oh-plan

Create GitHub issues ready for oh-task agents. The skill has two modes:

1. Task mode: Investigates codebase, asks clarifying questions, creates issues
2. Session mode: Reads existing .oh/<session>.md context, skips investigation, creates issues

Invocation

# Task mode: investigate from scratch
/oh-plan "Add dark mode support to the dashboard"

# Session mode: use existing context from skills workflow
/oh-plan auth-refactor

Mode Detection

At the start, check if the argument matches an existing session file:

If .oh/<arg>.md exists AND contains ## Solution Space:
    → Session mode (skip investigation, use context)
Else:
    → Task mode (investigate, ask questions)

JIT References

If the session contains ## Problem Weave or ### S&T Selection, load [references/open-horizons-session.md](references/open-horizons-session.md). It defines the selection gate, issue hierarchy, S&T lineage block, session plan table, and workflow handoff.

Session Mode Flow

When .oh/<session>.md exists with solution-space context:

1. Read session file and extract:

- ## Aim → Goal for issues - ## Problem Space → Constraints, terrain, assumptions - ## Problem Weave → Recursive S&T steps, sufficiency groups, lineage, owners, and wiring candidates - ## Solution Space → Selected approach, trade-offs accepted - ## Problem Statement → Framing (if present)

1. Skip investigation - context already gathered by grounding skills

1. Skip clarifying questions - decisions already made during solution-space

1. Decompose the selected solution into right-sized issues:

- Start from the Solution Space recommendation - If Problem Weave exists, load the Open Horizons session reference and include only selected step IDs - Preserve the lineage fields named in that reference - Stop and report a planning gap when a selected child depends on an unselected parent or incomplete sufficiency group - Apply the decomposition rules below; aim for 1-4 hours of work per issue

1. Choose issue hierarchy:

- One selected tactic that is independently deliverable → one issue - Several selected tactics under one objective → use the hierarchy rules in the Open Horizons session reference - Do not create an empty parent issue for a single tactic

1. Create GitHub issues from the session context:

```bash gh issue create --title "<title>" --label "oh-planned" --body "$(cat <<'EOF' ## Goal

<from Aim section>

## Context

<from Problem Space: constraints, terrain> <from Solution Space: selected approach>

<if Problem Weave exists, insert the S&T Lineage block from the Open Horizons session reference>

## Acceptance Criteria

- [ ] <derived from solution trade-offs> - [ ] <derived from aim feedback signals>

## Trade-offs Accepted

<from Solution Space>

## Notes

<from Problem Space assumptions> EOF )" ```

1. Update the session file with created issues. When Problem Weave exists, use the lineage table in the Open Horizons session reference.

Session Mode Example

$ /oh-plan auth-refactor

Found session file: .oh/auth-refactor.md

Reading context...
- Aim: Users complete signup in under 60 seconds
- Problem: Current flow has 5 steps, competitors have 2
- Solution: Collapse to email + password, defer profile to post-signup
- Trade-offs: Less data upfront, need progressive profiling later

Decomposing into issues...

Created 2 issues:

#43: "Simplify signup to email + password only"
  - Remove name, company, role fields from signup
  - Add post-signup redirect to profile completion
  - Depends on: none

#44: "Add progressive profile completion flow"
  - Prompt for missing fields on first dashboard visit
  - Track completion percentage
  - Depends on: #43

Updated .oh/auth-refactor.md with issue links.
Done. Issues ready for /oh-task.

Task Mode Flow

When no session file exists (current behavior):

1. Explore the codebase to understand:

- Project architecture and patterns - Relevant files that would be touched - Existing similar implementations to follow - Potential complications or dependencies

1. Ask clarifying questions via AskUserQuestion:

- Scope boundaries (what's in/out of scope) - Technical preferences (if multiple approaches exist) - Priority and any constraints - Only ask questions that affect the plan

1. Assess complexity and determine decomposition:

- Single coherent task → create 1 issue - Multiple independent pieces → create multiple issues - Large with dependencies → create issues with "Depends on #N" links - Aim for issues that are 1-4 hours of work each

1. Create GitHub issue(s) with the oh-planned label:

```bash # Create label if it doesn't exist (one-time) gh label create oh-planned --description "Created via oh-plan skill" --color "0E8A16" 2>/dev/null || true

# Create issue gh issue create --title "<clear, actionable title>" --label "oh-planned" --body "$(cat <<'EOF' ## Goal

<1-2 sentences describing what we're trying to achieve>

## Context

<relevant background - files identified, patterns to follow, constraints>

## Acceptance Criteria

- [ ] Criterion 1 (specific, testable) - [ ] Criterion 2 - [ ] Criterion 3

## Notes

<any technical decisions, edge cases, or things to watch out for> EOF )" ```

1. Signal completion: Call signal_completion(status: "success") to notify the orchestrator.

Fallback: If signal_completion is not available, output COMPLETION: status=success as your final message.

Issue Quality Guidelines

Issues created by oh-plan should be:

• Self-contained: All context needed to start work without re-investigation
• Actionable: Clear title, specific acceptance criteria, "done" is obvious
• Right-sized: 1-4 hours of focused work (not too big, not trivial)
• Linked: Dependencies explicitly noted if multiple issues created
• Traceable: Session-mode issues retain selected S&T step ID, parent objective/step, sufficiency group, and acceptance signal

Good Issue Title Examples

• "Add session timeout detection with configurable threshold"
• "Refactor user auth to use JWT instead of sessions"
• "Fix race condition in concurrent task spawning"

Bad Issue Title Examples

• "Improve the code" (too vague)
• "Do the thing we discussed" (no context)
• "Bug fix" (what bug?)

Decomposition Strategy

Prefer one focused issue per coherent change. Split only when the pieces can be tested independently; do not create trivial issues.

When to Split

Split into multiple issues when ANY of these apply:

• Multiple modules with distinct concerns - Each coherent subsystem change can be its own issue
• Mix of backend and frontend - API changes and UI changes often work better as separate issues
• Testable in isolation - If parts can be tested independently, they're candidates for separate issues
• Different risk profiles - Risky changes shouldn't be bundled with straightforward ones

When NOT to Split

Keep as one issue when:

• Integration is trivial - Just wiring things together with a few lines of code
• Changes are tightly coupled - Splitting would create issues that can't be tested alone
• The "integration" is obvious - No complex error handling, no new edge cases
• Total scope is still reasonable - Even combined, it's 2-4 hours of work

Integration Issues

Consider a separate integration issue when connecting components involves:

• Non-trivial error handling at boundaries
• New edge cases that emerge from the combination
• Complex state coordination between systems
• A test surface that warrants explicit end-to-end coverage

Don't create integration issues for:

• Simple API calls with straightforward error handling
• Passing data from one component to another
• Standard CRUD wiring

Pattern when integration IS complex:

Issue #1: "Add theme persistence API" (Foundation)
Issue #2: "Add theme toggle component" (Feature)
Issue #3: "Wire theme system with fallback handling" (Integration - Depends on #1, #2)
  - Handle API failures gracefully
  - System preference detection fallback
  - Theme flicker prevention on load

Decomposition Example

Task: "Add dark mode support to the dashboard"

Over-decomposed (too many trivial issues):

Issue #1: Add preferences API
Issue #2: Add toggle component
Issue #3: Add CSS variables
Issue #4: Wire toggle to API
Issue #5: Wire CSS to components
Issue #6: Handle loading states

Under-decomposed (one big issue):

Issue #1: Add dark mode support (everything)

Right-sized:

Issue #1: "Add user preferences API with theme support" (Foundation)
  - GET/POST /api/preferences
  - Includes theme field

Issue #2: "Add dark mode theming system" (Feature - Depends on #1)
  - CSS custom properties for light/dark
  - Theme toggle in settings
  - Wire to preferences API
  - Apply theme on app load

Issue #3: "Handle theme edge cases" (Polish - Depends on #2)
  - Only if complex: system preference fallback,
    theme flicker prevention, etc.
  - Skip if straightforward

Dependency Notation

Use "Depends on #N" in issue body:

## Goal
Add dark mode theming system with toggle and persistence.

**Depends on:** #1 (preferences API)

What NOT to Do

• Don't create issues for trivial changes (typo fixes, single-line changes)
• Don't over-decompose: 10 tiny issues are worse than 2-3 focused issues
• Don't start implementation: oh-plan is for planning only
• Don't skip the clarifying questions if scope is ambiguous (task mode)
• Don't re-investigate when session context exists (session mode)
• Don't create issues for Problem Weave tactics that Solution Space did not mark selected
• Don't drop S&T IDs or lineage when converting selected tactics into issues

Exit Conditions

Task Mode Example

$ /oh-plan "Add heartbeat monitoring for tmux sessions"

Exploring codebase...
Found relevant files:
- src/tmux/sessions.ts (session management)
- src/state/sessions.ts (session state tracking)
- src/index.ts (main loop)

I have a few questions to scope this correctly:

Q: What should happen when a session stops responding?
[1] Mark as crashed, notify user
[2] Auto-restart the session
[3] Just log and continue

User: 1

Q: How often should we check for heartbeats?
[1] Every 30 seconds (Recommended)
[2] Every minute
[3] Every 5 minutes

User: 1

Creating issue...

Created issue #67: "Add heartbeat monitoring for tmux sessions"

## Goal
Detect when tmux sessions stop responding and notify the user.

## Context
- Sessions are tracked in src/state/sessions.ts
- Tmux interactions in src/tmux/sessions.ts
- Need to add periodic check in main loop

## Acceptance Criteria
- [ ] Check session liveness every 30 seconds
- [ ] Mark session as "crashed" if not responding
- [ ] Send Telegram notification when session crashes
- [ ] Add crashed sessions to /status output

## Notes
- Use tmux has-session to check liveness
- Consider: what if tmux server itself is down?

signal_completion(status: "success")
Done. Issue #67 ready for /oh-task.

Integration with Skills Workflow

When an Open Horizons session is present, load the session reference. Grounding skills gather context; oh-plan turns selected work into traceable issues without repeating the investigation.