A Long-Run Agent Wiki helps people and AI agents continue extended wiki work without dragging every prior message, source file, and historical note into hot context. It keeps the long task visible as Macro Context, keeps the active work small as Micro Context, folds completed work into reviewed summaries, and stages durable syntheses for human review before promotion.
Support boundary
This is local documentation, schema/profile guidance, browser-only wizard packet generation, starter-bundle template material, lint/preflight guidance, discovery routing, and a dogfood workflow. It is not a hosted service, model-training claim, benchmark claim, hardware claim, hidden memory system, public MCP server, write API, autonomous publisher, certification, endorsement, or conformance program.
Core pattern
| Term | Job | Rule |
|---|---|---|
| Context-Folded Agent Run | Extended agent work divided into small reviewed subgoals. | Keep current work explicit and fold completed work into summaries with source paths. |
| Folded Run State | The safe, reviewable packet that survives pauses and handoffs. | Record observable task state only; do not store private chain-of-thought or hidden reasoning. |
| Macro Context | Goal, scope, boundaries, owners, source map, and public/support limits. | Keep stable enough to guide the whole run. |
| Micro Context | The smallest useful page set, source set, checks, and blocker state for the active subgoal. | Refresh at each State Review instead of loading the whole wiki. |
| Active Subgoal | The one currently actionable unit. | Name done criteria, evidence, allowed files, tests, and stop conditions. |
| Completed Subgoal Summary | A compact record of finished work. | Include changed artifacts, source traces, checks, blockers, and follow-up; exclude raw scratch reasoning. |
Folded Run State packet
The packet is a steering surface for humans and agents. It should be small, explicit, and replaceable by the reviewer. It is not durable truth until accepted into reviewed wiki pages, issue state, release notes, or another governed record.
{
"profile": "llmwikis.long_run_agent_wiki_state.v1",
"macro_context": "Goal, scope, owner, authority boundary, and stop conditions.",
"active_subgoal": "The one work item currently being executed.",
"micro_context": {
"pages_to_read": [],
"source_records": [],
"files_to_touch": [],
"checks_to_run": []
},
"completed_subgoal_summaries": [],
"evidence_checked": [],
"state_review": {
"subgoal_complete": false,
"next_action": "",
"blockers": [],
"citations": [],
"stop_condition": ""
}
}State Review
- Confirm the active subgoal. If the request changed, rewrite the active subgoal before opening more files.
- Check evidence scope. Use the index, source map, trust labels, and named pages to choose the smallest useful page set.
- Decide completion. Mark complete only when artifacts, checks, source traces, and done criteria are present.
- Fold completed work. Replace bulky working history with a Completed Subgoal Summary that names outputs, evidence, tests, blockers, and next dependencies.
- Pick the next action. Continue, choose the next subgoal, ask for review, or stop on missing permission, private data, unsupported claims, or ambiguous authority.
Active Wiki Navigation
Retrieval Budget Checks: keep each run index-first and evidence-scoped by choosing the smallest useful page set for the active subgoal, then stop when a page, source trace, permission, or authority boundary is missing.
| Step | What to do | Failure to catch |
|---|---|---|
| Start at the index | Read README, index, source map, route manifest, log, or public llms.txt first. |
Overbroad retrieval that pulls the whole wiki into context. |
| Narrow by job | Open only pages needed for the active subgoal. | Repetitive loops, unrelated source expansion, and slow drift. |
| Inspect trust labels | Check owner, status, source status, sensitivity, last review, and agent-use rules. | Drafts, stale pages, or proposal memory treated as authority. |
| Cite local paths | Name the wiki pages, source records, and discovery files used. | Missing source trace or unsupported success claim. |
| Stop on gaps | Ask or block when permission, evidence, privacy, or authority is unclear. | Fabricated paths, fabricated tools, or unreviewed synthesis promoted as truth. |
Governed Agent Synthesis
Raw agent runs, notes, tool outputs, and drafts are evidence inputs. The review path is: preserve raw run artifacts in approved locations, fold useful results into completed subgoal summaries, stage candidate pages as Agent-Staged Pages, attach source traces and process-quality checks, then promote only after owner review.
- Agent-Staged Page content starts as proposal memory, not canonical memory.
- Every reusable synthesis needs a source trace and explicit review state.
- Private raw data, chat transcripts, secrets, local-only paths, and unrestricted tool output stay out of public pages.
- Reviewers can promote, revise, reject, archive, or split the staged synthesis.
Process-Quality Checks
- Fabricated paths/tools: fail when a route, file, script, command, package, or upload is claimed without evidence.
- Unsupported success claims: fail when publication, upload, route verification, certification, sync, hosted import, or automation is claimed without proof.
- Repetitive loops: fail when the run repeats discovery without a changed active subgoal, new evidence, or a blocker.
- Malformed packets: fail when Folded Run State is missing active subgoal, citations, blockers, or done criteria.
- Overbroad retrieval: fail when the agent scans broad archives or sibling namespaces without an explicit cross-site task.
- Missing source traces: fail when staged pages or summaries lack local page/source paths.
- Private raw data exposure: fail when sensitive raw material appears in public pages or discovery files.
- Unreviewed synthesis as authority: fail when Agent-Staged Pages are promoted without owner review and checks.
Dogfood trail
This guide was created from active LlmWikis improvement intake, but the public copy intentionally uses neutral terms and avoids report-specific labels, external research project names, model names, optimizer names, benchmark names, hardware claims, and performance claims. The accepted lesson is operational: long-running agent wiki work needs visible state, bounded retrieval, source traces, review gates, and process-quality checks.
Copy-ready agent prompt
Use the Long-Run Agent Wiki Strategy. Start from the index and source map. Build a Folded Run State packet with Macro Context, one Active Subgoal, Micro Context, completed subgoal summaries, evidence checked, blockers, checks, and a State Review. Do not include private chain-of-thought. Use Active Wiki Navigation to choose the smallest useful page set. Stage reusable syntheses as Agent-Staged Pages with source traces and review state. Stop on missing permission, private data, unsupported claims, fabricated paths/tools, overbroad retrieval, or unclear authority. End with changed artifacts, citations, checks run, skipped checks with reasons, blockers, and the next active subgoal.