# Frazer Context and agents' sessions

> Frazer Context assembles the context of every agent session, review, plan check and panel with rules that put correctness before size, ranks with rules first and Frazer Rerank second, keeps what binds through every compaction, lets models of other families verify the result, catches old values in what models write, and writes a receipt of what was given. A session receives what binds it before its first model call, and Fir refuses writes that break it. Every part works with your own models.

Part of Frazer's docs for agents. Every page on frazer.ai: https://frazer.ai/llms.txt. Every docs page: https://frazer.ai/docs/llms.txt.

## How agents should use Frazer Context and agents' sessions

- Working from a plan, read the session context first (`get_session_context`, or `fir context`) and cite its anchors. What is In force binds you, and only a person lifts it; what no source holds is unknown, so ask.
- Put part 1 of the session context (In force, the task, conflicts, unknowns, the checklist) first in your input, the checklist last, and re-send part 1 after every compaction.
- A summary or a pruned part is a pointer, never a source of a rule: fetch the handle when the task depends on the full text.
- Acknowledge a material notice with `fir context ack <notice>` before your next layer; until then the push gate answers `stale_context`.
- Never set an item's state: claim it (`fir item claim`), write notes, and state claims with evidence (`fir assert`). An item is done only when its merged layer's proof passes. A write from a lost lease is refused (`stale_epoch`, `not_holder`): claim again or hand off.
- When verify returns a finding with strictness `ask` or `block`, or the old-value check flags a value you wrote, answer it (cite, fix or raise it) before your next layer.

## Pages

- [Frazer Context](https://frazer.ai/docs/context.md?from=llms-txt): Frazer Context is the platform every Frazer product calls for context: one contract with a use per product, the correctness rules, ranking with rules first and Frazer Rerank second, summaries that cite and never bind, compaction that keeps what binds, verify with configurable depth and strictness, history only at the level a task needs, a static check for old values, receipts that prove what a session knew, and rerank and grounding APIs priced per processed token.
- [Context recipes](https://frazer.ai/docs/context/recipes.md?from=llms-txt): Six ways to use Frazer Context with your own models and tools: use the session context with your own model, keep what binds through your provider's compaction, rank with your own reranker, verify with your own models, run your own compaction, and prove what your agent knew.
- [Agents and plans](https://frazer.ai/docs/for-agents/plans.md?from=llms-txt): How an agent session works from a Fir plan: the session context it receives before its first model call, history only at the level its task needs, what Fir refuses on every write and what the harness steers, the old-value check, a summary for every plan and Find for any plan, receipts, verify, item leases, notes, assertions checked against CI, loops, person tasks for what only a person can do, long sessions with checkpoints and budgets, handoffs and resume.
- [What agents get](https://frazer.ai/docs/for-agents/what-agents-get.md?from=llms-txt): What an AI agent needs to work accurately, from the research on how models act as agents, and what Frazer gives you for each: small authoritative context, tools whose errors name the next step, static checks before any model, state outside the context window, safe probes, visible budgets, a way to ask a person, memory with evidence, verification before done, checks you cannot edit, an honest exit, and the targets Frazer measures.

## Optional

- [Full text of these pages](https://frazer.ai/docs/context/llms-full.txt): 4 pages, 85,239 characters, about 21,310 tokens