PDPP

Layer 2 coverage audit: ChatGPT, Claude Code, Codex

add-polyfill-connector-systemProject noteConnector notes
Created openspec/changes/add-polyfill-connector-system/design-notes/layer-2-coverage-chatgpt-claude-codex.mdView on GitHub →

Raised: 2026-04-19 Context: Layer 1 compares emitted fields against declared manifest fields. Layer 2 compares the manifest against the source's actual surface — what the platform exposes that a user might want and we omit. This audit covers the three AI-assistant connectors; none have been reviewed against the full upstream surface since v0.1.

ChatGPT

Current manifest scope

conversations, messages (flattened current-branch tree), memories. Profiles chats, full.

Source surface

Endpoints the browser UI hits on chatgpt.com/backend-api:

  • /conversations, /conversation/{id} (declared)
  • /memories?include_memory_entries=true (declared)
  • /gizmos/mine — user-authored Custom GPTs
  • /gizmos/bootstrap/projects, /public-api/conversations — Projects (group of conversations + instructions + files)
  • /shared_conversations — public share links the user created
  • /user_system_messages — custom instructions ("about me" / "how to respond" / traits)
  • /files/{id}, /files/{id}/download — uploaded attachments (message.attachment_ids resolves here; never joined)
  • /accounts/check, /me — account + workspace identity
  • /settings/data_controls — training opt-out, retention Voice transcripts land in the normal conversation tree. Canvas docs appear as content_type:"canvas" message parts.

Gaps

  • Stream custom_gpts (P0): User's own GPTs are a primary creative artifact. GET /backend-api/gizmos/mine?limit=24&cursor=... → {items:[{id, short_url, display:{name, description, prompt_starters, welcome_message, profile_picture_url}, author, instructions, tools, files, updated_at}]}. Equivalent to Notion omitting pages the user wrote.
  • Stream custom_instructions (P0): GET /backend-api/user_system_messages → {enabled, about_user_message, about_model_message, name_user_message, role_user_message, traits_model_message, disabled_tools}. Explicit in OpenAI's own takeout; comparable in weight to memories which we already have.
  • Stream shared_conversations (P1): GET /backend-api/shared_conversations?order=created → {share_id, conversation_id, title, create_time, update_time, is_anonymous, highlighted_message_id}. High recall-value — users rarely remember what they've published.
  • Stream projects + field project_id on conversations (P1): lift Projects out as their own stream; project_id is in the detail response today but dropped.
  • Stream files (P2): resolve attachment_ids via GET /backend-api/files/{id} → {file_name, mime_type, size, download_url, created_at}. Optional bytes fetch.
  • Field split on messages (P2): for content_type=user_editable_context, split user_profile/user_instructions out of the shared content column (or lift into custom_instructions).
  • Stream account (P2): email, plan, workspaces from /me.

Deliberately omitted

  • Bearer token + oai-device-id — auth.
  • Voice audio blobs — no distinct endpoint; transcripts are in messages.
  • Enterprise admin surfaces (audit logs, SCIM) — admin-scoped, not the individual user's.

Claude Code

Current manifest scope

sessions, messages, attachments, all from ~/.claude/projects/<encoded>/*.jsonl.

Source surface (~/.claude/ on this machine)

  • projects/<p>/*.jsonl — session transcripts (declared)
  • projects/<p>/<session>/subagents/*.jsonl — full subagent transcripts (the main jsonl only has sidechain stubs)
  • projects/<p>/<session>/tool-results/*.txt — full tool outputs (main jsonl stores ~500-char previews)
  • projects/<p>/memory/*.md — assistant-curated per-project memory (MEMORY.md + feedback notes)
  • skills/<name>/SKILL.md — user-installed skills (40+ here)
  • commands/*.md — user-authored slash commands
  • hooks/*.sh — user shell hooks
  • plugins/installed_plugins.json + plugins/marketplaces/ — plugin registry
  • settings.json, settings.local.json — permissions, enabled MCP servers, hooks
  • file-history/<uuid>/<hash>@v<n> — versioned snapshots of every file the agent edited
  • todos/*.json, tasks/<id>/, sessions/<id>/, paste-cache/*.txt, captured-insights.md, history.jsonl — various state
  • CLAUDE.md, CLAUDE.local.md, RTK.md — user-level system prompts
  • Each repo also carries ./CLAUDE.md + ./.claude/settings.json + ./.claude/commands/ — project-scoped behavior contracts

Gaps

  • Stream skills (P0): ~/.claude/skills/<name>/SKILL.md + supporting files. Record: {id, name, scope:"user"|"project", description, triggers, body, path, installed_from?}. Same weight as ChatGPT custom_gpts — user-authored customization.
  • Stream slash_commands (P0): ~/.claude/commands/*.md (+ per-repo ./.claude/commands/*.md). Frontmatter {description, argument-hint} + prompt body.
  • Stream memory_notes (P0): ~/.claude/projects/<p>/memory/*.md. The direct analog to ChatGPT memories — currently invisible. {id, project_path, filename, body, updated_at}.
  • Stream subagent_transcripts (P1): projects/<p>/<session>/subagents/*.jsonl. Sidechain stubs in the main jsonl reference but don't contain these; for heavy sessions they hold most of the model's work.
  • Stream tool_results_full (P1): resolve 500-char content_preview to the full .txt body via tool_use_id.
  • Stream file_snapshots (P1): metadata from ~/.claude/file-history/<uuid>/<hash>@v<n>. Opt-in for bytes.
  • Stream project_instructions (P1): per-repo ./CLAUDE.md + ./.claude/settings.json; scope derivable from session.cwd.
  • Stream settings singleton (P2): permissions, hooks, enabled MCP servers.
  • Stream todos (P2): ~/.claude/todos/*.json.

Deliberately omitted

  • .credentials.json — auth.
  • statsig/, telemetry/, debug/, stats-cache.json, mcp-needs-auth-cache.json — product telemetry.
  • shell-snapshots/, session-env/, plugins/cache/ — restoration caches / checked-out plugin code.
  • paste-cache/*.txt, history.jsonl — redundant with message content when actually used.
  • file-history/ bytes — size-prohibitive; metadata only unless user opts in.

Codex

Current manifest scope

sessions, messages, function_calls from ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl.

Source surface (~/.codex/)

  • sessions/YYYY/MM/DD/rollout-*.jsonl — rollouts (declared)
  • state_5.sqlite — thread index DB. threads has: id, rollout_path, created_at, updated_at, source, model_provider, cwd, title, sandbox_policy, approval_mode, tokens_used, has_user_event, archived, archived_at, git_sha, git_branch, git_origin_url, cli_version, first_user_message, agent_nickname, agent_role, memory_mode, model, reasoning_effort, agent_path. Also thread_dynamic_tools, thread_spawn_edges, agent_jobs, agent_job_items.
  • config.toml — model/effort/sandbox defaults + per-project trust_level map
  • prompts/*.md — user-authored slash-command templates
  • rules/default.rules — accumulated prefix_rule(pattern=[...], decision=...) trust decisions
  • skills/<name>/SKILL.md — user-installed skills (14 here)
  • memories/ — reserved for future memory feature (empty today)
  • logs_2.sqlite, shell_snapshots/, cache/, tmp/, models_cache.json, auth.json, installation_id, version.json — telemetry/auth/ephemeral

Gaps

  • Fields on sessions via state_5.sqlite#threads join (P0): title, archived, archived_at, tokens_used, first_user_message, sandbox_policy, approval_mode, agent_nickname, agent_role, memory_mode, git_origin_url. Users cannot reconcile their session list without title and archived. Either enrich sessions or add a threads stream — but the DB, not the rollout jsonl, is the truth for these fields.
  • Stream prompts (P0): ~/.codex/prompts/*.md. User-authored slash commands, same as Claude Code.
  • Stream skills (P0): ~/.codex/skills/<name>/SKILL.md. Same as Claude Code.
  • Stream approval_rules (P1): ~/.codex/rules/default.rules, accumulated prefix_rule decisions. {id, pattern, decision, updated_at}.
  • Stream config singleton (P1): config.toml — model default, effort, sandbox_mode, per-project trust_level map.
  • Stream agent_jobs (P2): Codex background-agent state. Low priority for v0.1.

Deliberately omitted

  • auth.json, installation_id — auth/telemetry.
  • logs_2.sqlite, shell_snapshots/, cache/, tmp/, .tmp/, models_cache.json — operational.
  • Encrypted reasoning traces — opaque by design.
  • AGENTS.md when symlinked out to dotfiles (user manages externally).

Cross-cutting observations

  1. User-authored customization is the systematic blind spot. All three connectors capture runtime transcripts but miss authored artifacts: Custom GPTs + custom instructions (ChatGPT), skills + commands + memory notes (Claude Code), prompts + skills + rules (Codex). These are the highest-leverage, lowest-volume pieces of user data and their absence would be the most embarrassing gap in a "true export" promise.

  2. "Sessions" is two tables on disk for the local connectors. Codex state_5.sqlite#threads is the truth for title/archived/tokens_used; the rollout jsonl is subordinate. Claude Code has projects/<p>/memory/*.md and projects/<p>/<session>/subagents/*.jsonl that single-file parsing never sees. Both manifests implicitly assume one-file-equals-one-session; the reality is hierarchical.

  3. Fan-out payloads are represented by IDs, not resolved. ChatGPT attachment_ids, Codex output_preview, Claude Code content_preview all leave the real payload behind. Acceptable for v0.1 but the manifest should advertise it so Layer 1 doesn't treat previews as missing fields.

  4. Filesystem scope is unstated. Claude Code / Codex both honor env overrides (CLAUDE_CODE_PROJECTS_DIR, CODEX_SESSIONS_DIR) but any new stream touching ~/.claude/skills, per-repo CLAUDE.md, or ~/.codex/rules requires a manifest-level declaration of filesystem footprint so reviewers can audit reach.

  5. Registry parity is the payoff. If the three manifests add customizations + authored memory, they converge on the same mental model — runtime transcripts + user-authored customization + user-curated memory — which is worth more than any single field.