sequential-read（顺序读取）

📖 Sequential Read

Read prose (novels, non-fiction, articles) by ingesting content in semantic chunks and building structured reflections iteratively. The output captures how your perspective developed over the course of reading — predictions that were wrong, questions that got answered, opinions that shifted — not just a retroactive summary.

Invocation

Execution Model

The pipeline runs in spawned sub-agents. Novel-length reads are a two-phase process: a main reader that handles the bulk of chunks, then a finisher that completes the remaining chunks and writes synthesis. This is the normal flow, not an error.

When the user invokes /sequential-read:

1. Parse the command to extract the file path and optional lens
2. Pre-create the session:

   python3 {baseDir}/scripts/session_manager.py create <source-file>

1. Spawn the main reader sub-agent:

   sessions_spawn with label: reader-{session-id}
   Tell the agent: "Session already exists at {session-id}. Do NOT create it again."

1. Tell the user the session has started and they'll be notified when it's done
2. When the main reader returns (whether it completed or died mid-read):

- Check session status: python3 {baseDir}/scripts/session_manager.py get <session-id> - Check how many reflections exist vs total chunks - If synthesis exists: Done. Present results. - If chunks remain or synthesis is missing: Spawn a finisher sub-agent (see below). This is the expected path for novels.

1. When the finisher returns, present the synthesis and session path.

The Two-Phase Pattern

For novels (~20+ chunks), the main reader typically handles ~17-20 chunks before its context fills and the session ends. This is expected behavior, not failure. The finisher picks up the remaining 2-5 chunks and writes the synthesis with full context of all prior reflections.

Spawning the finisher:

sessions_spawn with label: finisher-{session-id}, model: "opus"
Task: "Resume reading session {session-id} at {baseDir path}.
  Read reflections written so far to understand context.
  Continue from chunk N (the next unwritten chunk).
  Write remaining reflections, then run synthesis.
  Session path: {session-path}"

Do not wait or ask the user between the main reader and finisher. When the main reader returns without a synthesis, immediately spawn the finisher. The whole pipeline should be hands-off.

Script Paths

All Python scripts are in {baseDir}/scripts/:

• {baseDir}/scripts/session_manager.py
• {baseDir}/scripts/chunk_manager.py
• {baseDir}/scripts/state_manager.py

Templates are in {baseDir}/templates/:

• {baseDir}/templates/reflection_prompt.md
• {baseDir}/templates/synthesis_prompt.md

Commands

/sequential-read <path-to-file> [--lens <persona>]

1. Create or Resume Session

python3 {baseDir}/scripts/session_manager.py create <source-file> [--lens <persona>]

This command handles resume detection automatically:

• If an in-progress session exists for the same source filename, it prints the existing session-id and path
• Otherwise it creates a new session

Capture the session-id from the first line of output.

2. Check Session Status (for resumed sessions)

python3 {baseDir}/scripts/session_manager.py get <session-id>

Check the status field to determine where to resume:

3. Run the Pipeline

For a new session or preread status:

Run the preread sub-skill ({baseDir}/preread/SKILL.md) with:

• SESSION_ID = the session-id
• SOURCE_FILE = path to the source text
• BASE_DIR = {baseDir}

For chunked status (or after preread completes):

Run the reading sub-skill ({baseDir}/reading/SKILL.md) with:

• SESSION_ID = the session-id
• BASE_DIR = {baseDir}
• LENS = the lens value (or null)

For read status (or after reading completes):

Run the synthesis sub-skill ({baseDir}/synthesis/SKILL.md) with:

• SESSION_ID = the session-id
• BASE_DIR = {baseDir}

4. Present Results

After synthesis completes, send the user:

• The full synthesis text
• The session path: memory/sequential_read/<session-id>/
• A brief note: how many chunks, whether a lens was used

/sequential-read list

python3 {baseDir}/scripts/session_manager.py list

Print the output to the user.

/sequential-read show <session-id>

python3 {baseDir}/scripts/session_manager.py get <session-id>

If status is complete, read and display:

memory/sequential_read/<session-id>/output/synthesis.md

If not complete, show the session status and progress.

Model Guidance

The reading phase is the most demanding — it runs for many iterations and must sustain quality throughout. Choose the model based on source length:

When spawning the sub-agent, set the model explicitly: model: "opus" for novels.

Why this matters: Lighter models degrade over long reading sessions — reflections become stubs as context accumulates. The first test run of this skill on Sonnet with a 35-chunk novel produced 4 genuine reflections and 31 placeholders. Opus is required for novel-length works.

Chunk sizing: The structural chunker targets ~550 lines per chunk (range 200-700). For a typical novel (~10-12k lines), this produces ~20 chunks. Longer texts (15k+ lines) may produce 35+ chunks and will need a finisher session (see below).

The two-phase pattern is standard. For novel-length works (20+ chunks), always expect to spawn a finisher after the main reader. The main reader handles ~80-90% of chunks; the finisher handles the rest plus synthesis. For very long texts (35+ chunks), the main reader may only get ~25 chunks. Plan accordingly -- this is the normal pipeline, not error recovery.

Pre-create sessions: Always create the session with session_manager.py create BEFORE spawning the sub-agent. Tell the agent the session already exists and not to create it again. This avoids failures from duplicate creation attempts.

Reader Context (Optional)

If you maintain reader-mind files (accumulated reading context — character knowledge, thematic threads, critical framework), load them into the sub-agent's task prompt as preamble. This gives the reader continuity across books in a series.

Include context in the spawn task:

"Before you begin reading, here is your accumulated reader context:

=== READING CONTEXT ===
[contents of reader-mind file]

Now read [book title]..."

After synthesis, update reader-mind files with new character knowledge, thematic thread updates, and cross-reference observations. Revise rather than append. Keep under ~4000 words per file.

Important Guidelines for the Reading Agent

• No peeking. Each reflection must be written from the perspective of not knowing what comes next. Do not reference content from later chunks.
• Be honest. Confusion, boredom, excitement, disagreement are all valid reactions. Don't perform engagement.
• Be specific in revisions. "I was wrong about X because Y" beats "my view has evolved."
• The lens is a suggestion. If the lens feels forced for a particular chunk, note that in the reaction rather than straining to apply it.
• Run autonomously. Do not stop to ask the user questions between chunks. The entire pipeline is set-and-forget.
• Persist everything. Every reflection and state update is saved to disk before moving to the next chunk, enabling resume on interruption.

Post-Synthesis

After synthesis is complete, you can integrate the output into whatever workflow you prefer — blog posts, reading logs, knowledge graphs, series trackers, etc. The synthesis file at output/synthesis.md is self-contained and portable.