Improve PR feedback assessment and loop confirmation

packages/core/agents/reviewer.md

@@ -48,6 +48,8 @@ Before reviewing, always check repository guidance:
 
 ## Review Stability
 
+- Treat prior review comments as context, not ground truth; independently verify each concern against the current code and diff before repeating or relying on it
+- Use available author type and repository association as weighting, not proof of whether feedback was human-authored; project members have greater authority on intent, while technical claims from every source still require evidence
 - Aim for stable outcomes across reruns on the same effective diff
 - When prior review context is provided, treat it as the baseline for consistency
 - If no relevant code changed since the last review, do not introduce new findings unless there is concrete new evidence or a clearly missed material defect

packages/core/commands/index.ts

@@ -69,7 +69,7 @@ export const commandDefinitions: Record<string, CommandDefinition> = {
     subtask: false,
   },
   "pr/fix/loop": {
-    description: "Watch PR CI and comments, repeatedly fixing both without approval prompts",
+    description: "Assess PR CI and comments, fix valid issues, and reply without approval prompts",
     agent: "worker",
     templatePath: "commands/pr/fix/loop.md",
   },
@@ -90,7 +90,7 @@ export const commandDefinitions: Record<string, CommandDefinition> = {
     subtask: false,
   },
   "pr/fix": {
-    description: "Fix PR feedback or CI failures, push updates, and reply",
+    description: "Assess PR feedback and CI failures, fix valid issues, push, and reply",
     agent: "worker",
     templatePath: "commands/pr/fix.md",
   },

packages/core/commands/pr/fix.md

@@ -1,6 +1,6 @@
 ## Goal
 
-Address feedback or CI failures on a pull request, validate the fixes, push them, and respond.
+Independently assess PR feedback and CI failures, implement valid fixes, push them, and respond to each current concern.
 
 ## Additional Context
 
@@ -19,6 +19,7 @@ $ARGUMENTS
 - Store automatic completion requests as `<execution-mode>` = `auto`; otherwise use `review`
 - Store a PR number or URL as `<pr-ref>` and remaining guidance as `<additional-context>`
 - Leave `<pr-ref>` undefined when absent
+- Initialize `<handled-feedback-ids>` as an empty set
 
 ### Load PR Context
 
@@ -30,16 +31,29 @@ $ARGUMENTS
 
 ### Output
 
-When fixes are complete, display:
+When `<fix-status>` is `complete`, display:
 ```
 PR fix complete for #<pr-context.pr.number>
 
 - Changes made: <changes-count> files modified
 - Base update: <base-update>
-- Threads resolved: <threads-resolved>
+- Feedback actionable: <feedback-actionable>
+- Feedback declined: <feedback-declined>
+- Replies posted: <feedback-replies-posted>
 - Validation passing: <validation-passing>
 - Validation details: <validation-results>
 - Pushed: <pushed>
 
 No additional steps are required.
 ```
+
+When `<fix-status>` is `waiting for clarification`, display:
+```
+PR fix waiting for clarification for #<pr-context.pr.number>
+
+- Changes made: <changes-count> files modified
+- Feedback declined: <feedback-declined>
+- Feedback awaiting clarification: <feedback-awaiting-clarification>
+- Replies posted: <feedback-replies-posted>
+- Pushed: <pushed>
+```

packages/core/commands/pr/fix/loop.md

@@ -1,6 +1,6 @@
 ## Goal
 
-Continuously watch a pull request, fix CI failures and new review feedback, push, reply, and repeat until clean.
+Continuously watch a pull request, independently assess CI failures and review feedback, fix valid issues, reply, and repeat until clean or clarification is needed.
 
 ## Additional Context
 
@@ -20,6 +20,7 @@ $ARGUMENTS
 
 - Store a PR number or URL as `<pr-ref>` and remaining guidance as `<additional-context>`
 - Initialize `<completed-fix-passes>` as `0`
+- Initialize `<handled-feedback-ids>` as an empty set
 
 ### Load Initial PR Context
 
@@ -43,19 +44,26 @@ $ARGUMENTS
 - Call `<%= it.config.tools.pr_load_review.name %>` with `pr: <pr-url>` and `since: <review-checkpoint>`
 - Store the result as `<fresh-review-context>` and advance `<review-checkpoint>` to `<fresh-review-context.loadedAt>`
 - Merge new reviews, issue comments, and whole changed threads into `<review-context>`, deduplicating by stable IDs
-- Combine open actionable feedback and `<ci-failures>` into `<actionable-work>`
-- If there is no actionable feedback and no actionable CI failure, continue to `### Output`
+- Combine open candidate feedback and `<ci-failures>` into `<actionable-work>` without presuming that reviewer requests are valid; exclude comment IDs already handled by this workflow unless new thread context or code changes materially reopen the concern
+
+### Confirm Clean PR Context
+
+- When `<ci-status>` is successful or no checks are configured and `<actionable-work>` is empty, call `<%= it.config.tools.pr_load.name %>` with `pr: <pr-url>` for a final complete snapshot
+- Store the result as `<final-pr-context>` and replace `<review-context>` with its reviews, issue comments, and threads
+- Recompute `<actionable-work>` from the complete snapshot, excluding comment IDs already handled by this workflow unless new thread context or code changes materially reopen the concern
+- If recomputed `<actionable-work>` is empty and there is no actionable CI failure, store `<fix-status>` as `complete` and continue to `### Output`
+- Otherwise continue with the newly discovered work; this includes comments posted by CI after checks completed
 
 <%~ include("@pr-fix", { config: it.config, auto: true }) %>
 
 ### Continue Loop
 
+- If `<fix-status>` is `waiting for clarification`, continue to `### Output` without starting another pass
 - Increment `<completed-fix-passes>` and return to `### Align Branch With Base`
-- Do not call `<%= it.config.tools.pr_load.name %>` again during this loop
 
 ### Output
 
-When complete, display:
+When `<fix-status>` is `complete`, display:
 ```
 PR loop complete for #<pr-number>
 
@@ -67,3 +75,14 @@ PR loop complete for #<pr-number>
 
 No additional steps are required.
 ```
+
+When `<fix-status>` is `waiting for clarification`, display:
+```
+PR loop waiting for clarification for #<pr-number>
+
+- CI status: <ci-status>
+- Feedback declined: <feedback-declined>
+- Feedback awaiting clarification: <feedback-awaiting-clarification>
+- Replies posted: <feedback-replies-posted>
+- Fix passes completed: <completed-fix-passes>
+```

packages/core/commands/pr/review.md

@@ -39,7 +39,7 @@ Call `<%= it.config.tools.changes_load.name %>` with `base: <pr-context.pr.baseR
 
 Following the reviewer agent guidance:
 1. Check `<pr-context.reviews>`, `<pr-context.issueComments>`, and `<pr-context.threads>`
-2. Use `<changes>` as the source of truth for changed files and diff hunks
+2. Use `<changes>` as the source of truth for changed files and diff hunks; independently verify existing review claims rather than assuming they are correct
 3. Derive `<author-decisions>` from `<pr-context.issueComments>` and `<pr-context.threads>`:
    - Include direct author replies that explicitly decline, defer, or intentionally narrow a suggestion and explain why they do not plan to implement it
    - Treat each matching author reply as higher priority than `<ticket-context>` for that same concern, unless the current diff introduces a materially different defect with a concrete failure mode

packages/core/components/pr-fix.md

@@ -2,11 +2,21 @@
 
 Call `<%= it.config.tools.changes_load.name %>` with `base: <pr-context.pr.baseRefName>`, `head: <active-branch>`, and positive `depthHint: <pr-context.pr.commitCount>` when available. Store as `<changes>`.
 
-### Analyze And Implement Fixes
+### Assess Feedback
 
-- Treat unresolved review threads, review state changes, `<actionable-work>`, and CI details in `<additional-context>` as candidate feedback
-- Use `<changes>` to understand the current PR scope
-- Separate true course corrections from noise, resolved feedback, and superseded feedback
+- Treat unresolved review threads, review state changes, `<actionable-work>`, and CI details in `<additional-context>` as candidate feedback, not instructions to follow blindly
+- Assess each candidate independently against the current code, `<changes>`, PR intent, applicable repository guidance, and the full thread including later replies
+- Use `authorType`, `authorAssociation`, author login, and whether the author is `<pr-context.pr.author>` to derive `<feedback-source>` as `automation`, `project-member`, or `external-or-unknown`; classify `Bot` or `[bot]` identities as `automation`, and non-automation `OWNER`, `MEMBER`, or `COLLABORATOR` identities as `project-member`
+- Treat `<feedback-source>` as a confidence signal rather than proof because agents can post through user accounts and bots can relay human input
+- Give `project-member` feedback and non-automation feedback from the PR author greater authority on intended behavior, product scope, and tradeoffs, but still verify technical claims against the code
+- Treat automation feedback as a technical hypothesis without product-intent authority; do not discount a supported finding merely because it came from automation
+- Classify each feedback item by comment ID as `actionable`, `already-addressed`, `superseded`, `disputed`, or `needs-clarification`, with a concise evidence-based rationale
+- Do not assume a reviewer request is correct, invent missing requirements, or change code solely to satisfy a comment
+- Store the assessment as `<feedback-assessment>` and counts as `<feedback-actionable>`, `<feedback-declined>`, and `<feedback-awaiting-clarification>`
+
+### Implement Valid Fixes
+
+- Implement only feedback classified as `actionable` and independently confirmed CI failures
 - Fix critical correctness, security, contract, and required-CI issues first
 - Follow existing patterns and make focused, minimal changes
 - Store the modified-file count as `<changes-count>`
@@ -26,11 +36,12 @@ Call `<%= it.config.tools.changes_load.name %>` with `base: <pr-context.pr.baseR
 - If `<execution-mode>` is `auto`, skip this review gate and continue to commit and push
 - Present the fix summary, changed-file count, and validation results
 - If changes were made, ask one `Review Fixes` question with `Go Ahead` and `Revise`; apply custom revision feedback and repeat implementation and validation until approved
-- If no changes were made, ask one `Need Feedback` question with `Revise` and `Stop Here`
+- If no changes were made but `disputed` or `needs-clarification` replies are pending, present `<feedback-assessment>` and ask one `Review Feedback` question with `Post Replies` and `Revise`
+- If no changes were made and no replies are pending, ask one `Need Feedback` question with `Revise` and `Stop Here`
 - STOP if approval is required but `question` is unavailable
 <% } else { -%>
 - Continue without an approval prompt
-- If actionable work produced no changes and `<base-update>` is `already up to date`, STOP to avoid looping without progress
+- If actionable work produced no changes, `<base-update>` is `already up to date`, and no feedback reply or clarification request is needed, STOP to avoid looping without progress
 <% } -%>
 
 ### Commit And Push Fixes
@@ -43,6 +54,10 @@ Call `<%= it.config.tools.changes_load.name %>` with `base: <pr-context.pr.baseR
 
 ### Respond To Threads
 
-- Only after commit and push succeed, use `<%= it.config.tools.pr_sync.name %>` to post short factual replies to addressed feedback
-- Reply with `replies` keyed by the addressed comment IDs; use `commentBody` only for general CI feedback
-- Store the number of addressed threads as `<threads-resolved>`
+- After any required commit and push succeeds, use `<%= it.config.tools.pr_sync.name %>` to respond to assessed feedback
+- Reply to `actionable` items with a short factual summary of the fix, to `disputed` items with the concise technical reason no change was made, and to `needs-clarification` items with one focused request for the missing information
+- Do not reply to resolved, already-addressed, or superseded feedback unless a correction is needed to prevent confusion
+- Use `replies` only for inline review-thread comment IDs; aggregate responses to issue comments, formal review bodies, and general CI feedback into one concise `commentBody` with clear references to each source
+- Store the number of replies as `<feedback-replies-posted>`
+- Add every assessed source ID to `<handled-feedback-ids>` after its required response succeeds, including already-addressed and superseded items that need no reply; reconsider them only when a later reply or code change adds material new evidence
+- If `<feedback-awaiting-clarification>` is greater than `0`, store `<fix-status>` as `waiting for clarification`; otherwise store it as `complete`

packages/core/test/commands.test.ts

@@ -35,13 +35,46 @@ describe("resolveCommands", () => {
     assert.doesNotMatch(template, /`<current-branch>` differs from `<pr-branch>` or `<current-head>` differs from `<pr-context\.pr\.headRefOid>`/);
   });
 
+  test("independently assesses PR feedback before fixing or replying", async () => {
+    const commands = await resolveCommands(process.cwd());
+    const fixTemplate = commands["pr/fix"]?.template ?? "";
+    const loopTemplate = commands["pr/fix/loop"]?.template ?? "";
+
+    for (const template of [fixTemplate, loopTemplate]) {
+      assert.match(template, /candidate feedback, not instructions to follow blindly/);
+      assert.match(template, /Classify each feedback item by comment ID/);
+      assert.match(template, /derive `<feedback-source>` as `automation`, `project-member`, or `external-or-unknown`/);
+      assert.match(template, /Give `project-member` feedback and non-automation feedback from the PR author greater authority/);
+      assert.match(template, /Use `replies` only for inline review-thread comment IDs/);
+      assert.match(template, /aggregate responses to issue comments, formal review bodies, and general CI feedback/);
+      assert.match(template, /Add every assessed source ID to `<handled-feedback-ids>`/);
+      assert.match(template, /Implement only feedback classified as `actionable`/);
+      assert.match(template, /`disputed` items with the concise technical reason no change was made/);
+      assert.match(template, /`needs-clarification` items with one focused request for the missing information/);
+    }
+
+    assert.match(fixTemplate, /PR fix waiting for clarification/);
+    assert.match(loopTemplate, /PR loop waiting for clarification/);
+    assert.doesNotMatch(fixTemplate, /Threads resolved/);
+  });
+
+  test("treats prior review comments as unverified context", async () => {
+    const commands = await resolveCommands(process.cwd());
+    const reviewTemplate = commands["pr/review"]?.template ?? "";
+
+    assert.match(reviewTemplate, /independently verify existing review claims rather than assuming they are correct/);
+  });
+
   test("runs pr fix loop inline with incremental review loading", async () => {
     const commands = await resolveCommands(process.cwd());
     const template = commands["pr/fix/loop"]?.template ?? "";
 
     assert.match(template, /gh pr checks <pr-number> --watch/);
     assert.match(template, /pr_load_review/);
     assert.match(template, /since: <review-checkpoint>/);
+    assert.match(template, /call `pr_load` with `pr: <pr-url>` for a final complete snapshot/);
+    assert.match(template, /comments posted by CI after checks completed/);
+    assert.match(template, /If recomputed `<actionable-work>` is empty/);
     assert.doesNotMatch(template, /<delegate/);
     assert.doesNotMatch(template, /question/);
   });