[plan] 为mega-planner添加自动备份debate和consensus的功能，issue提交到fork

[Tonny-Gu]

Now I see -- there's no /mega-planner in the workflow commands mapping. This means mega-planner is likely a newer rename/replacement but the permission system still references the old ultra-planner name. The general allow-list at rules.py:100 does cover gh issue (list|view|create) so view and create work anywhere. But gh issue edit is only allowed for the ultra-planner workflow specifically.

Now I have all the context needed. Here's my synthesis:

---

Implementation Plan: Mega-Planner Artifact Backup and Fork-Targeted Issues

Agent Perspectives Summary

Agent	Core Position	Key Insight
Bold	Both features: plans/ archive directory + -R flag on all gh commands	Clean separation of ephemeral .tmp/ and persistent plans/ directories
Paranoia	Feature 2 only: reject backup, add -R flag minimally	Consensus plan is already backed up to GitHub issue body; artifacts are ephemeral
Critique	Both proposals break permission system patterns; Bold violates "no code changes" principle	Permission regex patterns in determine.py:568-570 will not match with -R flag inserted
Proposal Reducer	Converge both to hybrid: .gitignore negation + minimal backup + -R	Both proposals reduce to ~14 LOC when simplified; plans/ directory is unnecessary
Code Reducer	Hybrid at +14 LOC vs Bold's +100 LOC; Paranoia's +12 LOC ignores user request	Repository grows 2.7% vs 19.5% — hybrid is optimal

Goal

Add fork-targeted issue operations (via -R flag) to all gh issue commands in the mega-planner workflow, and optionally persist debate artifacts to git so they survive .tmp/ cleanup.

Out of scope:

• Refactoring the permission system broadly
• Changing the mega-planner's multi-agent debate structure
• Adding new workflow modes

Codebase Analysis

File changes:

File	Level	Purpose
.claude-plugin/commands/mega-planner.md	major	Add FORK_REPO extraction and -R flag to all gh issue commands
.claude-plugin/lib/permission/rules.py	minor	Add gh issue edit to general allow-list OR add -R optional match
.claude-plugin/lib/permission/determine.py	minor	Update ultra-planner patterns to handle optional -R flag
.gitignore	minor	(Only if backup option chosen) Allow .tmp/issue-* tracking

Implementation Steps

Note: Steps below cover the consensus portion — Feature 2 (fork targeting) where all agents agree. Disagreement on Feature 1 (artifact backup) and permission system approach are in the Disagreement sections.

Step 1: Add FORK_REPO extraction at Step 1 of mega-planner

• File: .claude-plugin/commands/mega-planner.md
• Changes: Add repo slug extraction before the mode-parsing logic

Code Draft

 ### Step 1: Parse Arguments and Extract Feature Description

 Accept the $ARGUMENTS.

+**Resolve fork repository for issue targeting:**
+
+```bash
+FORK_REPO=$(git remote get-url origin | sed -E 's#.*(github\.com[:/])##; s/\.git$//')
+```
+
+Use `-R "${FORK_REPO}"` on all subsequent `gh issue` commands.
+
 **Resolve mode (fast-path):** If `--resolve` is at the beginning:

Step 2: Add -R "${FORK_REPO}" to gh issue view in resolve mode

• File: .claude-plugin/commands/mega-planner.md (line 135)
• Changes: Insert -R flag

Code Draft

-   gh issue view ${ISSUE_NUMBER} --json body -q '.body' > ".tmp/${FILE_PREFIX}-issue-body.md"
+   gh issue view ${ISSUE_NUMBER} -R "${FORK_REPO}" --json body -q '.body' > ".tmp/${FILE_PREFIX}-issue-body.md"

Step 3: Add -R "${FORK_REPO}" to gh issue view in refine mode

• File: .claude-plugin/commands/mega-planner.md (lines 163, 183)
• Changes: Insert -R flag in both the description and code block

Code Draft

-2. Fetch the issue: `gh issue view ${ISSUE_NUMBER} --json title,body`
+2. Fetch the issue: `gh issue view ${ISSUE_NUMBER} -R "${FORK_REPO}" --json title,body`

-gh issue view ${ISSUE_NUMBER} --json title,body
+gh issue view ${ISSUE_NUMBER} -R "${FORK_REPO}" --json title,body

Step 4: Add -R "${FORK_REPO}" to gh issue view in from-issue mode

• File: .claude-plugin/commands/mega-planner.md (line 193)
• Changes: Insert -R flag

Code Draft

-gh issue view <issue-no> --json title,body
+gh issue view <issue-no> -R "${FORK_REPO}" --json title,body

Step 5: Add -R "${FORK_REPO}" to gh issue create in Step 3

• File: .claude-plugin/commands/mega-planner.md (line 240)
• Changes: Insert -R flag

Code Draft

 ISSUE_URL=$(gh issue create \
+    -R "${FORK_REPO}" \
     --title "[plan] ${FEATURE_DESC}" \

Step 6: Add -R "${FORK_REPO}" to gh issue edit in Steps 8 and 9

• File: .claude-plugin/commands/mega-planner.md (lines 470, 491, 497)
• Changes: Insert -R flag on all edit commands

Code Draft

-gh issue edit ${ISSUE_NUMBER} --body-file "${CONSENSUS_PLAN_FILE}"
+gh issue edit ${ISSUE_NUMBER} -R "${FORK_REPO}" --body-file "${CONSENSUS_PLAN_FILE}"

-gh issue edit ${ISSUE_NUMBER} --add-label "agentize:plan"
+gh issue edit ${ISSUE_NUMBER} -R "${FORK_REPO}" --add-label "agentize:plan"

-gh issue edit ${ISSUE_NUMBER} --remove-label "agentize:feat-request"
+gh issue edit ${ISSUE_NUMBER} -R "${FORK_REPO}" --remove-label "agentize:feat-request"

Step 7: Update permission patterns to handle -R flag

• File: .claude-plugin/lib/permission/rules.py (line 100)
• Changes: Add gh issue edit to the general allow-list (since it's also needed by mega-planner, not just ultra-planner)

Code Draft

-        ('Bash', r'^gh issue (list|view|create)'),
+        ('Bash', r'^gh issue (list|view|create|edit)'),

• File: .claude-plugin/lib/permission/determine.py (lines 568-570)
• Changes: Update patterns to optionally match -R flag before other flags

Code Draft

         issue_creation_patterns = [
-            r'^gh issue create\s+--title\s+.*--body\s+.*',  # gh issue create with title and body
-            r'^gh issue edit\s+\d+\s+--title\s+.*',         # gh issue edit with title
-            r'^gh issue edit\s+\d+\s+--body\s+.*',          # gh issue edit with body
+            r'^gh issue create\s+(-R\s+\S+\s+)?--title\s+.*--body\s+.*',  # gh issue create with title and body
+            r'^gh issue edit\s+\d+\s+(-R\s+\S+\s+)?--title\s+.*',         # gh issue edit with title
+            r'^gh issue edit\s+\d+\s+(-R\s+\S+\s+)?--body\s+.*',          # gh issue edit with body
         ]

Overall Recommendation

Disagreement Summary

#	Topic	Options	AI Recommendation
1	Artifact Backup Strategy	A (No backup), B (.gitignore negation + commit), C (plans/ directory)	Option 1B
2	FORK_REPO Extraction Method	A (git remote sed), B (gh repo view)	Option 2A

Suggested Combination

Suggested combination: 1B + 2A because .gitignore negation is minimal and avoids duplicate storage, while sed-based extraction is more reliable offline and avoids an extra API call.

Alternative combinations:

• All Conservative (1A + 2A): No backup, just fork targeting — fastest to implement, ignores user request
• All Aggressive (1C + 2B): Full plans/ directory + gh repo view — most infrastructure, cleanest abstraction

---

Disagreement 1: Artifact Backup Strategy

Agent Perspectives

Agent	Position	Rationale
Bold	Create plans/ directory with copied artifacts	Clean separation of ephemeral and persistent; follows project README convention
Paranoia	No backup at all	Consensus already on GitHub; .tmp/ persists locally; committing generated files pollutes git
Critique	Bold's approach violates "no code changes" principle; Paranoia ignores user request	Both have valid points; hybrid needed
Proposal Reducer	.gitignore negation to track .tmp/issue-* files directly	Eliminates duplication; same result with ~3 lines vs ~80 lines
Code Reducer	Hybrid at +3 LOC for backup vs Bold's +88 LOC	plans/ has zero downstream consumers

Resolution Options

Option 1A: No Backup (Conservative)

Summary: Reject artifact backup entirely. Consensus plan is already backed up to GitHub issue body. Individual agent reports remain ephemeral in .tmp/.
Source: Paranoia

File Changes:

File	Level	Purpose
(none)	—	No changes needed

Implementation Steps:

No additional steps beyond the consensus fork-targeting changes.

Risks and Mitigations:

Risk	Likelihood	Impact	Mitigation
Lost debate context on .tmp/ cleanup	M	M	Re-run --from-issue to regenerate
User expectation not met	H	L	Document that consensus is persisted via GitHub issue

Option 1B: .gitignore Negation + Minimal Commit (Balanced)

Summary: Modify .gitignore to allow .tmp/issue-* files to be tracked, then git add and commit them after consensus. No new directories, no file copying.
Source: Proposal Reducer + Code Reducer (hybrid)

File Changes:

File	Level	Purpose
.gitignore	minor	Add negation pattern for .tmp/issue-*
.claude-plugin/commands/mega-planner.md	minor	Append 3 lines to Step 9

Implementation Steps:

Step 1: Update .gitignore

• File: .gitignore
• Changes: Add negation pattern

Code Draft

 .tmp/*
+!.tmp/issue-*
 trees/

Step 2: Append artifact commit to Step 9

• File: .claude-plugin/commands/mega-planner.md (after line 497)
• Changes: Add git add/commit after label operations

Code Draft

 gh issue edit ${ISSUE_NUMBER} -R "${FORK_REPO}" --remove-label "agentize:feat-request"
 ```

+**Archive debate artifacts to git:**
+
+```bash
+git add .tmp/issue-${ISSUE_NUMBER}-*.md
+git commit -m "[plan] Archive debate artifacts for #${ISSUE_NUMBER}"
+```
+
 **Expected output:**

Risks and Mitigations:

Risk	Likelihood	Impact	Mitigation
Violates "no code changes" principle	M	L	Archival commit is metadata, not code change; add clarifying comment
Git history growth	L	L	~9 files per planning session, each <50KB
.tmp/issue-* pattern too broad	L	M	Pattern only matches issue- prefixed files which are all mega-planner artifacts

Option 1C: Dedicated plans/ Directory (Aggressive)

Summary: Create plans/issue-{N}/ directory structure with copied artifacts and a README. Full separation of working and archived files.
Source: Bold

File Changes:

File	Level	Purpose
.claude-plugin/commands/mega-planner.md	major	Add new Step 10 with copy + commit logic (~40 LOC)
plans/README.md	minor	New file documenting directory purpose (~30 LOC)

Implementation Steps:

Step 1: Add Step 10 to mega-planner

• File: .claude-plugin/commands/mega-planner.md
• Changes: New step after Step 9 with artifact copy and commit

Code Draft

+### Step 10: Backup Debate Artifacts to Git
+
+```bash
+mkdir -p "plans/issue-${ISSUE_NUMBER}"
+
+for ARTIFACT in context bold paranoia critique proposal-reducer code-reducer debate consensus history; do
+    SRC=".tmp/issue-${ISSUE_NUMBER}-${ARTIFACT}.md"
+    if [ -f "$SRC" ]; then
+        cp "$SRC" "plans/issue-${ISSUE_NUMBER}/${ARTIFACT}.md"
+    fi
+done
+
+git add "plans/issue-${ISSUE_NUMBER}/"
+git commit -m "[#${ISSUE_NUMBER}][plan] Archive debate artifacts"
+```

Step 2: Create plans/README.md

• File: plans/README.md
• Changes: New file

Code Draft

# Plans Directory

Archived debate artifacts from the mega-planner workflow.

## Organization

```
plans/
  issue-{N}/
    context.md          - Understander context summary
    bold.md             - Bold proposer report
    paranoia.md         - Paranoia proposer report
    critique.md         - Critique agent report
    proposal-reducer.md - Proposal reducer report
    code-reducer.md     - Code reducer report
    debate.md           - Combined multi-agent debate report
    consensus.md        - Final consensus plan
    history.md          - Selection and refine history
```

Risks and Mitigations:

Risk	Likelihood	Impact	Mitigation
Duplicate storage (.tmp/ + plans/)	H	L	Accept as cost of clean separation
Violates "no code changes" principle	M	M	Frame as post-planning archival action
+100 LOC for a backup mechanism	M	L	Directory is self-documenting
No downstream consumer	H	L	Future --resolve could read from plans/

AI Recommendation: Option 1B because it fulfills the user's backup request with minimal code (+6 LOC across 2 files), avoids duplicate storage, and doesn't introduce new directories or workflow steps. The .gitignore negation pattern is the standard git approach for exceptions.

---

Disagreement 2: FORK_REPO Extraction Method

Agent Perspectives

Agent	Position	Rationale
Bold	sed on git remote get-url origin	Matches Python server pattern; works offline
Paranoia	sed on git remote get-url origin (same, slightly different regex)	Minimal, no API calls
Critique	Bold's regex is less robust than Paranoia's	Both work for this repo; Paranoia handles edge cases better
Proposal Reducer	gh repo view --json nameWithOwner	Simpler, no regex parsing
Code Reducer	gh repo view --json nameWithOwner	One canonical command, no regex

Resolution Options

Option 2A: sed-based extraction (Conservative)

Summary: Parse git remote get-url origin with sed. Works offline, no API call, matches the pattern already used in python/agentize/server/github.py.
Source: Bold + Paranoia (both agree on mechanism, differ in regex)

File Changes:

File	Level	Purpose
.claude-plugin/commands/mega-planner.md	minor	Add 1-line extraction

Implementation Steps:

Step 1: Add extraction

• File: .claude-plugin/commands/mega-planner.md
• Changes: (already included in consensus Step 1)

Code Draft

FORK_REPO=$(git remote get-url origin | sed -E 's#.*(github\.com[:/])##; s/\.git$//')

Risks and Mitigations:

Risk	Likelihood	Impact	Mitigation
Regex doesn't handle non-GitHub remotes	L	H	This project is GitHub-only
Fails if no origin remote	L	H	Error message from git is descriptive enough

Option 2B: gh repo view extraction (Aggressive)

Summary: Use gh repo view --json nameWithOwner to get the repo slug. Cleaner, no regex, but requires network access and authenticated gh session.
Source: Proposal Reducer + Code Reducer

File Changes:

File	Level	Purpose
.claude-plugin/commands/mega-planner.md	minor	Add 1-line extraction

Implementation Steps:

Step 1: Add extraction

• File: .claude-plugin/commands/mega-planner.md
• Changes: Replace sed-based extraction

Code Draft

FORK_REPO=$(gh repo view --json nameWithOwner -q '.nameWithOwner')

Risks and Mitigations:

Risk	Likelihood	Impact	Mitigation
Requires network for extraction (before any gh commands)	L	L	All subsequent commands also need network
gh repo view may resolve to upstream (same priority as gh issue)	M	H	Would defeat the purpose; needs testing
Extra API call at workflow start	L	L	One additional call is negligible

AI Recommendation: Option 2A because gh repo view has the same remote-resolution priority problem as gh issue — it may resolve to upstream rather than origin, defeating the purpose. The sed approach explicitly targets the origin remote, which is guaranteed to be the fork.

---

Success Criteria

• gh issue create targets the fork repository (origin), not upstream
• gh issue edit commands target the fork repository
• gh issue view commands target the fork repository
• Permission patterns in determine.py match commands with -R flag
• General allow-list in rules.py covers gh issue edit
• (If backup chosen) Debate artifacts persist in git after mega-planner completes

Risks and Mitigations

Risk	Likelihood	Impact	Mitigation
Permission patterns block -R flagged commands	H	H	Update regex patterns (consensus Step 7)
origin remote doesn't exist or isn't GitHub	L	H	Git error message is descriptive; this is a GitHub-only project
gh issue edit added to general allow-list too broadly	M	M	Only list/view/create/edit — no destructive operations
Workflow name mismatch (mega-planner vs ultra-planner)	M	M	Verify which workflow name triggers the permission path

Selection History

Timestamp	Disagreement	Options Summary	Selected Option	User Comments

Refine History

Timestamp	Summary

Option Compatibility Check

Status: VALIDATED

All option combinations (1A/1B/1C with 2A/2B) are architecturally compatible. No conflicting file modifications or design decisions detected between the disagreement points.