@razroo/iso-trace

Local observability for AI coding agent transcripts.

agentmd scores per-rule adherence, isolint lints prose portability, iso-harness fans out the harness layout, and iso-eval scores task success on synthetic workspaces. All of those work off signals you authored. Once the agent is in a real user's hands, the rest of the chain goes blind — and that's the gap this package closes.

iso-trace parses the transcript files Claude Code, Cursor, Codex, and OpenCode already write to disk, normalises them into one event model, and lets you ask the questions the rest of the chain can't:

• Which rules ever actually fire in production? (candidates to remove)
• Which tools does my agent reach for, in what order? (pattern drift)
• Which captured sessions would make good regression eval fixtures?
• Did the tool-call mix change after the last prompt edit?

Supported local sources today:

• Claude Code JSONL sessions in ~/.claude/projects/<encoded-cwd>/*.jsonl
• Cursor JSONL agent transcripts in ~/.cursor/projects/<encoded-cwd>/agent-transcripts/<chat-id>/<chat-id>.jsonl
• Codex JSONL sessions in ~/.codex/sessions/**.jsonl
• OpenCode sessions discovered from ~/.local/share/opencode/opencode.db and loaded via opencode export <session-id> for full-session reads. OpenCode model-score queries the SQLite store directly instead of replaying every export.

The event model is harness-agnostic, so consumers do not need to care where a session came from.

Zero upload. The tool never opens a network connection. Everything is local reads and stdout / user-specified file output.

Install

npm install -D @razroo/iso-trace

Event model

Every source is normalised into the same Session → Turn → Event[] shape:

type Event =
  | { kind: "message",    role: "user" | "assistant" | "system", text: string }
  | { kind: "tool_call",  id: string, name: string, input: unknown }
  | { kind: "tool_result",toolUseId: string, output: string, error?: string }
  | { kind: "file_op",    op: "read"|"write"|"edit"|"list"|"search", path: string, tool: string }
  | { kind: "token_usage",input: number, output: number, cacheRead: number, cacheCreated: number, model?: string };

file_op events are derived from the file-touching tools each harness emits (Read/Write/Edit, Codex apply_patch and common shell reads, OpenCode read/grep/glob, and so on). The original tool_call is always preserved alongside — nothing is lossy.

Claude Code thinking blocks are deliberately dropped on parse to keep exported data safer to share. Everything else round-trips.

CLI

iso-trace sources                # show detected transcript roots + parser status
iso-trace where                  # print the raw paths iso-trace scans

iso-trace list                   # recent sessions across every detected root
iso-trace list --since 7d
iso-trace list --cwd .           # only sessions whose cwd matches this project
iso-trace list --json            # machine-readable

iso-trace show <id-or-prefix>                       # full normalised event stream
iso-trace show cc_abcd1234 --events tool_call,file_op
iso-trace show cc_abcd1234 --grep "H3"              # regex across message + tool input

iso-trace stats                  # aggregate across all discovered sessions
iso-trace stats cc_abcd1234 cc_abcd5678
iso-trace stats --since 7d --cwd .
iso-trace stats --source path/to/sample.jsonl       # one file, no discovery needed

iso-trace model-score --cwd . --harness opencode --tool read
iso-trace model-score --since-hours 24 --harness opencode --tool read --fail-on-schema
iso-trace model-score --since 7d --tool Bash --json # model success/error scorecard

iso-trace export <id> --format jsonl --redact > session.jsonl
iso-trace export <id> --format json --redact-regex 'gho_[A-Za-z0-9]+'

iso-trace export-fixture <id> --out fixtures/my-task/         # lift a session into an iso-eval fixture
iso-trace export-fixture --source path/to/sample.jsonl --out fixtures/my-task/
iso-trace export-fixture <id> --out fixtures/my-task/ --runner claude-code --harness-source dist --edit-checks exists-only --run

Session IDs are 8-char prefixes derived from path + first-line hash, so they're stable across reads and unambiguous by design. Prefix matching follows git's semantics: unique prefix wins, multiple matches errors.

model-score — which models are using tools correctly?

model-score is currently available for Claude Code, Codex, and OpenCode. Cursor transcripts are parsed and exportable, but they do not yet expose stable model IDs or tool-result metadata, so Cursor is intentionally excluded from scorecards for now.

model-score groups tool calls by the model that emitted them and reports call volume, success/error counts, schema-error counts, and the latest observed timestamp.

For OpenCode, --tool read also breaks out the observed input shape, so you can spot schema drift immediately:

iso-trace model-score --cwd /path/to/project --harness opencode --tool read

If a weak route is sending read({ path: ... }) instead of read({ filePath: ... }), the scorecard will show that in the path / file_path columns together with the schema-error rate.

OpenCode windows are filtered by the actual tool event timestamp, not just the session start time, so long-running sessions do not hide fresh regressions near the end of a conversation.

For routing and schema guardrails, model-score can also fail non-zero:

iso-trace model-score \
  --cwd /path/to/project \
  --harness opencode \
  --tool read \
  --since-hours 24 \
  --fail-on-schema \
  --fail-on-model openrouter/minimax/minimax-m2.5:free

--fail-on-model is repeatable, so you can block multiple routes in one check.

export-fixture — turn an observed session into a regression fixture

When a real session does something you want to lock in as an iso-eval regression test, export-fixture lifts it into a fresh suite:

fixtures/my-task/
├── task.md      — the first user message from the session
├── workspace/   — empty placeholders for every file the agent read
└── checks.yml   — runner metadata plus file_exists per write; edits can
                   emit file_contains placeholders or exists-only checks

This is a seed, not a perfect replay — iso-trace can't know the agent's starting workspace, and it can't guess what "success" should assert. Review checks.yml and fill in any baseline workspace files, then drop the directory into an iso-eval suite and run iso-eval run fixtures/my-task/checks.yml.

For a faster path from "observed failure" to "rerunnable suite":

• --runner <fake|codex|claude-code|cursor|opencode> persists the intended eval runner.
• --harness-source <path> persists which generated harness files should be staged into each rerun workspace.
• --edit-checks exists-only avoids REPLACE_ME placeholders so the exported suite can smoke-run immediately.
• --run invokes iso-eval run directly after export.
• --redact and --redact-regex scrub the exported task text and paths before anything is written.

Library API

import {
  discoverSessions, loadSessionFromPath, filter, stats, exportSession,
} from "@razroo/iso-trace";

// discovery: autodetects Claude Code, Cursor, Codex, and OpenCode local sources
const refs = await discoverSessions({ since: "7d", cwd: process.cwd() });

// load + normalise one session
const session = loadSessionFromPath(refs[0].source.path);

// query
const bash = filter(session, (e) => e.kind === "tool_call" && e.name === "Bash");

// aggregate
const s = stats([session]);
// → { sessions, turns, durationMs, tokens, toolCalls, fileOps, filesTouched }

// export
process.stdout.write(exportSession(session, "jsonl"));

Bring your own roots (e.g. a CI agent that captures Claude/Codex JSONL under an arbitrary directory, or a host with OpenCode's default DB):

await discoverSessions({ roots: ["/path/to/captured/transcripts"] });

How this fits the rest of the monorepo

agent.md → agentmd → isolint → iso-harness → CLAUDE.md / AGENTS.md / .cursor/rules

                (user runs the agent in production — hours / days / weeks)
                                     │
                           transcripts on disk
                                     │    iso-trace
                                     ▼
                     normalised Session / Event stream
                 ┌───────────────────┼───────────────────┐
                 ▼                   ▼                   ▼
        iso-trace stats       export             agentmd
        (firing rates,        → iso-eval         adherence from
         tool patterns)        regression        production traces
                               fixtures

Other packages depend on iso-trace; the reverse never holds. A bare Claude Code user with none of the other @razroo/* packages installed still gets value from iso-trace alone.

Privacy posture

• Transcripts may contain code snippets, shell commands, and occasionally secrets. iso-trace never transmits them over the network.
• show and export write to stdout or user-specified paths only. The tool will never edit, move, or delete a transcript.
• --redact scrubs source paths, cwd/home prefixes, common API keys, bearer tokens, and private-key blocks. --redact-regex lets you add project-specific denylist patterns.
• export-fixture --redact applies the same scrubbing before it writes task.md, workspace/, and checks.yml.
• thinking blocks are stripped on parse so exports don't accidentally publish internal reasoning.
• Inspect exports before sharing anyway; redaction reduces risk, it does not make arbitrary transcripts safe by magic.

License

MIT — see LICENSE.