Skills now leave model and effort to the per-harness rules under `rules/claude/` and `rules/codex/`. Three skill sites lose their fallback model and effort wording and the Claude Code-specific "Agent tool" and `subagent_type=Explore`. Each keeps only what the sub-agent delivers (facts or a conclusion), which is what those rules select on. Without a rule, the harness default picks the model.

In 16 test runs, agents given the Claude rule picked `opus` at `xhigh` for conclusions and `medium` for facts. Agents given the Codex rule picked `gpt-6-astra` at `high` and `gpt-6-sol` at `medium`.

Two neighboring lines also changed. `improve-codebase-architecture` now applies the deletion test with the outcomes the skill defines. In `review-agent-instructions`, the bar on editing a retest no longer ends once the retest passes.

Stacked on #91. Retarget to `main` after #91 merges.

Reviewed-on: #92

@@ -0,0 +4,4 @@
+
+- `gpt-6-luna` at `high` for mechanical work with a clear output and little judgment: running a given command and reporting its output, finding every call site of a symbol, extracting fields from files, applying a specified bulk edit. The small model needs `high` even for mechanical work.
+- `gpt-6-sol` at `medium` for work that takes trial and error to finish: implementing a scoped change, getting a failing test suite green, browsing and scraping pages, collecting evidence against given criteria.
+- `gpt-6-astra` at `high` for work whose deliverable is a conclusion: reviewing work, synthesizing findings into a report, diagnosing a root cause, choosing an approach, weighing tradeoffs, designing a plan.

@@ -0,0 +6,4 @@
+- `gpt-6-sol` at `medium` for work that takes trial and error to finish: implementing a scoped change, getting a failing test suite green, browsing and scraping pages, collecting evidence against given criteria.
+- `gpt-6-astra` at `high` for work whose deliverable is a conclusion: reviewing work, synthesizing findings into a report, diagnosing a root cause, choosing an approach, weighing tradeoffs, designing a plan.
+
+Route a task that sits between two tiers to the higher one, and move a task that needs more reasoning up a tier rather than raising its effort. To set `model` and `reasoning_effort`, use `fork_turns: "none"` or a positive integer string and pass the child the task context it needs. A full-history fork inherits the parent model and effort and rejects overrides; use it when preserving that context matters more than selecting a different tier.

@@ -0,0 +6,4 @@
+- `gpt-6-sol` at `medium` for work that takes trial and error to finish: implementing a scoped change, getting a failing test suite green, browsing and scraping pages, collecting evidence against given criteria.
+- `gpt-6-astra` at `high` for work whose deliverable is a conclusion: reviewing work, synthesizing findings into a report, diagnosing a root cause, choosing an approach, weighing tradeoffs, designing a plan.
+
+Route a task that sits between two tiers to the higher one, and move a task that needs more reasoning up a tier rather than raising its effort. To set `model` and `reasoning_effort`, use `fork_turns: "none"` or a positive integer string and pass the child the task context it needs. A full-history fork inherits the parent model and effort and rejects overrides; use it when preserving that context matters more than selecting a different tier.

@@ -39,3 +39,3 @@
 Follow its loading procedure: a root `CONTEXT-MAP.md` means the repo has multiple contexts — read the touched contexts' `CONTEXT.md` files, the root `docs/adr/` (system-wide decisions), and each touched context's own `docs/adr/` — otherwise the root `CONTEXT.md` and the ADRs in the area you're touching.
 
-Then use the Agent tool with `subagent_type=Explore` — a fast executor model at moderate reasoning effort, since the walk delivers facts, not conclusions — to walk the codebase. The sub-agent starts with none of your context, so the brief must carry everything the walk needs: the friction questions below, the glossary terms the project docs define, and the [LANGUAGE.md](LANGUAGE.md) definitions of **shallow**, **seam**, and **locality**, with the instruction to report findings as `file:line` evidence in that vocabulary. Don't impose rigid heuristics — have it explore organically and note friction:
+Then spawn a read-only exploration sub-agent to walk the codebase. The walk delivers facts, not conclusions. The sub-agent starts with none of your context, so the brief must carry everything the walk needs: the friction questions below, the glossary terms the project docs define, and the [LANGUAGE.md](LANGUAGE.md) definitions of **shallow**, **seam**, and **locality**, with the instruction to report findings as `file:line` evidence in that vocabulary. Don't impose rigid heuristics — have it explore organically and note friction:

@@ -0,0 +2,4 @@
+
+When choosing a sub-agent's model explicitly, use one of three settings:
+
+- `gpt-6.1-sol` at `low` is the regular tier, and most sub-agents run on it: mechanical work with a clear output and little judgment, such as running a given command and reporting its output, finding every call site of a symbol, extracting fields from files, or applying a specified bulk edit.

@@ -0,0 +2,4 @@
+
+Launch every sub-agent at one of three settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier, and most sub-agents run on it: routine work with a clear goal, such as running a given command and reporting its output, finding every call site of a symbol, extracting fields from files, applying a specified bulk edit, or implementing a scoped change.

@@ -0,0 +6,4 @@
+- `opus` (Opus 5.5) at `medium` for work that takes trial and error to finish: getting a failing test suite green, browsing and scraping pages, collecting evidence against given criteria.
+- `opus` at `xhigh` for work whose deliverable is a conclusion: reviewing work, synthesizing findings into a report, diagnosing a root cause, choosing an approach, weighing tradeoffs, designing a plan.
+
+Stay on the regular tier unless the task clearly needs trial and error or delivers a conclusion. `opus` at `xhigh` is the ceiling: split a task that needs more, or report the limit. Set the model and effort explicitly wherever they can be set: workflow `agent()` calls and agent definition frontmatter.

@@ -47,3 +47,3 @@
 - Which parts of the codebase are untested, or hard to test through their current interface?
 
-The judgment calls stay with you: from the returned facts, decide what is shallow and apply the **deletion test** to anything you suspect — would deleting it concentrate complexity, or just move it? A "yes, concentrates" is the signal you want.
+The judgment calls stay with you: from the returned facts, decide what is shallow and apply the **deletion test** to anything you suspect. Complexity that vanishes marks a pass-through, the candidate you want; complexity that reappears across callers means the module is earning its keep.

@@ -32,3 +32,3 @@
 ## Judge independently
 
-For a clean-context review, use a fresh reviewer with the strongest available model at its deepest reasoning setting. This reviewer need not be the model under test. Give it, by path and reading order:
+For a clean-context review, use a fresh reviewer, which delivers a conclusion and need not be the model under test. Give it, by path and reading order:

@@ -0,0 +2,4 @@
+
+Launch every sub-agent at one of three settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier, and most sub-agents run on it: routine work with a clear goal, such as running a given command and reporting its output, finding every call site of a symbol, extracting fields from files, applying a specified bulk edit, or implementing a scoped change.

@@ -0,0 +6,4 @@
+- `opus` (Opus 5.5) at `medium` for work that takes trial and error to finish: getting a failing test suite green, browsing and scraping pages, collecting evidence against given criteria.
+- `opus` at `xhigh` for work whose deliverable is a conclusion: reviewing work, synthesizing findings into a report, diagnosing a root cause, choosing an approach, weighing tradeoffs, designing a plan.
+
+Stay on the regular tier unless the task clearly needs trial and error or delivers a conclusion. `opus` at `xhigh` is the ceiling: split a task that needs more, or report the limit. Set both explicitly in workflow `agent()` calls and agent definition frontmatter. A direct Agent tool call sets only the `model`; its effort comes from the agent type's definition.

@@ -32,3 +32,3 @@
 ## Judge independently
 
-For a clean-context review, use a fresh reviewer with the strongest available model at its deepest reasoning setting. This reviewer need not be the model under test. Give it, by path and reading order:
+For a clean-context review, use a fresh reviewer, which delivers a conclusion and need not be the model under test. Give it, by path and reading order:

@@ -0,0 +2,4 @@
+
+Launch every sub-agent at one of three settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier, and most sub-agents run on it: exploring and searching code, researching how a codebase works, running commands, watching for changes such as a CI run or deploy, and routine edits up to a scoped change.

@@ -0,0 +6,4 @@
+- `opus` (Opus 5.5) at `medium` for work that takes trial and error to finish, such as getting a failing test suite green or driving a multi-step browser flow.
+- `opus` at `xhigh` for work whose deliverable is a conclusion: reviews, diagnoses, designs, and decisions between approaches.
+
+Stay on the regular tier unless the task clearly needs trial and error or delivers a conclusion. `opus` at `xhigh` is the ceiling: split a task that needs more, or report the limit. Set both explicitly in workflow `agent()` calls and agent definition frontmatter. A direct Agent tool call sets only the `model`; its effort comes from the agent type's definition.

@@ -32,3 +32,3 @@
 ## Judge independently
 
-For a clean-context review, use a fresh reviewer with the strongest available model at its deepest reasoning setting. This reviewer need not be the model under test. Give it, by path and reading order:
+For a clean-context review, use a fresh reviewer, which delivers a conclusion and need not be the model under test. Give it, by path and reading order:

@@ -0,0 +2,4 @@
+
+Run every sub-agent at one of two settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and edits up to a scoped change.

@@ -0,0 +3,4 @@
+Run every sub-agent at one of two settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and edits up to a scoped change.
+- `opus` (Opus 5.5) at `xhigh` for work that needs original thought: reviews, opinions, root-cause diagnoses, designs, architecture, and choosing between approaches.

@@ -0,0 +5,4 @@
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and edits up to a scoped change.
+- `opus` (Opus 5.5) at `xhigh` for work that needs original thought: reviews, opinions, root-cause diagnoses, designs, architecture, and choosing between approaches.
+
+Use the higher tier only when the deliverable is a judgment the caller will rely on. `opus` at `xhigh` is the ceiling: split a task that needs more, or report the limit. Set both in workflow `agent()` calls and agent definition frontmatter. A direct Agent tool call sets only the `model`; to give it the tier's effort, use an agent type whose definition sets that effort.

@@ -0,0 +2,4 @@
+
+Run every sub-agent at one of two settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and edits up to a scoped change.

@@ -0,0 +3,4 @@
+Run every sub-agent at one of two settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and edits up to a scoped change.
+- `opus` (Opus 5.5) at `xhigh` for work that needs original thought: reviews, opinions, root-cause diagnoses, designs, architecture, and choosing between approaches.

@@ -0,0 +5,4 @@
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and edits up to a scoped change.
+- `opus` (Opus 5.5) at `xhigh` for work that needs original thought: reviews, opinions, root-cause diagnoses, designs, architecture, and choosing between approaches.
+
+Use the higher tier only when the deliverable is a judgment the caller will rely on. `opus` at `xhigh` is the ceiling: split a task that needs more, or report the limit. Set both in workflow `agent()` calls and agent definition frontmatter. A direct Agent tool call sets only the `model`; to give it the tier's effort, use an agent type whose definition sets that effort.

@@ -0,0 +2,4 @@
+
+Run every sub-agent at one of two settings:
+
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and implementing scoped changes.

@@ -0,0 +5,4 @@
+- `sonnet` (Sonnet 5.5) at `low` is the regular tier for most work: exploring and searching code, researching how a codebase or tool works, running commands and test suites, getting a failing test suite green, watching for changes such as a CI run or deploy, browsing and scraping pages, collecting evidence against given criteria, and implementing scoped changes.
+- `opus` (Opus 5.5) at `xhigh` is the higher tier for work that needs original thought: plans, designs, architecture, reviews, opinions, root-cause diagnoses, and choosing between approaches.
+
+Use the higher tier only when the deliverable is a judgment the caller will rely on. `opus` at `xhigh` is the ceiling: split a task that needs more, or report the limit. Set both in workflow `agent()` calls and agent definition frontmatter. A direct Agent tool call sets only the `model`; to give it the tier's effort, use an agent type whose definition sets that effort.

@@ -32,3 +32,3 @@
 ## Judge independently
 
-For a clean-context review, use a fresh reviewer with the strongest available model at its deepest reasoning setting. This reviewer need not be the model under test. Give it, by path and reading order:
+For a clean-context review, use a fresh reviewer, which delivers a conclusion and need not be the model under test. Give it, by path and reading order: