@tangle-network/harness-sessions

@tangle-network/harness-sessions

One reader for every coding-agent harness's native session.

Each harness persists its session in its own store: Claude Code and Pi append JSONL, Codex appends a rollout JSONL, and OpenCode writes SQLite. This package declares each store's layout once and parses each format into one normalized HarnessSession. The session holds the messages, the tool calls with their inputs and results, every model call with the model that served it and its usage, and the error the session ended with.

npm install @tangle-network/harness-sessions

Where this fits

• The catalog serves STORES and SessionFormatId for capture consumers. The sidecar's harness session contract (agent-container packages/cli-agent-registry) currently mirrors these declarations; its native-version adoption remains a separate consumer change.
• The capture path reads captures through listCapture / readCapture.
• Discovery's trace gate, agent-eval, agent-record and the blog read sessions only through this package.
• Its dependencies are @tangle-network/agent-core, @tangle-network/agent-trace-contract and Node built-ins (node:sqlite for OpenCode).

Formats

Harness	Format	Store under HOME	Served model from	Usage
claude-code	claude-code.projects-jsonl	.claude/projects/*/{id}.jsonl, subagents at */{id}/subagents/**/*.jsonl	provider response (message.model)	per response, counted once per message.id
codex	codex.rollout-jsonl	rollout-*-{id}.jsonl under sessions/ and archived_sessions/ of every Codex home (below)	turn context (the configured model)	per response (last_token_usage); session total from total_token_usage
opencode	opencode.sqlite	.local/share/opencode/opencode.db (shared by every session)	session record (providerID/modelID)	per step (step-finish)
pi	pi.session-jsonl	.pi/agent/sessions/**/*_{id}.jsonl	session record (provider/model)	per assistant message
kimi-code	kimi.session-dir	.kimi-code/sessions/wd_<hash>/{id}/agents/main/wire.jsonl; retained .kimi/sessions/*/{id}/wire.jsonl and child wires	requested model recorded separately; served identity unknown	per step, excluding mirrored and wrapped child events
factory-droids	factory.session-jsonl	.factory/sessions/*/{id}.jsonl plus settings	configured model in settings (turn-context)	session total in settings

A HOME can hold several Codex homes, and codexReader.locate reads all of them (codexHomes(home) lists them):

• .codex and each sibling .codex-<name>;
• each acct account home, .config/agent-accounts/codex/<account>;
• home itself when it is a Codex home by its path (isCodexHome), such as a fleet seat's codex directory passed as a store root;
• for this process's own HOME only, $CODEX_HOME and $ACCT_CONFIG/codex/<account> in place of the default accounts directory. Another HOME, such as an archived host's store root or a capture, is read by its layout alone.

Homes that resolve to one directory count once. A session copied into several homes (acct move) is one ref, its largest copy. A lookup by id takes the rollout whose name ends in the id; when no name does, it reads the session_meta of each rollout modified since the id's UUIDv7 time.

listReaders() and the exported catalog own the complete installed format inventory.

Kimi uses the full session_<uuid> directory as its native ID, not the agent directory main. kimiReader.stores (also KIMI_STORES from the catalog entry) declares both layouts; the existing STORES['kimi.session-dir'] remains the legacy declaration for consumers that still run that producer. A store's optional globs overrides the reader's default file globs, including when reading a v2 capture whose payload names are flat. No state or configuration sidecar is needed to read the current wire.

The current flat wire records prompts and loop steps alongside appended message snapshots, usage mirrors and a token-counting context summary. The fold takes prompt input and loop content/tools/step usage once; it preserves separately recorded injected user context. llm.request supplies requestedModel, without promoting configured or usage labels into served identity. Explicit native endings map completed to completed, cancelled/prompt-aborted to aborted, and failed/blocked to error; an assistant-looking prefix stays open. A later matching prompt ending preserves the turn's structured error. Unknown ending reasons stay open with an integrity gap. Recorded step counters of zero remain zero, while absent usage remains unknown. The retained fixture proves the completed and prefix cases; failure/abort mappings follow the official producer schema and await real failure captures. Nonzero accounting and new-format child-agent discovery remain unproved.

Child execution evidence

Kimi records a child agent_id and parent_tool_call_id on each wrapped SubagentEvent, and writes the child's own wire under the parent's subagents/<agent_id>/ directory. kimiReader.locate(home, { nativeSessionId: parentId }) returns the parent and these child refs. A child-ID lookup selects that child. The reader keeps wrapped child events out of the parent's messages and usage; read each returned child ref for its own recorded work. kimiRefForFile preserves the same parent binding when a child wire is read directly. A missing child wire remains missing execution evidence, even when its ID appears in the parent.

Pi's session-header parentSession is fork/clone/handoff lineage; it does not prove a delegated subagent. The native reader does not turn it or conversation-entry parentId into subagent execution. Extension workflow execution remains a separate recorded source. Factory Droid also needs an explicit producer-recorded parent/child binding before this reader can expose delegated children. Neither reader guesses relationships from a tool name, working directory, timing, or current project.

A bounded local capture inspection on 2026-10-25 found explicit Kimi child events and separate child wires. Its Pi and Factory Droid samples contained no explicit subagent binding: the latest 20 files per harness held 7,724 and 245 records respectively; an additional size-stratified selection read seven Pi files (8,229 records) and 20 Factory files (252 records). These are sample limits, not a claim that every producer version lacks the metadata. The remaining Pi/Factory child coverage is unknown. Closing it requires a retained parent/child capture with an explicit source binding, followed by a reader-to-OTLP integration proof; the graph-corpus integrating owner owns that follow-up.

Use

import { readerFor, listCapture, toChatMessages, toOtlpSpans, toTurns } from '@tangle-network/harness-sessions'

// Sessions in a HOME, through the harness's declared store
const reader = readerFor('pi')
for (const ref of await reader.locate('/home/agent')) {
  const summary = await reader.summarize(ref) // constant memory; a 537 MB rollout reads in ~200 MB RSS
  const session = await reader.read(ref)      // full fidelity: nothing is truncated
}

// Sessions in a capture (provider retention layout today, raw-evidence archive v2 when given)
const { sessions, missing } = await listCapture(captureDir)

Projections trim; the reader never does:

• toChatMessages(session): agent-eval's chat-with-tools messages.
• toOtlpSpans(session): contract spans (@tangle-network/agent-trace-contract), one LLM span per model call and one TOOL span per tool call.
• toTurns(session): the blog's turns, with the model, usage and error of each assistant turn.

For callers that are not JavaScript:

harness-sessions read <file|home|capture> [--harness h] [--session id] [--summary] [--json]

Rules the readers keep

• The served model comes only from the provider response or the session record, never from the profile or the request. A failed request has no served model; requestedModel says which model it was sent to.
• usage is null when the harness recorded none; it is never reported as zero. input excludes cache reads and cache writes, which have their own fields.
• A session with zero model calls is not the trace of a turn.
• locate(home, { cwd }) keeps a session whose recorded cwd is cwd or lies under it on a path boundary (withinPath): /w/agent matches /w/agent/sub, never /w/agent-dev-container. Every store is resolved against the home given; only the Codex environment overrides above depend on the process.
• An unparsable record is counted in integrity.unparsedRecords and never stops a read (corruption: 'strict' throws instead). A record cut off at the end of a live copy is reported as a torn tail, not as corruption.
• ending.status is completed, error, aborted or open. open means the records stop mid-turn: the turn was still running or the process died.
• An OpenCode store is opened read-only where it lies, with no copy: SQLite never writes the store or its write-ahead log. With a log present it reads the log's committed rows and keeps only its own shared-memory index (-shm) current; without one the store is opened immutable and nothing is created beside it. A store SQLite cannot open in place (a log with no -shm in a directory it cannot write, such as a capture on read-only storage, or any store under Node before 22.15, whose node:sqlite takes no URI file names) is read from a private copy. integrity.sourceFiles names the session as <store>#session=<id> with a digest of the rows read for it.

Tests

The baseline fixtures are sessions that the real harness CLIs wrote while talking to the trace-proof scripted model; the tests assert that script. The Kimi child fixture retains complete existing parent and child wire records with all content removed, IDs pseudonymized, and source hashes, record counts, redaction version and token counters recorded in test/fixtures/kimi/provenance.json. Its integration test exercises discovery, full/summary reads, in-memory input and contract-span projection, preserving parent and child usage separately. The current Kimi fixture in test/fixtures/kimi2/ retains all 33 actual records through agent-eval's share pipeline and content minimization, with provenance alongside it. Its real-file integration covers both store layouts, direct-file CLI identity, v2 capture joins, projections, and incomplete prefixes without invoking a native harness or model. The proof against the readers this package replaces is a comparison over stored sessions, recorded with each release.