Skip to content

feat(skills): ship dashboard-manager and retire the template picker - #18755

Merged
iamwhatever merged 1 commit into
mainfrom
feat/18642-dashboard-manager-skill
Oct 14, 2026
Merged

iamwhatever merged 1 commit into
mainfrom
feat/18642-dashboard-manager-skill

Conversation

@chenmingwei23

@chenmingwei23 chenmingwei23 commented Oct 10, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: one dashboard skill that tells an agent how to write the Dashboard tab page somebody asked for, instead of telling it to pick one and that it may not write one.

The shipped dashboard skill teaches a catalog flow -- search, preview a template_id, apply -- and states in so many words "You cannot write the page yourself." The RFC amendment of 2026-10-12, merged to main in #19287, retires that flow and records that a crewmate composes its own page, and the base now serves an authored page through the frame. So that body describes a flow the product has left and forbids the one it has.

The rest an agent needs is scattered: honesty rules in the artifacts skill, the frame contract in a module docstring, the fold catalogue in a dev-only skill, theme in the widgets skill.

Why it matters

This skill is what every member prompt and the dashboard subagent charter send an agent to read. While it ships, the honest answer to "show me where this project stands" is "pick one of these twelve", and an agent that reads it refuses work it can do.

Not a goal

What changed (motivation -> approach -> change)

Goal: the agent writes the page, and one document tells it how.

Approach. A new skill under a new name, with this one retired, was the earlier shape of this PR and was worse: it deleted a skill two shipped prompts still name and needed a _RELOCATED_SKILLS migration row. Rewriting in place keeps the name, the existing triggers and every caller.

  • builtin_skills/dashboard/SKILL.md rewritten, same name: and two added triggers, now carrying the whole job in seven sections: (1) ownership, the stage -> preview -> approve -> apply gate, and the layout lock; (2) design freedom bounded by the honesty rules; (3) the manager view's five questions and one-screen limits; (4) the frame contract -- data-dashboard-field sets text only, window.kirocrew.fields for a chart, window.kirocrew.subscribe(name, cb) for anything drawn, connect-src 'none', the no-script fallback under an admin ceiling, the staleness band; (5) which values exist, through dashboard_types, the generated fold list and dashboard_fields; (6) theme and reach via the widgets contract; (7) the gallery.
  • builtin_skills/dashboard/GALLERY.md (new) lists gallery/'s whole packages beside the pages that shipped and the crew-drawer boards, with the reader question each answers. Every entry is optional and nothing resolves one by name at run time.

Section 1 teaches the gate that amendment's Decision 1 records rather than the behaviour that predates it: a save STAGES a page, it cuts no version and is fed no values, and the live path is stage -> preview -> a person approves -> apply, where apply takes no page argument so what lands is provably the page the person read. The undo is artifact_revert on the package artifact; there is no dashboard-specific rollback verb beside it. dashboard_save, the tool that stages an authored page, is named with the PR that ships it, #18851, so an agent reading this today knows which half is in its hands. The section also records why the gate may not simply be dropped: the no-wrapper exemption of 2026-10-11 was granted on it.

Section 1 also states what a save does on this base while dashboard_save is unmerged, because the earlier revision of this paragraph got it backwards. A save does not land the page: mcp_tools/artifacts.py tells the caller that saving a kind='dashboard' package BINDS the crewmate's page to it "while the page still draws from its template instance", and applied_package_version (dashboard/handlers/member_dashboard.py) serves only a version a PERSON cut. So the page an agent saved is not yet the page the person sees, and agentic field writes are refused package_display_pending until the cut happens.

The base's layout lock is carried into that same section rather than dropped by the rewrite, and it now names the gap an agent would otherwise walk into. The lock is a ceiling over the crewmate: a locked tab ignores a bound package and serves the template page, and dashboard_apply and dashboard_rollback answer dashboard_locked. artifact_revert, the undo this skill names, is NOT lock-aware -- it succeeds while nothing on the tab changes and no error says why -- so the skill tells the agent to read the [DASHBOARD] LOCKED block rather than infer the lock from a tool that answered yes. The base paragraph that described the "Dynamic Dashboard" Feature Preview switch is gone from main already, and this rewrite does not reintroduce it.

The dispatch tree is one gallery row on the same footing as the others, with a paragraph saying it is never the default, never assumed by a prompt or a test, and that a page showing no tree is correct.

check_builtin_skill_scope.py refuses a src/kiro_crew/... path or a repository URL inside a shipped skill, so the gallery uses skill-relative paths like gallery/dispatch-tree/package.json and names the shipped pages by their directory name.

Backwards compatibility

Compatible: the skill keeps its name: and every existing trigger, so every prompt and every trigger match resolves as before. It adds two triggers: build a dashboard, dashboard gallery. No field, type, validator, tool or public function changes, and nothing is deleted or relocated.

It does take sentences away from a shipped skill -- the catalog flow, and the base's own gallery section, whose substance moves into GALLERY.md and sections 4 and 7 -- so every reader of those, found by grepping the base:

  • Reader: agent.py's dashboard manager charter, which loads this skill -- entry: subagent -- still resolves, because the skill keeps its name. Rewritten in the prompt PR.
  • Reader: context_assembly/member.py item 7, injected on every member turn -- entry: chat -- same: the name it loads is unchanged. That item still names the catalog verbs; the prompt 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 is where it changes.
  • Reader: the dashboard_templates / dashboard_preview / dashboard_apply tools -- entry: subagent, crew page -- all three stay registered and granted; the skill stops teaching their flow, the verbs still resolve.
  • Reader: dashboard_rollback -- entry: crew page -- stays registered and granted; the skill now names artifact_revert as the page's undo, which is what Decision 1 of the merged amendment records.
  • Reader: the layout lock's dashboard_locked answer -- entry: crew page -- unchanged and still taught, moved into section 1 from the base's own section 4, with artifact_revert's lock blindness stated beside it.
  • Reader: an install that edited its own copy -- entry: release -- the mtime-preserving sync protects a locally edited body, so this rewrite does not clobber it.
  • Reader: the dev fold catalogue -- pointed at, not copied, so it stays the one list true today.

Tests

No test file changes. The rewrite is held by the guards builtin_skills/ already carries, and three of them bound this work: check_builtin_skill_scope.py, which is why the gallery cites no repository path; test_builtin_skill_index_line.py, which pins every packaged description at or under 300 characters because the session-start index truncates there, and this one measures 285; and test_skill_home_paths.py, whose LEGACY_HOME pattern \.kirocrew(?![-.]) reads a bare `window.kirocrew` as the pre-move data home, so section 4 names the object only in its dotted forms.

Run on this head with -n 2. Every command exited 0.

files result
test_skill_home_paths.py test_docs_lint_fact_checks.py test_builtin_skill_index_line.py test_builtin_skill_packaging.py test_builtin_skill_scope.py test_skill_budget.py test_builtin_skill_sync_safety.py test_skills.py test_skills_frontmatter.py test_skills_composition_contract.py test_brand_name_gate.py test_dashboard_gallery_dispatch_tree.py test_dashboard_gallery_real_tree.py test_dashboard_structure_lock.py 710 passed, 1 skipped
scripts/docs_lint.py exit 0, "All documentation checks passed"
scripts/check_builtin_skill_scope.py exit 0, "no repository markers in a shipped skill outside kirocrew-dev/"
scripts/check_brand_name.py exit 0

0 binaries, 0 non-ASCII bytes in either skill file.

Manual verification

Every claim was read off this base rather than written from memory: the frame's three value paths, the gallery's one package and twelve shipped pages, the layout lock's refusal code, and each tool the skill names. window.kirocrew.subscribe resolves on the base in dashboard_frame.py and artifact_store/dashboard_package.py, and gallery/dispatch-tree/package.json is on disk -- both arrived on main with the frame-host PR, which is why this PR no longer needs to point at an unmerged branch for either. PREVIEW_DASHBOARD and "Feature Preview" are 0 hits on the base, matching the rewrite.

Screenshots / video

N/A -- the diff is two markdown documents. Nothing under website/src/ is modified, so there is no rendered delta.

Related Issues

Refs #18642

The T3 one-pager for #18642 was approved, so this line opens its PRs.

Based on main. The two earlier bases this PR stood on are both merged: the frame host that puts window.kirocrew.subscribe and the gallery/dispatch-tree package on the tree, and the RFC amendment in #19287 that records the flow this skill teaches.

Checklist

  • At most two commits (one is the norm), with a Conventional Commits title (feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)
  • Existing tests pass and new tests added for new functionality
  • Self-review completed; code follows project style guidelines
  • Documentation updated (if applicable)
  • No secrets, credentials, or internal references in the diff

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.

@chenmingwei23
chenmingwei23 requested review from a team and KiroCrewCaptain as code owners October 10, 2026 06:58
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Oct 10, 2026
@github-actions

github-actions Bot commented Oct 10, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — ⏭️ skipped

Revision 88efc9122a1bb5a327cf606af6529b62c5173aa5 touches no user-facing surface (no changes under website/ or committed screenshots), so the UX review was skipped. Advisory — does not block merge.

@github-actions

github-actions Bot commented Oct 10, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — 🟡 CONCERNS

Premise-level review of 88efc9122a1bb5a327cf606af6529b62c5173aa5 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

First-Principles-Verdict: CONCERNS

The rewrite teaches tools this base doesn't have yet (dashboard_save, an agent-driven artifact_revert undo), while the prompts that load this skill still say the opposite.

Not justified as shipped

  • Item 1 — symptom-level, premise risk: the skill says "dashboard_save … ships in feat(dashboard): dashboard_save stages a composed page an agent builds #18851". On this base, saving only binds the page and agentic writes get package_display_pending. Meanwhile context_assembly/member.py:331-336 and agent.py:5789-5810 still say "You cannot write a page yourself" and send the agent to this skill. The RFC's second amendment says "Until the change that carries it lands, the catalog still serves". So on base, the request "show me another one" now ends in a page that never shows.
    Subtraction: defer the rewrite of sections 1 and 7 until feat(dashboard): dashboard_save stages a composed page an agent builds #18851 and the prompt PR land, and keep the base's catalog sections until then.
    Clears when: feat(dashboard): dashboard_save stages a composed page an agent builds #18851 is merged on the base, then push or rebase.
  • Item 3 — premise risk: "Going back is artifact_revert". But applied_package_version (member_dashboard.py:1213) only counts a cut whose actor is by == "user". An agent's revert is stamped agent, so it doesn't move the tab, and dashboard_rollback, which still works, is no longer taught.
    Subtraction: keep dashboard_rollback as the taught undo until the gate module ships.
    Clears when: an agent artifact_revert is shown changing the served page on base.
  • Item 5 — inherited: "One screen at 1280x800 with no scroll … limits rather than advice". I grepped both dashboard RFCs and docs/decisions/ and found no provenance. It also contradicts "The design is yours".
    Subtraction: delete the "Hard limits" list in section 3.
    Clears when: the list is deleted, or a recorded decision mandates it.

What this change ships

Inventory (7 items) — 4 justified

Intent: an agent writes the Dashboard page somebody asked for instead of picking one from a catalog. This is an ADDITION (it rewrites a shipped skill).

  1. The skill now tells the agent to write and stage its own page — symptom-level (the tools aren't on base)
  2. The catalog search/preview-by-id flow is no longer taught — justified
  3. The taught undo moves from dashboard_rollback to artifact_revert — symptom-level (the premise doesn't hold on base)
  4. A new GALLERY.md lists reference pages and what each one answers — justified
  5. A mandatory manager view: five questions, 1280x800, no scroll, headline of ten words or fewer — inherited
  6. Frame contract, honesty rules and theme rules are gathered into one skill — justified
  7. Two new triggers: build a dashboard, dashboard gallery — justified

[FIRST-PRINCIPLES-REVIEWED] 88efc91

Description read

[DESCRIPTION-READ] 84036d4f0572566af42f972668fe95dd380d10263ba672e311658706c0fbc23c

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

github-actions Bot commented Oct 10, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — 🟡 CONCERNS

Design-level review of 88efc9122a1bb5a327cf606af6529b62c5173aa5 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: CONCERNS

This skill ships before the tools it teaches: on this base the tools refuse an agent-written page, and the catalog flow that does work is no longer taught.

Watch

  • No working path in between. The skill's gate is "stage -> preview -> a person approves -> apply", and it admits that "dashboard_save … ships in feat(dashboard): dashboard_save stages a composed page an agent builds #18851". On this base, dashboard_preview still says "You CANNOT preview a page you wrote yourself" and takes only template_id (mcp_panel.py:237-252). So an agent following the skill can't run steps 1, 2 or 4. Its only fallback is "leave the cut to them". The catalog flow, which is the only one that works end-to-end today, has been removed from the skill. Result: a request like "show me another one" gets a description of a page instead of a working preview, until feat(dashboard): dashboard_save stages a composed page an agent builds #18851 and the prompt PR both land.
    Clears when: this PR merges after #18851 is on main, or the skill keeps a short "until dashboard_save ships" catalog path (templates/preview/apply) as the working route.
  • The prompt and the tools contradict the skill. The skill says "there is no dashboard-specific rollback verb", but dashboard_rollback is still registered and granted. context_assembly/member.py:332-333 still tells every member turn to "follow it: dashboard_templates to search … dashboard_rollback to go back". The Backwards-compatibility list names these readers, but its reason ("the verbs still resolve") skips the problem: on every member turn the agent gets two sets of instructions that disagree.
    Clears when: the prompt PR on #18642 lands in the same release as this rewrite, or the skill's undo wording matches what the base's tools actually do.

[DESIGN-REVIEWED] 88efc91

Description read

[DESCRIPTION-READ] 612aa95d2c4aa6a0c07fc5698e012ff8a61d7aee579c1e226559aeeb3533eb5e

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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 10, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 88efc9122a1bb5a327cf606af6529b62c5173aa5 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 88efc91

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 88efc9122a1bb5a327cf606af6529b62c5173aa5: <one-sentence reason>

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Oct 10, 2026
@github-actions

github-actions Bot commented Oct 10, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed 88efc9122a1bb5a327cf606af6529b62c5173aa5 — this comment is updated in place on each push.

Review details

No findings that block the merge; three advisory findings show the rewritten skill pointing the agent at tools that don't exist in this tree.

FINDING — src/kiro_crew/builtin_skills/dashboard/SKILL.md:37 — When a person says "I want one that shows X", step 1 tells the agent "A save writes the package as a STAGED copy", but dashboard_save does not exist in this tree. The only save is artifact_create with kind="dashboard", which BINDS the crewmate's page before anyone has approved it. From that point every dashboard_write on the existing page is refused with package_display_pending (dashboard_agentic.display_pending_refusal) → Fix: describe the save as binding, and tell the agent not to save a package until the person has said yes, until a staging tool lands.

FINDING — src/kiro_crew/builtin_skills/dashboard/SKILL.md:51 — "Go back" is sent to artifact_revert, and the skill says "there is no dashboard-specific rollback verb beside it". But dashboard_rollback is registered (mcp_panel.py:284), it is lock-aware, and it is the tool that changes the template instance the tab actually draws. Following the skill, the agent reverts an artifact and nothing on the tab changes → Fix: keep "go back" pointed at dashboard_rollback(to_version) while it is registered.

FINDING — src/kiro_crew/builtin_skills/dashboard/GALLERY.md:31 — For shipped pages it says "Nothing resolves one by name at run time", yet dashboard_preview takes exactly those ids (project-report, office, …, found in dashboard_templates/builtin/). The skill also no longer mentions dashboard_templates or dashboard_preview, the only path that can preview a page, since dashboard_preview refuses pages the agent wrote itself. So "show me another one" now leads to a preview step no tool can perform → Fix: keep the dashboard_templates → dashboard_preview → dashboard_apply flow documented, and correct the "resolves by name" claim.

[OPUS-REVIEWED] 88efc91

Verdict parsed from the review's SHA-scoped output markers for commit 88efc9122a1bb5a327cf606af6529b62c5173aa5.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 88efc9122a1bb5a327cf606af6529b62c5173aa5: <one-sentence reason>

@chenmingwei23
chenmingwei23 force-pushed the feat/18642-dashboard-manager-skill branch from 9946759 to dbf6234 Compare October 10, 2026 08:42
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Oct 10, 2026
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Oct 10, 2026
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-dashboard-manager-skill branch from dbf6234 to cbc0b1c Compare October 10, 2026 10:07
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Oct 10, 2026
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-dashboard-manager-skill branch from cbc0b1c to 9f3cf8e Compare October 10, 2026 12:18
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention merge conflict Branch has merge conflicts with its base — author must resolve before merge and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Oct 10, 2026
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-dashboard-manager-skill branch from 9f3cf8e to 9a9b0c8 Compare October 11, 2026 00:08
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Oct 11, 2026
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-tree-fold branch 3 times, most recently from 65b3b63 to de29606 Compare October 11, 2026 20:57
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-tree-fold branch 7 times, most recently from ef8c4cf to af93d98 Compare October 12, 2026 04:58
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-dashboard-manager-skill branch from 1601642 to ff8a1a2 Compare October 12, 2026 05:08
@github-actions github-actions Bot removed the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Oct 12, 2026
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-dashboard-manager-skill branch from ff8a1a2 to f96fe22 Compare October 12, 2026 05:25
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Oct 12, 2026
@github-actions github-actions Bot added the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Oct 12, 2026
@chenmingwei23
chenmingwei23 force-pushed the feat/18642-tree-fold branch 5 times, most recently from b29a25a to 6eb2109 Compare October 12, 2026 09:21
… start from

The `dashboard` skill taught a catalog flow -- search, preview a
`template_id`, apply -- and stated that the agent cannot write the page.
The RFC's amendment of 2026-10-12 retires that flow and records that a
crewmate composes its own page, and the base serves an authored page
through the frame, so that body described a flow the product has left and
taught none of the one it has.

It is rewritten in place, keeping its name and its triggers, and now
carries the whole job in seven sections: who owns the page and the gate it
goes through; design freedom with the honesty rules that bound it; the
manager view's five questions and one-screen limits; the frame contract,
including `window.kirocrew.subscribe(name, cb)` for anything the page
draws; which values exist, through `dashboard_types` and the generated
fold list; the theme and reach contract; and the gallery.

Section 1 teaches the gate the amendment's Decision 1 records: a save
STAGES a page, and the live path is stage -> preview -> a person approves
-> apply, with `artifact_revert` on the package artifact as the undo and
no dashboard-specific rollback verb beside it. `dashboard_save`, the tool
that stages an authored page, is named with the pull request that ships
it, so the section describes the recorded contract rather than the
pre-gate behaviour. It also states what a save does on this base in the
meantime: it BINDS the page, the tab keeps drawing the template instance,
no version is cut, and agentic field writes are refused
`package_display_pending` until a person cuts one -- so the page an agent
saved is not yet the page the person sees.

The base's layout lock is carried into the same section rather than
dropped by the rewrite, and it names the gap an agent would otherwise walk
into: the lock is a ceiling over the crewmate, so a locked tab ignores a
bound package and serves the template page, `dashboard_apply` and
`dashboard_rollback` answer `dashboard_locked`, and `artifact_revert` is
NOT lock-aware -- it succeeds while nothing on the tab changes and no error
says why. Read the `[DASHBOARD]` block rather than inferring the lock from
a tool that answered yes.

`GALLERY.md` lists `gallery/`'s whole packages beside the pages that
shipped, with the reader question each answers. Every entry is optional
and nothing resolves one by name at run time: they are source to read,
copy and adapt. The dispatch-tree package sits among them on the same
footing, never the default and never assumed, and the three things worth
taking from it are named: bind a cell or subscribe, build rows with
`textContent`, and declare every field the markup binds.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants