[Bug]: Conductor parks workflow prematurely claiming 'context is heavy' at 37% usage (63% free)

[tejasavora]

Description

The AI-DLC conductor (the orchestrator persona running the forwarding loop) repeatedly claims "context is getting heavy" or "context is heavy" and parks the workflow at inter-stage boundaries, forcing the user to resume in a new session. However, /context shows only 37% of the 1M context window is used — 62.9% (629k tokens) remains free.

This creates an extremely disruptive experience: the user must /aidlc --resume every 1-2 stages, losing flow and wasting time on the unpark/resume handshake.

Steps to Reproduce

1. Start an enterprise-scope workflow (32 stages) on Claude Code with Opus 4.6 (1M context)
2. Complete stages normally through ideation + inception
3. Observe the conductor adding text like "Context is getting long — let me park here" or "Parking here — context is heavy" after each stage approval
4. Run /context immediately after parking
5. Observe: only 37% used, 63% free

Expected Behavior

The conductor should NOT park due to context pressure when >50% of the context window remains free. Parking for context should only happen when genuinely approaching limits (e.g., >85% usage). The conductor should continue running stages until either:

• The user asks to stop
• Context genuinely approaches capacity (>85%)
• The workflow reaches a natural done

Actual Behavior

The conductor parks after nearly every stage with messages like:

• "Context is getting heavy — let me park here"
• "Parking here — context is heavy and the next stage is substantial"
• "Context is getting long, let me park at this clean boundary"

Meanwhile /context shows 370.9k/1M (37% used). The conductor has NO mechanism to check actual context usage — it's guessing based on conversation length, not actual token count.

AI-DLC Version

v2 (alpha)

Release / Commit

Latest (installed via .claude/ workspace shell, July 2026)

AI-DLC Phase

Not phase-specific — happens across Ideation and Inception stages equally.

Platform / IDE

Claude Code

AI Model

global.anthropic.claude-opus-4-6-v1[1m] (Opus 4.6, 1M context)

Environment

• OS: macOS Sequoia (Darwin 25.5.0)
• Shell: zsh
• Terminal: Warp
• Claude Code: latest

Additional Context

Screenshot showing the behavior:

• Conductor says "Requirements approved. Parking here — context is heavy and User Stories (next) is another substantial stage with a reviewer loop."
• Immediately after, /context shows 370.9k/1M tokens (37%) with 629.1k free space

The root cause appears to be that the conductor persona has NO access to actual context metrics. It's making a judgment call based on perceived conversation length rather than querying the real token count. The SKILL.md instructs to park "when you are running low on context mid-loop" but provides no mechanism to CHECK context — the conductor is guessing conservatively and wrong.

Possible fixes:

1. Expose a /context-equivalent API the conductor can call before deciding to park
2. Add a guardrail: "never park for context reasons until you've been explicitly told context is >80%"
3. Remove the conductor's discretion to park for context entirely — only park when the user asks or the system actually compresses