@lvlup-sw/exarchos 2.8.1 → 2.8.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lvlup-sw/exarchos",
-  "version": "2.8.1",
+  "version": "2.8.3",
   "description": "Structure for agentic development — durable SDLC workflows, convergence gates, agent teams, and audit trails for Claude Code",
   "type": "module",
   "bin": {

package/skills/claude/shepherd/SKILL.md

@@ -93,7 +93,26 @@ Review the returned `actionItems` and `recommendation`:
 
 ### Step 2 — Fix
 
-Address each blocking action item from the assessment. Consult `references/fix-strategies.md` for detailed strategies per issue type.
+Before iterating over individual action items, classify them so the loop
+knows which to fix inline vs. delegate. Call `classify_review_items` on
+the assessment's `actionItems` (the comment-reply subset is what the
+classifier groups by file; CI-fix and review-address items are passed
+through unchanged):
+
+```typescript
+mcp__plugin_exarchos_exarchos__exarchos_orchestrate({
+  action: "classify_review_items",
+  featureId: "<id>",
+  actionItems: <actionItems from assess_stack>
+})
+```
+
+The result returns `groups: ClassificationGroup[]` with a `recommendation`
+per group: `direct` (handle inline), `delegate-fixer` (spawn the fixer
+subagent for batched/HIGH-severity work), or `delegate-scaffolder`
+(cheap subagent for doc nits). Iterate the groups in order, applying
+per-group strategy, then consult `references/fix-strategies.md` for
+detailed per-issue-type instructions.
 
 **Remediation event protocol (FLYWHEEL):**
 
@@ -130,7 +149,7 @@ These events feed `selfCorrectionRate` and `avgRemediationAttempts` metrics in C
 | Type | Strategy |
 |------|----------|
 | `ci-fix` | Read logs, reproduce locally, fix, commit to stack branch |
-| `comment-reply` | Read context from `actionItem.context`, compose response, post via GitHub MCP |
+| `comment-reply` | Use `actionItem.reviewer`, `normalizedSeverity`, `file`, `line`, and `raw` (full original comment) to compose a response. Provider adapters under `servers/exarchos-mcp/src/review/providers/` populate the input fields per #1159 — no manual tier parsing needed. **Posting:** PR-level summary comments use the provider-agnostic `add_pr_comment` orchestrate action; per-thread inline replies currently require the platform-specific MCP (e.g. `mcp__plugin_github_github__add_reply_to_pull_request_comment` for GitHub) until `VcsProvider` gains a thread-reply primitive — see [#1165](https://github.com/lvlup-sw/exarchos/issues/1165) for tracking. |
 | `review-address` | Fix code for CHANGES_REQUESTED, reply to each thread |
 | `restack` | Run `git rebase origin/<base>`, verify with `exarchos_orchestrate({ action: "list_prs" })` |
 | `escalate` | Consult `references/escalation-criteria.md` |

package/skills/claude/shepherd/references/fix-strategies.md

@@ -4,14 +4,16 @@ How to address common issues found during shepherd assessment.
 
 ## Decision: Fix Directly vs. Delegate
 
-| Condition | Approach |
-|-----------|----------|
-| Single file, < 20 lines changed | Fix directly in the stack branch |
-| Multiple files, contained concern | Fix directly if < 5 files |
-| Cross-cutting or architectural | Route to `/exarchos:delegate --fixes` for subagent dispatch |
-| Test changes needed | Fix directly (keep TDD cycle tight) |
+The `classify_review_items` orchestrate action owns this decision (#1159).
+Pass it the `actionItems` from `assess_stack` and consume the
+`recommendation` field on each returned group:
 
-**Default to fixing directly** — delegation adds overhead. Only delegate when the fix scope warrants it.
+- `direct` — handle inline in the shepherd loop
+- `delegate-fixer` — spawn the fixer subagent (batched / HIGH severity)
+- `delegate-scaffolder` — cheap scaffolder dispatch for doc nits
+
+Test changes still warrant inline handling regardless of recommendation —
+keep the TDD cycle tight rather than delegating test edits.
 
 ## Remediation Event Emission
 
@@ -154,52 +156,23 @@ For each comment, determine the appropriate response:
 | Already fixed (outdated) | Reply confirming | "Fixed in [commit/PR description] — [brief explanation]." |
 | False positive | Reply explaining | "[Explanation of why this doesn't apply in this context]." |
 
-### Sentry Comments
-
-Sentry's `[bot]` leaves **bug predictions** — AI-generated analysis of potential runtime issues. These appear as inline review comments with severity tags (CRITICAL, MEDIUM, etc.).
-
-**Sentry comments deserve careful attention because they often identify real bugs** (field name mismatches, type coercion issues, null reference risks).
-
-How to handle:
-1. Read the full comment body — Sentry includes a "Suggested Fix" section
-2. Evaluate whether the bug is real:
-   - Check if the code path is actually reachable
-   - Check if the field names/types match what the data actually provides
-   - Check existing tests — does any test exercise this path?
-3. If real: fix the bug, add a test, reply confirming
-4. If false positive: reply explaining why (e.g., "This path is guarded by X" or "The field is validated at Y before reaching this code")
-
-**Common Sentry findings:**
-- Field name mismatches between producers and consumers
-- Missing null checks on optional fields
-- Type mismatches (string vs. enum, array vs. object)
-- Unreachable error paths due to upstream validation
-
-### CodeRabbit Comments
-
-CodeRabbit leaves detailed code review suggestions with severity indicators. It re-reviews automatically on push, so code fixes may auto-resolve threads.
-
-How to handle:
-1. Read all CodeRabbit comments, noting severity (Critical, Major, Minor)
-2. Critical/Major: Must address — fix or provide strong rationale for not fixing
-3. Minor: Fix if low-effort, otherwise acknowledge
-4. CodeRabbit marks threads as "Addressed in commits" when it detects the code changed — but always verify with a reply
-
-**Common CodeRabbit findings:**
-- Error handling gaps (missing try/catch, bare catches)
-- Code duplication (DRY violations)
-- Style/naming suggestions
-- Performance optimizations
-- Security concerns
-
-### Human Reviewer Comments
-
-Human comments require the most careful handling:
-1. Read comments carefully — understand the full context
-2. For required changes: fix the code, reply confirming
-3. For questions: answer directly on the PR
-4. For suggestions: discuss or implement, reply with decision
-5. For approval with minor nits: fix nits, note the approval
+### Per-reviewer parsing (Sentry, CodeRabbit, Human, GitHub-Copilot)
+
+Severity normalization and per-reviewer comment parsing live in the
+provider adapters under `servers/exarchos-mcp/src/review/providers/` (#1159).
+`assess_stack` dispatches each PR comment through the adapter registry
+and attaches a normalized `ActionItem` (with `normalizedSeverity` and
+`reviewer` fields) to each unresolved comment. Use that signal when
+deciding response strategy below; you do not need to re-parse tier
+markers in the shepherd loop.
+
+If a *recognised* reviewer (e.g. CodeRabbit) ships a new severity tier
+that the adapter does not match, the `provider.unknown-tier` event
+surfaces the unrecognised tier marker for follow-up — the comment is
+processed as MEDIUM in the meantime. Unknown *reviewers* (authors that
+don't match any typed adapter) are routed silently to the `unknown`
+adapter and never trigger this event; their comments are also processed
+as MEDIUM by default.
 
 ### GitHub Actions Bot Comments
 

package/skills/codex/shepherd/SKILL.md

@@ -93,7 +93,26 @@ Review the returned `actionItems` and `recommendation`:
 
 ### Step 2 — Fix
 
-Address each blocking action item from the assessment. Consult `references/fix-strategies.md` for detailed strategies per issue type.
+Before iterating over individual action items, classify them so the loop
+knows which to fix inline vs. delegate. Call `classify_review_items` on
+the assessment's `actionItems` (the comment-reply subset is what the
+classifier groups by file; CI-fix and review-address items are passed
+through unchanged):
+
+```typescript
+mcp__exarchos__exarchos_orchestrate({
+  action: "classify_review_items",
+  featureId: "<id>",
+  actionItems: <actionItems from assess_stack>
+})
+```
+
+The result returns `groups: ClassificationGroup[]` with a `recommendation`
+per group: `direct` (handle inline), `delegate-fixer` (spawn the fixer
+subagent for batched/HIGH-severity work), or `delegate-scaffolder`
+(cheap subagent for doc nits). Iterate the groups in order, applying
+per-group strategy, then consult `references/fix-strategies.md` for
+detailed per-issue-type instructions.
 
 **Remediation event protocol (FLYWHEEL):**
 
@@ -130,7 +149,7 @@ These events feed `selfCorrectionRate` and `avgRemediationAttempts` metrics in C
 | Type | Strategy |
 |------|----------|
 | `ci-fix` | Read logs, reproduce locally, fix, commit to stack branch |
-| `comment-reply` | Read context from `actionItem.context`, compose response, post via GitHub MCP |
+| `comment-reply` | Use `actionItem.reviewer`, `normalizedSeverity`, `file`, `line`, and `raw` (full original comment) to compose a response. Provider adapters under `servers/exarchos-mcp/src/review/providers/` populate the input fields per #1159 — no manual tier parsing needed. **Posting:** PR-level summary comments use the provider-agnostic `add_pr_comment` orchestrate action; per-thread inline replies currently require the platform-specific MCP (e.g. `mcp__plugin_github_github__add_reply_to_pull_request_comment` for GitHub) until `VcsProvider` gains a thread-reply primitive — see [#1165](https://github.com/lvlup-sw/exarchos/issues/1165) for tracking. |
 | `review-address` | Fix code for CHANGES_REQUESTED, reply to each thread |
 | `restack` | Run `git rebase origin/<base>`, verify with `exarchos_orchestrate({ action: "list_prs" })` |
 | `escalate` | Consult `references/escalation-criteria.md` |

package/skills/codex/shepherd/references/fix-strategies.md

@@ -4,14 +4,16 @@ How to address common issues found during shepherd assessment.
 
 ## Decision: Fix Directly vs. Delegate
 
-| Condition | Approach |
-|-----------|----------|
-| Single file, < 20 lines changed | Fix directly in the stack branch |
-| Multiple files, contained concern | Fix directly if < 5 files |
-| Cross-cutting or architectural | Route to `/exarchos:delegate --fixes` for subagent dispatch |
-| Test changes needed | Fix directly (keep TDD cycle tight) |
+The `classify_review_items` orchestrate action owns this decision (#1159).
+Pass it the `actionItems` from `assess_stack` and consume the
+`recommendation` field on each returned group:
 
-**Default to fixing directly** — delegation adds overhead. Only delegate when the fix scope warrants it.
+- `direct` — handle inline in the shepherd loop
+- `delegate-fixer` — spawn the fixer subagent (batched / HIGH severity)
+- `delegate-scaffolder` — cheap scaffolder dispatch for doc nits
+
+Test changes still warrant inline handling regardless of recommendation —
+keep the TDD cycle tight rather than delegating test edits.
 
 ## Remediation Event Emission
 
@@ -154,52 +156,23 @@ For each comment, determine the appropriate response:
 | Already fixed (outdated) | Reply confirming | "Fixed in [commit/PR description] — [brief explanation]." |
 | False positive | Reply explaining | "[Explanation of why this doesn't apply in this context]." |
 
-### Sentry Comments
-
-Sentry's `[bot]` leaves **bug predictions** — AI-generated analysis of potential runtime issues. These appear as inline review comments with severity tags (CRITICAL, MEDIUM, etc.).
-
-**Sentry comments deserve careful attention because they often identify real bugs** (field name mismatches, type coercion issues, null reference risks).
-
-How to handle:
-1. Read the full comment body — Sentry includes a "Suggested Fix" section
-2. Evaluate whether the bug is real:
-   - Check if the code path is actually reachable
-   - Check if the field names/types match what the data actually provides
-   - Check existing tests — does any test exercise this path?
-3. If real: fix the bug, add a test, reply confirming
-4. If false positive: reply explaining why (e.g., "This path is guarded by X" or "The field is validated at Y before reaching this code")
-
-**Common Sentry findings:**
-- Field name mismatches between producers and consumers
-- Missing null checks on optional fields
-- Type mismatches (string vs. enum, array vs. object)
-- Unreachable error paths due to upstream validation
-
-### CodeRabbit Comments
-
-CodeRabbit leaves detailed code review suggestions with severity indicators. It re-reviews automatically on push, so code fixes may auto-resolve threads.
-
-How to handle:
-1. Read all CodeRabbit comments, noting severity (Critical, Major, Minor)
-2. Critical/Major: Must address — fix or provide strong rationale for not fixing
-3. Minor: Fix if low-effort, otherwise acknowledge
-4. CodeRabbit marks threads as "Addressed in commits" when it detects the code changed — but always verify with a reply
-
-**Common CodeRabbit findings:**
-- Error handling gaps (missing try/catch, bare catches)
-- Code duplication (DRY violations)
-- Style/naming suggestions
-- Performance optimizations
-- Security concerns
-
-### Human Reviewer Comments
-
-Human comments require the most careful handling:
-1. Read comments carefully — understand the full context
-2. For required changes: fix the code, reply confirming
-3. For questions: answer directly on the PR
-4. For suggestions: discuss or implement, reply with decision
-5. For approval with minor nits: fix nits, note the approval
+### Per-reviewer parsing (Sentry, CodeRabbit, Human, GitHub-Copilot)
+
+Severity normalization and per-reviewer comment parsing live in the
+provider adapters under `servers/exarchos-mcp/src/review/providers/` (#1159).
+`assess_stack` dispatches each PR comment through the adapter registry
+and attaches a normalized `ActionItem` (with `normalizedSeverity` and
+`reviewer` fields) to each unresolved comment. Use that signal when
+deciding response strategy below; you do not need to re-parse tier
+markers in the shepherd loop.
+
+If a *recognised* reviewer (e.g. CodeRabbit) ships a new severity tier
+that the adapter does not match, the `provider.unknown-tier` event
+surfaces the unrecognised tier marker for follow-up — the comment is
+processed as MEDIUM in the meantime. Unknown *reviewers* (authors that
+don't match any typed adapter) are routed silently to the `unknown`
+adapter and never trigger this event; their comments are also processed
+as MEDIUM by default.
 
 ### GitHub Actions Bot Comments
 

package/skills/copilot/shepherd/SKILL.md

@@ -93,7 +93,26 @@ Review the returned `actionItems` and `recommendation`:
 
 ### Step 2 — Fix
 
-Address each blocking action item from the assessment. Consult `references/fix-strategies.md` for detailed strategies per issue type.
+Before iterating over individual action items, classify them so the loop
+knows which to fix inline vs. delegate. Call `classify_review_items` on
+the assessment's `actionItems` (the comment-reply subset is what the
+classifier groups by file; CI-fix and review-address items are passed
+through unchanged):
+
+```typescript
+mcp__exarchos__exarchos_orchestrate({
+  action: "classify_review_items",
+  featureId: "<id>",
+  actionItems: <actionItems from assess_stack>
+})
+```
+
+The result returns `groups: ClassificationGroup[]` with a `recommendation`
+per group: `direct` (handle inline), `delegate-fixer` (spawn the fixer
+subagent for batched/HIGH-severity work), or `delegate-scaffolder`
+(cheap subagent for doc nits). Iterate the groups in order, applying
+per-group strategy, then consult `references/fix-strategies.md` for
+detailed per-issue-type instructions.
 
 **Remediation event protocol (FLYWHEEL):**
 
@@ -130,7 +149,7 @@ These events feed `selfCorrectionRate` and `avgRemediationAttempts` metrics in C
 | Type | Strategy |
 |------|----------|
 | `ci-fix` | Read logs, reproduce locally, fix, commit to stack branch |
-| `comment-reply` | Read context from `actionItem.context`, compose response, post via GitHub MCP |
+| `comment-reply` | Use `actionItem.reviewer`, `normalizedSeverity`, `file`, `line`, and `raw` (full original comment) to compose a response. Provider adapters under `servers/exarchos-mcp/src/review/providers/` populate the input fields per #1159 — no manual tier parsing needed. **Posting:** PR-level summary comments use the provider-agnostic `add_pr_comment` orchestrate action; per-thread inline replies currently require the platform-specific MCP (e.g. `mcp__plugin_github_github__add_reply_to_pull_request_comment` for GitHub) until `VcsProvider` gains a thread-reply primitive — see [#1165](https://github.com/lvlup-sw/exarchos/issues/1165) for tracking. |
 | `review-address` | Fix code for CHANGES_REQUESTED, reply to each thread |
 | `restack` | Run `git rebase origin/<base>`, verify with `exarchos_orchestrate({ action: "list_prs" })` |
 | `escalate` | Consult `references/escalation-criteria.md` |

package/skills/copilot/shepherd/references/fix-strategies.md

@@ -4,14 +4,16 @@ How to address common issues found during shepherd assessment.
 
 ## Decision: Fix Directly vs. Delegate
 
-| Condition | Approach |
-|-----------|----------|
-| Single file, < 20 lines changed | Fix directly in the stack branch |
-| Multiple files, contained concern | Fix directly if < 5 files |
-| Cross-cutting or architectural | Route to `/exarchos:delegate --fixes` for subagent dispatch |
-| Test changes needed | Fix directly (keep TDD cycle tight) |
+The `classify_review_items` orchestrate action owns this decision (#1159).
+Pass it the `actionItems` from `assess_stack` and consume the
+`recommendation` field on each returned group:
 
-**Default to fixing directly** — delegation adds overhead. Only delegate when the fix scope warrants it.
+- `direct` — handle inline in the shepherd loop
+- `delegate-fixer` — spawn the fixer subagent (batched / HIGH severity)
+- `delegate-scaffolder` — cheap scaffolder dispatch for doc nits
+
+Test changes still warrant inline handling regardless of recommendation —
+keep the TDD cycle tight rather than delegating test edits.
 
 ## Remediation Event Emission
 
@@ -154,52 +156,23 @@ For each comment, determine the appropriate response:
 | Already fixed (outdated) | Reply confirming | "Fixed in [commit/PR description] — [brief explanation]." |
 | False positive | Reply explaining | "[Explanation of why this doesn't apply in this context]." |
 
-### Sentry Comments
-
-Sentry's `[bot]` leaves **bug predictions** — AI-generated analysis of potential runtime issues. These appear as inline review comments with severity tags (CRITICAL, MEDIUM, etc.).
-
-**Sentry comments deserve careful attention because they often identify real bugs** (field name mismatches, type coercion issues, null reference risks).
-
-How to handle:
-1. Read the full comment body — Sentry includes a "Suggested Fix" section
-2. Evaluate whether the bug is real:
-   - Check if the code path is actually reachable
-   - Check if the field names/types match what the data actually provides
-   - Check existing tests — does any test exercise this path?
-3. If real: fix the bug, add a test, reply confirming
-4. If false positive: reply explaining why (e.g., "This path is guarded by X" or "The field is validated at Y before reaching this code")
-
-**Common Sentry findings:**
-- Field name mismatches between producers and consumers
-- Missing null checks on optional fields
-- Type mismatches (string vs. enum, array vs. object)
-- Unreachable error paths due to upstream validation
-
-### CodeRabbit Comments
-
-CodeRabbit leaves detailed code review suggestions with severity indicators. It re-reviews automatically on push, so code fixes may auto-resolve threads.
-
-How to handle:
-1. Read all CodeRabbit comments, noting severity (Critical, Major, Minor)
-2. Critical/Major: Must address — fix or provide strong rationale for not fixing
-3. Minor: Fix if low-effort, otherwise acknowledge
-4. CodeRabbit marks threads as "Addressed in commits" when it detects the code changed — but always verify with a reply
-
-**Common CodeRabbit findings:**
-- Error handling gaps (missing try/catch, bare catches)
-- Code duplication (DRY violations)
-- Style/naming suggestions
-- Performance optimizations
-- Security concerns
-
-### Human Reviewer Comments
-
-Human comments require the most careful handling:
-1. Read comments carefully — understand the full context
-2. For required changes: fix the code, reply confirming
-3. For questions: answer directly on the PR
-4. For suggestions: discuss or implement, reply with decision
-5. For approval with minor nits: fix nits, note the approval
+### Per-reviewer parsing (Sentry, CodeRabbit, Human, GitHub-Copilot)
+
+Severity normalization and per-reviewer comment parsing live in the
+provider adapters under `servers/exarchos-mcp/src/review/providers/` (#1159).
+`assess_stack` dispatches each PR comment through the adapter registry
+and attaches a normalized `ActionItem` (with `normalizedSeverity` and
+`reviewer` fields) to each unresolved comment. Use that signal when
+deciding response strategy below; you do not need to re-parse tier
+markers in the shepherd loop.
+
+If a *recognised* reviewer (e.g. CodeRabbit) ships a new severity tier
+that the adapter does not match, the `provider.unknown-tier` event
+surfaces the unrecognised tier marker for follow-up — the comment is
+processed as MEDIUM in the meantime. Unknown *reviewers* (authors that
+don't match any typed adapter) are routed silently to the `unknown`
+adapter and never trigger this event; their comments are also processed
+as MEDIUM by default.
 
 ### GitHub Actions Bot Comments
 

package/skills/cursor/shepherd/SKILL.md

@@ -93,7 +93,26 @@ Review the returned `actionItems` and `recommendation`:
 
 ### Step 2 — Fix
 
-Address each blocking action item from the assessment. Consult `references/fix-strategies.md` for detailed strategies per issue type.
+Before iterating over individual action items, classify them so the loop
+knows which to fix inline vs. delegate. Call `classify_review_items` on
+the assessment's `actionItems` (the comment-reply subset is what the
+classifier groups by file; CI-fix and review-address items are passed
+through unchanged):
+
+```typescript
+mcp__exarchos__exarchos_orchestrate({
+  action: "classify_review_items",
+  featureId: "<id>",
+  actionItems: <actionItems from assess_stack>
+})
+```
+
+The result returns `groups: ClassificationGroup[]` with a `recommendation`
+per group: `direct` (handle inline), `delegate-fixer` (spawn the fixer
+subagent for batched/HIGH-severity work), or `delegate-scaffolder`
+(cheap subagent for doc nits). Iterate the groups in order, applying
+per-group strategy, then consult `references/fix-strategies.md` for
+detailed per-issue-type instructions.
 
 **Remediation event protocol (FLYWHEEL):**
 
@@ -130,7 +149,7 @@ These events feed `selfCorrectionRate` and `avgRemediationAttempts` metrics in C
 | Type | Strategy |
 |------|----------|
 | `ci-fix` | Read logs, reproduce locally, fix, commit to stack branch |
-| `comment-reply` | Read context from `actionItem.context`, compose response, post via GitHub MCP |
+| `comment-reply` | Use `actionItem.reviewer`, `normalizedSeverity`, `file`, `line`, and `raw` (full original comment) to compose a response. Provider adapters under `servers/exarchos-mcp/src/review/providers/` populate the input fields per #1159 — no manual tier parsing needed. **Posting:** PR-level summary comments use the provider-agnostic `add_pr_comment` orchestrate action; per-thread inline replies currently require the platform-specific MCP (e.g. `mcp__plugin_github_github__add_reply_to_pull_request_comment` for GitHub) until `VcsProvider` gains a thread-reply primitive — see [#1165](https://github.com/lvlup-sw/exarchos/issues/1165) for tracking. |
 | `review-address` | Fix code for CHANGES_REQUESTED, reply to each thread |
 | `restack` | Run `git rebase origin/<base>`, verify with `exarchos_orchestrate({ action: "list_prs" })` |
 | `escalate` | Consult `references/escalation-criteria.md` |

package/skills/cursor/shepherd/references/fix-strategies.md

@@ -4,14 +4,16 @@ How to address common issues found during shepherd assessment.
 
 ## Decision: Fix Directly vs. Delegate
 
-| Condition | Approach |
-|-----------|----------|
-| Single file, < 20 lines changed | Fix directly in the stack branch |
-| Multiple files, contained concern | Fix directly if < 5 files |
-| Cross-cutting or architectural | Route to `/exarchos:delegate --fixes` for subagent dispatch |
-| Test changes needed | Fix directly (keep TDD cycle tight) |
+The `classify_review_items` orchestrate action owns this decision (#1159).
+Pass it the `actionItems` from `assess_stack` and consume the
+`recommendation` field on each returned group:
 
-**Default to fixing directly** — delegation adds overhead. Only delegate when the fix scope warrants it.
+- `direct` — handle inline in the shepherd loop
+- `delegate-fixer` — spawn the fixer subagent (batched / HIGH severity)
+- `delegate-scaffolder` — cheap scaffolder dispatch for doc nits
+
+Test changes still warrant inline handling regardless of recommendation —
+keep the TDD cycle tight rather than delegating test edits.
 
 ## Remediation Event Emission
 
@@ -154,52 +156,23 @@ For each comment, determine the appropriate response:
 | Already fixed (outdated) | Reply confirming | "Fixed in [commit/PR description] — [brief explanation]." |
 | False positive | Reply explaining | "[Explanation of why this doesn't apply in this context]." |
 
-### Sentry Comments
-
-Sentry's `[bot]` leaves **bug predictions** — AI-generated analysis of potential runtime issues. These appear as inline review comments with severity tags (CRITICAL, MEDIUM, etc.).
-
-**Sentry comments deserve careful attention because they often identify real bugs** (field name mismatches, type coercion issues, null reference risks).
-
-How to handle:
-1. Read the full comment body — Sentry includes a "Suggested Fix" section
-2. Evaluate whether the bug is real:
-   - Check if the code path is actually reachable
-   - Check if the field names/types match what the data actually provides
-   - Check existing tests — does any test exercise this path?
-3. If real: fix the bug, add a test, reply confirming
-4. If false positive: reply explaining why (e.g., "This path is guarded by X" or "The field is validated at Y before reaching this code")
-
-**Common Sentry findings:**
-- Field name mismatches between producers and consumers
-- Missing null checks on optional fields
-- Type mismatches (string vs. enum, array vs. object)
-- Unreachable error paths due to upstream validation
-
-### CodeRabbit Comments
-
-CodeRabbit leaves detailed code review suggestions with severity indicators. It re-reviews automatically on push, so code fixes may auto-resolve threads.
-
-How to handle:
-1. Read all CodeRabbit comments, noting severity (Critical, Major, Minor)
-2. Critical/Major: Must address — fix or provide strong rationale for not fixing
-3. Minor: Fix if low-effort, otherwise acknowledge
-4. CodeRabbit marks threads as "Addressed in commits" when it detects the code changed — but always verify with a reply
-
-**Common CodeRabbit findings:**
-- Error handling gaps (missing try/catch, bare catches)
-- Code duplication (DRY violations)
-- Style/naming suggestions
-- Performance optimizations
-- Security concerns
-
-### Human Reviewer Comments
-
-Human comments require the most careful handling:
-1. Read comments carefully — understand the full context
-2. For required changes: fix the code, reply confirming
-3. For questions: answer directly on the PR
-4. For suggestions: discuss or implement, reply with decision
-5. For approval with minor nits: fix nits, note the approval
+### Per-reviewer parsing (Sentry, CodeRabbit, Human, GitHub-Copilot)
+
+Severity normalization and per-reviewer comment parsing live in the
+provider adapters under `servers/exarchos-mcp/src/review/providers/` (#1159).
+`assess_stack` dispatches each PR comment through the adapter registry
+and attaches a normalized `ActionItem` (with `normalizedSeverity` and
+`reviewer` fields) to each unresolved comment. Use that signal when
+deciding response strategy below; you do not need to re-parse tier
+markers in the shepherd loop.
+
+If a *recognised* reviewer (e.g. CodeRabbit) ships a new severity tier
+that the adapter does not match, the `provider.unknown-tier` event
+surfaces the unrecognised tier marker for follow-up — the comment is
+processed as MEDIUM in the meantime. Unknown *reviewers* (authors that
+don't match any typed adapter) are routed silently to the `unknown`
+adapter and never trigger this event; their comments are also processed
+as MEDIUM by default.
 
 ### GitHub Actions Bot Comments
 

package/skills/generic/shepherd/SKILL.md

@@ -93,7 +93,26 @@ Review the returned `actionItems` and `recommendation`:
 
 ### Step 2 — Fix
 
-Address each blocking action item from the assessment. Consult `references/fix-strategies.md` for detailed strategies per issue type.
+Before iterating over individual action items, classify them so the loop
+knows which to fix inline vs. delegate. Call `classify_review_items` on
+the assessment's `actionItems` (the comment-reply subset is what the
+classifier groups by file; CI-fix and review-address items are passed
+through unchanged):
+
+```typescript
+mcp__exarchos__exarchos_orchestrate({
+  action: "classify_review_items",
+  featureId: "<id>",
+  actionItems: <actionItems from assess_stack>
+})
+```
+
+The result returns `groups: ClassificationGroup[]` with a `recommendation`
+per group: `direct` (handle inline), `delegate-fixer` (spawn the fixer
+subagent for batched/HIGH-severity work), or `delegate-scaffolder`
+(cheap subagent for doc nits). Iterate the groups in order, applying
+per-group strategy, then consult `references/fix-strategies.md` for
+detailed per-issue-type instructions.
 
 **Remediation event protocol (FLYWHEEL):**
 
@@ -130,7 +149,7 @@ These events feed `selfCorrectionRate` and `avgRemediationAttempts` metrics in C
 | Type | Strategy |
 |------|----------|
 | `ci-fix` | Read logs, reproduce locally, fix, commit to stack branch |
-| `comment-reply` | Read context from `actionItem.context`, compose response, post via GitHub MCP |
+| `comment-reply` | Use `actionItem.reviewer`, `normalizedSeverity`, `file`, `line`, and `raw` (full original comment) to compose a response. Provider adapters under `servers/exarchos-mcp/src/review/providers/` populate the input fields per #1159 — no manual tier parsing needed. **Posting:** PR-level summary comments use the provider-agnostic `add_pr_comment` orchestrate action; per-thread inline replies currently require the platform-specific MCP (e.g. `mcp__plugin_github_github__add_reply_to_pull_request_comment` for GitHub) until `VcsProvider` gains a thread-reply primitive — see [#1165](https://github.com/lvlup-sw/exarchos/issues/1165) for tracking. |
 | `review-address` | Fix code for CHANGES_REQUESTED, reply to each thread |
 | `restack` | Run `git rebase origin/<base>`, verify with `exarchos_orchestrate({ action: "list_prs" })` |
 | `escalate` | Consult `references/escalation-criteria.md` |