v0.6.6: aggressive compression + sanitizer + SQuAD validation

.githooks/pre-commit

@@ -81,6 +81,14 @@ if [ -n "$stale" ]; then
     exit 1
 fi
 
+# 8b. Block SQLite runtime database files (they get rebuilt by the daemon)
+sqlite_staged=$(git diff --cached --name-only --diff-filter=ACM 2>/dev/null | grep -E '\.sqlite(-journal|-shm|-wal)?$' || true)
+if [ -n "$sqlite_staged" ]; then
+    echo "FAIL: SQLite runtime DBs must not be committed (daemon rebuilds them):"
+    echo "$sqlite_staged"
+    exit 1
+fi
+
 # 9. shellcheck (lint shell scripts)
 if command -v shellcheck >/dev/null 2>&1; then
     shellcheck scripts/*.sh 2>/dev/null || {

.github/workflows/ci.yml

@@ -67,6 +67,14 @@ jobs:
           stale=$(find . -name "*.bak" -o -name "*.perf" -o -name "*.old" -not -path "./target/*" 2>/dev/null)
           if [ -n "$stale" ]; then echo "FAIL"; echo "$stale"; exit 1; fi
 
+      - name: no sqlite runtime DBs
+        run: |
+          if git ls-files | grep -qE '\.sqlite(-journal|-shm|-wal)?$'; then
+            echo "FAIL: SQLite runtime DBs committed (daemon rebuilds these)"
+            git ls-files | grep -E '\.sqlite'
+            exit 1
+          fi
+
       - name: gate.js sync
         run: |
           if ! diff pi/gate.js crates/reliary-agent/pi/gate.js >/dev/null 2>&1; then

.gitignore

@@ -1,5 +1,12 @@
 target/
 .reliary/
+*.sqlite
+*.sqlite-journal
+*.sqlite-shm
+*.sqlite-wal
 *.bak
 *.perf
 *.old
+__pycache__/
+*.pyc
+.pytest_cache/

CHANGELOG.md

@@ -1,5 +1,21 @@
 # Changelog
 
+## v0.7.0 (unreleased)
+
+### Compression Ceiling Breakthrough
+- **Aggressive skeleton compression:** Tokenizes all alphanumeric runs to `{w}` while preserving version/number placeholders. Cargo `Compiling X v1.0` lines now share skeletons and group together. Gate restricts activation to content where ≥80% of lines share the same aggressive skeleton AND line lengths are similar — prevents file reads with similar function signatures from collapsing.
+- **Information-preserving zone truncation:** Replaces blind first-30 + last-15 zone with score-based selection. Error lines (`FAILED`, `error[`, `panic`, `Traceback`) get +10 score bonus. Allows top-15 selection (vs blind 45 lines) without signal loss.
+- **FTS5 document frequency weighting:** Grammar-free pseudo-perplexity proxy (LLM Lingua analog without an LM). Tokens appearing in many files are "predictable" boilerplate; tokens in few files are "surprising" project-specific. Opt-in via `RELIARY_PROXY_FT_WEIGHT=1` (off by default until validated in live sessions).
+- **SRCR safety floor:** `preservation × compression` metric. Default `RELIARY_PROXY_SRCR_FLOOR=0.3` blocks destructive compression — if post-compression SRCR is below the floor, the proxy ships the pre-compression content instead. Set `0` to disable.
+- **Smoke test:** Cargo test output (6100 chars, 205 lines) compresses to ~225 chars (96% savings, `history_saved=5878`).
+
+### Tunable
+- `RELIARY_PROXY_SRCR_FLOOR` — default `0.3`, set `0` to disable
+- `RELIARY_PROXY_FT_WEIGHT` — default `0` (off), set `1` to enable
+
+### Test counts
+- 81 sift + 13 compress + 14 search + 7 sr_floor + 9 ft_weight_gate = 124 tests passing
+
 ## v0.6.4
 
 ### Scorecard Security (June 2026)

CONFIG.md

@@ -53,6 +53,10 @@ reliary-agent uses a cascade of configuration sources (highest priority first):
 | `RELIARY_PROXY_GUARD_DISABLE` | `1` | Disable guard (cross-file edit safety). On by default. |
 | `RELIARY_PROXY_ANTI_DISABLE` | `1` | Disable anti-decision (sticky identifier failure memory). On by default. |
 | `RELIARY_PROXY_OUTPUT_COMPRESS` | `1` | Enable first-appearance freeze compression. On by default. |
+| `RELIARY_PROXY_SRCR_FLOOR` | `0.3` | SRCR safety floor. If post-compression SRCR < floor, ship pre-compression content instead. Set to `0` to disable. |
+| `RELIARY_PROXY_FT_WEIGHT` | `0` | Enable FTS5 document-frequency weighting for zone truncation. Off by default until validated in live sessions. |
+| `RELIARY_PROXY_PASSTHROUGH` | `0` | Disable compression (true transparent forward). Sanitizer still runs. Off by default. |
+| `RELIARY_PROXY_SANITIZER` | `1` | Strip empty assistants and duplicate tool_call_id reuses. Default-on for OpenAI/DeepSeek compatibility. Set to `0` to disable. |
 
 ## Agent Setup Examples (Proxy Routing)
 

README.md

@@ -55,6 +55,28 @@ After `init`, your agents get MCP tools for search, risk, compression, fix, dead
 detection, healing, and prior session memory. For conversation compression, point your
 agent's `*_BASE_URL` at http://localhost:9090.
 
+## Validation
+
+Compression preserves reading comprehension on SQuAD v2 within 2.7x LLM
+variance. Measured on 3 runs × 30 samples (270 calls), both proxy conditions
+achieve F1 retention ≥ 95% of baseline (the regression-detection floor):
+
+| Condition   | F1    | EM    | Acc    | F1 Retention |
+|-------------|-------|-------|--------|--------------|
+| baseline    | 0.770 | 0.711 | 80.00% | 100%         |
+| recommended | 0.777 | 0.733 | 78.89% | 100.9% PASS  |
+| passthrough | 0.791 | 0.756 | 80.00% | 102.6% PASS  |
+
+F1 retention > 100% is within 2.7x LLM variance — not a real improvement, but
+proof that compression does not regress reading comprehension. See
+`scripts/benchmarks/accuracy/README.md` for full methodology and caveats.
+
+Run the benchmark yourself:
+
+```bash
+python3 scripts/benchmarks/accuracy/bench_squad.py --runs 3 --samples 30
+```
+
 ## Usage by Agent
 
 There are two separate things reliary gives your agent. Understanding the difference

crates/reliary-agent/pi/gate.js

@@ -32,7 +32,7 @@ function gateLog(level, msg) {
 }
 
 // ── Binary discovery (env var → PATH → hardcoded paths) ──
-let RELIARY_BIN = process.env. || null;
+let RELIARY_BIN = process.env.RELIARY_BIN_PATH || null;
 if (!RELIARY_BIN) {
   // Check PATH first
   try {
@@ -120,7 +120,7 @@ function daemonCmd(cmd) {
 }
 
 // ── Daemon health check: verify binary exists (TCP check would require daemon running) ──
-let  = true; // assume healthy — CLI fallback handles gracefully
+let DAEMON_HEALTHY = true; // assume healthy — CLI fallback handles gracefully
 
 function cacheRead(path, hash, len) {
   return daemonCmd(`cache-read ${path} ${hash} ${len}`);