aether-colony 5.0.0 → 5.2.1

package/.aether/utils/midden.sh

@@ -0,0 +1,520 @@
+#!/bin/bash
+# Midden (failure tracking) utility functions — extracted from aether-utils.sh
+# Provides: _midden_write, _midden_recent_failures, _midden_review, _midden_acknowledge
+#
+# These functions are sourced by aether-utils.sh at startup.
+# All shared infrastructure (json_ok, json_err, atomic_write, acquire_lock,
+# release_lock, LOCK_DIR, DATA_DIR, SCRIPT_DIR, error constants) is available.
+
+_midden_try_write() {
+    # Helper: write updated JSON to midden file with retry
+    # Usage: _midden_try_write <updated_json> <midden_file>
+    # Returns: 0 on success, 1 on failure
+    local mtw_json="$1"
+    local mtw_file="$2"
+    local mtw_tmp="${mtw_file}.tmp.$$"
+
+    if ! { printf '%s\n' "$mtw_json" > "$mtw_tmp" && mv "$mtw_tmp" "$mtw_file"; }; then
+      # Silent retry (once)
+      if ! { printf '%s\n' "$mtw_json" > "$mtw_tmp" && mv "$mtw_tmp" "$mtw_file"; }; then
+        echo "Warning: Midden write failed after retry -- entry may not have been saved." >&2
+        return 1
+      fi
+    fi
+    return 0
+}
+
+_midden_write() {
+    # Write a warning/observation to the midden for later review
+    # Usage: midden-write <category> <message> <source>
+    # Example: midden-write "security" "High CVEs found: 3" "gatekeeper"
+    # Returns: JSON with success status and entry details
+
+    mw_category="${1:-general}"
+    mw_message="${2:-}"
+    mw_source="${3:-unknown}"
+
+    # Graceful degradation: if no message, return success but note it
+    if [[ -z "$mw_message" ]]; then
+      json_ok "{\"success\":true,\"warning\":\"no_message_provided\",\"entry_id\":null}"
+      return 0
+    fi
+
+    mw_midden_dir="$COLONY_DATA_DIR/midden"
+    mw_midden_file="$mw_midden_dir/midden.json"
+    mw_timestamp=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
+    mw_entry_id="midden_$(date +%s)_$$"
+
+    # Create midden directory if it doesn't exist
+    mkdir -p "$mw_midden_dir"
+
+    # Initialize midden.json if it doesn't exist
+    if [[ ! -f "$mw_midden_file" ]]; then
+      printf '%s\n' '{"version":"1.0.0","entries":[]}' > "$mw_midden_file"
+    fi
+
+    # Create the new entry using jq for safe JSON construction
+    mw_new_entry=$(jq -n \
+      --arg id "$mw_entry_id" \
+      --arg ts "$mw_timestamp" \
+      --arg cat "$mw_category" \
+      --arg src "$mw_source" \
+      --arg msg "$mw_message" \
+      '{id: $id, timestamp: $ts, category: $cat, source: $src, message: $msg, reviewed: false}')
+
+    # Append to midden.json using jq with locking
+    if acquire_lock "$mw_midden_file" 2>/dev/null; then
+      mw_updated_midden=$(jq --argjson entry "$mw_new_entry" '
+        .entries += [$entry] |
+        .entry_count = (.entries | length)
+      ' "$mw_midden_file" 2>/dev/null)
+
+      if [[ -n "$mw_updated_midden" ]]; then
+        _midden_try_write "$mw_updated_midden" "$mw_midden_file"
+        release_lock 2>/dev/null || true
+        mw_total=$(jq '.entries | length' "$mw_midden_file" 2>/dev/null || echo 0)
+        json_ok "$(jq -n --arg entry_id "$mw_entry_id" --arg category "$mw_category" --argjson midden_total "$mw_total" \
+          '{success: true, entry_id: $entry_id, category: $category, midden_total: $midden_total}')"
+      else
+        release_lock 2>/dev/null || true
+        json_ok "{\"success\":true,\"warning\":\"jq_processing_failed\",\"entry_id\":null}"
+      fi
+    else
+      # Lock failed — graceful degradation, try without lock
+      echo "Warning: Midden write completed without lock -- if another write happened at the same time, one entry may be missing." >&2
+      mw_updated_midden=$(jq --argjson entry "$mw_new_entry" '
+        .entries += [$entry] |
+        .entry_count = (.entries | length)
+      ' "$mw_midden_file" 2>/dev/null)
+
+      if [[ -n "$mw_updated_midden" ]]; then
+        _midden_try_write "$mw_updated_midden" "$mw_midden_file"
+        json_ok "$(jq -n --arg entry_id "$mw_entry_id" --arg category "$mw_category" \
+          '{success: true, entry_id: $entry_id, category: $category, warning: "lock_unavailable"}')"
+      else
+        json_ok "{\"success\":true,\"warning\":\"jq_processing_failed\",\"entry_id\":null}"
+      fi
+    fi
+}
+
+_midden_recent_failures() {
+    # Extract recent failure entries from midden.json
+    # Usage: midden-recent-failures [limit]
+    # Returns: JSON with count and failures array
+
+    limit="${1:-5}"
+    midden_file="$COLONY_DATA_DIR/midden/midden.json"
+
+    if [[ ! -f "$midden_file" ]]; then
+      echo '{"count":0,"failures":[]}'
+      return 0
+    fi
+
+    # Extract failures from .entries[], sort by timestamp descending, limit results
+    result=$(jq --argjson limit "$limit" '{
+      "count": ([.entries[]?] | length),
+      "failures": ([.entries[]?] | sort_by(.timestamp) | reverse | .[:$limit] | [.[] | {timestamp, category, source, message}])
+    }' "$midden_file" 2>/dev/null)
+
+    if [[ -z "$result" ]]; then
+      echo '{"count":0,"failures":[]}'
+    else
+      echo "$result"
+    fi
+    return 0
+}
+
+_midden_review() {
+    # Review unacknowledged midden entries grouped by category
+    # Usage: midden-review [--category <cat>] [--limit N] [--include-acknowledged]
+    # Returns: JSON with unacknowledged_count, categories summary, and entries array
+
+    mr_category=""
+    mr_limit=20
+    mr_include_ack=false
+
+    while [[ $# -gt 0 ]]; do
+      case "$1" in
+        --category)             mr_category="${2:-}"; shift 2 ;;
+        --limit)                mr_limit="${2:-20}"; shift 2 ;;
+        --include-acknowledged) mr_include_ack=true; shift ;;
+        *) shift ;;
+      esac
+    done
+
+    mr_midden_file="$COLONY_DATA_DIR/midden/midden.json"
+
+    if [[ ! -f "$mr_midden_file" ]]; then
+      json_ok '{"unacknowledged_count":0,"categories":{},"entries":[]}'
+      return 0
+    fi
+
+    # Build jq filter based on options
+    mr_result=$(jq \
+      --arg category "$mr_category" \
+      --argjson limit "$mr_limit" \
+      --argjson include_ack "$mr_include_ack" \
+      '
+      # Start with all entries
+      [.entries // [] | .[] |
+        # Filter acknowledged unless --include-acknowledged
+        if $include_ack then . else select(.acknowledged != true) end |
+        # Filter by category if specified
+        if ($category | length) > 0 then select(.category == $category) else . end
+      ] |
+      # Sort by timestamp descending
+      sort_by(.timestamp) | reverse |
+      # Compute categories before limiting
+      . as $all |
+      # Apply limit
+      ($all | .[:$limit]) as $limited |
+      # Group $all by category for counts
+      ($all | group_by(.category) | map({key: .[0].category, value: length}) | from_entries) as $cats |
+      {
+        unacknowledged_count: ($all | length),
+        categories: $cats,
+        entries: $limited
+      }
+      ' "$mr_midden_file" 2>/dev/null)
+
+    if [[ -z "$mr_result" ]]; then
+      json_ok '{"unacknowledged_count":0,"categories":{},"entries":[]}'
+    else
+      json_ok "$mr_result"
+    fi
+    return 0
+}
+
+_midden_ingest_errors() {
+    # Ingest entries from errors.log into midden
+    # Usage: midden-ingest-errors [--dry-run]
+    # Returns: JSON with count of ingested entries
+    # After ingestion, moves errors.log to errors.log.ingested
+
+    mie_dry_run=false
+    while [[ $# -gt 0 ]]; do
+      case "$1" in
+        --dry-run) mie_dry_run=true; shift ;;
+        *) shift ;;
+      esac
+    done
+
+    mie_errors_file="$COLONY_DATA_DIR/errors.log"
+
+    # No errors.log → nothing to ingest
+    if [[ ! -f "$mie_errors_file" ]]; then
+      json_ok '{"ingested":0}'
+      return 0
+    fi
+
+    # Empty file → nothing to ingest
+    if [[ ! -s "$mie_errors_file" ]]; then
+      json_ok '{"ingested":0}'
+      return 0
+    fi
+
+    mie_count=0
+
+    # Read line by line (avoid pipe-to-while subshell)
+    while IFS= read -r mie_line; do
+      # Skip blank lines
+      [[ -z "$mie_line" ]] && continue
+
+      # Parse timestamp from [YYYY-...Z] prefix
+      mie_timestamp=""
+      mie_message="$mie_line"
+      if [[ "$mie_line" =~ ^\[([^\]]+)\]\ (.*) ]]; then
+        mie_timestamp="${BASH_REMATCH[1]}"
+        mie_message="${BASH_REMATCH[2]}"
+      fi
+
+      mie_count=$((mie_count + 1))
+
+      if [[ "$mie_dry_run" == "false" ]]; then
+        _midden_write "error_log" "$mie_message" "error-handler" >/dev/null 2>&1 || true
+      fi
+    done < "$mie_errors_file"
+
+    # Move the file (not dry-run only)
+    if [[ "$mie_dry_run" == "false" && "$mie_count" -gt 0 ]]; then
+      mv "$mie_errors_file" "${mie_errors_file}.ingested"
+    fi
+
+    json_ok "{\"ingested\":$mie_count}"
+    return 0
+}
+
+_midden_search() {
+    # Search midden entries by keyword match in message field
+    # Usage: midden-search <query> [--category <cat>] [--source <src>] [--limit N] [--include-acknowledged]
+    # Returns: JSON with query, match_count, and entries array
+
+    ms_query=""
+    ms_category=""
+    ms_source=""
+    ms_limit=10
+    ms_include_ack=false
+
+    # First positional arg is the query
+    if [[ $# -gt 0 && "$1" != --* ]]; then
+      ms_query="$1"
+      shift
+    fi
+
+    while [[ $# -gt 0 ]]; do
+      case "$1" in
+        --category)             ms_category="${2:-}"; shift 2 ;;
+        --source)               ms_source="${2:-}"; shift 2 ;;
+        --limit)                ms_limit="${2:-10}"; shift 2 ;;
+        --include-acknowledged) ms_include_ack=true; shift ;;
+        *) shift ;;
+      esac
+    done
+
+    ms_midden_file="$COLONY_DATA_DIR/midden/midden.json"
+
+    if [[ ! -f "$ms_midden_file" ]]; then
+      json_ok "{\"query\":$(printf '%s' "$ms_query" | jq -Rs .),\"match_count\":0,\"entries\":[]}"
+      return 0
+    fi
+
+    ms_result=$(jq \
+      --arg query "$ms_query" \
+      --arg category "$ms_category" \
+      --arg source "$ms_source" \
+      --argjson limit "$ms_limit" \
+      --argjson include_ack "$ms_include_ack" \
+      '
+      [.entries // [] | .[] |
+        # Filter acknowledged unless --include-acknowledged
+        if $include_ack then . else select(.acknowledged != true) end |
+        # Filter by category if specified
+        if ($category | length) > 0 then select(.category == $category) else . end |
+        # Filter by source if specified
+        if ($source | length) > 0 then select(.source == $source) else . end |
+        # Filter by keyword match in message (case-insensitive)
+        if ($query | length) > 0 then
+          select(.message | ascii_downcase | contains($query | ascii_downcase))
+        else
+          .
+        end
+      ] |
+      sort_by(.timestamp) | reverse |
+      . as $all |
+      {
+        query: $query,
+        match_count: ($all | length),
+        entries: ($all | .[:$limit])
+      }
+      ' "$ms_midden_file" 2>/dev/null)
+
+    if [[ -z "$ms_result" ]]; then
+      json_ok "{\"query\":$(printf '%s' "$ms_query" | jq -Rs .),\"match_count\":0,\"entries\":[]}"
+    else
+      json_ok "$ms_result"
+    fi
+    return 0
+}
+
+_midden_tag() {
+    # Add or remove a tag from a midden entry's tags array
+    # Usage: midden-tag --id <entry_id> --tag <tag_name>
+    #    OR: midden-tag --id <entry_id> --untag <tag_name>
+    # Returns: JSON with entry_id, tags array, and action
+
+    mt_id=""
+    mt_tag=""
+    mt_untag=""
+
+    while [[ $# -gt 0 ]]; do
+      case "$1" in
+        --id)    mt_id="${2:-}"; shift 2 ;;
+        --tag)   mt_tag="${2:-}"; shift 2 ;;
+        --untag) mt_untag="${2:-}"; shift 2 ;;
+        *) shift ;;
+      esac
+    done
+
+    # Validate: need --id
+    if [[ -z "$mt_id" ]]; then
+      json_err "$E_VALIDATION_FAILED" "midden-tag requires --id"
+    fi
+
+    # Validate: need --tag or --untag (but not both)
+    if [[ -z "$mt_tag" && -z "$mt_untag" ]]; then
+      json_err "$E_VALIDATION_FAILED" "midden-tag requires --tag or --untag"
+    fi
+
+    if [[ -n "$mt_tag" && -n "$mt_untag" ]]; then
+      json_err "$E_VALIDATION_FAILED" "midden-tag requires --tag or --untag, not both"
+    fi
+
+    mt_midden_file="$COLONY_DATA_DIR/midden/midden.json"
+
+    if [[ ! -f "$mt_midden_file" ]]; then
+      json_err "$E_FILE_NOT_FOUND" "midden.json not found"
+    fi
+
+    # Check entry exists
+    mt_exists=$(jq --arg id "$mt_id" '[.entries[]? | select(.id == $id)] | length > 0' "$mt_midden_file" 2>/dev/null || echo "false")
+    if [[ "$mt_exists" != "true" ]]; then
+      json_err "$E_RESOURCE_NOT_FOUND" "Midden entry '$mt_id' not found"
+    fi
+
+    # Acquire lock with trap-based cleanup
+    acquire_lock "$mt_midden_file" || {
+      json_err "$E_LOCK_FAILED" "Failed to acquire lock on midden.json"
+    }
+    trap 'release_lock 2>/dev/null || true' EXIT
+
+    if [[ -n "$mt_tag" ]]; then
+      # Add tag — create tags array if absent, append if tag not already present
+      mt_updated=$(jq \
+        --arg id "$mt_id" \
+        --arg tag "$mt_tag" \
+        '
+        .entries = [.entries[] |
+          if .id == $id then
+            . + {tags: ((.tags // []) | if contains([$tag]) then . else . + [$tag] end)}
+          else
+            .
+          end
+        ]
+        ' "$mt_midden_file" 2>/dev/null)
+      mt_action="added"
+    else
+      # Remove tag — remove from tags array if present
+      mt_updated=$(jq \
+        --arg id "$mt_id" \
+        --arg tag "$mt_untag" \
+        '
+        .entries = [.entries[] |
+          if .id == $id then
+            . + {tags: ((.tags // []) | map(select(. != $tag)))}
+          else
+            .
+          end
+        ]
+        ' "$mt_midden_file" 2>/dev/null)
+      mt_action="removed"
+      mt_tag="$mt_untag"
+    fi
+
+    if [[ -z "$mt_updated" ]]; then
+      trap - EXIT
+      release_lock 2>/dev/null || true
+      json_err "$E_INTERNAL" "Failed to update midden.json"
+    fi
+
+    atomic_write "$mt_midden_file" "$mt_updated"
+
+    trap - EXIT
+    release_lock 2>/dev/null || true
+
+    # Read back the updated tags for the entry
+    mt_tags=$(jq --arg id "$mt_id" '[.entries[]? | select(.id == $id) | .tags // []] | .[0] // []' "$mt_midden_file" 2>/dev/null || echo "[]")
+
+    json_ok "$(jq -n \
+      --arg entry_id "$mt_id" \
+      --argjson tags "$mt_tags" \
+      --arg action "$mt_action" \
+      '{entry_id: $entry_id, tags: $tags, action: $action}')"
+    return 0
+}
+
+_midden_acknowledge() {
+    # Acknowledge midden entries by id or by category
+    # Usage: midden-acknowledge --id <entry_id> [--reason <reason>]
+    #    OR: midden-acknowledge --category <cat> --reason <reason>
+    # Returns: JSON with acknowledged=true, count, and reason
+
+    ma_id=""
+    ma_category=""
+    ma_reason=""
+
+    while [[ $# -gt 0 ]]; do
+      case "$1" in
+        --id)       ma_id="${2:-}"; shift 2 ;;
+        --category) ma_category="${2:-}"; shift 2 ;;
+        --reason)   ma_reason="${2:-}"; shift 2 ;;
+        *) shift ;;
+      esac
+    done
+
+    # Validate: need either --id or --category
+    if [[ -z "$ma_id" && -z "$ma_category" ]]; then
+      json_err "$E_VALIDATION_FAILED" "midden-acknowledge requires --id or --category"
+    fi
+
+    ma_midden_file="$COLONY_DATA_DIR/midden/midden.json"
+
+    if [[ ! -f "$ma_midden_file" ]]; then
+      json_err "$E_FILE_NOT_FOUND" "midden.json not found"
+    fi
+
+    ma_now=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
+
+    # Acquire lock with trap-based cleanup
+    acquire_lock "$ma_midden_file" || {
+      json_err "$E_LOCK_FAILED" "Failed to acquire lock on midden.json"
+    }
+    trap 'release_lock 2>/dev/null || true' EXIT
+
+    if [[ -n "$ma_id" ]]; then
+      # Acknowledge single entry by id
+      ma_exists=$(jq --arg id "$ma_id" '[.entries[]? | select(.id == $id)] | length > 0' "$ma_midden_file" 2>/dev/null || echo "false")
+      if [[ "$ma_exists" != "true" ]]; then
+        trap - EXIT
+        release_lock 2>/dev/null || true
+        json_err "$E_RESOURCE_NOT_FOUND" "Midden entry '$ma_id' not found"
+      fi
+
+      ma_updated=$(jq \
+        --arg id "$ma_id" \
+        --arg now "$ma_now" \
+        --arg reason "$ma_reason" \
+        '
+        .entries = [.entries[] |
+          if .id == $id then
+            . + {acknowledged: true, acknowledged_at: $now, acknowledge_reason: $reason}
+          else
+            .
+          end
+        ]
+        ' "$ma_midden_file" 2>/dev/null)
+
+      ma_count=1
+    else
+      # Acknowledge all entries matching category
+      ma_count=$(jq --arg cat "$ma_category" '[.entries[]? | select(.category == $cat and .acknowledged != true)] | length' "$ma_midden_file" 2>/dev/null || echo "0")
+
+      ma_updated=$(jq \
+        --arg cat "$ma_category" \
+        --arg now "$ma_now" \
+        --arg reason "$ma_reason" \
+        '
+        .entries = [.entries[] |
+          if .category == $cat and .acknowledged != true then
+            . + {acknowledged: true, acknowledged_at: $now, acknowledge_reason: $reason}
+          else
+            .
+          end
+        ]
+        ' "$ma_midden_file" 2>/dev/null)
+    fi
+
+    if [[ -z "$ma_updated" ]]; then
+      trap - EXIT
+      release_lock 2>/dev/null || true
+      json_err "$E_INTERNAL" "Failed to update midden.json"
+    fi
+
+    atomic_write "$ma_midden_file" "$ma_updated"
+
+    trap - EXIT
+    release_lock 2>/dev/null || true
+
+    json_ok "$(jq -n --argjson count "$ma_count" --arg reason "$ma_reason" \
+      '{acknowledged: true, count: $count, reason: $reason}')"
+    return 0
+}

package/.aether/utils/oracle/oracle.md

@@ -0,0 +1,168 @@
+You are an **Oracle Ant** -- a deep research agent in the Aether Colony.
+
+## Mission
+
+Research a topic thoroughly. Each iteration targets knowledge gaps and deepens
+understanding. You are working through a structured research plan with tracked
+sub-questions, confidence scores, and accumulated findings.
+
+## Phase Directive
+
+Your current research phase and specific instructions were provided above this
+prompt. Follow them. The phase determines your strategy for this iteration.
+
+## Steering Signals
+
+If steering signals appear above this prompt, they were emitted by the user:
+
+- **REDIRECT** signals are HARD CONSTRAINTS. You MUST follow them. If a REDIRECT
+  conflicts with your planned approach, change your approach.
+- **FOCUS** signals indicate priority areas. When choosing your target question,
+  prefer questions related to focus areas. If no questions match, fall back to
+  default targeting (lowest-confidence).
+- **FEEDBACK** signals are gentle adjustments. Incorporate them where appropriate
+  into your research approach and output style.
+
+If no steering signals appear, follow default targeting as described in Instructions.
+
+## Instructions
+
+### Step 1: Read State Files
+
+Read these files to understand the current research state:
+
+- `.aether/oracle/state.json` -- Session metadata (topic, scope, iteration, phase, confidence)
+- `.aether/oracle/plan.json` -- Sub-questions with status, confidence, and iteration history
+- `.aether/oracle/gaps.md` -- Current knowledge gaps and contradictions
+- `.aether/oracle/synthesis.md` -- Accumulated findings organized by question
+
+Note the current `iteration` and `phase` from state.json. Your phase directive
+above tells you how to behave this iteration.
+
+### Step 2: Identify Target
+
+Select your target question based on the current phase:
+
+- **Survey phase:** Target questions with empty `iterations_touched` arrays first.
+  If all questions have been touched, target the lowest-confidence non-answered question.
+- **Investigate / Synthesize / Verify phases:** Target the lowest-confidence
+  non-answered question.
+
+If all questions are "answered", proceed to Step 5.
+
+### Step 3: Research
+
+**Before writing ANY finding:** READ existing findings for your target question
+in synthesis.md. Know what has already been discovered.
+
+**Your new findings MUST contain information NOT already in synthesis.md.** If you
+cannot find new information beyond what exists, write "No new findings beyond
+existing research" and target the next-lowest-confidence question instead.
+
+Acceptable new information includes:
+- Specific details not yet captured (numbers, dates, names)
+- Concrete examples or case studies
+- Source citations (URLs, documentation references, code paths)
+- Edge cases and limitations
+- Contradictions with existing findings
+
+Use available tools:
+- **Codebase:** Glob, Grep, Read for local files and source code
+- **Web:** WebSearch, WebFetch for external sources and documentation
+
+**Source Tracking (MANDATORY):**
+For every new finding, you MUST record:
+- The source URL (or file path for codebase research)
+- The source title/description
+- The date you accessed it
+
+Register sources in plan.json under the `sources` object using sequential IDs
+(S1, S2, S3...). Reuse existing source IDs if citing the same URL again.
+
+Source types: "documentation", "blog", "github", "academic", "codebase", "forum", "official"
+
+For codebase research, use file paths as URLs with type "codebase":
+```json
+"url": "src/components/Button.tsx",
+"title": "Button component source",
+"type": "codebase"
+```
+
+### Step 4: Update State Files
+
+After researching, update these files:
+
+**plan.json:** Update the target question:
+- Set `status` to "partial" (useful info but gaps remain) or "answered" (thoroughly addressed)
+- Update `confidence` (0-100) based on evidence quality -- see Confidence Scoring Rubric below
+- Add findings as OBJECTS (not strings): `{"text": "finding text", "source_ids": ["S1", "S2"], "iteration": <current>}`
+- Every finding MUST have at least one source_id
+- Add new sources to the top-level `sources` registry
+- Reuse existing source IDs for the same URL
+- Add current iteration number to `iterations_touched` array
+- If a question is IRRELEVANT to the topic, REMOVE it from the questions array entirely
+- Do NOT add new questions -- work through the original plan
+- Write the COMPLETE updated plan.json (not a partial update)
+
+**gaps.md:** Rewrite with current state:
+- List remaining open questions with confidence levels under "## Open Questions"
+- Note any contradictions discovered under "## Contradictions"
+- Update "## Last Updated" with current iteration number and timestamp
+
+**synthesis.md:** Update findings for the question you worked on:
+- Keep the "## Findings by Question" structure
+- Add new findings under the relevant question heading
+- Include question status and confidence in the heading
+- Do NOT duplicate existing findings -- add only new information
+- Do not remove findings from other questions
+
+**state.json:** Update:
+- `last_updated` to current ISO-8601 UTC timestamp
+- `overall_confidence` to the average of all remaining questions' confidence values
+- Do NOT change `iteration` or `phase` (oracle.sh manages these)
+
+### Step 5: Assess and Complete
+
+State your assessment: "Confidence: X% -- {brief reason}"
+
+If `overall_confidence` >= `target_confidence` (from state.json) OR all remaining
+questions are "answered": output `<oracle>COMPLETE</oracle>`
+
+Otherwise, end normally for another iteration.
+
+## Confidence Scoring Rubric
+
+Use this rubric when scoring question confidence. Anchor scores to evidence quality.
+
+| Score | Level | Criteria |
+|-------|-------|----------|
+| 0-20% | Unexplored | No research conducted on this question |
+| 20-40% | Surface level | General information only, no specific details or sources |
+| 40-60% | Partial understanding | Specific details from 1-2 sources, some gaps remain |
+| 60-80% | Good understanding | Multiple sources agree, edge cases identified and documented |
+| 80-95% | Thorough | Primary sources verified, contradictions resolved, limitations known |
+| 95-100% | Exhaustive | All reasonable angles explored, high-quality sources confirmed |
+
+**Do NOT inflate confidence.** One blog post = 30%, not 70%. A single source
+without corroboration caps at 50% regardless of detail level.
+
+**Do NOT deflate confidence to keep research going.** Score honestly based on
+the evidence you have. If the question is well-answered, say so.
+
+**Source-backed confidence rules:**
+- 0 sources: Finding is UNSUPPORTED -- do not record it
+- 1 source: Single-source claim, capped at 50% contribution to question confidence
+- 2+ sources: Multi-source claim, full confidence contribution
+- The overall question confidence should reflect the source backing of its findings
+
+## Important Rules
+
+- Target ONE question per iteration
+- Write COMPLETE JSON files, not partial updates (prevents corruption)
+- Do NOT add new sub-questions -- work through the original plan
+- Remove irrelevant questions entirely -- do not mark them as "skipped"
+- Reference existing findings BEFORE writing new ones -- no restatements
+- Do NOT modify any code files or colony state
+- Only write to `.aether/oracle/` directory
+- If this iteration is labeled "SYNTHESIS PASS" in the directive above, follow those instructions instead of the normal research flow -- consolidate and organize existing findings rather than researching new information
+- In synthesis passes, use inline citations [S1], [S2] linking findings to sources. Include a ## Sources section listing all sources with IDs, URLs, titles, and access dates.