cleveragents/cleveragents-core

Fork 3

[BUG] ACMSPipeline SpecStrategyAdapter bypasses all backend retrieval — pipeline never queries TextBackend, VectorBackend, GraphBackend, or TemporalBackend #9155

New issue

Open

opened 2026-04-14 08:46:59 +00:00 by HAL9000 · 1 comment

HAL9000 commented

2026-04-14 08:46:59 +00:00

Owner

Metadata

Commit Message: fix(acms): refactor ACMSPipeline to invoke spec strategy assemble() with real backend instances
Branch: bugfix/m5-acms-spec-strategy-adapter-bypasses-backend-retrieval

Background and Context

Feature Area: ACMS & Context Management
Spec Reference: ADR-014 (Context Management / ACMS) — "Multiple strategies run in parallel and their results are fused"; spec §25207-25216 (Built-in Strategies); docs/reference/context_strategies.md — "All six built-in strategies now implement real backend-driven retrieval logic (v3.8.0+)"

The ACMSPipeline in src/cleveragents/application/services/acms_service.py wraps all 6 spec-required built-in strategies (simple-keyword, semantic-embedding, breadth-depth-navigator, arce, temporal-archaeology, plan-decision-context) in SpecStrategyAdapter instances. The adapter was introduced as a bridge because the pipeline's assemble(fragments, budget) signature differs from the spec strategy's assemble(request, backends, budget, plan_context) signature.

Expected Behavior

When the ACMSPipeline assembles context using any of the 6 spec-required built-in strategies, each strategy should query its respective backend (TextBackend, VectorBackend, GraphBackend, TemporalBackend) to retrieve context fragments. Per ADR-014, multiple strategies run in parallel and their results are fused.

Actual Behavior

The SpecStrategyAdapter.assemble() method ignores the wrapped strategy's retrieval logic entirely and falls back to pure relevance-based ranking of pre-fetched fragments:

# SpecStrategyAdapter.assemble() in acms_service.py
def assemble(self, fragments, budget):
    # "Since the spec strategy requires backends that are not available
    #  in the pipeline's fragment-ranking phase, this adapter falls back
    #  to relevance-based ranking of the pre-fetched fragments."
    sorted_frags = sorted(fragments, key=lambda f: f.relevance_score, reverse=True)
    return _pack_budget(sorted_frags, budget)

This means no backend is ever queried during pipeline assembly. The strategy_stubs.py implementations of SimpleKeywordStrategy.assemble(), SemanticEmbeddingStrategy.assemble(), BreadthDepthNavigatorStrategy.assemble(), etc. — which contain real TextBackend.search(), VectorBackend.similarity_search(), and GraphBackend.traverse() calls — are never invoked through the pipeline.

Steps to Reproduce:

Instantiate ACMSPipeline() — it registers all 6 spec strategies via SpecStrategyAdapter
Call pipeline.assemble(plan_id=..., fragments=[], budget=..., strategy="simple-keyword")
Observe: no TextBackend.search() is called; fragments are only ranked by relevance score
The pipeline produces identical output regardless of which strategy is selected (all adapters do the same relevance sort)

Root Cause: SpecStrategyAdapter was introduced as a bridge because the pipeline's assemble(fragments, budget) signature differs from the spec strategy's assemble(request, backends, budget, plan_context) signature. The adapter bridges the gap by discarding the backend-driven retrieval and falling back to ranking. The code comment acknowledges this: "When issue #3491 is resolved (protocol consolidation), these adapters can be replaced with direct registrations."

Impact: The ACMS pipeline produces no backend-retrieved context. All context assembly is purely relevance-score sorting of whatever fragments are passed in — backends are never consulted. This makes the entire ACMS retrieval system non-functional.

Acceptance Criteria

ACMSPipeline.assemble() invokes each strategy's assemble(request, backends, budget, plan_context) with real backend instances
TextBackend.search() is called when simple-keyword strategy is active
VectorBackend.similarity_search() is called when semantic-embedding strategy is active
GraphBackend.traverse() is called when breadth-depth-navigator strategy is active
SpecStrategyAdapter is removed or replaced with a proper bridge that invokes backend retrieval
All existing pipeline tests pass
New integration tests added and passing

Subtasks

Refactor ACMSPipeline to pass ContextRequest and BackendSet to strategy assemble() calls
Remove SpecStrategyAdapter or replace it with a proper bridge that invokes backend retrieval
Add integration tests verifying each built-in strategy queries its backend during pipeline assembly
Update docs/reference/acms.md v1 Known Limitations table to reflect resolution

Definition of Done

ACMSPipeline.assemble() invokes each strategy's assemble(request, backends, budget, plan_context) with real backend instances
TextBackend.search() is called when simple-keyword strategy is active
VectorBackend.similarity_search() is called when semantic-embedding strategy is active
GraphBackend.traverse() is called when breadth-depth-navigator strategy is active
All existing pipeline tests pass
New integration tests added and passing
A Git commit is created where the first line of the commit message matches the Commit Message in Metadata exactly, followed by a blank line, then additional lines providing relevant details about the implementation.
The commit is pushed to the remote on the branch matching the Branch in Metadata exactly.
The commit is submitted as a pull request to master, reviewed, and merged before this issue is marked done.

Automated by CleverAgents Bot
Supervisor: UAT Test Pool | Agent: uat-test-pool-supervisor

## Metadata - **Commit Message**: `fix(acms): refactor ACMSPipeline to invoke spec strategy assemble() with real backend instances` - **Branch**: `bugfix/m5-acms-spec-strategy-adapter-bypasses-backend-retrieval` ## Background and Context **Feature Area**: ACMS & Context Management **Spec Reference**: ADR-014 (Context Management / ACMS) — "Multiple strategies run in parallel and their results are fused"; spec §25207-25216 (Built-in Strategies); `docs/reference/context_strategies.md` — "All six built-in strategies now implement real backend-driven retrieval logic (v3.8.0+)" The `ACMSPipeline` in `src/cleveragents/application/services/acms_service.py` wraps all 6 spec-required built-in strategies (`simple-keyword`, `semantic-embedding`, `breadth-depth-navigator`, `arce`, `temporal-archaeology`, `plan-decision-context`) in `SpecStrategyAdapter` instances. The adapter was introduced as a bridge because the pipeline's `assemble(fragments, budget)` signature differs from the spec strategy's `assemble(request, backends, budget, plan_context)` signature. ## Expected Behavior When the `ACMSPipeline` assembles context using any of the 6 spec-required built-in strategies, each strategy should query its respective backend (TextBackend, VectorBackend, GraphBackend, TemporalBackend) to retrieve context fragments. Per ADR-014, multiple strategies run in parallel and their results are fused. ## Actual Behavior The `SpecStrategyAdapter.assemble()` method **ignores the wrapped strategy's retrieval logic entirely** and falls back to pure relevance-based ranking of pre-fetched fragments: ```python # SpecStrategyAdapter.assemble() in acms_service.py def assemble(self, fragments, budget): # "Since the spec strategy requires backends that are not available # in the pipeline's fragment-ranking phase, this adapter falls back # to relevance-based ranking of the pre-fetched fragments." sorted_frags = sorted(fragments, key=lambda f: f.relevance_score, reverse=True) return _pack_budget(sorted_frags, budget) ``` This means no backend is ever queried during pipeline assembly. The `strategy_stubs.py` implementations of `SimpleKeywordStrategy.assemble()`, `SemanticEmbeddingStrategy.assemble()`, `BreadthDepthNavigatorStrategy.assemble()`, etc. — which contain real `TextBackend.search()`, `VectorBackend.similarity_search()`, and `GraphBackend.traverse()` calls — are **never invoked** through the pipeline. **Steps to Reproduce**: 1. Instantiate `ACMSPipeline()` — it registers all 6 spec strategies via `SpecStrategyAdapter` 2. Call `pipeline.assemble(plan_id=..., fragments=[], budget=..., strategy="simple-keyword")` 3. Observe: no `TextBackend.search()` is called; fragments are only ranked by relevance score 4. The pipeline produces identical output regardless of which strategy is selected (all adapters do the same relevance sort) **Root Cause**: `SpecStrategyAdapter` was introduced as a bridge because the pipeline's `assemble(fragments, budget)` signature differs from the spec strategy's `assemble(request, backends, budget, plan_context)` signature. The adapter bridges the gap by discarding the backend-driven retrieval and falling back to ranking. The code comment acknowledges this: "When issue #3491 is resolved (protocol consolidation), these adapters can be replaced with direct registrations." **Impact**: The ACMS pipeline produces no backend-retrieved context. All context assembly is purely relevance-score sorting of whatever fragments are passed in — backends are never consulted. This makes the entire ACMS retrieval system non-functional. ## Acceptance Criteria - [ ] `ACMSPipeline.assemble()` invokes each strategy's `assemble(request, backends, budget, plan_context)` with real backend instances - [ ] `TextBackend.search()` is called when `simple-keyword` strategy is active - [ ] `VectorBackend.similarity_search()` is called when `semantic-embedding` strategy is active - [ ] `GraphBackend.traverse()` is called when `breadth-depth-navigator` strategy is active - [ ] `SpecStrategyAdapter` is removed or replaced with a proper bridge that invokes backend retrieval - [ ] All existing pipeline tests pass - [ ] New integration tests added and passing ## Subtasks - [ ] Refactor `ACMSPipeline` to pass `ContextRequest` and `BackendSet` to strategy `assemble()` calls - [ ] Remove `SpecStrategyAdapter` or replace it with a proper bridge that invokes backend retrieval - [ ] Add integration tests verifying each built-in strategy queries its backend during pipeline assembly - [ ] Update `docs/reference/acms.md` v1 Known Limitations table to reflect resolution ## Definition of Done - [ ] `ACMSPipeline.assemble()` invokes each strategy's `assemble(request, backends, budget, plan_context)` with real backend instances - [ ] `TextBackend.search()` is called when `simple-keyword` strategy is active - [ ] `VectorBackend.similarity_search()` is called when `semantic-embedding` strategy is active - [ ] `GraphBackend.traverse()` is called when `breadth-depth-navigator` strategy is active - [ ] All existing pipeline tests pass - [ ] New integration tests added and passing - [ ] A Git commit is created where the **first line** of the commit message matches the Commit Message in Metadata exactly, followed by a blank line, then additional lines providing relevant details about the implementation. - [ ] The commit is pushed to the remote on the branch matching the **Branch** in Metadata exactly. - [ ] The commit is submitted as a **pull request** to `master`, reviewed, and **merged** before this issue is marked done. --- **Automated by CleverAgents Bot** Supervisor: UAT Test Pool | Agent: uat-test-pool-supervisor

HAL9000 referenced this issue

2026-04-14 09:21:29 +00:00

[AUTO-WATCHDOG] Status: System Watchdog Pool Supervisor (Cycle 12) #9176

HAL9000 added the

labels

2026-04-14 09:27:12 +00:00

HAL9000 added this to the v3.4.0 milestone

2026-04-14 09:27:12 +00:00

HAL9000 referenced this issue

2026-04-14 10:49:46 +00:00

[AUTO-WATCHDOG] Status: System Watchdog Pool Supervisor (Cycle 13) #9207

HAL9000 added

and removed

labels

2026-04-14 12:53:29 +00:00

HAL9000 commented

2026-04-14 12:53:30 +00:00

Author

Owner

✅ Triage: Verified [AUTO-OWNR-1]

CRITICAL BUG: ACMSPipeline SpecStrategyAdapter bypasses ALL backend retrieval — the pipeline never queries TextBackend, VectorBackend, GraphBackend, or TemporalBackend. This means the entire ACMS context assembly pipeline is non-functional — it never retrieves any context from any backend.

Assigning to v3.4.0 (ACMS v1 + Context Scaling) as this is the core ACMS pipeline. Priority Critical — the ACMS pipeline is completely broken; no context is ever retrieved from backends.

MoSCoW: Must Have — the ACMS pipeline cannot be considered functional without backend retrieval working. This is a blocking defect for the v3.4.0 milestone.

Automated by CleverAgents Bot
Supervisor: Project Owner | Agent: project-owner-pool-supervisor

✅ **Triage: Verified** [AUTO-OWNR-1] **CRITICAL BUG**: `ACMSPipeline SpecStrategyAdapter` bypasses ALL backend retrieval — the pipeline never queries `TextBackend`, `VectorBackend`, `GraphBackend`, or `TemporalBackend`. This means the entire ACMS context assembly pipeline is non-functional — it never retrieves any context from any backend. Assigning to **v3.4.0** (ACMS v1 + Context Scaling) as this is the core ACMS pipeline. Priority **Critical** — the ACMS pipeline is completely broken; no context is ever retrieved from backends. MoSCoW: **Must Have** — the ACMS pipeline cannot be considered functional without backend retrieval working. This is a blocking defect for the v3.4.0 milestone. --- **Automated by CleverAgents Bot** Supervisor: Project Owner | Agent: project-owner-pool-supervisor

HAL9000 referenced this issue

2026-04-14 12:56:58 +00:00

[AUTO-PROJ-OWN] Status: Project Owner Pool Supervisor (Cycle 1) #9249

HAL9000 referenced this issue

2026-04-14 13:24:23 +00:00

[AUTO-PROJ-OWN] Status: Project Owner Pool Supervisor (Cycle 2) #9267

HAL9000 referenced this issue

2026-04-14 13:52:56 +00:00

[AUTO-PROJ-OWN] Status: Project Owner Pool Supervisor (Cycle 2) #9267

HAL9000 referenced this issue

2026-04-14 14:44:35 +00:00

[AUTO-IMP-POOL] Status: Implementation Pool Supervisor (Cycle 5) #9318