Continue

continue

Skill metrics: Emit a start event to the activity log:

echo "{\"type\":\"skill\",\"event\":\"start\",\"skill\":\"continue\",\"ts\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\"}" >> docs/sessions/.activity-log.jsonl

Resume development. Execute this smart continuation workflow:

Current Git State (auto-fetched)

• Branch: !git branch --show-current
• Status: !git status --short
• Open PRs: !gh pr list --author @me --state open --limit 5 2>/dev/null || echo "No PRs or gh not configured"

0. Project Health Scan

Before session recovery, assess project maturity by checking key artifacts:

Present as a health dashboard before the session state:

### Project Health
- ✅ Architecture documented
- ✅ Ground rules established (5 rules)
- ⚠️ No backlog stories — consider `/ideate`
- ⚠️ Architecture doc stale (Last Verified: 45 days ago) — consider `/architecture-check`
- ❌ Test command not configured

Only flag items that are missing or incomplete — don't clutter with all-green checks unless it's the first session.

0.5. Check for Failure State

Check for an interrupted skill session:

cat docs/sessions/.failure-state.md 2>/dev/null

If .failure-state.md exists, a previous skill was interrupted mid-execution. This is the highest-priority context source. The file uses YAML frontmatter with structured fields:

---
status: active          # "active" means interrupted workflow
skill: story-cycle      # which skill was running
phase: "3"              # phase number
phase_name: "Execution" # human-readable phase name
started_at: "..."       # when the workflow started (ISO-8601)
story: "..."            # story/task being worked on
branch: "..."           # git branch at time of interruption
next_action: "..."      # what to do next to resume
files_modified: [...]   # files changed so far
---

## Context
[Free-form notes about the interrupted state]

Parse the YAML frontmatter to extract structured data. Present it prominently:

### Interrupted Session Detected
- **Skill:** [from `skill` field]
- **Phase:** [from `phase` + `phase_name` fields]
- **Story:** [from `story` field]
- **Branch:** [from `branch` field]
- **Next action:** [from `next_action` field]
- **Files modified:** [from `files_modified` field]

Check started_at — if the failure state is older than 4 hours, flag it as potentially stale:

"This failure state is [N] hours old. It may be from a previous session that was abandoned. Confirm you want to resume this workflow."

Validate story status: If the failure state references a story ID, check its status in the epic file (docs/reference/backlog/E*.md). If the story status is in-progress, recovery is valid. If it's done or review, the story may have been completed in another session — inform the user.

Recommend the user resume the interrupted skill (e.g., /story-cycle or /debug-session) with the recovery context from the next_action and ## Context section.

0.6. Read Latest Session Handoff

Check for the most recent session file:

ls -t docs/sessions/session-*.md 2>/dev/null | head -1

If a session file exists, read it as the primary context source. It contains: completed work, pending items, next steps, files to load, test status, and warnings.

0.7. Team Context Check

# Recent commits by other authors (last 7 days)
git log --since="7 days ago" --format="%an: %s" --no-merges | head -10
# Other open PRs that might conflict
gh pr list --state open --limit 5 2>/dev/null

If other PRs touch files related to your current story, flag potential agentic drift — semantically incompatible changes that merge cleanly but encode different assumptions. See docs/reference/TEAM_WORKFLOW.md for the full team coordination workflow.

1. Assess Git State

Branch Scenarios:

• On feature branch with changes: Mid-sprint, continue work
• On feature branch, clean: Sprint may be ready for PR or needs more work
• On main with changes: Should not happen - create branch first
• On main, clean: Between sprints, ready to start new work

1.5. Load Project Context

Load the project context knowledge base for deep project understanding (use the context-prime micro-component from .claude/prompts/context-prime.md):

• Read docs/context/ files in priority order: overview + tech first, patterns + structure second, product last
• Skip files that are still template placeholders (contain only <!-- filled by --> comments)
• This provides persistent project knowledge that compounds across sessions

1.6. Reload Working Context

If a session file was found, use its "Files Accessed" section to reload context efficiently:

• Modified files: Read all — they contain your changes from last session
• Read (context-relevant): Read these only if continuing the same story
• Investigated (can skip): Skip unless the user specifically asks about them

This avoids re-exploring files that were already investigated last session.

2. Assess Project State

• Read @docs/progress.md for current sprint goal, story statuses, and notes
• Metrics warnings: Check docs/progress.md → ## Metrics table for any 🟡 or 🔴 status. If found, surface in the Project Health section (step 0):

  "⚠️ [Metric name]: [status] ([value], target [target]) — [sprint note context]" This warns the user about quality degradation before they choose what to work on. Skip if all 🟢 or no data.

• Read @docs/reference/BACKLOG_INDEX.md for current story status

2.5. Load Sprint Context

If on a sprint branch (branch name matches sprint-*):

1. Find the matching sprint spec: docs/sprints/sprint-<number>.md
2. Read it and extract:
   • Sprint goal — display prominently in the continuation summary
   • Stories table — current status of all sprint stories
   • Capacity — sessions available vs sessions consumed (count ✅ stories by size: S=1, M=2, L=4)
   • Decisions log — any decisions made in prior sessions that constrain current work
   • Boundaries — out of scope items and risks
3. If a session handoff file exists, cross-reference sprint_capacity from its frontmatter

This ensures the sprint goal and constraints survive across sessions — without this, resumed sessions lose the "why" behind the current work.

3. Determine Continuation Point

If on a feature branch:

• This is an active sprint
• Check if PR already exists: gh pr view
• If PR exists and approved: offer to merge
• If PR exists awaiting review: wait or continue work
• If no PR: continue sprint implementation

If on main:

• Find current IN_PROGRESS story in backlog
• If found: switch to or create its feature branch
• If none: identify next TODO story for new sprint

4. Handle Pending PRs

If there are open PRs awaiting merge, offer to:

• Check review status
• Merge approved PRs
• Address review feedback

5. Quick Verification

Run the project's test command (from CLAUDE.md Commands section) to verify everything works:

# Use the project's test command from CLAUDE.md

If mid-sprint: Resume from last checkpoint If starting new sprint: Execute sprint-start workflow

5.5. Health Dashboard

Present a quick status pulse:

# Last commit time and branch
git log -1 --format="%cr on %D"
# Uncommitted file count
git status --short | wc -l
# Session file age
ls -lt docs/sessions/session-*.md 2>/dev/null | head -1

### Session Health
- **Tests:** [passing/total] ([green/red])
- **Last commit:** [time ago] on [branch]
- **Open changes:** [count] files
- **Session file:** [age — e.g., "2 days ago" or "none found"]

6. Present Options

Based on analysis, present relevant options:

• "Continue current sprint on branch [branch-name]"
• "Complete sprint: run agents, update docs, create PR"
• "Merge approved PR #[number] and start next story"
• "Address review feedback on PR #[number]"
• "Start next story [ID]: [title] (will create new branch)"
• "Commit and push current changes to feature branch"

7. Wait for Direction

Skill metrics: The analysis is complete once findings are ready — emit a completion event BEFORE presenting them (the turn ends waiting for the user, so anything after the wait may never run):

echo "{\"type\":\"skill\",\"event\":\"end\",\"skill\":\"continue\",\"outcome\":\"success\",\"ts\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\"}" >> docs/sessions/.activity-log.jsonl

Present findings and wait for user choice before proceeding.