feed-the-machine 1.0.0

package/ftm-retro/SKILL.md

@@ -0,0 +1,189 @@
+---
+name: ftm-retro
+description: Post-execution self-assessment skill. Automatically triggered after ftm-executor completes a plan. Scores execution across 5 dimensions, identifies what went well and what was slow, writes structured report with improvement suggestions. Use when user says "retro", "retrospective", "how did that go", "execution review", "self-assessment", "ftm retro".
+---
+
+## Events
+
+### Emits
+- `experience_recorded` — when a task outcome, fix attempt, or blocker is written to the blackboard experience log
+- `pattern_discovered` — when a recurring pattern is identified from accumulated experiences and promoted to patterns.json
+- `task_completed` — when the retro report is saved and the self-assessment session concludes
+
+### Listens To
+- `task_completed` — micro-reflection trigger: record the task outcome as a structured experience entry
+- `error_encountered` — failure analysis: record the error context as a failure experience for pattern learning
+- `bug_fixed` — success recording: record the fix details as a positive experience (what worked, what the root cause was)
+
+# FTM Retro — Post-Execution Self-Assessment
+
+Structured retrospective system for ftm-executor plans. Scores execution across 5 evidence-based dimensions, surfaces bottlenecks with specifics, and builds a cumulative pattern library that makes each future execution smarter.
+
+## Why This Exists
+
+Execution without reflection is a loop with no exit. FTM-retro closes the feedback cycle: every plan run generates a scored report, every report feeds a pattern library, and recurring issues get escalated until they're fixed. The goal is measurable improvement across executions, not vibes.
+
+## Operating Modes
+
+### Mode 1: Auto-triggered by ftm-executor (Phase 6.5)
+
+FTM-executor calls this skill after all waves complete and the final commit is made. It passes execution context directly:
+
+- Plan title and absolute path
+- Task count and wave count
+- Total agents spawned
+- Per-task audit results: pass/fail/auto-fix counts per phase
+- Codex gate results per wave (pass/fail + any failures found)
+- Total execution duration
+- Errors, blockers, or manual interventions that occurred
+
+When invoked in this mode, proceed directly to scoring — all data is available.
+
+### Mode 2: Manual (`/ftm retro`)
+
+When invoked without execution context:
+
+1. Search the current project for the most recent `PROGRESS.md` file. Read it fully to reconstruct what ran.
+2. If no `PROGRESS.md` exists, check `~/.claude/ftm-retros/` for the most recent `.md` file and ask the user which execution they want to review.
+3. Once context is established, proceed to scoring.
+
+Never ask the user to provide data you can find yourself. Read the files.
+
+---
+
+## Scoring Flow
+
+Score execution across 5 dimensions. Read `references/protocols/SCORING-RUBRICS.md` for the full rubric for each dimension (scale breakpoints, evidence requirements, formula for Dimension 4).
+
+**The 5 dimensions:**
+
+1. **Wave Parallelism Efficiency** — were independent tasks actually dispatched in parallel?
+2. **Audit Pass Rate** — what percentage of tasks passed ftm-audit on the first attempt?
+3. **Codex Gate Pass Rate** — what percentage of waves passed the ftm-codex-gate on the first attempt?
+4. **Retry and Fix Count** — how many total review-fix cycles were needed? Lower is better.
+5. **Execution Smoothness** — evidence-grounded assessment of blockers, ambiguities, and manual interventions.
+
+Every score requires a citation to specific data. If data for a dimension is unavailable, note the gap and score conservatively.
+
+---
+
+## Report Generation
+
+### Step 1: Create retro directory
+
+```bash
+mkdir -p ~/.claude/ftm-retros/
+```
+
+### Step 2: Check for past retros
+
+Before writing anything, check whether any `.md` files exist in `~/.claude/ftm-retros/`. If they do, read them all. You will use them for the Pattern Analysis section.
+
+### Step 3: Write the report
+
+Read `references/templates/REPORT-FORMAT.md` for the exact output template, slug generation rules, and section format.
+
+Save to: `~/.claude/ftm-retros/{plan-slug}-{YYYY-MM-DD}.md`
+
+---
+
+## Key Behaviors
+
+### Improvement specificity
+
+"Improve parallelism" is not an improvement proposal. "Add a dependency pre-check step to ftm-executor Phase 2 that flags tasks with no declared dependencies as parallelizable, and warn when they are dispatched serially" is an improvement proposal. Every proposed improvement must be concrete enough that a future session could implement it from the description alone.
+
+### Pattern escalation
+
+Recurring issues that have appeared in 3+ retros without being addressed should be flagged with `[ESCALATED - 3+ occurrences]` and moved to the top of the Proposed Improvements list. These are systemic problems, not one-off noise.
+
+---
+
+## Output
+
+After saving the retro file, print to the user:
+
+```
+Retro saved: ~/.claude/ftm-retros/{filename}
+
+Overall: {score}/50
+Top issue: {single most impactful bottleneck in one sentence}
+Top suggestion: {single highest-value proposed improvement in one sentence}
+```
+
+Do not print the full report to the terminal — it lives in the file.
+
+---
+
+## Micro-Reflection Mode
+
+Micro-reflections are lightweight experience entries recorded after significant actions — not just full executor runs. The mind triggers this mode via the `task_completed`, `error_encountered`, and `bug_fixed` events.
+
+### Trigger Events
+- `task_completed` — any task completion (micro through large)
+- `bug_fixed` — a bug was resolved
+- `error_encountered` — an unexpected error during execution
+- `code_committed` — a meaningful commit was made
+- `plan_generated` — a plan was created from brainstorming
+- `user_correction` — the user corrected the mind's approach
+
+### Reflection Format (Verbal RL)
+
+For each trigger, generate a structured reflection:
+
+"I [succeeded/failed/partially succeeded] at [task description] because [specific reason].
+Next time I should [concrete actionable adjustment].
+Confidence: [low/medium/high]"
+
+### Experience Entry Creation
+
+Write a structured experience entry to `~/.claude/ftm-state/blackboard/experiences/YYYY-MM-DD_task-slug.json`.
+
+Key fields:
+- `task_type`: derived from the task
+- `description`: 1-2 sentence summary
+- `approach`: what was tried
+- `outcome`: success/partial/failure
+- `lessons`: concrete, actionable takeaways — the verbal RL reflection above, decomposed into individual lesson strings
+- `complexity_estimated` vs `complexity_actual`: track both for calibration
+- `capabilities_used`: skills, MCPs, and agent types activated
+- `tags`: searchable labels
+- `confidence`: low for first-time observations, medium for confirmed patterns
+
+### Pattern Extraction
+
+After writing an experience, check for pattern promotion:
+
+1. Read `experiences/index.json`
+2. Count entries with overlapping `task_type` AND `tags` that share the same lesson theme
+3. If 3+ similar experiences exist with the same lesson → promote to `patterns.json`:
+   - Choose the appropriate category (codebase_insights, execution_patterns, user_behavior, recurring_issues)
+   - Set `confidence: "low"` for newly promoted patterns (3 occurrences)
+   - Raise to `"medium"` at 5+, `"high"` at 8+
+
+### Pattern Decay
+
+Patterns that are not reinforced within 30 days should have their confidence reduced:
+- `high` → `medium`
+- `medium` → `low`
+- `low` → remove from patterns.json
+
+Check for decay when reading patterns.json during any blackboard operation.
+
+### Cold-Start Behavior
+
+During the first ~10 interactions (when `experiences/index.json` has `total_count < 10`):
+- Record EVERY completed task, even trivial ones
+- Set `confidence: "low"` on all entries
+- Prioritize breadth of recording over depth of analysis
+
+## Blackboard Write
+
+After completing, update:
+1. `~/.claude/ftm-state/blackboard/context.json`:
+   - Set current_task status to "complete"
+   - Append decision summary to recent_decisions (cap at 10)
+   - Update session_metadata.skills_invoked and last_updated
+2. Write experience file to `~/.claude/ftm-state/blackboard/experiences/YYYY-MM-DD_task-slug.json`
+3. Update `experiences/index.json` with the new entry
+4. Emit `task_completed`

package/ftm-retro/references/protocols/SCORING-RUBRICS.md

@@ -0,0 +1,89 @@
+# Scoring Rubrics — All 5 Dimensions
+
+Detailed rubrics for ftm-retro scoring. Each dimension is scored 0–10 with a citation to specific data. Do not estimate without evidence — if data is missing, note it and score conservatively.
+
+---
+
+## Dimension 1: Wave Parallelism Efficiency (0–10)
+
+Were independent tasks actually dispatched in parallel? Could more tasks have been parallelized?
+
+- **10**: Every task that could run in parallel did. No serial bottlenecks where parallelism was possible.
+- **7–9**: Minor serial steps that could have been parallel (e.g., final post-processing tasks run sequentially).
+- **4–6**: Significant parallelism opportunities missed. Tasks that had no dependencies ran serially.
+- **1–3**: Nearly all tasks ran serially despite having no dependencies on each other.
+- **0**: Everything was serial regardless of dependency structure.
+
+Evidence to cite: wave structure from PROGRESS.md, task dependency graph, agent dispatch timestamps.
+
+---
+
+## Dimension 2: Audit Pass Rate (0–10)
+
+What percentage of tasks passed ftm-audit on the first attempt?
+
+- **10**: 100% first-pass. No task needed a fix cycle.
+- **8**: 90%+ first-pass. One or two tasks needed minor fixes.
+- **6**: 75–89% first-pass.
+- **4**: 50–74% first-pass. Roughly half the tasks needed audit remediation.
+- **2**: Below 50% first-pass.
+- **0**: Every single task failed audit on the first attempt.
+
+Evidence to cite: per-task audit results (pass/fail counts, auto-fix counts, manual-fix counts).
+
+---
+
+## Dimension 3: Codex Gate Pass Rate (0–10)
+
+What percentage of waves passed the ftm-codex-gate on the first attempt?
+
+- **10**: All waves passed on first gate run.
+- **7–9**: One wave needed a fix-and-retry.
+- **4–6**: Multiple waves needed retries.
+- **1–3**: Most waves failed the gate at least once.
+- **0**: Every wave failed the gate.
+
+Evidence to cite: codex gate results per wave (pass/fail, failure types).
+
+---
+
+## Dimension 4: Retry and Fix Count (0–10)
+
+How many total review-fix cycles were needed across all tasks and waves? Lower is better.
+
+Formula: `score = max(0, 10 - (total_retries / task_count) * 5)`
+
+- **10**: Zero retries.
+- **8**: Fewer than 0.5 retries per task on average.
+- **6**: 0.5–1.0 retries per task.
+- **4**: 1–2 retries per task.
+- **2**: 2–3 retries per task.
+- **0**: More than 3 retries per task on average.
+
+Evidence to cite: total retries, broken down by type (audit fix, codex gate retry, manual intervention).
+
+---
+
+## Dimension 5: Execution Smoothness (0–10)
+
+Subjective but evidence-grounded assessment. Were there blockers, ambiguous plan steps, confusing errors, or required manual interventions?
+
+- **10**: Fully autonomous from start to finish. No blockers, no ambiguity, no manual steps.
+- **7–9**: Minor friction — one clarification needed, one unexpected error handled gracefully.
+- **4–6**: Moderate friction — multiple ambiguities, one blocker that paused execution, one manual intervention.
+- **1–3**: Significant friction — repeated blockers, unclear plan steps that caused wrong-direction work, multiple manual interventions.
+- **0**: Execution could not proceed without constant human steering.
+
+Evidence to cite: error log entries, any manual interventions recorded in PROGRESS.md, plan ambiguities encountered.
+
+---
+
+## Scoring Principles
+
+### Evidence-first scoring
+
+Every score needs a citation. "Tasks passed audit" is not a citation. "12/14 tasks passed audit on first attempt; Tasks 3 and 9 each needed one auto-fix cycle" is a citation. If the data to score a dimension is genuinely unavailable, note the gap explicitly and score conservatively (assume worst case for that dimension).
+
+### No vibes
+
+Do not write "the execution felt smooth" or "agents seemed efficient." Write "0 manual interventions were required and all errors were caught and auto-resolved by ftm-audit Phase 2." The report is read by future executions that need to calibrate behavior, not by humans looking for encouragement.

package/ftm-retro/references/templates/REPORT-FORMAT.md

@@ -0,0 +1,109 @@
+# Report Format — Retro Output Template
+
+This is the exact format for all retro report files saved to `~/.claude/ftm-retros/`.
+
+---
+
+## File Naming
+
+Save to: `~/.claude/ftm-retros/{plan-slug}-{YYYY-MM-DD}.md`
+
+### Slug Generation
+
+Take the plan title, lowercase it, replace spaces with hyphens, strip all non-alphanumeric characters except hyphens.
+
+Examples:
+- "FTM Ecosystem Expansion" → `ftm-ecosystem-expansion`
+- "Fix Auth Bug + Rate Limiting" → `fix-auth-bug-rate-limiting`
+- "v2.0 API Refactor" → `v20-api-refactor`
+
+---
+
+## Report Template
+
+```markdown
+# Retro: {Plan Title}
+
+**Date:** {YYYY-MM-DD}
+**Plan:** {absolute path to plan file}
+**Duration:** {total execution time, e.g. "47 minutes"}
+
+## Scores
+
+| Dimension | Score | Notes |
+|-----------|-------|-------|
+| Wave Parallelism | X/10 | {1-sentence justification with data} |
+| Audit Pass Rate | X/10 | {N}/{total} tasks first-pass |
+| Codex Gate Pass Rate | X/10 | {N}/{total} waves first-pass |
+| Retry/Fix Count | X/10 | {total retries} across {N} tasks |
+| Execution Smoothness | X/10 | {1-sentence justification} |
+
+**Overall: {sum}/50**
+
+## Raw Data
+
+- Tasks: {N}
+- Waves: {N}
+- Agents spawned: {N}
+- Audit findings: {N} total ({N} auto-fixed, {N} manual)
+- Codex gate results: Wave 1: pass | Wave 2: fail → pass | Wave 3: pass
+- Errors/blockers: {list any, or "none"}
+
+## What Went Well
+
+{2–4 specific observations, each grounded in a data point or task number.}
+
+Example format:
+- **Task 3 (auth middleware)** completed in a single commit with zero audit findings. The agent prompt had clear acceptance criteria and a scoped file list — the agent never wandered.
+- **Wave 2 parallelism** was fully utilized: all 4 tasks dispatched simultaneously, cutting estimated serial time from ~32 minutes to ~9 minutes.
+
+## What Was Slow
+
+{2–4 specific bottlenecks with timing data or retry counts where available.}
+
+Example format:
+- **ftm-audit Phase 1 (knip)** repeated full project analysis for each task in wave 3, even though tasks only touched 2–3 files each. Added ~40s × 5 tasks = ~3.5 minutes of unnecessary scanning.
+- **Task 7 needed 3 audit fix cycles** due to an import path that kept regenerating incorrectly. The agent prompt did not specify the alias configuration in tsconfig.paths.
+
+## Proposed Improvements
+
+{3–5 specific, actionable suggestions. Each must identify: which skill to change, what to change exactly, and why it would help.}
+
+Format each as:
+**N. {Short title}** — {Skill to change} — {Specific change} — {Expected impact}
+
+Examples:
+1. **Cache knip results within a wave** — ftm-audit — In Phase 1, check whether knip results are already cached for the current wave (via a temp file at `/tmp/ftm-knip-cache-{wave-id}.json`). Only re-run knip if the cache is missing or if the files changed by this task differ from cached scope. Expected: 3× speedup for ftm-audit on large projects with many tasks per wave.
+2. **Dispatch Instrumentor and Researcher in parallel** — ftm-debug — These two agents have no shared state and currently run sequentially. Dispatch them simultaneously. Expected: ~40% reduction in ftm-debug total runtime.
+3. **Add tsconfig.paths to agent context for TypeScript projects** — ftm-executor — When generating agent prompts for TypeScript tasks, include the relevant `paths` aliases from `tsconfig.json`. Expected: eliminates the import-alias regeneration loop that caused 3 retries on Task 7.
+
+## Pattern Analysis
+
+{Only include this section if past retros exist in ~/.claude/ftm-retros/}
+
+### Recurring Issues
+
+{List problems that appeared in 2 or more retros. Format: "Issue description — appeared in: retro-slug-1, retro-slug-2"}
+
+### Score Trends
+
+{Compare overall scores across retros. Are they improving, declining, or stable? Cite actual numbers.}
+
+Example: Overall scores: 32/50 → 38/50 → 41/50 across the last 3 retros. Parallelism and smoothness improving; audit pass rate stuck at 6/10 for all three runs.
+
+### Unaddressed Suggestions
+
+{List proposed improvements from past retros that have not yet been implemented. These get escalated — flag them explicitly.}
+
+Format: "**[ESCALATED]** {suggestion} — first proposed in {retro-slug-date}, appeared {N} times"
+```
+
+---
+
+## Improvement Specificity Standard
+
+"Improve parallelism" is not an improvement proposal. "Add a dependency pre-check step to ftm-executor Phase 2 that flags tasks with no declared dependencies as parallelizable, and warn when they are dispatched serially" is an improvement proposal. Every proposed improvement must be concrete enough that a future session could implement it from the description alone without asking clarifying questions.
+
+## Pattern Escalation Standard
+
+Recurring issues that have appeared in 3+ retros without being addressed should be flagged with `[ESCALATED - 3+ occurrences]` and moved to the top of the Proposed Improvements list. These are systemic problems, not one-off noise.

package/ftm-retro.yml

@@ -0,0 +1,2 @@
+name: ftm-retro
+description: Post-execution self-assessment skill. Automatically triggered after ftm-executor completes a plan. Scores execution across 5 dimensions, identifies what went well and what was slow, writes structured report with improvement suggestions. Use when user says "retro", "retrospective", "how did that go", "execution review", "self-assessment", "ftm retro".

package/ftm-routine.yml

@@ -0,0 +1,4 @@
+---
+name: ftm-routine
+description: Execute named, recurring multi-step workflows from YAML definitions. Use when user says "routine", "run routine", "ftm-routine", "morning triage", or names a known routine. Routines are reusable workflows stored in ~/.ftm/routines/ that execute through the normal plan-approve flow.
+---

package/ftm-state/blackboard/context.json

@@ -0,0 +1,23 @@
+{
+  "current_task": {
+    "id": null,
+    "description": null,
+    "type": null,
+    "started_at": null,
+    "status": null
+  },
+  "recent_decisions": [],
+  "active_constraints": [],
+  "user_preferences": {
+    "communication_style": null,
+    "approval_gates": null,
+    "default_model_profile": null
+  },
+  "session_metadata": {
+    "started_at": null,
+    "last_updated": null,
+    "conversation_id": null,
+    "messages_count": 0,
+    "skills_invoked": []
+  }
+}

package/ftm-state/blackboard/experiences/index.json

@@ -0,0 +1,9 @@
+{
+  "entries": [],
+  "metadata": {
+    "total_count": 0,
+    "last_updated": null,
+    "max_entries": 200,
+    "pruning_strategy": "remove_oldest_low_confidence"
+  }
+}

package/ftm-state/blackboard/patterns.json

@@ -0,0 +1,6 @@
+{
+  "codebase_insights": [],
+  "execution_patterns": [],
+  "user_behavior": [],
+  "recurring_issues": []
+}

package/ftm-state/schemas/context.schema.json

@@ -0,0 +1,130 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "context.schema.json",
+  "title": "Blackboard Context",
+  "description": "Schema for the blackboard session context state file",
+  "type": "object",
+  "required": [
+    "current_task",
+    "recent_decisions",
+    "active_constraints",
+    "user_preferences",
+    "session_metadata"
+  ],
+  "additionalProperties": false,
+  "properties": {
+    "current_task": {
+      "type": "object",
+      "description": "The task currently being executed in this session",
+      "required": ["id", "description", "type", "started_at", "status"],
+      "additionalProperties": false,
+      "properties": {
+        "id": {
+          "type": ["string", "null"],
+          "description": "Unique identifier for the task"
+        },
+        "description": {
+          "type": ["string", "null"],
+          "description": "Human-readable description of the task"
+        },
+        "type": {
+          "type": ["string", "null"],
+          "description": "Category or type of task (e.g. feature, bugfix, refactor)"
+        },
+        "started_at": {
+          "type": ["string", "null"],
+          "format": "date-time",
+          "description": "ISO 8601 timestamp when the task was started"
+        },
+        "status": {
+          "type": ["string", "null"],
+          "enum": [null, "pending", "in_progress", "complete", "failed"],
+          "description": "Current execution status of the task"
+        }
+      }
+    },
+    "recent_decisions": {
+      "type": "array",
+      "description": "Log of decisions made during this session (capped at 10)",
+      "maxItems": 10,
+      "items": {
+        "type": "object",
+        "required": ["summary", "timestamp", "skill"],
+        "additionalProperties": false,
+        "properties": {
+          "summary": {
+            "type": "string",
+            "description": "Brief description of the decision made"
+          },
+          "timestamp": {
+            "type": "string",
+            "description": "When the decision was recorded"
+          },
+          "skill": {
+            "type": "string",
+            "description": "The skill that produced this decision"
+          }
+        }
+      }
+    },
+    "active_constraints": {
+      "type": "array",
+      "description": "List of constraints currently in effect for this session",
+      "items": {
+        "type": "string"
+      }
+    },
+    "user_preferences": {
+      "type": "object",
+      "description": "Preferences established or inferred during the session",
+      "required": ["communication_style", "approval_gates", "default_model_profile"],
+      "additionalProperties": false,
+      "properties": {
+        "communication_style": {
+          "type": ["string", "null"],
+          "description": "Preferred tone and verbosity level (e.g. concise, detailed)"
+        },
+        "approval_gates": {
+          "type": ["string", "null"],
+          "description": "Approval gate setting (e.g. always, on_destructive, never)"
+        },
+        "default_model_profile": {
+          "type": ["string", "null"],
+          "description": "Model profile to use by default when not overridden"
+        }
+      }
+    },
+    "session_metadata": {
+      "type": "object",
+      "description": "Metadata tracking the lifecycle and activity of this session",
+      "required": ["started_at", "last_updated", "conversation_id", "messages_count", "skills_invoked"],
+      "additionalProperties": false,
+      "properties": {
+        "started_at": {
+          "type": ["string", "null"],
+          "description": "ISO 8601 timestamp when the session began"
+        },
+        "last_updated": {
+          "type": ["string", "null"],
+          "description": "ISO 8601 timestamp of the most recent write to this file"
+        },
+        "conversation_id": {
+          "type": ["string", "null"],
+          "description": "Identifier for the conversation this session belongs to"
+        },
+        "messages_count": {
+          "type": "integer",
+          "minimum": 0,
+          "description": "Number of messages exchanged in the session so far"
+        },
+        "skills_invoked": {
+          "type": "array",
+          "description": "Ordered list of skill names invoked during this session",
+          "items": {
+            "type": "string"
+          }
+        }
+      }
+    }
+  }
+}

package/ftm-state/schemas/experience-index.schema.json

@@ -0,0 +1,77 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "experience-index.schema.json",
+  "title": "Blackboard Experience Index",
+  "description": "Schema for the experience index file that tracks all stored experience entries",
+  "type": "object",
+  "required": ["entries", "metadata"],
+  "additionalProperties": false,
+  "properties": {
+    "entries": {
+      "type": "array",
+      "description": "Index records pointing to individual experience files",
+      "items": {
+        "type": "object",
+        "required": ["id", "file", "task_type", "tags", "timestamp", "confidence"],
+        "additionalProperties": false,
+        "properties": {
+          "id": {
+            "type": "string",
+            "description": "Unique identifier for this experience entry"
+          },
+          "file": {
+            "type": "string",
+            "description": "Relative path to the experience JSON file (e.g. YYYY-MM-DD_task-slug.json)"
+          },
+          "task_type": {
+            "type": "string",
+            "description": "Category of task this experience was recorded from"
+          },
+          "tags": {
+            "type": "array",
+            "description": "Labels for filtering and retrieval",
+            "items": {
+              "type": "string"
+            }
+          },
+          "timestamp": {
+            "type": "string",
+            "description": "ISO 8601 timestamp when the experience was recorded"
+          },
+          "confidence": {
+            "type": "number",
+            "minimum": 0,
+            "maximum": 1,
+            "description": "Confidence score for this experience (0.0 = low, 1.0 = high)"
+          }
+        }
+      }
+    },
+    "metadata": {
+      "type": "object",
+      "description": "Administrative metadata about the experience index",
+      "required": ["total_count", "last_updated", "max_entries", "pruning_strategy"],
+      "additionalProperties": false,
+      "properties": {
+        "total_count": {
+          "type": "integer",
+          "minimum": 0,
+          "description": "Total number of experience entries in the index"
+        },
+        "last_updated": {
+          "type": ["string", "null"],
+          "description": "ISO 8601 timestamp of the most recent write to this index"
+        },
+        "max_entries": {
+          "type": "integer",
+          "default": 200,
+          "description": "Maximum number of entries to retain before pruning is triggered"
+        },
+        "pruning_strategy": {
+          "type": "string",
+          "description": "Strategy used to remove entries when max_entries is exceeded (e.g. remove_oldest_low_confidence)"
+        }
+      }
+    }
+  }
+}