feat(compact): reactive compaction with direct input support

[crstrn13]

Summary

Completes the remaining work items from #30:

• Use previous_usage for token estimation — should_compact() checks the stored usage.total_tokens from the rehydrated response before falling back to local tiktoken counting
• Configurable summary prefix — summary_prefix config option (default: [Previous conversation summary]) carried through to compaction items and both translation paths
• Hide compaction items from input_items API — normalize_input_items() filters out {"type": "compaction"} items so clients never see internal state
• Persist compaction as a stored response — after summarization, the compacted response is written back to the response store with the compaction item in its input
• Explicit compact endpoint — POST /v1/responses/{id}/compact triggers compaction on a previously stored response

Behavior details

• Invalid config is rejected up front — an invalid compact_threshold (e.g. below MIN_COMPACT_THRESHOLD) returns 400 invalid_request_error before the compactability gate, for both streaming and non-streaming requests, and for direct-input requests
• Summaries cover the full conversation — should_compact() summarizes the entire rehydrated history including the current turn, so the summarizer has complete context. The current turn is not lost or duplicated: replace_messages() preserves it verbatim after the compaction item via a tail split
• Direct input passes through — requests with inline input (no rehydration) are not reactively compacted

Test plan

• cargo test -p praxis-ai-apis — unit tests for all work items
• cargo test -p praxis-ai-filters — no filter regressions
• cargo test -p praxis-tests-integration — example integration tests (mock backends), incl. compact.rs
• vLLM functional SDK tests (test_openai_responses_vllm.py) — invalid-threshold rejection and over-threshold rehydrated-history compaction, against live vLLM + OGX
• make lint, rustdoc, unicode-safety, DCO/signoff all green

[praxis-bot]

PR Review

Summary: Completes five compaction features from #30: previous_usage token estimation fast-path, configurable summary_prefix, hiding compaction items from input_items, persisting compaction responses, and an explicit POST /v1/responses/compact endpoint.

Overall: Solid implementation with good integration tests covering both the rehydrated and direct-input paths. The summary prefix plumbing is clean and consistent across both translation paths. A few issues below around blocking in async context, stale docstrings, and missing unit test coverage for the new endpoint.

Severity	Count
Large	3
Medium	3

Findings without inline placement

[Large] Stale docstrings — The struct-level doc comment on CompactFilter (~line 96 of mod.rs) says compaction "only applies to multi-turn requests where openai_responses_rehydrate has loaded stored conversation history". The generated docs page (docs/filters/openai_responses_compact.md, Configuration Notes section) says the same. Both are now incorrect — this PR adds a direct-input compaction path that runs without rehydration. Update both to match the module-level doc comment (lines 14-20), which was correctly updated.

[Large] Missing unit tests for explicit compact endpoint — The new pure/mostly-pure functions parse_compact_request_body, extract_stored_messages, is_explicit_compact_request, and ensure_compactable_state have no unit tests. Per project convention each testable function should have coverage. Suggested cases:

• parse_compact_request_body: empty body, invalid JSON, missing response_id, valid request with optional fields
• extract_stored_messages: empty array, non-array messages, valid array
• is_explicit_compact_request: POST to correct path, GET to correct path, POST to wrong path
• ensure_compactable_state: no ResponsesState, rehydrated state, non-rehydrated with compaction config, non-rehydrated without compaction config

[praxis-bot]

Re-review (Aug 18 commits)

Previous review had 6 findings. The new commits addressed 4 fully (async conversion of fetch_response/build_and_persist_compaction, _body rename, response_id emptiness validation, tiktoken overhead restoration). One partially addressed (unit tests added for parse_compact_request_body and extract_stored_messages, but is_explicit_compact_request and ensure_compactable_state still lack coverage). One unaddressed (stale docstrings).

One new finding below from the rewritten explicit compact endpoint.

Severity	Count
Medium	1

[aslakknutsen]

Outside the PR diff, but

ai/apis/src/openai/responses/store/filter.rs

Line 464 in bc1e5bd

fn should_init_store_for_request(ctx: &HttpFilterContext<'_>) -> bool {

"ResponseStoreFilter::should_init_store_for_request() returns false for POST /v1/responses/compact because the path is not is_responses_create and the classifier does not set has_previous_response_id for sub-resource routes. The store Arc is therefore never registered into ResponseStoreRegistry before CompactFilter::resolve_store_and_tenant() runs. On a freshly built pipeline (or after reload), the first explicit compact call fails with "response store not available" even when the backend is healthy. Extend should_init_store_for_request (or equivalent) to initialize the store for POST /v1/responses/compact, mirroring how rehydrate paths are handled."

[praxis-bot]

PR Review (re-review)

The Aug 20 commits fixed the response_object finding from the previous re-review (now includes model, created_at, and output).

One new finding from the direct-input compaction path.

Severity	Count
Medium	1

[praxis-bot]

Re-review (Aug 24 commits)

The two new commits fix the summarization scope (exclude current input from summarization text in the rehydrated path), add is_compactable unit tests, add an explicit compact endpoint integration test, fix the stale docstrings on CompactFilter and the docs page, and add store initialization for the compact endpoint.

Previous findings addressed:

• Stale docstrings: fixed (struct doc and docs page updated)
• Missing is_compactable / ensure_compactable_state tests: fixed (extracted is_compactable and added 5 tests)

One new finding below.

Severity	Count
Medium	1

[leseb]

@crstrn13 please rebase

[praxis-bot-app]

Missing Signed-off-by: e849dfe. All commits require sign-off (via git commit --signoff).

[leseb]

last one on_failure: open returns raw input messages inside CompactResource.output. Those lack required output fields such as type, id, and status, so typed OpenAI clients may reject the response.

[leseb]

final:

Persisted response.compaction objects have no status, but rehydration requires status == "completed". Referencing the returned ID therefore fails as status 'unknown', despite the new code claiming follow-up continuity.

Accept response.compaction records during rehydration and add an actual follow-up test.

[leseb]

Thanks for your patience!