Skip to content

Release history

What’s new

What shipped in the flagship trusty-tools crates, generated from the crates’ own CHANGELOG.md files.

trusty-memory

Memory palace storage

45 releases. crates/trusty-memory/CHANGELOG.md on GitHub is the source this section is generated from.

1.0.1 2026-11-16

Fixed

  • Linux: start and restart pass RUST_LOG_BUFFER to the daemon they spawn, so its in-memory log buffer keeps the filter a shell-started daemon would use (#11085).

Changed

  • The Linux lifecycle restart tests scan an injected process table of their own processes, so a host trusty-memory daemon no longer changes their result (#11085). restart behaviour is unchanged.

1.0.0 2026-11-15

Fixed

  • memory_recall and memory_recall_deep open a cold palace on the blocking pool and wait at most TRUSTY_RECALL_OPEN_WAIT_SECS (15 s); past that they answer a retryable palace_opening error (JSON-RPC code -32007, error.data with kind, retryable, palace, waited_ms) while the open goes on, so a retry finds the palace (Refs #11255)
  • a memory_recall_all scope: "all" walk stops at TRUSTY_RECALL_ALL_DEADLINE_SECS (20 s) and returns the results gathered so far; every response carries deadline_exceeded and palaces_not_reached, unreached palaces count under the deadline reason, and no detached reopen runs after a deadline hit (Refs #11255)
  • the chat recall_memories tool and the per-message chat context use the same bounded open: the tool answers {error, kind: "palace_opening", retryable: true, palace, waited_ms}, and the context says the palace is still opening (or that the recall failed) instead of reading as "no relevant memories" (Refs #11255)
  • recalls of one palace that miss the cache together share one open, so callers and retries no longer each park a blocking thread on it; a failed open reaches every caller waiting on it and the next call tries again, and a caller whose shared open hit the open-queue bound waits out its own open wait and answers palace_opening instead of an internal error (Refs #11255)
  • a zero, negative or non-numeric TRUSTY_RECALL_OPEN_WAIT_SECS / TRUSTY_RECALL_ALL_DEADLINE_SECS falls back to its default; the open wait is capped at three quarters of TRUSTY_OPEN_QUEUE_TIMEOUT_SECS and the deadline at 3600 s, so no value can panic the daemon (Refs #11255)

Changed

  • trusty-memory 1.0.0: stability release; the public API surface, wire formats and on-disk formats are now covered by Semantic Versioning, so a breaking change needs a major version from here on (owner ruling Dnx)

0.36.8 2026-11-14

Added

  • memory.palace_unload {palace_id}, a socket method (not an MCP tool) that releases one palace's open handles and caches and keeps its data on disk (#10989). It waits up to 10 s for an in-flight write, flushes the palace's BM25 index, drops its session store and registry handle, and answers {palace_id, canonical_id, exists, was_resident, unloaded, data_kept}. A busy palace is refused (-32006) and nothing is released; a palace not on disk answers exists: false. The next use reopens the palace from disk. tm's project decommission calls it.
  • memory.palace_unload answers unloaded only after the palace's redb files are closed, waiting up to the same 10 s (#10989). The next use then reopens the palace writable, never on a read-only snapshot. A file still open after the wait is refused (-32006).
  • On Linux, trusty-memory start, stop and a new restart run on the shared Linux lifecycle (#11085): stop stops only the daemon serving this data dir's socket and fails while a process survives SIGKILL; restart stops, confirms the stop, then starts the installed binary, and starts a daemon that is not running. The daemon they start gets a cleared environment that keeps every non-credential setting it reads (multi-tenant mode, embedder model and caches, BM25, timeouts) and leaves API keys out. A non-UTF-8 or blank TRUSTY_DATA_DIR_OVERRIDE is refused. macOS is unchanged.

Fixed

  • the room-counts console metrics test waits for the evicted palace's KgWriter task before it reads the disk, which removes a flaky failure (Refs #11080)
42 earlier releases
  • 0.36.7 2026-11-10
  • 0.36.6 2026-11-09
  • 0.36.5 2026-11-07
  • 0.36.4 2026-10-31
  • 0.36.2 2026-10-30
  • 0.36.1 2026-10-30
  • 0.36.0 2026-10-26
  • 0.35.0 2026-10-12
  • 0.33.0 2026-10-10
  • 0.32.1 2026-10-10
  • 0.32.0 2026-10-10
  • 0.31.0 2026-10-09
  • 0.30.0 2026-10-09
  • 0.29.6 2026-10-08
  • 0.29.5 2026-10-08
  • 0.29.4 2026-10-08
  • 0.29.3 2026-10-06
  • 0.29.2 2026-10-06
  • 0.29.1 2026-10-06
  • 0.29.0 2026-10-06
  • 0.28.2 2026-09-28
  • 0.28.1 2026-09-27
  • 0.28.0 2026-09-26
  • 0.27.2 2026-09-26
  • 0.27.1 2026-09-25
  • 0.27.0 2026-09-25
  • 0.26.2 2026-09-23
  • 0.26.1 2026-09-18
  • 0.25.6 2026-09-02
  • 0.25.5 2026-08-31
  • 0.23.1 2026-08-15
  • 0.22.0 2026-07-27
  • 0.21.2 2026-07-26
  • 0.21.1 2026-07-24
  • 0.21.0 2026-07-23
  • 0.19.2 2026-07-09
  • 0.20.0 2026-07-21
  • 0.17.0 2026-06-25
  • 0.15.5 2026-06-16
  • 0.15.2 2026-06-09
  • 0.15.1 2026-06-05
  • 0.15.0 2026-06-03

What each of these changed is in crates/trusty-memory/CHANGELOG.md.

trusty-mpm

Multi-agent orchestration

75 releases. crates/trusty-mpm/CHANGELOG.md on GitHub is the source this section is generated from.

1.14.9 2026-11-18

Fixed

  • tm pm-guard no longer refuses unparseable skill or agent frontmatter whose only backslashes cannot spell a hooks key, such as "C:\Users\me" or "match \d+". A hex escape (\x, \u, \U) or an escaped line break still fails closed.
  • docs.rs builds the trusty-mpm docs again: two module-level intra-doc links that newer rustdoc cannot resolve now name their full path (#11451).

Changed

  • A fresh trusty-mpm session now launches on the default tmux server. TM_WORKSPACE_TMUX=1 opts in to the per-workspace tm-ws-* server.
  • Resuming a stopped session whose workspace server holds no live pane recreates its pane on the default tmux server, keeping its project, account and Claude conversation. A live pane is never moved, and TM_WORKSPACE_TMUX=1 keeps a session on its workspace server.

1.14.8 2026-11-18

Added

  • tm hook --pm-guard gains a stale-copy guard (#10574): a [stale_copy] table in .trusty-mpm.toml lists registered (paths) and locked artifacts. A Read of a registered file records its observed version (path, SHA-256, saved copy) under .trusty-mpm/stale-copy/, bounded at 32 MiB and 512 records. A replacing Write/Edit or Bash redirect is denied STALE_INPUT_HASH (GP-052) when the file changed since it was read, was never read, or the record store cannot be read; any write to a locked file is denied LOCKED (GP-053). An absent, empty or unreadable config leaves the guard off.
  • PM sessions register tm hook --completion-evidence on Stop (#10575). It warns, never blocks, when the final message claims completion with no evidence map (verified / helper-reported / unfinished / untested), names an incomplete map, sets a size target with no OWNER-REQUESTED-LIMIT: "<owner quote>" line (INVENTED_TARGET), or carries that line without a quote.
  • A bundled owner-plain output style, selected with [style] active = "owner-plain" or tm launch --style owner-plain: the default PM with plain-word, regularly timed reporting to a non-technical owner. tm install deploys it and the scaffold .gitignore block covers its file (#10576).
  • PM sessions now register tm hook --turn-task on UserPromptSubmit. Each turn gets Task: <id>@<MMDD> as hook context when datg's session record names an open task in its current field; the hook is silent otherwise and never writes datg state.
  • New tm hook --pm-guard-shadow: a minimal pm-guard that judges each Bash, Write and Edit call on parsed facts (tree-sitter-bash argv and canonical paths), not command text. It never blocks. Each would-refuse appends one JSON line to ~/.trusty-mpm/guard/shadow.jsonl; the line holds the rule, argv facts and a command SHA-256, never the command text. Rules: R1 trust-anchor writes, R2 destructive git, discarded work, raw-disk writes and protected-path deletes, R3 cross-repo push, R5 production deletes (per-session allow list in ~/.trusty-mpm/guard/shadow.toml). PM and subagent sessions run all four rules; the Architect runs R1 only. Every session registers the hook on PreToolUse, independent of [pm_guard] enabled.

Fixed

  • An Architect launch, resume or in-place relaunch whose project hook write fails is now refused with an error naming .claude/settings.json, instead of starting without the exec-form tm hook --pm-guard entry. A PM session keeps the previous best-effort behaviour (#10612).
  • tm hook --pm-guard no longer refuses every Write and Edit when one skill or agent file under ~/.claude/agents or ~/.claude/skills cannot be read for lack of permission, or has frontmatter that does not parse and carries no plausible hooks key. The file is skipped with a warning on stderr. Frontmatter that does not parse but declares a hooks key still refuses (#10937).
  • tm hook --pm-guard keeps the --body/--title/--message/--note text-payload exemption when the payload holds backticks or $( inside single quotes, so Markdown code spans in a gh comment no longer read as secret file names. A substitution outside single quotes still voids the exemption (#10974).
  • Regression tests pin that a shell redirect glued to a script word does not hide the script from the pm-guard rules, and that a benign script under the same redirect is allowed (#11262).
  • tm sessions ls reads whether a session's runtime still runs from the panes on its own workspace tmux server, so it agrees with tm tui instead of showing a running session stopped (#11071).
  • A SessionEnd hook no longer marks a session on a tm-ws-* workspace tmux server Stopped while its pane still runs the runtime, so tm projects decommission stops that session and its pane (#11175).
  • tm session ls and tm session resume read a session's liveness on its own tm-ws-* workspace tmux server, so a live per-workspace session no longer reads stopped (#11222).
  • The poller's watchdog now relaunches a dead Architect with tm architect up, the bound launch path, instead of session-relaunch.sh, so the new claude has a launch record and the pm-guard keeps its anchor exemptions. The relaunch fails closed unless up exits 0, launches into the Architect's own session, and a following up --dry-run sees it live and bound. TM_BIN overrides the tm lookup; the 1800 s debounce and 2-failure cap are unchanged.

Changed

  • tm pr open --rung N now applies the label rung:N to the PR, creating it if missing.
  • tm pr queue-check refuses a PR with no rung:N label, with two rung labels, or with a malformed one, and refuses a rung:5 or rung:6 PR unless a trusted APPROVE or WARN is pinned to the head. Its code-critic BLOCK check now uses the same reader as tm pr merge, so only OWNER, MEMBER and COLLABORATOR verdicts count (owner ruling Dr4).

1.14.7 2026-11-16

Breaking

  • trusty-mpm 1.14.7 ships a public-API break under the internal numbering policy (owner ruling 2026-09-26); recorded in scripts/semver-accepted-breaks/trusty-mpm-1.14.7.txt (Refs #8699).

Added

  • Model resolution recognizes the Fable tier (fable alias, [models.tiers] fable, default claude-fable-5-2) and reserves it for PM sessions and engineer and research roles. Any other agent that asks for Fable, through --model, [models.agents], frontmatter or [models] default, runs on Opus, and the substitution is logged with the agent, requested and resolved models. See #11353.
  • tm hook --pm-guard refuses a PM Agent/Task dispatch that would run any role other than PM, engineer (any *-engineer, including data-engineer) or research on a Fable model — whether the model comes from CLAUDE_CODE_SUBAGENT_MODEL, the dispatch's model argument, the deployed agent's frontmatter, or an inherit from a Fable PM session. When the model is unknown — inherit or no model: with no recorded parent model and no non-Fable [models] default, or an unreadable agent file — the dispatch is refused and asks for an explicit model: "opus"; model: "inherit" counts as inherit. tm launch and the Architect launch export the model they pass as --model as TRUSTY_MPM_LAUNCH_MODEL, which the guard reads after the statusline's session-model record, so a managed session's parent model is known before the first statusline render and in a headless session. New guard policy GP-051 fable-cap. See #11353.
  • tm architect up brings the Architect up from any state (no tmux server, no session, dead pane, pane at a shell) and attaches; bare tm architect does the same. It is idempotent, holds <dir>/.trusty-mpm/architect-up.lock, never stops a live claude (an unbound one is reported), and inside Claude Code only ensures a bound Architect. Flags --no-attach, --dry-run, --json; tm architect attach attaches only (#11364).
  • tm architect restart restarts the Architect only behind two gates: a fresh Intent Ledger checkpoint (else exit 4; a ledger it cannot read also refuses) and an idle box for 3 reads 5 s apart (else exit 2, or keep trying with --wait <s>). It then relaunches claude bound with resume, restores the poller, types one guarded resume line and appends one line to inbox/restart.log. It holds the tm architect up lock for the whole run and exits 1 at once if up or another restart holds it. Refused inside Claude Code; no bypass flag (#11365).
  • A dispatch of an ops agent (role: ops) now carries a ## Project accounts block listing each service's account name, scope and read-only verify command from [accounts.<service>] in .trusty-mpm.toml, plus the rule to name the account and run its verify command before any permission request (#11390).
  • .trusty-mpm.toml accepts [accounts.<service>] with account, scope and verify. A value outside the allowed characters, shaped like a credential, or under an unknown key rejects only the account map, never the rest of the file, and the block then says the map was rejected (#11390).

Fixed

  • tm hook now compresses chained Bash commands whose every segment is safe: cargo build && cargo test, git diff; git log -3, cd <dir> && … chains and a trailing | head/tail/wc/grep/sort/uniq/cut/tr. A destructive or unrecognised segment, ||, a background &, a subshell, a substitution, a redirect or a heredoc still leaves the command uncompressed, and isolation worktrees still skip chains. Compressed chains write savings rows, so the 💸 footer shows a figure instead of — (#10973, #11037).

Changed

  • The built-in opus tier default is now claude-opus-5-5 (was claude-opus-4-5), held in one constant OPUS_DEFAULT_MODEL_ID that also backs the Fable cap's fallback. See #11353.
72 earlier releases
  • 1.14.6 2026-11-16
  • 1.14.5 2026-11-14
  • 1.14.0 2026-11-13
  • 1.13.0 2026-11-11
  • 1.12.2 2026-11-10
  • 1.12.1 2026-11-09
  • 1.12.0 2026-11-07
  • 1.11.1 2026-10-30
  • 1.11.0 2026-10-30
  • 1.10.0 2026-10-30
  • 1.9.0 2026-10-26
  • 1.8.2 2026-10-10
  • 1.8.1 2026-10-10
  • 1.8.0 2026-10-10
  • 1.7.11 2026-10-05
  • 1.7.10 2026-10-05
  • 1.7.9 2026-09-27
  • 1.7.8 2026-09-26
  • 1.7.7 2026-09-26
  • 1.7.6 2026-09-26
  • 1.7.5 2026-09-25
  • 1.7.4 2026-09-24
  • 1.7.3 2026-09-24
  • 1.7.2 2026-09-24
  • 1.7.1 2026-09-23
  • 1.7.0 2026-09-23
  • 1.6.3 2026-09-18
  • 1.6.0 2026-09-16
  • 1.5.36 2026-09-13
  • 1.5.27 2026-09-10
  • 1.5.26 2026-09-10
  • 1.5.25 2026-09-10
  • 1.5.17 2026-09-06
  • 1.5.16 2026-09-02
  • 1.5.15 2026-09-01
  • 1.5.14 2026-08-31
  • 1.5.13 2026-08-31
  • 1.5.12 2026-08-31
  • 1.4.1 2026-08-14
  • 1.4.0 2026-08-14
  • 1.3.6 2026-08-12
  • 1.3.5 2026-08-10
  • 1.3.4 2026-08-03
  • 1.3.3 2026-08-03
  • 1.3.2 2026-08-03
  • 1.3.1 2026-07-31
  • 1.2.3 2026-07-28
  • 1.2.0 2026-07-27
  • 1.0.2 2026-07-25
  • 1.0.1 2026-07-24
  • 1.0.0 2026-07-24
  • 0.21.0 2026-07-23
  • 0.20.0 2026-07-21
  • 0.19.29 2026-07-21
  • 0.19.28 2026-07-20
  • 0.19.27 2026-07-19
  • 0.19.26 2026-07-19
  • 0.19.25 2026-07-18
  • 0.19.24 2026-07-17
  • 0.19.23 2026-07-17
  • 0.19.22 2026-07-17
  • 0.19.4 2026-07-09
  • 0.14.0 2026-07-01
  • 0.13.0 2026-06-30
  • 0.12.0 2026-06-27
  • 0.11.0 2026-06-24
  • 0.10.0 2026-06-17
  • 0.9.0 2026-06-16
  • 0.8.2 2026-06-16
  • 0.5.0 2026-05-28
  • consolidation 2026-05-26
  • 0.4.0 and prior

What each of these changed is in crates/trusty-mpm/CHANGELOG.md.

trusty-analyze

Code analysis sidecar

23 releases. crates/trusty-analyze/CHANGELOG.md on GitHub is the source this section is generated from.

0.13.0 2026-11-07

Breaking

  • trusty-analyze 0.12.7 ships a public-API break under the internal numbering policy (owner ruling 2026-09-26); recorded in scripts/semver-accepted-breaks/trusty-analyze-0.12.7.txt (Refs #8699).

Added

  • trusty_analyze::lsp: an on-demand stdio language-server client and child manager, off unless ANALYZE_LSP=1 (DOC-72 §2). It spawns rust-analyzer once per workspace root in its own process group, frames LSP messages with Content-Length, correlates requests with a per-request timeout, versions documents read from disk (didOpen at 1, didChange only when the content hash changes), and returns typed errors (lsp_disabled, lsp_not_installed, lsp_initialize_failed with the stderr tail, lsp_timeout, lsp_hung, lsp_stale, …). An idle server is killed after 2 minutes (ANALYZE_LSP_IDLE_SECS); every write is bounded by the request timeout, and a server that stops reading its stdin is killed as hung; a crashed or hung one is reported, not respawned in a loop; and no child outlives the daemon's idle exit, SIGTERM, drop or panic. No socket method uses it yet (Refs #9078).
  • With ANALYZE_LSP=1, every analyze.lsp_* call pushes two events to the trusty-events bus as source analyze (DOC-72 §4): ActionEvent::Tool with phase started, then finished or errored, sharing one call_id. The tool names are lsp.textDocument/references, lsp.publishDiagnostics and lsp.analysed_at. The Started event's parent_id is the call's parent_event_id; the terminal event's parent_id is the Started event. Events carry metadata only, in attrs: language, position, workspace, document version, snapshot, result_count, duration_ms and outcome (ok or the lsp_* error code). They carry no source text, no diagnostic messages and no absolute path under $HOME (shown as ~/...). Pushes only buffer, bounded and counted, so a missing or slow bus never delays or fails a call; a flush runs every second and once, bounded to 1 s, on idle exit, on SIGTERM and when serve --mcp ends on stdin EOF (the daemon task is now stopped through its exit path instead of aborted). With the flag off nothing is pushed and the bus socket is not resolved (Refs #9078).
  • analyze.lsp_diagnostics, analyze.lsp_references and analyze.lsp_analysed_at on the analyze socket, with the MCP tools lsp_diagnostics, lsp_references and lsp_analysed_at (DOC-72 §5). Positions are path plus zero-based line and column; every answer carries document_version and a <git head>+<dirty hash> snapshot, computed under a 10 s deadline and a 2,000-file / 64 MiB budget and refused as lsp_snapshot_failed past either; a version other than the file's current one is refused as lsp_stale; a file's diagnostics, and the lsp_analysed_at barrier, come from a pull (textDocument/diagnostic) after the server has loaded, because rust-analyzer pushes a set only when it changes; the pull is merged with the cargo check push (rust-analyzer's pull omits check findings such as E0382), counted only for the opening version because cargo check runs only at startup and rust-analyzer stamps a push with the version current when it publishes; without it the answer is complete: false with reason: "check_not_run"; a newer sync landing while the pull is in flight is refused as lsp_stale; pushed sets for other files that arrive without a version are marked unversioned; untracked files inside new directories are part of the snapshot; scope: workspace takes a file and answers complete: false (a directory is refused); lsp_analysed_at blocks until a version is analysed and answers ready or timed_out with version_seen, capped at the MCP client timeout minus 30 s. Every read waits for rust-analyzer's experimental/serverStatus to report quiescent, so a loading server's empty answer is never returned; past the request timeout the call is refused as lsp_not_ready, and a server that reports health: "error" (it could not load the workspace, for example a broken Cargo.toml) is refused as lsp_workspace_failed without echoing its message. An optional parent_event_id is accepted. Off unless ANALYZE_LSP=1; every refusal is error code -32006 with data.kind (lsp_disabled, lsp_not_ready, lsp_file_not_found, lsp_outside_workspace, lsp_unsupported_language, lsp_timeout, lsp_hung, …), which the MCP tools keep in their error text. A server answering "content modified" (-32801) while it loads is retried with backoff inside the request timeout (Refs #9078).
  • analyze.complexity_hotspots, analyze.smells and analyze.quality accept an optional scope: paths (changed files) and/or path_prefix (one directory). A scoped call walks only search.chunks.list pages filtered by each entry, never the whole corpus, and answers with scoped: true and scope: {paths, path_prefix, unmatched}, where unmatched names each entry that selected no chunks. A trusty-search that ignores path_prefix is refused rather than allowed to page the corpus. An entry naming the index root is invalid_params: ., /, a/.., and any absolute spelling that resolves to the index root_path — /repo, /repo/, /repo/., or a symlink to it. An absolute entry containing .. is also invalid_params, because trusty-search resolves a symlink before .. and /elsewhere/link/.. can land on the root. A scoped walk whose first page reports as many chunks as the whole index is also refused as invalid_params, before it pages the corpus; omit the scope for the corpus-wide result. More than 500 entries is invalid_params; paths: [] selects nothing (Refs #10781).
  • Unscoped calls to the same three methods are served from a per-index cache instead of walking the corpus in the request. An entry is current while search.index.status reports the same last_indexed, chunk_count and status it was computed against, and for at most 15 minutes. Otherwise it is served with cache.status: "stale" and a stale_reason while one background refresh runs. On a cold cache a request of the pre-#10781 shape waits for the corpus walk and returns its result, as before, and the walk fills the cache; concurrent cold callers share one walk. A caller that passes the new optional field allow_warming: true instead waits up to 5 s, then gets the new retryable error code -32007 with data: {"kind": "analysis_warming", "retryable": true}. A warm or stale cache answers at once either way. A failed refresh is an upstream error, and no new refresh starts for 30 s. An index trusty-search reports as not found answers not_found and its cache slot is dropped; a failed status read creates no slot. Every unscoped response adds scoped: false and a cache object; the existing fields are unchanged. The cache keeps the top 1000 hotspots, so a top_n above 1000 also returns hotspots_capped_at: 1000 (Refs #10781).
  • A chunk page that fails in transport (for example a timeout) is retried up to 4 times, halving the page limit down to 100, and the smaller limit is kept for the rest of the walk. A walk that still fails names the cursor, the attempt count and the last limit. A page that trusty-search answers with an error is not retried, and the #6043 short-export refusal still applies (Refs #10781).

Fixed

  • analyze.lsp_analysed_at no longer ends its bus event with outcome: ok when the barrier timed out. The terminal event now carries state (ready or timed_out), plus version_seen on a timeout, and a timeout reports outcome: timed_out; the ready case is unchanged and the RPC answer is unchanged (Refs #10087).
  • TRUSTY_ANALYZE_SOCKET now sets the socket serve binds and every client dials; a --socket flag still wins, and a relative path or a missing parent directory is an error (#10365).

0.12.7 2026-10-10

Breaking

  • TrustySearchClient::new takes the trusty-search socket path instead of an HTTP base URL; TrustySearchClient::from_env builds one at the standard path (#9214).
  • TrustySearchClient::base_url() is removed; socket_path() names the socket the client calls (#9214).
  • The --search-url flag and the TRUSTY_SEARCH_URL environment variable are removed. trusty-analyze reaches trusty-search only over its Unix socket, with no TCP fallback (#9214).

Added

  • analyze.smells rows carry a smells array naming what the detector found. The method selected chunks by running the detector and then discarded its output, so a row said "this chunk smells" without saying of what — the only other string on it, match_reason, is the SEARCH daemon's hit reason and says nothing about code quality. Consumers that group by smell had nothing to group by: the analyze dashboard's Smells view rendered "No smells detected for this index" over a corpus whose own quality card reported thousands. quality::smelly_chunks_with_smells keeps the detector output the existing smelly_chunks throws away; that function is unchanged, so no caller of it moves (#6155).

Fixed

  • trusty-analyze mcp now answers initialize and tools/list at once instead of waiting up to 30 s for the daemon to start, so a loaded host no longer times out the console's handshake. The daemon auto-start runs once in the background; the first tool call waits for it, and an auto-start failure comes back as that call's error instead of exiting the bridge (#8279).
  • trusty-analyze start now spawns the daemon through the shared detached helper, so it starts in its own session and survives a group kill or Ctrl-C aimed at the caller (#8783).
  • trusty-analyze now honours TRUSTY_SEARCH_SOCKET: every trusty-search call goes to that socket, or the standard path when it is unset. A missing socket is an error naming the path (#9214).

Removed

  • The admin UI is gone from this crate: the ui/ Svelte tree, the committed ui/dist/ bundle, build.rs and its Vite step, the build/exclude keys in Cargo.toml, and the rust-embed and mime_guess dependencies. The Svelte source moved to crates/trusty-console/ui-analyze, and the console serves the dashboard at /tools/analyze/ over this daemon's existing analyze.* socket methods. Nothing had referenced these assets since #6287 deleted this crate's embed and its HTTP listener, so no public item changes shape and no version bump is owed on that account (#6155).

Documentation

  • Retired stale POST /review/github-pr mentions in the GithubPrRequest doc comment and the PRD/COMPONENTS specs; they now name the analyze.review_github_pr JSON-RPC method served over the Unix socket (#5176).

0.12.6 2026-09-03

Changed

  • The MCP dispatcher dials the daemon through the shared trusty_mcp::daemon_bridge_json_rpc. mcp::rpc_client::call built its own JSON-RPC frame, called trusty_common::uds::send_framed_request_capped and unpacked an RpcResponse by hand — the same transport trusty-memory's stdio bridge carried a second copy of. Both now run one implementation (#6316)
    • The 32 MiB response budget and core::mcp_client_timeout() are unchanged; they are passed to the bridge rather than to the client. No request rewriter: the dispatcher has already built the exact params each analyze.* method expects
    • mcp::stdio::run is untouched. This crate's MCP surface is a tool translator with its own tools/list and its own #917 response-size guard, not an envelope forwarder, so the stdio loop stays where it is
    • A transport failure and a daemon-side JSON-RPC error still both surface as DispatchError::Transport naming the failing method. The message now carries the daemon's error code as well: <method> over <socket>: <message> (<code>)
20 earlier releases
  • 0.12.5 2026-09-02
  • 0.9.0 2026-08-10
  • 0.8.0 2026-07-27
  • 0.7.4 2026-07-27
  • 0.7.3 2026-07-09
  • 0.7.2 2026-06-16
  • 0.7.0 2026-06-09
  • 0.6.0 2026-06-09
  • 0.5.1 2026-06-07
  • 0.5.0 2026-06-03
  • 0.4.2 2026-06-02
  • 0.4.1 2026-06-01
  • 0.3.0 2026-06-01
  • 0.2.1 2026-05-31
  • 0.2.0 2026-05-29
  • 0.1.10 2026-05-22
  • 0.1.6 2026-05-20
  • 0.1.5 2026-05-20
  • 0.1.2 2026-05-11
  • 0.1.0 full Phase 1 + Phase 2 static analysis engine

What each of these changed is in crates/trusty-analyze/CHANGELOG.md.

trusty-review

LLM code review

52 releases. crates/trusty-review/CHANGELOG.md on GitHub is the source this section is generated from.

0.42.0 2026-11-16

Fixed

  • A review_pr miss on an index the trusty-search daemon lists as parked now says the index is parked and points at trusty-search ensure_project, instead of telling the operator to index the checkout (#11417). SearchClient gains list_parked_index_ids, which defaults to none.

Changed

  • MCP review_pr on a repo with no trusty-search index now returns a DEGRADED diff-only review (banner, degraded_reason, search unavailable in context_sources) instead of an error, unless TRUSTY_REVIEW_REQUIRE_SEARCH=true. The webhook and CLI GitHub-PR runs still fail closed (#11409).

0.41.0 2026-11-11

Breaking

  • ParsedReview gains the public questions field and is not #[non_exhaustive], so a caller that builds one with a struct literal must add questions. This is why the release is 0.41.0 and not 0.40.3. cargo-semver-checks against 0.40.2 reports constructible_struct_adds_field (#10782).

Added

  • Each finding in the review_diff / review_pr result and the CLI JSON now carries quoted_snippet, the code it quotes from its cited file at line. The reviewer schema requires it; a reply from a reviewer that omits it takes the first code the finding body quotes (#10782).
  • The reviewer schema has a top-level questions array for "confirm / check / verify X" notes with no code to quote. The result reports them as questions, absent when empty; no gate, grade or verdict reads them (#10782).
  • A finding with no quote is withheld before grading and the line-citation gate, under the new reason class no_quote_pregate, so withheld_by_reason and the withheld headline still count it. An APPROVE with no surviving finding whose only withholds are no_quote_pregate reads verdict_status: parsed, not all_withheld, and carries the new pregate_only_approve: true (absent when false). Any gate withhold, alone or beside pregate drops, keeps all_withheld on APPROVE; a blocking review keeps suppressed_reject whatever the reason. Each review logs one info event with the pregate and gate-withhold counts (#10782).
  • The line-citation gate places a finding that carries quoted_snippet on that quote alone: the cited line must hold it, or the finding moves to its one occurrence, or is withheld. A quote absent from the file is withheld, and every other snippet the finding quotes must still be in the file. A quote that is only a bare identifier (step_4, run()) or a path is no quote, by the same rule as a backtick span: the finding takes its body's quote, or is withheld no_quote_pregate (#10782).

Fixed

  • The line-citation gate and the citation-integrity check no longer withhold a finding for a backtick span in its body or consequence that it labels as an exploit payload, an example input or a proposed fix ("a payload such as ' OR '1'='1' --", "e.g.", "suggest", "replace with", "should be", "Fix:", "… x instead"), or that equals its own suggested_replacement. The finding still needs a quote of the cited file that places it; a payload-only finding, an unlabelled quote absent from the file, and a cue in another sentence still withhold it (Refs #10964).
  • Citation integrity drops two findings on one line as a mutual contradiction only when they make the same defect claim (title) and their file quotes conflict: quoted_snippet when present, else the unlabelled backtick spans. Apostrophes in prose and quote marks inside SQL are no longer read as citations, and an unterminated quote mark no longer opens a span to the end of the text (Refs #10964).
  • The reviewer schema now says backticks in body quote the file, and that a proposed fix goes in suggested_replacement or a payload is labelled (Refs #10964).

0.40.2 2026-11-08

Fixed

  • Inline review comments no longer drop a hunk when a removed line starts with -- (a SQL or Lua comment, shown as --- …) or an added line starts with ++ (shown as +++ …). The diff walker now reads each hunk by its @@ line counts before it looks for file headers, so findings on the later lines of such a hunk anchor on the correct side instead of reading off_diff (#10762).
49 earlier releases
  • 0.40.1 2026-11-07
  • 0.40.0 2026-11-07
  • 0.38.5 2026-10-10
  • 0.38.4 2026-10-09
  • 0.38.3 2026-10-09
  • 0.38.2 2026-10-09
  • 0.38.1 2026-10-09
  • 0.38.0 2026-10-07
  • 0.37.0 2026-09-30
  • 0.36.1 2026-09-26
  • 0.36.0 2026-09-13
  • 0.35.0 2026-09-06
  • 0.34.0 2026-09-04
  • 0.33.1 2026-09-03
  • 0.33.0 2026-09-03
  • 0.32.0 2026-09-02
  • 0.31.0 2026-08-31
  • 0.30.0 2026-08-31
  • 0.20.0 2026-08-19
  • 0.15.0 2026-08-12
  • 0.14.1 2026-08-11
  • 0.14.0 2026-08-10
  • 0.13.0 2026-08-10
  • 0.12.0 2026-08-10
  • 0.11.0 2026-07-27
  • 0.10.1 2026-07-23
  • 0.10.0 2026-07-21
  • 0.9.2 2026-07-17
  • 0.9.1 2026-07-16
  • 0.9.0 2026-07-11
  • 0.8.1 2026-07-11
  • 0.8.0 2026-07-10
  • 0.6.1 2026-06-25
  • 0.6.0 2026-06-24
  • 0.5.0 2026-06-24
  • 0.4.0 2026-06-18
  • 0.3.16 2026-06-18
  • 0.3.15 2026-06-18
  • 0.3.10 2026-06-16
  • 0.3.8 2026-06-09
  • 0.3.6 2026-06-07
  • 0.3.5 2026-06-03
  • 0.3.4 2026-06-03
  • 0.3.3 2026-06-03
  • 0.3.2 2026-06-03
  • 0.3.1 2026-06-03
  • 0.3.0 2026-06-03
  • 0.2.0 2026-06-03
  • 0.1.0 2026-05-28

What each of these changed is in crates/trusty-review/CHANGELOG.md.

The rest of the workspace

Only the flagship crates are published here. Every other crate keeps its changelog alongside its source — see crates/ in the repository. Nothing on this page is hand-written: it is generated at build time from the same files, and the repository remains the source of truth.