Repository navigation
feat(dashboard): the prompts tell a crewmate to compose its page, not pick one - #18758
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS The description says the per-milestone Watch
[DESIGN-REVIEWED] e8cdd18 Description read[DESCRIPTION-READ] 746c07cbefae7f40d52724d2146589d88e14ddb42aaacbd473bc0f4334bcac3a That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The description says the milestone artifact request "is gone", but the diff keeps it at all three sites, and one added clause contradicts the RFC it cites. Not justified as shipped
What this change shipsInventory (5 items) — 3 justifiedIntent: a conductor on a board tab stops writing a duplicate full-board artifact and reads one dashboard skill. ADDITION, implementing RFC
[FIRST-PRINCIPLES-REVIEWED] e8cdd18 Description read[DESCRIPTION-READ] c217c7593718ca4c67e5c5d81dbb50baf4a4a2cbc1d648a2b15d641051640de5 That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
a578a3e to
37605c1
Compare
|
GPT 6.1 F1 (upheld by adjudication) is fixed at the root on The finding was correct. The earlier head replaced the grant-derived universal What changed:
The adjudicator's other option, removing the three verbs from the grant tuple, Both arms were shown able to fail before the pass was trusted. Dropping Two CI failures on the previous head are fixed in the same push, both real and
1430 tests across four batches on the pushed tree, every command exit 0, plus The description is also rewritten so the whole |
37605c1 to
01244b4
Compare
01244b4 to
da29550
Compare
da29550 to
ee1fdab
Compare
b3848d1 to
ecbf089
Compare
ecbf089 to
95afe32
Compare
95afe32 to
a0074a5
Compare
a0074a5 to
75a51c4
Compare
Three prompt sites still described the view layer the way the catalog flow worked: a descendant publishes a `task-dashboard` HTML artifact and updates that slug at milestones. With the page now the owner agent's own HTML carried by the single `dashboard` skill, those sentences send the reader to the wrong skill for the page's rules. The rewrite branches exactly where `rfc-conductor-chat-board-on-tab.md` branches. Its "The `task-dashboard` artifact" bullet keeps the artifact on the no-tab branch -- "There it is still the only full board" -- so the artifact is dropped only where the `[DASHBOARD]` block names a board page (`goal-board` or `work-kanban`), and the request for it stays where none is named. A reader whose Dashboard tab shows no board therefore keeps the only full board they have, and no default changes. `agent.py`'s widget instruction keeps the widget as the milestone summary and names the Dashboard page as the full board on a board-page surface, loaded through the `dashboard` skill. The two conductor skills open their dashboard section with the one line the design gives them -- you own your dashboard page; it shows ledger evidence; workers report to you and do not publish that page -- defer the page, its flow and its honesty rules to that skill, and keep the sentence naming who publishes the artifact for the surface that still needs it. Both keep what is theirs: the no-file-writing role, the non-delegable jobs, the two columns of truth, and the rule that no dashboard justifies bypassing a tool approval or raising permission mode. The spec-byte goldens move for the two conductor specs those skills compose into; no other spec's digest or fileset changes.
75a51c4 to
e8cdd18
Compare
buluoray
left a comment
There was a problem hiding this comment.
Approve at e8cdd184b7.
What this brings users: a conductor whose Dashboard page is already the full board (goal-board / work-kanban) stops also writing a second full board as a task-dashboard artifact at every milestone, so the reader no longer sees every task twice and each milestone stops paying for that duplicate. Conductors without such a page keep today's artifact behaviour.
Verified against source
- The branch only fires where a board page really exists.
[DASHBOARD]comes fromdashboard_agentic.turn_block(src/kiro_crew/dashboard_agentic.py:722-754), injected only into a crewmate's member block (src/kiro_crew/context_assembly/member.py:544). It names the adopted template, or thedefaulttemplate when nothing was adopted (agent_panel.py:129,dashboard_templates/instance.py:450). Sogoal-board/work-kanbanare named only for a crewmate that adopted one; a plain session or an unadopted crewmate gets no board name and keeps writing the artifact, as before. - Nothing a reader relied on is lost. No panel reads a
task-dashboardartifact on main (docs/system-specs/modules/pipeline-conductor.md:25,artifacts.md:136), and with the Feature Preview off the tab draws thepanel_publishrecord instead (builtin_skills/dashboard/SKILL.md), which the goal-conductor skill still tells the conductor to publish. - Prompt-only: the other changes are the golden digests for the changed prompt bytes. The
dashboardskill is bundled underbuiltin_skills/, so the pointer resolves on every install.
CI on this head: 85 success, 13 skipped, 0 failure. I did not run tests locally; this is a prompt change, checked against source.
Non-blocking follow-ups
- Until #18755 lands, the
dashboardskill teaches how to change the page (search, preview, ask, apply), not "the page, its flow and its honesty rules" these prompts promise. A conductor following the pointer today gets the switching flow. [DASHBOARD]names the board even when the reader has the Feature Preview off, so "that page is the full board" can point at a page the reader cannot see. Thepanel_publishboard covers it, but saying so in the conductor prompt would avoid a conductor telling the user to "open the Dashboard page".
Problem / Motivation
Goal: one document tells a conductor how its Dashboard page works -- the
dashboardskill -- instead of three prompts describing a flow of their own.Three prompt sites still describe the view layer the way the old catalog flow worked: an authorized descendant publishes a
task-dashboardHTML artifact and updates that slug at each milestone. They are_CONDUCTOR_SYSTEM_PROMPTinagent.py,goal-conductor/SKILL.mdandpipeline-conductor/SKILL.md. Two of them also send the reader to theartifactsskill for the page's rules, so the rules live somewhere the dashboard work no longer maintains.Why it matters
A conductor asked for a board writes a second full board as an artifact while its Dashboard tab already shows one, and
rfc-conductor-chat-board-on-tab.md(in-progresson main) measures that cost: "the reader sees every task twice, and each milestone spends output tokens on a full widget plus atask-dashboardartifact update that the tab already shows." A reader who follows the pointer into theartifactsskill then gets page rules that are no longer the ones being developed.Not a goal
dashboardskill itself teaches is feat(skills): ship dashboard-manager and retire the template picker #18755's scope, not this PR's.dashboard_savedoes not exist on main, and nothing here promises it.task-dashboardtag, its route and its frame stay exactly as they are.What changed (motivation -> approach -> change)
Goal: a conductor reads one description of its own dashboard.
Approach. The alternative was to restate the page contract inside each prompt, which is how it drifted into three versions in the first place. Pointing each prompt at the one skill instead means the save and host PRs on #18642 change one document, not four.
Product shape: the decision to stop writing the artifact when the tab is the full board is recorded on main in rfc-conductor-chat-board-on-tab.md, status
in-progress, section 3 and its "Thetask-dashboardartifact" bullet.agent.py,_CONDUCTOR_SYSTEM_PROMPT: the chat widget stays the milestone summary, and the full board is now the Dashboard page, loaded through thedashboardskill. The request for a secondtask-dashboardartifact at every milestone is gone.goal-conductor/SKILL.mdandpipeline-conductor/SKILL.md: each dashboard section opens with the one line the design gives it -- you own your dashboard; a page shows ledger evidence; workers report, never publish -- and defers the page, its flow and its honesty rules to thedashboardskill. Each keeps what is its own: the no-file-writing role and the four non-delegable jobs in the goal conductor, the two columns that remain the source of truth in the pipeline conductor, and in both the rule that no dashboard justifies bypassing a tool approval or raising permission mode.test_agent_refactor_spec_bytes.py: 12 golden digest pairs move, forkirocrew-conductor.jsonandkirocrew-ledger-conductor.jsononly.Backwards compatibility
Compatible: no field, type, validator, tool, route or public function changes. Prompt text and two specs' measured bytes, nothing else.
The diff does take something away -- the instruction to publish a
task-dashboardartifact, at three of the four sites that carry it, and the prose pointer into theartifactsskill -- so every reader found by greppingorigin/main:website/src/pages/chat/command-center/useCommandCenter.ts:17(TASK_DASHBOARD_TAG, behindGET /api/artifacts?tag=task-dashboard) -- entry: crew page, chat -- the tag, the route and the frame are untouched, andartifacts/SKILL.md:58still teachesartifact_save(name=..., kind="html", tags=["task-dashboard"]), so an artifact that exists still renders. With none, the Overview shows the automatic card, which is already its documented behaviour (docs/feature-map/README.md:295).website/src/pages/chat/command-center/CommandCenterPanel.tsx:110andTaskDashboardFrame.tsx:31-- entry: crew page -- unchanged, behind thePREVIEW_DASHBOARDFeature Preview, default off.website/src/test/commandCenterPanel.test.tsxneeds no edit.src/kiro_crew/builtin_skills/artifacts/SKILL.md:29-58-- entry: chat, subagent -- the section and the skill'sname:are unchanged, so anything that loads it still gets the whole thing. What changes is that the two conductor skills stop pointing at it in prose; the section is left in place because the drop-templates PR on Dynamic Dashboard v3: drop built-in templates, compose a dashboard from a data-type catalog and store it as a dashboard artifact #18642 owns its removal.docs/system-specs/modules/pipeline-conductor.md:25-- entry: release -- its Components paragraph still describes the sentence this diff removes, including "published by an authorized descendant", which the new line inverts.docs_lint.pyexits 0 because nothing ties the two automatically. Correcting that paragraph is a one-paragraph edit and is the next push on this branch.Tests
test/test_agent_refactor_spec_bytes.pylocks the exact bytes of every packaged agent spec, so any prompt edit that reaches a spec has to be re-measured. 12 digest pairs moved, all of themkirocrew-conductor.jsonorkirocrew-ledger-conductor.json; no other spec's digest moved and no fileset changed, which is what says the two skills reached only the specs they compose into.No new test is added: the diff adds no behaviour to assert. The existing guards are what hold it, and they ran on this head with the repository venv,
--color=no,-p no:cacheprovider,KIROCREW_HOMEat a scratch directory outside the data home.test_agent_refactor_spec_bytes.pytest_skills.pytest_conductor_skills_steer_corrections.pytest_dashboard_agents_composition_contract.pyGates, each read by its own exit status, all 0:
flake8,black --check,check_builtin_skill_scope.py,docs_lint.py, andcheck_comment_history.pyrun after its own--testself-test, so the clean result comes from a checker proven able to see a violation.Manual verification
Both skills were read whole, not diffed alone, to confirm each keeps its own rules; every reader above was grepped on
origin/mainrather than inferred.Screenshots / video
N/A -- the diff is prompt text and one test file. Nothing under
website/src/is modified, so there is no rendered delta and no frontend surface changes.Related Issues
Refs #18642
The T3 one-pager for #18642 was approved, so this line opens its PRs.
The skill these prompts point at is rewritten in #18755.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
I confirm that this contribution is made under the terms of the Apache-2.0 license, per the Licensing section of CONTRIBUTING.md.