@dna113p/machines 0.3.4 → 0.3.5

package/skills/machine-delegation/SKILL.md

@@ -62,6 +62,16 @@ Workers do not inherit this conversation. Supply the context needed to act:
 - Applicable constraints, permitted effects, and what should produce a blocked outcome.
 - A result location readable by the coordinator, and what evidence it must contain.
 
+State these five fields for each assignment:
+
+1. Its goal.
+2. The paths it may change.
+3. Its acceptance criteria.
+4. How it will be verified.
+5. What is forbidden.
+
+A field that cannot be filled means the unit is not scoped yet.
+
 Follow the selected Machine's input contract. MCP/Pi accept JSON-serializable
 input; CLI positional input is text, while `--input-file <file|->` accepts JSON.
 Keep task details in input rather than hardcoding each assignment into a new
@@ -103,6 +113,15 @@ polling continuously. Interpret the current snapshot:
 Preserve existing user authorization when answering Human requests, and surface a
 new decision when the request requires one. A desire to finish is not an answer.
 
+A `blocked` or parked run is a completed run to read, not a failure to retry blindly.
+Design stuck unattended runs to end with work kept and a written note of what a
+person must decide, instead of waiting in a Human state; every stop keeps both.
+Read the note, answer it or take it to the user, then start a fresh run from the
+kept work and note when resolved, or let a person continue from them.
+
+Tell an editing Agent that checks it cannot run in its environment are the
+workflow's to run, so a sandboxed Agent returns its work instead of `blocked`.
+
 To follow a run this session did not start, such as a ticket a daemon runs, use
 `machine watch <ticket-or-run-id>` where run status is published (`machine runs`
 lists what can be watched; `--all` adds runs that already ended). Watching only
@@ -126,3 +145,7 @@ verified. Read the report and inspect relevant changes or checks, then integrate
 the result into the user's task. If evidence is missing, treat that as incomplete
 work even if the run says `completed`. Report the useful result, any blocker, and
 artifact paths; a run ID alone is not the delegated deliverable.
+
+Read the reviewer's structured report: each acceptance criterion met or not,
+findings tagged by severity. Let an Operation, not the reviewer's verdict word,
+decide what happens next.

package/skills/machine-project-setup/SKILL.md

@@ -29,6 +29,19 @@ types, quality gates, recurring failure modes, and decisions needing a Human.
 Include domain-specific work, not just generic coding stages. Use only relevant,
 available project context; distinguish observed needs from speculative ones.
 
+Inventory the project's deterministic checks before designing any workflow, by
+class: lint, types, tests, build, and a live check that drives the real artifact
+(a packaging smoke, an install test, a CLI or service run against a scratch
+directory). Take them from the project's own scripts and CI, not from a fixed
+list, and run each once to learn its duration and that it leaves the checkout
+clean. Record the inventory in the map, naming the classes the project has no
+command for; a missing class is a finding for the owner, not something to
+invent. Put the resulting gate list in one project-owned place that every
+route reads (a configuration file the launchers load, not a copy in each
+launcher), so an Agent is held to the same gates a person runs. Where the
+project's checks stop at the first failure, prefer running every gate and
+reporting all failures at once: one repair round then sees them all.
+
 Inspect the existing orchestrator instructions and any `.machines/README.md`.
 Run `machine list` and `machine agents` in the actual work directories, or use
 `machine_list` / `machine_agents` with the correct workspace. Read candidate
@@ -95,11 +108,16 @@ instead of presenting same-model review as independent.
 
 ## Build the smallest useful set
 
-When the Machine Shop starter library is available, inspect its `docs/catalog.md`
-and candidate `defaults/` sources before rebuilding a common software process.
-Select or adapt relevant entries with the project-local installer; do not install
-the entire catalog by default. Its check commands, Agent bindings, and task inputs
-remain project-specific. Do not assume that library is globally installed.
+Machine Shop is an optional starter library of common software-task Machines:
+implementation, bug fixing, review, investigation, and model-free verification
+among them. Nothing here requires it, but it is a good set of defaults to try
+before rebuilding a common software process. It is a separate repository,
+`https://github.com/dna113p/machine-shop`, and is not installed with the runtime;
+when no checkout is present, tell the user about it or clone it to try it out.
+Read its `docs/catalog.md` and the candidate `defaults/` sources, then run one
+from source or copy selected entries with its project-local installer; do not
+install the entire catalog by default. Its check commands, Agent bindings, and
+task inputs remain project-specific.
 
 Map the major work to existing or missing workflows. Favor frequent delegation,
 important quality gates, and costly recurring mistakes. Reuse good definitions;

package/skills/machine-project-setup/references/project-workflows.md

@@ -19,6 +19,10 @@ For each major task type, record:
 - Expected final state, checks, result retrieval, and verification command.
 - Allowed effects, Human decisions, retry limits, and untested prerequisites.
 
+Record the gate inventory once, by class (lint, types, tests, build, live
+check): the command for each, where the one list that every route reads lives,
+and the classes the project has no command for.
+
 Mark deliberately direct, deferred, or blocked work with its reason. An existing
 setup may need no new Machine. Do not generate placeholders to fill categories.
 Include copyable invocations only for routes that exist. CLI positional input is