bot-bottle/docs/research/manifest-format-and-grouping.md

Manifest format and grouping

Two open questions for claude-bottle's manifest layer after PRD 0011:

1. Grouping. Keep bottles and agents in the same manifest file (today's shape), or split them — one file per bottle and one file per agent.
2. Format. Stay on JSON, switch to YAML, or move to a Markdown spec with YAML frontmatter. The Markdown option splits into two sub-flavors: reuse Claude Code's existing subagent format with bottle-specific extensions, or invent a claude-bottle-owned Markdown spec used for both agents and bottles.

The trust boundary from PRD 0011 — bottle infrastructure lives in $HOME, agents may live in $CWD — is largely orthogonal to both axes. But the choice of grouping and format changes how naturally that boundary expresses on disk, and how comfortable the manifest will be once a user has 5+ bottles and 10+ agents.

Why this matters

Current shape: one JSON file at $HOME/claude-bottle.json (and optionally $CWD/claude-bottle.json for cwd-defined agents). After PRD 0011, the home file owns bottles + home agents; the cwd file is agents-only.

The single-file shape works fine for the project's first 1-2 bottles. Real friction starts when:

• A user has 5-10 bottles for different projects, each carrying several cred_proxy.routes and a few bottle.git entries — the home file becomes hundreds of lines of nested JSON.
• Multiple humans share a $HOME manifest pattern (dotfiles repo, shared workstation, CI machine baseline) and want to compose pieces — JSON doesn't merge cleanly outside of the resolver.
• Per-agent prompts grow long. JSON forces them onto a single escaped line; multi-paragraph prompts become unreadable.
• Documentation (why does this bottle exist? what's the threat model for these credentials?) has nowhere natural to live in a JSON file; you end up with a sibling README that drifts from the config.

JSON's strengths (stable parser, machine-readable, stdlib-only) are real and shouldn't be thrown away lightly. The question is whether the inflection point has been reached.

Axis 1 — grouping

Option A: one file for both (current)

$HOME/claude-bottle.json contains bottles: and agents:. Cwd file (optional) contains agents: only.

Pros

• Zero new lifecycle. One file to discover, edit, version, diff.
• Trust boundary lives entirely in the resolver — the on-disk shape doesn't enforce or surface it.
• Atomic edits: changing a bottle and the agents that reference it is one commit, one save.

Cons

• Scales linearly with bottles + agents. A user with 8 bottles and 12 agents hits a ~600-line file even with terse formatting.
• Diff conflicts: two changes to unrelated agents touch the same file. Codeowners-style ownership doesn't apply cleanly.
• Discovery harder beyond a point: searching for one agent requires reading the whole file in a JSON parser, not filename-globbing.
• The trust boundary is invisible on disk — a reader can't tell at a glance which entries are home-trusted vs cwd-supplied; they have to know the resolver's rules.

Option B: file per thing

Bottles live as $HOME/.claude-bottle/bottles/<name>.<ext>. Agents live as $HOME/.claude-bottle/agents/<name>.<ext> (home agents) and $CWD/.claude-bottle/agents/<name>.<ext> (cwd agents). The resolver globs each directory.

Pros

• Scales to N bottles + N agents without any single file growing.
• Trust boundary expresses on disk: $HOME/.claude-bottle/bottles/ is the only place bottles can come from. $CWD/.claude-bottle/ can only contribute agents. No resolver logic needed to enforce it — the file paths are the enforcement.
• Aligns with Claude Code's existing model: each subagent already lives as ~/.claude/agents/<name>.md. Claude Code users will recognize the directory shape.
• Per-file ownership / codeowners / diff workflows just work.
• Per-agent prompts grow without affecting other files.
• Documentation per bottle/agent can live in the file itself (e.g. comments, Markdown body).

Cons

• More lifecycle: creating, renaming, deleting agents/bottles becomes file ops (mkdir, mv, rm) instead of editing one file. Power users prefer that; new users may not.
• Discovery requires ls, not "grep one file." Tooling helps (e.g. ./cli.py list) but the manifest is no longer a single artifact to email or ship.
• Atomicity: swapping a bottle name across agents touches multiple files. Git handles this fine; a one-shot text editor flow loses something.
• Backwards compatibility: existing users have one JSON file. Migration tool needed.

Interaction with trust boundary

Option A keeps the resolver-enforced boundary from PRD 0011 as the only enforcement.

Option B can express the boundary purely as filesystem layout: <home>/bottles/ is privileged; <cwd>/ directory only has an agents/ subdirectory. The resolver becomes "glob the dirs, parse each file, validate the cross-references." Strictly cleaner than the current parse-and-reject logic, and more obvious to a reader auditing the security posture.

Axis 2 — format

Option 1: stay on JSON

What we have today. The trust-boundary change in PRD 0011 preserves this format.

Pros

• Zero migration cost.
• Stdlib parser; no new dependency. The project's CLAUDE.md sets "low dependencies by default" as a guideline.
• Stable, predictable parse semantics. No type-coercion gotchas.
• Tooling everywhere — IDE support, linters, jq.

Cons

• No native comments. (JSONC, JSON5, _comment fields are all workarounds.)
• Multi-line strings become escaped one-liners. Agent prompts longer than a sentence become unreadable.
• Trailing commas are an error. Hand-editing punishes small typos.
• Verbose: every key + value gets quotes; nested structures grow indent.

Option 2: full YAML

$HOME/claude-bottle.yaml (or .yml). Parser pulls in PyYAML (or ruamel.yaml).

Pros

• Comments, multi-line strings (block scalars), anchors for repeated blocks (e.g. shared egress allowlists across bottles).
• Common config language for ops tooling (Kubernetes, GitHub/Gitea Actions, Docker Compose, pipelock's own config).
• Less syntactic noise than JSON for nested data.

Cons

• New runtime dependency. The project today uses zero third-party Python packages for production code; YAML parsing pulls in PyYAML. (CLAUDE.md: "Python, stdlib-first; low-deps by default.")
• YAML's footguns: indentation sensitivity, the Norway problem (country: NO → boolean False), implicit type coercion that's surprised non-trivial production projects.
• Specifying schemas in YAML is harder to validate strictly — parsers are forgiving where JSON is strict.
• No native escape hatch for executable content / templating, but users will reach for one (Jinja, Helm-style) and then we're in yaml-as-template-language territory.

Option 3: reuse Claude Code's subagent spec (Markdown + YAML frontmatter), with claude-bottle extensions

Claude Code already stores subagents at ~/.claude/agents/<name>.md with YAML frontmatter and a Markdown body. Frontmatter today carries fields like name, description, model, color, memory; the body is the system prompt. Adding fields like bottle: dev and a claude_bottle: sub-block to the same frontmatter would make each claude-bottle agent a drop-in addition to Claude Code's agent directory.

---
name: implementer
description: Implements features against PRDs in this repo
model: opus
bottle: dev
claude_bottle:
  skills: [init-prd]
---

You are a feature-implementation agent running inside an
ephemeral claude-bottle sandbox. The host has copied the user's
project into /home/node/workspace...

Bottles don't fit Claude Code's agent schema — they're infrastructure, not behavior. Either:

• (3a) Bottles stay JSON / YAML; only agents adopt the MD+frontmatter format. Mixed-format manifest.
• (3b) Bottles adopt MD+frontmatter too, using a claude-bottle-only schema. Then we're really doing option 4 for bottles + option 3 for agents. Two formats but one parser.

Pros

• Existing Claude Code users already know this format and have a directory full of these files. The mental model is "an agent is a Markdown file."
• Each agent's prompt lives naturally as Markdown body — long prompts read well, can use headings/lists/code blocks.
• File-per-thing falls out automatically (one MD per agent).
• Claude Code may eventually consume claude-bottle's agent files directly, doubling their utility.

Cons

• Coupling to Claude Code's spec. Anthropic owns that schema; field names and semantics can change. Today's model / description / memory are stable, but tomorrow's may not be. Our bottle: / claude_bottle: extensions could collide with future official fields.
• The agent file's frontmatter starts to carry two unrelated schemas: Claude Code's (model, description) and ours (bottle, skills, ...). One file, two owners.
• Bottles still need a format choice (3a vs 3b above) — we don't escape that decision.
• Parsing MD+frontmatter is more work than JSON. Either pull a frontmatter library (python-frontmatter) or hand-parse the --- block and feed it to PyYAML. Either way, a new dep.

Option 4: invent a claude-bottle MD spec, used for both agents and bottles

---
# $HOME/.claude-bottle/agents/implementer.md
bottle: dev
skills: [init-prd]
---

You are a feature-implementation agent running inside an
ephemeral claude-bottle sandbox...

---
# $HOME/.claude-bottle/bottles/dev.md
cred_proxy:
  routes:
    - path: /anthropic/
      upstream: https://api.anthropic.com
      auth_scheme: Bearer
      token_ref: CLAUDE_BOTTLE_OAUTH_TOKEN
      role: anthropic-base-url
egress:
  allowlist: [example.com]
---

The dev bottle. Backs my work on personal projects: Anthropic
OAuth, the gitea instance at gitea.dideric.is, and an npm token
for publishing scoped packages.

Pros

• Single format, two directory layouts. One parser, one mental model.
• Bottle files get an MD body that's a natural home for documentation (why does this bottle exist? what tokens does it hold? who owns the keys?).
• Not coupled to Claude Code's schema; we own the spec.
• Trust boundary on disk: $HOME/.claude-bottle/bottles/ is the only place bottles can come from; $CWD/.claude-bottle/agents/ is the only thing cwd contributes.
• Agent files in this spec are almost compatible with Claude Code's subagent format. If we keep the name / description conventions, the same files can drop into ~/.claude/agents/ with no friction — best of both worlds without the formal coupling.

Cons

• Invents a format. Users learn one more thing (small thing — MD with frontmatter is widely understood).
• Bottle file bodies have no built-in use case beyond documentation; users may leave them empty, which looks weird ("why is this file partly Markdown?").
• Still requires a YAML parser for the frontmatter, so the dependency cost is the same as option 3.

Synthesis

Combining axes:

	JSON	YAML	MD reuse (3)	MD new (4)
Grouped (A)	today	yaml monolith	not natural — MD wants per-file	not natural
Per-file (B)	dir of JSON files	dir of YAML files	best fit	best fit

Per-file × MD-with-frontmatter is the natural shape on both axes — the format wants to live one-per-file (the "MD doc with metadata" pattern doesn't lend itself to monoliths), and the file-per-thing grouping fits how users iterate on agents (write a prompt, save, launch).

Between option 3 (reuse CC spec) and option 4 (new spec): the appealing middle ground is "claude-bottle agents follow the CC subagent shape closely (name / description / model + bottle and skills extensions) so they drop into ~/.claude/agents/ as a side effect, while bottles use the same MD+frontmatter shape but with claude-bottle's own schema and live in a dedicated directory." This:

• gives agents both a claude-bottle launch story AND a Claude Code invocation story from the same file;
• keeps bottles entirely under our schema (no Anthropic dependency for the security-load-bearing config);
• uses one parser, one body-format, two directories.

Cost of moving:

• New runtime dep: a YAML parser (PyYAML or a hand-parse-the- frontmatter shim). PyYAML is the safest choice if we accept the dependency.
• Resolver rewrite: glob two directories, parse each file, validate cross-references. Roughly the same complexity as today's JSON merge once the boundary check is in place.
• Migration tool: a one-shot script that splits today's JSON into per-file MD docs. Five minutes of work for the tool, five minutes of work for the user.
• Docs: README's manifest section gets rewritten. Worth doing alongside the move.

Recommendation

Per-file MD with frontmatter (option B × option 4 with the option-3 agent compatibility). The format change clears the way for the per-file grouping (which is the bigger UX win), and the per-file shape is what makes the trust boundary self-documenting on disk.

The dependency cost (PyYAML) is the main thing that needs an explicit yes from the user — claude-bottle today has zero third-party Python deps for production code, and adopting one crosses a clean architectural line. If "low deps" stays a hard constraint, the alternative is to hand-parse the frontmatter block and feed it to a minimal YAML subset parser (the keys claude-bottle uses are all flat string/list/dict — no anchors, no multi-line block scalars, no implicit type coercion).

If we don't want to commit to the move yet, the next-cheapest option is keeping JSON but splitting into per-file (option B × option 1): $HOME/.claude-bottle/bottles/<name>.json + $HOME/.claude-bottle/agents/<name>.json. Most of the scaling wins; none of the body-prose or dependency story.

Open questions

• Does Claude Code object to extra frontmatter fields? Test: drop a file with bottle: in ~/.claude/agents/ and see if CC warns / ignores / breaks. If it warns, we'd want a different field name (e.g. claude-bottle-bottle) or a namespaced block.
• Migration story. Is the project willing to ship a one-shot ./cli.py migrate-manifest command that does the JSON → MD conversion? Or do users just rewrite by hand from the new docs?
• Bottle file body content. If most bottle .md files have an empty body, is the MD-with-frontmatter format still warranted? An alternative is YAML for bottles only (no body, but with comments) and MD+frontmatter for agents.
• Dotfiles vs not. $HOME/.claude-bottle/ or $HOME/claude-bottle/? The hidden dotfile shape matches dev conventions (.config/, .ssh/); the visible shape signals "this is a real thing you own."
• PyYAML hard dep, or minimal subset parser? Trade-off between "honest about the dependency" and "stay stdlib-only."