Address the second-pass review of the State Store ratification doc (beads archie-core-8cda.4.1):
- State migration-decisions.md section 4 remains authoritative and open; this PRD
does not replace it. The Task/Status/Source + interface relocation is flagged as
a minimal, pulled-forward Phase 2 prerequisite only.
- Fix stale WorkflowStore reference in service-decomposition.md Q4: the workflow
contract is the domain-owned workflow.Store, which supersedes internal/store's
WorkflowStore.
- Label internal/store as a TRANSITIONAL compatibility location (target
internal/infrastructure/store), matching dependencies-and-contracts.md.
- Call out the current-state outliers explicitly: workflow files/tests still import
internal/store, store.WorkflowStore is still in TaskContext, and the NATS storerpc
path is still wired -- the acceptance criterion is NOT yet met.
- Operationalize the agent gRPC handoff: who binds the in-process State Store
listener (daemon, mirroring serveGateway), Docker-bridge host-gateway address
selection (reuses the embedded-NATS container/network.go resolution), per-task
injection via containerEnv beside NATS_URL, per-incumbence bearer token lifecycle
+ gRPC-interceptor validation -> codes.Unauthenticated, and fail-closed on
non-loopback targets without TLS/token.
- Update beads archie-core-8cda.4.2/.4.5 descriptions to match the ratified scope
(type/interface relocation, RecordDispatch *sql.Tx drop, in-process serving,
agent gRPC handoff).
Address the review of the Phase 2 ratification doc (beads archie-core-8cda.4.1):
1. Domain-dependency contradiction: ownership splits by consumer layer. The
workflow domain owns a consumer-owned workflow.Store interface plus the
Task/Status/Source types (dependency rules #2/#7), so internal/domain/workflow
imports no internal/store — satisfying the Phase 2 acceptance criterion. The
daemon/webui store surfaces stay producer-owned in internal/store. The type +
interface relocation is a Phase 2 prerequisite pulled from migration-decisions
section 4 (store.Task -> WorkflowExecution).
2. Agent gRPC endpoint propagation: explicit boot handoff via STATE_STORE_URL /
STATE_STORE_TOKEN (mirrors the NATS handoff), added to agentworker.Settings and
the Transport.Store facade; per-call timeout + long-lived grpc client.
3. Transport security + error sanitisation: loopback-only + insecure default, agent
authenticates over the container bridge with a per-task bearer token; TLS/mTLS
is an operator choice for non-loopback/k8s. Sanitised public error details in
gRPC; full detail logged server-side only.
Also records RecordDispatch dropping *sql.Tx as a required interface migration
before .4.2. Cross-references updated in the service-decomposition PRD.
Define the authoritative Phase 2 State Store contract boundary and
transport design (beads archie-core-8cda.4.1): one StateStore gRPC
service (40 RPCs) behind the narrow typed store contracts, producer-owned
in internal/store, with a staterpc.Client remote adapter that satisfies
the same interfaces via explicit compile-time assertions. Records method
ownership, DTO/wire representation, error semantics (structured gRPC codes
rehydrated to store sentinels so errors.Is survives), limits, mode/config
shape (presence-based [services.state].target, default local), local-
adapter compatibility, and cutover/deletion rules. Fixes the RecordDispatch
wire blocker (*sql.Tx dropped; production passes nil) and keeps Close()
off the wire. References the new doc from the service-decomposition PRD.
* feat(curator): add embedding client and sampler wave-1 prerequisites
Bead archie-core-1786637500752-415-434e5bb9 (GitHub #436) was closed with
no linked PR/commits, and the surprisal-sampling bead
archie-core-1786637500776-416-600a66e5 (GitHub #437) found none of its
stated prerequisites -- an embedding client, a Sampler interface, or the
four wave-1 cheap sampling strategies -- actually exist. This builds them:
- internal/domain/embedding + internal/infrastructure/embedding: a narrow
typed embedding client, config-driven via a providers table + a role in
[models], mirroring the existing chat model role convention. Degrades to
capability-unavailable (never fails startup or a call) for a missing
role, unknown provider, unsupported class, or unresolved credential.
Wired as an optional capability in bootstrap.go's setupEmbeddings.
- internal/domain/sampling: the Sampler interface plus recency, random,
all, and staleness-proxy strategies, per docs/prds/curator-sampler-wave1.md.
Deterministic given fixed inputs, matching GitHub #437's stated
acceptance criterion for the embedding-backed strategy it will add
behind this same interface (not part of this change).
Table-driven tests throughout; httptest for the embedding client, no live
service. No curator wiring yet -- nothing consumes either capability in
this change.
* no-mistakes(document): docs(architecture): mark curator sampling prereqs as landed
Beads epic archie-core-8cda (with 8 phased child issues per the
resolved extraction order) has been filed for the service-decomposition
work; the PRD's completion-criteria "file a beads epic" line now points
to it. Filing the epic does not authorize implementation -- the PRD's
completion criteria still require the captain's separate explicit
go-ahead.
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
Decision for channel/forge/module action keying in eda-playbook-engine.md:
one durable dedup record per (action, event) pair via a playbook_dispatches
ledger in internal/store, reusing binding_dispatches' INSERT OR IGNORE +
sentinel conventions, with record-before-invoke ordering (at-most-once, the
reverse of the binding loop's ordering, justified by the external-side-effect
harm profile). Key = (playbook_id, playbook_version, event_id, action_id):
event_id is the TaskEnvelope.IdempotencyKey() value already carried on
DispatchInput.TaskID; the key is a fixed structural composition, deliberately
not a CEL expression (event stays dyn; determinism under redelivery is the
binding constraint). Redelivery of a recorded action skips it and continues
the run at debug level, never surfaced as an error. Rows live as long as the
playbook exists, no time-based expiry.
Extends event-sources-and-reactions.md's contract without contradicting it:
PublishUnique's dedup is publish-path only (Nats-Msg-Id + 2-minute
DedupWindow), so consumer redelivery of side-effecting reactions was never
covered; the ledger adds exactly that consumer-side piece, reusing the
TaskEnvelope identity and the typed-family shape. The gap-2 prerequisites
(Playbook ID/Version, DispatchInput.TaskID) already landed in 1e736ef.
Documentation only: no code changes. Keying implementation is separate
follow-up work.
Refs: archie-core-t2db.17
Resolves archie-core-t2db.18. States stop-on-first-failure as the
committed default when a playbook action fails mid-run, grounded in
workflow.Run's three-way stage-failure distinction
(internal/domain/workflow/workflow.go:294, 317-336) and the t2db.9-12
load-collision drop-and-report rule (eda-playbook-engine.md:155-159),
with explicit reasoning why the default does not violate
event-sources-and-reactions.md's producer-only rule. Gap 2 (execution-time
idempotency) remains open. Documentation only; no code changes.
Land Q2 (contract definition mechanism: buf v2 + grpc-go, committed
proto/ tree, buf breaking gate) and Q3 (in-process-by-default
migration mechanism per organisation.md's Process boundaries rule) as
committed file edits, matching the style of already-resolved #1/#4/#5.
Transcribes decisions from a prior scout pass whose edits were
discarded when its worktree was torn down; codebase facts (go.mod,
organisation.md, servicediscovery.go, storerpc.go,
sessionstore_conformance_test.go, api_chat.go) re-verified in this
worktree before writing.
Records the AWX pluggable-architecture and Ansible task-model inspiration
thread verbatim, the verified current-vs-ideal coupling analysis (webui's
in-process HTTP task-action detour, direct gateway internals access,
shared config.Holder), the external service-discovery research
(K8s-native primary + NATS fallback behind one ServiceRegistry seam), and
a Proposed-status PRD recording the target service shape and open
questions before any implementation work is filed.
Phase 1 (archie-core-wp9s) shipped /tasks and the /status queue-depth
aggregate. This documents the remaining ask -- broker connectivity,
provider reachability, channel state, container pool, last poll --
split into cheap wiring (nats.Conn.Status(), container.Pool's
already-tracked active count, a new last-poll timestamp) versus two
real product decisions (live-probe vs last-known-outcome for provider
reachability; the ChannelManager-backed shape for channel state).
Proposed resolution, awaiting Sam's sign-off before Phase 2 lands.
Drafted for archie-core-t2db.17 (idempotency) and .18 (mid-run
failure), following the same evidence-then-sign-off process the CEL
resolution used -- explicitly marked PROPOSED, not decided, pending
Sam's approval. Every claim checked against the real shipped code
during drafting, not assumed, catching two wrong assumptions before
they landed:
- Gap 1 (mid-run failure): proposes reusing workflow.Run's existing
three-way distinction (real error -> stop+report; shutdown
interruption -> not a failure; no-outcome -> definition-bug parked)
rather than inventing new semantics. States the producer-only
boundary and the J3 (condition-error) vs action-failure distinction
explicitly, and states plainly that rollback is not attempted.
- Gap 2 (idempotency): proposes reusing internal/store/bindings.go's
binding_dispatches at-most-once ledger shape (INSERT OR IGNORE +
sentinel error) rather than a new mechanism. Caught two real gaps
while drafting, both stated honestly rather than glossed over: (1)
t2db.15's shipped Playbook struct has no ID/Version to key a ledger
on -- Binding gets its identity from being a DB row, a playbook is
just a loaded file; (2) DispatchInput is a task-label trigger (same
vocabulary as workflow.Route()), NOT a webhook-capture trigger --
captured_events.id, the first instinct, does not apply to what
shipped, and DispatchInput doesn't yet carry any task identity to
key on either. Both are named as concrete prerequisites for whoever
implements this, not silently assumed solved.
Not marked as settled -- the status lines say "awaiting Sam's sign-off"
explicitly, matching the process this doc already used for the CEL
decision.
Self-review caught a real architecture bug in a doc I'd just written:
organisation.md's target structure moves forge/ under
internal/infrastructure/forge/; today's internal/forge is a flat,
unmigrated legacy package. A new internal/domain/eda package depending
on it directly would create a domain-to-infrastructure dependency that
breaks the moment forge migrates -- the same trap the workflow domain
package already solved for itself via its own narrow Forger interface
(CloseIssue/CreatePR/LinkBranch, already exactly the three proposed
operations), already satisfied by the daemon's real forge client today.
Depend on workflow.Forger, not internal/forge.
Investigation for t2db.19. Forge is ready to design: internal/forge.Forge
is already a typed, general-purpose interface with real implementations
(GitHub/Gitea) -- native-Go-by-name, no Yaegi/Module detour needed.
close-issue/link-branch are the safe first-slice operations; create-pr
is deferred (harder idempotency shape than a single-call operation).
Corrects the crew investigation's assumption about Channel: verified
directly that channels.Channel has no Send capability at all (only
Name/Start/Stop + config validation). The only generic outbound
mechanism in the codebase is the single-destination Notify webhook.
Real per-channel sending (telegramApprover) is deliberately
conversation-scoped, not a reusable registry lookup -- a documented
precedent to prevent reading a shared bot handle across chats, which
doesn't transfer to a playbook trigger with no chat turn to anchor to.
Flags two materially different things "Channel action" could mean and
recommends Sam decide before any Channel design work starts, rather
than guessing.
Seventh slice of docs/prds/eda-playbook-engine.md (t2db.15), consuming
t2db.14's expr package: the rich trigger+actions playbook document and
the coordinator that dispatches it.
- internal/domain/eda/playbook: Load(dir) reads *.yaml/*.yml playbook
documents (trigger + exactly one workflow-kind action + optional CEL
when), validating the HARD BOUNDARY at load: a playbook with 2+ actions
or a non-workflow action position is a reported load failure --
multi-action and Module/Channel/Forge actions stay blocked on the
unresolved execution-time gaps. Any malformed YAML, invalid trigger,
missing workflow name, or when compile failure fails the WHOLE load
(reject-at-load, same shape as routing collisions); a missing dir is an
empty store. Trigger reuses the existing workintake kind/label
vocabulary (no second label-matching mechanism). Dispatch matches
trigger, evaluates when (CEL, event context with actions: {} -- no
prior actions in single-action playbooks), and returns the named
workflow from the registry via the same lookup Route() uses -- no
duplicated dispatch logic. when eval error follows the resolved doc's
J3: false, skip, caller logs.
- Store.Dispatch(reg, DispatchInput) is the coordinator entrypoint; input
carries Labels/Kind/Event. Production task intake wiring
(daemon.go/pollNATS/publishTask) is explicitly NOT in this ticket -- a
follow-up; proven end to end via direct tests plus the shipped example
(examples/eda-playbooks/bug-tdd.yaml) loaded and dispatched through the
real path.
- Config: EDAPlaybookDir (eda_playbook_dir), documented as DISTINCT from
PlaybookDirs (flat binding files) -- a reader cannot confuse the two.
- Bootstrap: playbook.Load wired in loadWorkflows, log-and-abort on
failure matching every prior slice; nonexistent dir is an empty store.
- workflow.Route(), workflowForLabels, and the t2db.9-.12 binding loaders
are UNTOUCHED (verified: their full test suites pass, zero regressions).
Also carries small lint cleanups on t2db.13/.14 code flagged by the gate
run: expr Eval no longer type-asserts to ref.Val (Eval returns ref.Val
directly in v0.32.0; cel-go's own Eval wraps evaluation in recover --
verified in source); module decodeLogArgs extracted to clear the nestif
finding. golangci-lint: 0 issues.
Tests (10): single-action round trip, when-false skip, 2-action load
failure, non-workflow position load failure, when compile failure drops
playbook, empty trigger load failure, trigger mismatch skip, no-when
match, one-invalid-file fails whole load, missing dir empty store, plus
the shipped example load+dispatch. task check clean (113/113 UI, exit 0).
CEL (cel.dev/cel-go, pinned v0.32.0) is the single expression mechanism
covering both an action's `when` condition and its `args` values,
evaluated against a read-only context (`event`, `actions.<id>.result`).
Rejects the custom-gate precedent (Yaegi-in-YAML) as inverting the
design's own "YAML is data, Yaegi is logic" premise. Expressions
compile/type-check at playbook load, consistent with the existing
collision-handling reject-at-load rule; generated Module Result schemas
become CEL type declarations so result-path typos reject at load too.
Vetted against expr-lang/expr, gval, govaluate (archived), starlark-go:
CEL wins on structural safety (non-Turing-complete, CostLimit-bounded,
error values not panics, zero security advisories on record) despite
the largest binary cost (+13MB) of the candidates -- accepted because
the alternative is re-implementing CEL's type-checker on a bespoke
grammar for the same guarantee.
Sign-off (Sam, 2026-09-03): all judgment calls (J1-J5) accepted as
proposed, including J2's per-value CEL-literal quoting cost and the
+13MB binary size trade.
Cross-references in module-position.md updated to point at this
resolution instead of citing open question 1 as unresolved.
Fifth slice of docs/prds/eda-playbook-engine.md (t2db.13), implementing
docs/prds/module-position.md's recommended first slice: prove schema ->
go:generate -> Yaegi load -> ModuleRegistry end to end with a
side-effect-free action kind, before any kind with a real external effect
(notify, image-gen, forge writes) -- those stay blocked on the parent
doc's unresolved execution-time gaps (idempotency, mid-run failure).
- internal/domain/eda/module/log: hand-written schema source (Args/Result/
SchemaVersion per module-position.md) + logextract generated symbol
table via go:generate yaegi extract (same tooling as wfextract/
gateextract; scoped to ./internal/domain/eda/module/...).
- internal/domain/eda/module: ModuleRegistry with Register(kind, dir) /
Invoke(ctx, kind, rawArgs). No generic Module interface with an any
payload: each kind is its own tiny package with its own generated
contract; the registry's internal storage is the only place erasure
appears (exactly as inside wfextract). Invoke decodes rawArgs strictly
into the kind's Args before calling the interpreted function -- an
unknown key or wrong type is a reported failure, not a silent
zero-value fill (the 'schema defines the accepted message' rule).
Generalizes gateeval's fixed-signature Yaegi pattern via
yaegiutil.New/Resolve/Safe -- no new interpreter-construction or
panic-recovery mechanism.
- Config: ModuleDir mirroring PluginDir's shape and doc comment
(operator-trusted directory of Yaegi module implementations, one .go
per kind; empty means none loaded). Wired at daemon startup in
bootstrap.go alongside PlaybookDirs loading -- log-and-abort on load
failure (a broken module is a startup failure, not degrade-and-skip).
- examples/modules/log.go: operator-facing example (//go:build ignore,
mirroring examples/secret-engines/) proving the shipped contract.
Tests: registry (register/invoke, unknown kind, missing file, strict
arg-decode shape failures, malformed file, panic recovery), logextract
smoke test, and an example-load test that runs the shipped
examples/modules/log.go through the real registry (not a test fixture).
task check clean (113/113 UI, exit 0).
Playbook YAML action dispatch is deliberately out of this slice: a
minimal 'run this one module action, no chaining' dispatch still requires
naming how args values reach the action (parent doc open question 1 --
data-flow/interpolation syntax) and what the result feeds, both
unresolved by design. Dispatch is a follow-up ticket that consumes this
registry as-is; recorded in module-position.md.
Resolves eda-playbook-engine.md's deferred schema-generation question
for the Module action position. Generalizes the fixed-signature,
resolved-by-name Yaegi pattern already shipped three times (custom
gate, workflow stage plugins, skill plugins) instead of inventing a
new mechanism: one small hand-written schema package per action kind,
go:generate yaegi extract (the same command wfextract/gateextract
already use) producing its symbol table, a ModuleRegistry keyed by
kind rather than a generic any-typed Module interface -- keeping every
position individually compliant with the plugin engine rule instead
of reopening the generic-hook question event-sources-and-reactions.md
already closed.
Closes open question 3 (trust boundary: operator-installed, in-process,
same tier as PluginDir/SecretEngineDir -- not repository-supplied task
code). Recommends a side-effect-free first Module kind, since any kind
with a real external effect is still blocked on the parent doc's
unresolved execution-time gaps (idempotency, mid-run failure).
Fourth slice of docs/prds/eda-playbook-engine.md (t2db.12): the design
doc's pre-merge collision layer, as a gopls-shaped standalone binary
(cmd/archie-playbooks) rather than a Taskfile wrapper -- one binary, one
schema source, mode-based entrypoints.
- cmd/archie-playbooks: subcommand dispatch table from day one (lint now;
an lsp/serve mode later slots in without restructuring), flag parsing
only per organisation.md's cmd/ rule.
- internal/app/archieplaybooks: the lint mode's composition -- collects
directories, calls workflow.LoadPlaybookDirs (the daemon's own loader),
reports findings. No second validation path: the linter agrees with
runtime startup validation by construction.
- Reuses LoadKindWorkflowsYAML/LoadLabelWorkflowsYAML/LoadPlaybookDirs
directly; collision/malformed/invalid bindings exit non-zero.
- Known limitation recorded: findings are file-granular, not
line-granular (yaml.Unmarshal into map discards line numbers; a
file:line diagnostic needs a yaml.Node upgrade, out of scope here).
- Taskfile lint:playbooks target; PRD records the shipped shape and the
file-granular caveat.
Not a bundled-vs-standalone tradeoff: validation logic lives in
internal/domain/workflow with no knowledge of its caller, so archied
and agents already validate in-process, and the standalone cmd/
binary (t2db.12) exists only because CI/editors are out-of-process
consumers by nature. One implementation, two thin consumer shapes.
Third slice of docs/prds/eda-playbook-engine.md (t2db.11), completing open
question 2: instead of a single playbook dir, the config gains a LIST of
directories (playbook_dirs) of *.yaml/*.yml binding files, loaded at
startup as an additional input to workflow_routing_file and
workflow_labels_file, which remain supported unchanged.
- LoadPlaybookDirs(dirs []string) reads every yaml file across the
configured directories and merges kind + arbitrary-label bindings.
- This is what finally exercises the design doc's drop-and-report
collision rule: a single file cannot collide with itself, so the
cross-source path (two files in one dir, two dirs, or a dir plus a
single-file field) had never run in production. A binding key declared
by more than one source is a reported load failure, not a silent
pick-one by source order or version.
- Router-level collision handling: MergeKindWorkflows/MergeLabelWorkflows
fail on a key bound by both the single-file fields and the playbook
dirs, applied at boot.
- Config: PlaybookDirs []string, empty by default (no behavior change for
existing configs). Repo-scoping (typing dirs to Config.Repos owner/name)
is explicitly deferred to a separate later decision per ticket;
org/tenant keying is out of scope.
- PRD: records the list shape per Sam's correction (single dir would
contradict multiple independently-maintained playbook sources).
Second slice of docs/prds/eda-playbook-engine.md, complementing the
kind->workflow layer (WorkflowRoutingFile): the label vocabulary itself
becomes playbook data. A label outside the closed bug/feature/bootstrap
kind set (e.g. security) now dispatches to a registered workflow from an
optional WorkflowLabelsFile, instead of falling through to triage.
- LabelWorkflows + LoadLabelWorkflowsYAML/SetLabelWorkflows in
internal/domain/workflow/routing.go, mirroring the kind layer's
once-at-startup lifecycle. The closed Kind/NATS-subject set is
untouched; the map only extends binding authority for labels the kind
layer does not own.
- Collision rule per design doc: a label owned by the kind set, an empty
label, an empty workflow name, or a duplicate binding is a reported
load failure (dropped, logged, error returned to the caller) -- the
schema defines what is accepted; nothing is silently arbitrated by
schema/version.
- Route() precedence: explicit t.Workflow -> arbitrary-label binding ->
kind binding -> triage -> implement -> default.
- Config: WorkflowLabelsFile (workflow_labels_file), wired at boot
alongside WorkflowRoutingFile; documented in config.example.toml.
- PRD: records this slice's decisions (single-file shape, collision rule,
precedence) so later slices stay anchored.
Records the design for drop-in, data-driven workflows/labels/actions:
YAML playbooks as pure orchestration data over four typed logic
positions (Workflow/Module/Channel/Forge), go:generate-driven schemas,
and load-time collision handling (detect+log+report, no auto-arbitration).
Flags execution-time idempotency and mid-run failure handling as
unresolved gaps to close before Channel/Forge or multi-action playbooks
ship.
RedactPayload only redacted values under a fixed set of sensitive key
names, so a sender that names a secret field something else leaked it into
captured_events.body. Harden the heuristic:
- broaden the key-name markers (compound names like access_token,
client_secret, signing_key, x_api_key, webhook_secret, refresh_token)
and keep gateway.sensitiveParameterKey identical to preserve the
documented agreement between the two redaction paths.
- add value-shape detection: a compact JWT or a PEM private-key block is
redacted under any key, catching the 'sender named the field something
the heuristic misses' gap.
- keep it conservative so schema-by-example field mapping still works:
only unambiguous shapes (eyJ-prefixed JWTs, PEM key blocks) are
redacted by value, never ordinary content.
Best-effort per docs/prds/webhook-intake-security.md point 5.
Refs archie-core-t2db (captured_events.body redaction)
binding.Binding.Secret was persisted as plaintext in the bindings.secret
column, so anyone with the SQLite file saw sender-side HMAC secrets. Seal
it with AES-256-GCM transparently inside the store: encrypt-on-write,
decrypt-on-read.
- cipher: AES-256-GCM, 12-byte random nonce, SHA-256-normalised 32-byte
key resolved through the secret registry ([bindings].encryption_key).
- envelope 'arcie-binding:v1:<fingerprint>:<b64(nonce||ct||tag)>' so a
future cipher/KDF change is a version bump and old rows stay readable.
- rotation via previous_encryption_keys; key identity is the material's
SHA-256 fingerprint embedded in each row.
- legacy plaintext rows written before a key was configured stay readable;
encryption is not retroactive.
- nil cipher (no key configured) preserves the pre-t2db.7 plaintext path.
Deferred: captured_events.body redaction upgrade (separate concern, see
docs/prds/binding-secret-encryption.md).
Refs archie-core-t2db.7
Field metadata (label, section, type, editability, restart-required)
currently lives only in settings.js/config-row.js, disjoint from the
backend's ConfigView allowlist -- adding or changing a runtime-tunable
setting means editing Go and JS by hand in two places with nothing to
catch drift. Repo.ReviewEnabled never reaches the API at all, and
AllowConcurrent/MaxRetries reach it but the repositories card never
renders them.
Decision: keep the existing secret-free ConfigView allowlist as the
safety boundary (no reflection over config.Config, which is already
scheduled for dissolution), and hand-author a descriptor per allowed
field so the frontend renders sections/fields generically instead of
hardcoding rows.
Filed as archie-core-b6ew (epic) with five dependency-ordered children.
Route() had no content-aware fallback: any task with no explicit
workflow and no label match ran the full "implement" pipeline
regardless of content. Confirmed in production (task 6): a
chat-spawned task whose entire content was "this is just a test,
close it when you receive it" burned 669,421 tokens finding that out
the expensive way -- baseline, plan, and build all ran before the
plan stage's own conclusion (nothing needed to change) arrived too
late to save anything.
Adds a "triage" workflow: one cheap, read-only classification turn
(mirrors feasibility.go's decide-tool pattern) that either closes a
task needing no code change immediately, or sets Task.Workflow and
requeues under whichever workflow fits. Route()'s fallback now
prefers a registered "triage" over "implement", unchanged when
triage isn't registered. Labeled tasks are untouched -- a label is
already a free, reliable signal, so triage never fires there.
See docs/prds/dynamic-workflow-triage.md for the full design.
ai-sdk (already a dependency) has a provider-agnostic image.Provider,
core.GenerateImage, and registry.RegisterImage -- three backends
(azure/togetherai/xai) implement it, OpenAI doesn't yet. The next
bead should adapt that instead of a raw HTTP client, mirroring how
curatorLLMRunner adapts runtime.Runtime for chat. Noted the one real
gap (no edit/mask support in ai-sdk's request type) for that bead to
resolve at its own design step.
Adds internal/domain/image (Provider/Registry per the plugin engine
family shape, but no Lifecycle -- a provider is call-scoped, not
long-running) and config.ImageConfig with the same explicit
enabled=true cost gate MinimaxConfig already uses, since a hosted
provider spends real API credits and a local one can saturate the
host GPU. Design doc at docs/prds/image-capability-contract.md per
CLAUDE.md's rule for undesigned capabilities.
No provider implementation lands here -- that's the epic's next
child bead. Also re-links all 7 epic children that were unlinked
from archie-core-1786748942120-1-6636e629 (same orphaning pattern
found in the curator/memory epics earlier this session).
Closes archie-core-1786637499114 (persistent session-memory curator),
the last open item in the curator engine epic.
docs/prds/session-memory-curator.md records the decisions this needed
before coding, most notably: Observation.Identity is the session ID,
not a user or cross-session identity. migration-decisions.md#5's
memory-scope mapping (global/agent-wide/user-wide/agent-user
relationship) is still explicitly open, and this issue has no business
deciding it unilaterally -- session-as-identity is the narrowest
unambiguous scope available today and is forward-compatible with a
later, properly-decided identity resolution.
curator gains ConversationSource (RecentSessions/Messages), gated in
Registry.validateDeclared/filter exactly like Skills/MemoryEngines, plus
Manifest.Conversations. internal/infrastructure/sessioncurator/adapter.go
adapts gateway.SessionStore's full CRUD/branch/search surface down to
those two read-only methods -- curator must not import gateway.Message,
so role derivation (message.From == botUser -> assistant) is copied from
gateway.compressTurnHistory's existing logic, not shared.
sessioncurator.go: Check is a cheap RecentSessions bool. Pass takes each
session's last 40 messages, asks the model for a JSON array of durable
facts (capped at 5/session), and writes each through
Registrar.MemoryEngines.Get(name) -- resolved by name at call time, the
same trust model every other curator capability already uses. A
response that fails to parse degrades to "nothing extracted" for that
session (not a Pass error); a real model-call failure propagates as one
(these were originally conflated into the same swallowed-error path --
caught by TestPassPropagatesALLMFailure, fixed by splitting the model
call from the parse step). PassInput.Since zero resolves to a 24h
lookback, not the beginning of time, so a cold-start first pass can't
read a session's entire history.
Manifest.OnInput = true wakes it via the existing WakeOnPrimaryInput
wiring (events.KindTurnCompleted), which already excludes every
curator-derived event kind from the forwarded set -- the trigger-
accounting invariant (archie-core-1786637489768) needed no new code to
hold here.
Wired into bootstrap.go:setupCurators via a new curatorLLMRunner adapter
(main.go) over the shared ai-sdk runtime, and sessioncurator.NewAdapter
over b.chatSessionStore/b.cfg.BotUser. Model resolves from
b.chatModels.ActiveModel() -- not b.defaultChatIdentity, which names a
task-routing identity, not a model reference (the same source
telegram_setup.go's sendChatTurn already uses for a real chat turn).
task check is green (full Go suite + dashboard tests).
Part of archie-core-1786637499206 (skill curator reference
implementation), under the curator engine epic.
docs/prds/skill-curator.md records the design decision this needed
before coding (no settled design existed for what "review, prune, and
improve skill definitions" means as concrete Pass() behavior): v1 is
mechanical validation and safe whitespace normalization only. No
LLM-authored rewrites of skill content, no automatic deletion -- an
autonomous process editing or deleting a user's skill has no undo path
today, and the epic's actual purpose for this issue is proving the
curator contract holds for a second, structurally different consumer
(files-as-skills vs. observations-as-memory), not solving "what makes a
skill good."
internal/infrastructure/skillcurator/store.go: a curator.SkillStore
implementation over one root's .agents/skills/*/SKILL.md files. Not
internal/skill.Discover directly -- Discover parses every skill up front
and fails the WHOLE call if any one of them doesn't parse (correct for
"load the catalog to boot the daemon", wrong for "review each skill and
report what's wrong with it individually"). This store lists/reads raw
bytes per skill; the curator itself parses each one so one bad skill's
parse failure becomes one Action, not an aborted pass.
internal/infrastructure/skillcurator/skillcurator.go: the CuratorEngine.
Check is a cheap directory-listing bool. Pass classifies each skill into
exactly one of: skill.invalid (frontmatter parse failure, left
untouched), skill.incomplete (missing required name/description,
untouched), skill.normalized (trailing whitespace / missing final
newline, written back), or nothing (already clean). Manifest declares
only Skills -- no Tools, no MemoryEngine -- since nothing here needs
either.
Wired into internal/app/archied/bootstrap.go:setupCurators, rooted the
same way loadWorkflows already resolves skillsBase (cfg.SkillsDir falling
back to cfg.WorkDir) -- one local root, not the full multi-root catalog a
chat turn reads. Skills is now a shared registrar-level service
(alongside Events/MemoryEngines), narrowed out of a curator's view by
registry.filter when undeclared, same as every other capability.
task check is green (full Go suite + dashboard tests); go test ./...
is green.
The web chat is now a persistent slide-over drawer available on any page
(not just the /chat route), and the agent is told which dashboard page the
operator is on so it can point them to the relevant page with a clickable chip.
Backend (gateway):
- Message gains a per-message Page field; the system prompt renders it in
the <env> block (Current page), so Archie knows where the operator is.
- New dashboard_tools.go: a DashboardPages registry plus two web-only tools
— page_index (lists every page and what it shows) and dashboard_navigate
(validates a path, returns {path,label}; rejects unknown/empty paths).
- The web chat sink emits a dedicated navigate stream event when the model
calls dashboard_navigate, independent of ShowToolCalls.
Frontend (webui):
- chat sends currentPage() (location.hash) with each stream request; a
dashboard_navigate result renders as a clickable link chip (href=#/path).
- The chat is mounted once in a slide-over drawer with a topbar launcher,
scrim and Esc-to-close; /chat opens the drawer instead of a second page.
- ui/dist rebuilt (committed) with the bundled changes.
- .gitignore: ignore .impeccable/ (impeccable tool critique snapshots).
New capability (per scope discipline): designed in
docs/prds/dashboard-aware-chat-navigation.md.
Tests: dashboard_tools_test (registry, resolve/reject, web-only gating),
prompt page-line cases, webui Message.Page + navigate-frame capture, and
chat-tools chip tests. task check is green.
task_logs could only read the newest N entries with no way to walk the
rest, so a task's log beyond the page cap was unreachable and the agent
had no honest way to say what was missing. Add logging.Page -- a
byte-offset cursor over the tail-most 8 MiB scan window -- plus
Query.Since/Until time bounds, and wire the gateway's task_logs tool to
expose level/since/until/after_id with cursor and more_available in the
result.
The reader now reads the scan window into a single in-memory buffer and
walks it by exact slice offsets. Offsets into a contiguous buffer are
trivially correct, which an incremental reader over the file is not:
bufio.Reader buffers ahead, so its underlying file position runs one
buffer past its logical position and cannot be used as a cursor. The
same rewrite removes f.Read in favour of f.ReadAt (full-or-error), since
a silent short read would drop the newest entries and still report
success.
The scan window is a documented hard ceiling: for a log over 8 MiB the
oldest bytes are unreachable by any cursor -- that is what Truncated
reports, and send_file remains the path for a complete archive. The
tool description teaches the model the difference between more_available
(more this call did not return) and truncated (older entries the scan
cap hides).
Closes archie-core-vq92 (#625).
Section 1 of the adversarial self-review decision was superseded on
2026-08-24: it assumed daemon-owned in-process execution and a daemon-side
second Pool.Acquire, both removed by archie-core-q15y. It is now re-decided
rather than merely marked dead, so the document is Active again.
The reviewer is an agent.Subagent run from inside the task-scoped worker.
ai-sdk v0.1.22 is already pinned, so this needs no SDK bump, and it adds no
host executor, no second container acquire and no nested container lifecycle
-- the open question the supersession left.
Subagent.Run passes Prompt and never populates Messages, so conversation
isolation is a construction fact rather than prompt text. It does NOT give
filesystem isolation: the sub-agent shares the parent's process, working
directory and credentials, so the .git-stripped snapshot survives from the
original design and is load-bearing, and the reviewer's toolset must be
read-only, rooted at the snapshot, and must exclude the worktreerpc
publication grant -- otherwise a reviewer could push the branch it exists to
block. Run() not Tool(), so the implementer cannot elect to review itself.
Findings return through a closure-captured tool because Run yields only text,
and are read even when Run errors, since captured findings survive.
Also decides fail-closed: a review that fails to RUN parks the task rather
than opening an unreviewed PR. ReviewEnabled is opt-in, so an operator who
turned the gate on wants the gate. This is why h019.2 models "ran and found
nothing" separately from "did not run" -- collapsing them would make an
outage indistinguishable from a clean review.
h019.3 is unblocked and rewritten with the construction rules. h019.4's
description said the task parks as waiting_human, contradicting section 4's
decision of StatusParked; corrected, and the fail-closed rule added.
Autonomous workflows now always hand a complete task to a scoped archie-agent container. Embedded NATS is the default and external NATS changes only broker placement; legacy agent and containers.enabled inputs decode for migration but cannot select an execution path. The daemon flushes task-event subscriptions before a worker can publish, and runtime updates snapshot and restore the configured worker image.
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
archie-core-7d5u.5 (idempotency and poll/webhook parity): the dedup
mechanism (TaskEnvelope.IdempotencyKey + PublishUnique + JetStream
Nats-Msg-Id dedup) and the shared dispatch predicate were already fully
wired by 7d5u.4. What was missing: proof the dedup actually holds, and
observability for a wedged receiver.
- internal/infrastructure/eventbus/nats/dedup_test.go: regression test
proving PublishUnique suppresses a duplicate publish on the same
WorkQueuePolicy stream ARCHIE_TASKS uses in production.
- internal/forge/webhook/receiver.go: Receiver.Status() tracks
authenticated deliveries separately from successful publishes; a GET
to the receiver's listen address (GitHub only ever POSTs) now returns
it as JSON, so a wedged/non-delivering receiver is observable instead
of indistinguishable from an idle repo.
Poll/webhook parity (disabling webhook still polls) needed no code
change: Cycle()/Run() never reference cfg.Forge.Intake anywhere in
internal/daemon, only setupForgeWebhook gates the webhook HTTP server.
archie-core-h019.1 (design: adversarial self-review stage): one-page
decision at docs/prds/adversarial-self-review.md. Reviewer runs in a
second workspace pointed at a .git-stripped snapshot of the reviewed
commit, never the implementer's worktree, so isolation is a property of
what's mounted rather than a prompt instruction. Findings contract is
workflow.ReviewFinding, owned by internal/domain/workflow. A surviving
error-level finding sets StatusParked, not WaitingHuman -- the epic's
informal wording named two different states in the actual machine; this
resolves which.
Both changes reviewed by a fresh subagent with no memory of the
implementation, per CLAUDE.md's adversarial-verification requirement;
zero blocking findings, two minor coverage gaps closed.
Adds t2db.3: schema-by-example field mapping for the no-code playbook epic.
An operator picks a captured event, clicks through its payload to bind named
fields to JSON paths, and previews the resolution against a real event
before saving -- a missing, null, or type-changed path fails loudly (named
by field and event) rather than handing an agent a blank value.
- internal/domain/mapping: Mapping/Field types, path parsing/walking,
Resolve (per-field failure reporting), Validate (incl. duplicate names)
- internal/store/mappings.go: SQLite persistence, mirrors captures.go
- internal/webui/api_mapping.go: CRUD + preview HTTP API
- ui/src/mappings/: dashboard editor (clickable payload tree, live preview)
Design decisions recorded in docs/prds/payload-field-mapping.md. t2db.4
(binding a mapping + matcher to a workflow) is out of scope here.
Wire github.com/nats-io/nats-server/v2 into production startup so a
single-process deployment gets reaction delivery and task distribution
over an in-process server rather than a separate one
(docs/prds/embedded-nats.md).
Meaning change: [nats].url empty used to mean "no NATS, use the SQLite
ClaimNext flow"; it now means "start an embedded NATS server". The old
behaviour is the explicit opt-out [nats].mode = "off". mode resolves from
url when unset (url set -> external, url empty -> embedded), and
agent.mode="nats" / [containers] now require mode="external" (an embedded
loopback server cannot be reached by agent workers or containers).
The daemon owns the embedded server's lifecycle: StartEmbedded before
dialing the client, Shutdown after the client closes (LIFO cleanups). The
embedded server binds loopback on a random port, enables JetStream in a
store dir under the data directory, and leaves its logger nil so a startup
failure surfaces as a Connect error rather than an os.Exit.
Retention becomes a nats.Config field (default WorkQueuePolicy) so a future
reaction stream can fan out under LimitsPolicy instead of the hardcoded
competing-consumer policy that starves a second consumer on overlapping
subjects.
Adds POST /webhooks/capture/{source}: an unauthenticated endpoint (mounted
alongside /healthz, bypassing dashboard token auth by design) that persists
any inbound webhook archie has no binding for yet, so it can be inspected
and mapped later (t2db.2/t2db.3).
docs/prds/event-capture-storage.md decides SQLite over NATS JetStream for
storage: NATS is not unconditionally available until 7d5u.2 ships embedded
support, and capture's single-writer/single-reader access pattern doesn't
need JetStream's fan-out/replay machinery anyway.
internal/store/captures.go adds a captured_events table with prune-on-write
(retention age + row-count cap, both configurable). The HTTP handler applies
webhookguard redaction before persisting, and enforces two more independent
disk-bound mechanics before that: a per-request body-size cap and a
per-remote-address rate limit. New [capture] config section with documented
defaults; all fields optional, capture is on by default.
Adversarial review (manual, after the usual fresh-subagent pass hit a
session API limit mid-run) caught and fixed three real defects before this
landed, each with a regression test confirmed to fail against the pre-fix
code:
- the body-size cap silently no-op'd when CaptureMaxBodyBytes was left
unset, the one bound that limits a single payload's size rather than row
count or write rate
- the rate limiter was keyed by the request's own source path segment --
unregistered and attacker-chosen by design for this endpoint, violating
webhookguard.RateLimiter's documented bounded-key-space assumption and
letting a sender bypass the limit by rotating the segment; rekeyed to
remote address
- the retention prune compared time.RFC3339Nano strings, whose trimmed
variable-width fractional seconds break lexicographic ordering within the
same second; switched to fixed-width RFC3339
Closes archie-core-t2db.1.
archie-core-7d5u.1 -- closes with the decision recorded. Investigated three
mechanisms already in the tree rather than designing from scratch:
- No out-of-process extension boundary exists or is needed; in-process typed
families throughout, and plugins-and-extensions.md does not need amending.
- Reaction delivery reuses internal/eventbus's Publisher/Consumer contract,
but NOT the existing nats.Client unmodified: it hardcodes
jetstream.WorkQueuePolicy (client.go:65), the competing-consumer policy
correct for ARCHIE_TASKS and wrong for fan-out reactions -- two reactions on
overlapping subjects would hit the exact silent-starvation trap CLAUDE.md
already documents. Retention needs to become a Config field; reactions get
a second stream.
- Reactions in these two epics are producer-only (curator's WakeOnPrimaryInput
for cheap wake, or the existing pollNATS->publishTask shape for real work).
No veto/mutate-in-flight capability is being built here.
- No single generic Source interface. workintake.TaskEnvelope already is
forge intake's typed contract and needs no new type -- its IdempotencyKey is
keyed on owner/repo/number, not delivery source, so a webhook-sourced
envelope dedups against a poll-discovered one for free. What's genuinely
shared is a small cross-cutting package for HMAC/rate-limit/capture/approval
mechanics, owned by neither domain.
Retitled and rescoped archie-core-7d5u.3 to match: it was still asking for
'one registry for every kind of inbound caller,' which this decision
explicitly rejects. Removed a false blocking dependency of t2db.1 on 7d5u.4
(GitHub-specific receiver) -- they are siblings sharing the new cross-cutting
package, not a chain -- and linked them instead.
archie-core-t2db.5 -- decides the five non-negotiables its description named
before capture (t2db.1) or the intake family (7d5u.3) exist to build against
it: per-source HMAC auth reusing internal/channels/webhook's existing scheme,
a draft/pending_approval/armed binding state machine modeled on telegram's
dangerous-command approval flow, per-source rate limiting (precedent:
pairing.MaxPendingPerPlatform) that rejects rather than silently drops,
blast-radius bounding via the same gate/diff-cap/sandbox every task already
gets rather than a new relaxation, and 7-day retention with best-effort
redaction reusing gateway/stream.go's sensitiveParameterKey heuristic.
The bead stays open: its code-level acceptance criteria constrain beads that
haven't landed yet and cannot be satisfied by a document alone.
Fold the old docs/ into the VitePress site and rename docs-2 -> docs.
The 12 architecture PRDs become first-class pages at docs/architecture/
with sidebar nav, rather than being buried under an archive path.
Historical material stays at docs/archive/. Rewrites every docs-2 and
docs/prds/architecture reference across the repo. Site builds clean.
Checkpoint commit covering concurrent work across several archie-core-abg
epics to avoid losing in-progress state:
- Memory manager (18.7-18.16): background sync pipeline, prefetch with
timeout, memory-write notification, threat scanning, shutdown drain
with abandoned-write tracking, frozen snapshot pattern. Wired
ScanContent and NotifyMemoryToolWrite into HandleToolCall so scanning
and notification actually apply to writes, not just exist as unused
primitives.
- Fixed a literal unresolved merge-conflict that had been committed to
internal/memory/scrubber/scrubber.go.
- Tool provider/registry, MCP client/transport (stdio, HTTP, SSE),
multimodal results, disclosure.
- Pairing device auth store, rate limiting.
- Gateway compression, session store, tasks.
- Various workflow, plugin host, and agentexec updates.