yvgude/lean-ctx v3.9.13

on GitHub

Fixed

• MCP cancellation handling — call_tool now checks the rmcp CancellationToken
  after tool execution. If the client sent notifications/cancelled during execution,
  the result is dropped instead of replying to a stale JSON-RPC message ID (#1265).
• Delta response diff algorithm — find_sync_point() replaces the greedy
  line-diff that consumed too many old lines on single-line changes, fixing false
  positives in compute_delta().
• ctx_symbol stale index panic — start_line is now clamped to lines.len()
  with a graceful "stale index, reindexing" fallback instead of panicking (#1126).
• ctx_patch dry_run honored — dry_run=true now correctly prevents writes to
  disk for all op types including replace_unique (#1112).
• Secret redaction false positives — token == "" in Go source no longer
  corrupted into token =[REDACTED] — comparison operators are excluded (#1095).
• ctx_search global limit — max_results enforced globally across all files,
  preventing unbounded output that could crash clients (#1153).
• Shell chained command compression — &&/; chains now preserve output from
  all segments instead of summarizing only the first (#1130).
• stderr capture — stderr is now always captured even when stdout is empty (#1127).
• Stale AUTO CONTEXT — relevance scoring gates preload injection to prevent
  stale task state from previous sessions (#1131).
• Shadow mode Edit allow — shadow-mode instructions now explicitly allow native
  Edit/StrReplace operations (#1139).
• Preload usefulness guard — handler returns structured (body, usable) instead
  of grepping prose for dead-end literals (#1137).
• OCLA cached errors — provider errors no longer replayed as 200 from cache (#1169).
• OCLA budget enforcement — proxy requests now respect budget limits (#1167).
• rmcp response task drain — response tasks are drained to prevent RAM growth (#1160).
• Shell allowlist # comments — bare # comment lines no longer blocked (#1109).
• ctx_search symbol path — path= parameter now respected for graph root (#1108).
• ctx_read lines:N:M syntax — colon-separator error now returns clear error
  instead of silent failure (#1209).
• Cross-Source Hints scoping — hints block now scoped to requested range (#1247).
• ctx_shell archived JSON truncation — JSON preview no longer truncated
  mid-document (#1245).
• Anchored mode instruction fallback — no longer silently falls back to full
  on instruction files (#1238).
• Background job status output — long-running jobs now show partial output
  during execution (#1217).
• ctx_patch hash error message — clear guidance instead of cryptic "missing hash" (#1216).
• Dashboard verification chip — 0/0 no longer shows red 0% (#1268).
• Hook rewrite schema validation — output now passes Claude Code PreToolUse
  schema (#1264).
• ctx_shell code hallucination — git show output no longer compressed into
  valid-looking nonexistent code (#1263).
• Cache-hit detection — no longer sniffs rendered output for markers (#1133).
• Codex PreToolUse allow decision — emits supported decision format (#1116).
• Shell long-running monitor loops — 120s/8MB cap relaxed for background
  monitors (#1113).
• Dashboard net-of-injection — calculation no longer overstates overhead (#1104).
• ctx_read Edit precondition — satisfies read-before-write requirement (#1091).
• ctx_shell selective compression — no longer discards selected lines (#1084).
• Security advisories — remaining pi package advisories remediated (#1201).
• Bounce cost tracking — extra turns counted as negative savings (#1193).
• Routing feedback timing — decisions recorded after measured outcomes (#1170).
• OCLA capsule order — delta insertion order preserved (#1143).
• Shell literal-path redirects — outside project root follow-up (#1142).
• Carriage return resolution — progress output no longer arrives as glued lines (#1140).
• Shell case/esac allowlist — case/esac no longer blocked; ${p%% *}
  tokenization fixed (#1147).
• Wrap/init allowlist — permission seeds now match active profile tool names (#1289).
• CodeQL alerts — open code scanning alerts addressed (#1166).
• Dependency lockfiles — refreshed for Dependabot alerts (#1162).
• Background job retention — completed job retention bounded (#1182).

Added

• Benchmark Optimization Initiative (epic #1305) — 14 new modules for context
  efficiency, targeting -55% token delta:
  • Compression annotation control — CompressionAnnotation enum (Full/Quantized/None)
    with configurable threshold to eliminate noisy [compressed 0%] markers (#1317).
  • Progressive disclosure — auto-mode resolution tiers files as full →
    signatures → map based on line count thresholds (#1309).
  • Prefix stability audit — determinism tests ensuring savings footer and
    auto-mode resolver produce byte-stable output for provider cache (#1310).
  • Turn-level context budget — turn_fresh_limit caps fresh tokens per tool
    response with truncation and expand hints (#1306).
  • Cross-read deduplication — DeliveredRanges tracks delivered line ranges
    per file; overlapping re-reads emit only novel lines (#1313).
  • Session-level budget — cumulative token tracking with pressure tiers
    (Green/Yellow/Orange/Red) and progressive mode downgrades (#1307).
  • Marginal Information Gate — suppresses tool responses with <20% novel
    content relative to already-delivered context (#1308).
  • Negative knowledge tracking — remembers irrelevant files, dead-end queries,
    and absent features to prevent re-exploration (#1314).
  • Content-addressed handles — blake3-based handles for zero-cost re-reads of
    unchanged content with mtime staleness checks (#1315).
  • Delta responses — delivers only changed lines for re-reads after edits using
    sync-point diff algorithm (#1316).
  • Cache diagnostics — monitors provider-side prompt cache hit rates with
    warning/critical alerts (#1311).
  • Query-aware adaptive compression — classifies task intent (Explore/Implement/
    Debug/Review) and maps to appropriate compression levels (#1312).
  • ETPAO metric — cost-weighted token measurement accounting for fresh input,
    cached input, output, and reasoning token pricing (#1318).
  • Benchmark suite foundation — EtpaoReport for reproducible A/B comparisons
    between lean-ctx and baseline (#1319).
• Config profiles — different settings for local LLM vs cloud provider (#1159).
• Config parameter comments — explainer comments for each config parameter (#1158).
• Agent-config compression — automatic rule scoping and context diet (#1141).
• Qodercli support — integration with Qoder CLI (#1135).
• Response shaping — reduce LLM output verbosity (#1125).
• HTML content extraction — compress web content before injection (#1124).
• Conversation history compression — reduce token cost of long sessions (#1123).
• Proactive context expansion — auto-inject compressed data on relevance (#1122).
• Inline shell output — moderately-sized output inlined instead of archived (#1110).
• Runtime flags refactor — internal env mutation replaced with flags (#1107).
• Dashboard compression exclusions — intentionally-uncompressed tools excluded (#1105).
• CLI ledger push — manual context loading without AI (#1101).
• Per-call banner noise — 0 tokens saved banner removed on mode=full reads (#1085).
• Dashboard session presence — shows editor session instead of MCP transport
  count (#1177).
• Dashboard context contents — properly separates history from active usage (#1174).
• Dashboard ledger status — verified status scoped appropriately (#1175).
• USD-aufgeschlüsselte Savings — per-token-stream cost breakdown (#1191).
• Adaptive Compression Depth — compression backs off when it breaks cache (#1195).
• Output determinism benchmark — validates #498 provider-cache preservation (#1194).
• A/B Benchmark harness — proves compression quality non-degradation (#1192).
• Pack create refactor — orchestration refactored for maintainability (#1277).
• SKILL.md accuracy — core tools table corrected (#1288).
• README token footprint — updated from ~2.1K to actual overhead (#1267).
• Dual-arm self-verify — turn count corrected (#1266).

Upgrade

lean-ctx update                 # recommended (auto-downloads + refreshes shell hooks)
cargo install lean-ctx          # or
npm update -g lean-ctx-bin      # or
brew upgrade lean-ctx

Note: After upgrading via cargo/npm/brew, run lean-ctx setup to refresh shell aliases. lean-ctx update does this automatically.

Full Changelog: v3.9.13...v3.9.13