WIRE: the SSE transport's Stop contract already says it closes the
connection and waits for background readers, but the reconnect loop issued
its HTTP GET with the Start context, which Stop could not cancel, so Stop
blocked on readerWg.Wait() against an unreachable endpoint. Wire a
per-lifecycle cancel (mirroring the stdio transport's stopCh plus guarded
commit) so Stop interrupts the in-flight reconnect request.
WIRE: read and edit each declared a DefaultToolTimeout deadline and then
discarded it, so the documented contract (Tools that exceed this are
cancelled) was a no-op on the live tool-executor path. Wire the deadline
through runWithContext, the same helper write already uses, so a hung
filesystem cannot block the tool executor.
DELETE: mcp.Client.RegisterTools and its handlerFor/writeMultimodalResult
chain, plus the WithMediaDir and WithParallelToolCalls options, had no
production caller -- the live adaptation path is provider/mcp.Provider,
which now owns binary media persistence itself. A repo-wide search finds
no remaining reference outside the composition contract test that asserts
RegisterTools must not appear, so the two paths can no longer diverge.
WIRE: the live provider/mcp adaptation path is the only production MCP
consumer and its handlerFor detected image/audio/resource-blob blocks but
returned an empty MultimodalResult, so telegram_setup's multimodalMediaRefs
(which reads only result.URLs) had nothing to deliver. Decode each media
block into the provider's media root and populate Files, SubdirHint and
URLs with local-path refs the channel uploads.
RegisterPluginTool/SetPluginToolExecutor/UnregisterPluginTools and the
PluginToolDef/PluginToolExecutor types have no production caller.
Plugin/MCP tool registration flows through agentexec.pluginToolSet and
internal/tools.Registry, not this staging registry (tools-builtin-2,
duplicate of tools-core-8).
Registry.Discover, DefaultRegistry and the go/ast extraction helpers
have no production caller; static source discovery has no PRD and no
consumer. Production registration goes through bootstrap.go's
toolReg.Register and the provider Registry's RegisterBatch (tools-core-3).
DispatchSequential, DispatchConcurrent, ClassifyEntry and their
supporting types (dispatch.go) have zero non-test callers; the live
tool path is agentexec/toolset.go's BuildToolSetFrom. Removes the
dead engine and its tests, and updates the PRD status table that
referenced it (tools-core-1).
checkRecursiveDelete compared the raw segment name, `seg.Name != "rm"`, while
every sibling rule in the same table resolves the command through baseName --
checkMkfs, checkBlockDeviceWrite, checkKillAll, checkSystemPower,
checkSelfTermination and checkDaemonLaunch all call baseName(seg.Name), and
baseName's own doc says its purpose is that /sbin/reboot and reboot are judged
the same. The test suite already pinned that convention for /sbin/poweroff.
Because tokenise yields the raw first word, `/bin/rm -rf /` produced
seg.Name == "/bin/rm", the rule declined, no other rule matched, and Hardline
returned nil. Since provider/builtin calls command.Hardline on the shell tool's
command field before every execution, an absolute path defeated the
recursive-root-delete gate entirely. `/bin/rm -rf /etc` (any critical path, not
just the root) was equally reachable; an overlay probe confirmed both, and
confirmed the sibling /sbin/poweroff case was already correct.
Resolve through baseName, matching the rest of the rule table, and add table
cases for absolute-path rm plus an allow case for an ordinary absolute-path
recursive delete so the normalisation cannot over-block.
Found by an adversarial package sweep (temp tracker: tools-command-1,
.local/issue-tracker/).
Compose the base registry with per-turn extras into a turn-local registry and
let ContextPressureGate decide full-vs-bridge disclosure per chat turn. The
three bridge tools (tool_search, tool_describe, tool_call) are now wired into
the chat toolset and only disclosed under context pressure, keeping the LLM's
schema footprint small while preserving full arsenal access on demand.
Per-turn extras are identity-bound per channel and cannot live in the
process-wide registry (agentexec.BuildToolSetFrom invariant), so the bridge is
composed turn-locally rather than registered globally. Thread the model's
context window through TurnModel.Prepare (TurnPrepareContext) so the gate
activates with a real window.
Refs #167
- mcp client_test: assert both initialize handshake calls happened and
notifications/initialized has an empty ID, not their exact order
- telegram model_test: assert the resolved model has an openrouter/
prefix instead of scanning button text for the string; replace the
single-hardcoded-count pagination test with cases covering 0/1/8/9/100
models so pageBounds and nav-row logic are actually exercised
- webui api_chat_test: assert the actionable file path is present in the
streamed error instead of matching generic wording
- tomlwrite.go: split Apply into collectPending + an applier struct
(indexTables/resolveTable/matchActiveKeys/uncommentKeys/
appendMissingKeys/render). gocognit 82 -> under threshold.
- find.go: split findWalkVisitor's closure into findSkipEntry/
findMatchesFilters. gocognit 46 -> under threshold. This clears the
last complexity finding from archie-core-agui's original 31.
- richblocks.go: go fix modernization (strings.SplitSeq).
Behavior-preserving.
Extracts focused helpers out of applyEdits, grepFallback/grepDirFallback,
serveReadContent, handleChatStream, TDD, runFindGoFallback,
makeGrepExecutor and makeShellExecutor to bring them under the
project's cyclop/gocognit/funlen thresholds. Behavior-preserving.
Partial pass on archie-core-agui: richblocks.go (gocognit),
gateway.go dispatchLocal (gocyclo), and a couple of remaining
gocognit/nestif findings are left for a follow-up.
Follow-up to the cmd/archied -> internal/app/archied wiring extraction:
the architecture test that greps main.go for required wiring strings
still pointed at the old cmd/archied path.
Follow-up to 0b2398f, which claimed non-delivery is never silent and that
oversized files fail visibly. Both claims had gaps.
chatStreamSink.Media skipped any attachment with an empty URL, exactly as
liveReply.Media used to. The tool registry is shared between the Telegram
and dashboard chat setups, so send_file is offered on the dashboard too:
the model announced a sent file, the sink dropped the event, and nothing
reached the transcript. The dashboard cannot upload at all, so the fix is
to REPORT the file as undeliverable and name where it is on the host,
rather than to stay quiet. The skip was right while media meant a hosted
URL; it stopped being right when a host path could arrive here.
send_file applied one 50MB ceiling to every type while the sender enforces
10MB for photos, so an 11-50MB image passed the early check, was reported
sent, and was rejected by a layer the model cannot see. The early check is
now per-type and mirrors the sender's limits.
MediaType called every image/* an image, but a photo endpoint rejects
image/svg+xml and image/tiff outright. Photo-safe formats are now an
explicit set, so an svg goes as a document and arrives, which is what the
"unrecognised means document" reasoning intended all along.
Also tests the readability probe's own error path (mode 0000), which was
the one case it exists for and the one case not covered; verified it fails
with the probe removed.
Remaining review findings are filed, not fixed: fallback lines appended
after the live reply finalizes are dropped by a stopped renderer
(archie-core-atv2), confinement is symlink-blind and re-opens after
checking (archie-core-jkmu), and concurrent uploads are uncapped
(archie-core-3svc).
The agent could produce a file -- a transcript, a log dump, a screenshot --
and had no way to hand it over. Asked to send one it offered a "download"
link; nothing arrived and it had no signal that anything had failed. The
defect was not that it lied: success and silent non-delivery were
indistinguishable to it.
Three places made a local file undeliverable, and each is fixed here:
- liveReply.Media skipped any attachment with an empty URL, so a local
file was dropped before it ever reached a sender.
- telegramMediaSender built every attachment as models.InputFileString,
which tells Telegram to FETCH the string as a URL. A host path is not
fetchable, so the send did nothing while reporting success. It now
chooses by what the attachment carries: URL -> fetch, Path -> upload
via models.InputFileUpload over an open file.
- Nothing let the agent ask for a send at all. New send_file tool
(internal/tools/sendfile) validates the file and returns a
MultimodalResult naming it; the turn's channel delivers it.
Meaning changes, called out because the names did not:
- MediaAttachment gains Path, the alternative to URL. Exactly one is
meaningful for an outbound send.
- tools.MultimodalResult.URLs used to mean remote-hosted media only; it
now means media this result wants delivered, which may be local. JSON
name unchanged so existing producers keep working.
- errNoURL is now errNoSource: having neither source, not having no URL,
is what makes an attachment undeliverable.
Failures are visible at every layer, which is the actual fix. send_file
errors on an absent, irregular, unreadable, oversized or confinement-
refused path, so the model sees it and can report it. The sender checks
Telegram's ceilings (10 MB photos, 50 MB otherwise) before the request and
reports the real size rather than leaving a 400 nobody reads. A delivery
that still fails now appends "could not send <name>" to the reply instead
of an empty link, which read as a delivered file.
Exposing a public URL for local files was explicitly rejected: it trades a
missing feature for a hosting and access-control problem.
send_file resolves paths through builtin.ResolveReadable -- the read tool's
own policy, exported rather than reimplemented -- so sending cannot reach
what reading cannot, whatever [chat] unrestricted_filesystem is set to.
Autonomous workflows now have exactly one production boundary: a task-scoped archie-agent receives a full-task core-NATS request and runs the worker-local stage loop. The legacy [agent] section remains decode-only for migration; it no longer selects runtime behavior.
Render completed ToolCallEvents as Telegram-safe streaming blocks
with compact name/status and a short cleaned preview. Summarize
file contents and listings instead of dumping source or quoted
JSON. Collapse repeated legacy turn-budget failures.
Remove max_tokens turn gating and the 200000 turn_budget_chars
default so absent or commented config cannot restore an aggregate
cap. Keep per-result max_result_chars and spill.
Confine find/grep discovery to the active workspace/worktree.
Refs #604
internal/tools/minimax wraps MiniMax's v2/video_generation (submit)
and v2/query/video_generation/{id} (poll) endpoints. Generation is
asynchronous on MiniMax's side but archie's tool subsystem has no
async-tool flow control (ToolEntry.IsAsync carries no flow control
today), so GenerateAndWait blocks the calling goroutine -- a tool
call, not the model's own generating goroutine -- submitting and
polling until the job reaches a terminal status or MaxWait elapses.
The poll loop tracks elapsed time purely from its injected Sleep
calls rather than a real time.Now() deadline, so a test that fakes
Sleep as an instant counter also makes the loop terminate instantly
instead of spinning in real wall-clock time.
Text-to-video only for now; image-to-video (image_url content items)
is a schema-and-routing change this can grow into. The tool returns a
tools.MultimodalResult carrying the video as a URLs entry -- the
shared envelope the gateway wiring downstream turns into a delivered
attachment.
MultimodalResult only modeled local bytes written under SubdirHint
(Files). A tool whose media lives at a remote, already-hosted URL --
a generation API's result, for example -- had nowhere to put it.
MediaRef{Type, URL} and the new URLs field give it one, reusing the
existing envelope rather than forking a second convention for the
same concept.
Ephemeral task containers re-resolve and re-download npx-launched MCP
server packages from the npm registry on every task, since each task
gets a fresh container with no persistent cache -- a worse amplifier
of a registry outage than the daemon-restart case d4bb914b already
fixed.
Extract the daemon's npm-cache env logic into mcp.NpmCacheEnv, shared
by both the daemon's own long-lived npx MCP servers and archie-agent's
per-task ones. Mount a new archie-cache-mcp-npm Docker volume at
/data/cache/mcp-npm on every task container regardless of ecosystem
(MCP servers are daemon-wide config, not tied to the task repo's
language), and point npx-launched per-task stdio transports at it.
Closes archie-core-qk8m.
Six mechanical fixes from `golangci-lint run ./...` (30 findings total, 23
deliberately left -- see the next commit's message for the one that grew
into something bigger, and the following note for the rest):
- apply.go: %v -> %w for the wrapped yaml error, matching the identical
double-%w pattern already used in decode.go/discovery.go/loader.go in the
same package (errorlint).
- contract_test.go: filepath.Join("provider.go") with a single argument is a
no-op join; replaced with the literal (gocritic).
- inprocess_test.go, yaegi_test.go: two goroutines whose whole purpose is
timing (does Execute return promptly on cancellation / does Run block
forever), where the call's result is irrelevant -- explicitly discarded
with `_, _ =` and a comment saying why, rather than silently ignored
(errcheck).
- find_test.go: an unchecked os.MkdirAll now matches the file's own
createTestFile helper, which already checks and t.Fatals on the same
operation (errcheck).
Not touched, and why: 18 cyclop/gocognit/gocyclo/funlen/nestif findings
require extracting helpers to reduce complexity -- the exact shape of two
regressions already recorded in CLAUDE.md's Lint-driven changes section
(one from --fix silently dropping a predicate, one from a nestif extraction
that doubled data), so a blind sweep across 9 unfamiliar files was declined
rather than risk a third. 2 godox findings (storage.go, worktree.go) mark
real missing work; deleting the comment would hide debt, not fix it. 3
recvcheck findings (Duration, TaskEnvelope, Provenance) each mix a
mandatory-pointer-receiver method (UnmarshalText/UnmarshalJSON's interface
contract, or genuine mutation) with correctly-value-receiver read-only
methods -- forcing consistency would either break interface conformance or
force pointer receivers onto call sites where the value isn't addressable.
Adversarially reviewed by a fresh subagent: all six confirmed
behavior-neutral or behavior-improving, all touched packages pass -race,
golangci-lint confirms exactly 23 issues remain (not more, not fewer), and
none are new.
Adds a human-consent gate for tools marked RequiresApproval, exposes
session management as LLM-invokable chat tools, and gives the operator
/delete to retire conversations permanently.
**Tool-approval gate (internal/tools, internal/agentexec)**
- ApprovalRequester contract lives in internal/tools (not gateway) so
agentexec can import it without pulling in the chat-adapter stack.
- ToolSetOptions.Approval wires the approver through to toolExecute.
- BuildToolSetFrom omits RequiresApproval tools when no approver is
configured; dispatch paths (invokeTool, tool_call bridge) refuse
them too, so the invariant holds at every indirection.
- The gate fails closed: only Approved/PermanentlyApproved pass;
an unknown decision value from a faulty approver does not execute.
- Tool-level timeout (2 min) returns a plain error, not a context
cancellation, so the model can report "approval timed out" instead
of the turn aborting.
- BuildApprovalDescription callback lets a tool surface the specific
resource being acted on ("Delete session \"abc123\"") rather than
showing a static tool description.
**Session tools (internal/gateware/session_tools.go)**
- session_list (idempotent), session_resume (mutating), session_title
(mutating), session_delete (mutating + RequiresApproval).
- Scoped to the current platform + channel: GetByChannel replaces
List() so a model in chat A cannot list or manipulate chat B.
- session_title validates the session belongs to this channel before
writing; session_resume/session_delete resolve refs against the
channel-scoped list.
- Shared limit helper extracted to tools.ListLimit.
**Telegram approval adapter (internal/channels/telegram/approval.go)**
- telegramApprover implements ApprovalRequester via the three-button
inline keyboard (Approve/Permanent/Deny). 2-min token expiry.
- Own map + mutex + "approval:" callback prefix — separate from the
dangerous-command system.
- Permanent approval uses exact-match (not prefix) lookup keyed by
action + resource description, so permanently approving "delete
session abc123" only covers that one session, not every call to
session_delete for 24h.
**Session lifecycle**
- /delete <session-id>: removes a session and its history. Refuses
prefixes shorter than 4 chars even when unique. Exact ID wins over
an ambiguous prefix. sessionTracker.forget sweeps every channel+
thread pointing at the deleted session.
- Extracted resolveSessionRef and shortSessionID for /resume and
/delete to share.
**Wiring**
- submitTurn builds the approver onto ctx before the LLM turn.
- chatGenerateOptions reads approver from ctx, passes it through
ToolSetOptions.
- SessionTools built per turn on the current message, combined with
taskTools into a single toolset.
- Reword docs and comments to drop vague "legacy" qualifiers where they no longer describe deprecation status
- Remove unused code: RecentEvents, MessageEvent.ToLegacy, index setState, stderr redaction tool, telegramOnlyCommands
- Delete or comment out tests covering removed or superseded behavior
The WalkDir callback swallows every per-entry error and returns nil,
so the only error that can escape is ctx.Err() from the cancellation
select. The old guard compared that error against itself and never
fired, so a timeout or /stop fell through to whatever partial matches
had accumulated, or "no matches found" when it had none -- telling the
model a search found nothing when it never finished.
Refs archie-core-ts9
provider.Registry.Start was atomic for every provider: one failure
unregistered every other provider including the builtins, returned an
error up through the capability host, and exited archied. Under
Restart=on-failure that is a crash loop. Observed on carina 2026-07-30 --
mcp.desktop-commander could not start because the runtime image was
missing unzip, and the daemon restarted continuously, losing chat, the
gateway and every builtin tool over one third-party npm package.
Atomicity is right for a provider archied requires. It is wrong for an
external MCP server pulled from a registry at runtime, which is exactly
the category allowed to be absent -- the same reasoning already applied
to the forge, and to containers.
RegisterOptional marks a provider the daemon can run without. Its
failure to start, discover or index stops only itself and excludes it;
Register keeps today's rollback. Required/optional is a deployment fact
rather than a property of the code, so the registrar states it instead
of the engine declaring it in its own manifest -- which also keeps
plugin.Manifest metadata-only, per the plugin engine rule.
A skipped provider's dependents cannot run either: an optional dependent
is skipped with the reason, a required one is a hard failure, since it
declared it cannot work without something now absent.
Degradation is reported, not silent -- Health goes degraded naming each
provider and cause, Skipped() exposes the set, and the daemon logs one
error line per exclusion. Silence is how this went unnoticed until the
crash loop. A provider that half-started before failing is still
stopped, so no child process is left behind.
The regression test drives the real capability host, the path main.go
calls: with desktop-commander registered as required it fails with the
verbatim carina error, and as optional the daemon starts with its
builtins intact.
Closes archie-core-ky6
The rg package carried three platform binaries (14MB) inside the Go
binary and extracted the matching one to a temp file on every single
run, so a long-lived daemon left a trail of copies behind and paid the
write on each grep. Nothing depended on ripgrep's own output format.
grep now uses the Go implementation unconditionally.
The nestif extraction in 712388d pulled --provider handling out of
parseModelRest into a helper that took modelParts by pointer and
appended every surviving field to it. The caller's own loop appended
the same fields again, so every model name was doubled: /model gpt-5.6
asked the router for "gpt-5.6 gpt-5.6" and SetActiveModel rejected it
as unknown. The index loop becoming a range loop also stopped the
space-separated "--provider openai" value being consumed with its flag,
so it landed in the name too. Direct model switching over Telegram was
broken for everything but --refresh and provider-only switches.
splitProviderFlag is now the sole accumulator and returns the surviving
fields, so the shared-slice shape that allowed the double append cannot
recur (archie-core-0d6).
Found while auditing the remaining hunks of 6108fc5 (archie-core-ifi):
the naked return in the same function could not be regression-tested
until the function produced correct values. That audit is complete and
the other seven hunks are behaviour-neutral. Notably shell.go's
errorlint rewrite is not a repeat of the grep failure - it had no
companion exit-code condition to drop, and the ctx.Err() timeout check
already runs first - and find.go's errors.Is rewrite is safe because
errors.Is(err, nil) is false for any non-nil err.
Tests cover the return order of parseModelRest (two strings and three
bools, so a transposed pair compiles silently), the field count of the
model name, and shell's exit-code branch, which had no coverage at all.
The find fallback's habit of reporting a cancelled walk as "no matches
found" is filed as archie-core-ts9.
An errorlint autofix rewrote the ExitError check to match any exit
code, so a search that never ran (bad pattern, unreadable path) was
reported as "no matches found". Keep errors.As, restore the exit-code
check, and add a regression test.
Add internal/tools/command: an always-on floor under the shell tool,
enforced in the provider adapter so the lifted tau files stay diffable.
This is not a security boundary and does not pretend to be one. Anything
with a shell has countless ways past a rule list. What it buys is that
the failures with no recovery -- where you cannot intervene afterwards
because the machine, the filesystem, or the gateway itself is gone --
are not reachable by an ordinary mistake.
Rules: recursive delete of / and system directories, mkfs, raw block
device writes, fork bombs, kill -1, shutdown/reboot/init/systemctl
power, and sudo -S with no SUDO_PASSWORD configured. Two more are
archie's own: an agent must not stop the daemon that hosts /stop
(pkill/systemctl/docker), nor start a second one outside the supervisor.
The parsing is the actual work. Patterns are anchored to command
position, so `git commit -m "revert the rm -rf / change"` and
`echo "rm -rf /"` are ordinary commands, while `echo x; rm -rf /` and
`$(rm -rf /)` are not. Quoting, separators, ${VAR} braces, environment
assignments and privilege wrappers are all parsed rather than matched.
Probing the first version found two bypasses that made every rule
optional, both of which a model writes unprompted rather than only an
attacker: `bash -c "rm -rf /"` and `xargs rm -rf /` hide the command in
an argument. Segments are now followed into shell -c payloads, eval, and
command wrappers (xargs, timeout, nice, env, nohup, ...), bounded by a
nesting limit. Tests cover both directions -- the refusals and, at equal
length, the ordinary commands that must not trip them.
shell.go relied on exec.CommandContext's default cancellation, which
kills only the process it started. A shell's real work is a child of
that shell, so cancelling `go test ./...` reaped the shell and left the
test binary running, and `nohup x &` was never touched at all.
That made /stop a lie for the case it exists to cover: the turn
unwound, the user was told it stopped, and the command carried on.
Worse than expected, and caught by writing the test first: without a
process group the tool does not merely fail to kill the command, it
hangs. The backgrounded grandchild inherits stdout, so cmd.Wait blocks
until the shell timeout regardless of cancellation. Verified by
reverting the fix -- the new test fails on a 30s hang, not on the
survival assertion.
Run the command as its own process group, signal the group on
cancellation, and bound Wait with WaitDelay so an inherited pipe cannot
hold the turn open. setProcessGroup/signalProcessGroup were already
present in the lifted set, used by tau's agent.go but not its shell
tool; this extends them to the shell and clears the unused-code lint
that was flagging them.
Recorded as a mutation in shell.go's provenance header.
The chat agent had two tools -- memory_edit and skill_activate -- so it
could not read a file, search a tree or run a command. ai-sdk's runtime
already drives a tool loop for it (MaxSteps 8, toolExecute wired as the
executor), so the loop was never the missing piece: the registry was empty
of anything useful.
Lifted from tau internal/agent/tools at commit f5289ea, per the recorded
copy-and-mutate strategy: every file carries a provenance header naming its
tau source path, the commit, and the mutations made, so refreshing from
upstream is a diff rather than archaeology.
internal/tools/builtin/ the lifted tools, ~3.1k lines
internal/tools/provider/builtin/ the adapter to tools.ToolEntry
Six tools: read, write, edit, shell, grep, find. Tau's own tests came with
them and pass unmodified, which is the evidence that the lift is faithful.
Deliberately not lifted:
- agent.go (1214 lines): subagent spawning, and the only real coupling in
the package -- 10 tau packages including chat, sessions, store and
registry. Nothing the chat agent needs to do work.
- skill.go: depends on tau's skills package; archie has internal/skill and
already exposes skill_activate.
- docs.go: serves tau's own embedded documentation via tau/docs.
Not lifting the coordinator either. It is tau's agent loop (6.4k lines over
ten packages) and archie already has one in the ai-sdk runtime; porting it
would leave two loops competing. Its cancellation and tool events are worth
having, but they are a few hundred lines built against the existing path
rather than six thousand lifted onto a second one.
The adapter bridges two shape differences: archie hands handlers decoded
arguments where tau's executor takes raw JSON, and tau reports tool-level
failure in Result.IsError rather than as an error -- surfaced as a real
error so the runtime and the guardrail engine both see it. Tools are
classified idempotent or mutating for guardrails and budgets; that is
categorisation, not a jail.
chat.workspace roots the tools and defaults to empty, which registers none
of them. They read, write and execute, so the directory has to be a
deliberate choice rather than whatever the daemon started in -- in the
container that is /.
Verified: go build, go vet, gofmt clean; full suite passes.