[richard-peng-xia]

Summary

This PR holds the core MIRIX part of skill-bench-public-repro (104d601, by nanxingw). The evals/skill_bench harness (ALFWorld / WebShop, ~143k lines) is left out and stays on skill-bench-public-repro, the frozen reproduction branch its README already points to.

New in this PR:

• Conversation store and session lifecycle:
  • conversation_session tracks each session.
  • task sessions are stored for skill learning and never reach the meta agent.
  • A session ends on an explicit close, on idle, or by the sweeper.
• Skill learning:
  • The threshold trigger queues skill_learning_runs.
  • The experience distiller turns sessions into experiences, and the skill curator turns experiences into skills in token-budgeted batches.
  • Skill changes are versioned (skill_version / skill_object), and leftover experiences feed an experience pool.
• Retrieval:
  • BM25 + cosine fused with RRF, then a similarity gate.
  • Every retrieval writes a retrieval_log row.
  • Tenant guards on retrieve and search.
• Images: a content-addressed image store and remote image fetching with SSRF guards.
• Security fixes against main:
  • A deleted client API key kept working (the delete was never awaited).
  • file_uri parts led to local file reads on the server.
  • The retrieve endpoints did not check the user's organization.
• Migrations: 9 SQL migrations plus backfill and re-embed scripts.

Commits

1. Add session-based procedural skill learning: 104d601 minus evals/skill_bench, byte-identical otherwise, authored by nanxingw.

2. Fix image release scope and organization deletion: fixes found in review.

   • release_message_images deleted any files row a message referenced by id; it now deletes only image rows MIRIX wrote.
   • is_managed_image_path did not require the scoped layout to sit under the images root.
   • delete_organization_by_id failed on Postgres: the new session and learning tables block the users' cascaded delete through their foreign keys. Added tests/test_organization_delete_pg.py.

3. Decouple core from the skill_bench harness:

   • Drops the skill-bench extra, the harness test path and its .gitignore entries.
   • The PG e2e test checks the harness's OpenAPI contract only when the harness is in the checkout.

4. Fit the e2e curator batch in its context window: the e2e test's window goes from 20k to 21k. The long topic's edit step now fits, and the long session still chunks and the curator still batches.

Tests

• pytest (default config): 2970 passed, 52 skipped, 0 failed.
• PG integration via scripts/run_pg_integration_tests.sh: all pass (after commit 4 below; before it, tests/test_skill_learning_e2e_pg.py::test_task_sessions_become_retrievable_skills failed).
  • That test also fails on the original 104d601, including in a Python 3.11 env pinned to requirements-lock.txt.
  • Cause: one curator batch peaks at ~14.5k context tokens, above the 20,000 × 0.8 − 1,600 = 14,400 limit. The batch ends context_limit and is returned instead of consumed.
  • Most likely the curator prompt grew after the test was calibrated.

To do before merge

• Fix the e2e test above: commit 4 raises its CONTEXT_WINDOW to 21k. The test passes 4/4 in a py3.12 env and in a py3.11 env pinned to requirements-lock.txt.
• Add an upgrade runbook covering:
  • the order of the 9 migrations (today it exists only in _MIGRATION_HINTS);
  • the downtime migrate_skill_retrieval_fields.sql needs (it nulls every skill vector under an ACCESS EXCLUSIVE lock and is not safe for a rolling deploy);
  • then backfill_skill_baseline_versions.py and reembed_skill_retrieval.py.
• Decide the retrieval gate (0.62 / 0.90) for embedding models other than gemini-embedding-001. Today they get those values with only a warning, which can return zero skills.
• Release notes for changed defaults:
  • RETRIEVAL_LOG_ENABLED is on and stores raw queries for 30 days.
  • Curator max steps go from 15 to 100, with no per-run cost cap.
  • A session ends only on close or after 48 h idle.
  • Non-Gemini embeddings are chunked at 2,048 tokens.
  • The first auto_dream after the upgrade may distill the whole backlog.
  • Removed env vars (SKILL_EDIT_BUDGET_*, SKILL_EVOLVE_MAX_CHAINING_STEPS, …) are now ignored silently.
• Release notes for API changes:
  • /memory/add with session_tag=task requires a session_id and skips meta extraction.
  • /memory/retrieve/conversation always makes a topic-extraction LLM call and returns 422 when there is no text or image.
  • DELETE /memory/procedural/{id} is scoped to the user.
  • The skill_list and skill_delete tool signatures changed.

Follow-ups (not blocking):

• Inline trigger locking on every task write.
• LIKE '%id%' scans in image release.
• The startup anti-join over conversation_message.
• The size of procedural_memory_agent.txt (7 KB to 19.5 KB) and its benchmark-flavoured rules.

After this merges, rebase skill-bench-public-repro onto main so that it carries only the harness.

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ richard-peng-xia
❌ nanxingw
_{You have signed the CLA already but the status is still pending? Let us recheck it.}