feat(code-review)!: one skill, driving the engine instead of repeating it

CONTEXT.md

@@ -147,6 +147,13 @@ _Avoid_: run, invocation, job
 A named set of **Reviewer**s, with how many instances of each to run and whether findings are
 validated. One panel runs per review. Which one is a **Context**'s default, possibly raised by
 an **Escalation**.
+
+It may also name the **Judge** and the **Validator** that answer for it, instead of the ones the
+**Review manifest** declares. A panel is how one context's reviewers are chosen, so it is where
+the runs that reconcile them belong: the judge runs in every review, and without this a repo
+reviewing locally with one **Provider** and its pull requests with another could say so for its
+reviewers and not for its judge. Both are overrides — a panel naming neither uses the
+manifest's, so declaring them on one panel is never the price of declaring them on all.
 _Avoid_: profile, preset, tier
 
 **Context**:
@@ -176,6 +183,20 @@ detecting one is language knowledge that has to be tested somewhere other than a
 YAML. A repo names paths instead.
 _Avoid_: heuristic, marker, flag
 
+**Exclusion**:
+A changed file no **Reviewer** is shown, and which counts toward nothing a rule measures. Most
+are mechanical — the file changed, but nobody wrote the change: a lockfile, a vendored tree, a
+generated file, a binary, a pure rename, a symlink. That vocabulary is closed and ships with
+the binary, for the reason **Signal**'s is.
+
+A repo adds its own as globs, for the one thing detection cannot reach: source a person wrote
+that is not worth a review's budget. It only ever removes files, which is the opposite
+direction from an **Escalation**, so it is declared in the **Review manifest** rather than
+given on a command line — committed where anyone can read it, and read from the base ref so a
+branch cannot exclude itself. A repo's own reason is reported ahead of a mechanical one,
+because only it points at a line somebody can edit.
+_Avoid_: ignore, skip, filter, exemption
+
 **Finding**:
 One issue a **Reviewer** reports: a file, a line range, a severity, and a body. The unit
 **Judge**ment is applied to and the unit that becomes an inline comment.
@@ -378,14 +399,21 @@ _Avoid_: summary, header, footer, marker (bare)
 
 **Review manifest**:
 `.agents/code-review/manifest.yaml`: the single declaration of **Reviewer**s, **Panel**s and the
-prompt bodies they use. Read by both engines — the in-session skill and `agtk code-review` — so
-there is one roster and not two.
+prompt bodies they use. Read by `agtk code-review` and by nothing else. A roster a skill also
+carried would be a second one, and the two would disagree the first time either changed.
 _Avoid_: panels.json, roster file, review config
 
 ## Flagged ambiguities
 **"Marker"** — the bare noun is a **Signal** synonym to avoid; the HTML comment that carries a
 **Fingerprint** is a **Fingerprint marker**, always both words.
 
+**"Exclusion" vs "Suppression"** — both withhold, and they withhold different things at
+different ends of a run. An **Exclusion** is about a *file*, decided before any reviewer runs:
+the file is never shown, so no **Finding** about it exists. A **Suppression** is about a
+*finding* that was made, withheld from a **Review** because a **Comment thread** already
+carries its **Fingerprint**. An excluded file produces nothing to suppress, and a suppressed
+finding came from a file that was reviewed.
+
 **"False positive" vs "Suppression"** — both withhold something, and they are opposite acts.
 **Suppression** is `agtk`'s and mechanical: a **Finding** is not posted again because a thread
 already carries it. A **False positive** is a person's and is about the claim itself: the
@@ -396,9 +424,9 @@ finding is posted, stays posted, and is declared not to be a defect. Suppression
 has not yet passed a **Validator**. The memory sense owns the bare noun; in review, say
 "candidate finding" and never "candidate" alone.
 
-**"Panel"** — `deep-code-review` used it for a per-stack group of reviewers *within* one run,
-so a polyglot change had several. A **Panel** here is the entire roster for a run — one runs,
-named by a **Context**'s default and possibly raised by an **Escalation**. The per-stack sense
+**"Panel"** — reads as a per-stack group of reviewers *within* one run, so that a polyglot
+change would have several. A **Panel** is the entire roster for a run, and exactly one runs:
+the one a **Context** defaults to, possibly raised by an **Escalation**. The per-stack sense
 has no name because per-stack partitioning is not built.
 
 **"Review"** — the activity and the artifact. **Review** is the artifact posted to the PR; say

definitions/CONFIG-SCHEMA.md

@@ -97,6 +97,7 @@ The location is fixed rather than configurable: the manifest is configuration, a
 | `escalate` | `[]Escalation` | no | Rules that raise the panel above a context's default. Every rule is evaluated and the highest target wins, so their order carries no meaning. |
 | `approval` | `Approval` | no | What approving a reviewed head requires of a finding's severity. Absent means the default floor, AMBER. |
 | `conventions` | `[]string` | no | Documents holding this repo's own written rules, as paths from the repo root, read at the base ref and injected raw into every reviewer's prompt. Replaces the default list rather than adding to it. Absent means the defaults: CLAUDE.md, AGENTS.md, .claude/CLAUDE.md, CONTEXT.md, CONTRIBUTING.md, docs/ARCHITECTURE.md, docs/CODE_STANDARDS.md. |
+| `exclude` | `[]string` | no | Paths this repo does not want reviewed, as globs (** spans path segments, * and ? stay within one). Added to the built-in exclusions — lockfiles, vendored trees and generated files are already excluded and need no entry. An excluded file is reported with its reason and reaches no reviewer. |
 
 ### `reviewers` entry, `judge`, `validator` (`Runner`)
 
@@ -116,6 +117,8 @@ One configured model invocation. It says which CLI, which model and which prompt
 | `reviewers` | `[]string` | **yes** | Names from the manifest's reviewers map. A panel that names one this manifest does not declare cannot staff itself, and is refused. |
 | `quorum` | `int` | no | How many independent instances of each reviewer to run. Agreement between them is the confidence signal. Defaults to 1. |
 | `validate` | `bool` | no | Whether findings are put to the validator. Unset leaves it to the context, and a context that posts validates regardless: a false finding on a PR is published and blocks approval. |
+| `judge` | `Runner` | no | Judge for reviews this panel produces, instead of the manifest's. Unset uses the manifest's. |
+| `validator` | `Runner` | no | Validator for reviews this panel produces, instead of the manifest's. Unset uses the manifest's. |
 
 ### `defaults`
 

definitions/settings/skill-permissions.yaml

@@ -1,5 +1,5 @@
 name: skill-permissions
-description: Pre-approve the file and shell operations the bundled skills/agents need (memory-explorer store reads and candidate writes, backlog reporting, Serena memories, deep-code-review scripts, pr-code-review read-only gh calls) so they run without permission prompts.
+description: Pre-approve the file and shell operations the bundled skills/agents need (memory-explorer store reads and candidate writes, backlog reporting, Serena memories, panel-code-review's model-free code-review commands, read-only gh calls) so they run without permission prompts.
 platforms: [claude]
 value:
   permissions:
@@ -15,10 +15,18 @@ value:
       # pre-approval, so the only cost of a miss is a prompt, never a block.
       - "Read(**/.agents/memory/INDEX.md)"
       - "Write(**/.agents/memory/candidates/**)"
-      - "Bash(git checkout *)"
-      - "Bash(*/skills/deep-code-review/scripts/detect-parent.sh*)"
-      - "Bash(*/skills/deep-code-review/scripts/capture-diff.sh*)"
-      - "Bash(*/skills/deep-code-review/scripts/list-changed.sh*)"
-      - "Bash(*/skills/deep-code-review/scripts/find-convention-docs.sh*)"
+      # The code-review subcommands that start no model. `run` is absent
+      # because it costs money, and `approve` because a person types that one.
+      #
+      # A prefix rule cannot exclude one flag, so `explain*` grants
+      # `explain --pr` too — which reads the pull request over the API and
+      # fetches its head, under this machine's App registration. That is
+      # deliberate rather than overlooked: it is the command the skill runs on
+      # every pull-request review, it starts no model and spends nothing, and
+      # the expensive step it precedes is `run --pr`, which is not pre-approved
+      # and asks. A grant whose comment claimed these touch no network would be
+      # the real hazard.
+      - "Bash(agtk code-review panels*)"
+      - "Bash(agtk code-review explain*)"
+      - "Bash(agtk code-review signals*)"
       - "Bash(gh pr view *)"
-      - "Bash(gh pr diff *)"