feat: diagnose Repository Skill Set readiness

.agents/skills/golang-testing/SKILL.md

@@ -6,7 +6,7 @@ license: MIT
 compatibility: Designed for Claude Code or similar AI coding agents, and for projects using Golang.
 metadata:
   author: samber
-  version: "1.2.3"
+  version: "1.2.5"
   openclaw:
     emoji: "🧪"
     homepage: https://github.com/samber/cc-skills-golang
@@ -57,6 +57,8 @@ This skill guides the creation of production-ready tests for Go applications. Fo
 9. Keep unit tests fast (< 1ms), use build tags for integration tests
 10. Run tests with race detection in CI
 11. Include examples as executable documentation
+12. Test files MUST be named after the source file under test, not after the function or method being tested
+13. Test functions SHOULD appear in the same order as the functions/methods they test in the source file
 
 ## Test Structure and Organization
 
@@ -70,6 +72,20 @@ package mypackage
 package mypackage_test
 ```
 
+Name the test file after the source file it tests, not after the function or method under test. Go's convention is one test file per source file (`foo.go` -> `foo_test.go`), because tools (`go test`, coverage reports, IDE "jump to test" navigation, `gotests`) and reviewers all resolve tests by source file, not by symbol. A source file usually declares several functions/methods; splitting its tests by symbol name scatters them across many files and breaks that file-to-file mapping.
+
+```
+// ✓ Good — one test file per source file
+helloworld.go       -> helloworld_test.go   // contains TestHelloWorld, TestAbcd, TestXyz, ...
+
+// ✗ Bad — test file named after the function/method instead of the source file
+helloworld.go       -> abcd_test.go         // wrong: should be helloworld_test.go
+```
+
+Exception: very large source files MAY be split into multiple `_test.go` files by concern (e.g. `foo_test.go` + `foo_edgecases_test.go`), but each split file's name MUST still be derived from the source file name, never from an individual function name. Prefer keeping a single `_test.go` file per source file even when it grows large — splitting adds navigation overhead and is rarely worth it; reach for the exception only when a single file becomes genuinely unwieldy to browse or review.
+
+Within a test file, order test functions to match the order their tested functions/methods appear in the source file. A reader (human or agent) scrolling `foo.go` alongside `foo_test.go` can then find the matching test by position instead of searching; drift between the two orderings compounds every time either file grows.
+
 ### Naming Conventions
 
 ```go
@@ -124,6 +140,34 @@ func TestCalculatePrice(t *testing.T) {
 }
 ```
 
+## Common Pitfall: Assert Scope Leaking into Subtests
+
+Never create a testify `assert`/`require` instance in the parent test function and reuse it inside `t.Run` closures. `assert.New(t)` captures the exact `*testing.T` it was built with, so if that `t` belongs to the parent, every failure raised inside the subtest gets attributed to the *parent* test in `go test` output — the failing subtest itself still reports `--- PASS`, silently hiding which case broke. This happens whether or not the subtest calls `t.Parallel()`.
+
+```go
+// WRONG -- `is` is bound to the parent's t
+func TestCalculatePrice(t *testing.T) {
+    is := assert.New(t)
+    for _, tt := range tests {
+        t.Run(tt.name, func(t *testing.T) {
+            is.Equal(tt.expected, CalculatePrice(tt.quantity, tt.unitPrice)) // misattributed on failure
+        })
+    }
+}
+
+// RIGHT -- each subtest builds its own instance from its own t
+func TestCalculatePrice(t *testing.T) {
+    for _, tt := range tests {
+        t.Run(tt.name, func(t *testing.T) {
+            is := assert.New(t)
+            is.Equal(tt.expected, CalculatePrice(tt.quantity, tt.unitPrice))
+        })
+    }
+}
+```
+
+Verify with a deliberately-broken case: if `go test -v -run TestName` shows `--- FAIL: TestName` but every `--- PASS: TestName/subtest_name` line still says PASS, the assert scope is leaking.
+
 ## Unit Tests
 
 Unit tests should be fast (< 1ms), isolated (no external dependencies), and deterministic.

.agents/skills/roundfix/SKILL.md

@@ -33,15 +33,32 @@ Task Worktrees, and QA uses its own Agent Session after Tasks settle.
 Use the Doctor Command, `roundfix doctor`, to diagnose Run readiness without
 installing dependencies, writing config, or changing files. Doctor runs the
 shared Node.js, minimum-supported acpx, effective adapter, required Agent Selection
-Profiles, and codex runtime hygiene checks and prints one line per check with
-status `ok`, `failed`, or `skipped`. Adapter Readiness requires the effective
-Codex command to prove official `@agentclientprotocol/codex-acp` lineage at
-version `1.1.4` or newer; executable presence and a matching name are not
-proof. The `profiles:` line is the selection authority: it exact-proves every
-distinct Preferred Selection and fallback through disposable ACP Sessions and
-reports affected category references plus one deterministic next action.
-Doctor has no separate legacy `agent:` or `model:` authority. Failed checks
-include `next: <action>` when Roundfix knows the remediation.
+Profiles, Repository Skill Set, and codex runtime hygiene checks and prints one
+line per check with status `ok`, `failed`, or `skipped`. Adapter Readiness
+requires the effective Codex command to prove official
+`@agentclientprotocol/codex-acp` lineage at version `1.1.4` or newer;
+executable presence and a matching name are not proof. The `profiles:` line is
+the selection authority: it exact-proves every distinct Preferred Selection
+and fallback through disposable ACP Sessions and reports affected category
+references plus one deterministic next action. Doctor has no separate legacy
+`agent:` or `model:` authority. Failed checks include `next: <action>` when
+Roundfix knows the remediation.
+
+The blocking `skills:` line runs after, and independently from, `profiles:`.
+The running binary's embedded bundle is authoritative for Roundfix-owned
+skills; each required external skill must match its `computedHash` in
+`skills-lock.json`. Outside a Git repository, Doctor does not inspect the
+Repository Skill Set and prints
+`skills: failed (Repository Skill Set readiness requires a Git repository; next: run roundfix doctor from a Git repository)`.
+Surface a failed `skills:` line and its printed `next:` remediation before work
+continues. Owned failures print
+`roundfix skills install --target project`; external failures print
+`bunx skills experimental_install && bunx skills update -p -y`; mixed failures
+print
+`roundfix skills install --target project && bunx skills experimental_install && bunx skills update -p -y`.
+Doctor is diagnosis-only: it never runs these commands, accesses the network,
+installs or updates skills, or writes repository state. Apply remediation only
+after explicit workflow authorization, then rerun Doctor.
 On macOS, the codex hygiene check resolves `CODEX_PATH` first and then `codex`
 on `PATH`, inspects the `com.apple.quarantine` attribute (the real XProtect
 trigger), and verifies the binary's code signature (not `spctl --assess`, which
@@ -112,20 +129,22 @@ unauthorized target. A rejected Sol/high proof never becomes an offer to use
 model-managed reasoning.
 
 Use `roundfix doctor` when you only need a read-only readiness report. It runs
-the Node.js, minimum-supported acpx, effective adapter, required Agent Selection Profile,
-and codex runtime hygiene checks and exits nonzero if any check fails. Adapter
-failures name the effective command, package classification, and official
-install action. Profile failure names the exact runtime/model/reasoning tuple,
-every affected category, bounded adapter evidence, and the next
-`roundfix profiles configure` or `roundfix profiles validate` action. A
-rejected explicit `high` does not recommend model-managed reasoning. The
-command has no flags and mutates nothing.
+the Node.js, minimum-supported acpx, effective adapter, required Agent Selection
+Profiles, Repository Skill Set, and codex runtime hygiene checks and exits
+nonzero if any check fails. Adapter failures name the effective command,
+package classification, and official install action. Profile failure names the
+exact runtime/model/reasoning tuple, every affected category, bounded adapter
+evidence, and the next `roundfix profiles configure` or
+`roundfix profiles validate` action. A rejected explicit `high` does not
+recommend model-managed reasoning. The command has no flags and mutates
+nothing.
 
 ```text
 node: ok
 acpx: ok
 adapter: ok (npx -y @agentclientprotocol/codex-acp@1.1.4; package=@agentclientprotocol/codex-acp; version=1.1.4)
 profiles: ok (3 distinct tuples; 10 category references)
+skills: ok (39 required: 14 Roundfix-owned, 25 external)
 codex: ok
 ```
 

.coderabbit.yaml

@@ -0,0 +1,90 @@
+language: "en"
+early_access: true
+
+knowledge_base:
+  code_guidelines:
+    enabled: true
+    filePatterns:
+      - "AGENTS.md"
+      - "**/AGENTS.md"
+      - "CONTEXT.md"
+      - "docs/agents/**/*.md"
+      - "docs/adr/**/*.md"
+      - ".agents/skills/**/*.md"
+
+reviews:
+  request_changes_workflow: true
+  high_level_summary: true
+  poem: false
+  review_status: true
+  collapse_walkthrough: false
+  sequence_diagrams: true
+  profile: "assertive"
+  related_prs: false
+  tools:
+    golangci-lint:
+      enabled: true
+  path_filters:
+    # ----- Generated mirrors and local agent state -----
+    - "!.claude/**"
+    - "!skills/**"
+
+    # ----- IDE / editor / OS noise -----
+    - "!.vscode/**"
+    - "!.idea/**"
+    - "!**/.DS_Store"
+    - "!**/Thumbs.db"
+    - "!**/*.swp"
+    - "!**/*.sw?"
+
+    # ----- Build, release, log, and runtime artifacts -----
+    - "!bin/**"
+    - "!dist/**"
+    - "!tmp/**"
+    - "!**/.cache/**"
+    - "!**/logs/**"
+    - "!**/*.log"
+    - "!**/.tmp/**"
+    - "!**/*.db"
+    - "!**/*.db-shm"
+    - "!**/*.db-wal"
+    - "!**/*.digest"
+    - "!**/*.sha256"
+    - "!**/*.png"
+    - "!**/*.jpeg"
+    - "!**/*.pdf"
+
+    # ----- Secrets / env -----
+    - "!**/.env"
+    - "!**/.env.local"
+    - "!**/.env.*.local"
+    - "!**/*.local"
+    - "!**/*.pem"
+
+    # ----- Roundfix source, tests, contracts, docs, and configuration -----
+    - "**/*.go"
+    - "**/*.golden"
+    - "**/*.txt"
+    - "**/*.stdout"
+    - "**/*.stderr"
+    - "**/*.json"
+    - "**/*.jsonc"
+    - "**/*.yml"
+    - "**/*.yaml"
+    - "**/*.toml"
+    - ".github/**/*"
+    - ".agents/skills/**/*"
+    - ".coderabbit.yaml"
+    - ".roundfixrc.yml"
+    - ".gitignore"
+    - "AGENTS.md"
+    - "CHANGELOG.md"
+    - "CONTEXT.md"
+    - "LICENSE"
+    - "README.md"
+    - "go.mod"
+    - "go.sum"
+    - "skills-lock.json"
+    - "cog.toml"
+    - "Makefile"
+    - "**/Makefile"

.roundfixrc.yml

@@ -17,7 +17,7 @@ profiles:
       reasoning_effort: high
     fallbacks:
       - runtime: codex
-        model: gpt-5.5
+        model: gpt-5.6-terra
         reasoning_effort: xhigh
   backend:
     preferred:
@@ -26,7 +26,7 @@ profiles:
       reasoning_effort: high
     fallbacks:
       - runtime: codex
-        model: gpt-5.5
+        model: gpt-5.6-terra
         reasoning_effort: xhigh
   frontend:
     preferred:
@@ -44,7 +44,7 @@ profiles:
       reasoning_effort: high
     fallbacks:
       - runtime: codex
-        model: gpt-5.5
+        model: gpt-5.6-terra
         reasoning_effort: xhigh
   review:
     preferred:
@@ -53,7 +53,7 @@ profiles:
       reasoning_effort: high
     fallbacks:
       - runtime: codex
-        model: gpt-5.5
+        model: gpt-5.6-terra
         reasoning_effort: xhigh
 
 review_source:

CONTEXT.md

@@ -367,7 +367,7 @@ The support command that produces a Release Plan without editing release files,
 _Avoid_: Release Command, publish command, cut-release command
 
 **Doctor Command**:
-The support command that diagnoses a repository and machine's readiness for Roundfix Runs — Repository Skill Set, minimum-supported acpx, Adapter Readiness, Agent Selection Profile Readiness, and codex runtime hygiene — reporting the detected acpx version against the minimum, each check with a next action, and mutating nothing. Distinct from the Setup Command, which prepares the machine.
+The support command that diagnoses a repository and machine's readiness for Roundfix Runs — minimum-supported acpx, Adapter Readiness, Agent Selection Profile Readiness, Repository Skill Set, and codex runtime hygiene. It reports the detected acpx version against the minimum, gives each check a next action, and mutates nothing. It evaluates Repository Skill Set readiness after, and independently from, Agent Selection Profile Readiness; unlike the Doctor Command, the Setup Command prepares the machine.
 _Avoid_: Health check run, setup run, environment wizard
 
 **Archive Command**:
@@ -503,7 +503,7 @@ A shipped agent skill that teaches an external Agent how to start Roundfix or ho
 _Avoid_: Runtime, Review Source, plugin
 
 **Repository Skill Set**:
-The complete set of Roundfix-owned and externally managed Agent Skills a repository declares for its workflow, with every installed artifact matching its authoritative local source.
+The complete set of Roundfix-owned and externally managed Agent Skills required by a repository workflow. Every required installed artifact matches its authoritative local source; unrelated extras do not join the set.
 _Avoid_: Installed skills, plugin set, global skill set
 
 **Interactive Input**:

README.md

@@ -30,7 +30,7 @@ Then make the machine Run-ready and check it:
 
 ```bash
 roundfix setup     # proves adapters and generated profiles before writes
-roundfix doctor    # read-only adapter and profile readiness; mutates nothing
+roundfix doctor    # read-only profile and Repository Skill Set readiness
 ```
 
 ## Requirements
@@ -43,10 +43,32 @@ roundfix doctor    # read-only adapter and profile readiness; mutates nothing
 - Building from source additionally needs Go 1.26+ and `make`.
 
 `roundfix doctor` diagnoses all of it — including official Codex adapter
-identity, exact Agent Selection Profile proof, and macOS codex hygiene — with
-one line per check and a `next:` action on failures. Use `roundfix setup` to
-provision the minimum supported acpx or official adapter, or migrate a stale
-legacy override after authorization.
+identity, exact Agent Selection Profile proof, Repository Skill Set readiness,
+and macOS codex hygiene — with one line per check and a `next:` action on
+failures. The independent Repository Skill Set result follows `profiles:`:
+
+```text
+profiles: ok (3 distinct tuples; 10 category references)
+skills: ok (39 required: 14 Roundfix-owned, 25 external)
+```
+
+The running binary's embedded bundle is authoritative for Roundfix-owned
+skills, including the Roundfix Skill. Each required external skill must match
+its `computedHash` in the repository's `skills-lock.json`. Missing, outdated,
+or invalid required skill state prints `skills: failed (...)`, still runs the
+other Doctor checks, and makes Doctor exit `1`.
+
+Doctor checks these authorities locally, without network access or writes. It
+ignores unrelated extra skill directories and lock entries, never deletes or
+updates skills automatically, and only prints the applicable command:
+
+```bash
+roundfix skills install --target project
+bunx skills experimental_install && bunx skills update -p -y
+```
+
+Use `roundfix setup` to provision the minimum supported acpx or official
+adapter, or migrate a stale legacy override after authorization.
 
 ## Quickstart