bot_user names the agent that chat turns and the session curator attribute memory
to. The loader requires it of a single-identity deployment, but nothing covered
the two paths between that requirement and the consumers.
local-ollama-standalone.toml set no bot_user and defined no [[identities]], so a
profile AGENTS.md lists as supported could not load at all (archie-core-cx0d).
Give the profile the field, and make the regression guard enumerate
deployments/*.toml through the production loader instead of naming two files, so
a profile added later is covered the moment it is committed. dev.toml is
excluded because it is a development overlay layered over an installed config
rather than a standalone profile, and the exclusion is itself checked so a
renamed overlay is noticed instead of silently widening the rule.
The root BotUser was validated only when no [[identities]] block was present, so
a config with exactly one identity left it empty: the curator then skipped every
session with "no agent id" while the chat turn path ran with the same empty agent
(archie-core-4slp). applyIdentityDefaults now adopts a sole identity as the root
single-identity values -- which is what those fields mean -- and leaves the root
alone when two or more identities leave no single answer. The skip reason now
names bot_user, so the remaining case reads as configuration rather than as a
data problem.
Both regressions were red-first: the enumerated profile test failed on
local-ollama-standalone.toml before the profile was fixed, and the sole-identity
test failed before applyIdentityDefaults adopted it.
VerifyHMAC compared the supplied signature against the computed digest as
case-sensitive hex text, so a sender that emits an uppercase digest failed
verification -- and the failure is indistinguishable from a wrong secret. The
helper is shared by all three inbound surfaces (forge intake, the chat webhook
gateway, capture intake), so one such sender breaks whichever surface it
targets.
Decode both sides to raw bytes and compare those. hex.DecodeString accepts
either case, and comparing digests rather than their text keeps the comparison
constant-time -- strings.EqualFold over the encoded text would not be. A
signature that is not valid hex is rejected outright, and a digest of the
wrong length simply does not match.
archie-core-rl04
The adversarial self-review system was removed in d73e0748, but four pages
describing it survived: the architecture page, its design-decision PRD, and the
two superseded calibration/operator PRDs. The architecture page carried a
hand-written "removed, not shipped" banner and was kept only because live PRDs
linked to it, so every reader paid for a system the tree no longer has.
Delete all four and repoint the live referrers at the authority that replaced
them (docs/prds/pr-review-agent.md): inline-review.md's Builds-on line,
pr-review-remediation.md's isolation-contract citation, pr-review-agent.md's
supersede clause, CLAUDE.md's contributor rule, and three Go comments that cited
the deleted paths. Regenerate docs/data/generated/dev-docs.json.
History stays in git; the tree stops describing a system it does not have.
Gitea reports a review comment position as a uint64. Past the int range the conversion wrapped negative and would reach a reviewer as a line that cannot exist. Refused instead, with a test that fails when the guard is made inert.
l45's refactor removed the assigned-but-unread daemon handles and l44 had
already moved the guardrail wiring out of the daemon, so the rebase union
left internal/daemon/daemon.go importing internal/tools with no use and
internal/app/archied/bootstrap.go assigning a Guardrails field that no
longer exists. main did not build.
Caught by building the merged tree rather than trusting the merge; the
merge train merged on a failed gate instead of chaining on it, which is
the mistake this commit repairs.
pinTaskProfile parked every pinWorkflowDefinition failure as
taskstate.ParkTransient. That class documents that a requeue can succeed,
but the dominant pin failures are durable: a stored definition whose YAML
does not hash to the digest beside it, a definition that does not parse,
and a workflow or playbook binding resolving to an undefined workflow all
recur on every requeue until an operator repairs the row or the
configuration. A future retry pass reading transient as its input would
loop on them forever.
pinFailure now carries the class its cause needs, decided where the cause
is known, and pinParkClass reads it at the park site. A durable pin defect
parks operator-actionable; only a State Store read or persist failure --
the genuinely environmental case -- stays transient. The safe default for
an unclassified cause is ParkNeedsHuman, never a requeue that will fix it.
Red-first: TestDurablePinDefectIsNotParkedTransient failed at its
assertion before the fix (parked "transient"); TestTransientPinFailureStaysTransient
keeps the environmental case transient so the rule is per-cause, not a
blanket flip.
Three exported Daemon fields were wired by the composition root and read
nowhere: CapabilityHost (*plugin.Host), Curators (*curator.Registry) and
ToolRegistry (*tools.Registry). No field access exists anywhere in the
repo, and git history shows no read ever existed.
Each handle has a live owner outside the daemon: the curator
registry/runtime is boot-owned (the curator isolation invariant keeps it
off the agent loop), the capability host is lifecycle-owned by the
composition root and consumed by the control plane, and the tool registry
is consumed by chat, curator and MCP wiring. Keeping the daemon copies was
decorative wiring, so the fields, their composition assignments and the
now-unused plugin/curator imports go together.
The daemon architecture test now refuses the three names, and the plugin
host contract test refuses a CapabilityHost: handle on the daemon alongside
the PluginRegistry: one it already refused.
TaskContext.Guardrails was never assigned: the engine was built in
archied's composition root and held on Daemon.Guardrails, which no code
read after workflow execution moved to the full-task worker handoff.
archie-agent is a separate process, so a live engine cannot cross that
handoff. Build the engine at the agentworker composition root (the
process that runs agent stages), thread it through the task dependencies,
and set it on TaskContext; drop the now-dead archied-side construction.
Refs: archie-core-jd5v
The remediate round cap runs inside archie-agent and reads
TaskContext.Cfg.MaxRetries, but TaskConfig did not carry the global
Config.MaxRetries: ForTask and ToConfig both dropped it, so a default
configuration reached the cap as 0 and a remediation never parked. Carry
the value (a repo-level override still wins via EffectiveMaxRetries) and
pin it with a red-first cap test traced through ForTask/ToConfig.
Refs: archie-core-ad3p
Commit 14c4de77 moved the remediate resume into daemon worktree
preparation and deleted StageResumeWorktree together with the Dir/Resume
pair it called, but nothing in the workflow replaced the binding it did:
TaskContext.Dir and Branch stayed empty for a whole run, so
remediate-build ran the builder with no workspace and
remediate-commit-push called CommitAll/Push with an empty dir and branch.
Extend the workflow-side Trees seam with the prepare target, mirroring
the daemon's mode-aware worktree.Manager.Prepare: PrepareFresh lands on
base, a non-empty target is the branch a resume continues. Remediate now
stages StagePrepareWorktreeOnBranch before any agent stage, binding the
same directory the daemon prepared and the task's PR branch, and failing
closed when the task has no branch rather than preparing base. The
container adapter returns the requested resume target verbatim.
Refs archie-core-h4tk
A task hosts the same configured MCP servers as a chat turn, but the
worker's provider builder passed no sampling handler, so a task-scoped
server's sampling/createMessage request got a JSON-RPC method-not-found
while the identical request succeeded on the chat path. The worker runs
its own model runtime, so it can answer.
Thread a sampling handler built from the task's builder model into the
worker's MCP provider builder, beside the chat path's handler. Build the
task's model runtime once and share it with the stage runners instead of
creating a second client for sampling.
Move the sampling message mapping and max-token bound to agentexec, next
to the runtime both paths call models through, so a server gets the same
bound on either path. Carry the catalog's reasoning flag in
ModelLimits: a reasoning-class builder model must omit the max-token
bound its provider rejects, exactly as the chat path does.
ShippedSoul returned an empty SoulDocument.Legacy, so SoulSuperseded,
replaceSoul and SoulUpgraded had no production producer: the only body the
build ships is its current default, which classifies as SoulCurrent, and only
tests constructed a Legacy document. No earlier SOUL ever shipped (the file was
introduced with the starter in a80cca7d), so the upgrade path has nothing to
consume.
Remove the speculative machinery and keep the behaviour that is real: seed the
starter when no file exists, recognise the shipped default as current, and
preserve every other body, including a template-shaped edit. A build that
supersedes the starter reintroduces the upgrade together with the old body it
consumes.
ui/src/settings/validation.ts hand-mirrors three control-plane validators with
no shared source, and the model-role rule had already diverged: the dashboard
required a stricter regex while the server only asks for a slash, so it refused
values the server accepts.
Add a shared fixture (internal/app/controlplane/testdata/dashboard-validators.json)
that both sides run through their real implementations: a Go test over the
server validators and ui/test/control-plane-validator-parity.test.ts over the
dashboard ones. Either side changing a rule alone now fails the other's test.
Fix the diverged rule to mirror the server's predicate exactly (non-blank role,
model containing a slash), so an empty or whitespace-padded model is judged the
same way the control plane judges it. Regenerate ui/dist and the dev-docs
artifact.
Telegram had no generic question or choice widget in the agent-turn path,
so the interactive contract could not be carried by the only channel that
ships. telegramInteractor implements ClarifyRequester and PickerRequester
on the shared TextFallback and reports both capabilities honestly: the
capability is about carrying the interaction, not the widget, and the cost
is a typed answer. A reply typed while a prompt is pending is consumed by
the blocked interaction rather than dispatched as a new turn; a command is
never consumed, so /stop still cancels the turn.
The carrier is set on the turn context in the Messaging process. It does
not cross the gRPC boundary into the Gateway process, so in the extracted
messaging deployment the gateway turn has no interactive adapter and
ask_user is not offered -- the same gap that already makes
WithApprovalRequester unreachable there. That transport is a product
decision recorded on the lane, not silently worked around; the in-process
path, the capability report and the tests land here.
Red-first: TestInteractivePickUsesTextFallback and
TestInteractiveClarifyReturnsTypedAnswer (plus the command guard) failed at
the assertion before the interactor was implemented.
prepareTurn now registers an ask_user tool only when the turn context
carries an interactive adapter (WithInteractive), so a contract with no
caller has a real one and a channel that cannot ask registers no tool
rather than one that always fails. A channel reporting Picker gets the
native single-choice request; one reporting only Clarify still carries a
choice, rendered as numbered text and matched from the typed reply through
the domain's FormatPickerText/ParsePickerReply rather than erroring.
Red-first: InteractiveTools returned nothing and
TestInteractiveToolsRegisteredWithCarrier / TestAskUserUsesPickerWhenChannelCanPick
/ TestAskUserFallsBackToTextWhenChannelCannotPick failed at the assertion
before the tool was implemented. TestTurnRegistersAskUserOnlyWithCarrier
proves reachability from a real TurnRunner turn.
The contract could carry a question but nothing carried the ability to ask
one. Add Interactive, resolved from a channel adapter's own capability
report through InteractiveOf, with WithInteractive / InteractiveFromContext
as the per-turn carrier -- the same shape WithApprovalRequester already uses
for consent. The capability still wins over the method set, so a
mis-reporting adapter contributes nothing instead of being called.
Tests: TestInteractiveOfTrustsCapabilityReport and
TestInteractiveCarrierRoundTrip pin the resolution and the round trip.
Flip tool-settings to ApplyMode live (archie-core-zfb0.3): the daemon now
watches the kind, diffs the stored MCP server set against the running
providers -- connecting added servers, disconnecting removed ones and
reconnecting changed ones while an unchanged server keeps running -- and
rebuilds the web_fetch and minimax entries, whose config is captured at
construction. A changed server whose replacement fails to start is rolled
back to the one running, and the refusal is reported through apply status.
toolprovider.Registry grows running-state Add/Remove/Replace beside
Register, Start and Stop, isolating a provider's failure the way the plugin
host's Add/Replace do. The chat turn model resolves its tool-output policy
per turn instead of capturing it at construction, and the boot registration
reads the running holder so a stored update that lands during boot is not
lost.
The container runtime policies kind was restart-required because the pool
captured image, pull policy, network, max uptime and max concurrency at
construction and the dispatcher fixed its slot count at the start of a drain
pass. Both consumers now re-read the published config:
- the pool's settings come from a live source read per acquire, with the
boot-resolved registry credential and Docker construction values preserved;
a live default-image change is pulled rather than assumed present;
- the daemon owns one long-lived dispatcher that accepts a new limit: running
tasks keep their slots, a raise admits queued work immediately, and a lower
limit is applied as tasks finish;
- the control-plane descriptor is live, the kind joins runtimeResourceKinds,
and the apply/reload paths resize the dispatcher after publishing. The
ContainerConfig fields whose consumers now re-read are added to the reload
allowlist; registry_auth stays restart-required.
Regenerate the docs artifacts. contracts.json also carries the voice
media-type description a prior commit left stale.
The pre-mutation file checkpoint added in dcc97afd wrote a full copy of
every write/edit target into .git/archie-checkpoints, but nothing could
restore it: CheckpointStore.RestoreLatest (and its latest/restore
helpers) had no production caller, no rollback surface reaches the
workspace, and the store was never pruned, so it grew without bound in
every workspace.
A writer whose only reader is unreachable is dead weight and an
unbounded disk cost, so the write half goes with the read half: write
and edit no longer snapshot, and CheckpointStore is deleted. The
regression test pins the decision by asserting a write and an edit
leave no .git/archie-checkpoints behind.
archie-core-zfb0.4. Channel settings were restart-required as a whole: a
change to one channel needed a process restart that restarted every channel.
The Messaging Service now reconciles the stored channel-settings resource on
the apply-status restamp interval (docs/prds/plugin-settings-live.md), re-layers
it over the file document exactly as boot does, and restarts only the channel
whose transport settings changed. Every other channel keeps serving.
A restart is stop-and-start, not an in-place reload: each channel gets its own
supervisor and a rebuild factory, so email and webhook -- which have no reload
seam -- are restarted too. Telegram's Start still calls dropPendingUpdates
before it long-polls, so the rule holds on the reconcile path as on /restart. A
replacement is built before the running channel is touched, so a configuration
this build cannot run leaves the running channel serving.
The composed set and the bound listen addresses are process bindings: enabling
or disabling a channel, or changing chat.webhook_addr / chat.email.listen_addr,
is refused and reported for a restart rather than applied. A channel rebuild
that fails is not recorded as applied, so it is retried on the next tick. The
applied version reaches apply status, and ChannelSettingsKind flips to live.
Files: internal/app/archiemessaging/{compose,reconcile,supervise,run}.go,
controlplane ApplyMode, archied runtime watch, configuration.md, the config
skill matrix and the regenerated dev-docs artifact.
Verified: go test ./internal/app/archiemessaging/... (race clean); go test
./cmd/archie-messaging/... ./internal/app/controlplane/... ./internal/app/archied/...
./internal/config/...; go vet and golangci-lint run on those packages (0 issues);
go fix ./... and golangci-lint fmt produced no diff; task docs:artifact:check.
The Messaging Service built the voice-transcription client in-process
from [models].transcription and [providers.*], which the messaging
boundary PRD reserves to the process that owns the model runtime and its
provider credentials. The deletion gate only measured linked packages,
so the PRD violation shipped green: config reads are not imports.
Move the capability across the boundary:
- archiemessaging no longer reads [models] or [providers], builds no
transcription client, and drops the ResolvedConfig.Transcriber field.
- telegram classifies a voice note as messaging.MediaTypeVoice and stops
transcribing; it carries the attachment bytes and its note to the turn,
which is all a frontend without the credential can do.
- archied builds the transcriber beside the chat runtime from its own
model-role configuration (setupTranscriber) and gives it to every
gateway TurnRunner.
- the turn runner transcribes a speech attachment before recording the
inbound message, so stored history and the model context carry the
transcript; a nil/failed/empty transcription keeps the frontend note.
- messaging.TranscribedMessageText renders the transcript into the text a
frontend built from its media note, preserving the sender's caption.
Red first: TestTurnRunnerTranscribesSpeechOnTheModelOwningSide failed at
'Transcribe calls = 0, want 1' with the transcription call disabled.
review.precision_gate and review.approve_before_post were file-only, so the
dashboard had no way to write them: no control-plane resource carried them.
Add the review-settings kind (docs/prds/review-settings-resource.md), seed it
from the file's [review] section, layer it field-by-field into cfg.Review,
watch it live, and publish it through ConfigView.Review. The dashboard gets a
Review settings page bound to the kind through PAGE_RESOURCES, showing each
dial beside the value in force.
docs/prds/workflow-step-vocabulary.md named the repository-rules step type
`workflow.diff-rules` while the vocabulary a deployment resolves carries
`gate.diff-rules` (workflow.DiffRulesStepName), so a definition written to
the document's word is refused at save and at run. The same decision's
sibling, docs/prds/runtime-control-plane.md, still claimed "Plugins can add
step types", which this document's closed-set rule contradicts: an
interpreted plugin contributes none, and a step type arrives as a provider
in the compiled-in set.
Correct both documents, and hold the vocabulary PRD to the provider set with
a guard test: every step-type-shaped name the document identifies must
resolve in workflowsteps.Providers(). The test fails on the pre-fix document
("workflow.diff-rules" is contributed by no provider) and passes after.
Refs: archie-core-vwqc, split from archie-core-gcoj (closed as invalidated by
re-verification: the provider set exists and cross-side agreement is already
pinned).
A retry now carries the operator's worktree choice (refresh onto the base
branch or continue the work already pushed on the task branch) from the
dashboard through the task action to the daemon dispatch, and persists it
on the task row so it survives a reload. The daemon resolves the mode from
the task, never from the workflow; BeginRemediation records continue
explicitly so a remediation still resumes its pull request branch.
The control-plane catalog advertises identities, capture-events,
capture-mappings and capture-bindings as domain-managed: the domain API
validates their writes, not the generic resource store. Two of those
paths did not hold.
Identities: the file loader judged an identity name with its own
`name == ""` test, while the store derives the same string through
identity.New, whose rule rejects an empty-after-trim name. A
whitespace-only name passed the file and then failed State Store boot
(or was silently skipped by BootstrapIdentities). The loader now calls
identity.ValidDisplayName, the feature's own rule, and
BootstrapIdentities derives legacy names through
identity.FromLegacyName rather than dropping one it cannot build.
Capture mappings and bindings: mapping.Validate and binding.Validate
ran only in the archie-ui handlers, so any other client of the RPCs the
catalog names could store a mapping or binding the feature refuses.
The staterpc write handlers now apply the feature's rule before the
store, per runtime-control-plane.md ("the server validates the change,
then writes it"). The rule stays in the domain package; the handlers
only translate its refusal.
Capture-events has no file-side copy and no feature rule for a captured
event's content, so there is nothing to unify there.
Tests: identity_agreement_test.go ties the file loader's name rule to
the store constructor; domain_validation_test.go proves the mapping and
binding RPCs refuse an invalid value before it reaches the store.
A remotely-shipped task log entry was re-emitted through the sink with
slog.Logger.LogAttrs, which stamps time.Now(): the entry's own event time
was discarded and replaced with the daemon's receipt time. Because a
container's records arrive in bursts, a multi-minute run collapsed onto a
few milliseconds of receipt timestamps and its timing became
unrecoverable -- which stage ran when, how long an iteration took, whether
a stage stalled.
Write now builds the record from entry.Time (falling back to receipt time
only for an entry that carries none) and dispatches it through the same
handler, preserving the sink's level gate. Receipt time stays available
separately in the record's received_at field, so event time and arrival
time cannot be confused. The agent-side producer already carried the
record's own time; a test now pins that half too.
The OpenAI chat-completions provider in the pinned ai-sdk serialises a
non-zero output-token bound as `max_tokens`, which reasoning-class models
reject with HTTP 400 ("Use 'max_completion_tokens' instead"). Send no bound
at all for a model the catalog marks reasoning; classic models keep their
bound. Covers the chat turn path and MCP sampling, the two call sites that
set one today.
The reasoning flag rides TurnPrepareContext from the catalog details the
gateway already resolves, so archie-core keeps one notion of "reasoning
model": messaging.ModelDetails.Reasoning. A max_completion_tokens path in
ai-sdk is still needed to ever bound these models; that is filed separately.
prepareWorkspace chose between Manager.Prepare (refresh onto base) and a
separate Manager.Resume, and Manager.Resume only opened an
already-prepared clone. A remediation whose local worktree the terminal
cleanup removed, an expired volume dropped, or a different host never had
could therefore never be resumed, and the retry path had no way to select
the mode that continues pushed work.
Unify the two matched methods into Prepare(..., Target): Fresh resets onto
origin/<base>, a non-empty Target resets onto origin/<that branch>, and
either target clones a missing worktree. The daemon's retry path selects
the target from the task row and passes the persisted branch, so a resume
lands on the branch the open PR lives on and fails closed when it is not
on the remote.
Delete the container-mode no-op resume (hybridTrees.Resume and the
resumableTrees capability) now that the daemon owns it. The remediate
workflow no longer stages a resume, and the retired remediate.resume step
word stays resolvable as an inert stage: a stored definition is validated
as a whole collection, so a removed word would fail every workflow to
decode, not just its own.
A Telegram photo or document was dropped between the process that receives
it and the one that runs the turn. Telegram runs in archie-messaging, which
downloads the attachment; the turn runs in archie-gateway/archied, which
never holds the bot token. gatewayrpc stripped MediaAttachment.Data on the
inbound request, so by the time the Gateway's model seam ran, the
attachment had no bytes and no URL, and mediaParts dropped it -- the model
read only the "[photo]" placeholder.
- proto/gateway/v1: Media gains a `data` field, populated only on the
inbound request. Stored-history and outbound-event directions keep it
empty: a persisted record still strips the bytes, and a streamed media
event still carries metadata only.
- gatewayrpc: the inbound conversion maps Data; the stored and outbound
conversions deliberately do not.
- archied: the Gateway server raises its gRPC receive limit from the 4 MB
default to the messaging contract's inbound attachment ceiling plus
framing headroom, so a maximal (20 MB) attachment is accepted rather than
failing the RPC.
- messaging: MaxInboundAttachmentBytes is the single ceiling the Telegram
download gate and the Gateway receive limit both derive from.
Red first: a new TestInboundMediaBytesReachTheTurn asserted the bytes crossed
the wire and failed at that assertion before the mapping changed; it was then
consolidated into TestInboundMediaRoundTrip, which now pins metadata plus bytes.
The Telegram update-report relay verified every pending report against a
nil running-version map, so a successful update always ended with "I still
can't confirm which version is actually running" and read as a failure.
The 2026-09-19 extraction decision left RunningVersions nil because the
Messaging Service could obtain only another binary's stamp. ec0f17a7
(archie-core-k94o) then gave every host binary, archie-messaging included,
a buildinfo.Version self-report stamped with the archied release tag, so
the process now has a truthful source for its own running version: report
daemon = buildinfo.Version and leave the agent Unverified, since every
archie-agent process is task-scoped and observed only through
daemon.AgentStatus. A relaying process that the update did not replace
still reports its old build, which Verify reads as drift.
The deliver stage's notify() returned early when [notify].webhook was
unset, so a parked feasibility PRD left the process nowhere while both
its doc comments promised an issue comment that was never posted. A
default deployment sets no webhook: the PRD reached only the operator
who went looking.
notify now posts a one-way delivery notice on the task's forge issue
before the webhook, for a forge-backed task. The issue is the channel a
default deployment always has; the webhook remains an additional push.
Neither is a reply channel -- the decision surfaces stay chat and the
Web UI, so the notice carries the PRD and names where to decide rather
than inviting a reply, and no reply cursor is recorded (WatchCommentID
stays 0).
Both stale doc comments now describe what the code does.
docs/prds/plugin-settings-live.md: a `plugin-settings` change now applies
live for plugin_dir, module_dir and secret_engine_dir, and a file dropped
into any of those directories loads on the reconciliation tick.
- PluginSettingsKind moves to ApplyMode "live" and joins
runtimeResourceKinds, so a stored directory change re-layers immediately.
- A reconciler re-runs the directory loads and the layering on the
apply-status restamp interval. Files are content-hashed, so only new and
changed files load.
- plugin.Host grows Add and Replace for the running state (Register keeps
its pre-Start rule), and plugin.LoadFile / secret.Registry.LoadFile are
the single-file load paths the boot loader and the reconciler share.
- Yaegi cannot unload, so a removed file keeps running and the removal is
reported through apply status against the live version until restart; a
clean pass clears it.
- A plugin-settings document that changes skills_dir is refused before the
layering, since skills_dir has no live consumer.
- applystatus.Reporter.AppliedVersion lets the reconciliation report an
outcome without resetting the version to zero.
Add SlashAccessPolicy, a deny-by-default role policy (admin allowlist plus
per-user delegated commands), and enforce it in Router.route before any
local command dispatches. A refusal executes none of the handler, is
visible to the sender, logged, and recorded through an optional
SlashDenials sink.
No config field yet: the gatewayrpc wire drops Message.SenderID (only the
display Sender crosses), so a per-sender allowlist cannot match a Telegram
sender at the daemon Router until the layered auth mixin carries identity
(archie-core-263). Nil policy preserves today's behaviour.
The optional-capability pattern had media and delete but no interactive
UX: a channel could ask a human to approve an action through
ApprovalRequester, but there was no domain contract for posing a
clarifying question or presenting a single-choice picker, and no
capability flag a caller could read before attempting one.
Add ClarifyRequester and PickerRequester to internal/domain/messaging
alongside ApprovalRequester, with Clarify/Picker/Approval flags on
AdapterCapabilities and ClarifierOf/PickerOf/ApproverOf lookups that
treat the capability as the contract rather than the method set, so a
mis-reporting adapter degrades instead of panicking.
TextFallback is the shared text degrade: a channel with no native
question or selection UI still carries the interaction (numbered
prompt, parsed reply) instead of failing or silently dropping it.
FormatClarifyText/FormatPickerText/ParsePickerReply hold the rendering
and matching logic once.
The approval half reuses the existing decision model: Telegram's
telegramApprover already renders a native prompt, so it now reports the
Approval capability and deliberately does not claim Clarify/Picker it
has no native UI for. Gateway re-exports the contract.
Red-first: internal/domain/messaging/interactive_test.go drives the
test double adapters (native, degrade, mis-reporting, over-claiming)
and the text rendering/parsing; telegram/approval_test.go pins the
honest capability report.
The baseline stage treated every red gate as a pre-existing code defect and
dispatched a TDD repair agent. A failure the environment caused -- no Docker
daemon for the test harness, or a gate command that is not installed -- cannot
be fixed from the worktree, so the run spent a full agent budget and then
parked. A provider error during that repair run (observed: HTTP 402) replaced
the gate cause in the park reason entirely.
Classify at the producer: gateEnvironmentCause names a failure whose command
could not start (*exec.Error) or whose output names unavailable host
infrastructure (the Docker shapes from archie-core-71mq). That failure parks
with the cause and the gate output and starts no builder. A real code failure
still dispatches exactly one repair run, and every post-dispatch failure now
carries the gate output alongside the agent/provider error instead of masking
it.
Refs archie-core-rwy6.
internal/gateway/dashboard_tools.go mirrors the dashboard's route table so
the agent can navigate to every page. The new /settings/harness route needs
its entry or page_index would not offer it (archie-core-egkf.7).
The harness page lists each credential binding with its captured OAuth
state (expiry, scopes, capture time; never the tokens), and the setup
terminal route upgrades to a WebSocket and bridges it to a harness
session. The session is a narrow webui-owned HarnessTerminal seam: the
container, sandbox network and egress path belong to the daemon, so the
dashboard process cannot implement it and answers 503 until a transport
is wired (archie-core-egkf.7). Opening the terminal authorizes as the
update action on the secret, per the PRD.
The sandbox kit spec lets one credential@1 declare apiKey and oauth
together, meaning "whichever the host has bound": the published
docker/claude-code-kit does exactly that for anthropic, and codex for
openai. archie decided the kind from the Kit's declaration instead
(egress.IsOAuthManaged), so a dual-declared service always came out
OAuth-bound however the org had bound it:
- kitrun.resolveCredentials granted an API-key binding's service the
empty string, so the proxy had no key to inject and every request to
a domain the Kit named carried an empty credential. The reference
Claude Code Kit could not authenticate with an API key at all.
- an API-key-only service whose binding named no secret resolved to ""
through the zero SecretRef, presenting an empty credential rather than
staying unbound.
- the proxy fired an apiKey inject rule for an OAuth-bound service,
putting "Bearer " on the wire, and intercepted the token endpoint of
an API-key-bound one, whose only outcomes were a 502 on the missing
token set, a capture the org never bound, or -- for a required
dual-declared credential -- a refused launch when no token store was
configured.
- the container's SBX_CRED_*_MODE said "apikey" for an OAuth-bound
dual-declared service, and kitrun demanded a captured OAuth token for
an API-key-bound one whose Kit renders a credential file.
One root cause, fixed once: the org's binding decides the kind. It is
now a first-class fact (egress.CredentialKind: apikey or oauth, absent
meaning unbound), computed from the binding in resolveCredentials and
carried to every consumer of it -- the resolver grant, the container's
credential state (mode variable and credential file), and the proxy
session, whose rule sets are built from it so a rule can only fire for a
service the run carries as that same kind. A service the run does not
carry at all keeps both rule sets, so resolving it still refuses a
required credential and skips an optional one.
Declared and bound stay two independent facts, as the reported
declared-and-granted intersection requires: neither widens the other.
No secret value crosses into kit (Bound carries kinds only) and none is
logged or written to a task row.
A Telegram voice note was stored as the placeholder "[voice message]" and
nothing turned its audio into text, so the agent could not read it. Add a
channel-neutral Transcriber capability to the messaging media layer
(internal/domain/messaging), an ai-sdk Whisper implementation behind it
(internal/infrastructure/transcription), and wire it for Telegram.
The capability is config-driven like the other model roles:
models.transcription names a provider/model, the matching [providers.*]
entry carries the credential (api_key_env or an api_key secret ref), and
a missing role, provider, class or credential degrades to today's
behaviour with a log -- never an error that fails the turn.
A transcribed note replaces its placeholder with "[voice transcription]"
followed by the transcript, preserving any caption, so stored history
shows the agent it is reading a machine transcript.
An EphemeralReply is a status notice a channel delivers and then removes
once its TTL elapses, so a transient placeholder does not linger in the
chat history after it is superseded. The TTL lives on the reply, the
Delete capability joins AdapterCapabilities alongside Media, and
channels.EphemeralSender does the send-then-retract with an injected
clock. A sender that cannot delete still delivers the reply and logs the
missed retraction instead of failing the turn.
Telegram implements MessageDeleter through deleteMessage and honours the
mechanism for the "Archie reloaded." status notice; the pre-restart
"Reloading…" message remains the durable acknowledgement.
ImagePull was called with empty options, so a private registry answered 401
on every pull regardless of containers.pull_policy. docker login did not help
either: it writes ~/.docker/config.json, which only the CLI reads.
Add containers.registry_auth, a SecretRef reusing the credential shape the
rest of the file uses, whose resolved value is a Docker registry.AuthConfig
JSON document. archied resolves it at boot and the pool base64url-encodes it
into the X-Registry-Auth header ImagePullOptions carries.
A missing or unresolvable credential degrades to the anonymous pull that
existed before the field (with a warning), so a deployment with no private
registry is unaffected; a half-named reference is refused by
configuration.Validate. The field is file-owned and json:"-", so the
container-runtime-policies layering has to preserve it explicitly, the way it
preserves Profiles.
MCP servers could not ask the client for a completion: the transports
routed only responses by id and dropped every server-initiated request,
so a server calling sampling/createMessage hung waiting for a reply.
Route server->client requests through the transport readers to a handler
on the client. The client answers sampling/createMessage from a
configured SamplingHandler and maps a missing handler, malformed params
and handler failure to JSON-RPC errors for that request alone -- never a
silent drop or a failed session. stdio and SSE implement the routing;
the stateless HTTP transport cannot receive the request at all and says
so in its doc.
The daemon wires the handler to its own chat runtime and active model
role, so a sample is answered by the model a chat turn uses, capped by
the request's maxTokens and the model's output ceiling, and recorded on
/status like every other model call. No model configured is a refused
request with a clear error.
modelcatalog.Load ran once at boot, so a provider whose credential became
resolvable later, or a model models.dev published since, needed a restart.
The daemon now re-reads the catalog on an interval and republishes everything
derived from it, without a restart.
- `refreshModelCatalog` re-reads through the same options boot used, layers
the result with the process's own runtimeConfig, and publishes. A failed
read keeps the catalog already running: the model list must not go blank
because the catalog service had a bad minute.
- Boot now publishes the merged config into `config.Holder`. applyModelCatalog
only ever wrote into the boot struct, so the catalog's model limits and
discovered providers never reached the running config -- and therefore never
reached `config.ForTask`, which is what a workflow stage resolves model
limits from. The first live update was the first snapshot to carry them.
- The chat model manager's catalog is a contribution that is replaced, not
merged. A model the catalog no longer publishes stops being offered and
described instead of lingering from an earlier snapshot.
- The catalog snapshot and the refs it contributes live behind one lock: the
refresh goroutine replaces them while the config paths read them.
The daemon owns the published configuration, so the loop is the daemon's; the
standalone Gateway reads the catalog once at boot, as it does for the other
live-applied settings it does not watch.
The interval is deliberately not an operator setting: the catalog is derived
state rather than a stored resource, and a knob with no second consumer would
be a setting that parses and does nothing.
A key exported into the shell after the process started stays invisible to
every refresh, and cannot be otherwise: the process environment is fixed at
exec. A credential a secret engine starts resolving is visible, because the
catalog's credential lookup is the secrets registry.
Refs archie-core-zfb0.6
Descriptor.Validate duplicated only the forge-permission half of
Authority.Validate, so a descriptor could declare a blank credential
service, egress host, trigger or tool and be installed: the acceptance
record refused the same blank, leaving the descriptor the operator is
shown disagreeing with the record the package is bounded by. Delegate to
Authority.Validate, the one validation of a grant list, and cover it with
the blank cases the descriptor now has to refuse.
The built-in workspace write and edit tools replaced file content with
no recoverable pre-state. Each now snapshots its target immediately
before the write and refuses the mutation when the snapshot cannot be
taken, so a checkpoint is never silently skipped.
Checkpoints land in .git/archie-checkpoints inside the workspace, the
same per-task state directory that already holds the task brief and the
prepared sentinel: nothing under .git can be swept onto the task branch
by go-git's Add{All:true}. A checkpoint records either the file's bytes
or its absence, so creating a file restores by removing it;
RestoreLatest returns the most recent pre-state for a path.
Shell, Go-script, MCP and plugin tools are deliberately not covered:
the set of files they can touch is not enumerable from the call.