so a workspace's listed test commands fail on a TypeScript error the way CI does. This repo's own `cli`, `werkmeister-core` and `werkmeister-vscode` test commands now run `typecheck` too.
|
|
673
|
+
|
|
674
|
+
## 0.40.0
|
|
675
|
+
|
|
676
|
+
### Minor Changes
|
|
677
|
+
|
|
678
|
+
- 4fc26a8: SW-103: a workspace can create and track more than one PR per code repo. `werkmeister pr new <worktree> --slug <slug>` cuts a fresh branch from base after that worktree's PR merged and writes a draft `pr-<worktree>-<slug>.md` with a `worktree:` field; gate, open, status, adopt, close-out and the workspaces pane handle every PR file of a worktree. `pr open` refuses a branch name an earlier merged PR of the same worktree used. Existing `pr-<repo>.md` files behave as before.
|
|
679
|
+
|
|
680
|
+
## 0.39.0
|
|
681
|
+
|
|
682
|
+
### Minor Changes
|
|
683
|
+
|
|
684
|
+
- c0d2c42: SW-102: a machine with no Kubernetes cluster can now get a baseline.
|
|
685
|
+
|
|
686
|
+
- **A new `baseline_upstream` key.** Set `baseline_upstream: "host:port"` on a service in `services.yaml`. Its `{service}` references and its generated `<SVC>_URL` / `_HOST` / `_PORT` then resolve to that address instead of `k8s-forwards`. Before, they threw an error (or the env vars were left out) when the service had no `k8s_forward_port`. The key wins over `k8s_forward_port`, and a service that declares it gets no forward. `status` reports it as `baseline`, and the fragment schema accepts the key.
|
|
687
|
+
- **Baseline containers in `w9r-infra`.** Add `infra.baseline.<service>.image` to `~/.werkmeister/config.yaml`, pinned as `<image>@sha256:<digest>`; a tag-only image is refused. Optionally add `environment`. `connect` then runs the image in `w9r-infra` as `w9r-baseline-<service>`, reachable at `baseline-<service>` on the shared network. Point the service's `baseline_upstream` at that alias.
|
|
688
|
+
- **Removing a baseline.** Its container keeps running until `down infra`.
|
|
689
|
+
|
|
690
|
+
See _Without Kubernetes_ in `docs/topology.md`.
|
|
691
|
+
|
|
692
|
+
## 0.38.1
|
|
693
|
+
|
|
694
|
+
## 0.38.0
|
|
695
|
+
|
|
696
|
+
## 0.37.1
|
|
697
|
+
|
|
698
|
+
### Patch Changes
|
|
699
|
+
|
|
700
|
+
- 5b37069: SW-100: Caddy starts on a fresh clone, a git worktree or an npm install. The generated `features.json` was file-bound inside the read-only `proxy/ui` mount, so Docker needed a gitignored `apps/cli/infra/proxy/ui/features.json` to exist as the mountpoint, and only old checkouts with a stale copy had one. It now mounts at `/srv/features.json`, and the Caddyfile's `/__werkmeister/features.json` handle serves it from there. The next `connect` / `up infra` regenerates the Caddyfile and recreates Caddy on the new mount. A leftover `apps/cli/infra/proxy/ui/features.json` in an old checkout is no longer read and can be deleted.
|
|
701
|
+
|
|
702
|
+
## 0.37.0
|
|
703
|
+
|
|
704
|
+
### Minor Changes
|
|
705
|
+
|
|
706
|
+
- fa98c10: SW-95: the shared-infra stack no longer says "sandbox". The compose project is `w9r-infra`, the containers are `w9r-caddy` and `w9r-k8s-forwards`, and Caddy's volumes are `w9r-infra_caddy-data` / `w9r-infra_caddy-config`. The dashboard's Jaeger and RedisInsight links point at `*.w9r-infra.orb.local`. On a machine still running the old `sandbox-infra` stack, the next `connect` / `up infra` takes it down first (its containers hold the names and ports 80/443 the new one needs) and then starts the renamed stack, so the proxy restarts once. `down infra` takes the old stack down too. Its volumes are kept unless you pass `--volumes`. Leftover `sandbox-infra_*` volumes can go with `docker volume rm`. An opted-in RedisInsight starts with empty settings, because its volume moves to the new project. Caddy's data is safe to lose: certs are bind-mounted from `certs_dir`.
|
|
707
|
+
|
|
708
|
+
## 0.36.2
|
|
709
|
+
|
|
710
|
+
## 0.36.1
|
|
711
|
+
|
|
712
|
+
### Patch Changes
|
|
713
|
+
|
|
714
|
+
- 67637e6: SW-97: `werkmeister migrate` no longer tells the person to run an `infra` subcommand, which doesn't exist. The refusal while shared infra is running now names `werkmeister down infra`, and the success message names `werkmeister connect`. A new test fails the build if any backticked `werkmeister …` command in migrate's messages isn't a registered CLI command.
|
|
715
|
+
|
|
716
|
+
## 0.36.0
|
|
717
|
+
|
|
718
|
+
### Minor Changes
|
|
719
|
+
|
|
720
|
+
- 7fa838e: SW-93: the Docker-side names and default domains no longer say "sandbox". New containers carry `com.werkmeister.*` labels, compose projects and per-feature networks are `werkmeister-<feature>`, features join the `werkmeister-shared` network, the built-in build caches live in `werkmeister-<key>` volumes, and the sandbox repo's setup service is `werkmeister-install`. A workspace brought up before this keeps running under its old `sandbox-<feature>` project and `com.sandbox.local.*` labels: slots, `status`, `stop`, `restart`, `logs`, the VS Code tree and the idle reaper all still find it, and its next `up` takes the old project down (keeping its node_modules volumes) before starting the new one. `down` and `close-out` remove both. Caddy and k8s-forwards join both shared networks, and `cache prune` also reaches a leftover `sandbox-<key>` volume. A fragment still naming its setup `sandbox-install` gets the renamed startup script. The default `local_domain`/`remote_domain` are now `local.werkmeister.io`/`remote.werkmeister.io`; a machine whose `config.yaml` sets them (every `setup` run writes them) keeps its own. The X-Feature cookie, the `connect` tool links and the docs-site's allowed hosts (a new `{{local_domain}}` registry placeholder) follow the configured domain instead of a hard-coded `local.sandbox.com`. SW-11 removes the old-name fallbacks.
|
|
721
|
+
- 9dfc1cc: SW-92: `werkmeister migrate` moves `~/.sandbox` to `~/.werkmeister` and repoints every file that carries the old path: `config.yaml` values, `env.sh`, the `~/.zshenv` block, `services.yaml`, the pr-watch launchd agent (or crontab line) and each workspace's compose file. It refuses while a container that mounts the old home, or any workspace container, is running. A second run changes nothing. `--workspaces` also renames each workspace's `.sandbox/` to `.werkmeister/`, and `--check` reports without writing. Registry sync, `detect` and the pre-PR gate now read `<repo>/.werkmeister/services.yaml` first and fall back to `<repo>/.sandbox/services.yaml`. `detect` writes new fragments under `.werkmeister/`. Fragments can use `{{werkmeister_home}}` for the tool's home folder. This repo's fragment moved to `.werkmeister/services.yaml` and uses it, so **pull and rebuild the CLI before the next `registry sync`**: an older build skips the moved fragment. Docs, help text and generated text now say `~/.werkmeister`, with `~/.sandbox` named only as the legacy name.
|
|
722
|
+
|
|
723
|
+
## 0.35.5
|
|
724
|
+
|
|
725
|
+
## 0.35.4
|
|
726
|
+
|
|
727
|
+
### Patch Changes
|
|
728
|
+
|
|
729
|
+
- 8dd972e: SW-90: `close-out` is safe when two runs overlap. `closeOut` takes a per-workspace lock, so a second run says `close-out of '<name>' is already running` right away. A run that finds the folder already gone, at the start or by the `closeout.md` write, says the workspace is already closed out instead of throwing ENOENT. `sandbox close-out` without `--yes` and without a terminal refuses at once ("needs --yes when not run from a terminal"). Before, it read an empty stdin and exited 130 after the prompt with no message. Any no-TTY prompt that runs out of input now says so before it exits.
|
|
730
|
+
|
|
731
|
+
## 0.35.3
|
|
732
|
+
|
|
733
|
+
### Patch Changes
|
|
734
|
+
|
|
735
|
+
- e3c5ec9: The PR gate's on-demand start no longer asks compose for a service id only the branch registry knows: when a branch has renamed a service (e.g. `sandbox-core` → `werkmeister-core`) but the machine registry, and so `compose.yml`, still uses the old id, the gate starts the compose service that owns the same path, tests through it, and notes the drift with a pointer to `werkmeister registry` (SW-87).
|
|
736
|
+
|
|
737
|
+
## 0.35.2
|
|
738
|
+
|
|
739
|
+
## 0.35.1
|
|
740
|
+
|
|
741
|
+
### Patch Changes
|
|
742
|
+
|
|
743
|
+
- 6d24ff6: The PR gate's on-demand service start no longer aborts a cold workspace with `docker compose failed (exit 130)` (SW-86). The 300 s compose timeout was firing while `up -d --wait <svc>` waited on the setup container's install and build. Compose catches the timeout's SIGTERM and exits 130, which `runAsync` took for an ordinary exit. `runAsync` now reports any kill by its own timer as a timeout. `sandbox up` and the gate's service start get a 20-minute budget, and the timeout message states the real budget. A failed start also names the state the container was left in, e.g. `Failed to start cli (left created)`.
|
|
744
|
+
|
|
745
|
+
## 0.35.0
|
|
746
|
+
|
|
747
|
+
### Minor Changes
|
|
748
|
+
|
|
749
|
+
- 8666887: A spike workspace can now open a PR for a deliverable in a code repo, such as a write-up in `docs/research/` (SW-83). A template with no propose stage keeps its PR files in its last stage's `output/` (`stages/02_findings/output/pr-<repo>.md`), so `werkmeister pr gate`, `pr open`, `pr adopt` and close-out work for a spike the same as for any workspace. Spikes now get `_shared/pr-rules.md`, and the findings stage says when to write a PR file. A spike with no PR file closes out as before. A PR file can set `skip_exercise: <reason>` to skip the gate's exercise check when there is nothing to run. With no PR file, `pr gate` / `pr open` say which file to write, where the format is, and that the person approves it. The PR guard now denies only a real `gh pr create` at the start of a statement; the phrase inside quotes, a heredoc body or a grep pattern no longer trips it.
|
|
750
|
+
|
|
751
|
+
## 0.34.0
|
|
752
|
+
|
|
753
|
+
### Minor Changes
|
|
754
|
+
|
|
755
|
+
- 8c99ce9: Remove the SW-9 plugin-rename compatibility shims (SW-23). The old plugin path plugins/sandbox-tools, the link that let `CLAUDE_CODE_PLUGIN_DIRS` installs keep loading the plugin under its old path, is gone: point the setting at `.../plugins/werkmeister`. Generated permission rules no longer allow or deny the `mcp__plugin_sandbox-tools_sandbox-workspace__` prefix, and existing workspaces drop those rules on their next refresh. `.mcp.json` generation no longer deletes a leftover `sandbox` MCP entry from before the rename: remove it by hand if it is still there. `w9r setup` still retires an old `sandbox-tools@sandbox` install. The dashboard's old sandbox-connect tool was already renamed `connect` in SW-48 (0.21.0), so no new Cowork reinstall is needed; an artifact installed before 0.21 still calls the old name and needs `werkmeister:install-workspace-dashboard` run again.
|
|
756
|
+
|
|
757
|
+
### Patch Changes
|
|
758
|
+
|
|
759
|
+
- 157e9e8: Check `{@link}` references in core's comments with TypeDoc in CI, and fix the broken `resolveSandboxBin` link in `claude_cli.ts`.
|
|
760
|
+
|
|
761
|
+
## 0.33.2
|
|
762
|
+
|
|
763
|
+
## 0.33.1
|
|
764
|
+
|
|
765
|
+
## 0.33.0
|
|
766
|
+
|
|
767
|
+
## 0.32.0
|
|
768
|
+
|
|
769
|
+
## 0.31.0
|
|
770
|
+
|
|
771
|
+
## 0.30.0
|
|
772
|
+
|
|
773
|
+
### Minor Changes
|
|
774
|
+
|
|
775
|
+
- 62e969a: Pause and Resume a workspace (SW-70). `werkmeister pause` and the `pause` MCP tool freeze the workspace's running containers (`docker pause`, `-sandbox-install` included) and hold its Claude session: while `.werkmeister/paused.json` exists, the agent-state hook ends the turn at the next tool call and blocks new prompts. `werkmeister resume` and `resume` thaw exactly those containers, starting any stopped meanwhile, release the session and, when Pause cut a working turn, tell it to continue: headless through claude-send, a Desktop session through the Workspaces pane. Every card, `status` and `check-in` show `paused`, and a paused workspace is never a check-in candidate. `up` and `restart` refuse while paused; `down` ends the pause. The card bar on a running workspace is now `■ Stop · ⏸ Pause` with `↻ Restart` in ⋯; VS Code's Stop icon is now a square.
|
|
776
|
+
|
|
777
|
+
## 0.29.0
|
|
778
|
+
|
|
779
|
+
## 0.28.1
|
|
780
|
+
|
|
781
|
+
### Patch Changes
|
|
782
|
+
|
|
783
|
+
- cd828e1: SW-66: `createWorkspace` and `initWorkspace` take `build: false` to skip the image build, so the core suite's create test no longer leaves a crash-looping `demo-backend` container and `demo` slot on the real Docker daemon.
|
|
784
|
+
|
|
785
|
+
## 0.28.0
|
|
786
|
+
|
|
787
|
+
## 0.27.0
|
|
788
|
+
|
|
789
|
+
## 0.26.1
|
|
790
|
+
|
|
791
|
+
## 0.26.0
|
|
792
|
+
|
|
793
|
+
### Minor Changes
|
|
794
|
+
|
|
795
|
+
- 443da5d: SW-12: start an intercept from the Workspaces pane when Telepresence is installed but not connected.
|
|
796
|
+
|
|
797
|
+
- **Snapshot.** `sandbox status --intercepts --json` gains `installed`: true whenever the binary is on PATH, connected or not. `available` keeps its meaning (the daemon answered). When installed but disconnected, the snapshot now carries the rows, every one `intercepted: false`.
|
|
798
|
+
- **No daemon launch on a poll.** The intercept read runs `telepresence status` first and `list -i` only against a connected daemon. Before this, every poll's `list` implicitly connected, launching a user daemon each time.
|
|
799
|
+
- **Pane.** Installed but disconnected, an interceptable service shows `Intercept`. Its confirm line adds "Connects telepresence (docker) first." The pane starts and leaves intercepts in docker mode (`connect --docker`), since it has no TTY for host mode's sudo prompt. Not installed still reads `intercepts unavailable`. A core older than this falls back to `available`.
|
|
800
|
+
- **Two daemons.** Any unqualified telepresence call can start a host user daemon beside the docker one, and telepresence then refuses unqualified commands. Docker-mode connect checks, intercepts and leaves (detach, agent uninstall, quit) now pass `--use <context>-<namespace>-cn`. The snapshot reads whichever daemon is connected.
|
|
801
|
+
- **CLI / VS Code.** `sandbox status --intercepts` prints "not connected — `sandbox intercept` connects first." plus the rows. `sandbox intercept --leave --docker-daemon` leaves through the docker daemon. The VS Code tree keeps a disconnected service interceptable, noting "telepresence not connected".
|
|
802
|
+
|
|
803
|
+
Core API:
|
|
804
|
+
|
|
805
|
+
- `InterceptState` gains `installed` and `namespace`; `InterceptsSnapshot` and `InterceptsResult` gain `installed`.
|
|
806
|
+
- New `dockerDaemonName()`; `LeaveInterceptArgs` gains `dockerDaemon`.
|
|
807
|
+
|
|
808
|
+
## 0.25.0
|
|
809
|
+
|
|
810
|
+
### Minor Changes
|
|
811
|
+
|
|
812
|
+
- d4d5e01: SW-42: shared infra only when a workspace needs it.
|
|
813
|
+
|
|
814
|
+
- **k8s-forwards.** `sandbox up` checks or restarts the shared k8s-forwards container only when a local service needs the cluster. A service needs it when it has a `k8s_deployment`, an environment value naming a `k8s-*` alias or a `{svc}` ref to a non-local service, or a non-local `depends_on`. A werkmeister-only `up` (cli, library or extension services) no longer touches it.
|
|
815
|
+
- **Caddy.** `up` starts Caddy only for routed or cluster-backed workspaces. When Caddy is already running, `up` still reloads it.
|
|
816
|
+
- **Network.** `up` and the gate's on-demand start create the `sandbox-shared` network when it is missing.
|
|
817
|
+
- **jaeger and redisinsight.** They still start only from `sandbox connect` / `sandbox up infra`.
|
|
818
|
+
- **Telepresence.** Leaving an intercept, whether with `sandbox intercept --leave`, the MCP `intercept` leave or `sandbox down`, now uninstalls the traffic agent of each workload no other intercept still uses. It quits the local daemons when no intercept remains. If the intercept list can't be read, it does neither.
|
|
819
|
+
|
|
820
|
+
Core API:
|
|
821
|
+
|
|
822
|
+
- new `needsCluster()` (config), `ensureSharedNetwork()` (caddy), and `releaseIdleTelepresence()`, `buildUninstallArgs()` and `buildQuitArgs()` (telepresence);
|
|
823
|
+
- `UpResult.caddy` gains `"skipped"`.
|
|
824
|
+
|
|
825
|
+
## 0.24.1
|
|
826
|
+
|
|
827
|
+
## 0.24.0
|
|
828
|
+
|
|
829
|
+
### Minor Changes
|
|
830
|
+
|
|
831
|
+
- 61780d5: Per-workspace container vitals (SW-41). `sandbox status`, the Workspaces pane and the VS Code tree show each workspace's CPU and memory (one stats sweep, every 30 s in the pane and the tree), when its containers started and by what (`up`, `gate`, `restart` or `person`, from `.werkmeister/starts.json`), and a crash-looping container as an error with its last log line, within one 5 s poll. The pane and the tree also show a host line: load average and the total CPU of feature containers. Core: `WorkspaceStatus` gains `started` and `crashLoops`, `sandbox status --json` gains `host` and per-feature `cpu_percent`, `mem_mib`, `started_at`, `started_by` and `crash_loops`, and new `getWorkspaceResources()`.
|
|
832
|
+
- 1ed01fe: SW-54: the dashboard and MCP App workspace card gains `↗ Continue in Desktop` and `+ New session`, the top bar gains `+ New workspace` (the MCP App list opens the create form in place), and the list gains a `☐ Focused only` filter. `claude-resume` takes `new: true` to open a fresh Terminal session.
|
|
833
|
+
|
|
834
|
+
### Patch Changes
|
|
835
|
+
|
|
836
|
+
- 138b15e: Create form (SW-51): Work type and Workflow are preselected on every GUI, labels drop the terminal hints ("blank to skip", "Enter for default") and follow docs/ui-spec.md, VS Code's New Workspace runs the wizard from the workspaces root instead of ~, and the terminal wizard requires at least one service or repo.
|
|
837
|
+
- e132f03: SW-65: the `check-in` MCP tool reports each workspace's `lifecycle` and `prs` (repo, url, status, `release`), and names the shipped or abandoned workspaces with no Claude turn working as `readyToClose`. Before reporting, it ships merged PRs a release tag now contains, with the same rule and code as `werkmeister pr status` and no `gh` call.
|
|
838
|
+
|
|
839
|
+
## 0.23.0
|
|
840
|
+
|
|
841
|
+
### Minor Changes
|
|
842
|
+
|
|
843
|
+
- f9017c7: The dashboard, the Cowork artifact and the MCP App show the handoff workspace card, rebuilt in React in `werkmeister-ui` (SW-53). Each workspace is one always-expanded card: the agent chip (solid amber only for `● Needs you`, with the last question), the work type and task title, the issue, then Containers, PRs (state badge, dashed for Draft, and CI), Workflow (stage pips or the lifecycle step) and Agent rows. The action bar shows the primary action and two secondaries; the rest are in `⋯`, which also opens Chat, Usage and Logs. Close out and `✕ Abandon…` confirm inline in the footer, and Abandon needs a reason; "Click again to confirm" is gone. The page gains a loading skeleton, a restyled top bar, a Needs-you filter and a legend. The new tokens (`tokens.css`, light and dark) replace the old palette, with Barlow and IBM Plex Mono inlined as woff2, so no font is fetched.
|
|
844
|
+
|
|
845
|
+
`listWorkspaces()` (and the `list-workspaces` tool) adds `task`, `workType` and `lifecycle` to each workspace. `RepoPr` adds `draft`, and `workspace-prs` now declares `ci`, `failingChecks` and `draft`. The `logs` tool returns its lines as `output` in its structured content too.
|
|
846
|
+
|
|
847
|
+
## 0.22.0
|
|
848
|
+
|
|
849
|
+
### Minor Changes
|
|
850
|
+
|
|
851
|
+
- 3f0542a: `stop` and `restart` for a feature, for infra and for Caddy now go through the Docker Engine API instead of `docker compose stop`/`restart`. Containers are handled in `depends_on` order: dependents stop first, and dependencies start first. `restart` no longer re-runs a finished setup container. A failure names the container and the daemon's message. `composeStop`, `composeRestart` and `stopCaddy` are now async (SW-60).
|
|
852
|
+
|
|
853
|
+
### Patch Changes
|
|
854
|
+
|
|
855
|
+
- 95598c3: `werkmeister logs`, the MCP `logs` tool and the VS Code logs pane read container logs from the Docker Engine API instead of running `docker compose logs` (SW-61). The output is the same: compose's `name | ` prefix, coloured per service, coloured only onto a terminal for the CLI and always in the pane. An unknown service prints `no such service: <name>`. A followed tail now picks up containers that start or restart after it began. The VS Code pane no longer needs `docker` on PATH, so it works from a VS Code launched from the Dock. `composeLogsStream` returns a `{ stop(), done }` handle instead of a `ChildProcess`, and loses its `env` parameter; `composeLogsArgs` is gone.
|
|
856
|
+
|
|
857
|
+
## 0.21.0
|
|
858
|
+
|
|
859
|
+
### Minor Changes
|
|
860
|
+
|
|
861
|
+
- eb040c0: `werkmeister pr gate` starts one container for a touched repo that has none running, runs the tests through it, and stops it again, pass or fail. Containers that were already running are left alone. compose.yml now also lists the checked-out repos' other services behind the `on-demand` profile, which a plain `up` skips. `stop` and `down` include that profile. The gate's baseline check covers only api/web services. A workspace no longer needs every service listed to pass the gate (SW-34).
|
|
862
|
+
- 5f85354: Agent state tells Done, Needs you and Working apart (SW-46). Every workspace now gets a `werkmeister hooks agent-state` hook on UserPromptSubmit, PreToolUse, PostToolUse, Notification, Stop and SessionEnd. The hook records `{state, at, question?}` in `.werkmeister/agent-state.json`. `getWorkspaceStatuses()` returns `claude.agent` (`working` · `needs-you` · `done` · `idle`), `claude.at` and `claude.question` in place of `claude.waiting`, and falls back to the transcript tail when a workspace has no state file yet. Existing workspaces pick up the hooks on their next `werkmeister up`. The workspaces pane and the dashboard show `◌ Working`, `● Needs you` and `✓ Done`, so a finished run no longer reads "waiting for you".
|
|
863
|
+
|
|
864
|
+
## 0.20.0
|
|
865
|
+
|
|
866
|
+
### Minor Changes
|
|
867
|
+
|
|
868
|
+
- 002e00f: Every generated workspace container gets a CPU limit (`cpus` 1.0, 2.0 for setup services such as sandbox-install), overridable per service, setup service or `extra_services` entry with `cpus:` in services.yaml. `werkmeister stats` and `werkmeister optimize` show each container's CPU limit (SW-40).
|
|
869
|
+
|
|
870
|
+
## 0.19.0
|
|
871
|
+
|
|
872
|
+
### Minor Changes
|
|
873
|
+
|
|
874
|
+
- 1e7d8a7: `werkmeister setup` offers to schedule `werkmeister pr watch` every 10 minutes (launchd on macOS, crontab elsewhere); `werkmeister pr watch --install` / `--uninstall` add or remove it. Logs go to `~/.werkmeister/pr-watch.log` (SW-14).
|
|
875
|
+
|
|
876
|
+
## 0.18.0
|
|
877
|
+
|
|
878
|
+
### Minor Changes
|
|
879
|
+
|
|
880
|
+
- 6f712a9: `getWorkspaceStatuses()` returns every workspace-card field: ICM `stage`, JIT `phase`, `autonomous`, `autoStop` and `lastActivity`, next to the existing per-service health and `model`. `getWorkspacePRs()` returns `ci` (`pass` · `running` · `fail`) and `failingChecks` per PR from `statusCheckRollup`. New exports `icmStage()` and `ciOf()`; the pane now reads the stage from core (SW-47).
|
|
881
|
+
|
|
882
|
+
### Patch Changes
|
|
883
|
+
|
|
884
|
+
- bd22bce: The Workspaces pane and the VS Code tree no longer show an intercept control on local-only services such as CLIs and libraries, which Telepresence can never intercept. `sandbox status --intercepts --json` has a new `remote_services` list, present whether or not Telepresence is available, naming the registry services that have `k8s_deployment`, `k8s_namespace` and `k8s_port`. Remote services keep every state they had, including a disabled control when `k8s_port_name` is missing.
|
|
885
|
+
- 3936ebd: `sandbox pr watch` now resumes terminal sessions that ran `sandbox pr open` from a code worktree or other subfolder. `pr open` records the session's start folder (`session_cwd`, from its transcript), and delivery looks the session up and runs `claude -p --resume` from there. It still never falls back to another session.
|
|
886
|
+
- 9420fd0: The built-in compose fallback now names feature containers `<feature>-<svc>`, as the template does, instead of `sandbox-<feature>-<svc>`. Both paths and the PR gate's container lookup share one `containerOf(feature, svc)`, so a workspace generated through the fallback has containers the gate, restart lines and allow rules can find.
|
|
887
|
+
|
|
888
|
+
## 0.17.1
|
|
889
|
+
|
|
890
|
+
## 0.17.0
|
|
891
|
+
|
|
892
|
+
### Minor Changes
|
|
893
|
+
|
|
894
|
+
- 5ba6593: The workspaces MCP server has a new `check-in` tool: the check-in skill's triage in one call. It returns every workspace's labels, container and Claude state, proposed stage gates, and whether its transcript changed since the last check-in, filtered by `skip` / `focus`, plus the `candidates` that need a courier, and marks the check-in time in `<root>/.check-in-last`. It returns plain JSON, so a check-in no longer renders the workspaces table inline the way `list-workspaces` does.
|
|
895
|
+
|
|
896
|
+
## 0.16.0
|
|
897
|
+
|
|
898
|
+
### Minor Changes
|
|
899
|
+
|
|
900
|
+
- 844ccbf: The workspace's issue now moves by itself as the work does (SW-27, SW-24). `sandbox issue start`, run by the first ICM stage and JIT's `sb-research`, moves it to In Progress. `sandbox pr open` moves it to In Review. `sandbox pr status` moves it back to In Progress when a review asks for changes or a PR closes unmerged, and to Done once the work ships; close-out is the safety net. `close-out --abandon` comments the reason, applies `abandoned` when configured, and on GitHub closes the issue as not planned. Configure the statuses with `overrides.issue_statuses: { in_progress, in_review, merged, done, abandoned }` (null skips a point); `issue_review_status` and `issue_done_status` still work. Moves are forward only, post a comment naming the trigger, and a tracker refusal is a warning that never fails a stage, PR open or close-out. On GitHub, a `status:` label replaces the issue's other `status:` labels.
|
|
901
|
+
- 70125e3: Service containers of type `cli`, `library` and `extension` now idle by default (SW-32): `sleep infinity` under `init: true`, `restart: on-failure:3`, and no always-passing healthcheck, so an idle workspace uses almost no CPU or memory and `sandbox stop` returns fast. Their `command:` watchers run with `sandbox up --dev` (or `SANDBOX_DEV=1 docker compose up -d`). A service that depends on one of them waits for `service_started`. `sandbox up` now regenerates a `compose.yml` that predates the current template or registry instead of only warning, so existing workspaces drop their watchers on their next `up`. The generated ICM reload lines say to `docker exec` into idle services rather than restart them.
|
|
902
|
+
- 39eb67a: Feature containers no longer keep running after a PR (SW-39).
|
|
903
|
+
- `werkmeister pr open` stops the workspace's feature containers once a pass opens a PR and nothing in it failed. `werkmeister pr status` does the same when it records a PR as merged or shipped. Slot and volumes are kept, so `werkmeister up` resumes.
|
|
904
|
+
- The new `werkmeister stop --idle <minutes> [feature]` stops each feature nothing has used for that many minutes. Use means a busy Claude session in the workspace (or one that became idle less than N minutes ago), a container started in that time, or a gate `docker exec`, which the gate now records in `.werkmeister/last-exec`.
|
|
905
|
+
- Every PR watch pass, from the werkmeister mod's timer or from `werkmeister pr watch`, ends with one reaper pass. Its N comes from `overrides.idle_stop_minutes`: 30 by default, `0` for off.
|
|
906
|
+
- `auto_stop` is now on by default for `autonomous` workspaces, and `auto_stop: false` still turns it off. The create form's unticked box no longer writes `auto_stop: false` into a new workspace.
|
|
907
|
+
|
|
908
|
+
### Patch Changes
|
|
909
|
+
|
|
910
|
+
- fef4b99: Docs sweep for the Werkmeister rename (SW-10). The README and `docs/` now call the tool Werkmeister, show `werkmeister` (or `w9r`) in every command, and use `~/.werkmeister` and `<workspace>/.werkmeister/`, with a note that the old `~/.sandbox` and `.sandbox/` are still read. The clone steps no longer `cd sandbox`. No behaviour change. The `sandbox` bin, `SANDBOX_*` variables, container names and a repo's `.sandbox/services.yaml` fragment keep their names.
|
|
911
|
+
- 07e7d18: The workspace router row for a merged or released PR now says to run `sandbox pr status`, then ask the person before `sandbox close-out`, instead of reading like permission to close out right away (SW-13). The ICM propose-stage contract and JIT's `sb-pr` skill use the same wording. Existing workspaces get it on their next `sandbox init-workspace`.
|
|
912
|
+
- 8b4a73d: `sandbox pr gate` (and `pr open`) now tests a branch with its own `.sandbox/services.yaml` fragment: the gated repo's `test:` commands and path mappings come from the worktree, every other repo and the `services.local.yaml` overlay as `registry sync` builds them. A PR that renames packages and updates its own test filters can pass its gate without a machine-wide override, and the shared `~/.sandbox/services.yaml` is not touched (SW-25).
|
|
913
|
+
- 65ff7f9: `werkmeister up` no longer fails when the kube context exists but its cluster is switched off. The k8s port-forwards are skipped with a warning when the API server doesn't answer, so features that need nothing from the cluster come up normally (SW-28).
|
|
914
|
+
|
|
915
|
+
## 0.15.0
|
|
916
|
+
|
|
917
|
+
### Minor Changes
|
|
918
|
+
|
|
919
|
+
- f85e6af: Rename the Claude Code plugin `sandbox-tools` to `werkmeister` and its marketplace `sandbox` to `werkmeister` (install id `werkmeister@werkmeister`). The seven `sandbox-*` skills drop the prefix (`/werkmeister:diagnose`, `:workspace`, `:feature-test`, `:status`, `:caddy-reload`, `:routing-test`, `:starting-prompt`). The plugin's MCP server `sandbox-workspace` is now `workspaces` and runs `w9r dashboard`, so its tools are `mcp__plugin_werkmeister_workspaces__*`. The headless registry server `sandbox` is now `werkmeister`, and its tools drop the prefix (`sandbox-up` is `up`, `sandbox-status` is `status`, `sandbox-workspace` is `workspace`, and so on). The dashboard's `sandbox-connect` tool keeps its name.
|
|
920
|
+
|
|
921
|
+
An existing install does not move by itself. Run `w9r setup`: it installs the new plugin, then removes `sandbox-tools@sandbox` and the `sandbox` marketplace (`w9r setup --check` flags the old one). `plugins/sandbox-tools` stays as a link to the new directory for one release, and generated permission rules allow both the old and new tool prefixes for one release. Manual steps, including Claude Desktop, are in `plugins/werkmeister/README.md`.
|
|
922
|
+
|
|
923
|
+
## 0.14.1
|
|
924
|
+
|
|
925
|
+
## 0.14.0
|
|
926
|
+
|
|
927
|
+
### Minor Changes
|
|
928
|
+
|
|
929
|
+
- 6ff0f46: The Caddy feature selector is now served at `/__werkmeister` (SW-8). `/__sandbox` keeps working as an identical, deprecated alias until SW-11 removes it. The `X-Feature` cookie and header are unchanged.
|
|
930
|
+
- 99a2303: `sandbox pr open` no longer moves the issue to a hardcoded "In Development". It moves it to `overrides.issue_review_status` from `~/.sandbox/config.yaml`, defaulting to "In Review", which matches the usual To Do → In Progress → In Review → Done workflow. On GitHub, the label of that name is added as before.
|
|
931
|
+
- 7e0a144: Werkmeister rename, config side (SW-7): the tool now reads `~/.werkmeister` and each workspace's `.werkmeister/` first, falling back to `~/.sandbox` and `.sandbox/` when the new directory is missing. Nothing is moved, so existing installs and workspaces keep working unchanged. A fresh machine gets `~/.werkmeister`, and a new workspace gets `.werkmeister/`, along with generated docs that name it. Every `SANDBOX_*` variable the tool reads now also accepts a `WERKMEISTER_*` name, which wins when both are set. The names the tool exports (env.sh, compose and container env, `.env`, the generated MCP config) stay `SANDBOX_*` for now. The VS Code extension's `sandbox.root`/`sandbox.binPath` settings are now applied as `WERKMEISTER_ROOT`/`WERKMEISTER_BIN`, so they still beat an inherited env var.
|
|
932
|
+
|
|
933
|
+
### Patch Changes
|
|
934
|
+
|
|
935
|
+
- acf3e20: The GitHub repository is now `jbbz-sandbox/werkmeister` (SW-10). The package's repository URL, the sandbox-tool issue repo in generated issue-filing guides, the plugin marketplace source written by `setup`, the schema `$id`s and the docs all point there. GitHub redirects the old `jbbz-sandbox/sandbox` URLs, so existing clones and links keep working.
|
|
936
|
+
|
|
937
|
+
## 0.13.0
|
|
938
|
+
|
|
939
|
+
### Minor Changes
|
|
940
|
+
|
|
941
|
+
- 0853f96: Narrow the `autonomous` permission rules to the workspace's own feature. `Bash(sandbox|docker|npm|gh:*)` and the whole sandbox-workspace MCP server give way to per-command rules (`sandbox up|restart|logs|test <feature>`, `docker exec|logs|restart <feature>-<svc>`, `npm ci|run|test`, read-only `gh` and MCP tools), and a recorded `permissions.deny` blocks `sandbox proxy-reload|down|remove-workspace`, `docker stop|rm`, `npm publish|link` and global installs, `gh pr merge`, and the cross-workspace MCP tools. Rerunning `init-workspace` or `sandbox up` replaces the old rules.
|
|
942
|
+
|
|
943
|
+
## 0.12.0
|
|
944
|
+
|
|
945
|
+
### Minor Changes
|
|
946
|
+
|
|
947
|
+
- e12539a: PR callbacks (#236): `sandbox pr open` records the Claude Code session that opened each PR (`session`, plus `host_session` in the desktop app), and that session is told once when the PR merges, ships or closes. Desktop sessions are reached through the app by the sandbox-tools mod, which checks every 10 minutes; terminal sessions are resumed headlessly by the new `sandbox pr watch` (run it from cron). It never merges or closes out.
|
|
948
|
+
|
|
949
|
+
## 0.11.0
|
|
950
|
+
|
|
951
|
+
### Minor Changes
|
|
952
|
+
|
|
953
|
+
- 0f2892d: Workspaces now get `compose.yml`, the Compose v2 preferred name, instead of `docker-compose.yml` (#224). An existing workspace's file is renamed on the next `sandbox init-workspace` or `sandbox up`; until then `down`, `logs` and `close-out` still read `docker-compose.yml`, so a running stack is never orphaned (Compose keys it by project `name:`, which is unchanged).
|
|
954
|
+
|
|
955
|
+
## 0.10.1
|
|
956
|
+
|
|
957
|
+
### Patch Changes
|
|
958
|
+
|
|
959
|
+
- db54e57: `sandbox up` no longer exits mid-run with Node's "unsettled top-level await" warning while waiting for restarted k8s port-forwards. The redis/postgres probe now treats a peer that hangs up without replying (Docker's port proxy before the tunnel binds) as down instead of waiting forever, so the wait either sees the forwards return and goes on to start the containers, or fails with `FORWARDS_NEVER_READY` (#207). `sandbox status --k8s` and the dashboard's forwards check share the probe and can no longer hang the same way.
|
|
960
|
+
|
|
961
|
+
## 0.10.0
|
|
962
|
+
|
|
963
|
+
### Minor Changes
|
|
964
|
+
|
|
965
|
+
- 93adb9a: Close-out is the one end-of-workspace verb (#160). `sandbox close-out <ws> --abandon "<reason>"` ends a workspace whose work won't ship: it records the reason, leaves the issue, pushes every branch and keeps them, then removes the workspace. Both outcomes list terminal Claude sessions still open in the workspace. The pane shows one end button per card: ✓ Close out once shipped, otherwise ✕ Abandon… with a reason field. After a close-out, the pane and the `close-out` MCP tool archive the workspace's desktop sessions. The `remove-workspace` MCP tool, the pane's ✕ Remove and VS Code's Remove Workspace are gone; VS Code has Close Out Workspace. `sandbox remove-workspace` still works for one release but prints a deprecation notice.
|
|
966
|
+
|
|
967
|
+
`closeOut(name, opts, deps)` now takes `{ abandon }` as its second argument, and `deps` moves to the third.
|
|
968
|
+
|
|
969
|
+
### Patch Changes
|
|
970
|
+
|
|
971
|
+
- ca13ea7: Shared-infra `docker compose` calls now pass `-p sandbox-infra`, so a caller whose env sets `COMPOSE_PROJECT_NAME` (every workspace container) no longer addresses the shared stack as its own project, which used to report k8s-forwards unhealthy and then fail restarting it. `sandbox up`/`down`/`stop`/`restart` run inside the target feature's own container now refuse with `INSIDE_OWN_CONTAINER` instead of killing themselves with exit 137 (#198).
|
|
972
|
+
- 5247bbc: An aborted stats collection now cancels the container stats requests in flight instead of waiting out a host-wide snapshot, so it stops right away however many containers are running, and the `stats-collect` aborted-run test no longer times out on a loaded host (#182).
|
|
973
|
+
|
|
974
|
+
## 0.9.1
|
|
975
|
+
|
|
976
|
+
### Patch Changes
|
|
977
|
+
|
|
978
|
+
- 4d0219c: `sandbox close-out` and `sandbox remove-workspace` now remove the workspace's per-feature Docker volumes (`docker compose down --volumes`) when they delete it, instead of leaving `<feature>-node-modules` and `<feature>-tools` behind. Plain `sandbox down` still keeps them.
|
|
979
|
+
|
|
980
|
+
## 0.9.0
|
|
981
|
+
|
|
982
|
+
### Minor Changes
|
|
983
|
+
|
|
984
|
+
- 5d23026: `sandbox pr gate`: a repo's registry `repo:` block can list `no_service_paths` (e.g. `docs`, `.changeset`, `README.md`); changes under them touch no service, so a docs-only PR no longer needs `skip_tests` for every service. The sandbox repo declares `docs`, `.changeset`, `.gitignore` and `README.md`.
|
|
985
|
+
|
|
986
|
+
### Patch Changes
|
|
987
|
+
|
|
988
|
+
- bc09b75: An ICM workspace's tracked `_shared/standards.md` no longer holds machine paths: each row reads `<standards_dir>/<module>/STANDARD.md`, with a line saying `sandbox standards check` prints that root and `sandbox setup` grants Claude Code read access to it. Opening the record on a machine with a different home dir or `overrides.standards_dir` no longer makes `sandbox skills check` report the file as drifted. Existing workspaces get the portable form on their next `sandbox skills sync` (#163).
|
|
989
|
+
- 100e350: `sandbox up` now regenerates `<workspace>/.sandbox/supergraph-config.yml` on every run, not only when the slot changes, so a stale copy is removed when the registry has no GraphQL subgraph (#120).
|
|
990
|
+
|
|
991
|
+
## 0.8.0
|
|
992
|
+
|
|
993
|
+
### Patch Changes
|
|
994
|
+
|
|
995
|
+
- 88430c7: `sandbox init-workspace` no longer adds the current template's stage dirs beside an older ICM layout. In a workspace whose `stages/` holds dirs the current template does not render (e.g. `03_verify`/`04_ship` from before the rename to `03_exercise`/`04_propose`), init now warns and leaves `CLAUDE.md`, `CONTEXT.md`, `_shared/` and `stages/` as they are, and `skills check`/`sync` leave them alone too. Services are still reconciled into `docker-compose.yml`, `.env` and `.sandbox/sandbox-context.md`. To regenerate the pipeline from the current template, move `stages/` aside and rerun (#149).
|
|
996
|
+
- e40578c: `generateSupergraphConfig` no longer creates an empty `<workspace>/.sandbox/` when the registry has no GraphQL service and there is nothing to write (#119).
|
|
997
|
+
|
|
998
|
+
## 0.7.0
|
|
999
|
+
|
|
1000
|
+
### Patch Changes
|
|
1001
|
+
|
|
1002
|
+
- 5f1c77f: `sandbox pr open` no longer fails to move a GitHub issue to "In Development" when the repo has no such label. The GitHub tracker now runs `gh label create <status> --color 1D76DB --force` before adding a status label, so the label is created when missing and left as is (same color) when present (#156).
|
|
1003
|
+
|
|
1004
|
+
## 0.6.0
|
|
1005
|
+
|
|
1006
|
+
### Patch Changes
|
|
1007
|
+
|
|
1008
|
+
- 4485440: `SANDBOX_FEATURE` is no longer read as the `feature` input. Every workspace container exports it as its own feature name, so a workspace created or inited from inside one was prefilled with that name. Pass `--feature` or set `feature:` in `workspace.yml` instead. The `sandbox-install` setup service now runs `npm ci`, so `sandbox up` no longer rewrites `package-lock.json`, and the compose-logs smoke test targets its own workspace's stack (#148).
|
|
1009
|
+
|
|
1010
|
+
## 0.5.1
|
|
1011
|
+
|
|
1012
|
+
### Patch Changes
|
|
1013
|
+
|
|
1014
|
+
- 0bb624c: `discoverRepos` and `repoDirs` now share one rule for which workspace dirs are repos: dot-prefixed dirs are skipped and symlinked dirs count. Before, an icm workspace with a dot-prefixed or symlinked repo dir reported drift in `sandbox skills check` that `sync` could never clear (#110).
|
|
1015
|
+
|
|
1016
|
+
## 0.5.0
|
|
1017
|
+
|
|
1018
|
+
## 0.4.0
|
|
1019
|
+
|
|
1020
|
+
### Minor Changes
|
|
1021
|
+
|
|
1022
|
+
- `sandbox setup` asks for the issue tracker (`overrides.issue_tracker`: `jira` or `github`). With `github`, issues are per repo and the tracker spans every repo in the services registry under `github_org`: `sandbox workspace` asks for a GitHub issue (`repo#n`, `owner/repo#n` or an issue URL; `--github-issue`) instead of a Jira ID, `CLAUDE.md` links it, and the issues guide sends app-code issues to those repos' issues. With `jira`, the issues guide sends them to the workspace ticket's Jira project. Jira links now use `jira.base_url` instead of a hardcoded site.
|
|
1023
|
+
- The `sandbox workspace` terminal wizard now renders from the same RJSF form schema as the browser form and MCP UI, so every surface shows the same fields, choices and validation. Local-service choices are listed under the repo that owns them in both the wizard and the create form. Removed `--print-prompts` and `--print-form-schema` from `workspace` and `init-workspace`; the form schema is built in-process by `getFormSchema`.
|
|
1024
|
+
- Generated workspaces include an issue-filing guide — `_shared/issues-report.md` for icm, `.sandbox/issues-report.md` for jit — routed from `CLAUDE.md`. Every issue has a type (bug | feature | task) that sets its GitHub label (`bug`, `enhancement`, `task`) or Jira issue type (Bug, Story, Task) and its body template. Bugs this workspace's changes caused are fixed in place; bugs that were already there, and features or tasks outside the workspace's task, are filed on `jbbz-sandbox/sandbox` or where the app code's issues live.
|
|
1025
|
+
|
|
1026
|
+
### Patch Changes
|
|
1027
|
+
|
|
1028
|
+
- `init-workspace` no longer treats plain directories (ICM `_shared/`, `stages/`) as repo worktrees when the workspace root is itself a git repo. Before this, it refused with "Refusing to remove worktree(s) with uncommitted changes" or tried to remove them, and it gitignored them. A directory now counts as a repo only if it has its own `.git`. Rerun `sandbox init-workspace` to clean the stale `/_shared/` and `/stages/` lines out of the managed `.gitignore` section.
|
|
1029
|
+
|
|
1030
|
+
## 0.3.0
|
|
1031
|
+
|
|
1032
|
+
## 0.2.0
|
|
1033
|
+
|
|
1034
|
+
## 0.1.3
|
|
1035
|
+
|
|
1036
|
+
### Patch Changes
|
|
1037
|
+
|
|
1038
|
+
- `.sandbox/workspace.yml` gains a `context_strategy` field (`jit` or `icm`, defaulting to `jit`) that selects which context layer `init-workspace` generates and `up` refreshes. `sandbox workspace` and `sandbox init-workspace` gain a matching `--context-strategy` flag (env: `SANDBOX_CONTEXT_STRATEGY`).
|
|
1039
|
+
- `sandbox skills`, `sandbox notes context` and `sandbox hooks gate` now dispatch through the workspace's `context_strategy` instead of assuming jit. `skills check`'s success line reads `Context layer fresh (<id>).`, and the "lags" hint now fires whenever the strategy's `check()` reports anything (previously only on drift). The MCP `sandbox-skills` tool returns `{ drift: string[] }` for `check` and `{ wrote: string[] }` for `sync`; `sandbox-hooks` now refreshes the workspace's context layer (jit: re-merge hooks; icm: remove sandbox-managed hooks) and returns `{ strategy: string }` instead of a hook file path. `sandbox-workspace`/`sandbox-init-workspace` gain a matching `context_strategy` argument.
|
|
1040
|
+
|
|
1041
|
+
## 0.1.2
|
|
1042
|
+
|
|
1043
|
+
## 0.1.1
|