compact-optimizer

Compact Optimizer

One job: build the smallest possible hint that tells /compact exactly what must survive.

Research Grounding

Paper	Core Rule
Selective Context (EMNLP 2023)	Drop what is reconstructable from general knowledge. Preserve what is unique to this session.
LLMLingua (EMNLP 2023)	Compress reasoning/exploration 60–80%. Compress instructions/constraints ≤20%.
H2O MIT (NeurIPS 2023)	Always retain the last user correction + last known file state — they are attention sinks.

---

Survival Priority Hierarchy

Rank 1 — Never drop (session-unique, irreplaceable):

• Active branch name and working directory
• Files modified in this session (paths + what changed)
• User corrections, explicit rejections, aversion decisions
• Definition of done / success criteria
• Last error message or test failure state
• Approved plan steps and which are complete

Rank 2 — Compress heavily (preserve meaning, strip words):

• Current task description (1 sentence max)
• Key constraints stated by user
• Architecture decisions made

Rank 3 — Drop entirely (reconstructable, low self-information):

• Exploration text ("let me look at...", "I'll check...")
• Preambles and acknowledgments
• Reasoning narration that led to a decision already made
• Redundant tool call descriptions (keep results, drop scaffolding)
• Any content the model could reproduce from the codebase alone

---

Compact Hint Template

/compact Preserve: [branch/dir] | Files touched: [path1, path2] | Last correction: [exact user correction] | Current task: [1 sentence] | Done criteria: [condition] | Status: [step N of M complete, next: X]

Rules for the hint itself:

• Max 3 sentences. No preamble.
• Lead with the file state — it is always the highest-attention-sink.
• Include the last user correction verbatim if one exists.
• Name the next concrete action.
• Never include reasoning or exploration in the hint.

---

Per-Phase Templates

Research phase

/compact Preserve: research goal=[goal], key finding=[finding], next: [what to research next]

Implementation phase (most common)

/compact Preserve: branch=[branch], files=[paths], constraint=[constraint], last correction=[correction], next step=[action], done when=[criteria]

Review / debug phase

/compact Preserve: error=[error msg], file=[path:line], hypothesis=[current], last tried=[approach], next: [next attempt]

---

When to Call /compact

• Context bar >60% (tokenmaxxing mode) or >65% (standard mode): run /compact with hint — thresholds match the tier table in pre-tool-use.js getTier()
• Phase change (research→implement, implement→review): always /compact with phase-transition hint
• After a user correction: /compact immediately, include correction verbatim
• Never compact mid-tool-use or mid-plan without noting the interruption point

---

Anti-patterns (never do these)

• /compact with no hint → generic summary, drops constraints
• Hint longer than 3 sentences → defeats compression purpose
• Including reasoning in the hint → low self-information, wastes hint budget
• Compacting after a failed approach without writing an Aversion Memory node first

---

Aversion Memory Gate

Per the YURI Aversion Memory Protocol: if you are compacting after a failed branch, first write the failure reason to an Aversion Memory node, THEN compact. The compact hint should reference the aversion: last aversion: [reason].

Token Budget Policy

• Trigger /compact at 60%+ (tokenmaxxing) / 65%+ (standard) context. A long transcript (~40k lines) is a guideline for manual /compact — no hook enforces a transcript-line limit.
• Dirty repo + mid-sprint: capture scoped git status --short marker before compacting.
• Lane output arriving in main context: summarize to marker-only before injection.
• After /compact, re-read 2–3 key touched files to verify compact summary accuracy.

Evidence Pack Shape

Canonical compact payload for DeepSeek reinforcement and lane handoff:

TASK: <one sentence>
CONTEXT_PACK:
- fact1
- fact2  (max 5 facts)
EVIDENCE:
- ref/snippet1
- ref/snippet2  (max 3 refs)
BLOCKERS: <none | specific>
QUESTION: <specific question or "proceed">
OUTPUT_CAP: <80 lines research | 120 lines final report>

Never inject raw command dumps, full file contents, or unfiltered grep output into a compact payload.

Reversible Path (ccr-compress)

This skill builds a ONE-WAY hint — the tokens it tells /compact to drop are gone. When a section is too valuable to lose-permanently but too heavy to keep inline, route it through the reversible compaction mechanism instead:

• _SYSTEM/Scripts/ccr-compress.mjs — compress(payload) caches the BYTE-EXACT original under a content hash (sibling dir _SYSTEM/state/ccr-cache/, TTL-pruned, gitignored) and injects a retrieval sentinel (⟪CCR:<hash>:<type>:<bytes>⟫) in its place. retrieve(hash) pulls the original back. The cache is the undo button — that is what makes dropping a section safe.
• Content-typed + honest about loss: json/code/prose routing. Structural elision is lossless (round-trips byte-exact). Semantic elision is marked lossy:true — the cache still restores the source, but the inline placeholder is not self-reconstructable. Never claim semantic inline is reversible.
• Cache-prefix conscience: _SYSTEM/Scripts/cache-prefix-scan.mjs is a WARN-ONLY detector — it flags volatile tokens (UUID / ISO-timestamp / JWT / long-hex / epoch-ms) leaking into the KV-cache-hot prefix and NEVER mutates the prefix. Pairs with the "Token Caching Shape" rule in CLAUDE.md (keep volatile ids out of the stable preamble). Lane label: 04CP_CACHE_PREFIX_SCAN_X_COMMITTED.
• --self is hardened + diagnose-only: reads ONLY global.md + MEMORY.md, never invokes brain-inject (which reads the deny-listed .claude/state/cortex-state.json). It reports self-context headroom; it does not repair context.

Reach for ccr-compress when the answer to "can I get this back if I need it?" must be yes. Use the plain /compact hint above when the dropped content is genuinely reconstructable from the codebase.

Session Notes

2026-06-16

• session: 7m | peak ctx: 0% | compacts: 0
• tools: Bash×9, Read×6, Edit×5, WebSearch×3, Write×1, Skill×1, ToolSearch×1
• corrections: none
• errors: none

2026-06-15

• session: 11m | peak ctx: 0% | compacts: 0
• tools: Bash×19, Read×2, Edit×2, Skill×1, Write×1
• corrections: Base directory for this skill: /Users/marcelspatz/.claude/skills/cross-reference-navigation

Cross-Reference Navigation (XREF)

The GROUND step of the work loop, made reflexive. One question asked a | Base directory for this skill: /Users/marcelspatz/.claude/skills/quantum-hypothesis-simulation

Quantum Hypothesis Simulation

The quantum-probability layer for YURI's claim/pulse machinery. It mode | Base directory for this skill: /Users/marcelspatz/.claude/skills/cross-reference-navigation

Cross-Reference Navigation (XREF)

The GROUND step of the work loop, made reflexive. One question asked a

• errors: none

2026-06-14

• session: 101m | peak ctx: 0% | compacts: 0
• tools: Bash×455, Read×142, Edit×74, Write×44, WebSearch×31, WebFetch×19, StructuredOutput×17, ToolSearch×8, TodoWrite×5, Workflow×3, AskUserQuestion×2, Skill×2, TaskStop×1
• corrections: Base directory for this skill: /Users/marcelspatz/.claude/skills/cross-reference-navigation

Cross-Reference Navigation (XREF)

The GROUND step of the work loop, made reflexive. One question asked a | Base directory for this skill: /Users/marcelspatz/.claude/skills/quantum-hypothesis-simulation

Quantum Hypothesis Simulation

The quantum-probability layer for YURI's claim/pulse machinery. It mode | Base directory for this skill: /Users/marcelspatz/.claude/skills/cross-reference-navigation

Cross-Reference Navigation (XREF)

The GROUND step of the work loop, made reflexive. One question asked a

• errors: none

2026-06-13

• session: 9m | peak ctx: 0% | compacts: 0
• tools: Bash×45, Write×16, Edit×13, Read×10, TaskStop×1, Skill×1
• corrections: rick i have a fun little task for you. I will be giving you the task of going through trending repos on github, scanning them, compare yuri to those, see what we can adopt and rebuild better in yuri u
• errors: none

2026-06-12

• session: 97m | peak ctx: 100% | compacts: 3
• tools: Edit×67, Bash×47, Read×42, Write×5, WebFetch×3, Skill×3, TaskUpdate×2, WebSearch×2, TaskCreate×1, EnterPlanMode×1
• corrections: none
• errors: none

2026-06-11

• session: 70m | peak ctx: 100% | compacts: 3
• tools: Read×34, Edit×30, Bash×19, Write×5, WebFetch×3, Skill×3, TaskUpdate×2, WebSearch×2, TaskCreate×1, EnterPlanMode×1
• corrections: none
• errors: none

2026-06-10

• session: 847m | peak ctx: 62% | compacts: 1
• tools: Bash×776, Read×365, Edit×55, Write×27, StructuredOutput×12, TodoWrite×11, WebFetch×9, Agent×7, Workflow×3, ToolSearch×2
• corrections: none
• errors: none

2026-06-04

• session: 90m | peak ctx: 0% | compacts: 0
• tools: Bash×452, Read×180, StructuredOutput×34, WebSearch×19, Edit×13, Write×10, TodoWrite×8, WebFetch×7, ToolSearch×6, Workflow×4, mcp×1
• corrections: none
• errors: none

2026-06-03

• session: 147m | peak ctx: 0% | compacts: 0
• tools: Bash×798, Read×626, Write×163, StructuredOutput×140, WebSearch×84, Edit×75, ToolSearch×39, WebFetch×16, Workflow×7, AskUserQuestion×1, Agent×1
• corrections: none
• errors: none

2026-05-02

• session: 4m | peak ctx: 14% | compacts: 0
• tools: Bash×16, Read×4, Edit×4, Skill×1
• corrections: none
• errors: none

2026-04-27

• session: 1m | peak ctx: 43% | compacts: 0
• tools: Bash×15, Read×12
• corrections: none
• errors: none

2026-04-26

• session: 7m | peak ctx: 0% | compacts: 0
• tools: Bash×15, Read×9, Write×4, Agent×1, ToolSearch×1, ExitPlanMode×1, Edit×1
• corrections: none
• errors: none

2026-04-25

• session: 0m | peak ctx: 14% | compacts: 0
• tools: Read×9, Bash×4, Write×2, Edit×2
• corrections: none
• errors: none

2026-04-24

• session: 0m | peak ctx: 21% | compacts: 0 (mid-session)
• tools: Bash×3, Read×1
• corrections: none
• errors: exit code 1: command not found