Which coding-agent formats are supported?

Sessions Viewer supports the seven layouts below. Adapter review and synthetic command tests are not CLI end-to-end certification. This matrix describes the 0.6.0 implementation reviewed on 2026-10-05; it does not promise compatibility with every release.

Capability and evidence matrix

All seven have history browsing/search, export and terminal resume. Only Claude Code and Codex have in-app chat. Resume requires the corresponding installed/configured CLI; syntax is documented in the command comparison.

Agent	Format covered	In-app chat	Validation evidence	CLI release tested end-to-end
Claude Code	JSONL message.content strings/blocks	Yes	Synthetic extraction + Rust parser projection	Not validated in this docs pass
Codex	session_meta, event_msg, response_item	Yes	Synthetic cwd/text + Rust parser projection	Not validated in this docs pass
Grok Build	ACP updates in updates.jsonl	No	Synthetic object/array chunks	Not validated in this docs pass
Kimi Code	Main wire primary events; adapter has legacy fallback	No	Synthetic primary events only	Not validated in this docs pass
Pi	Session header and message/tree entries	No	Synthetic text from all stored branches	Not validated in this docs pass
Antigravity CLI	Step-based transcript*.jsonl	No	Synthetic multiline XML extraction	Not validated in this docs pass
opencode	SQLite session, message, part	No	Synthetic read-only SQL queries	Not validated in this docs pass

OS installers are provided for macOS, Windows and Linux; the shell examples use macOS/Linux syntax. This is not a per-OS GUI certification. Antigravity's supported format has no usage fields. Older layouts, custom roots, images, tools, branches and provider behavior need the relevant reference page and troubleshooting.

Download safe, reproducible examples

These are intentionally minimal synthetic extraction fixtures, not real conversations or full CLI-ready sessions. Optional image/tool fields are incomplete; do not import them as real history or resume them. Paths and prompts are invented. The JSONL/SQL files and expected output manifest are the same inputs checked by npm run docs:test.

Download a sample, then replace the filename in its agent reference's text-extraction command. Expected lines are listed in order:

Agent	Download	Expected text
Claude Code	claude-code.jsonl	Question, Answer, Legacy string
Codex	codex.jsonl	Question, Answer; metadata cwd is /synthetic/project
Grok Build	grok-build.jsonl	Question, Answer
Kimi Code	kimi-code.jsonl	Question, Answer, Legacy string
Pi	pi.jsonl	Question, Answer, Other branch (not active-branch-only)
Antigravity CLI	antigravity-cli.jsonl	Question, on two lines (one multiline prompt)
opencode	opencode.sql	user and Question from the text query

For example, after downloading claude-code.jsonl:

jq -r 'select(.type=="user" or .type=="assistant")
  | .message.content
  | if type=="string" then . else .[]? | select(.type=="text") | .text end' \
  claude-code.jsonl

Question
Answer
Legacy string

The opencode sample is SQL that creates a reduced schema. Load it into a new temporary database, never your actual opencode database:

temp_dir="$(mktemp -d)"
sqlite3 "$temp_dir/synthetic.db" < opencode.sql
sqlite3 "file:$temp_dir/synthetic.db?mode=ro" \
  "SELECT json_extract(data, '$.text') FROM part WHERE json_extract(data, '$.type') = 'text';"

Expected output is Question. The regression suite also verifies that parent-only session listing excludes the sub-agent, cost aggregation includes it, and read queries leave the synthetic database unchanged. It does not validate provider billing or a real opencode migration.

What do the offline parser tests prove?

Keep three verification scopes separate:

1. jq / SQL extraction — npm run docs:test runs the documented commands on the small samples above. It also checks content/SEO regressions and screenshot integrity. Command output is not the application's parser output.
2. Actual Rust file parsers — on 2026-10-06, two isolated tests passed on macOS for the 0.6.0 Claude Code and Codex parsers and their message post-processing. They compare roles, text, Claude thinking, tool names/arguments/IDs/results and message order with a public expected projection. They also check that parsing leaves the temporary input unchanged. Codex receives an empty title index instead of loading the user's real one. No home discovery, CLI, tool execution or provider request runs.
3. Real CLI / GUI end-to-end acceptance — still separate and not performed here. Image behavior, browsing/search/export/resume, installed CLI releases and provider behavior are not certified by those parser fixtures. No verified CLI version has been added to the matrix.

Offline parser fixture	Expected messages
claude-code-adapter.jsonl	User question → assistant thinking/text/tool call → user tool result → legacy user string
codex-adapter.jsonl	User question → assistant tool call → user tool result → assistant answer; duplicate response text and unrelated metadata are ignored

Download adapter-expected.json for exact checked fields. These richer synthetic fixtures are still not complete importable or resumable CLI sessions. Tool arguments, including echo synthetic, are data, not executed commands.

From the repository root, with Rust/Tauri build dependencies already cached:

npm run docs:test
npm run docs:test:adapters

The adapter command uses cargo test --manifest-path src-tauri/Cargo.toml --lib --locked --offline docs_public_fixture. It runs only these two tests and fails rather than fetching uncached crates. The existing Rust CI job includes them; the Node documentation job does not require Rust. This is reproducible local evidence, not proof of production deployment or indexing.

How to report a compatibility change

Include app version, CLI version, OS, record layout, custom-root configuration and the failing operation. Share a minimal synthetic/redacted shape, not credentials or a full private database. A CLI version is added to a verified-runtime matrix only after that version and operation actually run successfully.

Evidence: adapter source revision 22fefc6. See privacy, installation and project maintenance.