Read Claude Code History

Read Claude Code History

Read Claude Code evidence only. Do not resume the old process, edit its project, or turn a read request into a continuation task. When the user later asks to act, hand the verified evidence to daymade-claude-code:continue-claude-code-work.

Route by the requested result

找「我们之前做的那个 X」（产物类目标）的三个判据

按词搜不到时换这三个判据，别扩词硬搜：

1. 时间约束是第一筛子。用户给了时间窗（「前两天」「本周」）→ 先按 mtime 筛文件系统（find <root> -newermt '<日期>' -name '*.html'），候选面缩到几十条再按形态排。全库关键词搜索应该在时间筛之后，不是之前。
2. session scratchpad 必须在搜索面内（/private/tmp/claude-$(id -u)/**/scratchpad/，macOS 单用户通常是 claude-501）。Claude session 的默认产物落点在那，不在 workspace——「我们之前写的 X」有相当概率躺在 scratchpad。找产物类目标时显式包含它。
3. 恢复出的版本 ≠ 最终版。从 session jsonl 恢复文件（取 Write 的 content）只拿到某次全量写，后续增量 Edit 不会自动合入——恢复产物可能只有真实文件的零头（实测：28KB 恢复片段 vs 2.45MB 完整版）。打开/渲染后觉得「不像目标」时，先怀疑「我拿到的不是完整版」（原路径文件还在不在、session 里还有没有后续 Edit），不要直接排除目标。

战例（2026-09-18）：找「之前做的聚类网页」，十几轮全库关键词搜索无果；目标一直在 9-13 的 session scratchpad（favorites-ledger.html，2.45MB），恢复的 28KB 片段渲染成裸样式被误判排除，用户给出精确路径后才定位。

The requested output wins over the background story. If the user asks for a chronological table of their raw inputs, return that table; do not replace it with a topic analysis because their motivation mentions an incident.

Evidence surface and completeness

By default, discover every active Claude config home and every archive registered in ~/.claude/history-sources.json. De-duplicate physical copies by Session ID and content identity, and use record timestamps rather than file mtime. A result scoped to one explicit --home is a diagnostic slice, not a completeness claim.

Treat Claude's record labels as storage metadata, not authorship proof. A top-level type: user record can contain a command envelope, hook boilerplate, a whole pasted document, agent-voiced text, or a system placeholder. Human text typed while the assistant was busy can live in attachment.queued_command.prompt with origin.kind: human; do not lose those corrections by reading only user records.

Read references/session_file_format.md when interpreting schemas, authorship, sidechains, attachment records, compaction, or file-history snapshots. Read references/hybrid_history_recall.md before building or repairing the optional BM25/vector index. Read references/workflow_examples.md for indexed discovery, triage, and recovery examples. Read references/claude_session_format.md when you need the layout rather than the message schema — where sessions live on disk, how project paths are normalized into directory names, the sessions-index.json fields, and the compact_boundary markers that tell you a transcript was summarized rather than truncated.

Commands

Resolve every script relative to this SKILL.md; do not search the machine for a same-named helper or recreate a JSONL parser inline.

Indexed discovery

<skill-dir>/scripts/history_index.py status
<skill-dir>/scripts/history_index.py recall '<topic>' --provider claude --mode bm25

Expected output: candidate Session IDs and index coverage. This is topic discovery, not a complete recent-session inventory. If a complete list is needed, report that the current bundled inventory would scan raw bodies.

Exact Session evidence

<skill-dir>/scripts/read_claude_session.py --session <SESSION_ID>

# Add this only when the caller intentionally wants to restrict lookup to one workspace.
<skill-dir>/scripts/read_claude_session.py --session <SESSION_ID> --project <workspace>

Expected output: # Claude Code Session Evidence Briefing, session identity, compaction boundary, chronological user/assistant handoff, queued human prompts, end reason, unresolved calls, subagent state, files touched, memory, and current workspace state. The exact reader always parses every physical Session record, including records before compaction; --full only removes output character clipping. It checks active and registered-archive copies, accepts only identical or strict append-only supersets, and fails visibly on divergent copies, multiple Session identities, a missing record-level Session identity, malformed JSONL, or unreadable bytes. With an exact Session ID and no --project, it searches every project across the discovered active homes and registered archives; an explicit --project remains a strict scope. A filename alone never proves Session identity. Read Local runtime and working-directory evidence before interpreting task endings, local-command output, or cwd receipts.

Command behind a hook audit row

Hook audit logs often store sid=<SESSION_ID> cmd_sha256=<HEX> rather than the command text. Resolve a row to the tool call that produced it:

<skill-dir>/scripts/read_claude_session.py --session <SESSION_ID> --find-command-sha256 <HEX>

Expected output: a # Tool calls whose command has SHA-256 … heading, the number of transcript files and tool calls examined, the match count, and for each match the file and line, timestamp, record uuid, tool name and id, cwd, and the full command in a fence. Exit 0 means at least one match, 1 no match, 2 a malformed digest or missing --session. The digest is SHA-256 over tool_input.command as UTF-8. A hook event raised inside a subagent carries the parent Session ID, so the lookup also searches <SESSION_ID>/subagents/*.jsonl; a main-transcript-only search misses those calls. Unlike the briefing, the lookup skips and counts unparseable lines so it still works on a Session that is being written; a zero with skipped lines does not prove the call is absent.

Indexed content search

<skill-dir>/scripts/history_index.py status
<skill-dir>/scripts/history_index.py recall '<keyword>' \
  --mode bm25 --provider claude --exclude-session <CURRENT_ID>

Inspect the index's provider coverage and last indexed time. Open only matching Session IDs with the exact-session reader to verify original records. Indexed recall returns ranked prose candidates, not a census of thinking, tool results, attachments, or unindexed records. The raw analyze_sessions.py search entry is disabled for live stores: its date flags filter after reading the files.

Use --role user or --role assistant to narrow the stored role, and repeat --phrase '<literal text>' to require exact substrings. These filters constrain both BM25 and vector candidates; --terms constrains only the FTS leg. Returned record_key, source_kind, and human_authorship describe provenance, not an authenticated human speaker. Inspect the original record before saying “you said”. Codex tool results are outside this prose index; hand a selected Codex candidate to read-codex-history's original tool/record evidence route. Export original values without automatic content redaction. A result about a neighboring fact does not answer the requested proposition.

Human-input export

<skill-dir>/scripts/extract_user_messages.py \
  <persistent-output-base> --days 7 --group-by session

This produces Markdown and HTML. It separates storage pollution from human prose and recovers queued prompts. Preserve timestamps, duplicates, and Session boundaries; do not add a second thematic classification unless asked.

Deleted-content recovery

Recovery writes files, so keep it separate from ordinary reading. First run the recovery report against the exact Session file, review every proposed destination, then write only after the user asked to recover content. Never restore directly over the current project tree.

Skill usage ledger

Answers "which Skills do we actually use, and who starts them". It is an incremental index (~/.claude-history-index/skill-usage-v1.db): the first index parses every Claude and Codex session once, later runs re-parse only files whose size or mtime changed. Reports read the ledger, never raw history.

python3 scripts/skill_usage_ledger.py index            # --no-codex to skip Codex
python3 scripts/skill_usage_ledger.py report --override user-invocable-only
python3 scripts/skill_usage_ledger.py report <skill> [<skill> ...] --since 2026-06-01 --until 2026-09-24
python3 scripts/skill_usage_ledger.py status           # freshness + what is not covered

Columns: user_command (user typed /X, or Codex $X), model_named (the model invoked X and the latest human prompt contained X's name), model_auto (the model invoked X unprompted), last_model_use, blocked (Claude refused a model call because X is not model-invocable — the model reached for it and could not use it; not counted as use), and the Skill's current skillOverrides value. --override STATE also lists Skills in that state with zero recorded use. Rows group by bare name, which is what skillOverrides keys on; identities lists the qualified forms merged into a row. More than one namespace there (review, suite:review) means same-named Skills were counted together — rerun with --exact to split them.

Read these limits into every conclusion:

• A Skill that is user-invocable-only cannot be model-invoked, so its model_auto count stops at the day it was hidden. Compare use before and after that date, not the lifetime total.
• Codex model use is inferred from the model reading one .../X/SKILL.md; a session that edits or audits X's SKILL.md with plain shell reads looks the same. Several SKILL.md files read by one command are recorded as bulk_read and excluded.
• Not covered: Skills invoked inside Claude subagents, Kimi CLI, and Codex Skills followed from memory without a read. Zero recorded use means "not seen in the covered stores", not "never used".