Set the repository up for agents and Han skills, and write down the coding standards #189

Merged
ethan-hann merged 3 commits from agent-setup into main 2026-10-26 20:02:47 -05:00
Owner

What changed

  • Agent setup. AGENTS.md is reorganized around a Project Discovery section that Han skills read, a section on tracked work (every issue has a Kind/ and a Priority/ label, and a milestone when it belongs to one), and the multi-agent rules: one batch branch per coordinator, implementers run only their own test classes, and the full gate runs once on the merged batch. .han/config.md sends Han output to .han/out/, which git ignores. Agent material lives in .agents/, never in docs/.
  • Gate. gate.yml skips the gate when a change touches only .agents/, .claude/rules/, or .han/, the same way it skips documentation. GateWorkflowTests checks that the Linux and Windows jobs share one skip pattern, and which paths skip and which run. AGENTS.md now requires a passing local build/gate.ps1 before a pull request opens.
  • Coding standards. Seven accepted standards in .agents/coding-standards/, drawn from the pattern most of the code follows: Api results and errors, injected side effects in Core, process argument lists, platform-specific code, view models and screen wiring, test names and structure, and comments. Claude Code finds them through .claude/rules/coding-standards/ (core, app, tests, tools). AGENTS.md, the developer guide, and architecture.md now match them.

Why

So agents work from the same rules, the Han skills find the project's docs, decisions, and standards, and a change to agent instructions alone does not spend an hour of gate time.

Verification

  • GateWorkflowTests: 21 of 21 pass. The new skip-pattern tests failed first on the four agent paths, as expected, before the workflow change.
  • dotnet format --verify-no-changes is clean for the test file.
  • Every standard's snippets and cited symbols were checked against the code, and each verification search finds only the listed violations.
  • The full local gate was not run for this pull request; CI runs both gates. The local-gate rule applies from the next pull request on.

Follow-ups

The code that breaks a standard is tracked as #178 to #188 in the Code Cleanup milestone.

## What changed - **Agent setup.** AGENTS.md is reorganized around a Project Discovery section that Han skills read, a section on tracked work (every issue has a `Kind/` and a `Priority/` label, and a milestone when it belongs to one), and the multi-agent rules: one batch branch per coordinator, implementers run only their own test classes, and the full gate runs once on the merged batch. `.han/config.md` sends Han output to `.han/out/`, which git ignores. Agent material lives in `.agents/`, never in `docs/`. - **Gate.** `gate.yml` skips the gate when a change touches only `.agents/`, `.claude/rules/`, or `.han/`, the same way it skips documentation. `GateWorkflowTests` checks that the Linux and Windows jobs share one skip pattern, and which paths skip and which run. AGENTS.md now requires a passing local `build/gate.ps1` before a pull request opens. - **Coding standards.** Seven accepted standards in `.agents/coding-standards/`, drawn from the pattern most of the code follows: Api results and errors, injected side effects in Core, process argument lists, platform-specific code, view models and screen wiring, test names and structure, and comments. Claude Code finds them through `.claude/rules/coding-standards/` (core, app, tests, tools). AGENTS.md, the developer guide, and architecture.md now match them. ## Why So agents work from the same rules, the Han skills find the project's docs, decisions, and standards, and a change to agent instructions alone does not spend an hour of gate time. ## Verification - `GateWorkflowTests`: 21 of 21 pass. The new skip-pattern tests failed first on the four agent paths, as expected, before the workflow change. - `dotnet format --verify-no-changes` is clean for the test file. - Every standard's snippets and cited symbols were checked against the code, and each verification search finds only the listed violations. - The full local gate was not run for this pull request; CI runs both gates. The local-gate rule applies from the next pull request on. ## Follow-ups The code that breaks a standard is tracked as #178 to #188 in the Code Cleanup milestone.
- Clarified agent instructions in AGENTS.md and added project discovery details.
- Updated CLAUDE.md to reference AGENTS.md and its coding standards.
- Revised developer-guide.md to include agent instructions and documentation changes.
- Added tests in GateWorkflowTests.cs to ensure correct gate behavior for documentation and agent instruction changes.
- Introduced README.md for coding standards in .agents/coding-standards.
- Created .han/.gitignore and .han/config.md for Han skill output management.
A failure on the Windows VM costs far more clock time than a local run, so
build/gate.ps1 must pass locally before a pull request is opened or pushed to.
Both CI gates are now required checks on main, so a pull request waits while
the Windows runner is offline instead of merging on a local run.

When a coordinator dispatches implementer agents, their work merges into one
batch branch with a single pull request. Implementers run only the test classes
for the files they changed; the full gate runs once, on the merged batch
branch. The fixed cap of two implementers is gone: the coordinator decides how
much runs side by side from the work and the resources it touches.
Write down the coding standards the code already follows
All checks were successful
Docs / site (pull_request) Successful in 1m20s
Gate / linux (pull_request) Successful in 16m47s
Gate / windows (pull_request) Successful in 35m21s
Docs / site (push) Successful in 1m31s
ffb67b3760
Seven accepted standards in .agents/coding-standards/, each drawn from the
pattern most of the code follows, with real examples and the code that breaks
it: Api results and errors, injected side effects in Core, starting processes
with an argument list, platform-specific code, view models and screen wiring,
test names and structure, and comments. Claude Code finds them through index
files in .claude/rules/coding-standards/, one per area (Core, App, tests,
tools).

Where AGENTS.md, the developer guide, or architecture.md said something
different, they now match the standard: the full list of words a comment never
names, cancellation (reads take a token, a job is cancelled through
IJobsApi.CancelAsync and passes its own token down, a short mutation takes
none), TimeProvider instead of IClock, and the one allowed static for a
process-lifetime OS handle (D59). AGENTS.md also says a standard wins over the
other files, and that code breaking a standard is fixed.

The code that breaks a standard is tracked as issues #178 to #188 in the Code
Cleanup milestone.
Sign in to join this conversation.
No reviewers
No milestone
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
ethan-hann/SavePoint!189
No description provided.