v0.20.9

What's changed

Added

• preview_object — the agent can now read a bounded preview of an object's
  content. Previously the agent could enumerate keys and read metadata
  (head_object) but could not look inside an object. It now has a read-only
  preview_object(provider_id, bucket, key) tool: a single bounded Range GET
  (hard cap 1 MiB/call), text-only (binary/oversized objects are reported, not
  decoded), redaction-passed, never persisted, and bounded per turn (a few
  objects / a few MiB) so it can't be looped into a bulk download. This makes
  "what's inside this manifest / config / log object?" answerable inline.
  • Agent-native by bounds, not a gate: no per-call confirmation (that would
    ossify the loop) — safety is code-enforced caps + sanitization + audit, the
    same model as the other read-only probes.
  • Security rule #11 updated accordingly: from "no object bodies by default"
    to "no bulk body downloads, with preview_object as the one bounded,
    audited, per-turn-capped exception." Bulk/recursive/full-object downloads
    remain prohibited; evidence import (GB-scale) still requires confirmation.

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.9-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.20.7

What's changed

Fixed

• Clicking a suggested next-step no longer drops the literal text "None" into
  the composer. A proposal with an explicit null title/reason was stringified
  as Python str(None) → "None" in normalize_proposal (the get(k, "")
  default only applies to absent keys, not present-but-null ones), which then
  surfaced as the ask_user_for_context composer prefill. Null/None now coerces
  to "", so title falls back to the action-type label and reason becomes
  None (and the prefilled question is always a real sentence).

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.7-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.20.6

What's changed

Added

• Two StorageOps skills for gaps the tools already supported (catalog now 18):
  • storageops-inventory-analysis — how to read an inventory for capacity and
    object-shape (size/count, size histogram, prefix and storage-class
    distribution, small-object ratio, largest objects) via analyze_uploaded_file
    (attached file) or a confirmed plan_inventory_import (+ read_run_result).
    The fact layer beneath the lifecycle/cost decision.
  • storageops-account-posture — how to use survey_account for an account-wide
    landscape + config posture (logging / inventory / lifecycle / public-access-
    block per bucket) and where to look first, with read_run_result for a
    backgrounded survey. The no-error audit entry point (vs triage's error path).
  • Both are written agent-native: on-demand knowledge with adaptive decision
    trees and capability hints, not fixed pipelines (account-posture explicitly
    says not to reflexively review every bucket); app-native tool names only;
    guidance-only. eval-golden-cases gains a "coverage honesty" check (don't
    assert a feature absent when access_denied; snapshot ≠ trend; visible vs
    total buckets). Routing relies on the distinct catalog descriptions.

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.6-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.20.4

What's changed

Fixed

• A step-budget (max_turns) limit no longer breaks the session. Previously
  a complex investigation that exhausted the turn budget surfaced a hard
  "Max turns (16) exceeded" error, lost the whole turn, showed a misleading "open
  settings" action, and (because the failed stream fell back to the blocking
  turn) re-ran the entire agent a second time. Now, when the budget is reached,
  the agent makes one tool-less finalize call that synthesizes a grounded
  best-effort answer from the investigation so far (explicitly marked as possibly
  incomplete, with an offer to continue). The turn budget is unchanged and still
  bounded (N tool-loop turns + 1 tool-less finalize); the client never sees a
  max-turns error and never double-runs. The agent is also instructed to converge
  and checkpoint findings (record_finding / note_fact) as it works, so a
  "continue" follow-up resumes from real context.
• The model chip refreshes after first-run configuration. Adding the first
  model provider through the Settings drawer (e.g. via the first-run wizard)
  changed neither sidecar-readiness nor the active session, so the composer chip
  stayed on "Add model" until a session switch — even though chat already worked.
  The chip now re-fetches when the Settings drawer closes.

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.4-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.20.3

What's changed

Fixed

• The thread no longer looks frozen while the agent is generating. After the
  tool trace appears, the post-tools / between-rounds wait (often the longest,
  with no streamed text yet) showed only a lone blinking caret. It now shows an
  explicit animated "Working… (still running)" indicator until the answer starts
  streaming.
• Error-triage next-step chips survive a reload / session-switch. The
  deterministic safe_next_actions were only on the POST response, so reopening
  a session showed empty chips. GET /error-triage/{id} and
  GET /sessions/{id}/error-triage now re-derive them deterministically from the
  stored (already redacted) input — no new storage, no migration.

Changed

• Tool-name consistency (§2.4). The error-triage playbooks, docs/tools.md,
  and the CLAUDE.md whitelist now use the agent-facing tool names
  (test_addressing_style, inspect_endpoint_tls) that the SKILL.md bodies and
  agent instructions already use — so guidance never names a tool the agent
  can't call. (The underlying S3-layer functions keep their names:
  test_path_style_vs_virtual_host, inspect_tls.)
• read_run_result is now listed in the agent's main tool instructions, not
  only in the survey-timeout note — so the agent knows it can re-read a
  backgrounded survey/review/import result in a later turn instead of re-running.
• Stale docs/docstrings aligned to the single-agent model: architecture.md
  (removed "analysis narrators"; skill context is catalog + read_skill
  progressive disclosure, not eager 1–3 selection; triage flow has no "optional
  Agent interpretation"); skills/__init__.py, skills/context.py,
  skills/contract.py (no "triage Agent" / eager-injection framing);
  pyproject.toml (no "agent planner mode"); summary_builder.py comment
  (proposals are free-form, not a fixed allowlist).

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.3-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.20.2

What's changed

Post-v0.20 review cleanup — no behavior change beyond stronger redaction.

Security

• Shared redactor now scrubs model API keys (sk-…). Defense-in-depth: a
  model key pasted into the chat or echoed in a provider error is masked
  everywhere the shared redactor runs (session messages, audit logs, reports),
  not just on the triage path. Aligns with security rule #15.

Removed (dead code from the v0.20 single-agent migration)

• analysis/drilldown.py + its test — the bounded-aggregate tools whose only
  consumer (the deleted in-run analysis narrator) is gone.
• runs/analysis_report.py: agent_analysis_md + render_agent_report (the
  "Agent Interpretation" / "Agent mode" report sections) and the now-empty
  agent_section parameter on the dataset-report renderers.
• Frontend dead API: uploadDataset (run-scoped upload) and listDatasets.
• next_actions.ALLOWED_ACTION_TYPES dead back-compat alias.

Changed (stale docs / comments)

• docs/architecture.md: account_discovery description no longer claims an
  "Agent mode 422 / future phase" — it's the agent's survey_account tool.
• CLAUDE.md: dropped the dead optimization_report capability bullet.
• agent_runtime/__init__.py, guardrails.py, main.py: docstrings no longer
  describe an "agent planner mode" (there is one conversational agent).
• next_actions.normalize_proposal docstring: clarified it accepts any safe
  free-form action_type (not a fixed allowlist).
• Frontend RunEvent: removed the never-emitted plan / tool_selected types.
• Stripped historical "(Phase NN)" provenance tags from module docstrings
  (migration provenance comments kept).

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.2-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.20.1

What's changed

Fixed

• Empty-state subtitle no longer overpromises. "Read-only by default — I'll
  ask before running anything" became "Read-only and never destructive — I'll
  ask before moving any data" (zh equivalent): the agent runs read-only checks
  itself; only cloud data-moving work is confirmation-gated.
• Backgrounded survey/review now resumes via read_run_result. When an
  inline survey/review exceeds the time budget, the timeout note and the agent
  instructions now tell the agent to call read_run_result(run_id) in a later
  turn instead of re-running the survey.
• Triage safe_next_actions are now clickable. TriageCard renders the
  deterministic next-check proposals as one-click chips (same handoff as agent
  proposals), instead of dropping a field the API already returned.
• Doc residual: docs/security.md "Graded execution" no longer references
  the removed autonomous_readonly/assisted autonomy policy.

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.1-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.20.0

What's changed

Single-agent architecture. This release finishes the agent-native migration
by eliminating the dual-track design: there is now exactly one LLM in the
product — the conversational session agent. Everything under runs/ is pure
deterministic compute the agent invokes as a tool or saves as an auditable
report artifact. No second "run-planner" LLM, no in-run interpretation
narrators, no planner_mode switch. The deterministic engines, DuckDB, the S3
read-only whitelist, output sanitization, and the confirm gate on data-moving
work are all kept — they are the security floor.

Removed

• The run-planner agent. Deleted agent_runtime/tool_registry.py,
  prompts.py, context_builder.py, result_parser.py, and the
  agent_service.run_agent / ToolInvoker machinery. agent_service.py now
  keeps only build_agent / get_model_credentials for the conversational
  agent.
• In-run interpretation narrators. Deleted
  agent_runtime/analysis_agent.py (the access_log_analysis /
  inventory_analysis narrator + analysis/drilldown.py aggregate tools) and
  error_triage/triage_agent.py. Analysis and triage are deterministic-only;
  the conversational agent narrates the sanitized result if asked.
• planner_mode. Dropped from the API (RunCreate/RunSummary/RunDetail,
  ErrorTriageRequest), the run SSE run_started event, the frontend types, and
  the run-detail UI. run_service.run_sync always dispatches a run to its
  deterministic executor; the runs.planner_mode SQLite column is retained
  (defaulting to 'deterministic') only because the schema is append-only and is
  no longer read or written.
• The optimization_report run type (never implemented as a real executor);
  an unknown run_type is now a clean 422.

Changed

• Runs expose only their real tool trace, findings, and summary — no canned
  step "plan" event and no agent-authored prose section in reports.
• Evidence import is reached through the agent, not a separate panel —
  AccountProfilePanel is now a read-only profile view.

Added

• read_run_result(run_id) tool — lets the agent pick up a backgrounded
  survey/review/import result in a later turn (status + sanitized summary; only
  runs linked to the current session) instead of re-running.

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.20.0-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.19.29

What's changed

Cleanup pass resolving the verified-true items from a code/skills review — no
new behavior, all agent-native consistency, dead-code removal, and small fixes.

Fixed

• Slash /logs and /inventory now open the file picker (like the
  empty-state chips), instead of seeding a prompt the agent has no file to act on.
• The model chip recovers from a transient sidecar blip — refreshModel
  retries a few times instead of getting stuck on "Add model" until a refresh.
• Sending an ambiguous-type attachment gives feedback (a "choose a type"
  hint) rather than a silent no-op.

Changed

• skills_used is bound to skills actually loaded via read_skill this turn
  — the model can no longer claim a skill it never opened (keeps the report
  honest).
• Skill selection is robust to spacing/punctuation — a keyword like
  SignatureDoesNotMatch matches "Signature Does Not Match" / "access-denied"
  without a hard-coded error→skill map (still metadata-driven).
• read_skill has a per-turn budget (max 6 loads) so a loop can't pull every
  skill body into context.
• The deterministic session report labels its "next actions" as rule-derived
  suggestions, distinct from the agent's own proposals.
• Refreshed stale SKILL.md guidance (access-log, lifecycle-cost, performance,
  security-iam, migration, replication) to the current tools: local files →
  analyze_uploaded_file inline; config/account → review_bucket_config /
  survey_account; only cloud imports stay confirmed.

Removed

• Dead /sessions/{id}/actions/preview endpoint + preview() + the frontend
  ActionPreviewResult type.

Docs

• Rewrote docs/architecture.md to the agent-native model (no autonomy toggle,
  no new_run form, free-form proposals, origin='agent' runs hidden from the
  thread); fixed the session_agent module header (attached files analyzed
  inline) and the inline-survey timeout note.

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.19.29-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.

v0.19.28

What's changed

Completes the agent-native rebuild: the conversational agent is the sole
operating surface, and no rigid/ossified pipeline remains. Deterministic engines
survive only as the security/reproducibility floor the agent invokes (and as
opt-in auditable reports) — never a UI-fired flow or a card mid-conversation.

Changed

• No run card ever appears from a conversation. The agent's own read-only
  survey/review tools (survey_account, review_bucket_config) now record runs
  with origin='agent' (migration 15) that the thread filters out — the agent
  narrates the result inline instead. This removes the stray deterministic
  account_discovery card that could fire mid-chat (e.g. while analyzing an
  uploaded log).
• Retired the agent-autonomy toggle entirely. The agent is always a fully
  autonomous read-only investigator; the assisted/autonomous_readonly setting,
  its endpoint, and its Settings UI are gone. Read-only investigation always runs;
  cloud data-moving work still always requires confirmation.
• The agent stays on the user's request. New instructions stop it from firing
  cloud probes (credentials, account survey) for a local-file task — it analyzes
  the attached file and answers, touching the cloud only when asked.
• Removed the retired new_run form handoff from next-action proposals:
  investigation/diagnosis/config/account/analysis proposals route back to the
  agent conversationally; only evidence import, the saved report, and a context
  question get a purpose-built flow.

Fixed

• Uploading a file no longer loses it if the upload fails (the composer is
  cleared only after success).
• Forking a session now copies its uploaded datasets and their files on disk.
• Re-uploading the same filename reuses the dataset row instead of leaving
  duplicate records pointing at one overwritten file.
• A streamed turn that ends without a completion event now reconciles via the
  blocking fallback instead of showing an empty next-steps list.
• Empty-state "Analyze access logs" / "Inventory" chips open the file picker.

Removed

• Dead code: agent_runtime/autonomy.py, the /settings/autonomy endpoint, the
  frontend previewSessionAction, and stale docs/comments (README confirmation
  wording, composer "two modes", the M012 "OS keychain" note, the "Phase 17
  allowlist" comment).

Install

Builds are unsigned / ad-hoc (no Apple notarization, no Windows Authenticode). Verify downloads against the SHA256SUMS-*.txt files.

• macOS (arm64): open the .dmg, drag to Applications. First launch shows a Gatekeeper prompt — right-click the app → Open, or run xattr -dr com.apple.quarantine "/Applications/Storage Agent Workbench.app". It starts in a few seconds.
• Linux (x64): sudo apt install ./storage-agent-workbench-v0.19.28-linux-x64.deb
• Windows (x64): run the -setup.exe; if SmartScreen warns, choose More info → Run anyway.

Prompt-free signed builds: see docs/signing.md.