Anthropic native context-management requirements
These requirements track the live implementation of the Anthropic native compaction and context-editing specification.
| ID | Requirement | Implementation | Verification | Documentation | Evidence |
|---|---|---|---|---|---|
| ANCM-001 | AnthropicCompactionExtension and AnthropicContextEditingExtension independently own their feature config, UI, gating, request shaping, response interpretation, tests, and package documentation. |
Implemented | Automated | Documented | extension modules; test_context_management_unit.py; package context-management guide |
| ANCM-002 | Official selected-model descriptors are authoritative, explicit boolean capabilities can opt compatible endpoints in, and absent support emits no optional feature tag or UI. | Implemented | Automated | Documented | context-management tag tests; official model-catalog test; Ollama omission test |
| ANCM-003 | Enabled compaction appends compact_20260112 last, uses the configured trigger/instructions, requests pause_after_compaction: false, and composes compact-2026-01-12 through provider transport. |
Implemented | Automated | Documented | ordered request/beta tests; Core stack request assertions; hosted compaction test |
| ANCM-004 | Thinking retention and tool-use clearing map to ordered context-management edits and compose context-management-2025-06-27; raw request options remain the final complete override. |
Implemented | Automated | Documented | context-management unit tests; hosted interrupted-user bank profiles; package guide |
| ANCM-005 | Ordinary and streaming responses preserve complete compaction blocks, including compaction_delta, null summaries, opaque encrypted continuation data, and unknown additive fields. |
Implemented | Automated | Documented | reducer tests; deterministic stack replay; hosted null-block compaction continuation |
| ANCM-006 | Compaction blocks survive persistence/reload and are reused by a successful follow-up request without making the visible transcript canonical. | Implemented | Automated | Documented | deterministic exact-request replay and hosted streaming persistence/follow-up tests |
| ANCM-007 | Context-editing applied statistics are available for ordinary and streaming responses without deleting or rewriting local history. | Implemented | Automated | Documented | applied-edit metadata/reducer tests and latest-turn hosted bank profile |
| ANCM-008 | usage.iterations contributes complete normalized message/turn/session usage exactly once and exposes separate compaction token classes while preserving duration/TPS interoperability. |
Implemented | Automated | Documented | iteration double-count test; cross-feature stack test; hosted compaction usage assertion |
| ANCM-009 | Ordinary request cancellation or malformed completion cannot persist a partial compaction block, and public metadata does not expose encrypted continuation content. | Implemented | Automated | Documented | incomplete-stream and encrypted-metadata tests; existing hosted request cancellation |
| ANCM-010 | Deterministic, Core stack, local Ollama omission/regression, and selective hosted Anthropic tests prove the supported native context-management contract. | Implemented | Automated | Documented | 74-test fast suite; 12-test Ollama suite; hosted model, compaction, bank, and default-config tests |
| ANCM-011 | Managed pause_after_compaction: true automatic continuation is not exposed or claimed by the initial implementation. |
Implemented | Automated | Documented | compaction schema/request tests; package guide limitations |
| ANCM-012 | The shipped direct Anthropic configuration enables capability-gated compaction, keep-all thinking retention, and conservative tool-use clearing, while session UI exposes only the six primary controls and keeps expert policy fields configuration-only. | Implemented | Automated | Documented | default config variants; test_context_management_unit.py; default-config effective/UI tests; package context-management guide |