Own the context window

When

Hour two of a task, or any time the transcript is treated as the product.

Do

1. 01

   Lay out the window

   Pinned: goal, constraints, skill index (name+description), tool schemas. Working: last N already-compacted events. Recalled: k retrieved hits. Compacted: running summary.

2. 02

   Bulk to files

   HTML, CI logs, PDFs, JSON envelopes become files. The window gets a path and a byte count. Grep if needed.

3. 03

   Keep the prefix stable

   Cacheable system prompt + skill index. Do not shuffle a timestamp into the prefix. Never drop the goal first.

Don't

• Concatenate six PDFs 'just in case'.
• Use the full transcript as long-term memory.
• Silently truncate the goal to keep junk observations.

Context builder

def build_window(state, skills, events):
    pinned = [
        state.goal,
        state.constraints,
        skill_index(skills),
        tool_schemas(state.allowed_tools),
    ]
    working = events[-8:]
    recalled = retrieve(state.goal, k=5)
    compacted = [state.running_summary]
    return concat(pinned, compacted, recalled, working)

Refuse

• Context hoarding — Raw HTML, full CI logs, six PDFs, entire JSON envelopes stuffed 'just in case'. The goal is the first thing forgotten.
• Silent truncation — The goal is dropped to keep junk observations. The model continues confidently.
• Memory as unfiltered logs — Session transcripts stuffed into the next session. The user cannot inspect or delete them.

Load next

• Compact observationsEvery tool return. Compaction is not a later optimization.
• Write a skillExpertise should load sometimes, not always. The constitution (AGENTS.md) is too coarse; a prompt dump is too fat.
• Document / RAG agentUsers ask questions over private documents and you must not hallucinate a missing page.

Trigger tests