llm: showcase optional Allium behavioral-spec callouts across five plugins

[withinfocus]

🎟️ Tracking

Initial exploration given some discussions. No associated Jira ticket — this is a self-directed showcase of an integration idea, opened for discussion.

📔 Objective

Adds an optional, conditional callout to five plugins pointing at JUXT's Allium — a third-party, MIT-licensed behavioral-specification language (.allium files in when / requires / ensures form, with a Rust CLI + LSP). This is a draft to showcase the idea, not a commitment to adopt Allium.

Why. Allium emits nothing in this meta-repo — but our engineers run these skills in product repos, which is exactly where a .allium spec and its CLI would live. The behavioral-spec lifecycle maps onto skills we already ship: elicit/distill at requirements & breakdown time, propagate/weed at implementation, weed again at review. These callouts plant tier-1 "consider Allium here" pointers at those seams without reimplementing anything or taking on a hard dependency.

Plugin	Skill / file	Verbs surfaced
bitwarden-product-analyst	requirements-elicitation	/elicit, /distill
bitwarden-delivery-tools	developing-breakdown-spec	seed .allium from the breakdown WHAT
bitwarden-tech-lead	architecting-solutions	/weed (narrower fit, scoped as such)
bitwarden-software-engineer	AGENT.md	/distill, /propagate, /weed, /tend
bitwarden-code-review	performing-multi-agent-code-review	/weed as an optional pre-Step-2 input

Scope & honest caveats

• Value is contingent on a product repo actually adopting Allium (CLI + LSP installed, .allium file present). Absent that, every callout is a no-op by design.
• Allium is new (2026, v2) and third-party. Nothing here depends on it at runtime — the callouts degrade to silence.
• I did not touch the code-review numbered pipeline. The /weed pass is an optional input feeding Step 2's architecture agent; deeper wiring is deliberately deferred pending a real pilot on a Bitwarden repo.
• Per repo policy: MINOR version bumps + changelog entries for all five plugins.

Validation. pnpm run lint, validate-plugin-structure.sh, and validate-marketplace.sh all pass for the five plugins.

Open questions for the team

1. Is the callout pattern the right vehicle, or should this be a single shared "when to reach for a behavioral spec" reference that plugins link to?
2. Which of these five genuinely want it vs. which dilute their focus? (The tech-lead fit is the weakest and is scoped accordingly.)
3. Worth a tier-1 pilot — install Allium, run /distill + /elicit against a real Bitwarden repo — before merging any of this?

[github-actions]

Plugin Validation Summary — PR #154

Result: ✅ PASS — All structural, skill-quality, and security checks pass. No errors (must-fix) found. A small number of minor warnings (should-fix / housekeeping) are noted below; none block merge.

This PR is additive documentation only: each of five plugins gains an optional "Allium" behavioral-spec callout in a skill or agent file, plus a semantic version bump and a changelog entry. No executable code, hooks, settings, or MCP configuration changed.

Plugin	New version	marketplace.json	AGENT.md version:	Changelog (2026-06-29)
bitwarden-code-review	1.14.0	✅ 1.14.0	✅ 1.14.0	✅ [1.14.0]
bitwarden-delivery-tools	2.1.0	✅ 2.1.0	— (no agent)	✅ [2.1.0]
bitwarden-product-analyst	0.2.0	✅ 0.2.0	— (no agent)	✅ [0.2.0]
bitwarden-software-engineer	1.1.0	✅ 1.1.0	n/a (no version: field)	✅ [1.1.0]
bitwarden-tech-lead	2.4.0	✅ 2.4.0	— (no agent)	✅ [2.4.0]

---

1. Plugin Structure (plugin-validator) — ✅ PASS

• Manifests: All five plugin.json files have required fields (name, version, description, author); names are valid kebab-case; versions match the table above.
• Version consistency: plugin.json ↔ root .claude-plugin/marketplace.json ↔ AGENT.md frontmatter all agree.
  • plugins/bitwarden-code-review/agents/bitwarden-code-reviewer/AGENT.md:3 → version: 1.14.0 (bumped from 1.13.0). ✅
  • plugins/bitwarden-software-engineer/agents/AGENT.md carries no version: field by existing convention, so the bump correctly did not touch it.
• Changelogs: Every plugin has a top entry in Keep a Changelog format under ### Added, dated 2026-06-29, matching its version.
• Agent frontmatter: Both changed AGENT.md files valid — names lowercase-hyphen within 3–50 chars, model: opus, substantial system-prompt bodies. bitwarden-software-engineer has 4 well-formed <example>/<commentary> blocks and color: blue.
• Secret scan: No API keys, tokens, private keys, or hardcoded passwords in any changed file.

Minor (informational, no action required):

• plugins/bitwarden-code-review/agents/bitwarden-code-reviewer/AGENT.md — no optional color: field (pre-existing; agent uses default).
• plugins/bitwarden-software-engineer/agents/AGENT.md — no version: field (pre-existing convention). Add version: 1.1.0 only if uniform agent-level version traceability is desired.

---

2. Skill Review (skill-reviewer) — ✅ PASS

Four SKILL.md files received small additive "Optional … (Allium)" sections (2–9 lines each). No frontmatter changed, so triggering behavior is unaffected. All frontmatter valid (name + description present). All referenced references/ and examples/ files exist on disk. Word counts after additions: code-review ~2,900 (body), delivery-tools ~1,111, product-analyst ~856, tech-lead ~1,478 — all within/below the 1,000–3,000 target. Progressive disclosure respected: each addition is a short conditional pointer that defers mechanics to external commands rather than inlining detail; none warrants a new references/ file. The tech-lead addition is exemplary — it actively narrows applicability to avoid over-triggering.

Minor (should-fix / polish — non-blocking):

• plugins/bitwarden-code-review/skills/performing-multi-agent-code-review/SKILL.md:63 — added section uses second-person permissive voice ("you may invoke") vs. the file's imperative house style; consider "Optionally invoke…".
• plugins/bitwarden-code-review/skills/performing-multi-agent-code-review/SKILL.md:63 — "the numbered pipeline below" is distant from the actual numbered steps (appear ~L129+); consider "the numbered Code Review Process steps."
• plugins/bitwarden-code-review/skills/performing-multi-agent-code-review/SKILL.md — body (~2,900 words) is at the upper edge of the 3,000 target. Pre-existing; route future additions to references/.
• plugins/bitwarden-product-analyst/skills/requirements-elicitation/SKILL.md:3 — pre-existing double space in the description string (cosmetic; not introduced by this PR).
• Cross-cutting consistency — the four callouts use three heading styles (## Optional: Behavioral-Spec Drift (Allium), inline **Optional — seed a behavioral spec.**, ## Optional: Capture a Behavioral Specification (Allium), ## Optional: Behavioral Specifications (Allium)). Per-context tailoring is fine; align the ## headings if uniformity is wanted.

---

3. Security Validation (reviewing-claude-config) — ✅ PASS

Check	Result
Committed settings.local.json / settings files	✅ PASS — none in diff
Hardcoded secrets / API keys / tokens / passwords	✅ PASS — none
Permission scoping / broad Bash:* / Read://*	✅ PASS — no settings/permission files changed
Dangerous command auto-approvals	✅ PASS — no hooks/commands added
Hooks / executable code	✅ PASS — none added

The external-tool slash commands (/weed, /elicit, /distill, /propagate, /tend) are documentation references, not auto-approved shell commands, and are correctly gated — every occurrence conditions use on the team having adopted Allium and a .allium file actually being present, with explicit "optional / skip otherwise" guards.

Minor (should-verify before merge — non-blocking):

• External repository link (unverified) — five Bitwarden-maintained configs now point engineers/reviewers at the third-party repo https://github.com/juxt/allium and assert it is MIT-licensed, at:

  • plugins/bitwarden-code-review/skills/performing-multi-agent-code-review/SKILL.md:63
  • plugins/bitwarden-delivery-tools/skills/developing-breakdown-spec/SKILL.md:79
  • plugins/bitwarden-product-analyst/skills/requirements-elicitation/SKILL.md:77
  • plugins/bitwarden-software-engineer/agents/AGENT.md:77
  • plugins/bitwarden-tech-lead/skills/architecting-solutions/SKILL.md:79

  The repo's existence/ownership/license could not be verified in this environment (network fetch not permitted). Since these are trusted persona configs, a human reviewer should confirm juxt/allium exists, is the intended owner, and is genuinely MIT-licensed before merge — or soften the license claim to "per the upstream repo; verify before adopting." Soft supply-chain/prompt-injection surface only; no code-level fix needed.

• Prompt-injection hygiene — plugins/bitwarden-code-review/skills/performing-multi-agent-code-review/SKILL.md:63 feeds Allium /weed drift output into the architecture agent "as additional context." .allium files and their output are untrusted repo content; the skill already routes the result as context / a finding (data, not instructions), which is correct. Keep it framed that way if this section is expanded later.

---

Verdict

✅ Approve. The PR is structurally sound, version/changelog compliant, and introduces no secrets, permission changes, or executable code. All flagged items are minor polish or a recommended pre-merge sanity-check on the external juxt/allium link — none are blocking.

Note: the repository's validate-plugin-structure.sh / validate-marketplace.sh scripts could not be executed in this environment (interactive approval unavailable); the equivalent checks were performed via the plugin-validator agent and direct file inspection.

[github-actions]

🤖 Bitwarden Claude Code Review

Overall Assessment: APPROVE

Reviewed the five-plugin showcase adding optional, conditional Allium behavioral-spec callouts to requirements-elicitation, developing-breakdown-spec, architecting-solutions, the software-engineer AGENT.md, and performing-multi-agent-code-review. All changes are additive markdown plus MINOR version bumps and changelog entries. The callouts are consistently framed as external, third-party, and no-ops absent a .allium file, and the code-review change correctly stays outside the numbered pipeline. No security, correctness, or breaking-change concerns in the prose itself.

Code Review Details

• ♻️ : Agent frontmatter version left stale at 0.1.5 while the plugin was bumped to 0.2.0. Before this PR both read 0.1.5, so the drift is introduced here. scripts/bump-plugin-version.sh only rewrites files named AGENT.md, and this agent file is product-analyst.md, so the script skips it; no validation script checks agent frontmatter versions, so CI does not catch it. Update line 3 to version: 0.2.0.
  • plugins/bitwarden-product-analyst/agents/product-analyst.md:3