A pending update report used to be relayed verbatim. It is entirely
self-reported: scripts/archie-update-watchdog decides "passed" from
`curl -sf $HEALTH_URL/healthz`, which only proves that *a* daemon is
answering, then copies the version the installer *intended*
(ARCHIE_UPDATE_INSTALLED_GATEWAY) into the report. If the new binary
never made it into service, the previous daemon answers the probe just
as happily, so the message read "Update complete: daemon 1.9.8 -> 1.9.0.
Health check passed - archie-core is healthy on the new version" about a
version that was not running. That shipped: /update reported success and
the next task hit the exact bug the update was meant to fix (#468).
Report.Verify now checks the claims against the versions components
report for *themselves* -- archied's compiled-in gatewayVersion, which no
installer can talk it out of -- and partitions every claimed component
into Confirmed, Drift or Unverified.
What changed in MEANING:
- A passed health check used to mean "the new version is running". It
now means only "something answered"; the version claim is a separate,
separately-checked assertion.
- "Update complete ... healthy on the new version" used to be the only
success wording. There are now three outcomes, because collapsing them
is how the false success happened: confirmed, unconfirmed ("I could
not confirm which version is actually running"), and contradicted
("Update did NOT take effect").
- An unverifiable component is NOT a failing one. Absent, empty,
"unknown" (the watchdog's own default) and "dev" (an unstamped build)
all mean "we could not check", never "it is wrong" -- treating them as
drift would turn healthy updates whose bookkeeping went missing into
reported failures.
Verification covers the daemon only, deliberately. archied's compiled-in
runtimeVersion records what its own pipeline stamped, not what the agent
container runs, so vouching for the agent would be the same kind of
guess this change exists to remove. The agent is therefore Unverified on
every successful update, and the success message names it as unchecked
rather than letting the confirmed daemon silently vouch for it.
Drift detail is attached per component: one component can sit on the
version the update meant to replace while another is on some third
version, and a single explanatory sentence would state a falsehood about
one of them.
Regression tests were confirmed to fail against the pre-fix code,
reproducing the exact false-success string, and each new guard was
mutation-tested. The drift-ordering test's component count is
load-bearing and documented: Go rotates small maps rather than permuting
them, so 8 components let an unsorted implementation pass ~12% of the
time while 9 fails it outright.
Refs #522, #468
GitHub's issues-list API treats an empty label filter as no filter at
all, so dispatch.trigger = "label"/"either" with cfg.Label unset
silently returned every open issue in the repo. A live incident
(GH#445) queued 124 unrelated issues in one poll cycle after
[dispatch.labels] -- a different field mapping task states to their
own labels -- was configured while the actual trigger-match label was
left blank.
pollIssuesWithConfig's "label" branch now refuses to poll and logs an
error when cfg.Label is empty; pollEitherWithConfig skips only the
label-side poll in that case, still returning assignee matches.
validateDispatch rejects the same misconfiguration at config-load
time. A per-identity gap remains: identities have no Label field of
their own and share the root's, so an identity overriding
dispatch.trigger to "label" isn't caught by validation -- only by
the runtime guard, which is why the runtime guard matters even with
validation in place.
Fixes archie-core-1786637501009-424-5b8332dd / GH#445.
The chat agent labelled unchecked figures "Verified details" and
fabricated an entire "Hermes gateway" subsystem with invented file
paths, stating both as confirmed fact until directly challenged.
Reword core rule 1 to separate static tool-capability claims (fixed
by the tool list) from dynamic file/memory/action/result claims,
which now must be verified this turn rather than merely "in this
conversation" -- the looser wording let a stale claim be treated as
still verified. Add rules 8 and 9: never label something
verified/confirmed without a this-turn tool check, and never state a
specific system detail (path, subsystem, count, config value) with
confidence unless just retrieved.
Fixes archie-core-3jag / GH#524.
The /update command went silent between "starting" and one final message,
gave no version summary, and required manual intervention on failure.
CommandInstaller now streams the installer script's stdout live instead of
buffering it, and parses a structured result (previous/installed versions,
per-component status) from a sentinel line. Because the reference install
script hands the daemon restart off to a detached systemd unit specifically
so restarting archied.service doesn't kill the process that requested the
update, the outcome is reported in two phases: the synchronous build/install
result comes back through Install() itself, and the post-restart health
check/rollback outcome is written by a new watchdog script and relayed by
whichever archied process boots next, via a small JSON handoff file cleared
once relayed.
Also fixes three defects an adversarial review pass found in the first
draft: the caller's request/handler context was passed straight into the
install child process, so a concurrent /restart or a browser disconnect
could SIGKILL it mid binary-copy; Telegram and the web dashboard each
tracked "install in progress" with their own private flag despite sharing
one Service, allowing two installs to race the same .prev backups; and the
watchdog reported rolled_back:true even when no backup existed to restore,
which would have told an operator to stand down while the daemon was still
broken.
internal/webui gains an unauthenticated /healthz the watchdog polls after
restarting archied.
Agent containers run as root (no USER in the Dockerfile, no
userns-remap), so a commit the agent writes to the bind-mounted
worktree lands owned by UID 0 on the host. The daemon then reads
those same loose objects to push, running as its own non-root host
user, and fails with "openat objects/../..: permission denied".
containerEnv() now passes the daemon's own UID/GID as
WORKTREE_UID/WORKTREE_GID. hybridTrees.Push -- the single point
where a commit's objects are handed back for the actual push --
chowns the worktree to that owner first, via os.Lchown so a
legitimately-tracked-but-dangling symlink doesn't abort a push that
would otherwise have succeeded.
Closes archie-core-1786749700606-6-bf6aa32d.
Release() always slept the configured GracePeriod and always used
Docker's default ~10s graceful SIGTERM-then-SIGKILL timeout, even
when the release was happening because /stop had just cancelled the
task's context. A cancelled task's container -- and the in-flight
generation/tool calls inside it -- could sit doing nothing for up to
GracePeriod+~10s before actually dying, so /stop reported success
while work visibly continued for tens of seconds to a couple minutes.
releaseDecision(ctx, grace) now skips the grace sleep and requests an
immediate SIGKILL (Timeout: 0) when ctx is already cancelled; a live
ctx (normal completion) keeps the original behaviour byte for byte.
Closes archie-core-1786749700580-5-e3878224.
CLI-based, env-configured, cache-less engines resolving SecretRef values
through the existing Yaegi plugin surface. Fresh resolve per call, so
values updated in the backend are picked up without daemon restart —
closing the gap that motivated the remote secret-input epic (#504),
whose value-entry web form duplicated the Bitwarden app for bws.
- examples/secret-engines/: sops (SOPS_FILE), vault (VAULT_ADDR/TOKEN,
key "path/field" or VAULT_SECRET_PATH prefix), age (AGE_FILE/
AGE_IDENTITY, key=value lines). //go:build ignore keeps them out of
the Go toolchain; yaegi treats the tag as a comment and loads them.
- enginehost: host-side Run helper registered as a Yaegi symbol. yaegi's
interpreted os/exec cannot pass env to child processes, so plugins
shell out through host code; bounded by a 30s timeout like the bws
engine.
- secret.LoadDir seeds the interpreter with the host environment: yaegi's
interpreted os package keeps a private env otherwise, and plugin
os.Getenv would silently see nothing.
- cmd/archied wires enginehost + unrestricted symbols into the engine
plugin loader, matching skillscript's interpretation posture.
- tests load the real shipped plugin files with fake CLIs on PATH;
added CONTEXT.md glossary for the remote secret-input vocabulary.
Keep bare GitHub issue and pull-request links separate from following prose across one-shot sends, live edits, and long-message splitting. Preserve Markdown delimiters and newline boundaries while retaining plain-text fallbacks.
Format /status response with clear heading, visual task status icons, canonical lifecycle ordering, and separated runtime details with graceful fallbacks.
EvictStale sweeps the per-key hit map and removes entries whose hits
have all aged out of the sliding window, so a long-lived limiter
driven on a ticker doesn't accumulate one map entry per identity for
the life of the process. The package doc now states the map is
unbounded until something actually calls EvictStale, rather than
claiming it's bounded by a mechanism nothing yet wires up.
archie-core-467 wired tool-call visibility into the interactive chat
gateway, but a workflow task's own transcript never showed what its
tools actually did -- only stage/agent-finish events existed, so
there was no way to confirm an agent's summary matched what it really
ran.
Adds a ToolCallReporter, invoked once per completed call from
toolExecute (the single choke point every tool call passes through
before reaching a model), reported via a new ToolSetOptions.OnToolCall
hook. Runner.Run grows a report parameter to carry it in: InProcessRunner
wires it through, since it executes in the same process as its tools;
SubprocessRunner and NATSRunner accept and ignore it, since their tool
calls happen in a different process with nothing here to observe them.
The workflow engine turns each report into a tool_call event on the
task's own Bus (agent.go, implement.go), riding the same NATS
event-forwarding archie-core-518 already built to get container
events onto the daemon's timeline.
Scoped to the central tool registry (read/write/edit/grep/shell/...)
built via BuildToolSet; capture tools (structured finish output) and
run_go_script build their ToolSet entries directly and are not yet
covered.
On a stage-handle failure, the main loop logged the error, called
Nak (which requests redelivery), and looped straight back into
FetchStage with no delay -- unlike the fetch-error branch three lines
above, which already backs off for a second. A deterministic failure
(a broken gate, a malformed request) therefore refetched and failed
again immediately, flooding the log with one "handle failed" line per
loop iteration instead of per genuine retry -- hundreds of identical
lines within the same second in production.
Adds the same time.Second backoff after Nak.
SystemLogHandler and FeedHandler both json.Marshal Entry.Fields
directly rather than through slog's own JSON handler, so a logged
`error` value went straight into a map[string]any as the raw
interface. encoding/json has no Error()-aware special case, and the
concrete type behind almost every Go error (errors.errorString,
fmt.wrapError, ...) holds only unexported fields, so it marshalled to
`{}` and silently discarded the message everywhere it was logged --
including on every container/system-log line shipped to the daemon
and shown in the webui, which is where this was actually noticed: a
flood of `"err":{}}` with no way to see what had actually failed.
attrValue now resolves an error attr to its .Error() string before it
reaches the map. It also guards against the classic Go footgun where
a non-nil error interface wraps a nil concrete pointer (var e *T; var
err error = e) -- a naive err != nil check would call Error() on a
nil receiver and crash the one thing that must never crash: logging
itself.
refresh checked out the task branch without Force before hard-resetting
to base. A prior attempt's leftover uncommitted changes -- a killed
container, or a stage that edits files before ever reaching a commit --
made go-git's default MergeReset checkout refuse to reconcile them, even
with HEAD already on the correct branch and the target commit unmoved.
The fallback then made it worse: any checkout failure was assumed to
mean the branch didn't exist yet, so it retried with Create: true, which
failed with a much more confusing error since the branch did exist --
"a branch named ... already exists" -- masking the real cause and
parking the retry with the same unhelpful message every time.
Reproduced directly against a live deployment's actual broken worktree
(modified tracked file, untracked file, branch already checked out) and
confirmed retry then succeeds and correctly restores to base.
Force is safe unconditionally here: refresh hard-resets to base a few
lines later regardless of what checkout leaves behind.
Closes#523
/update could report success at the correct target version while the
archie-agent container running it never actually changed: the update
mechanism trusted the configured installer's exit code and reported
whatever version number the release claimed, with nothing verifying
the two agreed.
The first piece of the fix: a new internal/installtype package
reports how a binary was built for distribution (Type(), defaulting
to "unknown"), stamped once per release artifact via
"-ldflags -X .../installtype.buildType=<value>" rather than guessed
at runtime by probing the environment -- restic and Tailscale both
treat "which install shape is this" as a build-time decision for the
same reason: a probed signal is guessable, a compiled-in one is not.
releaseupdate.Service.Install now refuses to run the configured
Installer at all when InstallType is unset or installtype.Unknown, so
an unstamped or misconfigured build can no longer silently attempt an
update it has no way to know is even the right kind for how this
instance was deployed. Wired through every real archied build path so
the gate doesn't just exist in principle: Taskfile.yml's build task
(binary), Dockerfile.archied (container), and install.sh (binary) all
now stamp it; cmd/archied/telegram_setup.go wires
Service.InstallType from installtype.Type().
Configuration-page display of install type/version and post-update
verification against the actually-running reference (not just the
installer's reported version) remain open -- this lands the
correctness gate, not the full observability surface.
Refs #522
turn_completed fires on every chat turn so input-driven curators can
wake on it, but it carries no task correlation (TaskID always 0,
Detail a raw session ID) -- on the dashboard's "Live activity" panel
this drowned real task-lifecycle events: 27 of the last 30 rows in a
live sample were content-free turn_completed noise.
Excluded at the SSE boundary (sseVisible), not at publish, so
curators reading the Bus directly are unaffected. since still
advances past a filtered event exactly as it would for a sent one --
EventsSince only ever returns events with id > since, so leaving
since behind a filtered event would make the next catchUp refetch it
forever.
Closes#521
A workflow run through the container/NATS path never wired
workflow.TaskContext.Bus, so every tc.Emit call (stage_start,
stage_finish, outcome, parked) was a silent no-op. archied's own
in-process path sets Bus and always worked; anything dispatched to an
archie-agent worker did not, which meant the dashboard's per-task
timeline showed "No events yet" for every task regardless of status
or stage -- confirmed live across a Working task and two differently
parked tasks.
Bridges the gap the same way system logs already cross this boundary:
ForwardTaskEvents relays a per-task in-process *events.Bus to a new
NATS subject (SubjectForEvents), and a new daemon-side subscriber
(subscribeAgentEvents) republishes decoded events on the daemon's own
Bus, landing them in the same persist-and-broadcast pipeline an
in-process workflow already used.
Closes#518
handleSSE mixed connection registration, backlog catch-up, and live
broadcast relay into one function with three nested closures --
gocognit flagged it at 52 against a limit of 30. Extracted each
concern into its own top-level piece: sseSince (resume-point
resolution), registerSSEConn (subscribe/unsubscribe), and a new
sseStream type carrying the per-connection state (send, catchUp,
sendPage, drain) that the closures used to capture implicitly.
Behavior is unchanged -- catchUp/sendPage's page-boundary and
target-reached logic is the same control flow the original closure
had, just named and independently testable. handleSSE itself is now
a straight-line sequence of calls.
Added direct unit tests against sendPage/catchUp for the boundary
cases a full HTTP round-trip test doesn't exercise: a target reached
mid-page vs. at a page boundary, a target of zero (drain everything),
a write failure stopping the page immediately, and backlog spanning
more than one EventsSince fetch.
A daemon shutdown or /restart cancels turnCtx without any guarantee
that the goroutine owning a liveReply ever runs again to notice and
clean up after itself -- a full process exit gives it no chance at
all. Gateway now tracks every liveReply mid-turn in a registry and
Stop drains it, marking each as interrupted so nothing is left ending
in the "typing" cursor forever.
finalize and abandon now run their bodies inside a per-reply
sync.Once (the terminal field): a turn finishing naturally and a
concurrent Stop draining the same reply were reaching it from two
goroutines at once, both sending their own edit for the same message
with the outcome decided nondeterministically by which HTTP response
Telegram processed last -- including a complete real answer getting
silently overwritten by the restart marker. Only one side's body now
ever runs; the other blocks on the winner and then does nothing.
The drain itself is bounded (liveDrainTimeout, default 5s): a single
stalled Telegram API call must not hang gateway, and therefore
daemon, shutdown indefinitely. Entries still in flight past the
deadline keep running in the background rather than blocking Stop.
A liveReply created in the narrow window between Stop claiming the
registry and its own registration call is abandoned immediately
instead of being registered into (and lost from) a registry nothing
will drain again.
Regression tests cover the mutual-exclusion race directly (20
iterations of finalize and abandonRestarted racing on the same
reply, asserting exactly one outcome ever wins), the bounded drain
against a deliberately hung API call, and the post-stop registration
gap.
show_tool_calls gated Telegram's tool-activity narration only; the
web dashboard's SSE chat stream narrated every completed tool call
unconditionally with no way to turn it off, leaking tool names and
parameter shapes to anyone with dashboard access even when an
operator had already opted out for Telegram.
Moved ShowToolCalls from TelegramConfig to ChatConfig so it is one
setting for every channel, and wired the webui stream to it:
chatStreamSink now carries a showToolCalls flag (snapshotted once per
turn, mirroring how Telegram's liveReply snapshots it) and ToolCall
no-ops when it's off.
BREAKING: an existing [chat.telegram] show_tool_calls = true in
config.toml now parses as an unrecognized key under [chat.telegram]
and is silently ignored (TOML decoding is non-strict) rather than
erroring -- move it to the top-level [chat] table.
The comment claimed a mount was checked before touching the
operator's global gitconfig; the code only ever checked os.Stat
(existence + directory-ness). A directory that merely exists at the
mount path -- a stale leftover, never actually bind-mounted this run
-- already satisfied the guard and got added to safe.directory,
exactly what the comment said it prevented. A real mount check would
need platform-specific mountinfo parsing for an inert failure mode
(the entry does nothing unless a git repo later exists at that exact
path), so this corrects the contract to match the code and pins it
with a test instead of changing behavior.
The finished reply is delivered by editing the message created at the
first streamed frame, and editMessageText does not change a
message's date, so a long tool-using turn's answer keeps its
turn-start timestamp and can sort before conversation that produced
it. Delete-and-resend would fix the ordering but reintroduce the
flicker/reorder streaming exists to avoid, so this documents the
accepted trade-off rather than changing behavior.
chatStreamEvent gained a Failed bool, set from ToolCallEvent.Err !=
"" -- the field the rest of the codebase already treats as
authoritative -- rather than the browser sniffing summary.startsWith
"failed:" on rendered text. A successful tool whose own output starts
with that literal string (a grep hit, a log line read back) was being
styled as a failed step.
cmd/archied carried a copy of the none/off/disabled predicate because the
configuration package's version was unexported. Two copies of a validation
predicate drift: a fourth disabled alias added in one place would silently
leave the other accepting it.
Export ForgeDisabled in internal/infrastructure/configuration as the single
definition, referenced by both validateIdentities and resolveForge, and
delete the duplicated copy. TestForgeDisabled pins the exact alias set
(including case-variant and empty-string guards); the resolveForge
degradation test now asserts on the warn log so the silent disabled path is
distinguishable from the token-unavailable fallback (verified to fail under
a re-inlined 'none'-only copy).
Three more P2s from the same code-review pass: the live frame's
single clamped buffer mixed tool lines and answer text chronologically,
so early tool lines silently scrolled out of the frame once the answer
grew past liveBodyMaxRunes and only reappeared at finalize (which
already led with tools separately). Tool lines are now tracked
exclusively in toolLines; a new framedText helper keeps the tool block
whole and clamps only the answer tail, used consistently by the live
frame, an abandoned frame, and the finished message.
A tool-only turn with no reply text used to finalize as a bare,
immutable tool line indistinguishable from a real (if terse) answer;
finalText now appends an explicit "(no response)" marker in that case.
A provider error used to funnel into the same clean-partial path as a
deliberate /stop, so a failed turn was indistinguishable from a
successful short one. abandonFailed appends a "turn failed" marker
before dropping the cursor; /stop still goes through the unmarked
abandon path.
Regression tests cover all three paths plus the degenerate case where
a tool block alone exceeds the frame budget, verified red-then-green
against the pre-fix code.
Three P1s from code review of PR #476: finalize sent the authoritative
reply through the same cancellable turnCtx a /stop kills, so a stop
landing mid-finalize stranded the live message with its cursor
forever; abandon had the identical hazard, saved only by its one
caller wrapping the context correctly. Both now strip cancellation
internally (context.WithoutCancel) rather than trusting callers to
remember.
Tool activity lines were composed unescaped into the same Markdown
body as the model's own reply, so a stray '*', '_', or backtick in
tool output/parameters could unbalance the whole message and take the
reply's formatting down with it. ToolCall now runs name, parameters,
and summary through bot.EscapeMarkdown before joining the line.
splitLongMessage sliced oversized lines on byte offsets, cutting
mid-rune for any multi-byte UTF-8 line crossing the 4000-byte bound
(CJK, emoji) and producing invalid UTF-8 parts. It now counts and
slices on runes throughout.
Regression tests cover all four paths, verified red-then-green
against the pre-fix code.
cmd/archie-agent used to own worker construction, NATS lifecycle, task execution, MCP setup, and workflow wiring. It now owns process input only and invokes internal/app/agentworker, while transport-specific NATS, boot-brief, and git-safety mechanics live under infrastructure.
ai-sdk v0.1.21 wraps the cancellation cause in aborted-stream errors
("%w: %w") instead of formatting it, so errors.Is(err, context.Canceled)
is reliable again. The Telegram channel was compensating with a comment
explaining why errors.Is could not be trusted.
The code was already correct; only the comment was obsolete. It now
states the real reason for context.WithoutCancel(turnCtx): turnCtx is
dead on both stop and fault paths, so the cursor-drop edit needs a live
context. The runtime effect of the bump is that a /stop landing inside
the step loop is now logged as "chat turn stopped" rather than
"route failed".
Refs ai-sdk-0pn
Telegram streaming callbacks now enqueue cancellable rendering rather than performing network I/O while draining the authoritative model stream. Final edit failures fall back to normal split delivery.
Completed tool events now correlate inputs by call ID and expose only bounded, conservative parameter shapes. Tool-call visibility is snapshotted per reply so reloads cannot change an in-flight turn.
Chat returned only the final answer, so a wrong tool choice, a
hallucinated action, or a footgun like `systemctl stop` against
`restart` was invisible from the conversation. Only text deltas were
threaded through the turn, and tool calls executed inside Generate with
nowhere to report.
Replace the delta callback with gateway.TurnStream, which carries text
fragments and completed tool calls on one ordered path from the model
loop to the channel. The sequence is the point: a tool line that
overtakes the sentence it interrupted describes a turn that never
happened. Text-only callers use gateway.DeltaFunc and are unaffected.
A tool is reported on its result, never on its call -- the call part
carries no outcome yet, so reporting both would list every tool twice,
once with nothing to say.
Telegram renders it behind [chat.telegram] show_tool_calls, off by
default: this is debugging visibility, and it narrates every internal
step to provide it. The dashboard always shows it, being the operator's
own observability surface; its tool lines are live-only and are not
persisted with the conversation.
The Telegram live reply changes mechanism, and with it what a frame
MEANS. It used to stream through sendMessageDraft: drafts are ephemeral
30-second previews, so a tool call longer than that outlived the
preview, and successive frames were observed concatenated with their
cursors intact. It now sends one real message and edits it. An update
used to be "the next preview of the answer"; it is now "the whole
answer as it currently stands, replacing what is there". The cursor is
appended at render time and never stored, so exactly one exists and
always at the end, and a frame too long for one message is cut to its
tail rather than rejected outright.
A stopped or failed turn now edits the cursor away instead of leaving
it blinking on a message nothing will finish. That cleanup runs on a
live context because turnCtx is already dead, and it runs on both
failure branches: ai-sdk raises an aborted stream as ErrAborted with
ctx.Err() formatted in rather than wrapped, so errors.Is cannot be
trusted to tell a /stop from a fault.
Closes#467