Slice 1 of docs/prds/memory-engine-unification.md: persist the
participant, so a later curator can address agent-user memory scope.
- messages gains sender_id: schema, ALTER migration (idempotent via
duplicate-column detection), both inserts, and both history selects.
Pre-existing rows default to empty.
- curator.SessionSummary gains AgentID and ConversationMessage gains
SenderID; the sessioncurator adapter fills them from the session's bot
identity and the stored record.
The memory contract is untouched and nothing consumes the new fields
yet. Deriving a session's participant from its user-role messages, and
failing closed on ambiguity, stays with the curator rewrite.
Review found that a bare "showing 3/197 lines" inside ordinary output was
accepted as truncation metadata and replaced by a summary, hiding the real
result. Both built-in producers wrap the notice as "[truncated: …]"
(internal/tools/builtin/truncate.go), so the wrapper is now required rather
than tolerated: it is the only marker separating a notice the tool printed
from content that merely mentions a count.
Four findings from review, all valid:
1. Search detection is now gated on the tool name rather than applied to every
result. A compiler diagnostic ("file.go:42: message") has the same shape as
a search hit, so counting it reported a build error as "N matches" and hid
the diagnostic.
2. Only ripgrep match lines are counted. The hyphen context form
("path-N-text") and "--" group separators were being counted as matches, so
a search with context requested reported the match count plus every
surrounding context line.
3. The notice parser accepts shell's "showing last N/M lines" form as well as
the plain "showing N/M lines" one. Shell emits the "last" form, which failed
integer parsing and was not recognised as metadata -- so a truncated shell
result quoted the raw truncation notice as its preview, the exact defect this
change set out to remove. The summary now also states which end was kept
("97 lines total; showing last 66" for a tail truncation).
4. The renderer cases are consolidated into one table-driven test per the
development protocol in AGENTS.md, covering all four defects above plus
singular/plural, colons in paths, and module-cache noise.
Operator report: tool output in the Telegram channel was unreadable during
search-heavy turns -- a wall of fenced blocks each showing one arbitrary
match line ("/path/config.json-322- }") under a heading that said
only "grep -- done".
The renderer quoted the first non-module-cache line of a result and appended
a line count recomputed from its own truncated sample, which could not agree
with the total the tool had already reported: one search rendered the tool's
"194 more lines" beside our "… 3 more lines". Two different numbers for one
result set read as noise, and the quoted line carried no context at all.
- Describe, don't quote: a search is reported as "N matches", a truncated
result as "<total> lines total; showing <n>", and every other tool falls
back to its first informative line. Truncation notices are skipped, since
they are metadata rather than content.
- Never invent a count: the tool's own reported total is the only
trustworthy source, parsed from its truncation notice rather than
recomputed from the visible sample.
- Remove the now-dead compactToolPreview and countNonEmptyOrContentLines.
Also bump go-telegram/bot v1.25.0 -> v1.27.0, which fixes a streaming bug
that lands squarely on this path: rawRequest sent the multipart body through
an io.Pipe, which net/http cannot replay, so sendMessage/editMessageText
failed until the connection was dropped.
Remaining work from the report (collapsing repeated calls into one counted
entry, and a ring of recent activity lines with a capped frame) is tracked
in #837.
Resolves infra-misc-3.
WIRE: webhookguard.SensitiveKeyMarkers is now the single exported marker
list, and gateway.sensitiveParameterKey consumes it instead of inlining an
identical copy. webhookguard/redact.go documented that the two lists must
match "exactly so the two redaction paths agree"; that agreement is now
structural (one slice, two consumers) rather than a comment nothing
enforces. A new test asserts both consumers observe the same set, so a
future edit cannot silently desynchronise chat-transcript redaction from
webhook-capture redaction.
No /goal, /steer, or /queue command exists in messaging.localCommands, no
Router handler dispatches them, and BuildGoalPrompt was never merged into
the turn's system prompt. NewSessionState had no caller outside its own
test. Delete the entire mechanism rather than let a fully-built but
unreachable feature be mistaken for a live one.
All four production Router constructions left Dangerous nil; the only
Decide implementation was webui.DangerousService, used directly by the
webui HTTP handler, not through the Router. handleDeny and the typed
/approve branches always returned "not configured" in production. Delete
DangerousApprover, the Dangerous field, /deny, and the dead branches so
the router surface matches what is actually wired; the dashboard's
dangerous-action path is unaffected since it never routed through here.
internal/ratelimit was fully built (sliding-window limiter, ticker-driven
eviction) but had no consumer, leaving multi-identity deployments with no
inbound request budget on any channel (archie-core-6wxb / GH #797).
Adds Message.SenderID (Telegram numeric user ID, SMTP from address, or
webhook route path -- the closest thing webhook has to a caller identity),
threads it into Router.Route/RouteStream via a shared checkRateLimit, and
wires an opt-in [chat.rate_limit] config section through bootstrap with the
documented EvictStale ticker.
Completes the objective deletion gate for the UI Service
(archie-core-8cda.5.6): go list -deps ./cmd/archie-ui now links zero
banned packages -- no modernc.org/sqlite, no internal/store, gateway,
channels, memory, secret, agentexec, tools, skill, yaegiutil, curator,
or the workflow engine. The closure shrinks from 68 to 26 archie-core
packages. Each cut is a producer-side move; no consumer-side defensive
shims.
- Store contracts (TaskStore family, value types, sentinels) moved from
internal/store to internal/domain/storecontract -- the approved contract
location per state-store-contract rev. 2d, and NOT internal/contracts:
that tree is buf generate output and tools/proto.sh check deletes
hand-written files there. internal/store keeps aliases; sentinel
message strings are byte-identical and remain the wire contract.
- Chat contracts (ChatContract, sessions, turns, streaming, commands)
moved from internal/gateway to internal/domain/messaging; gateway keeps
aliases. ToolCallEvent rendering helpers become gateway functions
(methods cannot live on alias types).
- Task vocabulary (Task, statuses, mid-run Store, Definition) split into
internal/domain/workflow/task; the pipeline engine stays in package
workflow with aliases, so daemon-side callers are unchanged.
- overlay.DeniedKeys moved to internal/infrastructure/configuration:
webui referenced the constant and that dragged the overlay SQLite
store in.
- webui.Server's concrete runtime handles become webui-owned view
interfaces with daemon-side bootstrap adapters: channel lifecycle
(internal/channels/status), MemoryStatus, CuratorStatus, and a
SkillCatalog seam replacing the direct skill.DefaultRoots/CatalogRoots
calls. Each degrades to an empty page when unwired.
- config.SecretRef moved into internal/config (config vocabulary);
internal/secret keeps an alias and Registry.Resolve absorbs the
zero-value short-circuit.
Docs: state-store-contract rev. 2d relocation record; ui-service-boundary
deletion gate marked satisfied; ARCHITECTURE.md package map rows.
Add a channel-neutral review_pr command that runs the adversarial reviewer
against an existing pull request and returns structured findings. The Web UI
chat and Telegram invoke the same command; the reviewer model is models.reviewer
(falling back to models.builder), configurable independently of planner/builder.
- forge: PullRequest + PullRequestReader.GetPullRequest (github, gitea)
- worktree: Manager.CheckoutPR materialises a PR head for Diff/Snapshot
- agentworker: NewOperatorReviewer + ReviewerModel (single model-resolution rule)
- gateway: review_pr tool, ChatPRReviewer, wired into TurnRunner for every channel
- archied: prReviewer adapter with identity-scoped repo allow-list and
per-PR single-flight, so duplicate/concurrent reviews are deterministic
Closes archie-core-h019.8.
archie-core-8cda.5.4 requires /api/tasks/{id}/action to go through
ChatContract rather than a task-action service the UI process composes
itself. The contract could not carry the request: ApplyTaskAction's identity
is a scope, and the daemon's action service distinguishes a named identity
from a nil one, which is the authenticated operator acting across
identities. Sending "" would not have expressed it either -- that is a real
identity in a single-identity deployment (chatTaskProfiles).
ApplyOperatorTaskAction is that case, additively: a separate RPC, so buf
sees no cardinality change on a shipped field, and no request that can carry
both an identity and operator authority. ChatTaskActor's identity becomes a
pointer end to end, from the Gateway's local adapter through the NATS hop to
taskactions.Service.Apply, which has always taken one.
The conformance test now covers both scopes through the local adapter and
the gRPC hop; the chat tool test pins that a chat action is never unscoped.
No caller passes nil yet -- the dashboard handler switches over with the
rest of the cutover.
Replaced gateway.Message as the currency at the channel-facing boundary
with the canonical internal/domain/messaging.Message. Channel adapters
now build the record directly (RoleUser, ConversationID, Sender) and the
Router hands it back unchanged, so nothing converts between two shapes in
process. gateway.Message is deleted; the dashboard route that had no home
in the record lives in a new Inbound envelope (Message + transport-only
Page), which reaches the system prompt and is never persisted.
Role is now set at construction instead of derived at the store boundary,
and the assistant reply is a direct messaging.Message (RoleAssistant,
Sender=botUser). The gRPC proto is unchanged: pb.Message still carries no
role, and both sides keep deriving it from the owning session's bot
identity via gateway.RoleForSender (ToStoredMessage/FromStoredMessage are
deleted).
Adds tests that the old shape could not express: Page reaches the system
prompt but not the persisted record, and a channel-built Inbound round-trips
with the Role the channel set.
Refs: archie-core-9g5e
89d8dc0 landed with golangci-lint red: scanSearchMessages was an exact copy
of scanMessages bar two error strings (dupl), and the pre-role schema
migration test seeded its fixture with db.Exec (noctx). task check has been
failing on main since.
Collapse the two scanners into one -- SearchMessages already selects the same
column order the history reads do, which the deleted function's own comment
said -- and seed the fixture through ExecContext.
No behaviour change: the gateway suite passes unchanged.
SessionStore message history now stores/loads domain/messaging.Message:
sender attribution, derived role, and conversation address are canonical
record fields. SQLite gains a role column with open-time backfill for
pre-migration databases. gateway.Message remains the channel-facing
transport; conversion lives in one boundary module. ChatContract,
gatewayrpc proto, and dashboard JSON are unchanged; the RPC client
splits into ChatContract (Client) and SessionStore (StoreClient) views.
Sessions stay SessionContext: UUID keys and platform/bot/title/branch
metadata have no lossless home in Conversation; keyed migration tracked
separately once Agent/user/binding ownership settles.
Refs: archie-core-d9dv
* refactor(workflow,store): relocate Task/Status/Source and define workflow.Store
Move store.Task, store.Status*/store.Source*, and the 3-method workflow
store interface out of internal/store into internal/domain/workflow as
workflow.Task/workflow.Status/workflow.Source plus a consumer-owned
workflow.Store, superseding store.WorkflowStore. internal/domain/workflow
now imports zero internal/store symbols; internal/store persists the
domain type instead (infrastructure depends on domain), and its Store
satisfies workflow.Store.
Updates every production and test caller of the relocated symbols across
daemon, webui, gateway, agentworker, agenttransport/nats, storerpc,
taskrun, and worktreerpc, and removes TestRecordDispatchViaExplicitTx per
docs/prds/state-store-contract.md (rev. 2c) sections 4 and 12 step 2.
Two workflow-package integration tests that exercise the real *store.Store
move to an external workflow_test package (store_integration_test.go) to
avoid a workflow(test)->store->workflow(prod) import cycle now that store
depends on workflow.
* no-mistakes(document): docs(state-store): mark §4.8 relocation done in state-store-contract
Stands up archie-gateway as its own binary serving ChatContract over gRPC
and makes it the daemon's only transport for webui chat (Q3 part 5's
atomic flip -- the in-process serving path is deleted, not deprecated).
- contract seams + protobuf toolchain (Phase 0)
- single taskactions.Service across webui, daemon NATS responder and the
standalone gateway's client
- services.gateway.target is the only config surface left; mode is gone
now that remote is the only transport
- task dev launches archie-gateway alongside archied + Vite
Verified end-to-end against live infra: real GitHub token (zero repos
configured, so no repo was polled or touched) and the real
@heyarchie_home_bot Telegram bot, both processes booted together against
shared NATS/store, chat and /status round-tripped successfully.
# Conflicts:
# go.sum
The chat agent had no way to know where it lives without probing the
filesystem (pwd, find, git remote), wasting turns on every orientation.
Per docs/architecture/identity.md the <env> block is Layer 1 invariants,
so the daemon's own facts belong there, driven from live configuration.
Render into the prompt's <env> block:
- workspace path (chat.workspace)
- repositories under management with forge host and default branch
- operator identity (chat.operator, which was dead plumbing)
Unconfigured values render explicitly as unknown / none configured
rather than a plausible-looking default, and a repo with no forge
reports forge unknown, so the model never claims a fact the daemon
cannot confirm. Repos are resolved per active chat identity, mirroring
chatTaskProfiles.
gateway: add RepoEnv, SystemPromptConfig Workspace/Repos/Operator,
TurnRunnerConfig Workspace/Repos.
app: chatRepoEnv resolver wired through newChatTurnRunner.
tests: prompt rendering (incl. explicit-unknown + hostile-SOUL guard)
and chatRepoEnv identity resolution. go.sum gains module-graph
/go.mod checksum lines.
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
/status conflated two unrelated concerns: a thin per-status task
count that duplicated information belonging to a task view, and
runtime config already visible via /model. Worse, a bare status
count actively misled -- "Running: 1" read the same whether that
task was making progress or stuck in a failure loop, because a count
carries no age or identity.
/status now reports one aggregate queue-depth line (in-flight work:
running/waiting/parked, health signal only) instead of an itemized
breakdown -- queued and terminal states are excluded as neither
actionable nor a health concern.
/tasks is new: the live, per-task view with enough identity and
state to act on it (id, title, status, workflow/stage, age since
last transition, attempt once a retry has happened, and a parked
task's reason). Age comes from Task.UpdatedAt, already persisted and
now threaded through ChatTaskSummary and chatTaskListerAdapter --
no new instrumentation needed. Tasks sort actionable-first (running,
waiting, parked, ...) ahead of terminal states, stable within a
status group so the lister's own recency order survives.
Wired via the existing configureTaskCommands call sites (web,
telegram, email, webhook) using the same chatTaskListerAdapter
already built for the task_list LLM tool -- no parallel data path.
Note: internal/gateway's separate /agents command (AgentReader) is
NOT touched here and remains dead in production (Router.Agents is
never wired by any composition-root call site) -- tracked as its own
finding, not folded into this fix.
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)
Persona/SOUL text was interpolated raw at the top of archie.md.tpl,
so a hostile SOUL could inject its own <core_rules> or
<instruction_precedence> tags ahead of the real ones. Render it
through the existing xml-escaping helper into a labeled
trust="data" <soul> block, and tell instruction_precedence
explicitly that the block is reference data, not an authority.
Closes archie-core-1786637490097-12-bf415ffb
channels.Manager.MarkRunning was never called for telegram, email, or
webhook -- every start path only called MarkStarting then MarkFailed
on error, so the dashboard showed a working channel stuck on
"Starting" forever (confirmed live: Telegram polling and delivering
messages while /api/channels still reported "starting").
Adds gateway.Lifecycle{Starting, Running func()}, threaded through
Gateway.Start(ctx, router, lifecycle). Email and webhook report
Running once their listener actually binds -- webhook moves off
http.Server.ListenAndServe's internal bind to an explicit
net.ListenConfig.Listen so binding is observable before Serve.
Telegram reports Running after StartWebhook launches in webhook mode,
or after the first successful getUpdates response in long-poll mode,
observed via a pollReadinessClient http.Client wrapper since the
go-telegram/bot library keeps polling errors inside its own worker
goroutine. Restart re-reports Starting/Running each cycle.
turn.go: extract checkConfigured, replayPriorReply, claimTurn,
recordInboundMessage, prepareTurn, generateAndComplete out of Run
(gocognit 53 -> under threshold, gocyclo brought under 20 then under
cyclop's 15 too).
gateway.go: split dispatchLocal's second half into dispatchLocalMisc
(cyclop 18 -> under 15, on top of the earlier gocyclo fix).
Behavior-preserving; internal/gateway lint clean.
- toolset.go: extract requireApproval out of toolExecute (nestif)
- gateway.go: extract handleDeny/handleVersion/handleUpdateCheck/
handleRestartAdapter out of dispatchLocal (gocyclo 26 -> under 20)
- cronstore/store.go: extract applyPatch/applyScheduleChange out of
Update (nestif, gocognit)
Behavior-preserving.
Funcs() was passed a *strings.Replacer value instead of its Replace
method, panicking gateway's init(). Fixing that surfaced a second bug:
the template still referenced .Operator, removed from promptData by
5465335, which made every render fail and silently fall back to a
generic hardcoded prompt.
Also replaces a real Telegram chat ID hardcoded across several test
fixtures with a placeholder value.
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.
An operator could only reject work that was waiting on a human decision, so a
task sitting in review (PR open) or running had no way to be refused even
when it was clearly dangerous.
- expose reject in taskstate.Actions for every non-terminal state (queued,
running, waiting_human, parked, pr_open) so both the dashboard and the chat
tool surface it; terminal states still only expose archive
- make rejectTask transition from the task's actual status to closed_wont_do
and interrupt running execution first, so nothing keeps working against a
task being closed out
- keep the distinct stop (recoverable → parked) semantics; reject is the
terminal escape hatch
- update the lifecycle contract tests to the new intent
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).
Task runs now preserve real model context limits, provider usage, baseline-repair spend, commit identity, durable agent-finish telemetry, and useful timeline details without imposing a token cap. Dashboard forge links are canonical per owning GitHub or Gitea identity. Chat turn claims acquire SQLite writer intent before reading, eliminating the concurrent SQLITE_BUSY snapshot race.
The agent could produce a file -- a transcript, a log dump, a screenshot --
and had no way to hand it over. Asked to send one it offered a "download"
link; nothing arrived and it had no signal that anything had failed. The
defect was not that it lied: success and silent non-delivery were
indistinguishable to it.
Three places made a local file undeliverable, and each is fixed here:
- liveReply.Media skipped any attachment with an empty URL, so a local
file was dropped before it ever reached a sender.
- telegramMediaSender built every attachment as models.InputFileString,
which tells Telegram to FETCH the string as a URL. A host path is not
fetchable, so the send did nothing while reporting success. It now
chooses by what the attachment carries: URL -> fetch, Path -> upload
via models.InputFileUpload over an open file.
- Nothing let the agent ask for a send at all. New send_file tool
(internal/tools/sendfile) validates the file and returns a
MultimodalResult naming it; the turn's channel delivers it.
Meaning changes, called out because the names did not:
- MediaAttachment gains Path, the alternative to URL. Exactly one is
meaningful for an outbound send.
- tools.MultimodalResult.URLs used to mean remote-hosted media only; it
now means media this result wants delivered, which may be local. JSON
name unchanged so existing producers keep working.
- errNoURL is now errNoSource: having neither source, not having no URL,
is what makes an attachment undeliverable.
Failures are visible at every layer, which is the actual fix. send_file
errors on an absent, irregular, unreadable, oversized or confinement-
refused path, so the model sees it and can report it. The sender checks
Telegram's ceilings (10 MB photos, 50 MB otherwise) before the request and
reports the real size rather than leaving a 400 nobody reads. A delivery
that still fails now appends "could not send <name>" to the reply instead
of an empty link, which read as a delivered file.
Exposing a public URL for local files was explicitly rejected: it trades a
missing feature for a hosting and access-control problem.
send_file resolves paths through builtin.ResolveReadable -- the read tool's
own policy, exported rather than reimplemented -- so sending cannot reach
what reading cannot, whatever [chat] unrestricted_filesystem is set to.
Adds the task_action tool to internal/gateway/task_tools.go and wires it via
chatTaskActorAdapter in cmd/archied. This enables operators in chat to execute
lifecycle actions (abandon, archive, retry, stop, cancel, approve, reject) on
tasks using the same action path and state machine as the dashboard.
Enforces identity scoping so an instance cannot mutate another identity's
tasks, verifies status permissions via internal/taskstate, and classifies the
tool as mutating with required human approval.
Pre-existing code `task fmt` rewrites on every run: a manual clamp to
`max`, and a `strings.Split` range to `strings.SplitSeq`. Both are
semantically identical; committing them keeps the gate from reporting a
dirty tree on unrelated work.
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
Carry the revalidated update snapshot through the installer contract so the shell adapter receives an explicit per-component plan. Build, back up, install, and roll back only selected components, and pass the runtime release version into agent image builds.\n\nCovers gateway-only, runtime-only, combined, and no-op updates.\n\nRefs #606\nBead: archie-core-462t
The concurrent regression retried SQLITE_BUSY immediately, allowing the losing goroutine to exhaust all attempts before the winning transaction was scheduled to commit on GitHub Actions. Yield for one millisecond between bounded busy retries; non-busy errors still fail immediately and exhausted contention still fails the test.