release: v2.0.5 — honest numbers (impact endpoint, mistake-avoided lie, shape drift)

Three correctness bugs hiding behind quiet metrics:

1. /api/v1/impact crashed on a fresh DB. ImpactEvent lives in
   engine/impact.py, but init_db() only imports storage/models.py
   before Base.metadata.create_all(). The impact_events table never
   existed on a fresh DB and any code path that hit record_impact
   died on `no such table`. Force-import memee.engine.impact inside
   init_db() before create_all so the table is present everywhere.

2. MISTAKE_AVOIDED was being credited for warnings the agent
   IGNORED, as long as outcome=success. The CLI hook line read
   warnings_avoided=N over numbers that were always
   warnings_violated. New ImpactType.WARNING_INEFFECTIVE for
   violated-but-task-passed; MISTAKE_AVOIDED is now reserved for
   evidence-backed behaviour change. CLI line renamed to
   warnings_violated, which is what it always was.

3. get_impact_summary returned 8 keys when empty, 18 otherwise —
   forced every caller into defensive .get() branching. Now: one
   shape with zeros / {} / [] in the empty case.

Plus: doc drift — README, CLAUDE.md, launch-copy and the historical
review-fixes doc all claimed 24 MCP tools; the real number is 19
(the five research_* tools were removed in v2.0.0). hooks_config.py
comments described /dev/null redirect and stderr writes that the
commands don't actually do; comments rewritten to match the
commands.

Six new tests under test_v2_0_5_fixes.py cover the three correctness
fixes. Existing test_learn_auto updated for the renamed CLI field.
368/368 trimmed-suite tests pass; the v2.0.4 ledger guard confirms
the suite leaves ~/.memee/packs.json byte-identical.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Tomas Pflanzer

CHANGELOG.md

@@ -7,6 +7,64 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+
+## [2.0.5] — 2026-04-28
+
+The "honest numbers" patch. v2.0.4 had three correctness bugs hiding
+behind quiet metrics. v2.0.5 makes the numbers say what actually
+happened.
+
+### Fixed
+
+- **`/api/v1/impact` worked on populated DBs and crashed on fresh
+  ones.** ``ImpactEvent`` is defined in ``engine/impact.py`` (so
+  ``record_impact`` lives next to the metric logic), but ``init_db``
+  only imports ``storage.models`` before calling ``Base.metadata.
+  create_all``. On a fresh DB the ``impact_events`` table never got
+  created and any code path that hit ``record_impact`` died on
+  ``no such table``. v2.0.5 force-imports ``memee.engine.impact``
+  inside ``init_db`` before ``create_all`` so the table is present
+  on every fresh DB. Regression test seeds a brand-new DB and
+  asserts ``inspect(engine).get_table_names()`` contains it.
+
+- **`MISTAKE_AVOIDED` lied when the agent ignored the warning.**
+  ``post_task_review`` recorded ``MISTAKE_AVOIDED`` for any warning
+  the agent *violated* as long as the task ended in ``outcome="success"``.
+  The metric read like a win for behaviour the agent had ignored.
+  v2.0.5 introduces a new ``ImpactType.WARNING_INEFFECTIVE`` for that
+  exact case and reserves ``MISTAKE_AVOIDED`` for evidence-backed
+  behaviour change. ``MISTAKE_MADE`` (violation + failure) is
+  unchanged. The CLI hook line that surfaced this as
+  ``warnings_avoided=N`` now reads ``warnings_violated=N``, which is
+  what the underlying number always was.
+
+- **`get_impact_summary` returned a different key set when empty.**
+  Eight keys on empty, eighteen otherwise — every consumer was
+  forced into defensive ``.get(key, 0)`` branching. v2.0.5 returns
+  the same key set in both cases, with zeros / ``{}`` / ``[]``
+  instead of missing keys.
+
+- **Docs claimed 24 MCP tools; the real count is 19.** README,
+  CLAUDE.md, launch copy and the historical review-fixes doc all
+  carried the v1.x number through the v2.0.0 deletion. The
+  ``research_*`` MCP tools that lived under "Research:" were
+  removed in v2.0.0 with the autoresearch engine; v2.0.5 lists
+  the 19 tools that actually ship.
+
+- **`hooks_config.py` comments described behaviour the commands
+  did not have** — Stop "redirects to /dev/null" (it does not),
+  UserPromptSubmit "writes to stderr" (it writes to stdout, which is
+  what Claude Code surfaces). Comments were rewritten to match the
+  commands; commands themselves were not changed because the current
+  behaviour is the one we want.
+
+### Tests
+
+- ``tests/test_v2_0_5_fixes.py`` — six new cases covering the four
+  fixes above. Plus the autouse ``_isolate_pack_ledger`` fixture
+  from v2.0.4 means none of these tests can pollute the developer
+  ledger.
+
 ## [2.0.4] — 2026-04-27
 
 The "tests stop polluting the developer's ledger" patch.

CLAUDE.md

@@ -181,8 +181,8 @@ Memee ships as two packages, with clear licence separation:
 
 | Package | Licence | What it adds |
 |---|---|---|
-| **`memee`** (this repo) | MIT | Full single-user product: every engine module, MCP server, CLI, CMAM adapter, dashboard. No users, no teams, no scope enforcement. |
-| **`memee-team`** (private repo, licence-gated) | Proprietary (EULA) | `User` + `Team` SQLAlchemy models, `scoping.py` engine (personal → team → org promotion), SSO (SAML/OIDC), audit log export, RBAC, multi-user dashboard auth, licence key verification. |
+| **`memee`** (this repo) | MIT | Full single-user product: every engine module, MCP server, CLI, CMAM adapter. No users, no teams, no scope enforcement. |
+| **`memee-team`** (private repo, licence-gated) | Proprietary (EULA) | `User` + `Team` SQLAlchemy models, `scoping.py` engine (personal → team → org promotion), SSO (SAML/OIDC), audit log export, RBAC, licence key verification. |
 
 `memee-team` plugs into OSS via `memee.plugins` hooks
 (`current_user_id`, `visible_memories`, `promote`, `can_promote`, `on_record`).
@@ -203,21 +203,22 @@ Pricing model reflects "Memee is memory, not model" — flat per-team
 (like Supabase, Vercel, Plausible), not per-seat (like Copilot, Cursor).
 Value scales sublinearly with headcount: one canon serves the whole team.
 
-## MCP Tools (23)
+## MCP Tools (19)
 
-Core: memory_record, memory_search, memory_suggest, memory_validate,
-memory_invalidate, decision_record, antipattern_record, antipattern_check
+Core: memory_record, memory_search, search_feedback, memory_suggest,
+memory_validate, memory_invalidate, decision_record, antipattern_record,
+antipattern_check
 
 Intelligence: propagate_patterns, predict_warnings, inherit_knowledge,
 run_dream, review_code, get_briefing, post_task_feedback
 
-Research: research_create, research_log, research_status, research_meta,
-research_complete
-
 Analytics: learning_status, canon_list
 
 Delivery: sync_to_cmam (push canon to Claude Managed Agents Memory)
 
+(The five ``research_*`` tools that lived here through v1.x were removed
+in v2.0.0 along with the autoresearch engine.)
+
 ## CMAM Bridge (Claude Managed Agents Memory)
 
 Anthropic's managed memory is a filesystem-style store at `/mnt/memory/` inside
@@ -276,11 +277,11 @@ MCP tool `sync_to_cmam` lets agents trigger the push themselves.
 | File | Purpose |
 |------|---------|
 | `src/memee/cli.py` | 25+ Click commands (incl. `cmam sync`/`cmam status`) |
-| `src/memee/mcp_server.py` | 24 MCP tools |
+| `src/memee/mcp_server.py` | 19 MCP tools |
 | `src/memee/adapters/cmam.py` | Claude Managed Agents Memory bridge |
 | `src/memee/storage/models.py` | 15 SQLAlchemy models |
 | `src/memee/storage/database.py` | DB init, FTS5, WAL mode |
-| `src/memee/api/routes/dashboard.py` | Chart.js dashboard |
+
 | `src/memee/api/routes/api_v1.py` | REST API (12+ endpoints) |
 | `src/memee/installer.py` | Interactive setup wizard |
 | `src/memee/doctor.py` | Health check + auto-configure AI tools |
@@ -304,5 +305,5 @@ test_real_impact (A/B with/without), test_perf_simulation (9 scenarios)
 - 33 commits on feat/initial-setup
 - 63 Python files, 18,899 lines of code
 - 201 tests passing
-- 16 engine modules + CMAM adapter, 24 MCP tools, 12+ API endpoints (GET-only dashboard API)
+- 16 engine modules + CMAM adapter, 19 MCP tools, 12+ API endpoints (GET-only dashboard API)
 - MIT licence (OSS `memee`), proprietary EULA for `memee-team`

README.md

@@ -125,7 +125,7 @@ Full methodology + per-repo file sizes: [docs/benchmarks.md](docs/benchmarks.md)
 
 ## Using it with Claude, GPT, Gemini
 
-An MCP server with 24 tools ships with the install. Drop this into `~/.claude/settings.json` — or the Cursor / Continue / any MCP-capable client equivalent:
+An MCP server with 19 tools ships with the install. Drop this into `~/.claude/settings.json` — or the Cursor / Continue / any MCP-capable client equivalent:
 
 ```json
 {

docs/launch-copy.md

@@ -99,7 +99,7 @@ Char count: 271. Hook in 4 words. One verb (writes). One outcome (every model re
 > memee setup
 > memee doctor
 >
-> MCP server with 24 tools ships in the box. Drop it into Claude Code, Cursor, Continue, anything MCP-shaped. Memory lives in ~/.memee/memee.db.
+> MCP server with 19 tools ships in the box. Drop it into Claude Code, Cursor, Continue, anything MCP-shaped. Memory lives in ~/.memee/memee.db.
 >
 > MIT. github.com/gizmax/memee
 

docs/review-fixes.md

@@ -2,6 +2,13 @@
 
 A thorough internal review (2026-04-24) measured Memee on impact tests, token calculations, OrgMemEval, large simulations, and a handcrafted 12-memory retrieval benchmark. The numbers confirmed the headline claims (71 % less time, 65 % fewer iterations, 96 % token reduction) but surfaced seven concrete problems. All seven were addressed in a single session using parallel fix-agents. This document records what changed, with before/after numbers, so future reviewers can see our work.
 
+> **Historical note (v2.0.5):** the dashboard panels referenced below
+> (`Retrieval health`, the impact-panel three-card layout) describe the
+> v1.x web dashboard. That dashboard was removed in v2.0.0 along with the
+> autoresearch engine. The same metrics still live in the underlying
+> tables and are surfaced in `memee status` and the JSON API; the panels
+> themselves do not exist any more.
+
 ## Summary
 
 | # | Problem | Before | After | Where |

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "memee"
-version = "2.0.4"
+version = "2.0.5"
 description = "Cross-model shared memory for AI agents. One canon. Every model, every project, every teammate."
 readme = "README.md"
 requires-python = ">=3.11"

src/memee/__init__.py

@@ -1,3 +1,3 @@
 """Memee — Institutional memory for AI agent companies."""
 
-__version__ = "2.0.4"
+__version__ = "2.0.5"

src/memee/cli.py

@@ -1318,8 +1318,12 @@ def learn(auto, project, diff_text, outcome, agent, model):
         # Silent unless something was learned. A noisy hook gets disabled.
         if patterns_followed == 0 and warnings_violated == 0:
             return
+        # Report what actually happened. ``warnings_violated`` is exactly
+        # that — counts of warnings the agent ignored. v2.0.4 and earlier
+        # surfaced this number as ``warnings_avoided``, which read like a
+        # win. It wasn't.
         click.echo(
-            f"memee learn: ok (warnings_avoided={warnings_violated}, "
+            f"memee learn: ok (warnings_violated={warnings_violated}, "
             f"patterns_followed={patterns_followed}, new_patterns={new_patterns})"
         )
     else:

src/memee/engine/feedback.py

@@ -99,23 +99,32 @@ def post_task_review(
                 continue
 
             # A violation means the agent wrote code that matches a known
-            # anti-pattern. If the task succeeded anyway we still credit
-            # MISTAKE_AVOIDED (the warning surfaced in review was visible and
-            # the agent's work passed). If the task failed — the warning was
-            # ignored AND there's a real negative outcome — it's a MISTAKE_MADE.
-            impact_kind = (
-                ImpactType.MISTAKE_AVOIDED.value
-                if outcome == "success"
-                else ImpactType.MISTAKE_MADE.value
-            )
+            # anti-pattern. We map outcome to one of three honest states:
+            #   - failure → MISTAKE_MADE     (warning ignored, real damage)
+            #   - success → WARNING_INEFFECTIVE (warning ignored, got lucky)
+            # We deliberately do NOT credit MISTAKE_AVOIDED here — that
+            # would be a metric that lies about what happened, and bad
+            # numbers ruin the trust users have to put in this dashboard
+            # for it to be worth running.
+            if outcome == "success":
+                impact_kind = ImpactType.WARNING_INEFFECTIVE.value
+                outcome_text = (
+                    "Warning ignored; task succeeded anyway. The warning "
+                    "did not change the agent's behaviour."
+                )
+            else:
+                impact_kind = ImpactType.MISTAKE_MADE.value
+                outcome_text = (
+                    f"Warning ignored. Task outcome: {outcome}"
+                )
             record_impact(
                 session, mem.id,
                 impact_kind,
                 agent=agent, model=model, project_id=project_id,
                 trigger="Post-task review detected anti-pattern violation",
                 memory_shown=mem.title,
                 agent_action=f"Violated warning: {warning.get('title', '')}",
-                outcome=f"Warning ignored. Task outcome: {outcome}",
+                outcome=outcome_text,
                 severity_avoided=warning.get("severity", "medium"),
             )
 

src/memee/engine/impact.py

@@ -23,6 +23,7 @@
 class ImpactType(str, Enum):
     MISTAKE_AVOIDED = "mistake_avoided"       # Agent got warning, changed approach
     MISTAKE_MADE = "mistake_made"             # Agent ignored warning, repeated mistake
+    WARNING_INEFFECTIVE = "warning_ineffective"  # Agent ignored warning but task succeeded anyway — warning didn't help, didn't hurt
     TIME_SAVED = "time_saved"                 # Pattern reuse saved iteration time
     DECISION_INFORMED = "decision_informed"   # Decision made with historical context
     CODE_CHANGED = "code_changed"             # Code diff proves behavior change
@@ -235,18 +236,11 @@ def get_impact_summary(session: Session) -> dict:
     investment_minutes = total_memories * 5
     roi = total_time_saved / investment_minutes if investment_minutes > 0 else 0
 
-    if not events and warnings_shown == 0:
-        return {
-            "total_events": 0,
-            "warnings_shown": 0,
-            "warnings_shown_unique": 0,
-            "warnings_acknowledged": 0,
-            "warnings_acknowledged_unique": 0,
-            "mistakes_avoided": 0,
-            "mistakes_avoided_unique": 0,
-            "mistakes_made": 0,
-        }
-
+    # Single shape for both empty and populated cases. v2.0.4 and earlier
+    # returned a subset of keys when there was nothing to report; that
+    # forced every consumer (CLI, dashboard, tests) into defensive
+    # branching. Now: same key set, zeros / empty containers when no
+    # data — callers can read the shape without ``.get(key, 0)``.
     return {
         "total_events": len(events),
         "total_time_saved_minutes": round(total_time_saved, 1),