[BUG] "Clear Context" No longer removes old context fully

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I executed a /clear command immediately preceding a query regarding an issue (as I've done numerous times before), and despite the clear, was at nearly 80% context usage nearly immediately. When I asked Claude about why I would be suddenly at such a high level, the response is below - of which I would have expected points 2 & 3 to have not been included (ie the insight report and the associated plan):
A few things stacked up even after the clear:

System prompt is never cleared — CLAUDE.md plus the three @-referenced docs (collaboration.md, architecture.md, dev-standards.md) are injected on every turn. They're substantial.

The /insights skill output was carried forward — the system-reminder marked it as "skill invoked earlier in this session" and re-injected a large JSON payload into this conversation even though you had cleared. That alone was probably several thousand tokens.

The previous plan file content — plan mode injected the old plan (the long one covering items #1 and #38) via system-reminder at the start of the turn.

The Explore agent response was verbose — it returned full file contents for five files in one block.

Multiple file reads — dependencies.py, invoices.py, auth.py, test_pdf_endpoint.py all read into context in addition to the agent output.

The /clear clears your conversation history but the runtime continues re-injecting system-level context (CLAUDE.md, memory, skill state, active plan) on every turn. The practical fix for future sessions: after finishing a large workflow like the BACKLOG run, start a genuinely fresh session rather than clearing mid-session, which avoids the carried-over skill/plan reminders.

What Should Happen?

Clear should reset the context instead of maintaining old context (that's what compact is for).

Error Messages/Logs

Steps to Reproduce

1. run /insights
2. run following to generate BACKLOG.md: Spawn 6 parallel audit subagents over my repo, each with a single focus: (1) tenant_id isolation correctness, (2) circuit breaker and decorator ordering across the codebase, (3) test coverage gaps in service layer, (4) DB migration safety, (5) frontend/backend contract drift, (6) dead code and unused exports. Each agent writes findings to audits/<name>.md with severity, file:line refs, and a suggested fix. Then a synthesis agent merges them into BACKLOG.md ranked by risk × effort, with auto-drafted implementation prompts I can paste back to you.
3. have Claude Code address the issues one at a time from the BACKLOG.md (mine was ~46 items)
4. manually compact along the way
5. once finished run clear
6. ask Claude to investigate an issue, see the context is not actually reset

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.78

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_