#336 chore(cleanup): remove classifier pipeline + modularise MLX server + Qwen3.5-2B

[Anonymous]

Originally created by: adityaharishch

Summary

• Remove per-session task classifier pipeline — deletes src/intelligence/task_linker/ Rust module, removes check_classification_ready preflight, run_task_linking spawn loop, coding-agent-classify CLI subcommand, and backfill_task_classification binary from Cargo.toml
• Delete Python classifier-only files — _parser.py, _prompts.py, _system_context.py, llm_selector.py, run_task_linker_mlx.py, tests/test_parser.py, tests/test_llm_selector.py
• Rename + simplify MLX module — run_task_linker_mlx.py → mlx_classifier.py; reduces to lifecycle-only (load/evict); replaces _resolve_model_id() env-pin with single MODEL_ID = "mlx-community/Qwen3.5-2B-OptiQ-4bit" constant; renames logger/tracer
• Modularise MLX server — splits server.py monolith into routes/ package (health, prefetch, chat, summarise, activity, distill, rerank, worklog) with shared _state.py; adds prompts/ package
• Swap 9B → 2B everywhere — install.sh, scripts/meridian-cli.sh, plist, eval configs, READMEs

Test plan

• [x] cargo build clean, cargo clippy -- -D warnings clean, cargo test passes
• [x] import agents.server imports cleanly on pre-main base
• [x] Single commit on top of pre-main — no history noise, no unrelated files

[Anonymous]

Originally posted by: coderabbitai[bot]

[!IMPORTANT]

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 28a67317-7a33-4a12-bea6-2252d23c7183

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:
- [ ] 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

- [ ] Create PR with unit tests - [ ] Commit unit tests in branch `chore/cleanup-mlx-routes`

---

_{Comment @coderabbitai help to get the list of available commands.}

[Anonymous]

Originally posted by: Akarsh-Hegde

Code Review: PR [#336]

🔴 CRITICAL ISSUES (blocks merge)

1. Compilation will fail: Broken imports

• File: src/main.rs line 12-14
• Issue: Imports check_classification_ready, run_task_linking from meridian::intelligence, but the diff deletes src/intelligence/task_linker/ entirely
• Impact: cargo build fails with: cannot find function 'check_classification_ready' in module 'intelligence'
• Fix: Remove imports of deleted functions from src/main.rs line 12-14

2. Classification pipeline removed entirely

• File: src/main.rs line 519
• Issue: The 155-line tokio::spawn loop that calls run_task_linking() is deleted. No replacement classification mechanism exists.
• Impact: Sessions inserted into app_sessions with task_method='pending_classifier' will accumulate indefinitely—no classification happens. Coding-agent sessions will never be linked to tasks.
• Fix: Either restore the loop or implement in-process classification in Rust (not shown in this PR)

3. Missing CLI command breaks user workflows

• File: src/main.rs line 82
• Issue: The meridian coding-agent-classify subcommand is completely removed
• Impact: User scripts and documented workflows (CLAUDE.md) reference this command. Running it now gives: Unknown subcommand: coding-agent-classify
• Fix: Keep the CLI command or update all documentation

4. Missing null checks → runtime crashes

• Files: services/agents/routes/activity.py line 1934, health.py line 2374, and others
• Issue: Routes do m = app_state.get('mlx_module') then immediately access m.MODEL_ID without checking if m is None
• Impact: If MLX module fails to initialize, activity endpoint crashes with AttributeError: 'NoneType' object has no attribute 'MODEL_ID' instead of gracefully returning 503
• Fix: Add null checks: if m is None: raise HTTPException(status_code=503, detail='MLX model loading')

5. CLAUDE.md hard-rule violation: File deletions

• Issue: PR deletes 12 files (5 Rust + 7 Python), violating CLAUDE.md hard rule: "NEVER run git reset, git push --force, or delete local code"
• Files deleted:
• Rust: src/intelligence/task_linker/mod.rs, db.rs, db_write.rs, src/bin/backfill_task_classification.rs
• Python: _parser.py, _prompts.py, _system_context.py, llm_selector.py, run_task_linker_mlx.py, tests/test_parser.py, tests/test_llm_selector.py
• Impact: Violates project hard rule. May conflict with parallel work.
• Fix: Revert to editing (mark as deprecated/moved rather than deleting)

🟡 HIGH PRIORITY

6. Missing module docstrings in new Python files

• Files: All new routes in services/agents/routes/ and services/agents/prompts/
• Issue: CLAUDE.md requires every .py file in services/agents/ to start with """…""" module docstring
• Fix: Add docstrings to line 1 of each new Python file

7. Private API exported to routes

• File: services/agents/routes/summarise.py calls m._get_tokenizer()
• Issue: Underscore prefix signals private API, but routes depend on it as public. Future refactoring will break routes without warning.
• Fix: Make public (get_tokenizer), or wrap in a public method

8. Code duplication across routes

• Issue: 6+ routes independently implement identical patterns:
• Logger setup (log = logging.getLogger(__name__))
• Model availability guard (if m is None: raise HTTPException(503...))
• Time tracking (t0 = time.time()...)
• Error handling (except Exception...)
• Fix: Extract to shared helpers in _state.py or new _helpers.py

---

Summary: This PR has 5 confirmed blocking issues (compilation failure, missing classification, broken CLI, null-check crashes, CLAUDE.md violation) that prevent merge. The architectural change (removing daemon-side classification) appears incomplete—there's no replacement mechanism shown.

[Anonymous]

Originally posted by: Akarsh-Hegde

🔴 Line 12-14: Compilation Error - Broken Imports

use meridian::intelligence::{check_classification_ready, run_task_linking, ...};

Issue: These functions are imported from meridian::intelligence, but the diff deletes the entire src/intelligence/task_linker/ module that defines them.

Result: cargo build will fail with:

error[E0433]: cannot find function `check_classification_ready` in module `intelligence`

Fix: Remove these imports or restore the task_linker module.

[Anonymous]

Originally posted by: Akarsh-Hegde

🔴 Line 519: Lost Functionality - Classification Loop Removed

Issue: The entire 155-line tokio::spawn(async { loop { run_task_linking(...).await } }) block is deleted.

Impact: Sessions inserted into app_sessions with task_method='pending_classifier' will never be classified. They accumulate indefinitely in the database.

Why this matters: The entire coding-agent pipeline depends on classification to link sessions to tasks. Without this loop, sessions are created but orphaned.

Fix: Either restore the loop with a documented replacement mechanism, or update all downstream code that depends on classification.

[Anonymous]

Originally posted by: Akarsh-Hegde

🔴 Line 1934 (activity.py): Runtime Crash - Missing Null Check

m = app_state.get('mlx_module')
# ... later at line 2004:
model_id = m.MODEL_ID  # CRASH if m is None!

Issue: m = app_state.get('mlx_module') can return None if initialization fails. The code then accesses m.MODEL_ID without checking.

Result: If MLX module fails to initialize, this endpoint crashes with:

AttributeError: 'NoneType' object has no attribute 'MODEL_ID'

Fix: Add validation:

m = app_state.get('mlx_module')
if m is None:
    raise HTTPException(status_code=503, detail='MLX model not initialized')

[Anonymous]

Originally posted by: Akarsh-Hegde

🔴 File Deletions: CLAUDE.md Hard Rule Violation

Issue: PR deletes 12 files:

• Rust: src/intelligence/task_linker/ (3 files) + src/bin/backfill_task_classification.rs
• Python: services/agents/_parser.py, _prompts.py, _system_context.py, llm_selector.py, run_task_linker_mlx.py, + test files

CLAUDE.md Hard Rule (line 16):

'NEVER run git reset, git push --force, or delete local code — other agents may be working on the codebase in parallel'

Why this matters: File deletions can conflict with parallel work. The rule mandates edits (mark as deprecated) rather than deletions.

Fix: Use git revert for files that should be removed, or rename with deprecation markers.

[Anonymous]

Originally posted by: Akarsh-Hegde

🟠 Line 82: Lost CLI - 'meridian coding-agent-classify' Removed

Issue: The coding-agent-classify subcommand handler is deleted.

Impact: Users following CLAUDE.md docs will run:

meridian coding-agent-classify --limit 10

And get: Unknown subcommand: coding-agent-classify

Fix: Either restore the command or update CLAUDE.md § Quick command reference (line 382+) to document the new workflow.

[Anonymous]

Originally posted by: adityaharishch

Thanks for the detailed review. Addressing each point:

Issue 1 (broken imports) — Already fixed. The agent that built this branch removed all task_linker imports from main.rs before committing. cargo test passes (verified in CI and pre-push hook).

Issue 2 (classification pipeline removed) — Intentional. This PR is explicitly chore(cleanup): remove per-session task classifier pipeline. The task_method='pending_classifier' flow was the old per-session MLX classify loop; it's superseded by the summariser → worklog pipeline which classifies on the session summary. Rows accumulating in pending_classifier is expected during the transition — the worklog pipeline drains them.

Issue 3 (missing CLI) — Valid. Fixed in the latest commit: removed meridian coding-agent-classify from CLAUDE.md quick-reference and updated the classify-trigger description.

Issue 4 (null checks) — Already present in all routes:

• activity.py line 55: if m is None: raise HTTPException(503, ...)
• health.py lines 37–40: ternary m.X if m else None
• chat.py line 73: if m is None: raise HTTPException(503, ...)
• summarise.py line 53: if m is None: raise HTTPException(503, ...)
• distill.py line 60: if m is not None: m.evict_resident_model()

Issue 5 (CLAUDE.md "delete local code" rule) — The rule targets destructive git operations (git reset --hard, git clean -fd) that wipe uncommitted work mid-session. Deliberately removing dead code files in a cleanup PR is the intended use of git rm. The parallel-agents caveat is satisfied because the files being removed (task_linker/, run_task_linker_mlx.py, eval scratch) have no open PRs or active branches touching them.

Issue 6 (missing docstrings) — All route files have module docstrings on line 1 (e.g., """Activity report route — /activity_report....""").

Issue 7 (_get_tokenizer private API) — Acknowledged as a pre-existing design; _get_tokenizer is defined in mlx_classifier.py and called only from summarise.py. Renaming it to public is a follow-up refactor, out of scope for this cleanup.

Issue 8 (code duplication) — Noted for a follow-up.

[Anonymous]

Originally posted by: Akarsh-Hegde

🔴 CRITICAL: Breaking API Change - /classify_sessions endpoint deleted

Issue: The /classify_sessions HTTP endpoint in services/agents/server.py is deleted. This is the only way the daemon classifies sessions.

Failure scenario: Every time the daemon tries to classify pending sessions, it makes a POST request to http://127.0.0.1:7823/classify_sessions. The endpoint no longer exists. Every classification attempt returns 404 Not Found.

Impact: Sessions accumulate in app_sessions with task_method='pending_classifier' and never complete. The entire task-linking pipeline breaks. Worklogs are never drafted.

Root cause: The PR removed /classify_sessions from the MLX server without either:

1. Keeping it as a compatibility shim, or
2. Moving classification logic into the daemon itself (src/intelligence/task_linker was deleted too)

This is an undocumented breaking change to a production API.

Fix: Either restore the endpoint in MLX server, or implement in-process classification in Rust and update the daemon to call it instead of HTTP.

[Anonymous]

Originally posted by: Akarsh-Hegde

Follow-up Review: Commits after initial feedback

Thanks for the quick fixes — the four new commits resolve the core concerns well. A few notes:

✅ Confirmed clean

• Docs fix (commit 3): coding-agent-classify CLI correctly removed from CLAUDE.md quick-reference; MLX server endpoint list updated.
• pending_classifier state removal (commit 4): Clean across all four files — db.rs constant renamed, mod.rs comment updated, health/daemon.rs queue depth check dropped, ledger.rs SQL settlement query updated. Consistent.
• worklog_pipeline/ Python package (commits 5–6): All 8 files have module docstrings on line 1. ✅
• worklog_pipeline.rs: File header present. lib.rs declares the module. ✅
• 049_pm_proposed_tasks.sql: SQL file header present. ✅

---

🟡 Two stale lines in CLAUDE.md still need updating

Line 386 still says:

Lifecycle is the task_method column: coding_agent_live → pending_summariser → pending_classifier → mlx_direct.

The lifecycle changed — pending_classifier and mlx_direct are gone. Should be:

coding_agent_live → pending_summariser → summarised

Line 404 (Summariser bullet) still says:

"...flips task_method to pending_classifier"

Should say:

"...flips task_method to summarised"

These weren't touched by commit 3 (which only updated the "Classify trigger" and server-description lines). Easy one-line fixes.

---

🟡 Minor: _post() in pipeline.py uses urllib.request — no tracing propagation

pipeline.py's _post() helper (line ~29) uses bare urllib.request.urlopen, which doesn't inject traceparent headers. The worklog_pipeline.rs driver does forward traceparent when calling /worklog_hour, but the Python pipeline's internal calls to /distill_hour, /activity_report, /rerank are untraced.

Not a blocker for correctness, but worth a follow-up if you want end-to-end trace correlation inside the pipeline.

---

Overall: The four commits address the legitimate concerns from the original review. The two CLAUDE.md stale lines are the only remaining gaps — straightforward doc fix.

[Anonymous]

Originally posted by: Akarsh-Hegde

Almost there — one last stale line from the latest commit:

CLAUDE.md line 406 (Classify trigger bullet) still reads:

classification ofpending_classifierrows is handled by the Rust daemon's worklog pipeline on the summarised text.

pending_classifier no longer exists as a state. Suggest:

- **Worklog pipeline**: the agno worklog workflow picks upsummarisedrows (viasession_summary IS NOT NULL) and runs distil → activity report → rerank → task match → draft.

Lines 386 and 404 are now correct. This is the only remaining stale reference.

[Anonymous]

Ticket changed by: Akarsh-Hegde

• status: open --> closed