skill-context-detection

Host: Codex CLI — This skill was designed for Claude Code and adapted for Codex. Cross-reference commands use installed skill names in Codex rather than /octo:* slash commands. Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it. For host tool equivalents, see skills/blocks/codex-host-adapter.md.

Context Detection - Internal Skill

Purpose

This skill provides automatic context detection to determine whether the user is working in a Development context (code-focused) or Knowledge context (research/strategy-focused). This replaces the manual /octo:km toggle with intelligent auto-detection.

Detection Algorithm

When a workflow skill activates, detect context using these signals:

Step 1: Check for Explicit Override

If user has explicitly set mode via /octo:km on or /octo:km off, respect that setting.

# Check if knowledge mode is explicitly set
if [[ -f ~/.claude-octopus/config/knowledge-mode ]]; then
  EXPLICIT_MODE=$(cat ~/.claude-octopus/config/knowledge-mode)
  if [[ "$EXPLICIT_MODE" == "on" ]]; then
    echo "knowledge"
    exit 0
  elif [[ "$EXPLICIT_MODE" == "off" ]]; then
    echo "dev"
    exit 0
  fi
fi
# If "auto" or not set, proceed with auto-detection

Step 2: Analyze Prompt Content (Strongest Signal)

Knowledge Context Indicators (check prompt for these terms):

• Business/strategy: "market", "ROI", "stakeholders", "strategy", "business case", "competitive"
• Research: "literature", "synthesis", "academic", "papers", "research question"
• UX: "personas", "user research", "journey map", "pain points", "interviews"
• Deliverables: "presentation", "report", "PRD", "proposal", "executive summary"

Dev Context Indicators (check prompt for these terms):

• Technical: "API", "endpoint", "database", "function", "class", "module"
• Actions: "implement", "debug", "refactor", "test", "deploy", "build"
• Artifacts: "code", "tests", "migration", "schema", "controller"

Scoring:

• Count knowledge indicators in prompt
• Count dev indicators in prompt
• Higher count wins
• If tied, check project context (Step 3)

Step 3: Analyze Project Context (Secondary Signal)

Dev Project Indicators:

• Has package.json, Cargo.toml, go.mod, pyproject.toml, pom.xml
• Has src/, lib/, app/ directories with code files
• Recent files are .ts, .js, .py, .go, .rs, .java

Knowledge Project Indicators:

• Has docs/, research/, strategy/, reports/ directories
• Majority of files are .md, .docx, .pdf, .pptx
• No code package managers detected

Step 4: Default Fallback

If signals are ambiguous or equal:

• In a git repo with code files → Default to Dev Context
• No code files detected → Default to Knowledge Context

Context Output Format

Return detected context as a structured object for use by workflow skills:

{
  "context": "dev" | "knowledge",
  "confidence": "high" | "medium" | "low",
  "signals": {
    "prompt_indicators": ["API", "endpoint", "database"],
    "project_type": "node_typescript",
    "explicit_override": false
  }
}

How Workflow Skills Use Context

flow-discover (Research)

Aspect	Dev Context	Knowledge Context
Research Focus	Technical implementation, library comparison, code patterns	Market analysis, academic synthesis, competitive research
Primary Agents	Codex (implementation), Antigravity (ecosystem)	Antigravity (analysis), research-synthesizer
Output Format	Code examples, API comparisons, tech recommendations	Reports, frameworks, strategic recommendations
Visual Banner	🔍 [Dev] Discover Phase: Technical research	🔍 [Knowledge] Discover Phase: Strategic research

flow-develop (Build)

Aspect	Dev Context	Knowledge Context
Build Focus	Code generation, implementation, architecture	PRDs, strategy docs, presentations
Primary Agents	Codex (code), backend-architect, tdd-orchestrator	product-writer, strategy-analyst, exec-communicator
Output Format	Source files, tests, migrations	Documents, frameworks, action plans
Visual Banner	🛠️ [Dev] Develop Phase: Building code	🛠️ [Knowledge] Develop Phase: Building deliverables

flow-deliver (Review)

Aspect	Dev Context	Knowledge Context
Review Focus	Code quality, security, performance	Document quality, argument strength, completeness
Primary Agents	code-reviewer, security-auditor	exec-communicator, strategy-analyst
Quality Gates	OWASP, test coverage, maintainability	Evidence quality, clarity, actionability
Visual Banner	✅ [Dev] Deliver Phase: Code review	✅ [Knowledge] Deliver Phase: Document review

Visual Indicator Update

When context is detected, update the visual banner to show context:

Dev Context:

🐙 **CLAUDE OCTOPUS ACTIVATED** - Multi-provider research mode
🔍 [Dev] Discover Phase: Researching OAuth implementation patterns

Providers:
🔴 Codex CLI - Technical implementation analysis
🟡 Antigravity CLI - Ecosystem and library comparison
🔵 Claude - Strategic synthesis

Knowledge Context:

🐙 **CLAUDE OCTOPUS ACTIVATED** - Multi-provider research mode
🔍 [Knowledge] Discover Phase: Researching market entry strategies

Providers:
🔴 Codex CLI - Data analysis and modeling
🟡 Antigravity CLI - Market and competitive research
🔵 Claude - Strategic synthesis

Implementation in Workflow Skills

Each flow skill should:

1. Before executing workflow, run context detection
2. Show detected context in visual banner
3. Adjust behavior based on context:
   • Agent selection
   • Prompt framing for external CLIs
   • Output format expectations
   • Quality gate criteria

Example Integration (Pseudocode)

When this skill activates:

1. **Detect context**
   - Analyze user's prompt for knowledge vs dev indicators
   - Check project type (code repo vs doc-heavy)
   - Check for explicit override (~/.claude-octopus/config/knowledge-mode)
   - Determine: "dev" or "knowledge" with confidence level

2. **Show context-aware banner**

🐙 CLAUDE OCTOPUS ACTIVATED - Multi-provider [research|implementation|validation] mode [Phase Emoji] [Context] [Phase Name]: [Description]

Detected Context: [Dev|Knowledge] (confidence: [high|medium|low])

3. **Execute workflow with context-appropriate behavior**
- Frame prompts for external providers based on context
- Select appropriate synthesis approach
- Apply context-specific quality gates

Override Mechanism

Users can still explicitly set context when auto-detection is wrong:

# Force knowledge mode
/octo:km on

# Force dev mode
/octo:km off

# Return to auto-detection
/octo:km auto

When explicit override is set, context detection respects it until user resets to "auto".

Confidence Levels

• High: Strong signals in prompt AND project context agree
• Medium: Signals in prompt OR project context (not both)
• Low: Ambiguous signals, using fallback default

When confidence is "low", consider briefly mentioning the detected context to user:

"I detected this as a [dev/knowledge] task. If that's wrong, you can use /octo:km to override."

Testing Context Detection

To verify context detection is working:

1. In a code repository, ask "octo research caching patterns" → Should detect Dev Context
2. In same repo, ask "octo research market opportunities" → Should detect Knowledge Context
3. With /octo:km on set, ask "octo research API patterns" → Should use Knowledge Context (explicit override)

Proactive Skill Suggestions

When detecting the user's work stage, surface relevant command suggestions:

Detected Context	Suggestion
Brainstorming / exploring ideas	Consider /octo:brainstorm for structured ideation
Reviewing a plan or strategy	Consider /octo:plan for strategic planning
Debugging errors or failures	Consider /octo:debug for systematic investigation
Writing or running tests	Consider /octo:tdd for test-driven development
Code review before merge	Use Claude-native /review for ordinary review; suggest /octo:review for multi-AI escalation
Ready to deploy or ship	Consider /octo:deliver for quality-gated delivery
Researching a topic	Consider /octo:research for multi-source synthesis
Working on security	Use Claude-native /security-review for ordinary security review; suggest /octo:security for escalated OWASP or adversarial audit

Suggestion Format

Suggestions should be non-intrusive, appended as a brief note:

💡 Tip: You appear to be debugging — `/octo:debug` provides systematic investigation with multi-AI support.

Persistent Opt-Out

• If user says "stop suggesting" or "no more tips": set OCTO_PROACTIVE_SUGGESTIONS=off in .claude-octopus/preferences.json
• If user says "be proactive" or "turn on tips": set OCTO_PROACTIVE_SUGGESTIONS=on
• Check preference at suggestion time — never suggest when opted out
• Respect current mode: dev mode suggestions differ from knowledge work suggestions

Re-Enable Suggestions

Users who previously opted out can re-enable suggestions at any time:

• Say "be proactive", "turn on tips", or "enable suggestions"
• Manually edit ~/.claude-octopus/preferences.json and set OCTO_PROACTIVE_SUGGESTIONS to on
• Default state (no preference set) is suggestions enabled

Detection Signals

Detect work stage from:

• Recent tool usage (many Bash calls = likely implementing/debugging)
• File types being edited (.test.ts = testing, .md = documentation)
• Error patterns in recent output (stack traces = debugging)
• Git state (uncommitted changes = implementing, clean tree = ready to review/ship)