Long tool-heavy session: instruction/command recognition drifts, formula-before-reference, coarse collision checks
Context
Long working session in Claude Code (Sonnet 5) driving a Figma MCP server (figma-console) through many sequential figma_execute tool calls to rebuild a design-system component's geometry to match a hand-built reference frame the user provided.
What happened
After a long chain of tool calls (dozens of figma_execute round-trips across a compacted/summarized session), the user had to explicitly say "ты тут даже команды не распознаешь уже" (you're not even recognizing commands anymore) — i.e. my behavior had drifted from her actual latest instruction, likely because I was continuing an internal task queue built up over many turns rather than re-grounding in her most recent literal message after each phase of tool-heavy work.
Related failure modes from the same session, self-diagnosed and filed as feedback:
- Given a hand-built reference example as ground truth ("rebuild to match this exact frame"), I first inferred a general geometric rule and applied it, instead of reading the reference's literal per-instance values and copying them — produced 2-3 wasted rebuild iterations before I extracted exact numbers.
- A geometric-overlap check used a composited instance's full bounding box instead of its actual visible leaf sub-nodes, producing 8 false-positive collision findings before being corrected to test real geometry.
- A Figma plugin API
resize()/resizeWithoutConstraints()call silently no-ops on one specific node with no error and no state change, cause undiagnosed within the session (separate from the above, filed for completeness — may be a figma-console MCP / Figma plugin API interaction rather than a Claude Code issue per se).
Ask
No specific repro steps to hand over (this is qualitative session-quality feedback, not a crash), but wanted to flag: after long tool-call chains (especially spanning a context-compaction boundary), the model's grounding in the user's actual latest message appears to degrade even though context isn't technically lost. If there's a known mitigation (more aggressive re-anchoring after compaction, some existing setting), pointers welcome.
Filed at the user's explicit request, as a self-reported bug rather than a user complaint.