Skip to content

CHANGELOG.md

Latest commit

 

History

History
353 lines (246 loc) · 45 KB

CHANGELOG.md

File metadata and controls

353 lines (246 loc) · 45 KB
Raw

Changelog

All notable changes to Agent Platform Advisor are documented in this file.

The format is based on Keep a Changelog, organized by repository commit date.

2026-08-19 (later)

Changed

  • Restored the approved conditional runtime-ownership experience. The custom-agent path remains five scored questions with a 15-point maximum. Only when Copilot Studio and Foundry are the top two viable platforms within 2 points does the advisor ask who should operate the runtime. Microsoft-managed operation prefers Copilot Studio without changing raw scores; engineering-owned runtime hard-disqualifies Agent Builder and Copilot Studio. This replaces both the temporary scored q4f runtime option and the universal optional governance-constraints screen.
  • Preserved shared-link compatibility. Links without q9 remain valid, legacy q9b/q9c normalize to the Microsoft-managed answer, and temporary q4f links normalize to complex orchestration plus engineering-owned runtime.
  • Kept Copilot Studio harness guidance independent from platform scoring. Copilot Studio winners still receive GitHub Copilot, workflow, Copilot chat, or standard harness starting guidance based on the five functional answers.

2026-08-19

Added

  • Copilot Studio harness starting guidance. When Copilot Studio wins the existing five-question assessment, the recommendation card now identifies the most relevant implementation starting point without changing scores: GitHub Copilot harness for adaptive multi-step orchestration and content creation, Copilot Studio workflow for deterministic triggered automation, Copilot chat harness for internal Microsoft 365 Copilot knowledge extensions, and standard harness for predictable topic-driven conversations. The guidance includes harness immutability, preview/lifecycle, and Copilot Credit cautions and is covered by Playwright tests.

Changed

  • Reconciled the harness guidance with the v1.4 scoring model. Runtime ownership remains represented by the scored q4f code-first/custom-runtime option rather than adding a second conditional governance question. This preserves the current 2,304-combination calibration while sharpening the recommendation after Copilot Studio wins.

Fixed

  • Stabilized the markdown export end-to-end test under full-suite concurrency. The test waited for and read the downloaded file before checking a two-second confirmation label, so a busy parallel run could correctly complete the download but inspect the label after it had reset. It now verifies the immediate click feedback before doing slower file I/O.

2026-08-12 (later)

Changed

  • Refreshed .github/copilot-instructions.md. The changelog convention it documented had gone stale in a way that would actively undo work: it told future sessions that work in progress belongs under ## Unreleased, a section that was deliberately removed, so anyone following the instructions would have recreated it. It now documents the real convention (dated by commit date, newest first, (later) for a second same-day batch) and says not to add ## Unreleased. Added a "Share links are untrusted input" subsection to the share-link contract, since two security bugs have now come from treating URL params as trusted — per-question option-ID validation (hard rules fire on option ID alone, so an ID smuggled into the wrong question changes the winner), the ^\d{8}$ guard on d=, the rule that any path setting fastTrack must also force card mode, and a preference for DOM nodes over innerHTML wherever a URL value is involved. Also recorded that og:description must equal the on-page welcome copy, that meta.last_updated is hand-maintained and must be bumped when content changes, that npm test needs PyYAML before Playwright ever starts, and the two recurring test traps (loose regexes that pass on generic text, and probes outside tests/e2e/ that Playwright silently ignores). Refreshed the spec list, which named 7 of the 16 files.

Removed

  • Deleted .github/improvements.md. It was a productization wishlist, and every item on it has since shipped: the version strip and footer "Last updated" indicator, the loading/welcome/error states, the stepped progress bar, markdown export, the "Did this help?" feedback loop, the prominently linked public changelog, and a README that leads with why the tool exists. Of the few items never built, three were deliberately declined rather than forgotten — a vanity domain would make a Microsoft-org tool read as unofficial third-party work, email-self needs a backend this app deliberately does not have, and a separate landing page would add a click before value when the welcome card already names the product, the problem, and every destination. The remaining item, whether to claim an aka.ms link, is a process question rather than code. Keeping a stale wishlist around implies work is outstanding when it is not. TODO.md P2.5 and P2.7 still record the items and their provenance.

Changed

  • Fixed the social/SEO metadata, which was still serving the old build-framed pitch. The welcome copy was corrected on the page, but og:description kept the previous sentence ("we'll recommend the best Microsoft agent platform for your scenario"). That is the text that unfurls in Teams and Slack and appears in search results, and since share links are a core feature here, more people meet this app through the unfurl than through the start page. og:description now carries the on-page sentence verbatim. Two further problems surfaced in the same block: name="description" was declared twice with identical content, so crawlers picked one non-deterministically, and og:title read "CAT Agent Platform Advisor" while both <title> and og:site_name read "Agent Platform Advisor" — a leftover from the shared Power CAT tooling whose logo and favicon were already stripped in an earlier release. Guarded by tests/e2e/meta-tags.spec.js, which pins the description meta to exactly one, asserts og:description equals the rendered welcome copy, requires the metadata to name both the "get work done" and "build" paths, and holds og:title/og:site_name to the real product name. All four proven non-vacuous against the old head.

  • Reworded the welcome description so it matches the two paths the app actually offers. It read "Answer a few questions about what you want from agents, and we'll recommend the best Microsoft platform for your agent needs" — build-framed copy sitting directly above a preview grid split into Use agents and Build agents, and one screen ahead of a prescreen that offers "find the right place to get work done" as its first option. The page contradicted the element beneath it and quietly presupposed the reader was building something. It now reads "Answer a few questions and we'll point you to the right place to get work done with an agent, or the right Microsoft platform to build one," which keeps "platform" on the build half where it is accurate (Agent Builder, Copilot Studio, Foundry are platforms; Microsoft 365 Copilot, Cowork, and Scout are destinations) and reuses the verbs already used by the prescreen options and the grid labels, so the start page stops introducing a third vocabulary for the same fork.

Security

  • Fixed a DOM XSS in the temporal-change banner (CodeQL js/xss, high). The banner was built with innerHTML while interpolating formatDateDisplay(originalDate), and originalDate is the d= share-link parameter. formatDateDisplay parsed the month and day as integers but returned the year raw via substring(0, 4), so four attacker-chosen characters reached the DOM as markup. Reproduced in Chromium: ?d=<img0101 injected a real IMG element, and because the injected tag absorbed the following text as attributes, it swallowed the "Retake assessment" link — the banner lost its only control. Share links are pasted around by people who did not build them, which is exactly the delivery path this needs.

    Fixed in three layers, so no single mistake reopens it: d= is rejected at the URL boundary unless it matches ^\d{8}$; formatDateDisplay now validates the whole string, range-checks the month and day, rejects dates that don't exist (20260231), and never returns input-derived text; and the banner is assembled from DOM nodes with textContent plus a real anchor, so the date is never parsed as markup. The inline onclick became an addEventListener. Guarded by tests/e2e/temporal-banner-xss.spec.js, which asserts the anchor is the only element the banner may contain, that the retake link survives every payload and still works, and that impossible dates are refused. All four tests were proven non-vacuous by restoring the vulnerable code.

2026-08-12

Added

  • "Last updated" date in the footer, directly below the Created by credit. Driven by a new meta.last_updated key in apa.yaml rather than hardcoded in the markup, so the date sits with the content it describes. It is distinct from guidance_verified, which states when the guidance was last checked against Microsoft Learn. Rendered via <time datetime="..."> for machines and a long-form label for people; the ISO string is parsed as UTC because new Date('2026-08-12') renders as August 11 in any timezone west of Greenwich. Covered by tests/e2e/footer-last-updated.spec.js, which asserts the visible label matches the datetime value and that the stamp sits between the credit and the doc links.

Fixed

  • The markdown export could leave HTML tags in its output. Tag stripping was a single replace(/<[^>]+>/g, '') pass, which CodeQL flagged as incomplete multi-character sanitization. Two bypasses exist and only one is closed by looping: a nested tag (<scr<b>ipt>) exposes a fresh tag once the inner one is removed, while an unterminated tag (<b><script) has no closing >, so the pattern never matches it and it survives any number of passes. The unterminated case is the one CodeQL reported. The new stripHtmlTags() removes complete tags until the string stops changing, then sweeps orphan angle brackets, so output can never contain < or >. Not exploitable as shipped — the input is apa.yaml content authored in this repo and the sink is a downloaded .md file, never innerHTML — but the replace was failing at its actual job of reducing display markup to plain text. Guarded by both bypass shapes in tests/e2e/coverage-gaps.spec.js.

  • A crafted share link could silently change the recommendation. Option ids from the URL were validated against one flat set of every option in apa.yaml, but hard rules key off the option id alone — getZeroedPlatforms iterates Object.values(answers) with no question context. So ?q1=q3f was accepted as an answer to q1: it scored zero (q.options.find returns undefined for the wrong question) yet still fired q3f's disqualification. Measured live, that turned "Agent Builder — Strong fit" into "Copilot Studio — Good fit" with no visible sign anything was wrong. Each q* param is now checked against only its own question's option set, and a mismatch routes into the existing schema-drift path, so the user sees the drift notice instead of a confidently wrong answer.

  • ?ft=1&mode=wizard let M365 Copilot into the scored wizard. The legacy fast-track branch returned whatever mode the URL asked for while leaving fastTrack = true — and fastTrack is what suppresses zeroed['m365_copilot']. Measured, getZeroedPlatforms returned {} on that path, so a platform that is never supposed to compete in the scored wizard could win it. ft=1 now forces card mode, matching how dt= already behaved. Both fixes are guarded by tests/e2e/share-link-integrity.spec.js.

  • npm test no longer fails cryptically on a fresh machine. scripts/golden_paths.py imports PyYAML, which the README mentioned but nothing installed, so a new clone hit a bare ModuleNotFoundError before Playwright ran. Added requirements.txt and an import guard that names the fix. Corrected the README's stale "8 spec files" count (13).

  • The "Was this helpful?" confirmation is now announced and keeps focus. Activating Yes or No removes the buttons, so focus dropped to <body> (measured) and the replacement text had no live region — a screen-reader or keyboard user got no confirmation that anything happened, and the next Tab restarted from the top of the document. The confirmation is now role="status", and it is unhidden before its text is written, since a live region only announces changes made while it is already in the accessibility tree. Focus moves to the confirmation only when the button was activated from the keyboard (:focus-visible), so a mouse click is not hijacked. Covered by tests/e2e/feedback-a11y.spec.js across the mouse, keyboard, and No-path cases.

  • Explore cards are clickable edge to edge. Each card on "Explore ways to use or build agents" lifted and brightened on hover — the universal signal for "click me" — but only the "Explore resources →" text at the bottom was a link. Measured live, the highlight covered 166,076px² while the click target was 6,415px², a 26× gap, and cursor stayed auto, so nothing corrected the promise except the click failing. All 9 cards have exactly one destination and no competing action, so the anchor now stretches over the card via ::after. The anchor stays the focusable element, so keyboard order and screen-reader link text are unchanged; :focus-within gives keyboard users the same highlight mouse users get. New tests/e2e/explore-card-links.spec.js probes each card's center with elementFromPoint and asserts the single-link precondition the pattern depends on.

  • Entry-point involvement tip corrected. The tip read "many people start hands-on in Copilot Chat, then hand the finished multi-step job to Cowork" — you can't hand off a job that's already finished; "finished" describes what Cowork returns. It also asserted a usage habit ("many people") for a product that reached GA on 2026-06-16, two months before this build's guidance_verified date, and named "Copilot Chat" unqualified on the routing screen, where it reads as a fourth destination and collides with the separate Microsoft 365 Copilot Chat SKU. Now matches the phrasing apa.yaml already uses at lines 698 and 702. Golden-path test tightened to assert the scoped product name and the handoff direction, instead of passing incidentally on a substring.

  • Design review round 3 (v3): eight findings from /plan-design-review against the main...v3 diff, measured on the live page rather than by inspection.

    • The result card no longer argues against itself before justifying itself. On the Agent Builder result, four advisory blocks (the SharePoint callout plus three adjacent-path notes) rendered above "Why this was recommended", so the payoff screen read as a list of alternatives before it gave a single reason. Reasoning now precedes the adjacent notes; the one answer-specific conditional callout stays above it, since it fires precisely when the winner may be the wrong call.
    • Colored left-border cards removed — the pattern was back six times on one screen. .rec-callout, .rec-adjacent-note, .delegate-tip, and .rec-spotlight each used border-left: 3-4px solid <accent>, the anti-slop pattern d9e9e88 had just stripped from the prescreen cards. Because all six looked identical, none of them carried more weight than the others. Emphasis is now tint + label weight: accent-tinted for conditional callouts, neutral surface for adjacent notes.
    • Five controls shipped under the 44px touch-target minimum. Measured live: feedback Yes/No 35px, Start Over 36px, constraints Back 41px, Download markdown 43px. Root cause was .btn never having a min-height, so every new button on this branch inherited the gap while earlier rounds patched links and .btn-decision one at a time. Fixed on the base class.
    • The constraints multi-select stopped drawing radio buttons. Rows are role="checkbox" with correct aria-checked, but the indicator was border-radius: 50% — the universal "pick exactly one" shape — so sighted users had no reason to select a second constraint. Multi-select cards now use a square box with a checkmark.
    • Two buttons that did the same thing collapsed into one. The constraints step offered "None of these — continue" and "See recommendation"; with nothing selected they were identical. One CTA now relabels itself based on selection state.
    • "Download markdown" confirms it worked. The export triggered a blob download with no visible response, so on browsers that hide download chrome the only recourse was to click again and produce duplicates. The button now reads "Downloaded ✓" for two seconds.
    • The guidance version strip appears once. It rendered on welcome, results, and footer — the last two roughly 100px apart on the same viewport. Kept on the result card, where freshness actually changes a decision.
    • "Why this was recommended" binds to its own list. After the reorder the heading sat 16px below the callout and 21px above its list, floating between sections; now 24px above and 8px below.

  • Design review round 2 (v3): seven findings from /design-review, audited against docs/DESIGN.md.

    • Prescreen routing cards no longer carry off-system color. The three cards were tinted indigo (#5B5FC7/#9EA2F0), --success green, and --warning amber via icon badges and 3px colored left borders. Indigo isn't in the palette and violet-as-AI-signal is on the anti-slop list; --success/--warning are state colors, not routing colors; colored-left-border-on-card is itself an anti-slop pattern. Badges are now neutral and take the single blue signal on hover/focus.
    • Option card text is left-aligned. Cards inherited text-align: center from .card-content--centered, so copy was centered beside a left-aligned icon with no shared left edge to scan down.
    • Mobile header no longer wraps to three rows. At 375px the progress bar wrapped and pushed the theme toggle onto its own orphan row. Logo + toggle now hold one row, progress bar takes a full-width row below.
    • Entry-point question titles moved onto the type scale. .delegate-question-title used the --fs-body-lg body token (18px), leaving the question barely larger than the paragraph above it; now --fs-subhead. The scored wizard's .question-title keeps --fs-heading.
    • Last two off-token font sizes removed. .rec-platform-name (the winning platform, the most important string on the results page) was a literal 1.5rem, below the heading token — now --fs-heading. The mobile .question-title shrink to 1.375rem is deleted outright, since DESIGN.md permits mobile reduction only for display/title.
    • Per-question fit grid no longer encodes state by opacity alone. Strong/moderate/weak were one blue at 100/55/25% in same-size dots, labelled only by a title tooltip (invisible on touch, unreliable for AT). Dot size now varies (12/9/6px) and every cell carries role="img" + an aria-label naming question, platform, and score.
    • Prescreen prose capped at a readable measure. The entry-point intro ran ~114 characters per centered line against DESIGN.md's ~70ch cap; now max-width: 60ch.

  • Design review (v3): dark-mode accent used as text now uses --primary-text (#2B9AEE) so progress labels, Explore group titles, badges, and decision chrome meet WCAG AA (fills stay #0078D4). Theme toggle and logo hit targets ≥44px; guidance strips no longer repeat Changelog (footer nav only); stronger :focus-visible on path cards and decision buttons.

Added

  • TODO.md backlog for the mid-2026 guidance & scoring overhaul. Prioritized P0–P3 work from a deep analysis of copy, routing, and scoring against current Microsoft Learn.
  • P0 correctness wave implemented (routing, taxonomy, scoring, one-liners, docs).
    • Adjacent build paths in Explore and recommendation footnotes: SharePoint agents, Microsoft 365 Agents Toolkit (declarative), custom engine agents (apa.adjacent_build_paths + per-card adjacent_paths).
    • Q4f scored option: code-first multi-agent / custom runtime → Foundry strong; hard-zeros AB + M365.
    • Playwright coverage for recurring/event + M365 → Cowork, recurring + cross → Scout, always-on + M365 → Cowork, always-on + cross → Scout, Explore adjacent paths, and dt=cowork share links.
  • P1 decision-quality wave
    • Conditional result_callouts on primary scored cards (SharePoint agents tip; Agents Toolkit for pro-dev + M365 chat + actions; Toolkit when AB wins for pro-dev).
    • Golden-path harness: scripts/golden_paths.py + tests/e2e/golden-paths.spec.js (G01–G12); npm run test:golden / npm test runs the script first.
    • Entry-point delegate tips (Chat then Cowork; interactive specialized vs Cowork research package).
    • Richer winner mismatch notes (Foundry+maker, CS+biz+APIs/workflows, AB+department audience, Scout Frontier access on entry-point results).
  • P2 coverage / trust / maintainability
    • Optional governance constraints multi-select after the five scored questions (soft boosts only; share param c=id1,id2; never hard-zero).
    • Guidance version + Learn verified strip on the result card (meta.version 1.4, meta.guidance_verified); the Changelog link lives in the footer only.
    • Markdown export of recommendation + scores + share URL; lightweight Was this helpful? feedback (Clarity tags + GitHub issue deep link).
    • Cross-question notes: maker+custom RAG; pro-dev+simple Q&A → AB/Toolkit guidance.
    • Welcome copy frames use vs build lanes before Get Started.

Changed

  • Entry-point Cowork ↔ Scout routing rebuilt (P0.1). Cadence options are now one-shot / recurring-or-event / always-on Autopilot / unsure. Reach is primary for recurring and always-on work: M365-only → Cowork; desktop/browser/local/shell → Scout; undecided → both. Daily briefings and M365 event triggers no longer mis-route to Scout. Legacy option ids ondemand/continuous still accepted in resolveDelegateResult.
  • Agent Builder taxonomy clarified (P0.2). Builder is framed as no-code declarative agents inside Microsoft 365 Copilot only — no Actions; migration lines point to Copilot Studio, Agents Toolkit, and Foundry custom engine. Scoring still hard-zeros AB for actions/background/external/custom app/custom RAG without giving Builder Toolkit-only credit.
  • Copilot Studio vs Foundry rebalanced after multi-agent GA (P0.3). Q4d (low-code multi-agent / long-running business orchestration) is now CS:3 / Foundry:2; Q4f covers code-first multi-agent / custom runtime (CS:1 / Foundry:3). Foundry wins on runtime ownership, not the keyword “multi-agent.” Distribution re-documented (~2,304 combos: CS ~81.3% / Foundry ~16.1% / AB ~2.6%).
  • Start-page, Explore, and platform one-liners aligned (P0.4) with recommendation truth (M365 Chat+agents+Agent Store; Cowork schedules/triggers; Scout Frontier gates; CS actions/governance; Foundry managed runtime; CS computer use vs Scout personal Autopilot).
  • Docs sync (P0.5): docs/SCORING.md, docs/FLOWCHART.md, README.md, .github/copilot-instructions.md; fixed stale “score 0–5” comment in apa.js (Not recommended is 0–3).
  • Runtime signal via Q2 labels (P1.3): q2a/q2b wording emphasizes Copilot orchestrator vs own endpoints/models/runtime (no new Q6).
  • Licensing honesty (P1.4) on AB/M365/CS/Foundry/Cowork/Scout watch-outs (seats, Copilot Credits, Azure consumption, Frontier gates).
  • GA/preview honesty (P1.5): CS computer use GA; new experience/memory/skills preview; Foundry IQ connect preview; Agent Optimizer preview; Foundry Local / Azure Local wording.
  • Why-not / valid_pairs (P1.6): dropped dead scored-path pairs with always-zeroed m365_copilot; stronger CS+Foundry and AB+CS pair rationale + why_not templates in computeWhyNot.
  • Earlier mid-2026 content refresh (still accurate under P0): Cowork scheduled prompts, Foundry Agent Optimizer / Foundry Local, CS multi-agent GA callout, M365 Agent Store + MCP Apps UI — now extended by the routing and scoring changes above.

For contributors

  • Explore group titles and descriptions moved out of renderExploration() and into apa.yaml (exploration_groups, exploration_adjacent_group), matching the rule that user-facing platform copy never lives in JS. Also removed two dead paths found in review: the unreachable rec.spotlight branch in renderCard() (Explore cards settled on one link each) plus its 41 lines of orphaned CSS, and skipConstraints(), which had zero references. A new structural test asserts all three Explore lanes, their exact titles, non-empty descriptions, and per-lane card counts, so a missing YAML key fails loudly instead of silently rendering an empty lane.

2026-07-24

Changed

  • New favicon built from the app's own robot-face mark. The previous favicon.png was the shared CAT icon at an odd 232×193, declared as type="image/x-icon" though it was a PNG. The replacement is the header logo's robot redrawn as a solid white head on the #0078D4 tile, with the eyes and smile knocked out in blue and the side nubs kept as filled bars. Solid shapes were necessary: the outlined head reads as a robot at 32px but collapses into mush below ~24px, and dropping the outline entirely (so the tile acts as the head) reads as a generic smiley rather than an agent. Ships favicon.svg (scalable, used by Chrome/Firefox at every size), 16px/32px PNG fallbacks, and a full-bleed 180px apple-touch-icon.png with square corners, since iOS applies its own mask.
  • Repointed the Open Graph URLs at the live site. og:image, twitter:image, and og:url all pointed under microsoft.github.io/cat/agent-platform-advisor/, which is now only a meta-refresh redirect stub — the app is served from microsoft.github.io/AgentPlatformAdvisor/. The old og:image path 404s, so link previews had no image to fetch regardless of which file it named; og:url also advertised the stub as the canonical URL. All three now use the AgentPlatformAdvisor host.
  • Replaced the Open Graph preview image with a screenshot of the app. The card previously pointed at the shared powercattools.png CAT logo, which said nothing about this tool; links now unfurl to images/og-image.png, a 1200×630 capture of the start page showing the six platforms the advisor chooses between. Corrected og:image:height (was declared as 1200 for a 1200×1200 image), added og:image:type and a descriptive og:image:alt, and added twitter:card="summary_large_image" so X/Twitter renders the wide card instead of a thumbnail.
  • Refreshed README.md and .github/copilot-instructions.md for the merged Microsoft 365 Copilot destination: the entry-point wizard now lists three destinations plus a starting surface, the share-link section documents dt=m365_copilot / st=chat|agents and the legacy ft=1 / dt=copilot_chat links, and the test count is corrected to 39. The Copilot instructions also gained a local-dev note (the app must be served over HTTP or the apa.yaml fetch fails), a share-link parameter table, a spec-file map, single-test commands, and corrected design facts (IBM Plex fonts, warm charcoal canvas, rem --fs-* tokens with a 12px floor) that had drifted from docs/DESIGN.md.
  • Entry-point recommendation cards now open their accordions by default. Microsoft 365 Copilot, Cowork, and Scout are single-card results with nothing to compare against — the card is the page — so Best For, Important Considerations, and the capability list start expanded instead of hiding the substance behind three clicks. Scored platform cards (Agent Builder, Copilot Studio, Foundry) stay collapsed so the comparison stays scannable. Driven by a new ENTRY_POINT_PLATFORMS constant in apa.js, replacing the platformId === 'm365_copilot' special case.
  • Merged the copilot_chat destination into m365_copilot. Copilot Chat is a surface of Microsoft 365 Copilot — as is each built-in agent — so modeling it as a sibling destination to Cowork and Scout claimed a product boundary that doesn't exist (and, after the headline rename, produced two identically-titled cards). The entry-point wizard's hands-on path now always resolves to the single Microsoft 365 Copilot card.
    • The task type answer no longer picks a destination; it picks which surface the card tells you to Start Here with, rendered through the previously unused spotlight slot: general help → Copilot Chat, specialized task → built-in agents (Researcher, Analyst, Facilitator, Interpreter).
    • Added recommendations.m365_copilot.start_here.{chat,agents} to apa.yaml and folded the old copilot_chat guidance into m365_copilot's new best_for / watch_out_for; deleted the copilot_chat recommendation block.
    • resolveDelegateStart() in apa.js derives the surface; buildPlatformCard() takes a startKey and labels the spotlight Start Here (a static spotlight still renders as Featured Capability).
    • Share links carry the surface as &st=chat|agents. The legacy ?dt=copilot_chat link still works and resolves to Microsoft 365 Copilot with the Copilot Chat surface featured.
    • Cowork's and Scout's "use X instead" notes now point to Microsoft 365 Copilot (Copilot Chat) rather than a sibling product. Updated docs/SCORING.md, docs/FLOWCHART.md, the prescreen copy, and tests/e2e/delegate-path.spec.js (39 tests passing).
  • Renamed the copilot_chat destination headline to "Microsoft 365 Copilot" (was "Copilot Chat"), which makes it share a headline with the m365_copilot destination. Updated tests/e2e/delegate-path.spec.js so the two are no longer distinguished by headline substring alone: a new expectPrimaryCard() helper asserts the exact .rec-platform-name text plus a description phrase unique to each destination ("conversational Microsoft 365 Copilot experience" vs. "Built-in, permission-aware AI across Microsoft 365"), so the two paths can't silently swap.
  • Rebuilt the type scale on rem tokens and raised every size one step. An audit found the CSS had drifted a full step below the scale documented in docs/DESIGN.md — body copy shipped at 14px (20 rules) against a documented 15px, captions at 12px, and badges/eyebrows at 10-11px — and every one of the 74 font-size declarations used px, so a reader's browser font-size preference had no effect.
    • Added --fs-display … --fs-mono-sm rem tokens on :root and html { font-size: 100% }; all 74 declarations now reference a token.
    • Sizes move up one step: body 14 → 15/16px, caption 13 → 14px, mono 12 → 13px, and everything at 10-11px comes up to a 12px floor (.sc-badge, .pq-legend, .rec-spotlight-eyebrow, .exploration-card-spotlight-eyebrow).
    • Removed the two responsive overrides that shrank text on small screens (.progress-bar → 11px at 768px, .sc-badge → 10px at 480px).
    • Capped running prose at 70ch — the 1024px container was producing ~95ch lines.
    • Updated the docs/DESIGN.md type scale with the token names plus rules on rem-only sizing, the 12px floor, no mobile shrinkage, and measure.
  • Made the start-page platform tiles read as informational, not clickable. They were styled as bordered, filled cards with a hover border-color change, which implied they were selectable. Removed the hover treatment and the card chrome (filled background, full border, rounded corners) in favor of flat, center-aligned entries separated by a hairline top rule, plus cursor: default. Added a lead-in line — "Here's what the advisor chooses between — select Get Started below to find your fit." — so the section reads as a preview of the destinations rather than a menu.
  • Center-aligned the start-page platform previews — icon, title, and description all sit on a shared center axis (.platform-preview-icon is a centered flex box; .platform-preview is text-align: center).
  • Increased the platform preview card title from 16px to 20px, matching the subhead type token in docs/DESIGN.md, so platform names read as card titles rather than body copy.
  • Increased the delegate group label size from 11px to 13px (weight 500 → 600) so the "…" section dividers above the delegate platform grid are legible at a glance.
  • Swapped the teal signal color for Microsoft blue at the user's request: the accent used for selected options, progress, the winning platform, focus rings, and primary CTAs is now #0078D4 (hover #2B9AEE, dim #0B5187) in dark mode and #005A9E in light mode. Warm charcoal canvas, neutrals, typography, and the no-glow rule are unchanged; --success / --warning / --error semantics are untouched. Updated docs/DESIGN.md.

2026-07-22

Changed

  • Replaced the two-question delegate path with an entry-point wizard ("Where should you get this work done?") that helps end users choose between Copilot Chat, Microsoft 365 Copilot's built-in agents, Cowork, and Scout from work patterns instead of product names — addressing the pain point of Microsoft asking users to pick among too many entry points.
    • New first question — Involvement: stay hands-on and iterate vs. hand it off to an agent.
    • Hands-on now asks a follow-up — Task type: general help (→ Copilot Chat) vs. a specialized task (→ Microsoft 365 Copilot built-in agents: Researcher, Analyst, Facilitator, Interpreter, …).
    • Hand it off asks Cadence + Reach (→ Cowork, Scout, or both). The wizard uses progressive disclosure: each follow-up is revealed only when its branch is chosen (smooth grid expand, respects prefers-reduced-motion), and collapsed follow-ups leave the tab order — no dimmed/greyed dead content. Reach is now gated behind Cadence — selecting "Hand off the whole task" reveals only the Cadence question; the Reach question ("Where does it need to reach?") appears once a cadence is picked, so both delegate follow-ups no longer show at once.
    • Removed the standalone "built-in Microsoft 365 Copilot experience" prescreen tile — that destination is now reached through the wizard, eliminating the naming overlap with Copilot Chat. m365_copilot is reused as the wizard destination; the legacy ?ft=1 share link still resolves to the same card.
    • Reframed the prescreen entry to "Help me find the right place to get work done."
    • Added a copilot_chat recommendation block to apa.yaml with use X instead when… cross-references to the built-in agents, Cowork, and Scout (and vice versa).
    • Shareable via ?dt=copilot_chat and ?dt=m365_copilot.
    • Tightened the wizard option copy to echo the "Microsoft 365 Copilot: what to use and when" task vocabulary — general help (brainstorm, find info, catch up on email/meetings, draft & edit docs), on-demand (multi-step job / multiple artifacts in one go), and continuous (always-on, manage & coordinate my day) — so users self-identify faster.
    • Updated docs/SCORING.md and docs/FLOWCHART.md; extended Playwright coverage for the entry-point wizard and built-in-agents destination.
  • Refreshed the visual identity to the Warm Charcoal Instrument (via /design-shotgun): the previous near-black #0C0F14 canvas with a single cyan-blue signal glow read as the generic AI-tool aesthetic. New system keeps IBM Plex Sans/Mono but swaps to a warm matte charcoal canvas (#1A1714, no blue-black) with a single restrained teal signal (#17B0A7) and no colored glows.
    • Retokenized dark + light :root palettes; removed the two blue box-shadow glows and the blue canvas grid tint.
    • Updated docs/DESIGN.md (direction, color table, light-mode note, no-glow rule, Decisions Log). All 38 Playwright tests passing.

2026-07-20

Added

  • Added a content-analysis task option for lightweight document, chart, image, and data-analysis scenarios.
  • Added a web/uploaded-files data option for scoped web sources, PDFs, Office files, and embedded content.

Changed

  • Implemented Graphite Decision Instrument design system across the entire application:
    • Dark graphite canvas (#0C0F14) with subtle grid texture as default theme
    • IBM Plex Sans + IBM Plex Mono typography (replacing Segoe UI / Geist Mono)
    • Azure-cyan #2BA8FF accent for all signal colors, progress, and CTAs
    • Dark-first approach: dark is default, light mode is the alternate
    • Updated type scale: title 36px, heading 26px, subhead 20px, body 15px
    • Refined component styling: lit-edge cards, instrument-grade score bars, surface hierarchy
    • Maintained full WCAG AA contrast compliance
    • All 32 Playwright tests passing
  • Updated README.md to reflect the current evolution of the tool: built-in Copilot use, personal-agent delegation, ways to use or build agents, current platform positioning, delegate routing, and the latest test coverage.
  • Refreshed Agent Builder guidance and scoring for current capabilities: Microsoft 365 content, scoped web, uploaded files, Teams/Outlook/People knowledge, admin-enabled Microsoft 365 Copilot connectors, code interpreter, and image generation.
  • Repositioned Agent Builder beyond SharePoint/OneDrive-only scenarios and increased scores for small-team no-code knowledge, connector-backed, web/uploaded-file, and lightweight content/data-analysis scenarios.
  • Clarified that Agent Builder remains disqualified for external audiences, custom app deployment, background/event-driven execution, direct business system integrations, custom retrieval architectures, and multi-step action workflows that submit forms or update external systems.
  • Split internal audience guidance into small-team versus broad internal deployment and clarified connector-backed business systems, direct business integrations, and custom retrieval architectures.
  • Refreshed Copilot Studio guidance and scoring for the latest agent experience, generative orchestration, event-triggered workflows, computer use, MCP tools/resources, connected/child agents, A2A integrations, Microsoft IQ, Foundry IQ preview, per-user memory preview, model selection, built-in evaluation, monitoring, agent inventory, and Copilot Credits cost considerations.
  • Increased Copilot Studio scoring for background/event-triggered agents, Dataverse/custom connector/API integration, complex-but-low-code orchestration, and AI/ML personas, while preserving Foundry as the stronger fit for custom model training, arbitrary BYO model/runtime control, high-scale code-first orchestration, and custom retrieval architecture.
  • Split data-source guidance into business-system integration (Dataverse, custom connectors, direct APIs: Copilot Studio strongest) and custom retrieval architecture (custom RAG, Azure AI Search, private indexes, Foundry IQ, engineering-managed retrieval: Foundry strongest). The scored matrix now has 1,920 answer combinations.
  • Refreshed Microsoft Foundry guidance and scoring to reflect current Foundry Agent Service capabilities: prompt-based agents, hosted code agents, stable managed endpoints, publishing to Microsoft 365 Copilot and Teams, custom app/service integration, Foundry IQ, toolboxes, MCP, agent identities, RBAC, private networking, tracing, evaluation, optimization, monitoring, and Azure-scale production controls.
  • Increased Foundry scoring for custom app deployment, multi-surface deployment, Microsoft 365 Copilot/Teams publishing, developer-controlled content/data analysis, and Microsoft 365/web/file grounding, while keeping Foundry strongest for custom retrieval architecture and full-code production agents.
  • Removed over-specific Foundry memory-positioning language and replaced it with more conservative production-runtime guidance.
  • Refreshed Microsoft 365 Copilot guidance to cover Copilot Chat, Copilot Search, app-native Copilot experiences, Copilot Pages and Notebooks, and Microsoft-built agents as the built-in productivity layer for licensed internal users.
  • Added Copilot Search and Copilot Pages/Notebooks to the Microsoft 365 Copilot recommendation card, updated Researcher to use the official Learn page, and clarified Facilitator and Interpreter capabilities.
  • Added Cowork and Scout as their own cards on the Explore page.
  • Updated the recommendation-card accordion label so Microsoft 365 Copilot can show built-in capabilities alongside first-party agents.
  • Moved the "Use agents" home-page row above the "Build agents" row, with Microsoft 365 Copilot, Cowork, and Scout shown before Agent Builder, Copilot Studio, and Microsoft Foundry.
  • Reorganized the Explore page into a three-card desktop grid so the six cards form two rows of three.
  • Moved README, Changelog, Flowchart, and Scoring links from the header hamburger menu into a footer documentation row beneath the "Created by" credit.
  • Corrected the Explore starting-point copy to use the approved "ways to use or build agents" decision model.
  • Grouped the Explore page into "Use agents" and "Build agents" sections so the page reflects the current decision model instead of a flat gallery.
  • Replaced the design system with the Graphite Decision Instrument direction: dark graphite surfaces, IBM Plex typography, Azure-cyan signal color, decision rails, evidence panels, and stricter anti-slop rules.
  • Updated docs/SCORING.md and docs/FLOWCHART.md to match the current scoring matrix and recommendation guidance.

2026-07-14

Added

  • Added a separate Cowork and Scout delegate group on the landing page.
  • Added short captions beneath all platform and delegate tiles.

Changed

  • Centered delegate cards beneath the four build platforms under an "Or delegate" divider.
  • Renamed LICENSE.md to LICENSE.

2026-07-13

Added

  • Added Copilot Cowork and Microsoft Scout as delegate destinations in apa.yaml, with recommendation content and imagery.
  • Added a prescreen path for users who want a ready-made agent to do work for them.
  • Added cadence/reach routing for Cowork versus Scout: on-demand Microsoft 365 work routes to Cowork; continuous or cross-environment work routes to Scout; undecided signals show both.
  • Added shareable delegate result URLs via dt=cowork|scout|both.
  • Added end-to-end tests for delegate path routing and URL loading.

Changed

  • Documented the delegate path in docs/CHANGELOG.md, docs/FLOWCHART.md, and docs/SCORING.md.

2026-06-10

Added

  • Added .github/improvements.md.

Changed

  • Elevated Copilot Cowork across Microsoft 365 Copilot surfaces.
  • Added a dedicated Cowork spotlight on the Microsoft 365 Copilot recommendation card.
  • Updated Microsoft 365 Copilot recommendation, exploration, and start-page copy to lead with Cowork and first-party agents.

2026-05-09

Changed

  • Bumped fast-uri from 3.1.0 to 3.1.2 in package-lock.json.

2026-04-30

Changed

  • Updated recommendation and assessment text based on feedback.

2026-04-24

Added

  • Added the repository license.
  • Added bug report and feature request issue templates.

Changed

  • Updated issue templates.

2026-04-21

Changed

  • Auto-expanded Microsoft 365 Copilot accordions on the fast-track recommendation path.
  • Updated Microsoft 365 Copilot recommendation content.
  • Adjusted Cowork information in the Microsoft 365 Copilot surfaces.

Fixed

  • Fixed README formatting.

2026-04-20

Added

  • Added a hamburger documentation menu in the header with links to README, Changelog, Flowchart, and Scoring docs.
  • Added .github/copilot-instructions.md.

Changed

  • Updated README to reflect v2 features and current state.
  • Updated changelog structure.
  • Updated the landing page.
  • Refreshed the prescreen UI with icons, colored accents, and updated typography.

Removed

  • Removed CLAUDE.md after moving relevant instructions to .github/copilot-instructions.md.

2026-04-10

Changed

  • Removed the Agent Builder hard rule for connector-backed business systems (q3b), reflecting limited but non-zero connector capability.
  • Updated Agent Builder scoring for mixed Microsoft 365 plus connector-backed systems (q3d).
  • Updated docs/FLOWCHART.md and docs/SCORING.md for the scoring changes.

Removed

  • Removed the duplicate root CHANGELOG.md; docs/CHANGELOG.md became the single changelog source.

2026-04-08

Added

  • Added Clarity analytics.
  • Added documentation updates for scoring and flowchart behavior.

Changed

  • Updated scoring and result logic.
  • Moved the share-results button into the recommendation flow.
  • Updated apa.yaml, assets/apa.js, assets/apa.css, index.html, and Playwright tests for the revised logic.

2026-04-07

Added

  • Added the initial Agent Platform Advisor v2 static web app: apa.yaml, assets/apa.css, assets/apa.js, and index.html.
  • Added platform imagery for Agent Builder, Microsoft 365 Copilot, Copilot Studio, and Microsoft Foundry.
  • Added initial README and changelog files.
  • Added Microsoft-required SECURITY.md.
  • Added design, scoring, and flowchart docs under docs/.
  • Added Playwright configuration and end-to-end tests for fast-track, sharing, shared links, temporal changes, and wizard completion.
  • Added package manifest and lockfile.
  • Added favicon and image fixes.
  • Added a resources CTA button.

Changed

  • Refined initial UI, CSS, and app behavior.
  • Updated .gitignore.
  • Updated README and app content after the initial import.

Removed

  • Removed temporary TODOS.md.