Blackbird engine: pitch-only reconciliation + removal cycle (stacked on #34)

CLAUDE.md

@@ -235,19 +235,26 @@ authoritative.)
 ## BlackbirdBot (the scout_hub role)
 
 BlackbirdBot screens PI ideas against `data/Blackbird_initial_priorities-criteria_v1.pdf`.
-The rubric lives in **`profiles/private/blackbird.md`** (loaded per-agent from
-`profiles/private/{agent_id}.md` and injected under the `## Your Private Instructions`
-header that `Agent._compose_system_prompt` builds into every phase's system prompt); the
-per-phase behaviour lives in `prompts/roles/scout_hub/` and `src/agent/thread_guidance.py`.
+The rubric criteria and the `<assessment_json>` skeleton live directly in
+`prompts/roles/scout_hub/phase4-thread-reply.md`; the per-phase behaviour otherwise lives
+in `prompts/roles/scout_hub/` and `src/agent/thread_guidance.py`. As of the 2026-08-12
+removal cycle (private instructions + reply-only hub), there is no runtime "private
+profile" mechanism — `Agent._compose_system_prompt` no longer injects a `## Your Private
+Instructions` header, and nothing reads `profiles/private/{agent_id}.md` per-agent.
+**`profiles/private/blackbird.md`** is untracked and unread at runtime; archive-and-diff it
+against the tracked rubric text before any deploy in case it holds content that was never
+migrated (see the deploy checklist).
 
 - **Interview guidance is per-role Python**, not a prompt: `src/agent/thread_guidance.py`.
   The `pi_lab` strings there are byte-identical to the pre-refactor literals and are pinned
   by `tests/characterization/__snapshots__/test_agent_turn_gm.ambr` — do not reword them,
   and never run `pytest --snapshot-update` to make a mismatch go away.
-- **Assessments are durable.** A `:mag:` Opportunity Assessment must carry an
-  `<assessment_json>` sidecar (bare JSON, *no* ``` fence — a fenced block would be parsed
-  as the phase-5 action and silently no-op the post). It is stripped from the Slack body
-  and written to `opportunity_assessments`, visible at `/admin/assessments`.
+- **The hub is reply-only — it never makes a top-level post.** An Opportunity Assessment is
+  not a post type: it is an `<assessment_json>` sidecar carried inside the hub's CONCLUDING
+  reply in the interview thread (bare JSON, *no* ``` fence). It is stripped from the Slack
+  body before anything is posted and written to `opportunity_assessments`, visible at
+  `/admin/assessments`. `:mag:` names the sidecar, not a post label — it never appears on
+  anything a PI or another lab sees.
 - **`weighted_score` is computed**, never taken from the model:
   `src/services/blackbird_rubric.py`. `recommendation` (which may be
   `route-to-incubation`) comes straight from the model's verdict and the computed `band`
@@ -261,7 +268,3 @@ per-phase behaviour lives in `prompts/roles/scout_hub/` and `src/agent/thread_gu
   most specific terms when the full phrase misses — before that backoff existed, every
   production search ANDed in domain-generic words like "inhibitor" and returned zero hits,
   reported to PIs as clean novelty. An empty title search is never FTO.
-- **`retrieve_foa` is withheld** from this role by `prompts/roles/scout_hub/role.toml`'s
-  tool allow-list. `prompts/roles/scout_hub/agent-system.md` explicitly tells the agent it
-  does not have this tool (so it doesn't hallucinate calling it); the phase-4 template
-  (`phase4-thread-reply.md`) has no need to and does not mention it either way.

alembic/versions/0026_drop_grantbot_posted_foas.py

@@ -0,0 +1,46 @@
+"""Drop grantbot_posted_foas (GrantBot/FOA surface retired)
+
+Revision ID: 0026
+Revises: 0025
+Create Date: 2026-08-12 00:00:00.000000
+
+GrantBot and the whole funding/FOA leaf surface are being removed (branch-2
+engine reconciliation, Task 3). ``grantbot_posted_foas`` was its own dedicated
+FOA-dedup coordination table (see 0012), not a column on a shared table, so
+dropping it is a clean, isolated migration. No production data in this table
+has any value once GrantBot itself is gone — it recorded only "which FOAs
+were already posted."
+
+Downgrade recreates the table exactly as 0012's upgrade() built it (4 columns,
+no legacy-JSON backfill — that one-time seed step is not reproducible here).
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+
+from alembic import op
+
+revision: str = "0026"
+down_revision: Union[str, None] = "0025"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.drop_table("grantbot_posted_foas")
+
+
+def downgrade() -> None:
+    op.create_table(
+        "grantbot_posted_foas",
+        sa.Column("foa_number", sa.String(50), primary_key=True),
+        sa.Column(
+            "posted_at",
+            sa.DateTime(timezone=True),
+            server_default=sa.func.now(),
+            nullable=False,
+        ),
+        sa.Column("channel", sa.String(100), nullable=True),
+        sa.Column("title", sa.Text(), nullable=True),
+    )

docker-compose.prod.yml

@@ -109,33 +109,6 @@ services:
         awslogs-create-group: "true"
         awslogs-region: ${AWS_REGION:-us-east-2}
 
-  grantbot:
-    build:
-      context: .
-    restart: unless-stopped
-    command: ["python", "-m", "src.agent.grantbot", "scheduler", "--run-hour", "8", "--max-per-channel", "1"]
-    env_file: .env
-    environment:
-      DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-copi}:${POSTGRES_PASSWORD}@postgres:5432/${POSTGRES_DB:-copi}
-      # Fail fast at `docker compose up` if no secret is provided, and default
-      # the process to production so the insecure default SECRET_KEY is rejected.
-      SECRET_KEY: ${SECRET_KEY:?Set a strong random SECRET_KEY in .env}
-      ENVIRONMENT: ${ENVIRONMENT:-production}
-    volumes:
-      - ./profiles:/app/profiles
-      - ./prompts:/app/prompts
-      - ./data:/app/data
-    depends_on:
-      postgres:
-        condition: service_healthy
-    logging:
-      driver: awslogs
-      options:
-        awslogs-group: /copi/grantbot
-        tag: grantbot
-        awslogs-create-group: "true"
-        awslogs-region: ${AWS_REGION:-us-east-2}
-
   nginx:
     image: nginx:1.27-alpine
     restart: unless-stopped

docker-compose.yml

@@ -62,17 +62,5 @@ services:
     profiles:
       - agent
 
-  grantbot:
-    build: .
-    command: python -m src.agent.grantbot scheduler --run-hour 8 --max-per-channel 1
-    env_file: .env
-    volumes:
-      - .:/app
-      - ./profiles:/app/profiles
-      - ./prompts:/app/prompts
-    depends_on:
-      postgres:
-        condition: service_healthy
-
 volumes:
   pgdata: