Skip to content

IRMind v0.5.3

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 17 Nov 18:01
· 220 commits to main since this release
3a05890

IRMind v0.5.3 — Prompt-cache accounting and interrupted-plan recovery

Format impact: Optional cache metadata extends signed usage receipts and
optional cache prices extend signed configuration. Missing fields retain their
existing canonical bytes. Case, artifact, seal and export versions are unchanged.
Retain the creating executable for full reconstruction of its signed native decoder/mapping profiles.

  • Replace unmaintained calendar/duration dependencies with maintained date
    primitives and narrowly reviewed Arrow/Parquet/storage source patches. Current
    canonical and physical-format golden bytes are unchanged; advisory enforcement,
    explicit credentials, TLS, bounded reads and EVTX isolation remain enabled.

  • Publish native macOS Intel binaries alongside Apple Silicon, Linux x86_64,
    Linux aarch64 and Windows. Verify clean installers even when optional native
    signing is disabled. Sigstore, provenance and SBOM verification remain mandatory.

  • Refresh the complete offline presentation with the actual v0.5.3 executable.
    Native Developer ID/notarization and Authenticode remain unavailable.

  • Expand the controlled Elasticsearch/TLS rehearsal guidance and the optional
    HuntEval pilot runner. A local-model rehearsal does not replace analyst
    validation of investigation quality or a connector in a real test environment.

  • Recover an interrupted subagent plan explicitly with /subagents recover,
    after confirmation. Nonblocking interprocess activity/recovery locks protect
    live previews and children, remain held through reconciliation and prevent
    concurrent deletion. Recovery preserves committed work and uncertain usage
    without replaying provider requests or tool effects. Signed formats are unchanged.

  • Account for provider-reported prompt-cache reads and writes separately in
    /budget and the configurable status bar. Inclusive input still consumes the
    full token budget; estimates require every applicable price. Native Anthropic
    requests mark the trusted system/tool prefix; OpenAI and DeepSeek cache
    automatically. Compatible/local endpoints receive no new cache controls.
    Optional cache fields extend signed usage receipts and price configuration;
    absent fields preserve existing bytes. Case, seal and export versions are unchanged.

  • Show an actionable notice when a provider completes without a visible answer,
    while retaining its signed empty reply and reported usage.

  • Rebuild mapped evidence through its signed decoder/mapping profile in both CLI
    and TUI; connector decoders do not need automatic file-format detectors.

  • Verify private originals retained by the case at their signed location instead
    of requiring a second external duplicate of acquired connector responses. The
    same signed source location is used when including originals in an export.
    Missing, altered or unavailable retained originals still fail verification.

  • Confirmed Ctrl+C exits the application from an open case; canceling its modal
    leaves the conversation unchanged. Evidence-picker help matches the actual keys.

Evaluation limits

IRMind remains pre-1.0. Validate a configured model and connector against your own controlled test environment before relying on their investigative results. The local synthetic rehearsal does not establish investigation quality. Native macOS Developer ID/notarization and Windows Authenticode are not active; verify the Sigstore signatures, GitHub provenance and SHA-256 checksums. The external skills catalog is not available in this release; bundled skills remain available.

Full changelog · Verified installation and platforms · Controlled validation


Code signing policy: https://github.com/ns2switch/irmind/blob/main/CODE_SIGNING.md