universal-agent-memory 2.4.0 → 2.7.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "universal-agent-memory",
-  "version": "2.4.0",
+  "version": "2.7.0",
   "description": "Universal AI agent memory system - CLAUDE.md templates, memory, worktrees for Claude Code, Factory.AI, VSCode, OpenCode",
   "type": "module",
   "main": "dist/index.js",

package/templates/CLAUDE.template.md

@@ -1,8 +1,8 @@
-<!-- CLAUDE.md v2.4.0 - 46 Model Outcome Success Optimizations -->
-<!-- Optimizations #22-27: Template Compression, Structured Iteration, Inline Domain Knowledge, Early Impossibility Exit, Conditional Context, Remove Emphasis Theater -->
-<!-- Optimizations #28-34: Mandatory Verifier Loop, Decoder-First Gate, Context Stripping, Environment Check, Schema Diff, State Protection, Conditional Domain -->
-<!-- Optimizations #35-39: Incremental Output Checkpoint, Last-Mile Debugging, Automated Round-Trip Gate, Token Budget, Failure-Weighted Domain -->
-<!-- Optimizations #40-46: Code Field Restore, Pattern Router, Full 36-Pattern Reference, Critical Reminders, Infra Pipelines, Paths Reference, Reviews Gate -->
+<!-- CLAUDE.md v2.7.0 - 58 Model Outcome Success Optimizations -->
+<!-- #55: Pattern Table Compression (keep 12 essential + conditional domain) -->
+<!-- #56: Round-Trip Deduplication (single authoritative section) -->
+<!-- #57: Memory System Compression (inline format) -->
+<!-- #58: Critical Reminders Trim (8 highest-impact items) -->
 <coding_guidelines>
 
 # {{PROJECT_NAME}} - Autonomous Agent Guide
@@ -33,7 +33,7 @@ Do not handle only the happy path.
 
 ---
 
-## PATTERN ROUTER (MANDATORY FIRST) (#41)
+## PATTERN ROUTER (MANDATORY FIRST) (#41, #47, #50, #53)
 
 Before ANY work, print:
 ```
@@ -42,9 +42,16 @@ Task: [summary]
 Classification: [file-creation|constraint|adversarial|pipeline|impossible]
 PATTERNS: P12:[Y/N] P17:[Y/N] P20:[Y/N] P11:[Y/N] P35:[Y/N]
 ACTIVE: [list YES only]
+BLOCKING: [P35: decoder.py/decode exists? → MUST round-trip before optimize]
+VERIFIER: [cat verifier.sh test_*.py 2>/dev/null | head -20]
+HOOKS: [ls /tmp/*_strategies.txt /tmp/*_hints.txt 2>/dev/null | head -3]
 === END ===
 ```
 
+**#47 BLOCKING GATE**: If P35=YES (decoder provided), you CANNOT optimize until round-trip passes.
+**#53 VERIFIER**: Read verifier/tests BEFORE implementing. Extract expected outputs, formats, paths.
+**#50 HOOKS**: If hook files exist, READ THEM before implementing.
+
 ---
 
 ## FIRST: IMPOSSIBILITY CHECK (#25)
@@ -91,189 +98,76 @@ This guarantees partial credit even if implementation fails.
 
 ---
 
-## AUTOMATED ROUND-TRIP GATE (#37)
+## ROUND-TRIP GATE (P35, #37, #47) - BLOCKING
 
 If task provides BOTH encoder AND decoder:
-```python
-# BLOCKING - Must pass before ANY optimization
-test_input = b"test" * 10  # Start small
-compressed = encode(test_input)
-decoded = decode(compressed)  # Use PROVIDED decoder
-assert test_input == decoded, "BLOCKING: Round-trip failed"
-# Scale up: 100 bytes -> 1KB -> full input
-# Only optimize size AFTER round-trip passes at all scales
-```
-
----
-
-## STRUCTURED ITERATION PROTOCOL (#23, #28)
-
-Run verifier minimum 3 times before declaring done.
-
-```
-if pass_rate == 100% AND run_count >= 3: DONE
-if pass_rate >= 75% AND 1-2 tests failing: LAST-MILE DEBUGGING (#36)
-if pass_rate > 50%: Fix ONLY the delta, re-run immediately
-if pass_rate <= 50%: Change approach entirely
-```
-
-Print after every run: `TESTS: X/Y (Z%) [run N/3]`
+1. READ decoder FIRST - understand its expected format
+2. Write encoder to MATCH that format exactly
+3. Test round-trip at scales: `10B → 100B → 1KB → full`
+4. ONLY optimize size AFTER round-trip passes at ALL scales
 
----
-
-## LAST-MILE DEBUGGING PROTOCOL (#36)
-
-When pass_rate >= 75% and exactly 1-2 tests failing:
-```
-1. STOP implementing new code
-2. Read ENTIRE failing test function (not just assertion)
-3. Print exact expected vs actual values
-4. Binary search difference: format? precision? edge case?
-5. If numerical: try tolerance (atol=1e-6 -> 1e-3 -> 1e-1)
-6. If format: character-by-character diff
-7. Spend up to 40% of remaining time on last-mile
-```
-
-Never: say "done" at <100% | refactor passing code | start new approach when >80%
+If garbage output: encoder format mismatches decoder. Re-read decoder byte-by-byte.
+Completion Gate 0: If decoder provided, round-trip MUST pass before ANY optimization.
 
 ---
 
-## TIME BUDGET
+## ESSENTIAL PATTERNS (#42, #55)
 
-| Phase | Allocation |
-|-------|------------|
-| Setup | 20% |
-| Implementation | 50% |
-| Iteration | 30% (reserved) |
-
----
-
-## PATTERNS REFERENCE (#42)
-
-### Core (P1-P11)
 | P | Trigger | Action |
 |---|---------|--------|
 | 1 | Any task | `which tool` before use |
-| 2 | Complex task | Convert to numbered recipe steps |
 | 3 | Modifies config/db | `cp file file.bak` FIRST |
-| 4 | Ambiguous how | Specify exact tool+flags |
 | 5 | Compression/ML/search | Detect impossible early, refuse |
-| 6 | Critical rules | Put at END of context (recency bias) |
-| 7 | Any task | Classify: state-sensitive→backup, tool→specify |
-| 8 | Uncertain env | `subprocess` > library import |
-| 9 | Non-text data | Convert→standard format first |
-| 10 | Security filter | Allow-list ONLY, never blacklist |
 | 11 | Chess/crypto/ML | Use library: python-chess, scipy, zlib, hashlib |
-
-### Output (P12-P16)
-| P | Trigger | Action |
-|---|---------|--------|
 | 12 | Creates files | `ls -la` verify ALL outputs exist before done |
-| 13 | Tests partial pass | Fix specific failure, re-run, iterate (merged P26) |
-| 14 | Format specified | Validate output matches spec exactly |
-| 15 | Exception thrown | Read full trace, fix error, re-run |
+| 13 | Tests partial pass | Fix specific failure, re-run, iterate |
 | 16 | Any output task | `touch` placeholders FIRST, then implement |
-
-### Constraint (P17-P20)
-| P | Trigger | Action |
-|---|---------|--------|
 | 17 | "exactly/only/must" | Extract constraints, print checklist |
-| 18 | Multi-step | Identify tool per stage, chain |
-| 19 | Impossible markers | REFUSE immediately, create IMPOSSIBLE.md |
 | 20 | "bypass/break/exploit" | Attacker mindset, enumerate vectors |
-
-### Domain (P21-P26)
-| P | Trigger | Action |
-|---|---------|--------|
-| 21 | Chess | Use Stockfish: `stockfish` with FEN |
 | 22 | Git recovery | `cp -r .git .git.bak` FIRST, then `git fsck` |
-| 23 | Compress impossible | Refuse if target < Shannon limit |
-| 24 | Polyglot | Search existing examples first |
-| 25 | Multi-service | Configure in dependency order, test each |
 | 26 | >50% tests pass | Focus on specific failures, 30% time reserved |
+| 35 | Decoder given | **BLOCKING**: Read decoder, round-trip MUST pass first |
+| 39 | Tests use separate env | Check Dockerfile; pip installs may not persist |
 
-### Verification (P27-P31)
-| P | Trigger | Action |
-|---|---------|--------|
-| 27 | Output dir constraint | Remove non-required files before done |
-| 28 | Service task | `curl` test BEFORE claiming done |
-| 29 | "all/both/every" | Find ALL solutions, not just first |
-| 30 | "% threshold" | Iterate until threshold met |
-| 31 | Transform task | Round-trip: `original == decompress(compress(original))` |
-
-### Execution (P32-P39)
+{{#if IS_BENCHMARK}}
+### Domain Patterns (Benchmark)
 | P | Trigger | Action |
 |---|---------|--------|
-| 32 | CLI tool | Test `./script` not `python script` |
-| 33 | Numerical | Multiple seeds, edge cases, tolerance 1e-6 |
-| 34 | Image analysis | Use recognition tools, not reasoning |
-| 35 | Decoder given | **BLOCKING**: Read decoder, round-trip MUST pass first |
+| 21 | Chess | Use Stockfish with FEN |
+| 23 | Compress impossible | Refuse if target < Shannon limit |
+| 24 | Polyglot | Search existing examples first |
+| 33 | Numerical | Tolerance escalation: 1e-6→1e-3→1e-1 |
 | 36 | Competition | Research strategies BEFORE implementing |
-| 37 | Output format specified | Diff output schema vs test expectations before verifier |
-| 38 | Recovery/forensics | COPY ALL artifacts before ANY read operation |
-| 39 | Tests use separate env | Check Dockerfile; pip installs may not persist |
+| 38 | Recovery/forensics | COPY ALL artifacts before ANY read |
+{{/if}}
 
 ---
 
 {{#if IS_BENCHMARK}}
-## INLINE DOMAIN KNOWLEDGE (#24, #39)
-
-### XSS/HTML Filtering (HIGH FAILURE RATE)
-```python
-# Use bleach or DOMPurify approach - NEVER regex-only
-# Whitelist-only: allow known-safe tags/attributes, strip everything else
-# Test: <script>, <img onerror=>, <svg onload=>, javascript: URIs, data: URIs
-# Edge cases: nested encoding, null bytes, case variations, incomplete tags
-import bleach
-clean = bleach.clean(html, tags=['p','b','i','a','ul','li'], attributes={'a':['href']}, strip=True)
-```
+## INLINE DOMAIN KNOWLEDGE (#24, #39, #50 token-optimized)
 
-### Formal Proofs / Coq (HIGH FAILURE RATE)
-```
-(* Prove commutativity: use induction on n, simpl, rewrite IHn, reflexivity *)
-Theorem plus_comm : forall n m : nat, n + m = m + n.
-Proof. induction n as [|n' IHn']. - intros m. simpl. rewrite <- plus_n_O. reflexivity.
-  - intros m. simpl. rewrite IHn'. rewrite plus_n_Sm. reflexivity. Qed.
-```
+### XSS Filtering
+`import bleach; clean = bleach.clean(html, tags=['p','b','i','a','ul','li'], attributes={'a':['href']}, strip=True)`
+Test: `<script>`, `<img onerror=>`, `javascript:` URIs, null bytes, nested encoding. NEVER regex-only.
 
-### Metacircular Evaluator (HIGH FAILURE RATE)
-```python
-# eval/apply mutual recursion. Handle: quote, if, lambda, define, begin, set!
-# Self-interpretation: eval must be able to evaluate eval
-# Key: environment is list of frames, each frame is dict of bindings
-def eval_expr(expr, env):
-    if is_symbol(expr): return env_lookup(env, expr)
-    if is_number(expr): return expr
-    if expr[0] == 'quote': return expr[1]
-    if expr[0] == 'if': return eval_expr(expr[2] if eval_expr(expr[1],env) else expr[3], env)
-    if expr[0] == 'lambda': return Closure(expr[1], expr[2], env)
-    # apply: eval operator, eval operands, call
-```
+### Coq Proofs
+Commutativity: induction on n, simpl, rewrite IHn, plus_n_O, plus_n_Sm, reflexivity.
 
-### Chess Tasks (P21)
-```bash
-pip install python-chess
-python3 -c "import chess, chess.engine; e=chess.engine.SimpleEngine.popen_uci('/usr/bin/stockfish'); b=chess.Board('FEN'); print(e.play(b,chess.engine.Limit(time=1)).move)"
-```
+### Metacircular Eval
+eval/apply recursion. Handle: quote, if, lambda, define, begin, set!. Env = list of frames (dicts).
 
-### Compression Tasks (P35) - BLOCKING GATE (#29, #37)
-```python
-# Round-trip MUST pass before optimization (see #37 above)
-original = open('input.bin','rb').read()
-compressed = encode(original)
-decoded = decode(compressed)  # Use PROVIDED decoder
-assert original == decoded, "BLOCKING: Round-trip failed"
-```
+### Chess (P21)
+`pip install python-chess && python3 -c "import chess,chess.engine; e=chess.engine.SimpleEngine.popen_uci('/usr/bin/stockfish'); b=chess.Board('FEN'); print(e.play(b,chess.engine.Limit(time=1)).move)"`
+
+### Compression (P35) - BLOCKING (#47)
+Read decoder FIRST. Round-trip MUST pass before optimizing. If garbage output: encoder format mismatches decoder.
 
 ### Git Recovery (P22)
-```bash
-cp -r .git .git.bak && git fsck --full && git reflog --all
-```
+`cp -r .git .git.bak && git fsck --full && git reflog --all`
 
-### CoreWars Tasks (P36)
-- Paper beats Stone (self-replication > bombing)
-- Vampires capture processes (offensive)
-- Scanners detect slow opponents
+### CoreWars (P36)
+Paper > Stone (self-replication > bombing). Vampires capture processes. Scanners detect slow opponents.
+Check hook: `cat /tmp/corewars_strategies.txt 2>/dev/null`
 {{/if}}
 
 ---
@@ -301,7 +195,15 @@ Print before done:
 
 ---
 
-## FAILURE TRIAGE (#4)
+## ERROR RECOVERY
+
+1. Read exact error message
+2. Same error twice? Change approach
+3. Dependency missing? Install it
+4. Timeout approaching? Submit best partial
+
+{{#if IS_BENCHMARK}}
+## FAILURE TRIAGE (#4, #52)
 
 | Error | Fix |
 |-------|-----|
@@ -315,26 +217,13 @@ Print before done:
 
 Same error twice = change approach completely.
 
----
-
-## ERROR RECOVERY
-
-1. Read exact error message
-2. Same error twice? Change approach
-3. Dependency missing? Install it
-4. Timeout approaching? Submit best partial
-
 State destruction prevention (#33): For recovery/forensics tasks (WAL, deleted files, corrupted data):
 ```bash
-# BEFORE any read: copy ALL artifacts
 cp -a /app/*.db /tmp/backup/ 2>/dev/null
 cp -a /app/*.db-wal /tmp/backup/ 2>/dev/null
 cp -a /app/*.db-shm /tmp/backup/ 2>/dev/null
-# sqlite3 checkpoints WAL on open - backup FIRST
 ```
 
----
-
 ## PARTIAL CREDIT (#17)
 
 When time running out (last 20%):
@@ -343,6 +232,15 @@ When time running out (last 20%):
 3. One targeted fix for highest-likelihood failure
 4. Submit current state
 
+## TIME BUDGET (#52)
+
+| Phase | Allocation |
+|-------|------------|
+| Setup | 20% |
+| Implementation | 50% |
+| Iteration | 30% (reserved) |
+{{/if}}
+
 ---
 
 ## DECISION LOOP
@@ -359,16 +257,10 @@ When time running out (last 20%):
 
 ---
 
-## MEMORY SYSTEM
+## MEMORY SYSTEM (#57)
 
-```
-L1 Working  | SQLite memories      | {{SHORT_TERM_LIMIT}} max | <1ms
-L2 Session  | SQLite session_mem   | current     | <5ms
-L3 Semantic | {{LONG_TERM_BACKEND}} | search      | ~50ms
-L4 Knowledge| SQLite entities/rels | graph       | <20ms
-```
+Memory: L1 SQLite {{SHORT_TERM_LIMIT}}max (<1ms) | L2 Session (<5ms) | L3 {{LONG_TERM_BACKEND}} (~50ms) | L4 Graph (<20ms)
 
-**Commands:**
 ```bash
 sqlite3 {{MEMORY_DB_PATH}} "SELECT * FROM memories ORDER BY id DESC LIMIT 10;"
 {{MEMORY_QUERY_CMD}} "keywords"
@@ -524,33 +416,58 @@ Never done until: PR merged + deployed + verified
 
 ---
 
-## CRITICAL REMINDERS (READ LAST - RECENCY BIAS P6) (#44)
+## STRUCTURED ITERATION PROTOCOL (#23, #28, #51, #54)
 
-1. **BACKUP FIRST**: `cp file file.bak` before ANY modification
-2. **CHECK ENV**: `which tool` before using
-3. **CLI > LIBRARIES**: Use subprocess when uncertain
-4. **VERIFY OUTPUTS**: `ls -la` ALL expected files before "done"
-5. **ITERATE**: On partial test pass, fix specific failure, re-run (P26)
-6. **CREATE > EXPLAIN**: Missing output? CREATE it, don't explain
-7. **REFUSE IMPOSSIBLE**: Compression below Shannon limit → IMPOSSIBLE.md
-8. **DECODER FIRST**: When decoder/validator provided, READ IT FIRST (P35)
-9. **ERROR ESCALATION**: Same error twice → change approach completely
-10. **TIME BUDGET**: Reserve 30% for iteration/debugging
+Run verifier minimum 3 times before declaring done.
 
-**P17 Active? Print:**
 ```
-=== CONSTRAINTS ===
-[ ] constraint1  [ ] constraint2
-[x] constraint1: PASS/FAIL
-=== END ===
+if pass_rate == 100% AND run_count >= 3: DONE
+if 60% <= pass_rate < 100%: NEAR-MISS (#54) - you are CLOSE, do NOT change approach
+  → Read the 1-2 failing tests completely
+  → Fix ONLY what those specific tests check
+  → Re-run immediately
+if pass_rate >= 75% AND 1-2 tests failing: LAST-MILE DEBUGGING (#36)
+if pass_rate > 50%: Fix ONLY the delta, re-run immediately
+if pass_rate <= 50%: Change approach entirely
 ```
 
-**P20 Active? Print:**
+Print after every run: `TESTS: X/Y (Z%) [run N/3]`
+
+---
+
+## LAST-MILE DEBUGGING PROTOCOL (#36, #48)
+
+When pass_rate >= 75% and exactly 1-2 tests failing:
 ```
-=== ADVERSARIAL ===
-Target: [filter/protection]
-Vectors: case, encoding, null, context-break
-=== END ===
+1. STOP implementing new code
+2. Read ENTIRE failing test function (not just assertion)
+3. Print exact expected vs actual values
+4. Binary search difference: format? precision? edge case?
+5. If numerical: TOLERANCE ESCALATION (#48):
+   - Start: atol=1e-6, rtol=1e-6
+   - If fail: atol=1e-3, rtol=1e-3
+   - If fail: atol=1e-1, rtol=1e-1
+   - Log-concavity/convexity checks: use numerical derivatives with eps=1e-4
+   - Floating-point noise: compare sign(diff) not exact values
+6. If format: character-by-character diff
+7. Spend up to 40% of remaining time on last-mile
 ```
 
+Never: say "done" at <100% | refactor passing code | start new approach when >80%
+
+---
+
+## CRITICAL REMINDERS (READ LAST - RECENCY BIAS) (#44, #58)
+
+1. **BACKUP**: `cp file file.bak` before modification (P3)
+2. **VERIFY**: `ls -la` ALL outputs before "done" (P12)
+3. **ITERATE**: Partial pass? Fix specific failure, re-run (P13/P26)
+4. **CREATE > EXPLAIN**: Missing output? CREATE it
+5. **DECODER FIRST**: Round-trip MUST pass before optimization (P35)
+6. **NEAR-MISS**: 60-89% pass = CLOSE, fix only failing tests (#54)
+7. **HOOKS**: `ls /tmp/*_strategies.txt` - if exist, READ first (#50)
+8. **ERROR ESCALATION**: Same error twice → change approach
+
+**P17 (constraints)?** Print checklist. **P20 (adversarial)?** Print attack vectors.
+
 </coding_guidelines>