68 KiB
archied changelog
[1.36.0] - 2026-09-22
The dashboard declares tools an in-browser agent can call
- The dashboard registers seven WebMCP tools. Read-only:
list_tasks,get_task,recent_events,daemon_health. Recovery:retry_task,stop_task,cancel_task. An agent operating the dashboard now reads structured state and recovers stuck work instead of scraping the UI. - Every tool wraps the same
ui/src/lib/api.tsclient the dashboard itself calls, so an agent and a person exercise one path rather than two. archie-uicannot be attributed to the agent that called it: a page can invoke its own tools, so an agent's call and a page script's call reach the server identically. Tools therefore record the signed-in identity and a source, and never that an agent acted.
The four judgement verbs are deliberately absent
approve,reject,abandonandarchiveare not exposed.internal/domain/taskactionsrecords only a source string and stampsKindHumanApproved, so an agent's approval would be written down as a person's. An agent that can approve can approve its own work, and the one control this design exists to keep is a human accepting it. Those verbs land once task actions carry both the acting identity and the authorising principal.
consequentialHint does not gate anything, and the release says so
- Chrome 152 accepts
consequentialHintat registration and omits it fromgetTools(), so no client sees it and the browser cannot enforce what it does not surface. The three recovery tools instead require an explicitconfirm: true, which is a deliberate-action signal rather than a security control an agent can trivially satisfy. The boundary that holds is the session: the tools cannot exceed what the signed-in operator can already do.
Two ways this fails silently, now recorded next to the code
- A secure context is required. On a plain-HTTP origin
document.modelContextis absent even with the flag, and nothing reports why.mkcertcovers local development. archie-uiembedsui/distinto its binary, so rebuilding the distribution is not deploying it; a running process can serve a bundle with no tools while the built one has them.
Binaries built by the development build task can be updated
task buildpassed theinstalltypestamp for three of six commands, soarchie-messaging,archie-state-storeandarchie-uiwere built without it. An unstamped binary reportsunknownand the self-update path refuses rather than guesses, which is how an operator ends up unable to update withinstall type is unknownand no way forward. All six are stamped now, and a test reads the build task and fails naming any command line that omits it.
[1.35.0] - 2026-09-22
The Messaging Service is severed from the store and workflow runtime
cmd/archie-messagingno longer linksmodernc.org/sqlite,internal/store,internal/domain/workfloworinternal/agentexec. All four arrived through a single import ofinternal/app/controlplane; the client half now lives ininternal/infrastructure/controlplanerpc, and the app package keeps the server, the registry, the seeds, the validators and the schema derivation.- The PRD's guard now exists, and passes:
cmd/archie-messaging/architecture_test.gomeasures the linked binary against the banned list, withinternal/domain/workflow/taskas an exception checked in both directions. Before this the guard had never been written -- because it failed. A messaging process linking a SQLite driver and the workflow engine is a severance hole, not a design choice. - The resource documents are type-aliased back into the app package, so there is one definition read under two names rather than a copy per side.
Channel lifecycle reaches the dashboard
- The Messaging Service reports each channel's state, recorded per channel: one channel's failure
marks only itself, a channel that cannot start is marked failed with its reason rather than only
logged, and one that returns is marked stopped.
internal/channels/status.Managerfinally has a production writer. PutChannelStatus/ListChannelStatuscarry that to the dashboard, following the config-snapshot split. The store replaces the reported set, so a channel the reporter stops mentioning is deleted rather than left as a stalerunning; an empty report clears the table, because hosting no channels is a fact worth reporting; and an id-less row is refused, since it could never be replaced or removed by a later report.Service.ReloadChannelroutes to the channel's own seam and refuses a channel that cannot reload with a reason rather than accepting and ignoring it.ReloadSupportedwas previously declared by name, so a channel could advertise reload with nothing behind it; one function now decides both.
Action playbooks: a second playbook shape, with typed results
- A playbook is now one of two shapes: a workflow playbook (exactly one
workflowaction, unchanged) or an action playbook (one or moremoduleactions in order). A playbook does not mix them. argsandwhenshare one compilation path and one evaluation context, andactions.<id>is typed by the kind that action runs, so a misspelled result field, an unknown id, a map index,in, a comprehension, or a shape-foreign key (kindon aworkflowaction,workflowon amoduleaction) is a load failure rather than a dispatch failure.- Action playbooks load but do not run yet. Routing arrives with the first side-effecting position; the daemon warns at load rather than accepting them silently.
- The CEL environment is built per playbook, because result typing depends on the ids a playbook declares.
ModuleRegistry.DecodeResultturns a kind's flat map result into its typedResultbefore a later expression reads it. Without it a typed read evaluated against a shape the run never produces, and so silently read null.
[1.34.0] - 2026-09-22
Dashboard: an actionable authentication failure state
- An unauthenticated document request now receives a self-contained HTML page with a
paste-the-token form, instead of plain text. Previously the whole dashboard --
index.html, the bundle and every asset -- sat behind the token check, so a bookmarked bare URL returned atext/plain"unauthorised: open the dashboard URL archied logged at startup", the SPA never ran, and none of the dashboard's existing failure surfaces were reachable at all. - The response is still 401, with
Cache-Control: no-storeand a CSP. It is deliberately not a 200: a login page served as success would be cacheable, would read as healthy to anything that checks status alone, and a shared cache could hand it to the wrong person. - API and stream callers are unaffected.
application/json,*/*andtext/event-streamrequests keep the plain-text 401, with a test pinning that split so an XHR or anEventSourcereconnect can never start receiving a page. - The auth banner gains a Retry action that re-runs the dashboard's bootstrap reads, so a recovered session is actionable without a browser-level reload. It deliberately does not re-drive the SSE stream, which already reports its own reconnect state.
- The page's guidance stays generic: it points at the dashboard URL the daemon logs rather than naming where the token is stored, so the HTTP layer is not coupled to the token file's location.
[1.33.0] - 2026-09-22
EDA: playbook action arguments are compiled at load (J2)
- A playbook action may now declare
args, compiled as CEL expressions when the playbook loads.argsandwhenshare one compilation path and one evaluation context, so an expression written for one works in the other. - Behaviour change: an
argsexpression that does not compile is a load failure, not a runtime miss — the same reject-at-load rule already applied to a binding collision and to an unknownactions.<id>. In practice, a daemon loading a playbook with a malformedargsblock fails at startup rather than part-way through a dispatch. - Compiling several arguments now reports the first bad key deterministically (keys are sorted). Previously it ranged a map, so the reported bad key could differ between runs and the failure was not reproducible.
argsis deliberately inert for now: parsed, compiled and stored, but no shipped position consumes it yet. A consumer arrives with the first side-effecting position.
Not yet decided: how an args evaluation failure behaves at dispatch. The when predicate rule
(error, skip, log) does not obviously transfer, because args is data rather than a predicate, and
silently skipping an action whose arguments failed to evaluate would be a silent failure of exactly
the kind this release train has been removing. It currently propagates the error to the caller, and
that asymmetry is documented at the code. Ratifying it against a real consumer is tracked separately.
[1.32.1] - 2026-09-22
Host packaging: archie-messaging now ships and updates
cmd/archie-messagingis included in the distribution zip, andscripts/archie-update-installinstalls and restarts it.- Why this mattered. v1.30.0 moved Telegram, email and webhook out of
archiedand intoarchie-messaging, but nothing packaged the new binary. A host install that upgraded across v1.30.0 therefore ran on with no messaging process at all:archiedlogged that it stopped serving the telegram gateway, nothing replaced it, and Telegram stayed dead — with no error logged anywhere, because nothing was failing. The process simply did not exist. - The host component set was enumerated in four places — the zip's build loop, the zip's copy list,
the installer's service list, and the installer's binary list — and
archie-messagingwas missing from all four. The duplication was the defect: a single test now derives the host-run commands fromcmd/and asserts that all four lists agree with it, so adding or extracting a process fails until it is packaged. deployments/systemd-user-service.mdgains the messaging unit. That is load-bearing rather than cosmetic: the updater refuses an update when a service unit is missing and tells the operator to create one from that document, which previously had nothing to offer.
Upgrade note: a host that reached v1.30.0 without a messaging unit has been running with no
Telegram, email or webhook. Updating to this release installs and starts archie-messaging.
[1.32.0] - 2026-09-22
EDA: playbook action ids, and a dispatch idempotency ledger
- Action ids (J1). Every action a later action references must now carry an
id. Ids are validated at load -- unique and stable-identifier-shaped -- and awhen/argsexpression naming an unknownactions.<id>is a load failure, not a runtime miss, the same reject-at-load rule already applied to a binding collision. - Reference classification is strict at load. A map-index reference (
actions[expr]), a non-literal index, and a reference to an unknown prior id are all rejected when the playbook loads; references are classified by identifier versus static access, and the when-compilation helper is shared between the load-time and runtime paths. - Dispatch idempotency ledger. A
playbook_dispatchesrow records a dispatch against(playbook_id, playbook_version, event_id, action_id), written withINSERT OR IGNOREand surfaced through the sharedstore.ErrAlreadyDispatchedsentinel. Module, channel and forge action positions are therefore at-most-once, which is what unblocks the side-effecting positions that were previously rejected at load. - The ledger is wired into production.
stateStoreDepsconstructsDeps.PlaybookDispatcher, so the ledger RPCs no longer answerUnavailablein a running daemon. This defect was found by the pre-release review of this change rather than by its tests.
Behaviour change: a playbook whose actions reference actions.<id> without declaring that
id, declare a duplicate id, or index actions with a non-literal value now fails to load
instead of failing at dispatch time.
[1.31.0] - 2026-09-22
Dashboard replaced: Preact -> Vue 3 + shadcn-vue
The dashboard is now a Vue 3 application built on shadcn-vue. This is the large piece of this release; it touches every page.
- History routing, a grouped topbar built from components, the violet palette with self-hosted Geist, and a command palette with live tasks.
- Chat is a FAB-opened panel that behaves like a chat window and scrolls its
transcript with
MessageScroller. - Task actions and dangerous approvals are deliberate; the task detail page is a single head strip with honest states; the run page settles its head around an attempts pager, which is hidden for zero or one attempt.
- Appearance and System navigation were refined, with a four-column appearance grid on wide screens.
- A copy pass removed redundant tails and restating descriptions, empty states derive their own titles, and a stuck log read surfaces as an error rather than eternal loading.
- The capture pane is bound to the list it sits beside and reads beside it, not over it.
Events is one page with three tabs
- Inspector, Bindings and Mappings are now tabs of a single Events page.
- The nav entry is always live, the tabs prune by capability, and the three old paths survive as bookmarks that redirect to their tab.
- Tab selection lives in one module (
ui/src/events/tab-selection.ts) that the page uses, rather than a copy of it. feat(tools): a capture-intake sender exists for exercising the Events surfaces.
The channel travels with the message
SessionSource.Platform -- the first component of the session natural key, and
the column GetByChannel filters on -- was a constant "web", because one
Router serves every channel and named itself after its own. Two things followed:
the per-user identity policy was real code that no production call site reached
with a real channel name, and a curator pass's write to agent-user memory could
not be read back by the turn that wrote it, because that turn resolved no
UserID.
- The frontends now name the channel they carry (
telegram,email,webhook,web), the channel rides the message, and the Gateway stamps the session from it -- falling back to its own name when a sender names none, which is what every existing row holds. - Identity resolution reads the platform off the inbound, so
carriesPersonIdentitydecides something in production: telegram and email resolve a person, the dashboard and a webhook resolve nothing, and an unknown platform fails closed. - Behaviour change: a chat whose active session came from
/newor/branchre-resolves once, because session lookup now filters on the channel's real name. A chat without/newre-resolves onto the same deterministic id and heals silently. - Deliberately unchanged:
TurnRunnerConfig.Channelstill keys tool availability, so no channel gains or loses a tool in this release.
Control plane: runtime administration, history, and derived schemas
- The State Store's gRPC server now carries runtime administration, and each process reports which settings version it is running; the dashboard shows it.
- Resource history is exposed so edits can be reviewed and rolled back, and
resource schemas are derived from the documents they describe -- the document
editor reads the derived schema (labels from
title,format: duration->1m0s, hints fromdoc), and the abandoned schema-driven card is deleted. - A live settings replacement is checked before it is switched to, the daemon's state surfaces are folded into one boot step, and a SIGHUP reload re-applies database-owned settings.
- Forge intake settings that nothing reads are now rejected at load.
Configuration: judged where it is owned
- The runtime config overlay is deleted; database-owned settings are judged at the layer that owns them rather than on the file load path.
- A file overlay no longer drops the fields it omits inside a map entry, and a kind with no stored value leaves the file's value in force.
- The default config path is resolved in one place, and the channel document has its own shape.
State Store: offline backup, restore, validate and rollback
- New offline backup / restore / validate / rollback path.
validateruns boot's own gate; it no longer replaces the database, reads a kind the store does not hold as its seed, or touches the daemon's log.- The recovery control-plane server is built on the shared step vocabulary.
Workflow: one step vocabulary and a gate.diff-rules replacement
- The step vocabulary is owned by a family manager registered at each
composition root, replacing
.archie/gate.gowithgate.diff-rules, which fails closed on a change it never read. - Workflow roots are pinned and the dead vocabulary is removed; the workflow pin crosses the wire.
- The schedules resource skips the legacy cron jobs it cannot hold, and the scheduling policy carries the label its trigger pairs with.
EDA: the playbook coordinator reaches production selection
- The playbook coordinator is wired into production workflow selection, so a
matching playbook decides a real task's workflow at the daemon's definition
pin, inheriting idempotency from
TaskEnvelope.
Reliability and operations
cronstorewrites a schedule interval as a duration, not nanoseconds -- a persisted1mno longer round-trips as60000000000.- Worktrees are cut from the mapped base branch, and the Gateway's worktree
manager is built so it serves
review_pr. - A late max-uptime callback no longer resurrects teardown state.
- The state store dial installs client keepalive, and its server policy accepts the dialer's pings; control-plane watch streams re-establish themselves after they end.
/statusserves only the health facts a reply can carry; the daemon reports systemd readiness and loop-driven watchdog heartbeats.- A webhook route path no longer becomes a sender identity.
- A chat session compresses when a turn crosses its context budget.
- The MCP client supports per-server
parallel_tool_calls, carried through the tool-settings projection. - Resource request-ID idempotency keys are scoped to their kind.
- The session curator addresses memory by the deployment's bot user.
[1.30.0] - 2026-09-20
Messaging Service extraction completed
Telegram, email and webhook no longer run inside the daemon. They run in
archie-messaging and reach the daemon through the Gateway's ChatContract
over gRPC; the daemon builds no channel and no gateway.Router (#899).
- The operator surface is re-wired in the extracted service:
/update(built from its own[chat.telegram]commands), the update-report relay, theshow_tool_callsprojection, and/restart's config reload, which re-resolves the token and allowlist from this service's own file and refuses a reload that resolves to an empty token. /versionis sourced from the Gateway through a newChatSnapshot.Versionfield.cmd/archie-messagingdeliberately takes no version ldflags, so it cannot report a partial upgrade as a matched one.- Channel configuration is validated at composition, so a configured-but- invalid front-end fails the service instead of being silently dropped.
- The inbound rate limiter moves to the Gateway, which now owns the only Router and is therefore the one place a per-sender budget can apply to every channel's turns.
RunningVersions and ReleaseAnnouncements are deliberately left unwired:
both need a component's real running version, and the only versions this
process can obtain are another binary's stamps. See
docs/architecture/migration-decisions.md.
Configuration
diff_cap_lines = 0now switches the cap off, as the dashboard and the workflow step have always documented. It was a plainint, so the defaults pass could not tell an absent key from an explicit0and rewrote both to 400 -- the documented way to disable the cap was unreachable. Behaviour change: a config that set0was silently getting 400 and will now get no cap.- The same ambiguity was failing open per identity:
configForIdentityassigned the identity's value unconditionally, so an identity that merely omitteddiff_cap_linesoverwrote the shared cap with the zero value and disabled the safety rail entirely. An identity now overrides only when it sets the key. - A diff-cap park no longer tells the operator to "approve manually". A parked task's actions are retry, abandon and reject, so there is no approve path out of diff-cap; it now names the two things that work.
[health].dependency_timeoutmakes the daemon's readiness probe timeout configurable, matching the extracted services.
Chat rendering
/statusand/tasksare line-oriented reports but separated lines with a plain newline, which is CommonMark's soft wrap: the Telegram renderer joined them into one run-on paragraph. The renderer implements CommonMark's hard line break and the reports use it. Ordinary prose still soft-wraps.- A park reason is capped to its first line in the
/taskslist. It carries whatever the failing stage reported, which for a lint gate is a whole tool log; unbounded, one parked task pushed every other task out of the reply. The full reason is unchanged on the task, intask_listand on the dashboard.
Review
- Review findings are posted as line-anchored PR comments with suggested fixes, and a suggestion must be single-line -- GitHub applies a multi-line suggestion only up to its first line, silently corrupting the file (#891).
- The remediate workflow handles PR review comments and requested changes.
/status health surface
/status reports the live broker connection and the last-known outcome of this
process's chat-model calls, alongside the queue depth and runtime block (#892).
- A provider error is flattened and capped, so a multiline or oversized error cannot break the one-line-per-fact layout.
- Concurrent model calls no longer reorder: the completion time is taken before the lock, so the newer outcome wins rather than whichever goroutine stored last.
- Curator model calls are recorded too. They bypass the chat path, so a daemon whose only recent model traffic was curator work reported "no calls attempted yet" while the provider was demonstrably reachable.
- Title generation does not record health. It runs detached under its own bound and its error is swallowed, so a cosmetic timeout was overwriting the line with "failed".
Known limitation: the container, poll and channel sections cannot render. Those facts live in the daemon, but the only Router lives in the Gateway process, which has no container pool and no poll loop. The command description says what actually appears; restoring the rest needs a cross-process health contract.
Dashboard and containers
- The served lifecycle vocabulary reaches an already-painted page, the page
registry matches the routes, and the build verifies
ui/distis fresh instead of never looking at it (#890, #898). - The container reaper no longer deletes teardown bookkeeping early, so a released container is not torn down twice (#893).
Build
task fmtrunsgolangci-lint fmt-- the same engine and configuration the gate checks with. Standalonegofumptcannot reproduce gci's import grouping, so formatting and linting could not converge and no command satisfied the gate (#901).goimportsis deliberately absent from the formatter set: it re-reads and re-parses the module index once per file, ~135s across this tree.
Fixed
Config.Cloneleft the new diff-cap pointers aliased, so a clone could write through to the published snapshot (#902).- A curator asking for a completion on a daemon with no configured provider panicked the process; it now returns an error.
[1.29.0] - 2026-09-19
Security
google.golang.org/grpcbumped pastv1.84.0to patch GO-2026-6443, a server panic via a missing:authority/Hostheader reachable througharchied's gRPC gateway server. No stable release carries the fix yet, so this pins the pre-release commit that does (v1.85.0-dev.0.20260825072537-93e31b48545e) untilgrpc-gocutsv1.85.0.
Memory engine unification
One memory engine, addressed by four typed scopes (global, agent, user, agent-user), replaces the two disconnected stores that previously existed: a chat-tool-written store nothing ever read back, and a curator-written store with no scope model.
internal/domain/memoryis a CRUD contract (Create/Get/Query/List/Update/Forget/Revisions) with retained revisions; updates supersede rather than overwrite, so a crash mid-write leaves a recoverable extra history entry, never a lost prior state (#886).- The chat turn reads its subject's scopes synchronously before the prompt is
built, rendering recalled records into a
<memory>block; a failing or panicking engine degrades to no memory block rather than failing the turn (#887). - The model writes memory through four id-addressed tools
(
memory_create/update/delete/list), scoped to the turn's own resolved identity -- the model chooses the scope kind, never the ids, so it has no path to write into another user's memory (#888). - The session curator now derives the agent-user scope from the session's own user-role messages, instead of keying facts by a session id that stops meaning anything once the session ends.
- A channel with no resolvable user (dashboard, webhook) reads and writes only global and agent scope; a webhook's route path is never treated as a user identity.
- Fixed a review finding on the read path: the per-scope record limit was being applied as one shared total across all scopes rather than per scope, and the rendered block's byte cap was checked before template escaping rather than after, letting the escaped block exceed it.
Work intake
- feat(workintake): define review reaction envelope
- feat(workintake): authorize review reactions by owned PR
Cleanup
- Deleted the unwired plugin capability-host health surface
(
Host.Health/Host.Manifests/ModuleStatus) and theModule.Healthchain it fed -- including themcp/builtintool-provider health implementations and their aggregation intools/provider.Registry-- none of which had ever had a production caller (#882).
[1.28.0] - 2026-09-19
Task detail and attempt attribution
- New per-run task detail page, showing each attempt's own events, changes and configuration rather than a single merged view.
- Events, changes and config are attributed to the attempt that produced them;
EventcarriesattemptandReadTaskLogRequestcarriesstage. - The task detail page remounts when the task id changes, so switching tasks does not leave the previous one's data on screen.
Configuration
- Services are keyed by name through a registry, so the State Store and Gateway listen addresses are configurable rather than fixed.
- A partial overlay no longer clears the fields it omits — an overlay that set one key was wiping the rest.
- The Services map is cloned on publish, so an overlay can no longer mutate the snapshot other readers are holding.
- The service context is enforced at registration.
Logging and park reasons
- A parked run's reason is recorded in its own attempt log, in both the daemon and the workflow, so "why did this park?" is answerable from the task itself.
TaskRegistry.Writetakes the caller's context.
Worktrees and containers
- Retries recover from path-type conflicts instead of failing opaquely, and the underlying cause is no longer masked.
- The worktree is handed back on every sandbox exit.
- A container's max-uptime reaper is cancelled on release, the cap stays active through the grace period, and teardown is claimed once rather than by cancelling a timer.
Daemon and startup
- Every identity's forge and repos are swept at startup, including the root forge.
- A cancelled startup is no longer reported as an exhausted sweep budget; an exhausted budget names itself as the cause.
Other
grepcan use the workspace codesearch index.- Forge labels are matched when namespaced, so label-driven routing works for namespaced labels.
- Triage is given criteria for choosing between workflows, and a workflow is required when a code change is needed.
- The Telegram adapter restart callback is wired into the chat router.
- The
ARCHIE_REACTIONSfan-out stream is provisioned and bounded by age. archie-uiexits cleanly when graceful shutdown overruns its deadline.- The dashboard's event pump retries priming until the State Store answers.
- The update writer emits the agent version sidecar with a real newline.
- The UI finishes its Preact port; the imperative DOM layer is deleted.
- Collection fields decoded over the State Store gRPC contract are never nil
slices, so an empty collection marshals as
[]rather thannull.
[1.27.0] - 2026-09-17
Dashboard task logs
- Task logs are readable from the dashboard again, through a new State Store
read contract (
ReadTaskLog,StreamTaskLogContent) rather than the UI process reaching into the daemon's log files. Wired server-side by the standalone state-store process. - Added
GET /api/tasks/{id}/logs/download, serving an attempt's raw log as atask-<id>-attempt-<n>.logattachment. - A process that cannot read logs and an attempt that recorded no log are now distinguishable ("cannot read" answers 503, "no log" answers 404) instead of both rendering as a claim about a configuration setting.
Telegram output
- Tool results are summarised rather than quoted: a search reports how many matches came back, a truncated result reports the tool's own total. The renderer no longer states a count it computed from its own truncated sample.
- Repeated successful tool calls collapse into one counted entry.
- The tool-activity block is capped to a third of the live frame, keeps the
newest lines, and marks what it dropped with a
+N earlierindicator. Every line that fits is kept. - Every clamped answer marks its cut with an ellipsis, including the opening frame, which previously appeared to begin mid-sentence.
- An error response from the log download no longer ships as a
.logfile.
Readiness and shutdown
- The disk probe reports per-filesystem results and checks the root filesystem as well as the work directory, so a full root is no longer invisible.
- A path shared by several disk targets keeps the strictest requirement any target places on it, so an optional target can no longer mask a required one.
- The shutdown-watchdog test no longer races the Go scheduler, which was intermittently failing the quality gate and parking work unrelated to it.
Other
- Messages persist the channel-native sender id.
- go-telegram/bot bumped to v1.27.0, fixing a streaming bug where the multipart
body could not be replayed and
sendMessage/editMessageTextfailed until the connection dropped.
[1.26.0] - 2026-09-17
- fix(webui): publish per-identity forges so multi-identity task rows keep their links
- feat(config): add [scheduling] config block
- feat(configuration): translate [scheduling] into scheduling.EngineConfig
- feat(cronevents): adapt the events bus to scheduling.Sink
- fix(webui): count every chat front-end in the published projection
- fix(scheduling): respect configuration ownership
- feat(ratelimit): wire per-identity inbound rate limiting into the gateway
- fix(tools): judge the rm guard by base name so absolute paths cannot bypass it
- fix(store): keep an empty binding secret empty so a partial edit cannot erase it
- fix(archied): stop advertising review_pr when no reviewer can run
- fix(daemon): carry kind/label routing bindings to the process that routes
- feat(configuration): add archied setup, the caller the first-run flow was built for
- fix(configuration): stop letting a secret value be set from the command line
- feat(configuration): let setup reference a secret without carrying its value
- fix(archied): disable a provider whose credential is missing instead of refusing to boot
- refactor(install): install.sh delegates config generation, and tests stop reading hand-written configs
- fix(daemon): guard binding dispatch loop against nil dispatcher/creator
- fix(webui): enforce CSRF/origin gate on binding and mapping mutations
- fix(secret): recover a panicking secret engine instead of crashing
- fix(config): deep-copy Image.Hosted/Image.Local in Config.Clone
- fix(email): stop dropping the first paragraph of plain-text bodies
- fix(logging): keep the live descriptor open across a failed rotation
- feat(configuration): report config keys no decode target consumed
- feat(archied): warn at boot when the config file has unrecognised keys
- feat(archied): wire the scheduling ticker engine at boot
- fix(agentworker): wire configured tool-result policy into worker's LoopRunner
- refactor(staterpc): delete the superseded single-token auth interceptor
- refactor(memory): delete the unreachable fence scrubber
- fix(channels): delete unreachable Telegram webhook delivery mode
- fix(gateway): delete unwired Router.Dangerous approval path
- refactor(gateway): delete orphaned SessionState goal/steer/queue machinery
- refactor(archied): delete the daemon's unused webui.Server and UI adapters
- refactor(tools): delete unreachable dispatch/gating engine
- feat(channels): wire chat.webhook route config through to the gateway
- refactor(tools): delete unreachable AST-based tool discovery
- refactor(tools): delete unused plugin-tool registration API from builtin.Registry
- feat(channels): wire ConfigSchema/ValidateConfig into a startup validation loop
- refactor(state): delete the dead IncrementRetryCount RPC (breaking change, accepted)
- fix(curator): populate PassInput.Since with the prior pass time
- fix(memory): install the threat scanner the composition root never wired
- fix(staterpc): preserve marshal_error marker in eventDataJSON
- fix(gate): validate finding level and share the blocking threshold
- fix(scheduling): record a completed run so a recurring job stops re-firing every tick
- fix(releaseupdate): don't treat empty available as deferred
- fix(tools/provider/mcp): preserve MCP binary tool results for delivery
- refactor(gateway): share sensitive-key markers with webhookguard
- refactor(daemon): remove unreachable SQLite poll/drain path
- refactor(tools/mcp): delete unreachable MCP adaptation path
- fix(taskactions): resolve max_retries from the task's owning identity
- fix(daemon): forward configured MCP servers to task agents
- refactor(skill): delete dead Discover/DiscoverPlugins/Skill path
- refactor(skillbuild): delete dead AugmentRegistry path
- fix(builtin): bound read/edit filesystem work by DefaultToolTimeout
- refactor(binding): delete dead Matcher.Overlaps method
- refactor(forge): delete dead CreateIssue surface
- fix(mcp): cancel in-flight SSE reconnect on Stop
- fix(container): enforce MaxUptime lifetime cap with a timer
- refactor(store): use taskstate status constants in SQL
- fix(ui): unbreak binding approve and the two deletes, and centralise the client
- fix(tools): stop testing on the text of other files, and ungate contractaudit
- refactor(forge): delete the unwired RepliesAfter surface
- refactor(config): delete the two decoded-but-unread policy fields
This release is dominated by dead-code removal and the completion of the
scheduling capability. Roughly half the commits delete surfaces that no longer
had a caller: the unreachable SQLite poll/drain path, the unwired MCP
adaptation path, the AST-based tool discovery, the superseded single-token auth
interceptor, and several dead forge and binding methods. The
IncrementRetryCount State Store RPC is gone with them -- a breaking change to
the wire contract, accepted because nothing called it.
The cron/scheduling epic lands its remaining slices: the [scheduling] config
block, its translation into the engine's own settings, the events-bus sink
adapter, and the boot wiring that starts the ticker. A recurring job now records
a completed run, so it stops re-firing on every tick.
Operator-facing work: archied setup gives the first-run flow the caller it was
built for, configuration no longer accepts a secret value from the command line,
and a provider whose credential is missing is disabled rather than refusing to
boot. A secret engine that panics is now recovered instead of crashing the
daemon.
[1.25.0] - 2026-09-15
- fix(update): resolve releases without a checkout, and fail loudly when it cannot
- fix(update): back up the task store the daemon actually uses, or refuse
- fix(update): build each component from its own approved release tag
Components are independently versioned, so the newest archied/v* and
archie/v* tags usually sit on different commits. The installer required both
tags to resolve to one commit whenever both components changed, which refused
every update once the two newest releases diverged -- a deployment stuck on
1.22.0 could not take any update. Each component now builds from its own
approved tag.
[1.24.0] - 2026-09-12
Makes 1.23.0 deployable. That release split archied into separate processes,
but the self-updater still installed only the archied binary and restarted
only archied.service, so updating to it left the State Store, Gateway and UI
on the previous release. archied cannot boot without a reachable State Store,
so the update failed its health check and rolled back. Update to this release
rather than to 1.23.0.
If your host does not yet run the four units in
deployments/systemd-user-service.md, create them first: the updater now
refuses, naming what is missing, instead of installing binaries onto a host it
would leave broken.
- fix(update): install and cycle every process in the release
- fix(telegram): stop the relay corrupting identifiers and code
Also in this release, though outside the version-stamped package closure:
- ci(deploy): produce a complete distribution zip on every release. No zip was
built for v1.23.0 at all, and the zip omitted
archie-ui.
[1.23.0] - 2026-09-12
Archied is no longer a single process. The State Store (archie-state-store)
now owns archie.db outright, the Gateway serves its contract over
authenticated gRPC, and the dashboard ships as a standalone UI Service
(archie-ui). Upgrading is not a binary swap: these processes must be running
and addressable before archied will start clean. See deployments/ for the
supported profiles.
Service decomposition
- feat(contracts): add gateway service seams and protobuf toolchain
- feat(gateway): extract ChatContract service
- feat(gateway): delete in-process ChatContract path, make remote the only mode
- feat(gateway): authenticate the gRPC contract and allow off-loopback listeners
- fix(gateway): drop services.gateway.mode, remote is the only transport
- feat(state-store): generalize storerpc as the State Store gRPC contract (#770)
- feat(state-store): stand up archie-state-store binary and own archie.db (#771)
- refactor(state-store): migrate capture/mapping/binding consumers to State Store (#772)
- refactor(state-store): migrate workflow/intake/task lifecycle consumers to State Store (#773)
- refactor(state-store): cut daemon/Gateway over to the remote State Store (#774)
- fix(state-store): task-scope container credentials, stream large capture lists, validate Update
- feat(archieui): standalone UI process over remote Gateway and State Store
- refactor(archied): cut the dashboard over to the archie-ui process
- refactor(ui): sever the archie-ui binary from every daemon runtime package
- refactor(archied): move the update relay to the process that owns updates
- refactor(archied): move webhook capture intake to the process that owns work intake
- feat(archieui): mount the capture intake receiver over the State Store
- refactor(workflow): relocate Task/Status/Source and define workflow.Store (#754)
- refactor(store): move the producer-owned contracts to internal/contracts/store/v1
- feat(messaging): migrate gateway session persistence onto domain types
- feat(gateway): flip channel boundary onto messaging.Message
Pull request review
- feat(gateway): operator-triggered PR review with selectable reviewer model
- feat(forge): pull request review read surface (ListReviews/ListReviewComments/ReplyToReview)
- feat(workflow): surface adversarial review findings on the PR body
- feat(workflow): review calibration, checked properties and style nits out of contract
- fix(archied): share one PR reviewer across channels; harden forge and worktree
- fix(workflow): order findings before footer; test the report stash
Dashboard
- feat(webui): render configuration from a published snapshot
- feat(staterpc): carry the dashboard's configuration snapshot
- refactor(webui): take configuration off the dashboard server
- feat(webui): report which dashboard sections a process can serve
- feat(webui): deliver live dashboard events by polling the State Store cursor
- feat(webui): route dashboard task actions through the Gateway contract
- feat(gateway): give the dashboard operator its own task-action contract
- feat(webui): add /health and /health/detailed readiness endpoints (#717)
- fix(webui): distinguish refused vs broken task actions, handle 401 (#753)
Daemon, gateway and workflow
- feat(archied): add shutdown watchdog that force-exits a hung graceful shutdown (#716)
- feat(archied): serve the daemon's own health endpoint
- feat(drain): honour external drain marker with instantiation epoch (#718)
- feat(gateway): wire progressive tool disclosure (bridge tools)
- feat(gateway): render workspace, managed repos and operator into chat env
- feat(gateway): add session_transcript tool for full JSON export
- feat(gateway): consolidate task actions and repair daemon default path
- feat(eda): thread Playbook ID/Version and DispatchInput task identity
- feat(worktree): Resume re-syncs a worktree onto its branch tip
- fix(staterpc): attach the bearer token to State Store streams
- fix(archieui): resolve service tokens from env, without requiring a config file
- fix(archied): initialize conversation persistence on daemon startup
- fix(archied): preserve channel turn ledger with remote gateway
- fix(gateway): share NATS and session ownership
- fix(update): probe the dashboard's real address after a self-update
- fix(store): migrate binding owner and repo columns
[1.22.0] - 2026-09-03
- feat(telegram): support secret refs for bot tokens (#705)
- fix(update): make daemon rollback schema-safe (#704)
- fix(archied): preserve ephemeral gateway ports (#707)
- fix(webui): bound log limit parsing and add endpoint regression coverage (#706)
- fix(deps): upgrade x/crypto to v0.56.0 for security fixes (#708)
[1.21.0] - 2026-09-03
- feat(workflow): make kind-to-workflow routing YAML-overridable
- feat(workflow): route arbitrary forge labels to workflows via playbook YAML
- feat(workflow): load kind/label bindings from multiple playbook directories
- fix(extract): regenerate stale Yaegi symbol tables; fix wfextract directive
- feat(eda): Module position first slice -- registry + side-effect-free log kind
- feat(eda): CEL expression environment (compile/validate at load, evaluate at dispatch)
- feat(eda): playbook document + event coordinator, single-action workflow-kind dispatch
- fix(skillcurator): document and scope the intentional nilerr in reviewOne
- fix(binding): add Owner/Repo pin, fixing multi-repo dispatch
- feat(webui): bindings dashboard UI; fix app-wide Preact render regression
- fix(binding): split ValidateForUpdate so edits can preserve the secret
- refactor(ui): extract Pill to ui/src/base/, closing archie-core-mlro
- refactor(eda): extract helpers to satisfy cyclop/gocognit gates
- fix(webui): stretch HEALTH row cards to equal height
- fix(eda): make playbook key collision reporting deterministic
- fix(gateway): split /status (health) and /tasks (work) apart
[1.20.0] - 2026-09-03
- fix: baseline gate repair (go test ./... -count=1)
- fix(logging): propagate bound slog attrs to wrapped handler in FeedHandler
- feat(image): define provider-neutral image capability contract and config
- fix(channels): call MarkRunning once a channel is actually up
- fix(workflow): stop discarding baseline-fix commits as 'no changes'
- fix(webui): wire the Logs page to per-task attempt-log detail
- fix(workflow): net cache hits from token totals, trim gate-retry output
- feat(workflow): dynamic triage instead of defaulting every task to implement
- fix(gateway): stop SOUL from forging the prompt invariant tags
- fix(curator): stop curator actions from serializing a zero timestamp
- feat(webui): hand-author the configuration field descriptor catalog
- feat(webui): attach the config field schema to GET /api/config
- feat(webui): render scalar config sections generically from the schema
- feat(webui): surface allow_concurrent/max_retries/review_enabled on the repositories card
- feat(webui): inline editing for per-repo allow_concurrent/max_retries/review_enabled
- feat(webui): playbook bindings -- matcher + mapping + workflow, HMAC-gated
- fix: resolve golangci-lint funlen, nilerr, and empty branch issues
- feat(store): encrypt playbook binding secrets at rest
- feat(webhook): harden captured payload redaction
- fix(store): satisfy makezero when building cipher payload
- feat(webui): refactor chat module to Preact component and add unit tests
[1.19.10] - 2026-09-02
- feat(curator): session-memory curator reference implementation
[1.19.9] - 2026-09-02
- feat(memory): typed memory engine contract, registry, and registrar
- feat(memory): builtin engine implementing the domain/memory contract
- feat(config): config surface for memory engine selection
- feat(daemon): wire the domain/memory engine registry at boot
- feat(curator): bind curators to memory engines
- fix(config): reject memory.provider instead of silently ignoring it
- feat(curator): skill curator reference implementation
[1.19.8] - 2026-09-02
- fix(update): drop dev-facing wording from the unconfirmed-update message
- fix(telegram): stop adjacent rich blocks from rendering glued together
- fix(telegram): send undelivered-media fallback after finalize as its own message
[1.19.7] - 2026-09-02
- fix(logging): make Tail/Page allocation bound explicit for CodeQL
[1.19.6] - 2026-09-02
- fix(build): stamp gatewayVersion/runtimeVersion in the right package
[1.19.5] - 2026-09-01
- feat(webui): dashboard-aware chat — Archie sees the page, points you there
- fix(telegram): drop retry_after bandaid, bump go-telegram/bot to v1.24.0
- feat(curator): track and expose recent activity per curator
- fix: go.mod bumped to 1.27.0 and stupid tests removed
- refactor: remove unnecessary AI-generated scaffolding
- refactor: go modernisation fixes
- refactor: remove unused Operator param.
- fix(gateway): fix broken system prompt template and scrub real telegram ID from tests
- fix(plugin): add missing CLAUDE.md reference to plugin engine rule
- fix(app): repoint archied wiring contract test at internal/app/archied
- refactor(agentexec): use ai-sdk NewTypedTool for pluginToolSet and scriptToolSet
- fix(config): standardize Provenance to pointer receivers (recvcheck)
- refactor(lint): reduce cyclomatic/cognitive complexity in tools, workflow, webui
- fix(app): repoint capability-host wiring contract test at internal/app/archied
- refactor(lint): extract more helpers to clear nestif/gocyclo findings
- refactor(lint): split AgentStage.Stage into resolveModel/buildRequest/handleResult
- refactor(lint): split TurnRunner.Run and dispatchLocal further
- refactor(lint): extract markdownBlockParser and list/quote helpers in richblocks.go
- refactor(lint): extract Apply/findWalkVisitor helpers; go fix modernization
- fix(lint): noctx fixes and dead-code cleanup
- feat(workflow): adversarial self-review stage before PR open
[1.19.4] - 2026-08-28
- refactor(telegram): render replies as structured rich blocks
[1.19.3] - 2026-08-28
- feat(tasks): serve the lifecycle presentation catalog from /api/task-meta
- feat(dashboard): render task actions, statuses and filters from the server catalog
[1.19.2] - 2026-08-28
- feat(tasks): allow rejecting a task from any non-terminal state
[1.19.1] - 2026-08-27
- fix(telegram): keep Markdown headings/lists as blocks and strip Markdown on plain fallback
[1.19.0] - 2026-08-28
- feat(logging): cursor pagination and filtering for task_logs
- build(deps): upgrade github.com/samcharles93/ai-sdk to v0.1.29
[1.18.0] - 2026-08-25
- Fix concurrent chat-turn claims so SQLite writer contention no longer leaks as
SQLITE_BUSY. - Make task timelines operationally useful with stage outcomes, durations, retry context, and durable per-stage token economics.
- Return canonical repository, issue, and pull-request URLs for each task's owning GitHub or Gitea identity.
- Preserve per-identity commit attribution and real model context limits across daemon-to-worker dispatch.
[1.17.1] - 2026-08-25
- fix(update): remove source checkout dependency
- fix(gateway): make UUIDv7 session references unambiguous
[1.17.0] - 2026-08-25
- fix(webui): add nil receiver guard to Server.trustForwardedHeaders
- fix(chat): restore transcript and update replay
[1.16.0] - 2026-08-25
The agent can hand you a file, and the dashboard works behind a reverse
proxy. Media delivery was fetch-by-URL only, so a file the agent produced
locally — a transcript, a log dump, a screenshot — was passed to Telegram as
a URL it could not fetch: the send did nothing and reported success. The new
send_file tool plus upload-by-reader delivery closes that, and every layer
that cannot deliver now says so rather than staying silent. Dashboard
mutations behind a TLS-terminating proxy were refused as cross-origin;
forwarded headers are now honoured, but only under an explicit opt-in that is
off by default.
- feat(telegram): deliver a local file as an uploaded attachment, choosing
upload-by-reader or fetch-by-URL per attachment;
send_fileresolves paths under the same confinement policy the read tool applies, so sending cannot reach what reading cannot - feat(gateway):
task_actionchat tool for operator task management — abandon, archive, retry, stop, cancel, approve and reject from chat, on the same action path and state machine as the dashboard - fix(webui): validate mutation
Originagainst the effective external scheme and host.X-Forwarded-Proto/-Hostare honoured only whenweb.trust_forwarded_headersis set, which is off by default and safe only behind a proxy that overwrites or strips untrusted forwarded headers - fix(webui): report a local file the dashboard chat cannot upload as
undelivered instead of dropping the event, which had reproduced the exact
silent non-delivery
send_filewas built to end - fix(daemon): detach terminal state transitions and cleanup from a cancelled context, so a stopped task still records its outcome
[1.15.0] - 2026-08-24
Autonomous workflows now have exactly one production execution path. Every
task is handed complete to a scoped archie-agent container over core NATS;
archied no longer runs an agent loop in-process. Embedded NATS is the
default and generates a per-start token, binding to the resolved Docker bridge
so managed workers can reach it; external NATS changes only broker placement.
The legacy [agent] section and containers.enabled still decode for
migration but can no longer select an execution path, and a reloaded
nats.url is logged as requiring a restart. Operators upgrading from a
agent.mode = inprocess configuration get the managed-worker topology
automatically; see deployments/ for the supported profiles.
- feat(nats): expose the embedded broker to managed task containers behind a generated per-start credential
- fix(daemon): park before worktree access when the container pool or task transport is unavailable, rather than silently executing locally
- refactor(agentexec): remove the in-process, subprocess and single-stage execution paths
- fix(worktree): report a successful push as successful even when writing local upstream tracking metadata afterwards fails, so a published branch no longer parks the task as failed
- fix(worktree): honour context cancellation in
CommitAll, so a cancelled task can no longer stage or commit - fix(daemon): remove the worktree when a task completes with no changes, which previously leaked a full clone per no-change task; parked worktrees are still retained for post-mortems
- fix(worktree): derive the branch prefix from the issue title before falling back to labels, matching the documented contract
- fix(worktree): clean untracked files when
.gitis a gitdir file - fix(agentexec): recover agents that end a turn without the required finish call
- feat(scheduling): ticker engine with a configurable interval and pools
- fix(telegram): keep tool-call context when compacting tool output (#609)
- feat(workflow): add the adversarial-review findings contract, where only a confirmed finding blocks and "review found nothing" is distinct from "review did not run" (contract only; no review stage runs yet)
[1.14.0] - 2026-08-23
- fix(telegram): compact tool output and drop turn caps
- fix(update): install only approved components
[1.13.0] - 2026-08-23
- fix(telegram): render tool progress as bounded code blocks
[1.12.0] - 2026-08-23
- fix(webui): give dashboard-initiated updates a phase-2 report
- feat(agent): stamp archie-agent's build and report it back to the daemon
- feat(releaseupdate): enrich components with install type and reference
- feat(webui): surface per-component update status on /api/version and Configuration
[1.11.1] - 2026-08-23
- fix(update): wire Telegram personas and survive callback cancellation
- fix(webui): theme the Workflows 'Start work' form controls
[1.11.0] - 2026-08-23
- feat(forge): make webhook intake observable and deduplicate concurrent delivery
- fix(telegram): preserve code fences across message splits
- fix(gateway): replay completed-duplicate turns with their tool activity
- feat(storage): persist agent project state across sessions
- fix(container): persist npm cache for per-task MCP servers
- fix(worktree): bind publication to dispatch grants, reset retries to a pristine base, and reject diffs without a merge base
- feat(media): deliver generated media through Telegram with link fallback in the dashboard
- feat(tools): generate videos through MiniMax, gated by provider configuration
[1.10.0] - 2026-08-22
- feat(forge): GitHub webhook receiver for immediate issue dispatch
- feat(webhookguard): shared HMAC, rate-limit, and redaction mechanics
- feat(nats): run an embedded server when no external NATS is configured
- feat(capture): store unbound webhook events, retention-bounded
- feat(webui): event inspector -- see captured payloads in the dashboard
- feat(webui): payload field mapping -- bind JSON paths to named fields, preview before saving
- fix(webui): stop the capture live-event notification duplicating payloads into the unbounded events table
- fix(update): verify an update took effect instead of trusting the report (#530)
- fix(daemon): refuse to poll on an empty label-trigger match (#529)
- fix(gateway): stop persona from asserting unverified claims as fact (#528)
- fix(gateway): stop ClaimTurn's insert-race branch losing its own retry read
- fix: restore provider gate and dashboard token accounting
- fix: address simple findings from a repo-wide golangci-lint sweep
[1.9.11] - 2026-08-19
- fix(mcp): persist npx package cache across daemon restarts
[1.9.10] - 2026-08-19
- feat(update):
/updatenow streams live progress while an update installs, reports a clear version summary (previous/installed, daemon/agent) once the build succeeds, and automatically rolls back and reports the outcome if the restarted daemon fails its post-update health check
[1.9.9] - 2026-08-18
- fix(telegram): drain live renderer in concurrency test
[1.9.8] - 2026-08-17
- chore: no user-facing changes
[1.9.7] - 2026-08-17
- chore: no user-facing changes
[1.9.6] - 2026-08-16
- fix(update): stamp install type and bound retry backoff
[1.9.5] - 2026-08-16
- refactor(archied): decompose the daemon bootstrap into phased wiring (no behaviour change; splits the ~950-line run() into focused setup phases so complexity linting passes)
- refactor(archied): extract the config overlay boot helper and thread context through chat gateway setup
- refactor(tools): split the read executor and flatten nested config/tool fallback logic to lower complexity
- fix(tools): avoid predeclared
maxnames - fix(email): use context-aware test networking
- fix(indexing): remove the unused code search test runner
- fix(tests): use contexts and type assertions in overlay and gateway tests
- docs(architecture): document the memory engine family
[1.9.4] - 2026-08-15
- feat(gateway): polish /status command response format
- fix(telegram): preserve URL boundaries in responses
[1.9.3] - 2026-08-15
- fix(webui): scroll selected command into view on keyboard navigation
- fix(logging): resolve error attrs to their message string, not {}
- feat(agentexec,workflow): surface tool calls on the task timeline
[1.9.2] - 2026-08-15
- fix(worktree): force checkout so retry recovers from a dirty worktree
[1.9.1] - 2026-08-15
- fix(agentworker): forward workflow events across the NATS worker boundary
- fix(webui): filter turn_completed noise from the Live Activity SSE feed
- feat(releaseupdate): stamp install type at build time, fail closed on unknown
[1.9.0] - 2026-08-15
- fix(webui): style tool-call failure from a structured field, not a string prefix
- fix(config,webui): make show_tool_calls one cross-channel setting
- fix(telegram): guarantee no message is stranded on gateway stop/restart
- fix(telegram): use NewRequestWithContext to satisfy noctx
- refactor(webui): decompose handleSSE to cut cognitive complexity
[1.8.0] - 2026-08-15
- feat(archied): add -version so the installed build can be read from a shell
- fix(webui): rebuild the Configuration page layout so values stop colliding
- fix(webui): size Configuration key/value lists to their own content
- fix(webui): keep Configuration rows full width, hug only the label column
- fix(webui): cap and centre the Configuration page instead of its lists
- fix(webui): preserve encoded redirect path characters
- feat(chat): show tool calls inline and stream replies into one message
- fix(chat): harden inline tool call streaming
- refactor(agent): move worker runtime into application layer
- fix(telegram): harden live-reply delivery against /stop and multi-byte content
- fix(telegram): keep tool activity visible and mark failed/empty turns
- fix(config): de-duplicate disabled-forge predicate, export ForgeDisabled
[1.7.0] - 2026-08-10
- fix(build): install ripgrep where grep.go actually needs it on PATH
- fix(webui): Send button no longer passes its click event to sendMessage
- feat(webui): Enter sends, Shift+Enter inserts a newline
- fix(chat): stop showing raw 501 banner for unconfigured update-check
- fix(chat): give chat page a real height budget instead of a floor
- fix(workflow): carry gate output into the baseline park error
- feat(logging): add a per-task log sink
- feat(agentexec): publish agent-side logs to the task's system subject
- feat(daemon): consume agent system logs into per-task sinks and the dashboard feed
- feat(webui,gateway): add /api/tasks/{id}/logs endpoint and task_logs chat tool
- refactor(gateway): persist conversations in sqlite
- refactor(store): replace legacy persistence with SQLite
- fix(webui): tighten configuration KV row spacing
- fix(gateway): harden session recovery paths
- refactor(daemon,webui): make Daemon.Cfg and webui.Server.Cfg reload-safe
- feat(config): SIGHUP-triggered reload with requires-restart warnings
- fix(config): share one config Holder between daemon and dashboard; re-apply model catalog on reload
- fix(config): complete the reloadable allowlist and pin it against ForTask
- feat(config): runtime config overlay store, PATCH /api/config, and the
--no-config-overlayrecovery flag - feat(webui): inline config editing on the Configuration page
- feat(config): surface overlay state, shadowed keys, and per-row reset
- fix(config): close review blocker on the PATCH path; deep-copy snapshots; atomics for shared state
- fix(config): preserve nil slices in Clone to avoid spurious reload warnings
[Unreleased]
[1.6.0] - 2026-08-09
- feat(gateway): tool-approval gate, session tools for chat, and /delete command
- fix(webui): close open-redirect in dashboard token-exchange handler
- feat(gateway): make chat turns durable and idempotent
- feat(webui): lifecycle-safe task controls with atomic retry and archive
- feat(config,channels): format-neutral config resolution and truthful channel state
- feat(logging,workflow): daemon log feed and executable workflow registry
- fix(store,gateway): taskstate constants and mobile jump-to search
- feat(webui,gateway): wire config, channel, log-feed, and work-intake seams
- fix(ui): hide inactive chat command menu
[1.5.0] - 2026-08-08
- feat(curator): engine family contract, registry, and runtime
- feat(curator): trigger accounting — primary-input wake path
- feat(setup): interactive setup steps with a TTY-gated prompter
- feat(webui): rebuild the dashboard and add chat
- feat(daemon): identity-scoped isolation plus gateway session titles and /delete
- feat(logging): optional rotating log file
- feat(configuration): export Validate for the setup flow
- fix(gateway): durable message identity with a SQLite store
- fix(gateway): consistent message ordering and paged search
- fix(gateway): stop /compress destroying history it should have kept
- fix(gateway): do not answer a redelivered message twice
- fix(gateway): claim the session slot before writing it, not after
- fix(gateway): keep /topic off from poisoning the session cache
- fix(gateway): resolve the most recently active session, not an arbitrary one
- fix(gateway): give branch-inherited messages their own identity
- fix(gateway): make both backends agree on newest-first
- fix(store): honour the transition guard in production persistence
- fix(store): move ID counters away from the reserved metadata namespace
- fix(store): surface task timestamps and tag the summary structs
- fix(telegram): pause the update loop when rate limited
- fix(memory): fail startup on an unusable memory directory
- fix(tools): let an optional provider fail without taking the daemon down
- fix(tools): report a cancelled find walk as an error, not empty results
- fix(worktree): name the missing forge credential when a push fails
- fix(tomlwrite): recognise table headers carrying a trailing comment
- fix: close the duplicate-work loop and the dead paths around it
- perf(tools): drop the embedded ripgrep binaries
- refactor: one definition of what approving or declining a task means
- refactor: move workflow into its own domain
[1.4.1] - 2026-07-31
- fix(catalog): match SDK provider fallback
[1.4.0] - 2026-07-31
- fix(secrets): wire provider credentials through engines
[1.3.1] - 2026-07-31
- fix(telegram): omit empty model picker markup
[1.3.0] - 2026-07-31
- fix(memory): replace framework-specific names with generic equivalents
- feat: unify provider and model selection
[1.2.0] - 2026-07-31
- feat: add MCP client capabilities to archied and archie-agent
- fix: handleSummary ignores errors from WorkflowStats/StageStats/TokensByDay
- feat: establish Archie identity and use stable callback tokens
- refactor(telegram): keep one model selector command
- fix: handleSSE silently swallows EventsSince backlog fetch errors
- fix: Transition() ignores
fromstatus guard, allowing lost updates - fix: skillscript.Run ignores context, cannot be canceled/timed out
- feat(telegram): report installed component versions
- feat(telegram): add approved component updates
- fix: LoadDir: no timeout around Yaegi Eval can hang daemon startup
- feat(chat): real system prompt for Archie — identity, tools, env
- feat(indexing): port tau's codesearch workspace index
- fix(indexing): surface index config properly and stop silent degradation
- refactor(worktree): replace git shell-out with go-git v6
- fix(worktreerpc): bound handler contexts and drop the git binary from tests
- refactor(nats): rebuild as infrastructure/eventbus/nats with real boundaries
- refactor: route tasks on a typed kind, not forge labels
- fix(natsrpc): make RegisterAll wait for the server to see subscriptions
- refactor(config): move loading into infrastructure/configuration
- fix(telegram): publish all executable session commands in menu and help
- fix(telegram): publish the restored gateway commands
- refactor: add internal/eventbus as the broker-neutral messaging contract
- fix(chat): operator name from config, session wiring, prompt rules
- feat(tools): lift tau's file and shell tools into the tool registry
- feat: make /stop cancel the running chat turn
- fix: kill the whole process group when a shell command is cancelled
- feat: refuse unrecoverable shell commands
- feat: make /stop reach running agent tasks
- fix(chat): temper prompt brevity rules with warmth
- fix: use errors.Is/As for context errors and cleanup
- fix(grep): distinguish ripgrep exit 1 from a failed search
- fix(archied): disable the forge on a missing credential, not the daemon
- fix: restore /model parsing and close the 6d4e0be autofix audit
- feat: embed config.example.toml and add comment-preserving TOML writer
- fix(docker): install unzip and Node 24 so MCP servers can start
- fix(docker): slim both images and repair unreachable Go tooling
- fix(container): write the task boot brief under .git
- fix(telegram): drop pending updates before long polling
[1.1.0] - 2026-07-27
- feat: add the Telegram gateway with persistent sessions, topic-thread support, streamed rich replies, typing indicators, and Markdown rendering
- feat: add
/spawn,/approve, and/cancelfor chat-driven task control - feat: provide a comprehensive
/helpguide whose command list stays aligned with Telegram's published menu and Archie's executable command surface - feat: add
/providerand provider-filtered/modelselectors that switch the provider and model used by subsequent live chat turns - feat: include the active provider and full model reference in
/status - feat: notify authorized Telegram users once when a newly tagged component release is installed
- feat: add
/restartto reload and relaunch the Telegram gateway without interrupting in-flight daemon tasks - feat: add deny-by-default Telegram sender authorization and native command-menu publication
- feat: connect Desktop Commander through Archie's managed MCP client so chat can use workspace file and process tools
- feat: add multi-identity daemon configuration, persona routing, an email channel, and webhook platform support
- feat: add durable task, session, message, and memory persistence
- feat: add memory synchronization, prefetch, context scrubbing, threat scanning, and built-in memory tools
- feat: add tool guardrails, classification, budgets, availability filtering, and disclosure controls
- feat: add feature-based YAML configuration with
conf.doverlay loading - feat: add typed secret-engine registration and Yaegi extension support
- fix: reject unknown slash commands locally instead of allowing the LLM to fabricate command behavior
- fix: publish Telegram commands to default, private-chat, and group-chat scopes so stale commands inherited with the bot token cannot shadow Archie's menu
- fix: drain the SDK's full event stream so streamed Telegram replies cannot deadlock
- fix: use standard newline-delimited JSON framing for MCP stdio clients
- fix: make MCP stdio shutdown and timeout handling race-safe
- fix: restore legacy
[forge].token_envcompatibility and required container environment passthrough - fix: make the agent-container Docker network explicitly configurable
[1.0.0] - 2026-07-22
- feat: ship the resident forge-polling daemon, deterministic workflow engine, isolated worktrees, quality gates, Gitea integration, Docker/NATS worker split, plugin loading, and skills support
- feat: publish
archiedandarchie-agentcontainer images to the Gitea registry