The rate-limit budget's runtime owner is the channel-settings control-plane
resource, which the Web UI edits and which boot.loadRuntimeConfig layers over
the file before boot.startRateLimiter reads it. [chat.rate_limit] is that
resource's seed, not its authority.
1fe4691b's comments and commit message described the file key as the thing that
documents a webhook's per-route budget, which is the pre-migration framing.
Cite the resource and the layering instead, and say plainly that editing the
file does not change a stored value.
Three defects in one session landed the same way: a map field the file overlay
dropped (xrux), a field missing from the control-plane projection
(parallel_tool_calls), and per-identity forge.intake - accepted TOML that no
consumer reads, and which is not even validated. The entry names the shape,
gives the live example with file:line, and points at
docs/architecture/configuration.md as the config authority.
A webhook set SenderID to its configured route path. The turn path reads that
field through a per-channel resolver which refuses webhook, but the session
curator derives its participant from the raw field with no channel filter, so a
webhook session with exactly one distinct sender and one extracted fact wrote
agent-user memory addressed to User: "/hooks/...". Nothing reads it, and every
pass added more, because the facts are identical strings with no dedup.
docs/prds/memory-engine-unification.md criterion 7 requires that a route path
never become a user identity.
Clear it at the producer rather than teaching every consumer which channels are
lying: a route path is a source, not a person, so it no longer enters SenderID.
It still keys the inbound rate limit, where it is the documented per-source
budget for webhooks (config.RateLimitConfig), so it travels in the new
transport-only Inbound.BudgetKey -- carried on the wire beside Page, never
persisted, and never read as an identity.
pb.Message gains budget_key (field 9) for the same reason page (field 7) exists:
transport context that rides with an inbound message and is deliberately absent
from storedProto's rendering of a canonical record.
The curator needs no change. An empty SenderID already fails closed as an
unidentified user message, which is the path a dashboard session takes, and
sessionParticipant's existing skip covers it; the test name now says so.
Three claims the reconnect loop records were false, and one boundary was
unpinned either way.
The claim that a store which "stalls until its read times out" is charged and
climbs the ladder described a timeout this code does not impose: nextUpdate
waits on the stream and the loop's context, and never on a clock, so a peer
that accepts the stream and then never answers it ends nothing here and is
charged by nothing here -- a probe held the read open for 2s with no retry
logged at all, and the attempt that finally ended was read as healthy. What
ends that stream is the transport's client keepalive on
internal/infrastructure/staterpc/dial.go (Time 10s, Timeout 5s), which the
watch rides because bootstrap takes its control-plane client off the
*staterpc.Client that dial returned. The comment now states that, names the
file, and states this loop's own boundary: it charges only attempts that END.
"The rule reaches every shape" was false for the shape it was argued from. The
window is derived from the delay the attempt would be charged (four times the
minimum at the minimum rung), so a stall at or past it -- 1s and 1.1s were
measured -- is read as a store that answered on every attempt and pins the
ladder at 250ms, against [250ms 500ms 1s 2s] at a 500ms stall. The comment now
says the reach stops at the window, and why that is deliberate rather than a
gap to close: the stalled attempt lasts the window, so those reopens are under
one a second against the four a second the rule exists to charge.
TestKeepWatchTreatsAStallPastTheWindowAsAStoreThatAnswered pins it, at a
five-minimum stall against the four-minimum window (widening the window past
the stall fails it).
The rule's sentence said "outlived a window", which the operator (>=) does not
mean, and a verification pass confirmed >= -> > survived the whole package. The
comparison stays inclusive and the sentence now says so: an attempt still open
when the window closes is a store that answered. The load-bearing inequality is
the other one -- the window is a strictly wider multiple of the charge, so a
store that fails as slowly as the charge is still charged -- and
TestAttemptHealthyAtTheWindowBoundary pins both edges at every rung: equal to
the window is healthy, a nanosecond inside it is charged.
TestKeepWatchChargesAnAttemptThatEndedWithNothing's window case now stands open
past the window rather than for exactly it: its sleep starts before the loop's
attempt does, so an attempt measured against it could be judged a hair early,
and the exact edge is pinned without a clock by the new unit test.
TestBothWatchStreamsRunThroughTheReconnectLoop asserted only that the name
keepWatch appeared, so a launch site handing the loop a wait of its own was
invisible -- the persona site's wait has no behavioural test at all. It now
reads the wait argument of each keepWatch call (changing it fails the test, and
nothing else in the package does).
No upper bound was added for the wait itself: the injection seam replaces the
wait in every test that drives the loop, so a bound on waitFor's own duration
would have to be a wall-clock one, and the correct code overshot a 25ms timer
to 65ms on a saturated 16-CPU box (169/200 samples over 1.5x, 8/200 over 2x),
which is exactly the flake a bound tight enough to catch the 2x mutation buys.
The wait doc states what the injected wait can and cannot show.
The settings and persona watches were launched once, at boot, and ended on the
first stream failure: the goroutine returned, so every later stored change for
that kind was silently never applied and the process recovered only by
restarting (archie-core-yrmr).
Both watches now run one keepWatch goroutine per kind, which re-establishes the
stream in place after a bounded backoff, resuming after the last version an
update carried and stopping on ctx cancellation -- which the loop watches for
itself, so a stream that never closes cannot hold up shutdown. A first open that
fails still fails the boot that asked for it.
The reconnect delay is decided by the attempt that just ended, and only by that
(attemptHealthy, retryDelay): an attempt is healthy iff it delivered an update
this watch had not handled, or it outlived a window that is a strict multiple of
the delay it would otherwise be charged. A healthy attempt returns the delay to
the minimum, a failed one doubles it up to the cap, and the window moves with
the ladder instead of sitting at its floor. The acceptance of a reopen is the
wrong unit: gRPC takes the stream before the handler reads anything, and
controlplane.Server.Watch answers a store it cannot read by ending the stream,
so during the outage this reconnect exists for every reopen is accepted and
every stream ends carrying nothing -- measured with the real server and client
over bufconn, 4.0 watch RPCs a second across 2s of store outage with retry_in
pinned at 250ms when that acceptance is read as health. A window fixed at the
minimum is the same defect for the likelier shape of outage: a store that has
stopped answering does not error at once, it stalls until its read times out,
and an attempt that takes longer to fail than the minimum outlives that window
on every attempt -- measured in the end-to-end outage test before the change, a
300ms stall left the ladder flat at 250ms/250ms/250ms/250ms while the store
never answered a read. Both shapes are charged now, and the outage test drives
three of them (an immediate error, a 300ms stall, and a stall as long as the
delay that rung would be charged) up four rungs of the ladder, each logged retry
held to the delay actually waited.
The wait is injected (wait, production passes waitFor), because the ladder was
only ever asserted through the retry_in attribute: a loop that logged one delay
and waited another would have reported a growing ladder and reopened on a
schedule of its own. TestKeepWatchWaitsTheDelayItCharged reads the delays back
as values and asserts the ladder, and the outage test's gaps catch a wait that
returns without waiting. The retargeted reset test is joined by
TestKeepWatchChargesAReopenTheStoreMerelyAccepted: a delay reset by the
acceptance of a reopen charges every attempt of the outage from the minimum, and
nothing pinned that before.
The watch returns on cancellation without draining what the producer left for
it, so controlplane.Client's producers stop on the same context instead of
parking on a send no reader will ever take -- reproduced 40/40 times as a
goroutine still blocked on an undelivered send after keepWatch returned, one
goroutine and one channel leaked per kind at every exit. That cross-package
coupling is load-bearing in both directions and is stated as such; the two
producers now share the one stream reader that carries it.
A Recv failure is the store being unreachable, not a document this process
refused, and docs/prds/control-plane-apply-status.md ("What can be reported, and
what cannot") is what forbids reporting it as a failure of the process it
affected. The record is now written only for an update that carries a version,
which is what controlplane.Client emits for a document that would not validate
and never for a transport failure. Reporting the transport failure would also
overwrite the text of a standing refusal, replacing the one message that says
what to fix with "control-plane unavailable". The record is not un-clearable of
itself -- Report overwrites it on the next applied version -- but a store that
never changes again leaves the one stamp standing, and the process must not
write it at all. That emission is the whole of the failure path and now has the
test it lacked, from the producer's side:
TestWatchClientsDeliverARecvFailureAsAVersionlessUpdate pins that a Recv failure
is handed over carrying the mapped error and no version, and that a stream the
server closed and one that ended with the watch's context are not updates at all.
The refused version is deliberately not re-offered: the resume point moves past
it, so an edit only a newer build can decode cannot become a hot loop. Both ends
of that contract are pinned now: the watch stub filters its scripted stream by
AfterVersion as the server does, and controlplane.Server.Watch's own filter has
a test for the first time. The apply-status reporter's 30s re-stamp is unchanged
(TestRestampKeepsAReportedFailure).
The build task's only sources were **/*.go, go.mod and go.sum, while
ui/embed.go embeds the committed bundle with //go:embed all:dist. Task's
up-to-date check is a content checksum over files, so a change to ui/dist
-- including the one `task ui` writes in place -- left the task "up to
date" and the previously embedded dashboard in every binary.
Measured on this worktree, appending a byte to ui/dist/index.html:
git status M ui/dist/index.html
task build task: Task "build" is up to date
grep <marker> bin/archied 0 <- stale asset shipped
With ui/dist/** listed, the same probe rebuilds all six binaries and the
marker is present; a nested asset (ui/dist/assets/index.css) does too.
Reverting the probe rebuilds again, and a second build with nothing
changed still reports "up to date", so the content cache is intact.
touch-only changes do not retrigger, confirming the check is on content
rather than mtime.
deployment_contract_test.go gains a bounded test over the build task's
sources block, the shape the neighbouring fmt assertion already uses.
The version stamps in the same task stay out of reach of a file checksum:
git describe moves with HEAD while no file content does, and the git
metadata is not addressable from a worktree, where .git is a file. That
needs a status/stamp design decision, so it is split to archie-core-pdq2.
setupGatewayChat is the only production constructor of a Router and it runs
in the Gateway, which owns no container pool, no poll loop and no channel
manager. The producers wired to those three built report sections no reply
could carry, and /status' own description of itself claimed all five fields.
The health source now carries the broker connection and the chat-model
outcome, the two facts the answering process holds.
The tests for the deleted producers go with them, replaced by one that pins
the served shape with a pool and a daemon present in the boot.
Corrects docs/prds/status-health-surface.md, where the intro, the Phase 2
summary and the rendering bullet all claimed the five-line surface.
mcpServerSettings mirrors config.MCPServer field by field, and
runtimeToolConfigFrom rebuilds cfg.Tools from that mirror wholesale. The
parallel-tool-calls flag had no home in it, so on any deployment with a
control plane the file document's parallel_tool_calls = true was dropped and
every MCP server stayed serialized. The field is now carried through both
conversions -- the seed that writes the resource and the layering that reads
it back -- and the resource accepts it, so an operator can set it there too.
The field-by-field mirror is what let the field fall through, so the test
populates every MCPServer field by reflection and requires each one to be
projected and handed back, with an explicit allowlist for the fields the
projection deliberately leaves to the file document.
MCPServer gains ParallelToolCalls (default false). The client's
per-server callMu is now nil for a server that opted in, so CallTool
stops serializing that server's calls and the caller's own concurrency
decides; every other server keeps exactly one tools/call in flight.
The flag is threaded from config through the provider constructor -
the only place MCPServer configs become Client instances - into
NewClient, in both the daemon and the agent worker.
The keepalive agreement was guarded only by the 50s stream case, whose control
tears down 41s in: a revert of the server policy cost the whole window to
detect. stateStoreServerOpts's options are now checked in microseconds -- a
loopback listener installs exactly one option, the enforcement policy, and that
policy is 5s with PermitWithoutStream false, strictly below the dialer's ping
interval and at least a factor of two under it.
staterpc exports ServerKeepalivePolicy and ClientKeepaliveParams so both halves
of the agreement are read from the package that owns them rather than restated
in the test: changing either side alone (the floor, or the dialer's interval)
fails it. ServerKeepaliveOption is now the policy wrapped in an option.
The stream pair is unchanged and still proves the tear-down and the survival end
to end; its two cases already run concurrently (t.Parallel, 214f9bd9), so the
package pays the longer case (50s) rather than the sum (91s).
The dialer's keepalive pings a server running grpc's default enforcement policy (MinTime 5m), which counts every ping arriving less than five minutes after the previous one as a strike and answers the third with GOAWAY too_many_pings. The dialer pings every 10s, so a connection that carries no write for ~40s -- a watch while the watched resource's version is unchanged -- was torn down, where before the client keepalive it stayed up.
ServerKeepaliveOption returns MinTime 5s, half the client's ping interval, so every ping lands 5s clear of a strike; stateStoreServerOpts installs it on the loopback and token listeners alike.
The curator derived the agent half of its agent-user scope from the
session record's BotUser, which is the router's session-resolution match
key: empty when no [[identities]] are configured (the shipped default),
identities[0].Name when they are. The engine validates the scope and
refuses an empty agent, so every pass in a single-identity deployment
failed on its first session and wrote nothing, forever; in a
multi-identity deployment the write landed at an agent no chat turn
reads.
A turn addresses its own memory with cfg.BotUser, so the composition now
hands the curator that same value through NewAdapter. The record's
BotUser stays exactly what it is, the match key resolution compares
against.
A session the pass cannot address is now skipped with a memory.skipped
action naming why, at the point the scope is produced, instead of
aborting the pass and dropping the actions already collected: one
session with no address must not starve the rest. Real infrastructure
failures -- an unregistered engine, an unreadable store -- remain pass
errors.
One go fix pass applies each fixer once, to the parse it started with, so a
rewrite made by one fixer is only fixed up by the next pass: stditerators
turns `for i := range t.NumField()` into a field iterator that copies its
loop variable, and forvar deletes that copy on the following pass. A single
pass therefore left the copy behind for copyloopvar to reject, and no
fmt-then-lint sequence healed it. task fmt now applies passes until a dry run
reports nothing left to fix, and fails loudly instead of handing the lint
step a tree it must reject.
deployment_contract_test.go pins the fixpoint proof rather than the old
single-invocation literal.
Four binaries derived the default with os.UserConfigDir while the daemon
resolved it from its own config home. The two agree on Linux with an
absolute XDG_CONFIG_HOME and diverge everywhere else, so one deployment
booted from a different config file depending on which binary started.
Move the rule into the package that owns configuration resolution
(internal/infrastructure/configuration, which already owns xdgDataHome and
DefaultWorkDir) as DefaultConfigDir/DefaultConfigPath, and have every
-config default take it from there: archie-gateway, archie-messaging,
archie-state-store (serve and recovery) and archie-ui, plus the daemon's
own parseArgs, setup generator and drain-marker path.
Fence the claim at the source, since the resolved value is only observable
by starting a process: no source under cmd/ or internal/app/archied may
call os.UserConfigDir, and every -config registration must default to the
helper.
Three of its decisions have no implementation: J1 action ids, J2 args as CEL,
and generated Result schemas declared to CEL for field-level result typing.
That is why it stays Approved rather than Finalised, so the header names the
beads that close the gap.
PutResource replayed a prior write by request_id alone: the dedup
lookup and the schema's own UNIQUE constraint were both global, while
the resource API -- and every production caller -- is per kind. A
request ID reused across kinds therefore either replayed the first
kind's resource with the second kind's name attached to the call, or,
once the lookup was fixed, failed the insert on
UNIQUE(resource_history.request_id). The step-vocabulary agreement test
hit exactly this shape before its fixture was fixed to per-step IDs,
which is what elevated the hazard from latent to demonstrated
(archie-core-fcvd).
The dedup key is now (kind, request ID) in both places: resourceByRequest
filters on the kind, and schema v4 rebuilds resource_history with a
UNIQUE(kind, request_id) index instead of the column constraint.
Existing data satisfies the new shape -- a global constraint is strictly
stricter than the per-kind one -- so the rebuild copies every row and
drops nothing. Detection is the auto-index the old column constraint
created, so the step is a no-op on a database that already has the v4
shape and on one created fresh.
The playbook engine PRD had grown to 972 lines against CLAUDE.md's one-page
bar, because four documents were living in it.
The expression decision was the worst of it: an H2 nested inside a numbered
list item, 270 lines with twelve H3s and its own Status and Date fields. That
broke the outline and gave the document a second, non-canonical status. It is
now docs/prds/playbook-expression-syntax.md, named flat and linked by a Parent
field the way module-position.md already relates to this epic. The parent
keeps the decision and a pointer.
Repetition, against the state-each-fact-once rule: reject-at-load was stated
ten times, the plugin engine rule nine, producer-only seven. The producer-only
argument ran 32 lines and closed by restating itself; it is eight lines and
stated once. One heading read "confirmed non-issue (stated once)" above the
third statement of the point.
Open question 3 was closed by module-position.md and still listed here as
open. The section holding four decisions and one note is no longer called
"Open questions", and the note is a standing constraint, which is what it is.
Also removed: two list summaries written before their lists, ten dated
attributions, two progress claims, three orphan line citations, and the
undefined { Call: ... } notation.
Five findings in other PRDs came from tightening the guard, not from this
split.
The playbook engine PRD passed the guard while carrying a whole second PRD
inside a list item. Three separate blind spots, each with a test:
A second **Status:** field was invisible twice over. lintStatus returned on
the first status it found, and its pattern was anchored at ^, so an indented
one inside a list item was never read. It now checks every status line and
reports a second one as what it is: a nested document that belongs in its
own file.
dated-status matched verb-then-date only, so "(2026-08-22, decided)" and
"surveyed 2026-09-03" read as clean. It matches either order now, plus the
words that turn a date into provenance for a finding rather than a fact.
Rules were line-scoped, so a citation broken across a line wrap was two
innocent lines: "...Dedup section, lines" then "155-159, the rule". Each rule
now also tests the line joined to its successor, reporting the line the match
starts on.
Three defects the first sweep left behind, each caught by a rule the guard
did not yet have.
Statuses. RULES.md gives three; the tree had 25 values, and the free-text ones
were carrying progress. "Implemented (wiring landed)" and "Ratified (rev. 2e).
Rev. 2e adds the OpenTaskByPR lookup. Rev. 2d revised section 10..." were both
changelogs living in the one field the other rules exempt. Every PRD now
declares Draft, Approved or Finalised and nothing else; five that declared
nothing now declare one. A revision number moves to **Revision:**, a pointer
to the authoritative document to **Authority:**, and what a past revision
changed to git. CLAUDE.md's citation of state-store-contract rev. 2e still
resolves.
Stale cross-references. "Supersedes the 2026-09-03 open question", "this
bullet previously said", "an earlier draft of this document claimed" all point
at text the reader cannot see, because it was replaced. A supersede marker
survives only when it names a file, a section or a symbol that can be found.
Temporal hedges. 100 uses of today, currently and right now across 31 files.
Each dates its sentence to the day it was written: "[notify] today is one
webhook URL, read in one place" stops being true the moment this design lands,
while still reading as the premise it rests on.
The header-field exemption was also too wide. "**Date:** 2026-09-03, Channel
resolved 2026-09-22" smuggled a dated status into the one line the rules did
not read; the exemption is now a Date holding only a date and a Status holding
only one of the three.
The scheduling-policy resource carried the dispatch trigger but not the
trigger-matching label, so validateScheduling could judge only half of
the pairing the file layer's validateDispatch enforces on the effective
document (a label-requiring trigger with no label matches every open
issue -- the GH#445 guard). The write path therefore stored trigger
label/either with no label whenever the file's own label was unset, and
boot then refused the layered document: a value the store blessed,
archied unstartable, the operator left with the offline rollback as the
only remedy (archie-core-7pyj).
The resource now carries the label: the seed copies the file document's
label, RuntimeConfig layers it with the trigger (a policy stored before
the field existed decodes nil and leaves the file's label in force, the
same absence-means-inherited shape a kind with no stored value has),
and validateScheduling refuses a label-requiring trigger with no label,
the pairing judged on the data the store actually holds. Once the store
carries a label it owns it, so a file that later loses the label can no
longer break a stored policy that depends on one. The parity test
gains both directions, and a legacy policy without the field is pinned
as an absence, not an erasure.
golangci-lint ran ./... from the repository root, which stops at the root
module, so the tools module was never linted and had drifted unseen. task
check already closes this hole for tests with test:tools; lint now has the
same shape in lint:tools, which runs in the tools directory rather than
passing a path prefix that only rewrites output paths.
Fixes the six issues that had accumulated behind it:
stripJSComments was 21 cyclomatic complexity against a cap of 15, with the
comment-skipping and literal-copying loops inlined into one scanner. It is
four named functions now, and the De Morgan warning went with the rewrite. It
had no test despite deciding whether a commented-out call reads as a live
consumer, so it has one: literals that look like comments, comment markers
inside strings, escapes, and unterminated literals and block comments.
docsgen's missing-artifact error ended in punctuation, and three test type
assertions were unchecked, which panic rather than fail when the generated
artifact does not have the shape the case assumes.
183 findings across 28 of 46 PRDs, of which 150 were file:line citations. Two
were already hard-broken: one named a file that no longer exists, one pointed
past the end of its file. The rest resolved to a line that still exists,
which is the worse failure -- a citation that silently drifted onto unrelated
code still reads as authoritative.
Citations now name the file and the symbol. Dated status markers become the
decision they carried ("Status 2026-09-05: resolved" -> "Resolved"), since
whether a question is answered is design information and when it was answered
is not. Progress inventories become scope statements: mcp-tool-completion now
fences what is out of scope rather than reporting what shipped.
status-health-surface drops phase-completion from its headings and keeps the
rule that earned them, that a reported field must have a truthful source.
eda-playbook-engine.md was the worst file and my own t2db.23 status block was
part of why; it is rewritten as the resolution it records.
A PRD states a decision, a requirement, a constraint or evidence. Those
survive the code moving underneath them. A line number, a progress report or
a dated status claim does not: it is true on the day it is written and
silently wrong a month later, while still reading as authoritative.
prdlint rejects five shapes: file:line citations (including a document citing
its own line numbers), dated status claims, progress reports on the work
being designed, phase-completion narration, and records of how a finding was
produced. Each finding names the file, line, rule and fix.
The rules target decay, not subject matter. A PRD may describe a phased plan,
name a dependency that exists, and carry its own Status and Date header
fields; what it may not do is assert the state of the tree on the day it was
written. Fenced blocks are skipped, since a listing may specify a line
reference as its literal content.
Wired into task check next to docs:check. A sweep alone would not stop the
class recurring.
migrateLegacySchedules replaced the whole legacy cron/jobs.json list into
the schedules resource, whose validator refuses every job whose kind is
not 'workflow' -- and chat is the legacy default (an empty kind on
schema-v1 jobs means chat). One chat job therefore failed the replace,
the migration returned the error and RunStateStore aborted before the
server ever listened. The failure was permanent: the seed had already
landed at version 1, so every restart re-ran the migration and refused
again, until an operator renamed cron/jobs.json by hand -- a remedy no
recovery doc mentions (archie-core-q6vw).
The migration now moves only the jobs the schedules resource can hold
and reports the rest instead of refusing them: chat and unrecognised
kinds were already dead deliveries in the old engine (the delivery
router maps KindWorkflow alone). An unreadable or corrupt sidecar is
logged and left in place for the next start, the way ImportConfig skips
a seed the resource validator refuses -- a value this process cannot
use must not stop the one process every other process dials. A
successful replace is marked by renaming the file to jobs.json.migrated
(the skipped jobs stay readable in it); until the rename lands, the
replaying request ID keeps the replace idempotent across restarts.
Headings carried scheduling status the Blocked-on line already states, and
two paragraphs narrated how the document was written rather than what was
decided. Both violate the PRD rule that every sentence adds a decision,
requirement, constraint, evidence or verification step.
Notification is a contract in internal/domain/messaging, not a channel type.
A channel implementation satisfies it by importing its own parent, which it
already does, so channels.Channel gains no notification method and no import
and the dependency stays channels -> messaging. A notification has no
lifecycle, no inbound direction and no config schema, so modelling it as a
channel would force Start/Stop/ConfigSchema onto something with no meaning
for them.
Destinations are configured, never named by a playbook: [notify] grows from
one webhook URL into named destinations, each binding a name to a channel
already registered in the Messaging composition plus that channel's address.
An action names a destination. Playbook triggers are reachable from a public
intake surface, so an action that could address an arbitrary chat would be a
new exfiltration surface.
Host access needs no new process surface. archie-messaging dials out only and
deliberately has no NATS connection, so nothing can push to it today.
Delivery travels the Gateway connection every process already holds, as a new
server-streaming RPC on ChatService; the per-turn Stream RPC is not reused.
The 2026-09-03 section presenting two options is superseded. Its claim that no
send capability exists anywhere was too strong: telegram.SendPendingReport
already sends unsolicited to a stored chat id, with no interface for it.
Closes archie-core-t2db.19
Project skill for running a bounded agent fleet on this repository. It stays
thin on purpose: the method -- doctrine, pipeline shape, the interview, the
pi-subagents ceilings and traps, scripts/, templates/ -- is the user-scope
high-throughput-programming skill, which is normative for method. This skill
carries only archie's facts and is normative for those.
What it carries: the seven invariants every lane call must repeat (fail-closed
control plane, the webui/workflow deletion gate, single ownership of
<db_path>-tasks.sqlite, no .beads/ commits, generated assets as LAW, scratch in
/tmp because CommitAll stages untracked files, and the cmd -> app -> {domain,
infrastructure} direction); the canonical gate-evidence.sh invocation and why it
must run the module build and the linter, not just a scoped test; the scoped
test-command shape for a lane plus the fact that task check is the serial full
gate the merge train owns; the agent-discovery trap (a lane worktree is a
sibling, so .pi/agents must be symlinked into it or children die with "Unknown
agent: cp-writer"); bd conventions; and that .pi/control-plane-run/ is
disposable run evidence, not documentation.
Also moves the September 2026 control-plane run out of the user-scope skill's
references/worked-example/ and into references/worked-example/ here, so exactly
one copy exists and it is committed. The user-scope skill's References table and
normative-copy paragraph now point at this repository, and note that a run in
another repository uses templates/ plus its own filled-in briefs. The moved
README gains the worktree agent-symlink fix and loses its stale "everything here
is gitignored" claim; its paths to scripts/ and templates/ now name the
user-scope skill explicitly. The scripts warning is unchanged: they are the
snapshot's and must not be copied.
A loaded playbook now decides a real task's workflow. The wiring point is the
daemon's definition pin (Daemon.resolveWorkflowID, consumed by
pinWorkflowFromCollection), not pollNATS/publishTask as the design doc
assumed: routing moved out of the agent process when workflow definitions
became database rows pinned before dispatch, so the pin is the one place a
task's workflow is chosen.
Precedence: an explicit Task.Workflow (the waiting_human -> approved requeue)
still wins, then a matching playbook, then workflow.ResolveWorkflowID's
existing label/kind bindings. A playbook naming an undefined workflow is a
reported failure that parks the task rather than a silent fall-through to the
binding's choice -- the operator bound that trigger deliberately.
Store.Dispatch returns the selected workflow name plus the matching playbook's
id and version, instead of a compiled workflow.Workflow: the production caller
decides against the active definition collection, and the id/version are the
first two components of the resolved gap-2 idempotency key. The playbook
package no longer imports internal/domain/workflow.
The shipped example's `when` read event.priority, a field no intake supplies,
so it could never have fired once wired. Both examples now read the event
surface playbookInput really builds, and label-trigger.yaml documents the axis
an instance defines for itself: trigger.kind is the closed workintake
vocabulary because a kind addresses a NATS subject, while trigger.labels is
free-form and binds an operator's own taxonomy to a workflow.
Module/Channel/Forge positions stay rejected at load; they need the
playbook_dispatches ledger first (archie-core-t2db.24).
Closes archie-core-t2db.23
Two agents independently produced archie-core-eju6. Keeps the wider test (queried
plugin_dir as well as the watched budgets, the SIGHUP reload, a task-scoped grant
denied on the apply-status RPCs, real gRPC over bufconn behind TaskGrants) and
carries over the fix for the staleness leg both versions got wrong: the aged
fixture derived from the constant the reader honours, so it now uses the
documented 90s window instead.
Takes the wider of the two independently-written archie-core-eju6 tests: it also
covers the queried half of boot.runtimeConfig (plugin_dir, not just the watched
budgets), the SIGHUP reload re-applying the database layer, a task-scoped grant
denied on PutApplyStatus/ListApplyStatus, and a real gRPC transport over bufconn
behind TaskGrants.
Its staleness leg had the same defect the narrower test did: the aged fixture was
derived from applystatus.StaleAfter, the constant the reader honours, so the
fixture moved with any mutation and the assertion held for a window measured in
hours. The aged row now uses a documented-window literal (90s, the operator-facing
contract) so the reader is constrained from both sides.
Resolves archie-core-eju6. One end-to-end test drives a settings change through
the dashboard's control-plane API, observes it land in the audit trail as an
ordinary edit attributed to identity.SystemID, restarts the consuming process
against the same store, and asserts it came back on the stored value rather than
the file value, with the apply-status row reporting the applied version and an
aged row reading as 'not reporting' rather than current.
The clause-5 staleness reading aged its fixture by applystatus.StaleAfter, so
the fixture moved with the very constant the reader is supposed to honour and
the reading agreed with the reader by construction: mutating StaleAfter to
1000h left the test passing, which is a staleness assertion that never went
stale.
Age the row by the documented 90-second window (three 30-second re-stamps,
docs/prds/control-plane-apply-status.md) held as a fixture literal instead, so
a window that stopped expiring fails the reading, and say in the fresh half's
assertion that it is the other end of that age pair.
archie-core-eju6.
archie-core-eju6: docs/prds/runtime-control-plane.md ("First implementation")
lists an end-to-end restart test as a deliverable and there was none.
One test drives the whole path against a real State Store: the dashboard's own
HTTP adapter writes a task-limit change (attributed by the adapter to the System
identity), the audit trail records it as an ordinary edit above the migration
seed, the process that consumes it stops and starts again against the same
store and the same file document, and the restarted process is running the
stored limits rather than the file's. The apply-status row for that process
reports the version it applied, and the settings page reads an aged record as
not reporting instead of current.
task.WorkflowDefinitionVersion/Digest/YAML had no pb.Task field and no
taskProto/taskValue mapping, so over gRPC the pin was never persisted and
any pin already in the row was zeroed by the same Update. The production
daemon's Store is the gRPC client, so the digest guard in pinWorkflowDefinition
was dead there: WorkflowDefinitionYAML was never non-empty for a task read
back over the wire, and each retry re-resolved from the active control-plane
resource instead of the version it started with
(docs/prds/runtime-control-plane.md:95).
The SQL store already wrote and read all three columns; only the contract
dropped them. Adds proto fields 28-30 and the mapping both ways.
Field parity was untied, which is why this was invisible: the task still
crossed, just with those fields silently zeroed. TestTaskProtoCarriesEveryDomainField
fills every domain field and requires the round trip to return it, so the
next forgotten field fails by name. The conformance battery now carries the
pin through Update and back out of TaskByID, failing in grpc mode only.
Resolves archie-core-xrux and archie-core-i3qm. The -config-overlay file path
(and the overlay-directory path) now fold a partial map-valued entry instead of
replacing it wholesale, so an overlay that sets one sibling key no longer clears
services.state.target_token. The file load path stops judging settings the
control plane owns, so a stale TOML value in a database-owned field can no
longer stop archie-state-store starting; boot.runtimeConfig remains the one
layering call and still validates the merged document. A kind with no stored
value leaves the file's value in force.
The fix-closure review found the previous pass moved the failure instead of
ending it. A seed the resource validator refuses leaves the kind ABSENT, and
RuntimeConfig treated an absent kind as fatal, so archied and the Gateway
refused to start with "control-plane resource not found" -- a database-named
error that editing config.toml cannot clear. Reproduced before the fix: every
kind in the new table failed that way.
Two changes, one on each side of the seam:
- Client.RuntimeConfig (and RuntimeChatConfig) now treat a kind with no stored
value as "the file document's value stays in effect", which is what the
skip-and-report comment already claimed. The version map records no version
for such a kind, so nothing is reported as applied. A value the file layer
also refuses still fails closed, at configuration.Validate, with a message
naming the file's own field -- and the file is the fix.
- Which repository lists are valid can no longer differ between the layers.
configuration.ValidateRepositories is now the single rule set (owner and name
present, no duplicates, compilable test glob) and the repository-policies
resource validator delegates to it. The file layer had no duplicate rule
while the resource layer did, so a duplicate [[repos]] entry passed the loader
and then vanished into a skipped seed.
Also, from the same review: the scheduling validator's "0s" carve-out is gone,
so a zero poll interval is refused at write time as configuration.Validate
refuses it (there is no defaulting behind a stored value); applyOverlayFile's
doc now says which overlay forms fold (a conf.d extra does not, and nothing
reads cfg.Extra); and plugin-settings' no-op validator is a named function with
the reason it accepts everything.
Review of the two previous commits found the split was incomplete in both
directions: the checks moved off the file load path were still enforced
anywhere the value could enter the system, and the fold reached only one of the
three overlay paths.
A stale database-owned TOML value could still stop archie-state-store. The
process seeds every control-plane resource from the same file config on a fresh
database, and the seed runs the resource's own validator; three of the five
values the loader now accepts aborted the boot there (verified by booting
RunStateStore before the fix). A seed that does not validate is now SKIPPED and
reported instead of fatal: the resource stays absent, so the file's value is the
one in effect and editing config.toml is still the fix -- the daemon and the
Gateway are where it fails closed, at boot.runtimeConfig.
The resource validators were weaker than configuration.Validate for exactly the
fields the split reclassified, so a value the file path used to refuse at load
became storable in the database, and then un-fixable by editing config.toml.
The trigger enum and the test-glob rule now come from the configuration package
rather than being restated, the container image is required there as it is in the
file, and a table test ties the two layers together value by value.
The directory form of -config-overlay kept the archie-core-e2e2 failure: it
decoded each overlay file straight into the populated config, so a map-valued
entry it only partly addressed still cleared the fields it did not name. Both
the overlay layer's main file and its feature files now fold, which is what
Loader.Dir's documented parity with Loader.Overlay has always claimed.
The overlay file's key report is derived from the apply's own view of the file as
well as the file decode's, because TOML matches keys case-insensitively and the
fold's yaml decode does not: a key only the file decode matched was consumed by
the report and dropped by the apply, a silent no-op that worked in the base file.
Tests: the state-store boot is asserted against the real process, the directory
overlay and the case-mismatched key each have one, and ImportConfig pins that a
refused seed is not stored.
Loader.finalize ran the whole validation over the file document, so a stale
TOML value in a setting the control plane owns failed startup -- including
archie-state-store, which resolves its file config and opens the store without
ever layering a database resource over it. docs/prds/runtime-control-plane.md,
"Bootstrap, migration, and recovery": after migration, settings in TOML are
ignored and cannot block State Store startup.
The checks over the settings a control plane resource is seeded from and
replaces wholesale now run only on the effective document: providers
(provider-settings), poll_interval and dispatch (scheduling-policy), containers
(container-runtime-policies) and the repository lists (repository-policies).
They moved from validate to validateDatabaseOwned; validateBootstrap keeps the
checks over settings no resource carries (forge intake, identities, nats,
memory, image, capture). File sources pass validateBootstrap and
Loader.ApplyOverlay, which layers overrides onto an effective document, passes
Validate.
Nothing stops being judged: the daemon and the Gateway both run the full
Validate in boot.runtimeConfig, the one place database resources are layered
in, before any capability starts -- so a stored value that will not validate
still stops the process rather than being published. Cases that asserted the
load path rejects a database-owned value now assert both layers, and
TestRuntimeConfigFailsClosedOnAnInvalidDatabaseValue pins the second.
Loader.overlayFile decoded the overlay file straight into the config the
base file produced, and a decode replaces a map-valued entry wholesale:
[services.state] target = ... in an overlay cleared the target_token the
base file set. That is archie-core-e2e2 on the user-visible -config-overlay
path -- deployments/dev.toml is the documented argument and sets
[services.state] target/listen, while target_token lives in the installed
config.toml. An enabled image.hosted entry lost its class and api_key_env
the same way, silently turning the provider off.
apply.go's foldOverrides is the fix that path was missing, so the file
overlay now goes through the same fold as the map overlay: the overlay file
parses to its raw mapping (decodeFileMapping), the fields of each partially
addressed map entry are folded forward out of the base config, and the
completed mapping is decoded over it. Loader.ApplyOverlay, ApplyOverlayValues
and apply.go stay: the hazard they were written for is this one.
An overlay file still reports its own unknown keys, from a separate decode
into a scratch value.
Resolves archie-core-6zw0. Subcommands of cmd/archie-state-store (not a new
binary, per organisation.md#process-boundaries) operating on the task database
file directly with the State Store stopped: backup, restore, validate, rollback.
A revision carries its value, so restoring replays it through the ordinary
replace and no rollback RPC is added. The pre-flight backup in
scripts/archie-update-install now calls the subcommand; the watchdog's refusal
semantics are unchanged.
After the rebase onto main, controlplane.NewServer takes the step vocabulary
it resolves definitions against (controlplane.ResourceStore, *workflow.Manager)
and returns (Server, error). Both offline call sites still passed the store
alone, so internal/app/archied stopped compiling -- and with it
cmd/archie-state-store, cmd/archied and cmd/archie-gateway.
validate and rollback now build their server through
openStateStoreControlPlane, the root helper the served path already uses, so
the offline checks resolve the same shared provider set
(internal/infrastructure/workflowsteps) as the process whose verdict they
predict, and the constructor's error is returned as a refusal rather than
dropped. No second vocabulary call site: the fwmp pin on
openStateStoreControlPlane covers these paths unchanged.
The test's seeding helper builds its server from the same shared set
(workflowsteps.NewManager) because it cannot call the unexported root helper;
an empty manager there would seed values the served path never validates
against. That is not merely a claim: with an empty manager the three
validate tests fail with `workflow "bootstrap" step 1: unknown type
"bootstrap.prepare"`, so they already pin that the offline path resolves a
live, shared vocabulary.
Four defects from the re-run contract review, all in the recovery path.
A store missing a control-plane kind was refused while a store with no
resources table at all was accepted -- two stances inside validateStore for
states the daemon cannot tell apart. The State Store seeds every kind from
its own config before it serves (RunStateStore -> ImportConfig), so a kind
the file does not hold yet is one archied reads the seed for: an upgrade that
adds a kind, and a validate before the State Store has started again, exited
1 with "read provider-settings: resource not found" for a store the daemon
boots on fine. StoredRuntimeConfig now reads a missing kind as the seed the
State Store would write for it (Definition.seededValue, shared with
ImportConfig so the two cannot drift), and the missing-table branch is the
same stance taken to its extreme.
defaultConfigPath() derived the daemon's config with os.UserConfigDir, which
disagrees with archied's configHome on darwin and rejects a relative
XDG_CONFIG_HOME outright -- leaving validate with an empty path and no
verdict, on a host where archied boots from relative-dir/archie/config.toml.
It is now the daemon's own rule, exported as archied.DefaultConfigPath and
used by the daemon's own -config default too, with a test pinning the
equality.
-h exited 2, the code documented as "the command line was wrong", while the
serve path in the same binary exits 0 and main.go advertises -h for these
subcommands. ErrHelp now returns 0.
validate resolved the boot config through loadConfig -> setupLogging, which
created cfg.Log.File's directory, opened the log, and wrote a line stamped
component="daemon" -- a diagnostic editing the deployment it inspects, and
rotating a log the daemon is writing. boot.stderrLog keeps setupLogging on
the stderr logger for the recovery path: no feed, no task-log registry, no
durable copy.
Tests: TestRecoveryValidateRunsTheGateBootRuns now asserts the unseeded and
the missing-kind stores are accepted (both exited 1 before), while the
bogus-policy case still refuses; TestRecoveryHelpExitsZero,
TestRecoveryValidateLeavesTheDaemonLogAlone and
TestRecoveryDefaultConfigFollowsTheDaemonsConfigHome are new. Each was run
red first, and the seed fallback and stderrLog were mutation-checked.