feed-the-machine 1.6.1 → 1.7.0

package/ftm-mind/evals/promptfoo.yaml

@@ -1,142 +1,142 @@
-description: "ftm-mind routing accuracy"
-
-prompts:
-  - "You are ftm-mind. Given this user request, which ftm skill should handle it? Reply with just the skill name.\n\nUser: {{input}}"
-
-providers:
-  - id: "exec:claude --print"
-    label: "claude-code"
-
-tests:
-  - description: "debug request routes to ftm-debug"
-    vars:
-      input: "debug this flaky test"
-    assert:
-      - type: contains
-        value: "ftm-debug"
-
-  - description: "design thinking routes to ftm-brainstorm"
-    vars:
-      input: "help me think through auth design"
-    assert:
-      - type: contains
-        value: "ftm-brainstorm"
-
-  - description: "plan execution routes to ftm-executor"
-    vars:
-      input: "execute ~/.claude/plans/foo.md"
-    assert:
-      - type: contains
-        value: "ftm-executor"
-
-  - description: "trivial rename routes to micro or direct handling"
-    vars:
-      input: "rename this variable"
-    assert:
-      - type: regex
-        value: "micro|direct"
-
-  - description: "multi-model deliberation routes to ftm-council"
-    vars:
-      input: "what would other AIs think about this approach"
-    assert:
-      - type: contains
-        value: "ftm-council"
-
-  - description: "wiring audit routes to ftm-audit"
-    vars:
-      input: "audit the wiring"
-    assert:
-      - type: contains
-        value: "ftm-audit"
-
-  - description: "pause request routes to ftm-pause"
-    vars:
-      input: "pause"
-    assert:
-      - type: contains
-        value: "ftm-pause"
-
-  - description: "save state request routes to ftm-pause"
-    vars:
-      input: "save state"
-    assert:
-      - type: contains
-        value: "ftm-pause"
-
-  - description: "resume request routes to ftm-resume"
-    vars:
-      input: "resume"
-    assert:
-      - type: contains
-        value: "ftm-resume"
-
-  - description: "continue request routes to ftm-resume"
-    vars:
-      input: "continue where we left off"
-    assert:
-      - type: contains
-        value: "ftm-resume"
-
-  - description: "upgrade request routes to ftm-upgrade"
-    vars:
-      input: "upgrade ftm"
-    assert:
-      - type: contains
-        value: "ftm-upgrade"
-
-  - description: "config check routes to ftm-config"
-    vars:
-      input: "check my config"
-    assert:
-      - type: contains
-        value: "ftm-config"
-
-  - description: "ftm settings routes to ftm-config"
-    vars:
-      input: "ftm settings"
-    assert:
-      - type: contains
-        value: "ftm-config"
-
-  - description: "intent doc update routes to ftm-intent"
-    vars:
-      input: "update the intent docs"
-    assert:
-      - type: contains
-        value: "ftm-intent"
-
-  - description: "architecture diagram update routes to ftm-diagram"
-    vars:
-      input: "update architecture diagram"
-    assert:
-      - type: contains
-        value: "ftm-diagram"
-
-  - description: "browser task routes to ftm-browse"
-    vars:
-      input: "open browser and check the page"
-    assert:
-      - type: contains
-        value: "ftm-browse"
-
-  - description: "codex validation routes to ftm-codex-gate"
-    vars:
-      input: "run codex validation"
-    assert:
-      - type: contains
-        value: "ftm-codex-gate"
-
-  - description: "retrospective routes to ftm-retro"
-    vars:
-      input: "how did that go"
-    assert:
-      - type: contains
-        value: "ftm-retro"
-
-  - description: "retro keyword routes to ftm-retro"
-    vars:
-      input: "retro"
-    assert:
-      - type: contains
-        value: "ftm-retro"
+description: "ftm-mind routing accuracy"
+
+prompts:
+  - "You are ftm-mind. Given this user request, which ftm skill should handle it? Reply with just the skill name.\n\nUser: {{input}}"
+
+providers:
+  - id: "exec:claude --print"
+    label: "claude-code"
+
+tests:
+  - description: "debug request routes to ftm-debug"
+    vars:
+      input: "debug this flaky test"
+    assert:
+      - type: contains
+        value: "ftm-debug"
+
+  - description: "design thinking routes to ftm-brainstorm"
+    vars:
+      input: "help me think through auth design"
+    assert:
+      - type: contains
+        value: "ftm-brainstorm"
+
+  - description: "plan execution routes to ftm-executor"
+    vars:
+      input: "execute ~/.claude/plans/foo.md"
+    assert:
+      - type: contains
+        value: "ftm-executor"
+
+  - description: "trivial rename routes to micro or direct handling"
+    vars:
+      input: "rename this variable"
+    assert:
+      - type: regex
+        value: "micro|direct"
+
+  - description: "multi-model deliberation routes to ftm-council"
+    vars:
+      input: "what would other AIs think about this approach"
+    assert:
+      - type: contains
+        value: "ftm-council"
+
+  - description: "wiring audit routes to ftm-audit"
+    vars:
+      input: "audit the wiring"
+    assert:
+      - type: contains
+        value: "ftm-audit"
+
+  - description: "pause request routes to ftm-pause"
+    vars:
+      input: "pause"
+    assert:
+      - type: contains
+        value: "ftm-pause"
+
+  - description: "save state request routes to ftm-pause"
+    vars:
+      input: "save state"
+    assert:
+      - type: contains
+        value: "ftm-pause"
+
+  - description: "resume request routes to ftm-resume"
+    vars:
+      input: "resume"
+    assert:
+      - type: contains
+        value: "ftm-resume"
+
+  - description: "continue request routes to ftm-resume"
+    vars:
+      input: "continue where we left off"
+    assert:
+      - type: contains
+        value: "ftm-resume"
+
+  - description: "upgrade request routes to ftm-upgrade"
+    vars:
+      input: "upgrade ftm"
+    assert:
+      - type: contains
+        value: "ftm-upgrade"
+
+  - description: "config check routes to ftm-config"
+    vars:
+      input: "check my config"
+    assert:
+      - type: contains
+        value: "ftm-config"
+
+  - description: "ftm settings routes to ftm-config"
+    vars:
+      input: "ftm settings"
+    assert:
+      - type: contains
+        value: "ftm-config"
+
+  - description: "intent doc update routes to ftm-intent"
+    vars:
+      input: "update the intent docs"
+    assert:
+      - type: contains
+        value: "ftm-intent"
+
+  - description: "architecture diagram update routes to ftm-diagram"
+    vars:
+      input: "update architecture diagram"
+    assert:
+      - type: contains
+        value: "ftm-diagram"
+
+  - description: "browser task routes to ftm-browse"
+    vars:
+      input: "open browser and check the page"
+    assert:
+      - type: contains
+        value: "ftm-browse"
+
+  - description: "codex validation routes to ftm-codex-gate"
+    vars:
+      input: "run codex validation"
+    assert:
+      - type: contains
+        value: "ftm-codex-gate"
+
+  - description: "retrospective routes to ftm-retro"
+    vars:
+      input: "how did that go"
+    assert:
+      - type: contains
+        value: "ftm-retro"
+
+  - description: "retro keyword routes to ftm-retro"
+    vars:
+      input: "retro"
+    assert:
+      - type: contains
+        value: "ftm-retro"

package/ftm-mind/references/blackboard-protocol.md

@@ -0,0 +1,110 @@
+# Blackboard Loading Protocol
+
+Read the blackboard in this order:
+
+1. `context.json`
+2. `experiences/index.json`
+3. `patterns.json`
+
+Use these exact paths:
+
+- `~/.claude/ftm-state/blackboard/context.json`
+- `~/.claude/ftm-state/blackboard/experiences/index.json`
+- `~/.claude/ftm-state/blackboard/patterns.json`
+
+## context.json
+
+Use `context.json` for live session state only.
+
+Pull out:
+
+- `current_task`: does the request continue the active thread or branch away from it?
+- `recent_decisions`: what did we already decide this session?
+- `active_constraints`: no auto-commit, avoid production, stay terse, etc.
+- `user_preferences`: communication and approval preferences
+- `session_metadata.skills_invoked`: what workflow is already underway?
+
+Key heuristic:
+
+- trajectory matters more than isolated wording
+
+If the last sequence was brainstorm -> plan -> execute, then "go ahead" means something different than if the session began 10 seconds ago.
+
+## Experience Retrieval
+
+Experience retrieval must be concrete, not hand-wavy.
+
+Protocol:
+
+1. Read `~/.claude/ftm-state/blackboard/experiences/index.json`
+2. Parse `entries`
+3. Derive a current `task_type`
+4. Derive current tags from the request and codebase context
+5. Filter entries where:
+   - `task_type` matches the current task type, or
+   - there is at least one overlapping tag
+6. Sort filtered entries by `recorded_at` descending
+7. Load the top 3-5 matching experience files from:
+   - `~/.claude/ftm-state/blackboard/experiences/{filename}`
+8. Prefer lessons from entries with:
+   - `outcome: success`
+   - higher `confidence`
+   - recent dates
+9. Synthesize the lessons into concrete adjustments to the current approach
+
+Derive tags from:
+
+- language or framework names
+- domain nouns like `auth`, `poller`, `slack`, `database`, `deploy`, `calendar`, `jira`
+- task shape like `flaky-test`, `refactor`, `ticket-triage`, `plan-execution`
+
+Use retrieved experience for:
+
+- complexity calibration
+- known pitfalls
+- better sequencing
+- better routing
+- faster first checks
+
+Never use experience to blindly repeat an old approach when the live context has changed.
+
+## Pattern Registry
+
+Read `patterns.json` after experience retrieval.
+
+Scan all four sections:
+
+- `codebase_insights`
+- `execution_patterns`
+- `user_behavior`
+- `recurring_issues`
+
+Apply patterns only when they materially match the present case.
+
+Examples:
+
+- matching `file_pattern` on touched files
+- recurring issue symptoms that fit the current failure
+- user behavior that affects response style or approval expectations
+- execution patterns that suggest a proven sequence
+
+Patterns are promoted summaries. They should speed up orientation, not replace it.
+
+## Cold-Start Behavior
+
+Cold start is normal.
+
+When the blackboard is empty:
+
+- do not apologize
+- do not say capability is reduced
+- do not surface that memory is empty unless the user asked
+- operate at full capability using live observation, codebase state, MCP awareness, and base heuristics
+
+Warm start adds shortcuts. Cold start is still a smart engineer on day 1 at a new job.
+
+If `experiences/index.json` has no usable matches:
+
+- continue normally
+- lean harder on current repo state and direct inspection
+- record the resulting experience aggressively after completion