self-development: add review-derived agent conventions by kelos-bot[bot] · Pull Request #1249 · kelos-dev/kelos

kelos-bot · 2026-05-30T18:06:40Z

What type of PR is this?

/kind cleanup

What this PR does / why we need it:

Adds evidence-backed conventions and workflow safeguards to the project
instructions, shared Kelos development AgentConfig, and reviewer prompts.

Compare forwarded streams exactly. PR Stream agent output through kelos-capture instead of writing to /tmp #1189's
review comment
found that per-line substring assertions could not detect dropped, duplicated,
or reordered output. The printer/formatter test rule now also requires full
normalized equality for stream-forwarding tests.
Scope autonomous triggers to trusted authors and intended subjects. PR
Add kanon-development TaskSpawners for autonomous kanon development #1241 was corrected because a
review trigger lacked an author filter
and a
PR-only command also matched issue comments.
The TaskSpawner convention now requires both author allowlists and commentOn
scoping where applicable.
Classify request failures by cause. PR feat: add WebhookGateway CRD for per-channel webhook auth and multi-instance GitHub #1238's reviews found that
TaskSpawner list errors silently returned success
and
all gateway Get errors became 404s.
PR Support file attachments in Sessions #1630 then showed the broader form of the same mistake:
message matching mapped server-side storage failures to 400.
Error-to-status mappings must use typed or sentinel causes, reserve 4xx for
client-correctable failures, and return 5xx for lookup, RBAC, storage, and
transport failures.
Honor reduced-motion preferences. PRs
#1481
and
#1482
independently added continuous running-state animation without a reduced-motion
fallback. New nonessential animation must retain a distinguishable static state
when motion is disabled.
Protect credentials in configuration and transport. PR feat: Add native OpenTelemetry (OTLP) support for metrics and traces #1559's
review
found authorization headers rendered from plaintext Helm values, while PR feat: add GitHub App auth to HTTP context sources #1564's
review
found that an HTTP endpoint exposed a GitHub App JWT. Credential-bearing values
must use Secret references, and their endpoints must use HTTPS unless explicitly
opted into insecure transport.
Protect state-dependent transitions from stale work. PR Add section choosers to the Session server #1548's
review
found that an old dialog request could mutate a newly opened Session dialog. PR
Allow Session interruption while draining #1618's
review
similarly required an interrupt to be revalidated against the snapped active turn.
PR Use inline session section selector #1649 then left a
finding
because an async failure path rendered a captured pre-request Session after
polling had refreshed the selected Session. PR Add Session goal and shell commands #1650 exposed the interaction-level
form of the same problem: a
stale goal notification could complete a new message,
interrupt could race goal startup,
completion used provider-global goal state,
and
goal control matched the wrong active turn kind.
PR Keep kelos logs -f open after empty streams #1654 then exposed the same precedence problem across independently updated
lifecycle signals: a
failed Pod was treated as retryable while Task status lagged,
and a
non-zero terminated container was also treated as retryable.
Async completions and controls must re-read current state, bind to the exact
operation identity and kind, and coordinate startup and handoff so stale work
cannot render or mutate another operation. When lifecycle signals disagree or
lag, terminal and failure evidence must be inspected before returning a
transient or retry decision.
Test behavior through the layer that owns it and assert the primary effect.
PR Add section choosers to the Session server #1548's same
review
found that source-fragment assertions did not exercise stateful UI behavior. PR
feat: Add optional Prometheus Operator PodMonitor to Helm chart #1560's
review
found that a render test never exercised the non-default namespace branch it
claimed to cover. PR Display and manage pending Session messages #1648's
review
also found that an exact minified CSS-rule assertion was brittle while failing
to exclude declarations that would restore the regression. PR Support file attachments in Sessions #1630 had a
download fake that discarded the requested attachment ID,
PR Monitor pull request checks in Sessions #1639 had an
end-to-end fixture that omitted the changed check-rollup response,
and PR Reduce Session GitHub API usage #1652 had a
command fixture that accepted any GraphQL query and could not detect duplicate calls.
PR Add Session goal and shell commands #1650 also had an
interrupt test that checked journal status without checking process death,
while PR Reduce Session GitHub API usage #1652 had a
periodic test whose blocking callback leaked its goroutine.
PR Keep kelos logs -f open after empty streams #1654's
test review
also found that new terminal, transient, and unknown lifecycle branches were
left unvalidated.
Tests must execute the owning layer and important non-default branches, assert
the real routed or lifecycle effect rather than a bookkeeping proxy, and make
appropriate static assertions on behavior instead of serialization details.
State-dependent decision tests must cover each relevant terminal, transient,
and unknown class plus combinations where status sources can lag one another.
Fakes, fixtures, and command shims must exercise changed response fields,
validate complete request shapes, reject unexpected calls, and expose call
counts when batching or request reduction is part of the contract. Background
goroutines and child processes must also be cancellation-aware and verified
stopped before the test returns.
Keep all documentation surfaces complete and consistent. Recent reviews
repeatedly found partial documentation updates: PR Show sessions waiting for input #1604
claimed header styling that was not implemented,
PR Track sticky issue validity verdicts #1611 left
per-spawner sections inconsistent with its overview,
PR Paginate Session terminal history #1613
removed documentation for still-supported flags,
and PR Add Session turn duration separator #1616
omitted newly observable terminal behavior.
Agents must search related overviews, component sections, tables, setup examples,
and command/flag references and update all affected mentions together.
Separate historical replay from live event handling. PR Paginate Session terminal history #1613's
review
found that older terminal pages could be rendered as continuations of the active
live stream and that duplicate history requests reused a cursor. PR Paginate Session web chat history #1621's
review
found that web history replay could emit a false live interruption toast, omit
independently needed current file state, and leave viewport preservation untested.
PR Browse Session prompts in terminal chat and console #1739's review found two related failures: the web dialog
retried an expired cursor indefinitely,
while the terminal
reused cached prompt history and its cursor after reconnecting.
The author confirmed the
web reload fix
and
terminal reconnect fix,
and the merged diff includes executable tests for both paths.
Replay must keep its rendering, state, and side effects separate while preserving
valid cursors, request state, current state, order, and viewport behavior.
The history rule in AGENTS.md and self-development/agentconfig.yaml also
requires clearing stale cursors and cached pages when reconnecting to replacement
history or recovering from an expired cursor, so the next request starts at the
first page. Tests must cover recovery from expired cursors as well as reconnects.
Use lease-protected pushes after rebasing published branches. PR Add full spawner fleet for kelos-dev/open-actions #1622's
reviews independently found that the
worker existing-PR path
and
PR responder
rebased already-published branches and then attempted normal pushes, which fail
as non-fast-forward updates. The shared instructions now require
git push --force-with-lease, never an unleased force-push, after rebasing a
published branch.
Treat controller-generated Kubernetes names as durable identities. PR
Bound Session workload names #1629's reviews found that dropping the Session workload prefix would
collide with legacy workloads
and
abandon retained persistent workspaces.
The independent
Kelos Claude review
also found that a 63-character StatefulSet bound still left no room for
ControllerRevision and Pod-label suffixes. The
review on PR #1631
then showed that a Session name valid under DNS-1123 can still be invalid for
its governing Service under DNS-1035. Name derivation must validate the exact
grammar of every generated kind and account for all downstream suffix and label
limits, and controller-owned resources must not be renamed without a
collision-safe migration that preserves ownership and data.
Keep shared feature contracts consistent across clients. PR Support file attachments in Sessions #1630's
review found that the plain terminal
did not enforce the TUI's eight-file limit,
the TUI
accepted an empty /send that the plain terminal rejected,
and the plain terminal
omitted /send from command help.
The
independent Kelos Claude review
also identified the cross-client limit mismatch. Shared web, TUI, and CLI
features must apply the same limits, empty-input rules, failure semantics, and
discoverability, with each surface tested explicitly.
Preserve prompt contracts when consolidating instructions. PR Simplify session spawner prompts #1641
removed detailed worker and PR-responder prompts on the assumption that shared
configuration supplied the same behavior. Reviewers found the same missing-owner
failure across
Open Actions workers,
Agora workers,
the Kanon responder,
the Agora responder,
the Open Actions responder,
and the Kelos responder.
Separate comments also found the
persistent-workspace test contract
and
Agora documentation
out of sync. Before consolidating prompt instructions, agents must map every
removed requirement to a concrete source the spawned agent actually loads and
update contract tests and documentation in the same change.
Audit responsive rules when changing web controls. Three recent web
changes independently missed existing phone behavior:
PR #1630 left the attachment button below the mobile touch-target size,
PR #1638 resized the composer without resizing its adjacent attachment control,
and
PR #1647 omitted its new textarea from the mobile anti-zoom selector.
Agents must sweep every relevant selector group and breakpoint override for
new or renamed controls, preserving mobile font sizing, touch targets, and
sibling alignment.
Use production identities and collection cardinality in tests. PR
Handle rotated CRD CA bundles during install #1656 had a
multi-certificate review finding
because singleton-only expected bundles did not exercise whether every
required certificate was present. PR fix(slack): stop retrying replies rejected with cannot_reply_to_message #1660 then used an empty Task UID in
a
keyed-suppression test,
so the test passed without exercising the production identity path.
Fixtures must populate identity fields used as keys and use multi-element
inputs when behavior depends on collection membership or aggregation.
Test API mutability through updates. PR Make Session idle policy mutable #1665's
API review
and
general review
independently found that making Session.spec.idlePolicy mutable was tested
only for one non-nil replacement. A
separate inline finding
also found that create-only negative cases did not prove validation remained
enforced on updates. Mutability changes must cover the newly allowed presence
transitions and at least one invalid update.
Preserve keyboard operability in dynamic web controls. PR Add per-session actions menu to session server #1663's reviews
found that keyboard-invoked actions
lost focus after rerendering their trigger,
menu items
removed the visible focus ring,
and
Tab left the menu open after focus moved away.
PR Show Task logs in the Console #1673 was then independently flagged by
Cubic
and the
Kelos Claude reviewer
because roving tabindex made the inactive Manifest tab unreachable without
Arrow/Home/End handling. Custom widgets must implement their expected keyboard
interaction model; dynamic menus and dialogs must also retain a visible focus
indicator, dismiss on expected keyboard navigation, restore focus after success
and failure, and test those interactions through the browser behavior layer.
Make component renames upgrade-safe. PR Turn the Session server into the Kelos Console #1670's current-head reviews
found that the Console rename
dropped the persisted namespace selection
and
left legacy installed objects, uninstall RBAC, and Helm values without an upgrade path.
Component renames must inventory prior deployed identifiers, migrate stored
state, clean up resources that non-pruning installs leave behind, retain
uninstall cleanup for old names, and reject unsupported stale configuration
actionably. Upgrade-state tests are required alongside fresh-install tests.
Test every distinct behavior path independently. PR Add Session suspend action to console #1681 reviews found that the suspend tests depended on an earlier test global, omitted a runtime dependency and could not execute, and left the idempotent API branch untested. PR Visualize resource relationships in the console #1682 then had an independent review find that an exact-set relationship test skipped core derivation branches while a dependency edge was missing. Tests must now establish their prerequisites, include runtime dependencies, prove that the intended path is reached, and inventory relevant mapping variants.
Centralize behavior that must stay in lockstep. A
review on PR #1673
found duplicate authentication and error handling across two request helpers.
A
review on PR #1681
found duplicated suspend and resume bookkeeping that could drift, and a
review on PR #1683
found the Console build commands duplicated instead of routed through the
existing Makefile target. Code paths that intentionally implement the same
behavior and differ only by an input or action must share the implementation,
and repeated build or generation commands must use one Makefile target. The
rule explicitly excludes incidental similarity to avoid unnecessary
abstraction.
Preserve touch-accessible actions across breakpoints. The Claude review on PR #1686 found that the mobile breakpoint hid the only touch-usable path for Session section assignment while drag-and-drop was not reliable on touch devices. An independent inline review also found that the redesigned mobile tabs fell below the established 44px touch target. Responsive changes must keep every supported action available through an equivalent touch-operable path and test that path.
Preserve readable contrast across supported themes. A review on PR #1663 found that removing an explicit focus outline left only a weak dark-mode indicator, and PR #1686 reused a light-theme label color that produced roughly 2.5:1 contrast in dark mode. Color and theme changes must verify text, interactive labels, and focus indicators in every supported theme against the applicable WCAG AA contrast requirement.
Preserve Unicode semantics in text processing. A review on PR #1673 found that byte-offset truncation could split a multibyte rune and emit invalid UTF-8. A review on PR #1692 then found that len() enforced a documented character threshold in bytes, causing non-ASCII false negatives and misleading diagnostics. Character limits must count Unicode code points, byte-bounded truncation must preserve UTF-8 boundaries, and affected limits must be tested with non-ASCII input.
Preserve UI behavior when restructuring web views. PR Make console sidebar controls scrollable #1688 moved
Console sidebar scrolling to a new container but left
menu dismissal attached to the old non-scrolling list.
PR Keep the current request visible in console #1702 later let
hidden-view geometry updates clear current-request state without recomputing it on reveal.
Web view restructuring must migrate behavior bindings to the actual
containers, avoid deriving state from hidden layout geometry, recompute
derived state when a view becomes visible, and test the affected scroll and
hide/reveal interactions.
Bound work and persisted state on long-lived data and hot paths. PR Show Task logs in the Console #1673 reviews found
an entire accumulated WorkerPool log read into memory
and a later
fallback that still scanned the unbounded log.
PR Visualize resource relationships in the console #1682 then had a
quadratic TaskBudget-to-resource scan,
while PR Keep the current request visible in console #1702 performed
synchronous geometry reads and DOM writes on every scroll event,
independently echoed by its
Claude review.
The current-head API review on PR API: Add TaskPipeline for multi-stage orchestration #1698 then found that
copying every child output into CR status
could exceed Kubernetes object-size limits and wedge status updates. Agents
must bound long-lived input and persisted summaries, use references or
explicit truncation for high-cardinality status, prefilter nested matching,
and coalesce high-frequency UI work while avoiding unchanged DOM writes.
Revalidate preservation data during API conversion. The resolved
review on PR Match Slack spawner events by exclusion rules #1703 found that a user-writable preservation annotation could
bypass the destination field's pattern and item-count schema
after a syntactically valid JSON decode. PR Match GitHub webhook events by pull request author #1708 then repeated the same
omission for another preserved field, where a
current-head review found missing item-count and length validation.
Conversion code must treat preservation storage as user-controlled, enforce
the destination field's complete schema before restoring it, and test
syntactically valid but out-of-schema values as well as malformed encodings.
Keep configurable ServiceAccount identities in lockstep. PR feat: add WebhookGateway CRD for per-channel webhook auth and multi-instance GitHub #1238
repeatedly exposed incomplete handling of a Helm-configurable identity: the
gateway Pod could use a
ServiceAccount that lacked the required RBAC binding,
a second review found the
same custom-name drift across rendered resources,
and a later review found that
reserved chart-managed names rendered duplicate ServiceAccounts.
Helm-controlled ServiceAccount names must reach the workload, account, and
every RBAC subject; chart-owned collisions must be rejected and both valid
non-default and reserved values must be render-tested.
Keep auxiliary failures from breaking primary work. In three PRs, a
best-effort step broke work that would otherwise have succeeded. Each fix
degraded only the enhancement:
- PR Show merge queue status for Sessions #1642 made an added merge-queue lookup a
  hard prerequisite for Session pull-request status,
  which the
  Kelos Claude API review
  also flagged. The merged change logs the lookup failure and falls back to
  the pull request's checks.
- Two independent current-head reviews on PR Keep Console sessions awake while connected #1710 found that a
  single transient heartbeat write error tore down the live Console bridge,
  which the
  Kelos Claude review
  confirmed independently.
- PR Fix console workspace changes across agents #1779
  aborted Session runtime startup when Git inspection for the Changes tab failed,
  for example on a workspace with dubious ownership. The author
  confirmed the fix:
  directory configuration is still validated, but inspection failures are
  logged, reported in snapshot responses, and retried on later requests.
The rule in AGENTS.md and self-development/agentconfig.yaml previously
covered only lease refresh. It now covers any best-effort enhancement: keep
the primary result, log the failure, report it where the enhancement is shown,
and retry later, without aborting startup or tearing down a healthy
connection. Invalid configuration and required writes remain fail-fast. The
Kelos reviewer did not flag the Fix console workspace changes across agents #1779 startup abort at the head cubic reviewed
(dccd90f). Its checklist flags only silent degradation (as P1), so
kelos-reviewer.yaml now includes the converse check at P2.
Keep provider integrations isolated end to end. PR Add GitLab source and webhook support for TaskSpawners #1721 repeatedly mixed the new GitLab path with GitHub assumptions: spawner reporting received a GitHub-only token resolver, GitHub App keys could suppress GitLab token injection, and a GitLab-mode reporting controller could reconcile GitHub Tasks. A separate self-hosted URL finding, confirmed and fixed by the author, found that deriving the project path dropped the instance URL prefix. Provider additions must route credentials, configuration shapes, and work by explicit provider identity and test coexistence and self-hosted forms.
Verify downloaded artifacts fail closed. Two reviews on PR Add GitLab source and webhook support for TaskSpawners #1721 found that sha256sum --ignore-missing could install an unverified glab package when its checksum entry was absent, independently flagging the Cursor image and Gemini image. Artifact verification must select exactly one checksum for the downloaded filename and fail when no unambiguous match exists.
Synchronize state shared with test goroutines. PR Add GitLab source and webhook support for TaskSpawners #1721 introduced two HTTP-handler tests where handler goroutines wrote variables read unsafely by the test goroutine: the GitLab source request fixture and GitLab reporting assertions. Shared test state must use mutexes, atomics, or channels and remain clean under the race detector.
Validate collection semantics in generated CRDs. PR Add GitLab source and webhook support for TaskSpawners #1721 documented
allowed values for two list fields, but review found that the
generated CRD omitted both item enums.
The field-specific reviews confirmed that the markers for
gitlabWebhook.events
and
gitlab.types
did not constrain items, while another review found that
duplicate types caused duplicate discovery work.
Slice enums must be verified in generated YAML, and duplicate values must
be rejected when they have no semantic meaning. The API reviewer prompts
now check these generated collection semantics explicitly.
Keep CLI resource detail output aligned with configuration. PR Add GitLab source and webhook support for TaskSpawners #1721
repeatedly added TaskSpawner settings without exposing them to operators:
GitLab webhook excludeAuthors was omitted,
and the polling source hid both
reviewState and pipelineStatus.
New user-visible fields must be added to every applicable detail printer,
compared with sibling provider branches, and covered by distinct full-line
assertions. The general reviewer prompts now check this surface explicitly.
Page unbounded Kubernetes and API lists. PR Remove v1alpha1 API #1725 attempted to migrate
all stored Kelos resources with
one unpaginated list request,
which can fail once the API server response limit is exceeded. The existing
bounded-work rule now requires limits and continuation-token loops whenever
result cardinality is not contractually bounded.
Keep skill discovery metadata aligned with skill behavior. Reviews on PR
Narrow Kelos skill discovery scope #1720 found that narrowing the skill frontmatter
left supported documentation uses undiscoverable
and
removed distinctive kind names needed for reliable triggering.
Skill descriptions are now treated as discovery contracts: their positive
terms, exclusions, and body-level usage guidance must remain aligned.
Verify required filesystem outputs. PR feat: support optional skills packages #1726 accepted a successful
required skill installation when it
created an empty output directory,
then a follow-up review found that its relocation loop
silently omitted dot-prefixed entries.
Setup and install scripts must validate usable required artifacts rather than
only exit status or directory existence, preserve supported hidden entries,
and exercise those postconditions with executable tests.
Verify external contracts before reporting defects. Two review threads
on PR Add GitLab source and webhook support for TaskSpawners #1721 were resolved after the author corrected unsupported assumptions:
GITLAB_HOST accepts a URL,
and
head_pipeline requires the single-MR endpoint.
The GitLab CLI documentation,
pinned glab v1.116.0 host parser,
and merge-request response documentation
support those corrections. Both general reviewer prompts now require
verification of endpoint fields and accepted inputs against the relevant
official documentation or upstream implementation, a supporting source link,
and omission of unverified external-contract claims. The provider-isolation
evidence above uses the confirmed relative-URL parsing defect, so the rejected
host-only claim is no longer presented as a valid finding.
Use shared criteria for new source exclusion filters. The maintainer
requested the same rule-based exclusion model in the
Slack review on #1703
and the
GitHub review on #1708,
and the author
accepted the shared direction.
Both the
API review
and
general review
then confirmed at head 1661bf4 that reusing the matcher without validating
event-specific criteria could make a label exclusion reject every push;
explicitly empty criteria could also become an unintended match-all rule.
AGENTS.md and the Kelos development AgentConfig now require shared typed
criteria under source-level excludeFilters, AND within each rule,
rejection on any matching exclusion regardless of accepting filters, and
preservation of shipped positive-filter semantics. Agents must validate
empty and inapplicable criteria and test intended matches, missing attributes,
empty values, and exclusion precedence across supported trigger paths.
Keep stream control and teardown responsive under backpressure. In
PR Add interactive shell terminals to Console sessions #1736, review found that
blocked stdin prevented reading disconnect messages
and that
the bridge returned without stopping its workers.
The author confirmed the fixes for
bounded input and responsive control handling
and
cancellation, socket interruption, and worker completion;
both threads are resolved and the merged diff contains the fixes. PR Increase Session attachment limit to 100 MiB #1735's
download review
also verifies that closing a stream cancels exec and unblocks a pending pipe
write. AGENTS.md and the Kelos development AgentConfig now require control
handling to remain responsive when consumers stall, bounded queues with
explicit overflow handling, and bridge teardown that interrupts I/O and
waits for its workers. Executable bridge tests must exercise blocked
consumers, control delivery, overflow, and worker termination.
Preserve UI behavior across view and resource lifecycles. In PR
Add a native Session terminal tab with shell completion #1737, both the
inline review
and the independent Console review found that page departure disposed the
terminal while leaving its tab selected and Reconnect disabled. The author
confirmed the fix:
close the shell, select a usable view, and allow Terminal to reopen after
restoration. PR Add browser alerts to the console #1744 exposed another missing lifecycle transition:
browser notifications survived Session deletion or recreation.
The author
confirmed cleanup by namespace/name/UID on accepted Session refreshes,
covering selected and background Sessions. Both inline threads are resolved,
and the merged diffs contain the accepted fixes. The existing UI convention
in AGENTS.md and self-development/agentconfig.yaml now includes page and
Session lifecycles, cleanup of invalid resources, usable views and controls
after teardown, and tests for cleanup and restoration or reopening.
Verify child identity before reusing a generated resource. On PR API: Add TaskPipeline for multi-stage orchestration #1698,
both the API review
and general review
found that a matrix stage review and a single-Task stage review-0 could
generate the same Task name. The AlreadyExists path checked only the
pipeline owner, so the later stage reused the earlier stage's successful
Task and reported success without running. The maintainer
confirmed the fix and reconciliation test
in merged head d823e10: reuse requires matching ownership, stage, and index;
a different child's name collision fails explicitly. The existing generated-name
convention in AGENTS.md and self-development/agentconfig.yaml now requires
checking the intended child's identity as well as ownership and testing
collisions between children of the same parent. This extends the existing
identity guidance with a confirmed controller contract.
Preserve pending UI intent and keyboard dismissal. PR Move browser notifications into console Settings #1746's Cubic review found that a pending browser-permission render reset the user's switch, keyboard focus could leave the Settings popover without dismissing it (pending state, focus dismissal), and the browser test double dropped repeated listeners (test-double semantics). The shared rules now preserve in-progress UI intent during asynchronous actions, require focus-leaving dismissal or an intentional focus trap, and require browser test doubles to preserve relevant platform semantics.
Pin CSS/layout fixes with owning tests. PR Fix session conversation alignment with long headers #1747's Kelos Claude review found that the new .conversation layout declaration was not included in the existing style assertions. The shared testing rule now requires a focused assertion for each behavior-critical CSS declaration in the owning test.
Guarantee that waits and lifecycle gates can finish. Three PRs shipped
waits or lifecycle gates with no guaranteed exit. Each finding was fixed
before merge:
- PR Allow Session interruption while draining #1618:
  forced interruption waited without a bound on a provider that ignores cancellation,
  and a
  non-deadline interrupt error left a draining Session Interrupting indefinitely.
  Both were addressed in aba5bda.
- PR Allow Session service account updates #1774: a recreated runtime RoleBinding
  dropped the old Pod's access and left the drain waiting indefinitely.
  The merged code keeps drain access when it creates or repairs the binding.
- PR Preserve Session background replies across user turns #1780: cubic
  and the Kelos review at 1e5a9a1 found that completeInteraction left
  ready open, so interrupting a new goal stalled until timeout and then
  restarted Codex. The
  Kelos review
  was edited in place. Its versions at e0a02c6 and b9588e2 found two
  more problems: a background task that never ends kept the Session
  Active forever and blocked idle-policy and runtime-update drains, and
  Claude Code's persistent ambient monitor tasks were counted as
  background work. The merged commit dcf7b3a closes ready, skips
  ambient tasks, applies the task gate only to the idle drain, and
  documents that effect.
AGENTS.md and self-development/agentconfig.yaml now require:
- releasing every owned channel, pending flag, or waiter on every exit
  path;
- bounding waits on external work that may ignore cancellation, and
  escalating when the bound expires;
- gating idle, drain, and interrupt transitions only on signals that reach
  a terminal state, and limiting unbounded user-work gates to the
  transitions that need them, with the effect documented;
- keeping the access a pending drain needs;
- testing failed, canceled, and never-finishing paths.
The existing stale-work rule covers the wrong operation acting on shared
state, and the backpressure rule covers only stream bridges. Neither
requires a wait to finish.

Two findings of the same shape are not used as evidence:
- A Recover Sessions with StatefulSets #1455 finding that turn.accepted was never closed after a journal
  failure. Its thread is unresolved, and the fix could not be confirmed.
- A Fix console workspace changes across agents #1779 cubic finding that a request flag could stay set forever. The
  author rebutted it.
Clear runtime-derived UI state across Session suspension. PR Show background task progress between Session turns #1786 adds
a Console background-task indicator. Two reviews found live runtime state
that outlived the runtime it described:
- The Kelos review
  found that the count stayed visible after the Session was suspended or
  stopped being Ready. It noted that a stale active turn already had the
  same gap on main.
- cubic
  then found that resuming restarted the elapsed-time timer for a turn
  from the stopped runtime before fresh runtime events arrived.
The author fixed both
(suspension,
resume).
An unavailable Session now clears foreground activity and the
background-task count in the current view and its cache. Tests cover
resume before fresh runtime events and cached-view restoration. Both
fixes are in d3e6a82, the head that merged on October 1.

Item 40 applied the lifecycle rule to terminals and browser notifications
across page departure and Session deletion. It did not cover suspension or
live activity indicators. AGENTS.md and
self-development/agentconfig.yaml now also cover live runtime activity
and Session suspension/resume. They require clearing runtime-derived state
in current and cached views, and testing resume before fresh runtime events
arrive.
Grant only the access a component uses. Reviews on three merged PRs
found RBAC permissions or credentials that reached beyond the code that
used them. Each was fixed before merge:
- PR fix: grant list/watch on secrets for webhook RBAC and fall back to global GitHub token resolver #1363: the maintainer's
  P1 review
  found cluster-wide list/watch on Secrets when the webhook only
  needed named reads. The author
  replied "Done",
  and the merged diff contains no RBAC change.
- PR feat: add Task-native budget enforcement via TaskBudget and TaskRecord CRDs #1426: cubic found a
  TaskSpawner RBAC marker on the Task controller, which never reads TaskSpawners.
  cubic marked it addressed in 940e7ed, and the marker is absent on
  main.
- PR Add OIDC authentication and Kubernetes RBAC to Console #1787: cubic found that
  GH_TOKEN was inherited by the whole integration-test step,
  which runs pull-request code, and that
  the OAuth2 Proxy sidecar shared the Console ServiceAccount token.
  The author
  confirmed the sidecar fix,
  and main sets GH_TOKEN only on the download step.
The Kelos review on #1787
raised the sidecar token only as an optional P3 suggestion and did not
flag the CI token. AGENTS.md and self-development/agentconfig.yaml now
require RBAC scoped to the API calls the code makes and credentials exposed
only to the container, process, or CI step that uses them.
kelos-reviewer.yaml adds a matching security check: P1 when a credential
or Secret is exposed, P2 otherwise.
Recompute geometry-derived UI state on every geometry change. PR Fix console prompt jumps and add Jump to latest #1793
added a Jump to latest button whose visibility depends on distance from
the bottom of the conversation. Two independent reviews found that it was
refreshed only from scroll events:
- The Kelos review
  found that visibility was measured right after a bottom anchor was
  scheduled but before it ran. The anchor callback then moved
  scrollTop without refreshing the button, which flashed visible on
  open and on cached Session switches. The author
  fixed it
  by refreshing the button in the anchor callback before paint.
- cubic found that
  late image loads grew the conversation after the one-shot anchor
  without firing a scroll event, so the button stayed hidden away from
  the bottom. The author
  confirmed it with lazy images and fixed it
  by refreshing the button when image previews load.
Both fixes are in the merged commit dce3dd0. PR Keep the current request visible in console #1702 had the same
shape:
hidden-view geometry cleared current-request state and was not recomputed on reveal.
That is the only non-scroll case the rule from item 24 named. In
AGENTS.md and self-development/agentconfig.yaml, the rule now requires
recomputing geometry-derived state whenever the underlying geometry
changes, not only on scroll events. That covers reveal, deferred
programmatic scrolls (after they run, not when scheduled), and late
content growth such as image loads. It also requires testing those
changes. The Kelos review also noted that the harness mocked the anchor
and checked only the end state. The existing rule that browser test
doubles must preserve platform semantics already covers that.
Document the residual security risk of access changes. Kelos reviews
on three PRs found docs for new access or authentication features that
left out risks operators must manage. Each author fixed the gap by
documenting it:
- PR Allow Session service account updates #1774: the
  Kelos review
  found that making serviceAccountName mutable let a Session's own
  agent switch its Pod to another ServiceAccount. It asked for a fix or
  an explicit statement of the trust boundary. The merged commit
  4bcd679 documents the trust boundary in docs/reference.md.
- PR Add namespace membership and configuration administration to Kelos Console #1792: the
  Kelos review
  found that the chart README understated what a compromised Console
  could do. Its create rolebindings permission lets it bind roles it
  already holds, such as pods/exec, to any identity. The merged commit
  87f6447 states this.
- PR Add Console GitHub login with GHES endpoint support #1795: the
  Kelos review
  found that the GitHub login docs did not say that removing a user from
  the organization or team leaves existing Console logins active, or how
  to revoke them. The author
  documented both
  in bca8487.
The existing "Docs must match implementation" rule forbids overstated
guarantees, which covers Add namespace membership and configuration administration to Kelos Console #1792. It does not require documenting residual
risk the docs leave out, as in Allow Session service account updates #1774 and Add Console GitHub login with GHES endpoint support #1795. AGENTS.md and
self-development/agentconfig.yaml now cover changes that grant
permissions, authenticate users, or accept a trust boundary. Their docs
must state:
- what a compromised component or permitted identity can reach,
  including access it can grant to others;
- whether removing upstream access ends existing sessions, and how to
  revoke them.
kelos-reviewer.yaml is unchanged because the reviewer found all three.
Preserve latest-only fields through older-version conversion. Two
PRs added v1alpha2-only fields that a v1alpha1 read-modify-write could
drop:
- PR Add branchLock to let read-only Tasks skip the branch lock #1808 (open) adds branchLock and preserves it for TaskSpawner
  templates in a v1alpha1 annotation. cubic
  found
  that Task conversion still dropped it. Because Task.spec is
  immutable, writing a Task back through v1alpha1 would then fail
  validation. The author
  fixed it
  in 7fe5ffe by preserving the field on Tasks with the same annotation,
  and added round-trip tests.
- PR feat: add Task-native budget enforcement via TaskBudget and TaskRecord CRDs #1426 added Task status.usage and status.conditions. cubic
  found
  that a v1alpha1 status update could erase them, and the
  Kelos GLM review
  found the same gap. The maintainer
  called it valid
  and deferred it to a follow-up. internal/conversion/task.go on main
  still converts Task status with no preservation, and no tracking issue
  was found.
The existing latest-version rule only says older served versions must
stay "convertible". The preservation rule assumes preservation already
exists and covers only validation. In AGENTS.md, the latest-version
rule now requires preserving each latest-only field on every kind that
carries it (spec, status, and embedded templates) and testing the round
trip for each kind. That rule is only in AGENTS.md, which spawned
agents in this repository load, so self-development/agentconfig.yaml
is unchanged.

The reviewer prompts are unchanged. The GLM review that found the feat: add Task-native budget enforcement via TaskBudget and TaskRecord CRDs #1426
gap was later folded into the current reviewers by self-development: run every spawner on Claude Code #1762. No Kelos
review has run on Add branchLock to let read-only Tasks skip the branch lock #1808, so there is no evidence that the current
reviewers miss this.

PR Support files in inline AgentConfig skills #1841 then answered the October 13 question in the notes below:
should the rule allow a documented drop? Its first version preserved
the new v1alpha2-only skill files in a v1alpha1 annotation. cubic
found
that a valid 900 KiB bundle could exceed the annotation limit, and the
first Kelos API review rated the annotation overflow a P1. The maintainer
replied
"Let's not care about the compatibility of the v1alpha1." The author
removed the annotation
and its round-trip tests. They also added the sentence the
API review
asked for: reads through v1alpha1 omit files, and writes through it
remove them. Support files in inline AgentConfig skills #1841 merged that way on October 25. docs/reference.md
on main documents the same kind of drop for Add requestedReviewer to GitHub webhook filters #1813's
requestedReviewer and for the pullRequestAuthor, conclusion, and
checkName criteria. internal/conversion on main still has ten
preservation annotations, so preservation stays the default. The rule
now adds one exception. When the maintainer decides not to carry a
field through older versions, agents skip its preservation and state
the drop in the field's docs/reference.md entry. The API reviewer
prompt is unchanged, because the Support files in inline AgentConfig skills #1841 API review already asked for
that documentation.
Bound every field of size-limited messages and objects. Three PRs
had fields that could grow past a fixed size limit:
- PR Fix console workspace changes across agents #1779: the
  Kelos review
  found that the 1 MiB cap on workspace.changes applied only to diff
  text. The changed-file list had no limit, so a large workspace pushed
  the event over the Console bridge's 8 MiB line limit, and the Session
  reconnected in a loop. The author
  bounded the file count and encoded file-list size
  and added tests that include JSON escaping.
- PR Share Session read state across browsers #1822: cubic
  found
  that the Session kept one read-receipt annotation for every user. They
  would accumulate until the Kubernetes annotation size limit rejected
  every acknowledgment. The author
  added pruning of superseded receipts
  with a test, and it merged in 92e6926.
- PR Add explicit review scopes to console Changes tab #1828:
  cubic
  and the
  Kelos review
  independently found that a new branch list in the same
  workspace.changes event has no limit. The event's other fields were
  sized to fit the 8 MiB line. The author
  capped the list
  at 1,000 entries and 128 KiB of encoded JSON, reported truncation, and
  added tests that combine a bounded list with large escaped diffs. The
  fix merged in b717c57.
Item 25 cited PR API: Add TaskPipeline for multi-stage orchestration #1698, where copying every child output into CR status
could exceed the object size limit. The bounded-work rule covered only
CR status. When Fix console workspace changes across agents #1779 was reviewed, no other PR had a message-limit
finding, so no rule was added then. In AGENTS.md and
self-development/agentconfig.yaml, the rule now covers any message or
object with a fixed size limit, such as a streamed event line, CR status,
or a resource's annotations. It requires:
- keeping the payload within the limit at the maximum cardinality of
  every variable-length field;
- counting each field's encoded size in one shared budget;
- pruning or capping entries that accumulate per user or per child.
The existing requirement to test bounds with realistic input sizes still
applies. kelos-reviewer.yaml is unchanged: the Kelos reviewer flagged
Fix console workspace changes across agents #1779 and Add explicit review scopes to console Changes tab #1828 and missed only Share Session read state across browsers #1822.
Retry only failures that can clear on their own. Five PRs had a
component retry a failure that could not succeed until an input
changed:
- PR Add Workspace options to restrict the checkout and agent credentials #1831 (open): cubic
  found
  that the WorkerPool reconciler returned an error for an unsupported
  Workspace. It retried with backoff for a state that would not clear
  until the Workspace changed, and it skipped reconciling a Task with
  the pool's name. An
  earlier cubic finding
  noted that the rejection was only logged. The author
  now records a Warning event and returns without an error,
  and the Workspace watch re-enqueues the pool.
- PR fix(slack): stop retrying replies rejected with cannot_reply_to_message #1660 fixed a shipped defect. The Slack reporter retried a reply
  that Slack rejected with cannot_reply_to_message on every 30-second
  cycle, and production saw 198+ failed attempts in one hour. The fix
  treats that error as permanent, keeps HTTP-level failures retryable,
  and tests both.
- PR Share Session read state across browsers #1822: the
  Kelos review
  found that the Console re-sent read acknowledgments rejected with 403
  on every poll, scroll, and focus. The author
  stopped retrying 400, 403, and 404
  and kept conflicts, rate limits, server errors, and network errors
  retryable.
- PR Bound Session workload names #1629: the
  Kelos review
  noted, as an optional P3, that a foreign Service with a Session's
  name makes the reconcile return an error and retry forever, while a
  foreign StatefulSet fails the Session. ensureSessionService on
  main still returns that error.
- PRs Implement Axon controller #1 and Add Workspace CRD for reusable workspace configuration #37:
  Copilot
  found that a permanently invalid Task returned its Job build error
  and retried with backoff, and
  cubic
  found that a missing Workspace retried forever.
The October 14–16 window weighed Share Session read state across browsers #1822 against Bound Session workload names #1629 and Add Workspace CRD for reusable workspace configuration #37 and added
no rule, because those two were in reconcile paths and Share Session read state across browsers #1822 was a
client. Add Workspace options to restrict the checkout and agent credentials #1831 is a third reconcile-path finding, and a search of review
comments since January found Implement Axon controller #1 and fix(slack): stop retrying replies rejected with cannot_reply_to_message #1660. The request-failure rule
covered only the transient side: propagate a List/Get error so the
sender retries. In AGENTS.md and self-development/agentconfig.yaml,
a new rule after it covers the permanent side:
- report a failure that cannot succeed until its spec, configuration,
  or permissions change once, where operators see it, and stop retrying
  until that input changes;
- in reconcilers, return without an error and let a watch re-enqueue
  the object;
- keep conflicts, rate limits, server errors, and network failures
  retryable, and test both kinds.
A referenced object that does not exist yet is a wait, not a permanent
failure. main requeues a Task whose Workspace is missing after 2
seconds without an error, and Add Workspace options to restrict the checkout and agent credentials #1831's
fix
requeues a Task whose pool Workspace is missing. The rule exempts that
case and leaves 404 out of its examples. The reviewer prompts are
unchanged: the Kelos reviewer flagged Share Session read state across browsers #1822 and Bound Session workload names #1629, and no Kelos
review ran on Add Workspace options to restrict the checkout and agent credentials #1831.
Keep test outcomes independent of real time. Reviews on four PRs
found tests whose result depended on the real clock:
- PR Show individual background task durations in the console #1839: cubic
  found
  that a new test gave a pre-existing background task a hard-coded
  July 23, 2026 start time and compared it with a task started at
  time.Now(). A run before that date sorts the tasks the other way
  and fails. The author switched to a start time one minute before
  time.Now() in 953943e, which merged. The
  Kelos review
  at the same head (2580827) approved without this finding.
- PR Implement TaskSpawner CRD #22: cubic
  found
  that an idempotency check slept 500 ms instead of using
  Consistently, and
  Copilot
  found that a slow reconcile would let the check pass with
  idempotency broken. test/integration/taskspawner_test.go on main
  now uses Consistently.
- PR fix(worker-runner): kill task process groups #1579: cubic
  found
  that fixed sleeps could let the regression test pass before the
  background child forked, or fail on a loaded runner. The author
  replaced them
  with deadline-bounded polling in 5d8e8b3.
- PR feat(ghproxy): add singleflight to coalesce concurrent GET cache misses #868, which kelos-bot authored: cubic found two fixed sleeps,
  one that
  stood in for a wait condition
  and one that
  let cancellation run before the request reached upstream.
  Neither had a reply. The Kelos review approved at e8d95d2, which
  contains both sleeps, and the PR merged. cmd/ghproxy/main_test.go
  still sleeps 50 ms before releasing the upstream gate.
Copilot also flagged fixed sleeps on PR Implement Axon controller #1, and cubic found on PR Add cron source to TaskSpawner #40
that a future LastDiscoveryTime made a cron test check nothing.
Neither was confirmed fixed, so they are not used as evidence. No
existing rule covers this: base-agent.yaml says how to respond to
CI flakes, not how to keep a test independent of time. In AGENTS.md
and self-development/agentconfig.yaml, a new rule after the
test-goroutine rule requires:
- fixtures derived from time.Now(), or an injected clock, when code
  under test compares timestamps with the current clock;
- Eventually, Consistently, or a deadline-bounded poll on an
  observable condition instead of a fixed sleep.
The Kelos reviewer approved feat(ghproxy): add singleflight to coalesce concurrent GET cache misses #868 and Show individual background task durations in the console #1839 without these findings, so
kelos-reviewer.yaml now checks for them at P2.

Files updated:

AGENTS.md
self-development/agentconfig.yaml
self-development/kelos-api-reviewer.yaml
self-development/kelos-reviewer.yaml

self-development/base-agent.yaml is intentionally unchanged because none of
these changes are an upstream gjkim42/kanon-repo synchronization.

Which issue(s) this PR is related to:

N/A

Special notes for your reviewer:

The existing unassigned configuration PR was reused. It had become unmergeable
after #1762 deleted self-development/kelos-claude-reviewer.yaml and
self-development/kelos-claude-api-reviewer.yaml and folded their content into
kelos-reviewer.yaml and kelos-api-reviewer.yaml. This branch had made the same
additions to both files in each pair, so the branch was rebased onto main and
pushed with --force-with-lease. The rebase keeps the additions in the surviving
reviewers and drops the edits to the deleted files. The added lines in
AGENTS.md, self-development/agentconfig.yaml, and both surviving reviewers
are unchanged by the rebase.

Validation after the rebase:

make verify — passed
go test ./internal/examples/... — passed (self-development manifest, reviewer
prompt, and model/effort tests, including those rewritten by self-development: run every spawner on Claude Code #1762)
git diff --check origin/main HEAD — passed

Reviewed PR activity from September 17 through September 24. None of it
supported a new convention:

self-development: run every spawner on Claude Code #1762: the only inline finding was a cubic P3
about stale spawner counts left in the Open Actions README after sections were
deleted. The author fixed it in c982e7d and swept the sibling READMEs. The
existing "Keep documentation complete and internally consistent" rule already
requires updating all affected mentions together, and this is a single finding,
so no rule was added.
Render console Markdown while assistant responses stream #1752: the Kelos Claude reviewer approved with two optional P3 suggestions,
and cubic found no issues.
Use GPT-5.6 Luna for routine self-development tasks #1753, chore: ignore superpowers planning artifacts #1751, Document screenshot requirements for UI pull requests #1749: cubic found no issues.
feat: support optional skills packages #1726: the only new activity was a squash notice after earlier reviews, which
item 36 already reflects.
Shallow clone only the configured workspace branch #1711: a maintainer question about rebase compatibility, not a finding.
Add the TaskRouter CRD, controller, and TaskSpawner triggerMode #1761, Isolate Codex Session notifications by thread #1740: no review activity in this window.
Generated image-update PRs Update claude-code image to 2.1.278 #1754–Update cursor image to 2026.09.18-9a7762b #1758 and Update claude-code image to 2.1.281 #1763–Update cursor image to 2026.09.23-86fc751 #1767: no review comments.

Reviewed PR activity from September 24 through September 29. Only item 28's
broadening was supported by findings in more than one PR. The update is a
fast-forward commit; main has moved on only in files this PR does not touch,
so no rebase was needed. Validation for that commit:

make verify — passed
go test ./internal/examples/... — passed
git diff --check — passed