Skip to content

v1.16.0 — Phase D/E/F/G + Mutation Harden

Choose a tag to compare

View all tags
@muchiny muchiny released this 03 May 12:38
· 575 commits to main since this release
v1.16.0
b1be378

[1.16.0] - 2026-05-03

Summary

MCP interactive integration: Phase D/E/F/G across 38 handlers + mutation
testing harden across 27 modules. Long-running tools now stream
per-step progress, destructive tools confirm via elicitation, diagnostic
tools optionally enrich output via LLM sampling, and multi-step tools
emit structured logs through notifications/message. Test suite grew
from 6,934 to 6,973 tests; mutation kill rate is now 99 %+ on the
testable surface across 27 hot files.

Added

  • Phase D — Per-step progress notifications. Long-running multi-host
    tools (ssh_exec_multi, ssh_metrics_multi, ssh_runbook_execute,
    ssh_log_search_multi, ssh_log_tail_multi) now emit
    notifications/progress for each host or step, so clients with a
    progress UI can render real-time completion instead of a single
    black-box wait.
  • Phase E — Elicitation on destructive tools. 18 destructive
    handlers gain a security.require_elicitation_on_destructive
    opt-in: when enabled and the client advertises the elicitation
    capability, the server sends elicitation/create and short-circuits
    with a structured error if the user declines. Affected tools include
    ssh_cron_remove, ssh_file_patch, ssh_file_write,
    ssh_files_write, ssh_firewall_deny, ssh_group_delete,
    ssh_helm_uninstall, ssh_hyperv_vm_stop, ssh_k8s_delete,
    ssh_net_equip_config, ssh_process_kill, ssh_reg_delete,
    ssh_service_stop, ssh_template_apply, ssh_win_feature_remove,
    ssh_win_firewall_remove, ssh_win_process_kill,
    ssh_win_service_stop.
  • Phase F — Optional LLM sampling on diagnostic tools. 13
    diagnostic handlers add a summarize=true flag (with
    summary_max_tokens cap) that asks the client's LLM to analyse the
    raw output via sampling/createMessage and appends the summary
    alongside the raw data. Falls through to raw-only when the client
    doesn't advertise sampling. Tools: ssh_diagnose,
    ssh_security_audit, ssh_log_aggregate, ssh_compliance_check,
    ssh_compliance_score, ssh_compliance_report, ssh_vuln_scan,
    ssh_journal_query, ssh_compare_state, ssh_env_diff,
    ssh_env_drift, ssh_incident_correlate, ssh_incident_triage.
  • Phase G — MCP-side structured logging. 8 long-running tools
    (ssh_runbook_execute, ssh_ansible_playbook, ssh_helm_install,
    ssh_helm_upgrade, ssh_helm_rollback, ssh_terraform_apply,
    ssh_k8s_rollout, ssh_security_audit) emit per-step
    notifications/message so clients can stream the playbook /
    rollout / audit log to their UI instead of waiting for the bulk
    result.
  • Async pre-execute and enrich hooks on StandardTool so handlers
    can drive elicitation / sampling / logging without re-implementing
    the 16-step pipeline.

Changed

  • Tool count: 338 → 357 across 75 groups (was 74). The
    total_tools and total_groups fields on the mcp_list_tool_groups
    response, the DXT manifest, and the .well-known/mcp/server-card.json
    all reflect the new totals.
  • Metrics::render_token_summary arithmetic refactor. The two
    inner if before > 0 { saved * 100 / before } else { 0 } and
    if total_kind > 0 { *count * 100 / total_kind } else { 0 } checks
    are now removed — the outer guards already prove both denominators
    positive, so the dead branches are inlined as direct divisions.
    Behaviour identical, surface for arithmetic mutations reduced.

Tests

  • Mutation testing: 27 modules, ~99 % kill rate on viable mutants.
    Files now at 100 % kill rate (excluding provably-equivalent or
    feature-gated mutants): mcp/{elicitation, sampling, progress, logger, meta_tools, pending_requests, client_requester, instructions, protocol}.rs, domain/{output_kind, output_cache, diff, output_truncator, runbook, history, task_store, data_reduction}.rs, metrics.rs, security/{rbac, validator, audit, rate_limiter}.rs, ports/{tools, executor, protocol}.rs.
    Two mutants on mcp/standard_tool.rs (audit-log content + warn
    tracing) remain — killing them needs a tracing-subscriber and
    audit-log file harness, deferred.
  • Test count: 6 934 → 6 973 lib tests (+39 net new). New tests
    include tokio::time::timeout-bounded shortcuts on
    ToolContext::elicit_confirm / sample, async-cache twins of the
    truncator arithmetic invariants, two-kind percentage assertions on
    render_token_summary, single-field is_empty cases on
    ToolAnnotations, exact-byte 160/161-char truncation cases, and a
    threaded floor_char_boundary / ceil_char_boundary termination
    check that catches i /= 1 / i *= 1 infinite-loop mutants.
  • .cargo/mutants.toml exclude_re now lists ten classes of
    provably equivalent or untestable-without-Clock-injection
    mutations: ToolHandler::output_kind defaults, empty-vec
    scoped_paths, jq-feature-gated bodies, OutputCache TTL
    boundaries, floor/ceil short-circuit equivalents, infinite-loop
    arithmetic on the boundary scanners, ConfigWatcher fs-timing /
    filter mutations, AuditLogger TTL boundary, and the > -> <
    u64-zero comparison family in render_token_summary.

Full Changelog: v1.15.1...v1.16.0

Assets 20
Loading