feat: restore tier-3 eval coverage across eight plugins

plugins/ai-briefing/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "ai-briefing",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "description": "Build source-backed AI-industry briefings from official vendor publications, configured RSS/Atom feeds, GitHub releases, reputable secondary reporting, and user-supplied URLs. Deduplicate, rank, and present results as markdown or optional HTML/PPTX decks, with repository-owned profile, audience, and brand configuration. Automated X/Twitter collection is disabled; Playwright is used only for deterministic local rendering.",
   "author": {
     "name": "Melodic Software",

plugins/ai-briefing/CHANGELOG.md

@@ -0,0 +1,20 @@
+# Changelog
+
+All notable changes to the `ai-briefing` plugin are documented here. Format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/); this plugin uses semantic versioning.
+
+## [0.4.0]
+
+### Added
+
+- Engine behavioral evals in `skills/ai-briefing/evals/evals.json`, covering: `retro` action
+  routing and per-item acted/noted/skipped scoring against an archived briefing; `search`
+  action full-text matching across archives; markdown-only output when `--format slides`/`html`
+  is not explicitly requested; merge-not-append behavior when folding newly collected items into
+  an already-open briefing window; the apolitical filter, pragmatic-use ranking lens, and
+  profile-provided impact-lens annotation (via `references/audience-defaults.md`); and graceful,
+  visibly-surfaced degradation when an optional collection source is unreachable, without
+  aborting the run.
+- Three supporting fixtures under `skills/ai-briefing/evals/fixtures/`: `archive-sample.md`,
+  `open-window-sample.md`, and `candidate-items-sample.md` — neutral, synthetic AI-industry
+  content with no real company, person, or consumer-specific references.

plugins/ai-briefing/skills/ai-briefing/evals/evals.json

@@ -0,0 +1,86 @@
+{
+  "skill_name": "ai-briefing",
+  "evals": [
+    {
+      "id": 1,
+      "name": "retro-action-routing-and-scoring",
+      "prompt": "/ai-briefing retro --meeting 5\n\nThe archived briefing for meeting 5 is at evals/fixtures/archive-sample.md relative to the skill directory — read it, then run the retro.",
+      "expected_output": "An interactive retrospective that reads the meeting-5 archive, walks each item, captures an acted/noted/skipped signal per item, and produces a retro annotation for meeting 5 (e.g., a retro-5 record under the profile's state directory) — without launching a fresh collection run or reopening the window.",
+      "files": ["evals/fixtures/archive-sample.md"],
+      "expectations": [
+        "Routes to the retro action — scores already-archived items and does NOT start a fresh collection run (no vendor-blog/GitHub/RSS/secondary-reporting wave collection)",
+        "Extracts the individual item bullets (provider, headline, source URL) from the meeting-5 archive fixture rather than inventing items",
+        "Captures an acted / noted / skipped signal for each item (interactive per-item prompting, or a documented auto-mode heuristic inference)",
+        "Produces retro output (a per-item annotation plus a pattern summary), not a new briefing for a different window",
+        "Surfaces follow-list demote/promote recommendations derived from the per-item signals, rather than emitting a fresh collection briefing"
+      ]
+    },
+    {
+      "id": 2,
+      "name": "search-action-across-archives",
+      "prompt": "/ai-briefing search \"Cloud Agents\"\n\nTreat evals/fixtures/archive-sample.md (relative to the skill directory) as the archived-briefing corpus to search.",
+      "expected_output": "A full-text search result that returns the matching Northwind Cloud Agents item from the archive, grouped by meeting number with tier/signal context, without running any collection.",
+      "files": ["evals/fixtures/archive-sample.md"],
+      "expectations": [
+        "Routes to the search action — full-text search over archived briefings, no collection run and no new briefing emitted",
+        "Returns the matching item(s) that actually contain the query text in the fixture (the Northwind Cloud Agents GA item), grouped or labelled by meeting number",
+        "Reports match context such as tier (HIGH/MED/LOW) and/or cross-meeting recurrence count",
+        "Does not fabricate items absent from the archive — only real matches from the fixture are returned"
+      ]
+    },
+    {
+      "id": 3,
+      "name": "slides-never-auto-generated",
+      "prompt": "Catch me up on AI from the last two weeks so I'm ready for the team meeting.",
+      "expected_output": "A markdown briefing covering roughly the last 14 days. No PPTX or HTML deck is produced because the user did not opt into --format slides or --format html.",
+      "files": [],
+      "expectations": [
+        "Output is a markdown briefing, not a PPTX or HTML slide deck",
+        "The slide/HTML build pipeline (output/build/*.js) is not invoked without an explicit --format slides or --format html opt-in",
+        "The window resolves to roughly 14 days (interpreting 'last two weeks' / the --since default) rather than a 1-day or 30-day window"
+      ]
+    },
+    {
+      "id": 4,
+      "name": "merge-not-append-continuation",
+      "prompt": "The meeting-prep briefing already open for this window is at evals/fixtures/open-window-sample.md relative to the skill directory — the bottom section lists new items collected today. Fold the new items into the briefing and give me the updated consolidated document.",
+      "expected_output": "A single consolidated briefing where the new items are merged into the existing provider buckets under their HIGH/MED/LOW headings, prior items preserved, and no delta/supplement section appended.",
+      "files": ["evals/fixtures/open-window-sample.md"],
+      "expectations": [
+        "New items are merged into the existing provider buckets under their HIGH/MED/LOW headings (add-only = MERGE)",
+        "No '## Run N', '## Supplement', or 'New items collected today' delta/append section survives in the merged output — the body stays a single consolidated document",
+        "Existing items from the prior briefing are preserved in the merged output, not dropped",
+        "The previously-empty Northwind Systems bucket is populated by its new item rather than left as a stale 'no notable items' note",
+        "Each provider appears once with its items consolidated under it, not duplicated across multiple sections"
+      ]
+    },
+    {
+      "id": 5,
+      "name": "apolitical-filter-pragmatic-ranking-bucket-coverage",
+      "prompt": "I've already collected this raw batch of AI items — evals/fixtures/candidate-items-sample.md relative to the skill directory. The active profile's audience.md declares an impact lens over a typical web-application stack (a JS/TS runtime, a hosted database, an auth provider, and cloud hosting). Categorize, apply the filters, rank them, and lay out as the briefing.",
+      "expected_output": "A ranked briefing that drops partisan-only items, keeps genuine industry controversy even when a lawmaker is named, ranks usable tooling releases HIGH above research speculation, routes the robotics item to EXTRAS, and tags HIGH items with an impact annotation against the declared stack lens.",
+      "files": ["evals/fixtures/candidate-items-sample.md"],
+      "expectations": [
+        "Partisan-only items are dropped before ranking (the candidate-deepfake meme and the partisan horse-race thread do not appear anywhere in the briefing)",
+        "Genuine industry-wide controversy is kept even though a lawmaker is named (the VendorCo procurement-dispute item appears, under Legal & regulatory)",
+        "Tooling/IDE/CLI releases usable within a week are ranked HIGH (the Solstice Code, Meridian Agent CLI, and Northwind IDE items) ahead of research speculation",
+        "The research-speculation post and the hype post are ranked LOW/MED, not HIGH",
+        "The robotics item (the Anchorpoint Unit-7 warehouse demo) is routed to the EXTRAS section rather than a mainstream provider bucket",
+        "HIGH items carry an impact annotation assessed against the declared stack lens (e.g., \"impact: medium — ...\"), per the profile-provided impact lens"
+      ]
+    },
+    {
+      "id": 6,
+      "name": "unreachable-source-degrades-visibly",
+      "prompt": "/ai-briefing --since 3d\n\n(One of the profile's configured RSS/Atom feeds times out and is unreachable during this run; official vendor blogs and GitHub releases remain reachable.)",
+      "expected_output": "The briefing proceeds using the still-reachable source classes; the unreachable feed is surfaced as a visible warning/caveat rather than silently dropped, and the run is not aborted for one unreachable optional source.",
+      "files": [],
+      "expectations": [
+        "The run continues via the still-reachable source classes — one unreachable RSS/Atom feed does NOT block or abort the briefing",
+        "The unreachable source is surfaced as a visible warning (in the run report and/or the briefing's caveats section), not silently ignored",
+        "The run is not failed outright for a single unreachable optional source, matching the documented 'no black-box degradation' quality gate",
+        "Items collected from the still-reachable source classes are still deduplicated, ranked, and emitted normally in the resulting briefing"
+      ]
+    }
+  ]
+}

plugins/ai-briefing/skills/ai-briefing/evals/fixtures/archive-sample.md

@@ -0,0 +1,44 @@
+# AI Briefing — Meeting 5 (archived)
+
+Window: 2026-06-05 to 2026-06-19 · Runs: 3 · Source classes: official vendor blogs, GitHub releases, RSS/Atom feeds, secondary reporting
+
+## Solstice AI
+
+### HIGH
+
+- **Solstice Code v3.4 ships checkpoint rewind** — restore any prior turn's file state without external version control. Source: Solstice AI Engineering Blog, <https://blog.solstice.example/code-3-4-checkpoint-rewind> (2026-06-06).
+- **Solstice model pricing cut 20%** — input/output token cost drops across all paid tiers. impact: medium — lowers per-run inference cost. Source: Solstice AI Pricing Notes, <https://blog.solstice.example/pricing-update-june> (2026-06-09).
+
+### MED
+
+- Solstice AI publishes a research note on interpretability via feature steering. Source: Solstice AI Research Blog, <https://research.solstice.example/feature-steering> (2026-06-11).
+
+## Meridian Labs
+
+### HIGH
+
+- **Meridian Agent CLI adds MCP server support** — call local MCP tools from the terminal agent. Source: Meridian Labs Release Notes (GitHub), <https://github.com/meridian-labs/agent-cli/releases/tag/v1.8.0> (2026-06-12).
+
+### LOW
+
+- A Meridian Labs executive teases "a big announcement soon" with no specifics. Source: secondary reporting, <https://newswire.example/meridian-teaser> (2026-06-13).
+
+## Northwind Systems
+
+### HIGH
+
+- **Northwind Cloud Agents GA** — background agents run on Northwind-hosted infrastructure and open pull requests directly. Source: Northwind Systems Blog, <https://blog.northwind.example/cloud-agents-ga> (2026-06-15).
+
+## Legal & regulatory
+
+### HIGH
+
+- **Cross-border AI-governance framework finalized** — general-purpose model obligations take effect next quarter, industry-wide compliance impact. Source: secondary reporting corroborated by the regulator's official notice, <https://gov.example/ai-framework-final> (2026-06-17).
+
+## Other
+
+*No notable items this window.*
+
+## EXTRAS (Robotics, science, and novel applications)
+
+- Anchorpoint Dynamics' Unit-7 humanoid demo folds laundry unassisted in an unstructured home setting. Source: Anchorpoint Dynamics Blog, <https://blog.anchorpoint.example/unit-7-laundry-demo> (2026-06-18).

plugins/ai-briefing/skills/ai-briefing/evals/fixtures/candidate-items-sample.md

@@ -0,0 +1,16 @@
+# Raw candidate items — pre-categorize batch
+
+Unranked, unfiltered items collected this run via approved source classes (official vendor
+blogs, GitHub releases, RSS/Atom feeds, secondary reporting). Categorize, apply the apolitical
+filter, rank HIGH/MED/LOW by the pragmatic-use lens, and lay out as the briefing.
+
+1. Solstice Code v3.6 released: native subagent checkpointing, install via update. Engineers can use it this week. Source: Solstice AI Engineering Blog, <https://blog.solstice.example/code-3-6> (2026-06-23).
+2. Meridian Agent CLI 0.9 adds `--effort` flag and desktop app parity. Source: Meridian Labs Release Notes (GitHub), <https://github.com/meridian-labs/agent-cli/releases/tag/v0.9.0> (2026-06-23).
+3. Northwind IDE 4.1 ships inline Cloud Agent handoff from the editor. Source: Northwind Systems Blog, <https://blog.northwind.example/ide-4-1> (2026-06-23).
+4. A Meridian Labs research blog post speculates scaling laws may bend at very large parameter counts; no release, purely theoretical. Source: Meridian Labs Research Blog, <https://research.meridian.example/scaling-speculation> (2026-06-22).
+5. VendorCo and a national defense procurement office dispute a supply-chain security clause in a multi-year AI procurement contract; a lawmaker is quoted in the coverage. Affects federal AI access industry-wide. Source: secondary reporting, <https://newswire.example/vendorco-procurement-dispute> (2026-06-21).
+6. A deepfake video of a political candidate saying absurd things circulates online; pure partisan meme, high engagement. Source: secondary reporting, <https://newswire.example/deepfake-meme> (2026-06-20).
+7. A widely shared thread argues one political party uses AI campaign tools better than the other; partisan horse-race framing, no industry substance. Source: secondary reporting, <https://newswire.example/partisan-horse-race> (2026-06-20).
+8. An online influencer posts "AGI is basically here, feel the acceleration" with no supporting artifact. Source: secondary reporting, <https://newswire.example/agi-hype-post> (2026-06-19).
+9. Anchorpoint Dynamics' Unit-7 humanoid completes a full warehouse pick-and-pack shift autonomously. Source: Anchorpoint Dynamics Blog, <https://blog.anchorpoint.example/unit-7-warehouse-shift> (2026-06-19).
+10. Cobalt Dynamics posts open weights for its 4.1 open model with a minor context-length bump. Source: Cobalt Dynamics Blog, <https://blog.cobalt.example/weights-4-1> (2026-06-18).

plugins/ai-briefing/skills/ai-briefing/evals/fixtures/open-window-sample.md

@@ -0,0 +1,31 @@
+# AI Briefing — Meeting 6 (open window)
+
+Window: 2026-06-19 to present · Runs: 1 · Status: open (not yet closed out)
+
+## Solstice AI
+
+### HIGH
+
+- **Solstice Code v3.5 ships background tasks** — long-running agents survive session end. Source: Solstice AI Engineering Blog, <https://blog.solstice.example/code-3-5-background-tasks> (2026-06-20).
+
+### MED
+
+- Solstice AI's MCP registry adds 40 verified servers. Source: Solstice AI Blog, <https://blog.solstice.example/mcp-registry-40-servers> (2026-06-21).
+
+## Meridian Labs
+
+### HIGH
+
+- **Meridian Desktop app enters public beta.** Source: Meridian Labs Blog, <https://blog.meridian.example/desktop-beta> (2026-06-22).
+
+## Northwind Systems
+
+*No notable items this window.*
+
+---
+
+## New items collected today (uncategorized — fold these in)
+
+- Solstice Code v3.6 adds native subagent checkpointing (HIGH, tooling). Source: Solstice AI Engineering Blog, <https://blog.solstice.example/code-3-6-subagent-checkpointing> (2026-06-23).
+- Northwind IDE 4.1 ships inline Cloud Agent handoff (HIGH, tooling). Source: Northwind Systems Blog, <https://blog.northwind.example/ide-4-1-handoff> (2026-06-23).
+- Meridian Agent CLI 0.9 adds an `--effort` flag (MED, tooling). Source: Meridian Labs Release Notes (GitHub), <https://github.com/meridian-labs/agent-cli/releases/tag/v0.9.0> (2026-06-23).

plugins/codebase-audit/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "codebase-audit",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "description": "Repo-wide drift audit between docs, config, code, and architecture: verifies every factual claim against reality via parallel subagent fan-out, severity-rates findings, and fixes or presents for review. Audit dimensions are configurable through a tracked .claude/codebase-audit.md config file written by the setup skill.",
   "author": {
     "name": "Melodic Software",

plugins/codebase-audit/CHANGELOG.md

@@ -3,6 +3,16 @@
 All notable changes to the `codebase-audit` plugin are documented here. Format follows
 [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); this plugin uses semantic versioning.
 
+## [0.3.0]
+
+### Added
+
+- Eval covering the scope-boundary decline: declining claim-extraction fan-out over
+  `settings.json` / `.mcp.json` / hooks / permissions and routing to the adjacent
+  `claude-config-audit` plugin's `/claude-config-audit:settings-audit` skill (or stating
+  out-of-scope when that plugin is not installed) — behavior already documented in SKILL.md,
+  now regression-tested.
+
 ## [0.2.0]
 
 ### Added

plugins/codebase-audit/skills/codebase-audit/evals/evals.json

@@ -62,6 +62,19 @@
         "Presents a verified non-issues list showing what was checked",
         "When no discrepancies are found, states explicitly that all claims verified as correct rather than padding with marginal issues"
       ]
+    },
+    {
+      "id": 6,
+      "name": "boundary-route-to-settings-audit",
+      "prompt": "/codebase-audit:codebase-audit\n\nCheck my .claude/settings.json and .mcp.json for schema drift, and tell me whether my hooks and permissions are misconfigured.",
+      "expected_output": "Recognizes that settings.json / .mcp.json schema-drift and hook/permission validation is the adjacent claude-config-audit lane, not this skill's factual-claim-drift lane. Declines to run its own claim-extraction fan-out over the settings schema. When the claude-config-audit plugin is installed, routes the user to /claude-config-audit:settings-audit; otherwise states the request is out of scope.",
+      "files": [],
+      "expectations": [
+        "Identifies settings.json / .mcp.json / hooks / permissions schema-drift as out of scope for this skill's claim-verification lane",
+        "When claude-config-audit is installed, routes the user to /claude-config-audit:settings-audit rather than a bare or differently-named skill",
+        "When claude-config-audit is not installed, states the request is out of scope instead of routing to a nonexistent skill",
+        "Does not run its own claim-extraction fan-out over the settings/MCP schema"
+      ]
     }
   ]
 }

plugins/discovery/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "discovery",
-  "version": "0.4.1",
+  "version": "0.5.0",
   "description": "Structured discovery before changes: explore the local codebase (inline or in an isolated forked subagent) and run disciplined multi-source external research with source tiers, falsification, and recency gates — persisting EXPLORE.md / RESEARCH.md handoff artifacts.",
   "author": {
     "name": "Melodic Software",

plugins/discovery/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog — discovery plugin
 
+## 0.5.0 — 2026-07-15
+
+### Added
+
+- Research floor-scaling and broad-topic-minimums evals in `skills/research/evals/evals.json`:
+  `floor-scaling-single-product` pins that Phase 2 query count tracks the Phase 1 written gap
+  count rather than stopping at the 3-query floor (SKILL.md's "floor is a starting point, not a
+  target"); `broad-topic-triple-tool-comparison` pins that a 3-tool comparison topic fires the
+  doubled phase/query/source minimums (SKILL.md item 8, discipline.md's "Broad-topic
+  auto-detect").
+
 ## 0.4.0 — 2026-07-14
 
 Adopt the marketplace topic-docs convention, contract v1.0.0