cc-transcript 3.0.0

The Claude Code session-activity platform: a superset JSONL parser (Rust fast path, Python reference), typed tool calls with a cross-language content digest, turns/edits/evidence over one spine, context windows with labeled fidelity, unified decision and correction ledgers, LLM judging, and a transcript-investigation CLI.

cc-transcript

cc-transcript parses Claude Code's on-disk JSONL transcripts into a typed superset event model — every entry type preserved, nothing dropped — so you build on one faithful representation and apply your own semantic filtering on top.

The one property that makes it worth using: the parser is non-lossy. It never silently discards sidechains, synthetic turns, tool results, or unrecognized entry types; filtering is opt-in and lives in your code, not buried in the parser. It ships as a Python library, a uvx-runnable CLI, and a Claude Code plugin.

Install

uv add cc-transcript        # or: pip install cc-transcript
uvx cc-transcript --help    # CLI, no install needed

Quickstart

Discover the transcripts on disk, parse one, and look at the events:

import anyio

from cc_transcript import AssistantEvent, TranscriptDiscovery, UserEvent, parse_events_from_bytes

path = anyio.run(TranscriptDiscovery.find_transcripts)[0]
events = parse_events_from_bytes(path.read_bytes())

for event in events:
    match event:
        case UserEvent(text=text):
            print("user:", text[:80])
        case AssistantEvent(model=model, text=text):
            print(f"assistant ({model}):", text[:80])

Compose a filter from small builders and apply it. The builders return clauses, build_spec assembles them into a spec, and apply_spec yields the survivors:

from cc_transcript import apply_spec, build_spec, keep_only, drop_junk, drop_short

spec = build_spec(keep_only("user", "assistant"), drop_junk("structural"), drop_short(2))
clean = list(apply_spec(events, spec))

NOISE_SPEC is a ready-made spec for the universal structural noise (system reminders, local-command output, skill banners).

The CLI

Six commands — list, show, grep, stats, slice, digest — and every one runs as uvx cc-transcript ..., no install step. list finds transcripts, newest first:

$ uvx cc-transcript list --limit 3
2026-06-11 19:27    1.0MB ~/.claude/projects/-Users-yasyf-Code-captain-hook/d2ca206a-2561-4c2c-9a4c-3ecaac9f8443/subagents/agent-a804d9aea43a110b5.jsonl
2026-06-11 19:27   70.6KB ~/.claude/projects/-Users-yasyf-Code-cc-transcript/4c77d556-8694-4613-8f50-253d905da68e/subagents/agent-affd5dbe069a3660d.jsonl
2026-06-11 19:27  740.8KB ~/.claude/projects/-Users-yasyf-Code-cc-transcript/4c77d556-8694-4613-8f50-253d905da68e.jsonl
3 of 6608 transcripts under ~/.claude/projects

stats summarizes a session before you read any of it:

$ uvx cc-transcript stats ~/.claude/projects/-Users-yasyf-Code-cc-transcript/4c77d556-8694-4613-8f50-253d905da68e.jsonl
files        1
events       181
kinds        other 68 · assistant 53 · user 33 · mode 22 · system 5
models       claude-fable-5 53
tools        TaskCreate 10 · Agent 5 · Read 5 · TaskUpdate 5 · Bash 2 · ToolSearch 2 · AskUserQuestion 1 · ExitPlanMode 1
text         14.8KB
thinking     8.7KB
tool io      89.0KB
sessions     1
span         2026-06-12 01:07:55 → 2026-06-12 02:28:03
interrupts   0
tool errors  0
sidechain    0

show renders one compact line per event; --signal keeps the conversational spine, and the index column is the event's position in the raw file:

$ uvx cc-transcript show ~/.claude/projects/-Users-yasyf-Code-cc-transcript/4c77d556-8694-4613-8f50-253d905da68e.jsonl --signal --tail 4
  189 asst  02:30:49 [claude-fable-5] Bash(rg -A3 'name = "great-docs"' /Users/yasyf/Code/cc-transcript/uv.lock | head -6; echo ---; rg -n "cl…)
  194 asst  02:31:31 [claude-fable-5] "`cli:` support confirmed in the pinned great-docs. Checking the exact config shape before writing:"
  195 asst  02:31:31 [claude-fable-5] TaskUpdate(8)
  196 asst  02:31:32 [claude-fable-5] Bash(sed -n '40,60p;1750,1790p' /Users/yasyf/.cache/uv/git-v0/checkouts/a9f52a54772f9b4e/d318527/great_d…)

grep searches event content; hit indexes feed straight back into show --range:

$ uvx cc-transcript grep -i "filterspec" --kind user --max-matches 3 ~/.claude/projects/-Users-yasyf-Code-cc-transcript/4c77d556-8694-4613-8f50-253d905da68e.jsonl
== ~/.claude/projects/-Users-yasyf-Code-cc-transcript/4c77d556-8694-4613-8f50-253d905da68e.jsonl
   16 user  01:12:00 <-Agent (10161ch) ## Findings Report: cc-transcript Repository Based on a thorough exploration of `/Users/yasyf/Code/…
   29 user  01:16:29 <-? (1378ch) /Users/yasyf/Code/cc-transcript/cc_transcript/: total 8648 drwxr-xr-x@ 19 yasyf staff 608 Jun 11 17…
   69 user  01:36:17 <-Read (4247ch) 1 """Composable builder fragments for :class:`~cc_transcript.FilterSpec`. 2 3 Each fragment returns…
1 files, 3 matches

The output is compact by design — one line per event, hard truncation — so an agent triages a session in a few hundred tokens instead of paging through megabytes of JSONL.

The last two verbs serve programs, not people: slice emits per-tool-call JSONL for a session UUID and time window (the language-neutral bridge cc-review consumes), and digest generates and checks the cross-language fixture corpus for the tool-digest contract.

Claude Code plugin

Install the bundled plugin from inside Claude Code:

/plugin marketplace add yasyf/cc-transcript
/plugin install cc-transcript@cc-transcript

The plugin's skill teaches Claude to answer questions about its own history — "what did I ask yesterday", "find the session where we fixed the parser" — by funneling through the CLI's list, stats, grep, and show commands instead of reading raw JSONL.

What problems does this solve?

• One faithful parse. Anything reading Claude Code transcripts re-implements the same JSONL quirks (str-or-list content, tool results nested two ways, envelope-less mode markers). This is that parser, written once and typed strictly.
• Non-lossy by design. The event model is a superset: sidechains, <synthetic> turns, thinking blocks, and unrecognized entry types all survive parsing. You decide what to drop, via composable filter specs (build_spec).
• Incremental ingestion. FileStateStore tracks per-file mtimes in SQLite (WAL, safe across concurrent tasks) so re-runs only reparse changed files, and you compose your own writes in the same transaction.
• Two engines, one contract. A single Backend protocol with two implementations: RustBackend (PyO3 + rayon) is the default fast path, and PythonBackend is the readable reference — parity-asserted against each other. Filter specs are portable, so a spec built in Python runs Rust-side without giving up the fast path.
• One activity spine. SessionActivity lifts parsed events into turns, typed tool calls, and first-class edits; ContextWindow persists EventRefs plus labeled previews and hydrates back to full fidelity while the transcript lives; rendering is Budget-bounded in exactly one place.
• Mining, judging, and sentiment. cc_transcript.mining mines transcripts for user feedback — detectors, confidence calibration, candidate filtering, and the feedback store; cc_transcript.judge runs fidelity-aware LLM verdict passes over the mined corpus; cc_transcript.sentiment scores conversational sentiment per time-bucketed conversation window.
• Transcript investigation for agents. The CLI answers "what happened in that session" in a few hundred tokens, which is what makes the Claude Code plugin viable.

Docs

Each section of the docs site is a focused guide:

• Getting Started — install, parse, filter, persist.
• Filtering events — clauses, specs, and NOISE_SPEC.
• Scoring sentiment — the lexicon engine and score specs.
• Rust/Python backends & parity — the Backend protocol and parity testing.
• Compose your own policy — building a bespoke filtering policy.
• Mining feedback — detectors, confidence, candidates, and verdicts.
• The transcript CLI — list/show/grep/stats end to end.
• API reference — the complete typed surface.

License

PolyForm Noncommercial 1.0.0.