context-mgmt

Reasoning fidelity is the real metric, not token count. Quality degrades at ~60-70% fill, not at the limit. </purpose>

<core_principles>

1. COMPACTION: Summarize and reinitialize. Keep architecture decisions, discard noise.
2. STRUCTURED NOTES: AgentDB + active.md persist across context resets.
3. MULTI-AGENT: Delegate research to subagents. They explore, report summaries. Cap the return at ~1,000-2,000 tokens (conclusion + path:line evidence, no raw dumps); the exploration context stays in the subagent.
4. MINIMAL READS: grep/glob to find, then read specific sections.
5. TOOL RESULT CLEARING: Old tool output rarely needs to stay in context. When calling the Claude API directly (not via this harness), prefer the server-side context-management-2025-06-27 beta over hand-rolled truncation: clear_tool_uses_20250919 clears old tool results past a token trigger (client keeps full history, no sync needed), clear_thinking_20251015 clears old thinking blocks while preserving cache hits. Client-side SDK compaction (compaction_control, TypeScript/Ruby) is deprecated; use server-side compact_20260112 for summarization, which takes its own beta header compact-2026-01-12, not context-management-2025-06-27: trigger is input_tokens only (default 150000, minimum 50000), and custom instructions replace the default summarization prompt entirely, so restate the full framing (what to keep: goal, decisions, paths, errors) rather than appending. pause_after_compaction: true returns stop_reason: "compaction" with only the compaction block, so you can re-append recent messages or key instructions, then resend without the flag to continue. Ordering: clear_thinking_20251015 must precede clear_tool_uses_20250919 in edits. Set exclude_tools for results that must survive; clearing tool results invalidates the cached prefix, so set clear_at_least so each clear pays for itself. Pair with the memory_20250818 tool: Claude is warned as the threshold nears and saves state to memory files before results clear. clear_tool_inputs (default false) also drops call parameters. clear_thinking_20251015 keep defaults are model-specific (newer models keep all turns, older keep the last), so set it explicitly across model tiers. On Fable 5.1/Opus 5.5/Sonnet 5.5/Haiku 5.5, server-side edits never invalidate thinking blocks, but hand-editing earlier turns client-side can; for accounts created on/after 2026-08-31 a replayed invalidated block is rejected unless you opt into dropping it. Verify clearing fired from context_management.applied_edits in the response (cleared_tool_uses / cleared_thinking_turns and cleared_input_tokens; in streaming, the final message_delta). Preview savings by passing context_management to count_tokens: it returns original_input_tokens beside the post-clearing input_tokens. </core_principles>

<token_budget>

• CLAUDE.md: <150 lines (always loaded)
• rules/: <100 lines (always loaded)
• Skills: metadata only at startup, full content on demand
• Commands: only when invoked
• Reference docs: only when skill explicitly reads them </token_budget>

<compaction_protocol>

1. CHECK fill level via /context. (gate: <60% → continue; ≥60% → proceed to step 2)
2. Write critical state: agentdb write-end '{"did":"X","next":"Y","blocked":"Z"}'
3. Commit any uncommitted changes. (gate: clean working tree)
4. Generate handoff if complex work in progress: /kernel:handoff
5. Trigger compaction.
6. POST-COMPACT RESTORE: a. Read active.md for project context. b. Run agentdb read-start for failures, patterns, contracts. c. Check for pending checkpoints to review. </compaction_protocol>

<fidelity_signals> Watch for reasoning degradation BEFORE the token meter shows a problem:

• Hypothesis depth dropping: agent defaults to first idea, stops exploring alternatives
• Backtracking absent: errors patched in place instead of root-caused
• Step count contracting: multi-step plans collapse to one-shot attempts
• Cross-file awareness fading: agent forgets earlier files in the same session
• Inline checks disappearing: agent stops verifying assumptions before acting

Any one of these → compact or hand off NOW, even if /context shows <60% fill. </fidelity_signals>

<what_to_preserve> Never compact away:

• Current task context and goal
• Active decisions and their rationale
• Blocking issues and error states
• File paths currently being worked on
• Uncommitted changes description
• Contract IDs and branch names </what_to_preserve>

<what_to_compact> Aggressively compress:

• Exploratory searches → "searched X, found Y at path:line"
• File reads → "read file, key insight: Z"
• Successful operations → "completed: X"
• Debugging traces → "root cause: X, fix: Y"
• Tool call history → "N tool calls, outcome: X"
• Research context → offload to AgentDB, keep one-line summary </what_to_compact>

<failure_modes>

1. Agent tries too much at once → context exhausted mid-work Fix: Incremental progress with commits at each step.
2. Later agent assumes done prematurely → work incomplete Fix: Explicit done-when criteria with evidence required. </failure_modes>