Skip to content

Changelog

[v0.1.7-beta] - 2026-04-02

Fixed

  • fix(server): /api/status now reports degraded when Ollama health check explicitly returns false (sticky online fallback only for transient probe failures)
  • fix(tui): restored status bar visibility by keeping status bar CSS in widget-local StatusBar.DEFAULT_CSS
  • fix(version): align package export by re-exporting version from airecon._version in airecon/__init__.py
  • fix(quality): resolve Ruff F401 for package __version__ re-export

Improved

  • feat(mcp): expose and display total_tools for MCP servers to avoid misleading truncated counts
  • feat(tui): /mcp list <name> now shows first 10 tools only to keep UI/context lightweight
  • feat(tui): moved ConfirmDelete modal CSS to global styles.tcss with scoped selectors

Validation

  • tests: suite green at release time

[v0.1.6-beta] - 2026-03-17

Critical Security Fixes

Session Save Race Condition

  • fix(agent): add asyncio.Lock() to _schedule_token_usage_snapshot_save() to prevent concurrent session save corruption
  • fix(agent): acquire lock in stop() method before synchronous session save
  • fix(server): call agent.stop() in lifespan shutdown to ensure session saved on exit
  • fix(filesystem): check file_path.is_symlink() before create/read operations
  • fix(filesystem): resolve symlink and verify it points within workspace
  • fix(filesystem): use atomic writes via tempfile.mkstemp() + os.replace() to prevent partial writes

Command Injection Prevention

  • fix(validators): add 8 new dangerous patterns to DANGEROUS_PATTERNS:
  • \$\{[^}]+\} - variable expansion
  • <\([^)]+\) - process substitution
  • \$'[^']*' - ANSI-C quoting
  • `[^`]+` - backtick command substitution
  • \$\([^)]+\) - $(command) substitution
  • ;\s*(curl|wget|nc|...) - pipe to interpreter
  • \|\s*(bash|sh|python|...) - pipe to interpreter
  • /etc/(passwd|shadow|sudoers|ssh/) - sensitive file access
  • chmod\s+[0-7]*[4-7][0-7]{2} - setuid/setgid chmod

Stability Improvements

Session Persistence

  • fix(agent): save session on Ctrl+C/exit in AgentLoop.stop() method
  • fix(server): call agent.stop() in lifespan shutdown handler
  • fix(agent): log session data count (subdomains, live_hosts, vulns) on save

HTTP Timeout Protection

  • fix(fuzzer): propagate baseline failures with explicit logging
  • fix(fuzzer): skip unreachable params (status=-1) before fuzzing
  • fix(fuzzer): log skipped params due to baseline failures

Browser Resource Cleanup

  • fix(browser): add force kill fallback (pkill -9 chromium) on cleanup failure
  • fix(browser): surface screenshot failures to caller via screenshot_failure flag in result dict

CVE Validation

  • fix(reporting): tighten _CVE_RE regex from r'^CVE-\d{4}-\d{4,7}$' to r'^CVE-(19[89]|20\d{2})-\d{4,7}$'
  • fix(reporting): validate CVE year range 1989-2099 to prevent fake CVEs

Context Management

  • fix(models): make context limits config-based via _get_context_limits()
  • fix(models): calculate max_conversation_messages = ollama_num_ctx // 128
  • fix(models): add 5 new config keys: agent_max_conversation_messages, agent_compression_trigger_ratio, agent_uncompressed_keep_count, agent_llm_compression_num_ctx, agent_llm_compression_num_predict

Tool Result Truncation

  • fix(models): add _truncate_tool_result() helper function (50KB limit)
  • fix(executors): add _append_tool_history() helper with truncation on append
  • fix(models): keep legacy truncation in add_message() as safety net

Incremental Pruning

  • fix(loop): change _executed_cmd_hashes pruning from .clear() to FIFO keep newest 2500
  • fix(loop): log pruning: "incrementally pruned: X → 2500 entries"

Code Quality

  • refactor: remove 51 lines of verbose FIX comments
  • refactor: simplify docstrings across 12 files
  • test: update test_validation.py to reflect improved security
  • test: all 1369 tests passing (100% backward compatible)

Files Modified

  • airecon/proxy/agent/loop.py - Session lock, save on shutdown, incremental pruning
  • airecon/proxy/agent/validators.py - 8 new dangerous patterns
  • airecon/proxy/agent/models.py - Config-based context limits, truncation helper
  • airecon/proxy/agent/session.py - Vulnerability dedup fix
  • airecon/proxy/agent/executors.py - Tool result truncation helper
  • airecon/proxy/filesystem.py - Symlink TOCTOU protection, atomic writes
  • airecon/proxy/browser.py - Screenshot failure surfacing, force kill cleanup
  • airecon/proxy/fuzzer.py - Baseline failure propagation
  • airecon/proxy/reporting.py - CVE validation regex
  • airecon/proxy/server.py - Session save on shutdown
  • airecon/proxy/config.py - Bounds validation for all numeric fields
  • tests/proxy/test_validation.py - Updated for improved security
  • tests/proxy/test_fuzzer.py - Updated for baseline failure handling

Added

Phase 1 — Autonomous Recovery & Exploration Engine

  • feat(agent): watchdog forcing — LLM stuck in text-only loop forces execute tool (max 2x before abort)
  • feat(agent): anti-stagnation exploration — temperature boost when no new high-confidence evidence (≥0.65)
  • feat(agent): tool diversity tracking — same-tool streak detection via _recent_tool_names deque
  • feat(agent): per-phase exploration directives via _build_exploration_directive()
  • feat(agent): quality scoreboard — evidence 40%, reproducibility 35%, impact 25%
  • feat(agent): recovery state context injected after conversation truncation
  • feat(models): objective_queue (max 64) + evidence_log (max 200, dedup last 50)
  • feat(config): 6 new exploration config keys (agent_exploration_mode, agent_exploration_intensity, agent_exploration_temperature, agent_stagnation_threshold, agent_tool_diversity_window, agent_max_same_tool_streak)

Phase 2 — Skill Orchestration & Tool Budget

  • feat(agent): skill phase boost — _PHASE_SKILL_DIRECTORIES gives +2 score to phase-preferred skills
  • feat(agent): tool budget per phase — _PHASE_TOOL_BUDGETS with soft limits per tool per phase
  • feat(agent): budget warnings at 75% (warning), 100% (exhausted), 0 (discouraged)
  • feat(pipeline): phase skill hints injected into get_phase_prompt()

Ollama Stability — Context & VRAM Recovery

  • feat(agent): multi-level VRAM crash recovery — 4 escalation tiers persisted via _adaptive_num_ctx: Tier 1 (ollama_num_ctx_small, 80 msgs), Tier 2 (÷2, 50 msgs, 5s wait), Tier 3 (÷4, 30 msgs, 10s wait), Tier 4 (4096, 20 msgs, 30s wait)
  • feat(agent): proactive context monitoring — trims at ≥80% token usage, aggressive at ≥90%
  • feat(agent): dynamic compression interval (5/10/15 iters based on context fullness)
  • feat(agent): skip compress_with_llm when >65% context full (OOM prevention)
  • feat(agent): _cap_tool_result scales down dynamically with _adaptive_num_ctx
  • feat(agent): _adaptive_num_predict_cap limits token generation after VRAM crash
  • feat(ollama): complete() accepts options: dict (num_ctx, num_predict, temperature)
  • feat(models): compress_with_llm passes num_ctx=8192, num_predict=1024 to avoid OOM
  • feat(agent): session auto-saved after each VRAM crash recovery

Tested Endpoints Memory

  • feat(session): SessionData.tested_endpoints — LRU list (max 500) tracking "METHOD url" strings
  • feat(session): record_tested_endpoint(session, url, method) with dedup + LRU eviction
  • feat(agent): _record_tested_endpoint() auto-records from execute (curl), browser_action, fuzz tools
  • feat(agent): last 20 tested endpoints shown in _build_critical_findings_context after truncation

@/file and @/folder References

  • feat(agent): @/path resolver — copies local files/dirs to Docker workspace/uploads/ automatically
  • feat(agent): per-file try/except OSError in directory copy — single file errors no longer abort
  • feat(agent): detailed skip reporting (binary, too-large, OS-error) in copy summary

TUI — Slash Command Autocomplete

  • feat(tui): / prefix triggers slash command autocomplete in chat input
  • feat(tui): PathCompleter widget with proper error logging

Agent Intelligence

  • feat(agent): attack chain detection — links vuln evidence across phases
  • feat(agent): semantic dedup for objectives (Jaccard 0.70 threshold)
  • feat(agent): adaptive thinking with confidence floor 0.65 for meaningful evidence
  • feat(agent): cross-session memory — loads prior session findings on start
  • feat(agent): 6 hypothesis-driven vuln discovery improvements
  • feat(data): expand all correlation pattern files (major expansion)
  • feat(data): rename expert_testing_patterns.json → patterns.json
  • feat(zeroday): redesign zero-day patterns for realistic LLM discovery
  • feat(agent): smart fuzzer routing + dynamic URL correlation + injection-chain detection
  • feat(agent): data-driven injection points, port/tech hints, HTTP impact validation
  • feat(skills): add 22 new skills (frameworks, protocols, technologies, LLM coverage)
  • feat(skills): aggressive exploration mode + headless reverse/pwn skill loading

Fixed

  • fix(agent): [CONTEXT MONITOR] messages removed from TUI output (logged to file only)
  • fix(security): block $() and backtick command substitution in watchdog (has_dangerous_patterns())
  • fix(security): auth header propagation improvements
  • fix(validators): add auth browser actions: login_form, handle_totp, save_auth_state, inject_cookies, oauth_authorize
  • fix(agent): _executed_cmd_hashes pruned at >5000 entries to prevent memory leak
  • fix(agent): IDOR false positive reduction in correlation engine
  • fix(agent): phase timeout now counts iterations (not wall-clock time)
  • fix(agent): evidence truncation preserves high-confidence items
  • fix(agent): press_key dedup, DDG lock race, port-scan rerun block
  • fix(agent): harden LLM loop, subagent isolation, and command detection
  • fix(agent): subdomain workspace path, CTF false positives, LLM hallucination
  • fix(agent): watchdog extracts full multi-line bash scripts (not just first command)
  • fix(ollama): enforce thinking/native_tools invariant + guard max_retries
  • fix(ollama): improve detection-failure warning
  • fix(browser): add --ignore-certificate-errors for TLS cipher mismatch on pentest targets
  • fix(docker): fix 8+ binary name mismatches between tools_meta.json and installed binaries
  • fix(docker): fix race condition in docker force_stop
  • fix(patterns): fix all match-breaking issues across data pattern files
  • fix(reporting): _resolve_report_workspace_target() for URL/file/path resolution
  • fix(tui): PathCompleter.hide() bare except replaced with proper logging
  • fix(data): spawn_agent max iterations 200→100 in tools.json
  • fix(data): web_search updated to SearXNG preferred + DuckDuckGo fallback

Changed

  • refactor(ollama): remove name-heuristic capability detection
  • docs: condense README from 983 → 228 lines
  • docs: add airecon-skills community library reference
  • ci: add label-based project board routing workflow
  • chore: add coming_soon/ to .gitignore (local-only roadmap)
  • style: fix ruff E702 semicolons, unused imports, unused variables across codebase
  • test: 448 → 879 tests (96% growth); new test files for context recovery, tested endpoints, recon dedup, path completer, command parse, reporting helpers

[v0.1.5-beta] - 2026-03-05

Fixed

  • fix(core): unpack tuple return value from auto_load_skills_for_message to resolve unhashable list crash
  • fix(core): resolve correlation logger, fuzzer graceful degradation, and browser timeout bugs
  • fix(docker): fix race condition in docker force_stop and ollama model detection
  • fix(tui): initialize live output and remove unused reload override
  • fix(test): patch browser unpacking bug in unit tests

Added

  • test: implement comprehensive unit test suite covering proxy, agent, and TUI components

Changed

  • chore: remove .vscode from version control tracking
  • chore: add pycache and workspace/ to .gitignore
  • docs: update README badge version formats