feat(pr-review): add optional max_turns input to override turn budget

.github/actions/README.md

@@ -71,13 +71,17 @@ The `pr-review` action additionally accepts:
 - `review_depth` - optional, defaults to `standard`.
   Set it to `deep` to make the semantic-analysis stage fan out relevant review dimensions
   and adversarially verify the candidates before returning one structured result.
+- `max_turns` - optional, defaults to empty.
+  Overrides the per-invocation turn budget (`--max-turns`) that is otherwise derived from the
+  model tier and review depth. Set a positive integer to pin the main analysis ceiling; the
+  retry keeps its 1.5x headroom relative to it. A non-integer or non-positive value fails the run.
 - `premortem` - optional, defaults to `auto`.
   Automatic mode runs the independent production-failure analysis for initial and high-risk
   reviews, but skips it for ordinary incremental updates.
   `on` always enables it and `off` disables it.
 
 The semantic-analysis stage runs under a turn budget: 12 for a low-risk incremental review,
-36 for a strong-tier one, and 56 for `deep`.
+44 for a strong-tier one, and 56 for `deep`. Set `max_turns` to override any of these.
 Roughly ten turns go on mandated context — six pipeline files plus repo guidance — before the
 diff is read, and a small diff inside a large file spends many more paging through it, so the
 budget tracks files to understand rather than lines changed.
@@ -91,8 +95,8 @@ with:
   github_identity_token: ${{ steps.app-token.outputs.token }}
 ```
 
-The semantic stage is bounded to 12 turns for fast incremental reviews, 24 for standard
-full or high-risk reviews, and 40 for explicit deep reviews.
+The semantic stage is bounded to 12 turns for fast incremental reviews, 44 for standard
+full or high-risk reviews, and 56 for explicit deep reviews, unless `max_turns` overrides it.
 
 ### PR review pipeline
 

.github/actions/claude-pr-review/action.yml

@@ -40,6 +40,15 @@ inputs:
     description: >
       `standard` runs one lead review. `deep` asks the lead to fan out review
       dimensions and adversarially verify them before returning structured data.
+  max_turns:
+    required: false
+    default: ""
+    description: >
+      Optional override for the per-invocation turn budget (`--max-turns`).
+      Empty (default) uses the built-in budget derived from model tier and
+      review depth (fast 12, standard 44, deep 56). Set a positive integer to
+      pin the ceiling for the main analysis pass; the retry pass keeps its 1.5x
+      headroom relative to this value.
   debug_logs:
     required: false
     default: "false"
@@ -103,6 +112,7 @@ runs:
         INCREMENTAL_MODEL: ${{ inputs.incremental_model }}
         MODEL_TIER: ${{ steps.prepare.outputs.model_tier }}
         REVIEW_DEPTH: ${{ inputs.review_depth }}
+        MAX_TURNS_OVERRIDE: ${{ inputs.max_turns }}
         RUN_PREMORTEM: ${{ steps.prepare.outputs.run_premortem }}
       run: |
         set -euo pipefail
@@ -133,6 +143,17 @@ runs:
         if [[ "$REVIEW_DEPTH" == "deep" ]]; then
           max_turns=56
         fi
+        # An explicit max_turns input overrides the tier/depth-derived budget,
+        # letting a consumer repo tune the analysis ceiling without editing this
+        # action. Validate up front so a typo fails the step loudly instead of
+        # silently disabling the turn cap.
+        if [[ -n "$MAX_TURNS_OVERRIDE" ]]; then
+          if [[ ! "$MAX_TURNS_OVERRIDE" =~ ^[1-9][0-9]*$ ]]; then
+            echo "::error::max_turns must be a positive integer, got '$MAX_TURNS_OVERRIDE'" >&2
+            exit 1
+          fi
+          max_turns="$MAX_TURNS_OVERRIDE"
+        fi
         # Exhausting the budget is deterministic: retrying with the same one
         # burns a second full model run that cannot succeed. Give the retry
         # half again as many turns so it can actually finish.